text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Modelling of a Permanent Magnet Synchronous
Machine Using Isogeometric Analysis
Jacopo Corno
Prithvi Bhat, Herbert De Gersem Zeger Bontinck, Sebastian Schöps
arXiv:1708.02409v1 [cs.CE] 8 Aug 2017
Theorie Elektromagnetischer Felder
Technische Universität Darmstadt
Darmstadt, Germany
Theorie Elektromagnetischer Felder
Graduate School of CE
Technische Universität Darmstadt
Darmstadt, Germany
Abstract—Isogeometric Analysis (IGA) is used to simulate a
permanent magnet synchronous machine. IGA uses Non-Uniform
Rational B-splines (NURBS) to parametrise the domain and
to approximate the solution space, thus allowing for the exact
description of the geometries even on the coarsest level of
mesh refinement. Given the properties of the isogeometric basis
functions, this choice guarantees a higher accuracy than the
classical Finite Element Method (FEM).
For dealing with the different stator and rotor topologies,
the domain is split into two non-overlapping parts on which
Maxwell’s equations are solved independently in the context of
a classical Dirichlet-to-Neumann domain decomposition scheme.
The results show good agreement with the ones obtained by the
classical finite element approach.
Index Terms—Electromagnetic field simulation, Permanent
magnet machines, Numerical analysis, Finite element method,
Isogeometric analysis
I. I NTRODUCTION
Electric machines are usually modelled through the magnetoquasistatic approximation of Maxwell’s equations discretised
on the machine cross section, which requires the solution
of a 2D Poisson problem. Its numerical solution commonly
amounts to triangulating the domain and applying the FEM.
The acting electromotive forces and the torque can be determined by a post-processing procedure, e.g. invoking the
Maxwell stress tensor method. A drawback of the standard
approach is that a very fine mesh is needed to achieve
an acceptable accuracy. Moreover, the air gap needs to be
resolved properly and the solution is known to be extremely
sensitive to the used discretisation [1]. Furthermore, remeshing
or reconnecting mesh parts in order to account for the machine
rotation can introduce a spurious ripple on the solution for the
torque [2].
An exact parametrisation of the air gap is not possible
within a classical FEM framework since elements of any order
rely on polynomial mappings which are unable to represent
conic sections such as circles and ellipses. Furthermore, the
regularity of FEM solutions is typically limited by the C 0
continuity across the elements [3].
IGA [4] is able to overcome these issues. It chooses
Computer Aided Design (CAD) basis functions such as e.g.
B-splines and NURBS for the approximation spaces and is
thus able to represent CAD geometries exactly even on the
Theorie Elektromagnetischer Felder
Graduate School of CE
MOX Modeling and Scientific Computing
Politecnico di Milano, Italy
coarsest level of mesh refinement. A higher regularity at the
mesh interfaces can also be obtained by applying k-refinement,
which results in smoother basis functions compared to FEM
and increases the accuracy of the solution accordingly [5].
Another benefit of IGA is that it provides an elegant way to
do geometric optimisation [6] and to handle uncertainties [7].
This work aims at using IGA to model and simulate a
permanent magnet synchronous machine (PMSM). The stator
and rotor models parts are inevitably resolved by multiple
patches, which in turn necessitates the use of a domain
decomposition approach across the air gap. We employ a
classical iterative procedure based on a Dirichlet-to-Neumann
map [8].
In the following section, we introduce the 2D model electric
machine model and the IGA method used for its spatial discretisation. In section III, the iterative domain decomposition
scheme is explained and, finally, in section IV, we show the
results for the simulation of a PMSM.
II. S OLVING M AXWELL’ S E QUATIONS FOR A PMSM
A. Magnetostatics
In early design steps, it is sufficient to model electric
machines by the 2D magnetostatic approximation of the
Maxwell’s equations. Let Ω = Ωrt ∪Ωst (see Fig. 1) depict the
computational domain. One has to solve the following Poisson
equation
−∇ · (ν∇Az ) = Jsrc,z + Jpm,z ,
Az |Γd = 0,
(1)
Az |Γl = −Az |Γr ,
where ν = ν(x, y) is the reluctivity, Az = Az (x, y) is
the z-component of the magnetic vector potential, Jsrc,z =
Jsrc,z (x, y) is the source current density in the coils and
~ pm is the current
(0, 0, Jpm,z ) = (0, 0, Jpm,z (x, y)) = −∇ × H
~
~ pm (x, y) of
density according to the magnetisation Hpm = H
the permanent magnets. On Γd , a Dirichlet boundary condition
is applied and on Γl and Γr , an anti-periodic boundary
condition holds.
Applying the loading method [9] on the solution of (1)
gives us the spectrum of the electromotive force (EMF). The
first harmonic E1 is the EMF of the machines. The higher
Ω
Γd
1
Bi1 (ξ)
Ωst
0
Γl
Ωrt
Γag
Fig. 1: One pole of the PMSM. In red, the interface Γag
between the rotor and the stator is highlighted.
harmonics are used to calculate the total harmonic distortion
(THD), which is defined as
qP
∞
2
p=2 |Ep |
,
(2)
THD =
|E1 |
with p the harmonic order and Ep the EMF of order p.
B. Discretisation in the Isogeometric Framework
For the discretisation of Az , a linear combination of basis
functions wj = wj (x, y) is used, i.e.,
N
DoF
X
(3)
uj wj ,
j=1
where uj are the unknowns and NDoF depicts the number
of unknowns. Using the Ritz-Galerkin approach, a system of
equations is obtained, i.e.,
(4)
Kν u = jsrc + jpm ,
where Kν has the entries
Z
∂wi ∂wj
∂wi ∂wj
kν,ij =
ν
+ν
dΩ,
∂x ∂x
∂y ∂y
Ω
and the entries for the discretised current density are
Z
jsrc,i =
Jz wi dΩ,
(5a)
(5b)
Ω
Z
Ω
~ pm ·
H
2/3
1
0
1/3
2/3
1
Bi2 (ξ)
Γd
jpm,i =
1/3
1
Γr
Az =
0
−
∂wi ∂wi
,
∂y ∂x
dΩ.
(5c)
The choice of the basis functions depends on the method
one favours. The simplest case in the well established FEM is
the use of linear hat functions [10]. This paper proposes the
use of the IGA framework to model the electrical machine. In
this method, the basis functions are defined by NURBS [11].
Let p depict the degree of the basis functions and let
Ξ = ξ1 . . . ξn+p+1
(6)
be a vector that partitions [0, 1] into elements, where ξi ∈
Ω̂ = [0, 1]. Then, the Cox-de Boor’s formula [11] defines n
0
ξ
Fig. 2: B-spline basis functions of degree 1 and 2 on an open,
uniform knot vector Ξ = [0, 0, 0, 1/3, 2/3, 1, 1, 1].
B-spline basis functions {Bip } with i = 1, . . . , n. B-splines of
degree p = 1, 2 are shown in Fig. 2. NURBS of degree p are
then constructed as
ωi Bip
(7)
Nip = P
p,
j ωj Bj
with ωi a weighting parameter associated with the i-th basis
function. A NURBS curve is obtained by the mapping
n
X
Pi Nip ,
(8)
F=
i=1
with Pi control points in R3 and F : [0, 1] → R2 . From
the curves, surfaces can be constructed by using tensor products [11], i.e.,
p1 ,p2
Nij
= Nip1 Njp2 ,
i = 1, . . . , n1 ,
j = 1, . . . , n2 .
(9)
The number of one-dimensional basis functions along direction d = 1, 2 and their degrees are depicted by nd and
pd respectively. The mapping F is now re-defined, namely
F : Ω̂ → Ω, where Ω̂ = [0, 1]2 is the reference square and
Ω ∈ R3 is the physical domain. Due to the NURBS mapping,
the subdivisions constructed by the vectors Ξd are transformed
to a physical mesh on Ω.
Due to the complexity of the geometry, the cross section of
the machine cannot be represented by a single map F [12].
Multiple patches Ωk are constructed (Fig. 1), each one of them
defined as the image of the unit square through a parametrization Fk of the type (8) in such a way that ∪k Ωk = Ω and
Ωi ∩ Ωj = ∅ [12].
Using (7) and (8) enables an exact parametrization of e.g.
circles and arcs. This implies that the geometry of the machine
can be modelled exactly, which is particular beneficial in the
air gap region. In IGA, the same set of basis functions (9) is
also used to approximate the solution of (1). Hence, we can
rewrite (3) as
N
DoF
X
Az =
uj Nj ,
(10)
j=1
III. I TERATIVE S TATOR -ROTOR C OUPLING
Since the rotor and the stator have a different topology,
we parametrise them independently as multipatch NURBS
entities with non-conforming patches at the air gap interface
Γag . The idea is to follow a classical non-overlapping domain
decomposition approach based on a Dirichlet-to-Neumann
map (see e.g. [8]).
Let us consider the circular arc Γag in the air gap and split
the domain Ω such that Ωrt ∪ Ωst = Ω and Ωrt ∩ Ωst = Γag (see
Fig. 1). Given k = 1 and an initialisation λ0 , we solve:
−∇ · ν∇Ak+1
= Jz ,
z,rt
k+1
Az,rt |Γd = 0,
(11)
Ak+1
= −Ak+1
z,rt |Γl
z,rt |Γr
Ak+1
= λk
z,rt |Γag
−∇ · ν∇Ak+1
= Jz ,
z,st
k+1
Az,st |Γd = 0,
(12)
Ak+1
= −Ak+1
z,st |Γl
z,st |Γr
ν∇Ak+1
nag = ν∇Ak+1
nag ,
z,st |Γag · ~
z,rt |Γag · ~
where ~nst is a unit vector perpendicular to the air gap interface.
The two problems are solved iteratively with the update
k
λk+1 = αAk+1
z,st + (1 − α)λ ,
(13)
where α ∈ [0, 1] is a relaxation parameter (in general, the
method is not ensured to converge if α = 1 [8]).
As a stopping criterion for the method, the L2 error between
two successive iterations is required to be below a specified
tolerance both in the rotor and the stator, i.e.,
k
εrt = Ak+1
z,rt − Az,rt
εst =
Ak+1
z,st
−
L2 (Ωrt )
Akz,st L2 (Ω )
st
/ Ak+1
z,rt
< tol,
/
< tol.
L2 (Ωrt )
k+1
Az,st L2 (Ω )
st
The discretisation of problems (11)-(12) is straightforward
in the IGA framework presented above.
IV. R ESULTS
For testing the suitability of IGA for machine simulation,
we first consider the domain Ωrt and we solve a test problem
with homogeneous Dirichlet boundary conditions on Γd ∪ Γag
and anti-periodic boundary conditions connecting Γl and Γr .
The only source present in the model is the magnetisation of
the permanent magnets. The IGA solution for the magnetic
vector potential is compared to the one obtained using a
fine FEM discretisation with first order elements. We solve
IGA basis functions of degree 1 and 2 and for an increasing
mesh refinement, and we depict the difference to the FEM
10−4
p=1
p=2
kAz,IGA − Az,FEM k0
where we have removed the indices p1 , p2 for simplicity.
We will always assume the same polynomial degree in both
directions.
This choice of basis functions guarantees a higher regularity
of the solution across the elements. The continuity at the patch
boundaries reverts to the FEM case since only C 0 continuity
is imposed (see [12]).
10−5
102
103
NDoF
104
Fig. 3: Poisson problem for the magnetic vector potential
on Ωrt : L2 error between the IGA solution and a fine FEM
solution for increasing mesh refinement.
TABLE I: Numerical results for the EMF and the THD.
IGA (degree 2)
FEM
relative difference
NDoF
3224
59678
-
E1
30.6 V
29.8 V
2.5 %
THD
6.06 · 10−2 %
5.72 · 10−2 %
5.7 %
solution in L2 in Fig. 3. For both discretisation degrees, the
curves approach a fixed value, which is expected since the
IGA method is solving the problem for the exact geometry in
contrast to FEM which introduces an additional geometrical
error related to the triangulation of the geometry.
Secondly, the full problem is considered, incorporating
the Dirichlet-to-Neumann coupling between stator and rotor
introduced in the previous section. In Fig. 4, the convergence
of the algorithm is shown for a simulation with degree 2,
with approximately 3200 total degrees of freedom. After 29
iterations, both solutions show an incremental error below the
prescribed tolerance tol = 10−7 .
The simulation results are shown in Fig. 5, where the flux
lines in the machine are shown. In the post-processing the
spectrum of the EMF has been calculated. In Fig. 6, the first 32
modes of the spectrum of the EMF of the machine are shown.
Both methods result in similar spectra. This is also depicted
in Tab. I, where it is shown that the relative difference of the
THD is below 6%.
V. C ONCLUSIONS
Isogeometric Analysis (IGA) recently emerged as a promising alternative to the Finite Element Method (FEM) for highly
accurate electromagnetic field simulation. IGA is capable of
exactly resolving circles and, hence, avoids any geometric
approximations errors in contrast to FEM, which is of particular importance in the air gap region. Our IGA model for
εrt
εst
100
ε
EMF (V)
10−2
10−4
FEM
IGA
30
20
10
10−6
0
10−8
5
10
15
20
25
30
iteration
Fig. 4: Convergence of the Dirichlet-to-Neumann coupling for
the PMSM. Relative difference in L2 norm for the magnetic
vector potential between successive iterations on Ωrt (in blue)
and Ωst (in orange).
0
5
10
15
20
harmonic order
25
30
Fig. 6: Spectrum with the first 32 modes of the EMF of the
machine.
(grant nr. SCHO1562/3-1) and by the ’Excellence Initiative’ of
the German Federal and State Governments and the Graduate
School of CE at TU Darmstadt.
R EFERENCES
Fig. 5: The distribution of the flux lines obtained by modelling
with FEM.
a permanent magnet synchronous machine consists of two
separate multipatch IGA discretisations for stator and rotor,
glued together with a Dirichlet-to-Neumann map in the air gap.
IGA (degree 2) attains the prescribed accuracy with a number
of degrees of freedom which is substantially smaller (≈ 20
times) than for lowest order FEM. Therefore, IGA is a valuable
extension to an engineer’s simulation toolbox, especially when
fast and accurate machine simulation tasks are due.
ACKNOWLEDGMENT
This work is supported by the German BMBF in the context
of the SIMUROM project (grant nr. 05M2013), by the DFG
[1] D. Howe and Z.Q. Zhu, “The influence of finite element discretisation on
the prediction of cogging torque in permanent magnet excited motors”,
IEEE Trans. Magn., vol. 28, no.2, pp. 1080–1083, 1992.
[2] I. Tsukerman, “Accurate computation of ‘ripple solutions’ on moving
finite element meshes”, IEEE Trans. Magn., vol. 31, no. 3, pp. 1472–
1475, 1995.
[3] T. Tärnhuvud and K. Reichert, “Accuracy problems of force and torque
calculation in FE-systems”, IEEE Trans. Magn., vol. 24, no. 1, pp. 443–
446, 1988.
[4] T. J. R. Hughes, J.A. Cottrell and Y. Bazilevs, “Isogeometric analysis:
CAD, finite elements, NURBS, exact geometry and mesh refinement”,
Comput. Meth. Appl. Mech. Eng., vol. 194, pp. 4135–4195, 2005.
[5] J. A. Evans, Y. Bazilevs, I. Babuška and T. J. R. Hughes,“n-Widths, supinfs, and optimality ratios for the k-version of the isogeometric finite
element method”, Comput. Meth. Appl. Mech. Eng., vol. 198, no. 21–26,
pp. 1726–1741, 2009.
[6] A. Pels, Z. Bontinck, J. Corno, H. De Gersem and S. Schöps, “Optimization of a Stern-Gerlach Magnet by magnetic field-circuit coupling
and isogeometric analysis,” IEEE Trans. Magn., vol. 51, no. 12, pp. 1–7,
2015.
[7] J. Corno, C. de Falco, H. De Gersem and S. Schöps,“Isogeometric Analysis simulation of TESLA cavities under uncertainty,” Electromagnetics
in Advanced Applications (ICEAA), 2015 International Conference on.
IEEE, 2015.
[8] A. Quarteroni and A. Valli, “Domain decomposition methods for partial
differential equations”, Oxford University Press, 1999.
[9] M.A. Rahman and P. Zhou, “Determination of saturated parameters of
PM motors using loading magnetic fields,” IEEE Trans. Magn., vol. 27,
no. 5, pp. 3947–3950, 1991.
[10] S.J. Salon, Finite element analysis of electrical machines, Boston USA:
Kluwer academic publishers, 1995.
[11] L. Piegl and W. Tiller. “The NURBS book”, Monographs in Visual
Communication, 1997.
[12] J.A. Cottrell, T.J.R. Hughes and Y. Bazilevs, “Isogeometric Analysis:
Toward integration of CAD and FEA”, John Wiley & Sons, Ltd, 2009.
| 5cs.CE
|
Reduced Dimensional Optimal Vector Linear Index
Codes for Index Coding Problems with Symmetric
Neighboring and Consecutive Side-information
arXiv:1801.00406v1 [cs.IT] 1 Jan 2018
Mahesh Babu Vaddi and B. Sundar Rajan
Department of Electrical Communication Engineering, Indian Institute of Science, Bengaluru 560012, KA, India
E-mail: {vaddi, bsrajan}@iisc.ac.in
Abstract—A single unicast index coding problem (SUICP)
with symmetric neighboring and consecutive side-information
(SNCS) has K messages and K receivers, the kth receiver Rk
wanting the kth message xk and having the side-information
Kk = {xk−U , . . . , xk−2 , xk−1 } ∪ {xk+1 , xk+2 , . . . , xk+D }. The
single unicast index coding problem with symmetric neighboring
and consecutive side-information, SUICP(SNCS), is motivated by
topological interference management problems in wireless communication networks. Maleki, Cadambe and Jafar obtained the
symmetric capacity of this SUICP(SNCS) and proposed optimal
length codes by using Vandermonde matrices. In our earlier
work, we gave optimal length (U + 1)-dimensional vector linear
index codes for SUICP(SNCS) satisfying some conditions on K, D
and U [7]. In this paper, for SUICP(SNCS) with arbitrary K, D
U +1
-dimensional
and U , we construct optimal length gcd(K,D−U,U
+1)
vector linear index codes. We prove that the constructed vector
linear index code is of minimal dimension if gcd(K −D+U, U +1)
is equal to gcd(K, D −U, U +1). The proposed construction gives
optimal length scalar linear index codes for the SUICP(SNCS) if
(U + 1) divides both K and D − U . The proposed construction
is independent of field size and works over every field. We
give a low-complexity decoding for the SUICP(SNCS). By using
the proposed decoding method, every receiver is able to decode
its wanted message symbol by simply adding some index code
symbols (broadcast symbols).
I. I NTRODUCTION
AND
BACKGROUND
An index coding problem, consists of a transmitter and a
set of m receivers, R = {R0 , R1 , . . . , Rm−1 }. The transmitter
has a set of n independent messages, X = {x0 , x1 , . . . , xn−1 }.
Each receiver, Rk = (Kk , Wk ), knows a subset of messages,
Kk ⊂ X, called its side-information, and demands to know
another subset of messages, Wk ⊆ Kkc , called its Want-set or
Demand-set. The transmitter can take cognizance of the sideinformation of the receivers and broadcast coded messages,
called the index code, over a noiseless channel. The problem
of index coding with side-information was introduced by Birk
and Kol [3]. An index coding problem is called single unicast
[4] if the demand sets of the receivers are disjoint and the size
of each demand set is one.
Index coding with side-information is motivated by wireless broadcasting applications. In applications like video-ondemand, during the initial transmission by the transmitter, each
receiver may miss a part of data. A naı̈ve technique is to
rebroadcast the entire data again to the receivers. This is an
inefficient approach. The index coding problem, therefore aims
at reducing the number of transmissions by the transmitter by
intelligently using the data already available at the receivers.
In a vector linear index code (VLIC) xk ∈ Fpq k , xk =
(xk,1 , xk,2 , . . . , xk,pk ), xk,j ∈ Fq for k ∈ [0 : n − 1] and
j ∈ [1 : pk ] where Fq is a finite field with q elements. In
vector linear index coding setting, we refer xk ∈ Fpq k as a
message vector or a message and xk,1 , xk,2 , . . . , xk,pk ∈ Fq
as the message symbols. An index coding is a mapping defined
as
E : Fpq 0 +p1 +...+pn−1 → FN
q ,
where N is the length of index code. The index code C =
{(c0 , c1 , . . . , cN −1 )} is the collection of all images of the
mapping E. We call the symbols c0 ,c1 , . . .,cN −1 the code
(broadcast) symbols, which are the symbols broadcasted by
the transmitter. If p0 = p1 = · · · = pn−1 , then the index code
is called symmetric rate vector index code. If p0 = p1 = · · · =
pn−1 = 1, then the index code is called scalar index code. The
index coding problem is to design an index code such that the
number of transmissions N broadcasted by the transmitter is
minimized and all the receivers get their wanted messages by
using the index code symbols (broadcast symbols) broadcasted
and their known side-information.
Maleki, Cadambe and Jafar [1] found the capacity of
SUICP(SNCS) with K messages and K receivers, each receiver has a total of U + D side-information, corresponding
to the U messages before and D messages after its desired
message. In this setting, the kth receiver Rk demands the
message xk having the side-information
Kk = {xk−U , . . . , xk−2 , xk−1 } ∪ {xk+1 , xk+2 , . . . , xk+D }.
(1)
The symmetric capacity of this index coding problem setting
is:
1
if U + D = K − 1
(2)
C=
U+1
if
U + D ≤ K − 2.
K−D+U
where U, D ∈ Z, 0 ≤ U ≤ D.
In the one-sided side-information case, i.e., the cases where
U is zero, the kth receiver Rk demands the message xk having
the side-information,
Kk = {xk+1 , xk+2 , . . . , xk+D },
(3)
for which (2) reduces to
1
C=
1
K−D
if D = K − 1
if D ≤ K − 2
(4)
symbols per message.
Jafar [2] established the relation between index coding
problem and topological interference management problem.
The capacity and optimal coding results in index coding can
be used in corresponding topological interference management
problems. The symmetric and neighboring side-information
problems are motivated by topological interference management problems.
A. Review of Adjacent Independent Row (AIR) matrices
In [8], we constructed binary matrices of size m×n(m ≥ n)
such that any n adjacent rows of the matrix are linearly
independent over every field. We refer these matrices as
Adjacent Independent Row (AIR) matrices.
The matrix obtained by Algorithm 1 is called the (m, n)
AIR matrix and it is denoted by Lm×n . The general form of
the (m, n) AIR matrix is shown in Fig. 1. It consists of several
submatrices (rectangular boxes) of different sizes as shown in
Fig.1. The description of the submatrices are as follows: Let c
and d be two positive integers and d divides c. The following
matrix denoted by Ic×d is a rectangular matrix.
Id
Id
c
number of Id matrices
(5)
Ic×d = .
d
..
Id
and Id×c is the transpose of Ic×d .
Towards explaining the other quantities shown in the AIR
matrix shown in Fig. 1, for a given m and n, let λ−1 =
n, λ0 = m − n and
n = β0 λ0 + λ1 ,
λ0 = β1 λ1 + λ2 ,
λ1 = β2 λ2 + λ3 ,
λ2 = β3 λ3 + λ4 ,
..
.
λi = βi+1 λi+1 + λi+2 ,
..
.
λl−1 = βl λl .
(6)
where λl+1 = 0 for some integer l, λi , βi are positive integers
and λi < λi−1 for i = 1, 2, . . . , l. The number of submatrices
in the AIR matrix is l + 2 and the size of each submatrix is
shown using λi , βi , i ∈ [0 : l].
In [8], we gave an optimal length scalar linear index code
for one-sided SUICP(SNCS) using AIR encoding matrices. In
[7], we constructed optimal length (U +1) dimensional VLICs
for two-sided SUICP(SNCS) satisfying some conditions on
K, D and U . The VLIC construction in [7] does not use AIR
Algorithm 1 Algorithm to construct the AIR matrix L of size
m×n
Let L = m × n blank unfilled matrix.
Step 1
1.1: Let m = qn + r for r < n.
1.2: Use Iqn×n to fill the first qn rows of the unfilled part
of L.
1.3: If r = 0, Go to Step 3.
Step 2
2.1: Let n = q ′ r + r′ for r′ < r.
′
2.2: Use IT
q′ r×r to fill the first q r columns of the unfilled
part of L.
2.3: If r′ = 0, go to Step 3.
2.4: m ← r and n ← r′ .
2.5: Go to Step 1.
Step 3 Exit.
matrices. In [9], we gave a low-complexity decoding for onesided SUICP(SNCS) with AIR matrix as encoding matrix. The
low complexity decoding method helps to identify a reduced
set of side-information for each user with which the decoding
can be carried out. By this method every receiver is able to
decode its wanted message symbol by simply adding some
broadcast symbols.
B. Contributions
In a b-dimension vector index coding, the transmitter has to
wait for b-realizations of a given message symbol to perform
the index coding. In scalar index codes, the transmitter encodes
each realization of the message symbols separately. In a
delay critical environment like real time video streaming, it
may not be desirable to wait for b-realizations of a given
message symbol. Hence, in this paper, we focus on reducing
the dimension of VLICs without compromising on optimal
length of the code.
• For the two-sided SUICP(SNCS) with arbitrary K, D and
U+1
dimenU , we construct optimal length gcd(K,D−U,U+1)
sional VLICs. The proposed construction is independent
of field size and works over every field.
• We prove that the constructed VLICs are of minimal
dimension if gcd(K − D + U, U + 1) is equal to
gcd(K, D − U, U + 1). In other words, optimal VLICs
of lesser dimension do not exist.
• We give a low-complexity decoding for the two-sided
SUICP(SNCS). By this method every receiver is able
to decode its wanted message symbol by simply adding
some broadcast symbols.
• We give generator matrices for the proposed VLICs by
using AIR matrices.
The SUICP(SNCS) considered in this paper was referred as
symmetric neighboring antidote multiple unicast index coding
problem by Maleki, Cadambe and Jafar in [1].
All the subscripts in this paper are to be considered
modulo K.
n
In
n
n − λ1
m
λ1
Iβ1λ1×λ1
λ0 − λ2
IT
β0λ0×λ0
m−n
λ3
λ1 − λ3
λ2
λ0 − λ2
ITβ2λ2×λ2
λ2
S
S = Iλl ×βl λl if l is even and S = Iβl λl ×λl otherwise.
Fig. 1: AIR matrix of size m × n.
The paper is organized as follows. In Section II, we give a
construction of reduced dimension optimal VLICs for twosided SUICP(SNCS). In Section III, we give some special
cases of the proposed construction.
II. C ONSTRUCTION OF R EDUCED D IMENSION V ECTOR
L INEAR I NDEX C ODES
For the b dimensional vector linear index coding, we refer
xk ∈ Fbq as a message vector and xk,1 , xk,2 , . . . , xk,b ∈ Fq as
message symbols.
Theorem 1. Given arbitrary positive integers K, D and U ,
consider the two-sided SUICP(SNCS) with K messages
{x0 , x1 , · · · , xK−1 } and the receivers being K, and the receiver Rk for k ∈ [0 : K − 1] wanting the message xk and
having the side-information given by
with K messages and D side-information after and U sideinformation before is given by
[c0 c1 . . . cKa −∆a −1 ] =
KX
a −1
ys L s .
(11)
s=0
Proof. We prove that every receiver Rk for k ∈ [0 : K − 1]
decodes itswanted
message vector (xk,1 , xk,2 , . . . , xk,ua ).
Let s = ka . We have
k = as + r for some r ∈ [0 : a − 1].
(12)
Let
zs,i =
a−1
X
xas+j,i ,
s ∈ [0 : Ka − 1] and i ∈ [1 : ua ].
j=0
Kk = {xk−U , . . . , xk−2 , xk−1 } ∪ {xk+1 , xk+2 , . . . , xk+D }.
(7)
Let
ua =
U +1
,
a
∆a =
D−U
,
a
Ka =
K
,
a
(8)
where
a = gcd(K, D − U, U + 1).
(9)
Let xk = (xk,1 , xk,2 , · · · , xk,ua ) be the message vector
wanted by the kth receiver, where xk ∈ Fuq a , xk,i ∈ Fq for
every k ∈ [0 : K − 1] and i ∈ [1 : ua ].
Let
ys =
ua a−1
X
X
xa(s+1−i)+j,i
s ∈ [0 : Ka − 1].
(10)
i=1 j=0
Let L be the AIR matrix of size Ka × (Ka − ∆a ) and Ls
be the sth row of L for every s ∈ [0 : Ka − 1]. The optimal
length ua dimensional VLIC for the two-sided SUICP(SNCS)
(13)
We refer zs,i as extended message symbol and zs =
(zs,1 zs,2 . . . , zs,ua ) as extended message vector. The extended message symbol zs,i comprises of ′ a′ message symbols xas,i , xas+1,i , . . . , xas+a−1,i . Each of the Kua message
symbols xk,i for k ∈ [0 : K − 1] and i ∈ [1 : ua ] appear
exactly once in Ka ua extended message symbols zs,i for
s ∈ [0 : Ka − 1] and i ∈ [1 : ua ].
From (10) and (13), we have
ys =
ua
X
zs+1−i,i for s ∈ [0 : Ka − 1].
(14)
i=1
The symbol ys comprises of ua extended message symbols
zs,1 , zs−1,2 , . . . , zs−ua −1,ua . The Ka ua symbols zs,i for s ∈
[0 : Ka − 1] and i ∈ [1 : ua ] appear exactly once in Ka
symbols ys for s ∈ [0 : Ka − 1]. By combining (13) and (14),
the symbol ys comprises of aua = U + 1 message symbols
corresponding to U + 1 neighboring receivers.
•
•
•
•
The symbol ys comprises of (U + 1) message symbols
one from each of the message vectors xt for every t ∈
[as − (U + 1) + a : as + a − 1].
The symbol ys+1 comprises of U + 1 message symbols
one from each of the message vectors xt for every t ∈
[as − (U + 1) + 2a : as + 2a − 1].
The symbol ys+∆a comprises of U + 1 message symbols
one from each of the message vectors xt for every t ∈
[as − (U + 1) + a(∆a + 1) : as + a(∆a + 1) − 1].
The symbol vector [ys ys+1 . . . ys+∆a ] comprises of
(U + 1)∆a message symbols from the message vectors
xt for every t ∈ [as− (U + 1)+ a : as+ a(∆a + 1)− 1] =
[as − (U + 1) + a : as
+ D −{zU + a − 1}].
|
{z
} |
=h
=g
Note that h − g = D.
From the alignment of the symbols ys , ys+1 , . . . , ys+∆a
and the side-information given in (7), we can conclude the
following:
Observation 1. For every receiver Rk whose wanted message
symbol is present in ys , every other message symbol in ys is
in side-information of Rk .
Observation 2. For every receiver Rk whose wanted
message is present in ys , every message symbol in
ys+1 , ys+2 , . . . , ys+∆a which does not belong to the message vector xk is in side-information of Rk . That is, for
every receiver Rk whose wanted message is present in ys ,
every extended message symbol in ys+1 , ys+2 , . . . , ys+∆a
which does not belong to the message vector zs is in sideinformation
Pa−1of Rk (we refer the extended message symbol
zt,i = j=0 xat+j,i is in the side-information of Rk if every
message symbol xat+j,i for j ∈ [0 : a − 1] is in the sideinformation of Rk ).
Consider the one-sided SUICP(SNCS) with Ka messages
{y0 , y1 , · · · , yKa −1 }, Ka receivers and the receiver wanting
the message ys and having the side-information given by Ks =
{ys+1 , ys+2 , . . . , ys+∆a }, for every s ∈ [0 : Ka − 1]. Let C
be the optimal length index code given by (11). In [8], we
proved that the code C enables the decoding of Ka message
y0 , y1 , · · · , yKa −1 . That is, for every s ∈ [0 : Ka − 1], with
linear decoding there exists a linear combination of broadcast
symbols c0 , c1 , · · · , cKa −∆a −1 to get the sum of the form
Ss = ys + b(1)
+ b(2)
+ · · · + b(d)
s ys+a(1)
s ys+a(2)
s ys+as(d)
s
s
(15)
(1) (2)
(d)
where bs , bs , . . . , bs
(d)
. . . < as ≤ ∆a .
∈ Fq , d ≤ ∆a and 1 ≤
(1)
as
<
(2)
as
<
In the given two-sided
SUICP(SNCS), we have k ∈ [0 :
K − 1], s = ka ∈ [0 : Ka − 1]. From (13), the message
symbol xk,i for i ∈ [1 : ua ] is present in the extended message
symbol zs,i . For every message symbol xk,i for i ∈ [1 : ua ],
the decoding is performed as given below.
•
Rk first decodes the extended message symbol zs,i for
i ∈ [1 : ua ], where the message symbol xk,i is present.
•
Pa−1
In zs,i = j=0 xas+j,i , every message symbol present
is in the side-information of Rk . Hence, Rk decodes its
wanted message symbol xk,i from zs,i .
Step 1. Decoding of xk,ua
The message symbol xk,ua is present in zs,ua . Receiver Rk
first decodes the extended message symbol zs,ua as follows.
From (14), we have
ys+ua −1 =
ua
X
zs+ua −i,i = zs,ua +
uX
a −1
zs+ua −i,i .
i=1
i=1
Hence, the extended message symbol zs,ua is present in
ys+ua −1 . The symbol ys+ua −1 can be decoded from Ss+ua −1 .
The sum Ss+ua −1 obtained from (15) is given in (16) and (17).
In Ss+ua −1 , we have
(1)
(2)
(d)
1 ≤ as+ua −1 < as+ua −1 < · · · < as+ua −1 ≤ ∆a .
(18)
From (14) and (18), we have
ys+ua −1+t =
ua
X
zs+ua +t−i,i for s ∈ [0 : Ka − 1],
i=1
(d)
(2)
(1)
t ∈ {as+ua −1 , as+ua −1 , . . . , as+ua −1 },
and in ys+ua −1+t , the extended message symbols belonging to zs are not present. Hence, in Ss+ua −1 , only one
extended message symbol is present which belongs to zs .
From Observation 2, the symbols ys+ua −1+t for every t ∈
(1)
(2)
(d)
{as+ua −1 , as+ua −1 , . . . , as+ua −1 } are in the side-information
of Rk . From Observation 1, every message symbol present in
ys+ua −1 is in side-information of Rk except xk,ua . Hence,
by using Ss+ua −1 , receiver Rk decodes its wanted extended
message symbol zs,ua . From (12) and (13), we have
zs,ua =
a−1
X
xas+j,ua
j=0
=
a−1
X
xk−r+j,ua where r ∈ [0 : a − 1]
j=0
= xk,ua +
a−1
X
xk−r+j,ua .
j=0,j6=r
|
{z
}
side information to Rk
Hence, Rk decodes its wanted message xk,ua by using zs,ua
(Note that if Ss+ua −1 comprises h ≥ 2 extended message
symbols for some integer h which belong to zs and if Rk yet
to decode c of them for some integer c such that 2 ≤ c ≤ h,
then the c extended message symbols interfere and Rk can not
decode any of the c message symbols).
Step 2. Decoding of xk,ua −1
The message symbol xk,ua −1 is present in the extended
message symbol zs,ua −1 . Receiver Rk first decodes the extended message symbol zs,ua −1 as follows. From (14), we
Ss+ua −1 = ys+ua −1 +
d
X
(t)
bs+ua −1 ys+ua −1+a(t)
|
Ss+ua −1 =
+
zs,u
| {z a}
wanted extended message symbol toRk
|
uX
a −1
ys+ua −2 =
|
{z
zs+ua −1−i,i
i=1
ua
X
= zs,ua −1 +
zs+ua −1−i,i .
i=1,i6=ua −1
Hence, the extended message symbol zs,ua −1 is present in
ys+ua −2 . Receiver Rk uses the sum Ss+ua −2 to decode
zs,ua −1 . The sum Ss+ua −1 obtained from (15) is given in
(19) and (20).
In (19), we have
(1)
(2)
(d)
1 ≤ as+ua −2 < as+ua −2 < · · · < as+ua −2 ≤ ∆a .
(21)
(1)
Depending on the value of as+ua −2 , sum Ss+ua −2
comprises of either one extended message symbol
or two extended message symbols belonging to
(1)
zs (zs,1 , zs,2 , . . . , zs,ua −1 , zs,ua ). If as+ua −2 > 1, then
Ss+ua −2 comprises only one extended message symbol
(1)
(zs,ua −1 ) belonging to zs . If as+ua −2 = 1, Ss+ua −2 comprises
two extended message symbols (zs,ua , zs,ua −1 ) belonging to
zs . Receiver Rk has already decoded the extended message
symbol zs,ua in Step 1. From Observation 2, the symbols
(1)
(2)
(d)
ys+ua −2+t for every t ∈ {as+ua −2 , as+ua −2 , . . . , as+ua −2 }
are in the side-information of Rk . From Observation 1, every
message symbol present in ys+ua −2 is in side-information of
Rk except xk,ua −1 . Hence, by using Ss+ua −2 , receiver Rk
decodes its wanted extended message symbol zs,ua −1 . From
(12) and (13), we have
zs,ua −1 =
a−1
X
=
(16)
}
d
X
(t)
}
}
|
(17)
s+ua −1
t=1
{z
.
bs+ua −1 ys+ua −1+a(t)
{z
side information to Rk
}
Step l. Decoding of xk,ua +1−l for l ∈ [3 : ua ]
The message symbol xk,ua +1−l is present in the extended
message symbol zs,ua +1−l . Receiver Rk first decodes the
extended message symbol zs,ua +1−l as follows. From (14),
the extended message symbol zs,ua +1−l is present in ys+ua −l .
The symbol ys+ua −l can be decoded from Ss+ua −l . Define
(j)
(j)
(j)
Aj = {as+ua −l + 1, as+ua −l + 2, · · · , ua } if ua > as+ua −l ,
else Aj = φ for every j ∈ [1 : l − 1]. Let 1A be the
indicator function such that 1A (x) = 1 if x ∈ A, else it is
zero. The sum Ss+ua −l obtained from (15) is given in (22)
and (23). In this sum Ss+ua −l , zs,ua +1−l is the required ex(j)
tended message symbol and 1Aj (l)bs+ua −l zs,ua +1−l+a(j)
s+ua −l
for j ∈ [1 : l − 1] is the interference to the receiver
Rk from its wanted extended message vector zs . Receiver
Rk already knows the l − 1 extended message symbols
{zs,ua , zs,ua −1 , · · · , zs,ua +2−l } and thus the interference from
the extended message vector zs in the sum Ss+ua −l can
be cancelled. From Observation 2, the symbols ys+ua −l+t
(1)
(2)
(d)
for every t ∈ {as+ua −l , as+ua −l , . . . , as+ua −l } are in the
side-information of Rk . From Observation 1, every message
symbol present in ys+ua −l is in side-information of Rk except
xk,ua +1−l . Hence, by using Ss+ua −l , receiver Rk decodes its
wanted extended message symbol zs,ua +1−l . From (12) and
(13), we have
zs,ua −l =
a−1
X
xas+j,ua −l =
a−1
X
xk−r+j,ua −l
j=0
j=0
= xk,ua −l +
a−1
X
xk−r+j,ua −l .
j=0,j6=r
|
xas+j,ua −1
j=0
a−1
X
+
zs+ua −i,i
i=1
have
ua
X
{z
side information to Rk
side information to receiver Rk
ys+ua −1
.
s+ua −1
t=1
xk−r+j,ua −1 where r ∈ [0 : a − 1]
j=0
= xk,ua −1 +
a−1
X
xk−r+j,ua −1 .
j=0,j6=r
|
{z
side information to Rk
}
Hence, Rk decodes its wanted message xk,ua −1 by using
zs,ua −1 .
{z
side information to Rk
}
Hence, Rk decodes its wanted message xk,ua −l by using
zs,ua −l .
The decoding continues until the receiver Rk has decoded
its ua wanted extended message symbols zs,1 , zs,2 , · · · , zs,ua .
The receiver Rk decodes the extended message vector zs
successively in the following order:
zs,ua → zs,ua −1 → zs,ua −2 → . . . → zs,2 → zs,1 .
Hence, The receiver Rk decodes the wanted message xk
(1)
Ss+ua −2 = ys+ua −2 + bs+ua −2 ys+ua −2+a(1)
+
s+ua −2
Ss+ua −2 =
{z
(1)
bs+ua −2 zs−1+a(1)
|
{z
|
s+ua −2 ,ua
(1)
y
|
l−1
X
wanted extended message symbol to Rk
+
uX
a −1
{z
{z
{z
{z
}
}
++
d
X
zs+ua +1−l−i,i +
i=1,i6=ua +1−l
|
side information to Rk
+
a −l
}
t=1
|
(t)
bs+ua −2 ys+ua −2+a(t)
{z
|
}
}
|
(t)
.
(22)
s+ua −l
{z
}
side information to Rk
(t)
bs+ua −l ys+ua −l+a(t)
s+ua −l
{z
side information to Rk
(t)
bs+ua −l ys+ua −l+a(t)
s+ua −l
}
− 1At (l)zs,ua +1−l+a(t)
s+ua −l
{z
successively in the following order:
{z
(t)
s+ua −l+a
s+ua −l
.
(23)
}
side information to Rk
y
(20)
}
side information to Rk
bs+ua −l ys+ua −l+a(t)
d
X
.
s+ua −2
t=2
t=l
{z
l−1
X
+
d
X
t=l
|
ua
X
Interference from zs
|
s+ua −2 −i,i
a −l
(t)
bs+ua −l 1At (l)zs,ua +1−l+a(t)
s+u
|
zs+ua −1+a(1)
(t)
bs+ua −l ys+ua −l+a(t)
s+u
ys+ua −l
t=1
}
side information to Rk
+
zs,ua +1−l
| {z }
l−1
X
}
}
(1)
s+ua −2+a
s+ua −2
t=1
|
{z
i=1
}
Ss+ua −l = ys+ua −l +
Ss+ua −l =
{z
side information to Rk
zs+ua −1−i,i
(1)
+ bs+ua −2
interference from zs if as+ua −2 =1
|
(19)
side information to receiver Rk
ys+ua −2
+
.
s+ua −2
t=2
i=1,i6=ua −1
wanted extended message symbol to Rk
|
(t)
bs+ua −2 ys+ua −2+a(t)
|
ua
X
+
zs,ua −1
| {z }
d
X
}
proceeds to decode
xk,ua → xk,ua −1 → xk,ua −2 → . . . → xk,2 → xk,1 .
xk,ua −1 → xk,ua −2 → . . . → xk,1
The extended message symbols wanted by receiver Rk are
decoded from Si as given in Table I.
The number of broadcast symbols in the ua dimensional
VLIC C is Ka − ∆a . The rate achieved by C is
after cancelling the interference from already decoded message
symbols.
ua
=
Ka − ∆a
U+1
a
K
D−U
−
a
a
=
U +1
.
K −D+U
This is equal to the capacity mentioned in (2). Hence, the
constructed VLICs are capacity achieving.
Remark 1. The decoding procedure given in Theorem 1 uses
successive interference cancellation. That is, in the decoding, the receiver Rk decode its wanted message symbols
(xk,1 , xk,2 , . . . , xk,ua ) one at a time starting from xk,ua and
In the Lemma 1 given below, we prove that the constructed
optimal VLICs in Theorem 1 are minimal dimensional if
gcd(K, D − U, U + 1) is equal to gcd(K − D + U, U + 1).
Lemma 1. If gcd(K, D − U, U + 1) = gcd(K − D + U, U +
1), then the constructed VLICs in Theorem 1 are minimal
dimensional.
K−D+U
U+1
, b = gcd(K−D+U,U+1)
. We
Proof. Let a = gcd(K−D+U,U+1)
have gcd(a, b) = 1. For SUICP(SNCS), we have
C=
U +1
=
K −D+U
U+1
gcd(K−D+U,U+1)
K−D+U
gcd(K−D+U,U+1)
=
a
.
b
xk,ua
zs,ua
Ss+ua −1
xk,ua −1
zs,ua −1
Ss+ua −2
...
...
...
xk,ua +1−l
zs,ua +1−l
Ss+ua −l
...
...
...
xk,2
zs,2
Ss+1
xk,1
zs,1
Ss
TABLE I: Sequential decoding of message vector from VLIC
If a = 1, then the index code is a scalar linear index code and
it is the minimal dimensional index code. Hence, with out loss
of generality, we assume a > 1. Consider there exist an a − t
dimensional capacity achieving VLIC for some t ∈ [1 : a − 1].
This a−t dimensional VLIC is obtained by mapping (a−t)K
message symbols into b − s broadcast symbols for some s ∈
[1 : b − 1]. This code is capacity achieving and hence we have,
a−t
a
=
b−s
b
(24)
which gives as − bt = 0. From the extended Euclid algorithm
and Bezout coefficients, the equation of the form as − bt = 0
a
b
and t = k gcd(a,b)
for
has solutions of the form s = k gcd(a,b)
k ∈ Z>0 . We have gcd(a, b) = 1 and hence s = kb ≥ b and
t = ka ≥ a. This is a contradiction. Hence, the dimension
used in Theorem 1
a=
U +1
U +1
=
gcd(K − D + U, U + 1)
gcd(K, D − U, U + 1)
is minimal.
L8×7
1000000
0 1 0 0 0 0 0
0 0 1 0 0 0 0
0 0 0 1 0 0 0
=
0 0 0 0 1 0 0
0 0 0 0 0 1 0
0 0 0 0 0 0 1
1111111
by
C = {y0 + y7 ,
y4 + y7 ,
y1 + y7 ,
y5 + y7 ,
y2 + y7 , y3 + y7 ,
y6 + y7 }.
The VLIC for the given two-sided SUICP(SNCS) is obtained by replacing ys in C with xs,1 + xs−1,2 for s ∈ [0 : 7].
The VLIC is given by
C(2s) = {x0,1 + x7,2 + x7,1 + x6,2 , x1,1 + x0,2 + x7,1 + x6,2 ,
For all SUICP(SCNC) not satisfying the condition
gcd(K, D −U, U +1) = gcd(K −D +U, U +1), we conjecture
that the constructed optimal VLICs in Theorem 1 are minimal
dimensional.
A. Low Complexity Decoding
In [9], we gave a low-complexity decoding for one-sided
SUICP(SNCS) with AIR matrix as encoding matrix. The low
complexity decoding method helps to identify a reduced set
of side-information for each user with which the decoding
can be carried out. By this method every receiver is able to
decode its wanted message symbol by simply adding some
broadcast symbols. The low complexity decoding of one-sided
SUICP(SNCS) with Ka messages and ∆a side-information
(d)
(d)
explicitly gives the values of at and bt in (15) for t ∈
[0 : Ka − 1] and d ∈ [1 : ∆a ]. Hence, from the proof of
Theorem 1 we get a low complexity decoding for two-sided
SUICP(SNCS) with arbitrary K, D and U .
The following examples illustrate the construction of optimal VLICs and low complexity decoding for two-sided
SUICP(SNCS).
Example 1. Consider the two-sided SUICP(SNCS) with K =
8, D = 2, U = 1. For this SUICP(SNCS), gcd(K, D −
U, U + 1) = a = 1. The optimal length index code is a two
dimensional index code. For this two-sided SUICP(SNCS), we
have Ka = 8 and ∆a = 1. The AIR matrix of size L8×7 is
given below.
The scalar linear index code for the one-sided
SUICP(SNCS) with Ka = 8 and ∆a = 1 is given
x2,1 + x1,2 + x7,1 + x6,2 , x3,1 + x2,2 + x7,1 + x6,2 ,
x4,1 + x3,2 + x7,1 + x6,2 , x5,1 + x4,2 + x7,1 + x6,2 ,
x6,1 + x5,2 + x7,1 + x6,2 }.
Let τs be the set of broadcast symbols required to decode
ys from C. The values of τs and Ss are given in Table II.
Ws
y0
y1
y2
y3
y4
y5
y6
y7
τs
y0 + y7 , y1 + y7
y1 + y7 , y2 + y7
y2 + y7 , y3 + y7
y3 + y7 , y4 + y7
y4 + y7 , y5 + y7
y5 + y7 , y6 + y7
y6 + y7
y0 + y7
S0
S1
S2
S3
S4
S5
S6
S7
=
=
=
=
=
=
=
=
Ss
y0 + y1
y1 + y2
y2 + y3
y3 + y4
y4 + y5
y5 + y6
y6 + y7
y7 + y0
TABLE II: Decoding at each receiver for one-sided
SUICP(SNCS) with K = 8, ∆a = 1
Receiver Rk wants to decode xk,1 and xk,2 for k ∈ [0 : 7].
Rk first decode xk,2 from Ss+1 , then decodes xk,1 from Ss
for every k ∈ [0 : 7]. The decoding procedure at the receivers
R0 , R1 . . . R7 is shown in Table III.
The capacity of the given two-sided SUICP(SNCS) is 27 .
The VLIC C uses seven broadcast symbols to transmit two
message symbols per receiver. The rate achieved by C is 27 .
Hence, C(2) is an optimal length index code.
Example 2. Consider the two-sided SUICP(SNCS) with K =
22, D = 7, U = 3. For this SUICP(SNCS), gcd(K, D −
U, U + 1) = a = 2. The optimal length index code is a two
dimensional index code. For this two-sided SUICP(SNCS), we
have Ka = 11 and ∆a = 2. The AIR matrix of size L11×9 is
given below.
Rk
R0
R0
R1
R1
R2
R2
R3
R3
R4
R4
R5
R5
R6
R6
R7
R7
Wk
x0,2
x0,1
x1,2
x1,1
x2,2
x2,1
x3,2
x3,1
x4,2
x4,1
x5,2
x5,1
x6,2
x6,1
x7,2
x7,1
Sum from which xk,i is decoded
S1 = x1,1 + x0,2 + x2,1 + x1,2
S0 = x0,1 + x7,2 + x1,1 + x0,2
S2 = x2,1 + x1,2 + x3,1 + x2,2
S1 = x1,1 + x0,2 + x2,1 + x1,2
S3 = x3,1 + x2,2 + x4,1 + x3,2
S2 = x2,1 + x1,2 + x3,1 + x2,2
S4 = x4,1 + x3,2 + x5,1 + x4,2
S3 = x3,1 + x2,2 + x4,1 + x3,2
S5 = x5,1 + x4,2 + x6,1 + x5,2
S4 = x4,1 + x3,2 + x5,1 + x4,2
S6 = x6,1 + x5,2 + x7,1 + x6,2
S5 = x5,1 + x4,2 + x6,1 + x5,2
S7 = x7,1 + x6,2 + x0,1 + x7,2
S6 = x6,1 + x5,2 + x7,1 + x6,2
S0 = x0,1 + x7,2 + x1,1 + x0,2
S7 = x7,1 + x6,2 + x0,1 + x7,2
TABLE III: Decoding at each receiver for the two-sided
SUICP(SNCS) with K = 8, D = 2 and U = 1
L11×9
100000000
0 1 0 0 0 0 0 0 0
0 0 1 0 0 0 0 0 0
0 0 0 1 0 0 0 0 0
0 0 0 0 1 0 0 0 0
=
0 0 0 0 0 1 0 0 0
0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 1
1 0 1 0 1 0 1 0 1
010101011
The scalar linear index code for the one-sided
SUICP(SNCS) with Ka = 11, ∆a = 2 is given by
C = {y0 + y9 ,
y4 + y9 ,
y1 + y10 ,
y5 + y10 ,
y2 + y9 ,
y6 + y9 ,
y3 + y10 ,
y7 + y10 ,
Table V, we have
S7 = y7 + y8 + y9 = x14,1 + x15,1 + x12,2 + x13,2 +
{z
}
|
y7
x16,1 + x17,1 + x14,2 + x15,2 + x18,1 + x19,1 + x16,2 + x17,2 .
|
{z
} |
{z
}
y8
y9
In S7 , all other messages except x14,1 and x14,2 are in the
side information of R14 . R14 already decoded x14,2 and hence
it cancels the interference in S7 due to x14,2 and then decodes
x14,1 .
4
=
The capacity of the given two-sided SUICP(SNCS) is 18
2
.
The
VLIC
C
uses
nine
broadcast
symbols
to
transmit
two
9
message symbols per receiver. The rate achieved by C is 92 .
Note 1. For the two-sided SUICP(SNCS) given in Example 2,
the construction given in [7] gives four dimensional optimal
length VLIC, whereas the construction given in this paper
gives two dimensional optimal length VLIC.
Example 3. Consider the two-sided SUICP(SNCS) with K =
24, D = 11 and U = 2. For this SUICP(SNCS), we have
gcd(K, D − U, U + 1) = 3. The optimal length index code for
this SUICP(SNCS) is scalar linear index codes. For this twosided SUICP(SNCS), we have Ka = 8, ∆a = 3. AIR matrix
of size 8 × 5 is given below.
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
L8×5 = 0 0 0 0 1
1 0 0 1 0
01001
00111
The scalar linear code C for the one-sided SUICP(SNCS)
with Ka = 8 and ∆a = 3 is
C = {c0 , c1 .c2 , c3 , c4 }
= {y0 + y5 , y1 + y6 , y2 + y7 , y3 + y5 + y7 , y4 + y6 + y7 }.
y8 + y9 + y10 }.
The VLIC for the given two-sided SUICP(SNCS) is obtained by replacing ys in C with x2s,1 + x2s+1,1 + x2(s−1),2 +
x2(s−1)+1,2 for s ∈ [0 : 10]. The two dimensional VLIC for
the is given SUICP(SNCS) is given in Table IV.
Let τs be the set of broadcast symbols required to decode
ys from C. The values of τs and Ss for s ∈ [0 : 10] are given
in Table V.
Receiver Rk wants to decode xk,1 and xk,2 . Rk first decode
xk,2 from Ss+1 , then decodes xk,1 from Ss for every k ∈ [0 :
21]. The receiver R14 decodes x14,2 from S8 and decodes
x14,1 from S7 . From Table V, we have
S8 = y8 + y9 + y10 = x16,1 + x17,1 + x14,2 + x15,2 +
{z
}
|
y8
The scalar linear index code C(2s) for the given two-sided
SUICP(SNCS) is given by
C(2s) = {c0 , c1 , c2 , c3 , c4 } =
{ x0 + x1 + x2 + x15 + x16 + x17 ,
{z
} |
{z
}
|
y0
y5
x + x + x5 + x18 + x19 + x20 ,
| 3 {z4
} |
{z
}
y1
y6
x + x + x8 + x21 + x22 + x23 ,
} |
{z
}
| 6 {z7
y2
y7
x9 + x10 + x11 + x15 + x16 + x17 + x21 + x22 + x23 ,
|
{z
} |
{z
} |
{z
}
y3
y5
y7
x12 + x13 + x14 + x18 + x19 + x20 + x21 + x22 + x23 }.
|
{z
} |
{z
} |
{z
}
y4
y6
y7
x18,1 + x19,1 + x16,2 + x17,2 + x20,1 + x21,1 + x18,2 + x19,2 .
|
{z
} |
{z
} Note 2. For the two-sided SUICP(SNCS) given in Example 3,
y9
y10
the construction given in [7] gives three dimensional optimal
In S8 , all other messages except x14,2 are in the side length VLIC, whereas the construction given in this paper
information of R14 . Hence R14 decodes x14,2 from S8 . From gives optimal length scalar linear index code.
c0
x0,1 + x1,1 + x20,2 + x21,2 + x18,1 + x19,1 + x16,2 + x17,2
{z
} |
{z
}
|
c1
x2,1 + x3,1 + x0,2 + x1,2 + x20,1 + x21,1 + x0,2 + x1,2
{z
} |
{z
}
|
c2
x4,1 + x5,1 + x2,2 + x3,2 + x18,1 + x19,1 + x16,2 + x17,2
{z
} |
{z
}
|
c3
x6,1 + x7,1 + x4,2 + x5,2 + x20,1 + x21,1 + x0,2 + x1,2
{z
} |
{z
}
|
c4
x8,1 + x9,1 + x6,2 + x7,2 + x18,1 + x19,1 + x16,2 + x17,2
|
{z
} |
{z
}
c5
x10,1 + x11,1 + x8,2 + x9,2 + x20,1 + x21,1 + x0,2 + x1,2
|
{z
} |
{z
}
c6
x12,1 + x13,1 + x10,2 + x11,2 + x18,1 + x19,1 + x16,2 + x17,2
|
{z
} |
{z
}
c7
x14,1 + x15,1 + x12,2 + x13,2 + x20,1 + x21,1 + x0,2 + x1,2
|
{z
} |
{z
}
c8
x16,1 + x17,1 + x14,2 + x15,2 + x18,1 + x19,1 + x16,2 + x17,2 + x20,1 + x21,1 + x0,2 + x1,2
{z
} |
{z
} |
{z
}
|
y0
y9
y1
y10
y2
y9
y3
y10
y4
y9
y5
y10
y6
y9
y7
y10
y8
y9
y10
TABLE IV: Vector linear index code for the two-sided SUICP(SNCS) given in Example 2
Ws
y0
y1
y2
y3
y4
y5
y6
y7
y8
y9
y10
τs
y0 + y9 , y2 + y9
y1 + y10 , y3 + y10
y2 + y9 , y4 + y9
y3 + y10 , y5 + y10
y4 + y9 , y6 + y9
y5 + y10 , y7 + y10
y6 + y9 , y7 + y10 ,
y8 + y9 + y10
y7 + y10 , y8 + y9 + y10
y8 + y9 + y10
y0 + y9
y1 + y10
Ss
S0 = y 0 + y 3
S1 = y 1 + y 3
S2 = y 2 + y 4
S3 = y 3 + y 5
S4 = y 4 + y 6
S5 = y 5 + y 7
S6 = y 6 + y 7 + y 8
S7 = y 7 + y 8 + y 9
S8 = y8 + y9 + y10
S9 = y 0 + y 9
S10 = y1 + y10
TABLE V: Decoding at each receiver for one-sided
SUICP(SNCS) with Ka = 11, ∆a = 2
Lemma 2. Consider the two-sided SUICP(SNCS) with K
messages, U side-information after and D side-information
before the desired message. Let xk = (xk,1 , xk,2 , · · · , xk,ua )
be the message vector wanted by the kth receiver, where
xk ∈ Fuq a , xk,i ∈ Fq for k ∈ [0 : K − 1] and i ∈ [1 : ua ]. For
this two-sided SUICP(SNCS), an encoding matrix for optimal
length VLIC L(2s) of size Kua × (Ka − ∆a ) is given by
Ṽ0
Ṽ1
L(2s) = Ṽ2
..
.
ṼKa −1
B. Encoding Matrices for Optimal Vector Linear Codes
where Ṽk is defined in Definition 1.
In this subsection, we give encoding matrices for optimal
VLICs of two-sided SUICP(SNCS). with K messages, D sideinformation after and U side-information before the desired
message.
Proof. Let xk = [xk,1 xk,2 . . . xk,ua ]. Let x[a:b] =
[xa xa+1 . . . xb ]. From Theorem 1, the ua dimensional VLIC
for the given two-sided SUICP(SNCS) is given by
Definition 1. Let L be the AIR matrix of size Ka ×(Ka −∆a ).
Let Ls be the sth row of L for s ∈ [0 : Ka − 1]. Then, the
sth precoding matrix Vs is defined as
Ls
Ls+1
Vs = Ls+2 .
..
.
[c0 c1 . . .cKa −∆a −1 ] =
=
s=0
=
=
=
ua X
a−1
X
i=1 j=0
(
KX
a −1
s=0
(25)
KX
a −1
j=0
ys L s
s=0
KX
a −1 a−1
X
s=0
Ls+ua −1
The matrix Ṽs is defined as
Vs
Vs
Ṽs = . a number of Vs matrices
..
Vs
KX
a −1
KX
a −1
a−1
X
j=0
ua
X
xa(s+1−i)+j,i Ls
xa(s+1−i)+j,i Ls
i=1
!
xas+j Vs
x[as:as+a−1] Ṽs = x[0:K−1] L(2s) .
s=0
This completes the proof.
(2s)
L16×7
1 0 0 0 0 0 0
0 1 0 0 0 0 0
0 1 0 0 0 0 0
0 0 1 0 0 0 0
0 0 1 0 0 0 0
0 0 0 1 0 0 0
0 0 0 1 0 0 0
0 0 0 0 1 0 0
= 0 0 0 0 1 0 0
0 0 0 0 0 1 0
0 0 0 0 0 1 0
0 0 0 0 0 0 1
0 0 0 0 0 0 1
1 1 1 1 1 1 1
1111111
1000000
Fig. 2: Encoding matrices for two-sided SUICP(SNCS) given
in Example 4
Example 4. Consider the two-sided SUICP(SNCS) with K =
8, D = 2, U = 1. The optimal length VLIC for this SUICP
(2s)
is given in Example 1. The encoding matrix L16×7 for this
optimal length VLIC by using Lemma 2 is given in Fig 2.
Example 5. Consider the two-sided SUICP(SNCS) with K =
22, D = 7, U = 3. The optimal length VLIC for this SUICP
(2s)
is given in Example 2. The encoding matrix L44×9 for this
optimal length VLIC by using Lemma 2 is given in Fig 3.
Example 6. Consider the two-sided SUICP(SNCS) with K =
24, D = 11, U = 2. The optimal length VLIC for this SUICP
(2s)
is given in Example 3. The encoding matrix L24×5 for this
optimal length VLIC by using Lemma 2 is given in Fig 3.
III. S OME SPECIAL CASES OF PROPOSED CONSTRUCTION
In this section, we give four special cases of the proposed
construction. Consider the two-sided SUICP(SNCS) with K
messages, D side-information after and U side-information
before the desired message as given in (1). We have a =
gcd(K, D − U, U + 1). In Theorem 1, we combine U + 1
message symbols, one from each of the message vectors xt
for every t ∈ [as − (U + 1) + a : as + a − 1] as given below.
ys =
ua a−1
X
X
xa(s+1−i)+j,i
(2s)
L44×9
1 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0
0 0 1 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0
0 0 1 0 0 0 0 0 0
0 0 1 0 0 0 0 0 0
0 0 0 1 0 0 0 0 0
0 0 1 0 0 0 0 0 0
10000
0 0 0 1 0 0 0 0 0
1 0 0 0 0
0 0 0 1 0 0 0 0 0
1 0 0 0 0
0 0 0 0 1 0 0 0 0
0 1 0 0 0
0 1 0 0 0
0 0 0 1 0 0 0 0 0
0 1 0 0 0
0 0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0 0
0 0 1 0 0
0 0 0 0 0 1 0 0 0
0 0 1 0 0
0 0 0 0 1 0 0 0 0
0 0 1 0 0
0 0 0 0 0 1 0 0 0
0
0
0
1
0
0 0 0 1 0
0 0 0 0 0 1 0 0 0
0 0 0 1 0
0 0 0 0 0 0 1 0 0 (2s)
= 0 0 0 0 0 1 0 0 0 L24×5 = 0 0 0 0 1
0 0 0 0 1
0 0 0 0 0 0 1 0 0
0 0 0 0 1
0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 1 0
1 0 0 1 0
0 0 0 0 0 0 1 0 0
1 0 0 1 0
0 0 0 0 0 0 0 1 0
1 0 0 1 0
0 0 0 0 0 0 0 1 0
0 1 0 0 1
0 1 0 0 1
0 0 0 0 0 0 0 0 1
0 1 0 0 1
0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 1
0 0 1 1 1
0 0 0 0 0 0 0 0 1
0 0 1 1 1
00111
1 0 1 0 1 0 1 0 1
0 0 0 0 0 0 0 0 1
1 0 1 0 1 0 1 0 1
1 0 1 0 1 0 1 0 1
0 1 0 1 0 1 0 1 1
1 0 1 0 1 0 1 0 1
0 1 0 1 0 1 0 1 1
0 1 0 1 0 1 0 1 1
1 0 0 0 0 0 0 0 0
010101011
100000000
Fig. 3: Encoding matrices for two-sided SUICP(SNCS) given
in Example 5 and 6
i=1 j=0
for s ∈ [0 : Ka − 1].
(26)
A. Optimal scalar linear index codes for two-sided
SUICP(SNCS)
In the construction given in Theorem 1, if U + 1 divides
both K and D − U , then we have a = U + 1 and ua = 1. In
this case, (26) reduces to
ys =
1 X
U
X
x(U+1)(s+1−i)+j,i
i=1 j=0
=
U
X
j=0
x(U+1)s+j
for s ∈ [0 : Ka ].
(27)
Hence, in this case, the index code given by (11) is an
optimal length scalar linear index code.
Example 7. Consider the two-sided SUICP(SNCS) with K =
14, D = 3 and U = 1. For this SUICP(SNCS), we have
gcd(K, D − U, U + 1) = U + 1 = 2. The optimal length
index code for this two-sided SUICP(SNCS) is a scalar
linear index code. For this two-sided SUICP(SNCS), we have
Ka = 7, ∆a = 1. An optimal length encoding matrix for this
one-sided SUICP(SNCS) is a AIR matrix of size 7 × 6. AIR
matrix of size 7 × 6 is given below.
1 0 0 0 0 0
0 1 0 0 0 0
0 0 1 0 0 0
L7×6 = 0 0 0 1 0 0
0 0 0 0 1 0
000001
111111
The scalar linear index code C for the one-sided
SUICP(SNCS) with Ka = 11 and ∆a = 3 is given below.
The scalar linear index code C for this one-sided
SUICP(SNCS) is given below.
The VLIC C(2s) for the given two-sided SUICP(SNCS) is
obtained by replacing ys in C with xs,1 + xs−1,2 + xs−2,3
for s ∈ [0 : 10]. The VLIC C(2s) for the given two-sided
SUICP(SNCS) is given below.
C = {c0 , c1 , c2 , c3 , c4 , c5 } = {y0 + y6 ,
y2 + y6 ,
y4 + y6 ,
C = {y0 + y8 , y1 + y9 , y2 + y10 ,
y3 + y8 , y4 + y9 , y5 + y10 ,
y6 + y8 + y10 , y7 + y9 + y10 }.
y1 + y6 ,
y3 + y6 ,
y5 + y6 }.
The scalar linear index code C(2s) for the given two-sided
SUICP(SNCS) is obtained by replacing ys in C with x2s +
x2s+1 for s ∈ [0 : 6]. The scalar linear index code C(2s) is
given below.
C(2s) = {x0 + x1 + x12 + x13 ,
x4 + x5 + x12 + x13 ,
x8 + x9 + x12 + x13 ,
C(2s) = { x0,1 + x10,2 + x9,3 + x8,1 + x7,2 + x6,3 ,
{z
} |
{z
}
|
y0
y8
x1,1 + x0,2 + x10,3 + x9,1 + x8,2 + x7,3 ,
{z
} |
{z
}
|
y1
y9
x2,1 + x1,2 + x0,3 + x10,1 + x9,2 + x8,3 ,
|
{z
} |
{z
}
y2
x2 + x3 + x12 + x13 ,
x6 + x7 + x12 + x13 ,
x10 + x11 + x12 + x13 }.
B. Optimal (U + 1) dimensional VLICs for two-sided
SUICP(SNCS)
In the construction given in Theorem 1, if gcd(K, D −
U, U + 1) = 1, then we have a = 1 and ua = U + 1. In
this case, (26) reduces to
y10
x3,1 + x2,2 + x1,3 + x8,1 + x7,2 + x6,3 ,
{z
} |
{z
}
|
y3
y8
x4,1 + x3,2 + x2,3 + x9,1 + x8,2 + x7,3 ,
|
{z
} |
{z
}
y4
y9
x5,1 + x4,2 + x3,3 + x10,1 + x9,2 + x8,3 ,
{z
} |
{z
}
|
y5
y10
x6,1 + x5,2 + x4,3 + x8,1 + x7,2 + x6,3
{z
} |
{z
}
|
y6
ys =
U+1
0
XX
x(s+1−i)+j,i
y10
i=1 j=0
=
U+1
X
xs+1−i,i for s ∈ [0 : K − 1].
(28)
i=1
Hence, in this case, the index code given by (11) is an
optimal length (U + 1) dimensional VLIC.
Example 8. Consider the two-sided SUICP(SNCS) with K =
11, D = 5, U = 2. For this SUICP(SNCS), gcd(K, D −
U, U + 1) = a = 1. The optimal length index code is a
three dimensional VLIC. For this SUICP(SNCS), we have
Ka = K = 11 and ∆a = 3. The AIR matrix of size L11×8 is
given below.
L11×8
y8
+ x10,1 + x9,2 + x8,3 ,
|
{z
}
1 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0
0 0 0 0 0 1 0 0
=
0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 1
1 0 0 1 0 0 1 0
01001001
00100111
x7,1 + x6,2 + x5,3 + x9,1 + x8,2 + x7,3
{z
} |
{z
}
|
y7
y9
+ x10,1 + x9,2 + x9,3 }.
|
{z
}
y10
C. Optimal scalar
SUICP(SNCS)
linear
index
codes
for
one-sided
In the construction given in Theorem 1, if U = 0, then we
have, a = gcd(K, D − U, U + 1) = 1 and ua = U + 1 = 1.
In this case, (26) reduces to ys = xs for s ∈ [0 : K − 1].
Hence, in this case, the index code given by (11) is an
optimal length scalar linear index code.
Example 9. Consider the two-sided SUICP(SNCS) with K =
7, D = 2 and U = 0. For this SUICP(SNCS), we have
gcd(K, D − U, U + 1) = 1. The optimal length index code
for this SUICP(SNCS) is scalar linear index code. For this
SUICP(SNCS), we have K = Ka = 7, D = ∆a = 2. An
optimal length encoding matrix for this SUICP(SNCS) is a
AIR matrix of size 7 × 5. AIR matrix of size 7 × 5 is given
below.
L7×5
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
= 0 0 0 1 0
0 0 0 0 1
10101
01011
The scalar linear index code C(2s) for this SUICP(SNCS)
is given below.
C(2s) = C = {x0 + x5 ,
x3 + x6 ,
U+1
D. Optimal D−U
SUICP(SNCS)
x1 + x6 , x2 + x5 ,
x4 + x5 + x6 }.
dimensional
VLICs
for
two-sided
In the construction given in Theorem 1, if gcd(K, D −
K
U+1
, Ka = D−U
U, U + 1) = D − U, then we have ua = D−U
and ∆a = 1. Hence, in this case, the AIR matrix used is
always of dimension Ka × (Ka − 1) and the index code given
U+1
by (11) is an optimal length D−U
dimensional VLIC. In [6],
it is shown that the message probability of error in decoding
a message at a particular receiver decreases with a decrease
in the number of transmissions used to decode the message
among the total of broadcast transmissions. In this case, as the
AIR matrix size is always of the form Ka × (Ka − 1), the
low complexity decoding given in [9] enables each receiver
to decode its wanted message symbol by using atmost two
broadcast symbols.
Example 10. Consider the two-sided SUICP(SNCS) with
K = 12, D = 5 and U = 3. For this SUICP(SNCS), we
have gcd(K, D − U, U + 1) = D − U = 2. The optimal
length index code for this two-sided SUICP(SNCS) is a two
dimensional VLIC. For this two-sided SUICP(SNCS), we have
Ka = 6, ∆a = 1. An optimal length encoding matrix for this
one-sided SUICP(SNCS) is a AIR matrix of size 6 × 5. AIR
matrix of size 6 × 5 is given below.
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
L6×5 = 0 0 0 1 0
0 0 0 0 1
00000
11111
The scalar linear index code C for this one-sided
SUICP(SNCS) is given below.
C = {c0 , c1 .c2 , c3 , c4 } = {y0 + y5 ,
y3 + y5 ,
y1 + y5 ,
y2 + y5 ,
y4 + y5 }.
The scalar linear index code C(2s) for the given two-sided
SUICP(SNCS) is obtained by replacing ys in C with x2s,1 +
x2s−2,2 + x2s+1,1 + x2s−1,2 for s ∈ [0 : 5].
Example 11. Consider the two-sided SUICP(SNCS) with
K = 27, D = 8 and U = 5. For this SUICP(SNCS), we
have gcd(K, D − U, U + 1) = D − U = 3. The optimal
length index code for this two-sided SUICP(SNCS) is a two
dimensional VLIC. For this two-sided SUICP(SNCS), we have
Ka = 9, ∆a = 1. An optimal length encoding matrix for this
one-sided SUICP(SNCS) is a AIR matrix of size 9 × 8. AIR
matrix of size 9 × 8 is given below.
10000000
0 1 0 0 0 0 0 0
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0
L9×8 =
0 0 0 0 0 1 0 0
0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 1
11111111
The scalar linear index code C for this one-sided
SUICP(SNCS) is given below.
C = {y0 + y8 ,
y1 + y8 ,
y2 + y8 ,
y3 + y8 ,
y4 + y8 ,
y5 + y8 ,
y6 + y8 ,
y7 + y8 }.
The two dimensional VLIC C(2s) for the given two-sided
SUICP(SNCS) is obtained by replacing ys in C with x3s,1 +
x3s−3,2 + x3s+1,1 + x3s−2,2 + x3s+2,1 + x3s−1,2 for s ∈ [0 : 8].
ACKNOWLEDGMENT
This work was supported partly by the Science and Engineering Research Board (SERB) of Department of Science and
Technology (DST), Government of India, through J.C. Bose
National Fellowship to B. Sundar Rajan.
R EFERENCES
[1] H. Maleki, V. Cadambe, and S. Jafar, “Index coding – an interference
alignment perspective”, in IEEE Trans. Inf. Theory,, vol. 60, no.9,
pp.5402-5432, Sep. 2014.
[2] S. A. Jafar, “Topological interference management through index
coding,” IEEE Trans. Inf. Theory, vol. 60, no. 1, pp. 529–568, Jan.
2014.
[3] Y. Birk and T. Kol, “Coding-on-demand by an informed-source (ISCOD) for efficient broadcast of different supplemental data to caching
clients”, in IEEE Trans. Inf. Theory,, vol. 52, no.6, pp.2825-2830, June
2006.
[4] L Ong and C K Ho, “Optimal Index Codes for a Class of Multicast
Networks with Receiver Side-information”, in Proc. IEEE ICC, 2012,
pp. 2213-2218.
[5] Z. Bar-Yossef, Z. Birk, T. S. Jayram, and T. Kol, “Index coding with
side-information”, in IEEE Trans. Inf. Theory,, vol. 57, no.3, pp.14791494, Mar. 2011.
[6] A. Thomas, R. Kavitha, A. Chandramouli and B. S. Rajan, “Single
uniprior index coding with min-max probability of error over fading
channels,” in Proc. IEEE Trans. on Veh. Technology July 2017.
[7] M. B. Vaddi and B. S. Rajan, “Optimal vector linear index codes
for some symmetric multiple unicast problems,” in Proc. IEEE ISIT,
Barcelona, Spain July 2016, pp. 125-129.
[8] M. B. Vaddi and B. S. Rajan, “Optimal scalar linear index codes
for one-sided neighboring side-information problems,” In Proc. IEEE
GLOBECOM Workshop on Network Coding and Applications, Washington, USA, December 2016.
[9] M. B. Vaddi and B. S. Rajan, “Low-complexity decoding for symmetric, neighboring and consecutive side-information index coding
problems,” in arXiv:1705.03192v2 [cs.IT] 16 May 2017.
| 7cs.IT
|
arXiv:1701.02096v2 [cs.CV] 6 Nov 2017
Improved Texture Networks: Maximizing Quality and Diversity in
Feed-forward Stylization and Texture Synthesis
Dmitry Ulyanov
Skolkovo Institute of Science and Technology & Yandex
Andrea Vedaldi
University of Oxford
[email protected]
[email protected]
Victor Lempitsky
Skolkovo Institute of Science and Technology
[email protected]
Abstract
The recent work of Gatys et al., who characterized the
style of an image by the statistics of convolutional neural
network filters, ignited a renewed interest in the texture generation and image stylization problems. While their image
generation technique uses a slow optimization process, recently several authors have proposed to learn generator
neural networks that can produce similar outputs in one
quick forward pass. While generator networks are promising, they are still inferior in visual quality and diversity
compared to generation-by-optimization. In this work, we
advance them in two significant ways. First, we introduce
an instance normalization module to replace batch normalization with significant improvements to the quality of image
stylization. Second, we improve diversity by introducing
a new learning formulation that encourages generators to
sample unbiasedly from the Julesz texture ensemble, which
is the equivalence class of all images characterized by certain filter responses. Together, these two improvements take
feed forward texture synthesis and image stylization much
closer to the quality of generation-via-optimization, while
retaining the speed advantage.
(a) Panda.
(b)
(c)
(d) Style.
(e)
(f)
Figure 1: Which panda stylization seems the best to you?
Definitely not the variant (b), which has been produced
by a state-of-the-art algorithm among methods that take no
longer than a second. The (e) picture took several minutes
to generate using an optimization process, but the quality
is worth it, isn’t it? We would be particularly happy if you
chose one from the rightmost two examples, which are computed with our new method that aspires to combine the quality of the optimization-based method and the speed of the
fast one. Moreover, our method is able to produce diverse
stylizations using a single network.
1. Introduction
The recent work of Gatys et al. [4, 5], which used deep
neural networks for texture synthesis and image stylization
to a great effect, has created a surge of interest in this area.
Following an earlier work by Portilla and Simoncelli [15],
they generate an image by matching the second order moments of the response of certain filters applied to a reference
texture image. The innovation of Gatys et al. is to use non-
linear convolutional neural network filters for this purpose.
Despite the excellent results, however, the matching process is based on local optimization, and generally requires
a considerable amount of time (tens of seconds to minutes)
in order to generate a single textures or stylized image.
In order to address this shortcoming, Ulyanov et al. [19]
and Johnson et al. [8] suggested to replace the optimization process with feed-forward generative convolutional
networks. In particular, [19] introduced texture networks
The source code is available at https://github.com/
DmitryUlyanov/texture_nets
1
to generate textures of a certain kind, as in [4], or to apply
a certain texture style to an arbitrary image, as in [5]. Once
trained, such texture networks operate in a feed-forward
manner, three orders of magnitude faster than the optimization methods of [4, 5].
The price to pay for such speed is a reduced performance. For texture synthesis, the neural network of [19]
generates good-quality samples, but these are not as diverse as the ones obtained from the iterative optimization
method of [4]. For image stylization, the feed-forward results of [19, 8] are qualitatively and quantitatively worse
than iterative optimization. In this work, we address both
limitations by means of two contributions, both of which
extend beyond the applications considered in this paper.
Our first contribution (section 4) is an architectural
change that significantly improves the generator networks. The change is the introduction of an instancenormalization layer which, particularly for the stylization
problem, greatly improves the performance of the deep network generators. This advance significantly reduces the gap
in stylization quality between the feed-forward models and
the original iterative optimization method of Gatys et al.,
both quantitatively and qualitatively.
Our second contribution (section 3) addresses the limited diversity of the samples generated by texture networks.
In order to do so, we introduce a new formulation that
learns generators that uniformly sample the Julesz ensemble [20]. The latter is the equivalence class of images that match certain filter statistics. Uniformly sampling
this set guarantees diverse results, but traditionally doing
so required slow Monte Carlo methods [20]; Portilla and
Simoncelli, and hence Gatys et al., cannot sample from
this set, but only find individual points in it, and possibly
just one point. Our formulation minimizes the KullbackLeibler divergence between the generated distribution and
a quasi-uniform distribution on the Julesz ensemble. The
learning objective decomposes into a loss term similar to
Gatys et al. minus the entropy of the generated texture samples, which we estimate in a differentiable manner using a
non-parametric estimator [12].
We validate our contributions by means of extensive
quantitative and qualitative experiments, including comparing the feed-forward results with the gold-standard
optimization-based ones (section 5). We show that, combined, these ideas dramatically improve the quality of feedforward texture synthesis and image stylization, bringing
them to a level comparable to the optimization-based approaches.
2. Background and related work
Julesz ensemble. Informally, a texture is a family of visual patterns, such as checkerboards or slabs of concrete,
that share certain local statistical regularities. The concept
was first studied by Julesz [9], who suggested that the visual system discriminates between different textures based
on the average responses of certain image filters.
The work of [20] formalized Julesz’ ideas by introducing the concept of Julesz ensemble. There, an image is
a real function x : Ω → R3 defined on a discrete lattice Ω = {1, . . . , H} × {1, . . . , W } and a texture is a
distribution p(x) over such images. The local statistics
of an image are captured by a bank of (non-linear) filters
Fl : X × Ω → R, l = 1, . . . , L, where Fl (x, u) denotes the
response of filter Fl at location u on image x. The image x
is characterized
by the spatial average of the filter responses
P
µl (x) = u∈Ω Fl (x, u)/|Ω|. The image is perceived as a
particular texture if these responses match certain characteristic values µ̄l . Formally, given the loss function,
L(x) =
L
X
(µl (x) − µ̄l )2
(1)
l=1
the Julesz ensemble is the set of all texture images
T = {x ∈ X : L(x) ≤ }
that approximately satisfy such constraints. Since all textures in the Julesz ensemble are perceptually equivalent, it is
natural to require the texture distribution p(x) to be uniform
over this set. In practice, it is more convenient to consider
the exponential distribution
p(x) = R
e−L(x)/T
,
e−L(y)/T dy
(2)
where T > 0 is a temperature parameter. This choice is motivated as follows [20]: since statistics are computed from
spatial averages of filter responses, one can show that, in
the limit of infinitely large lattices, the distribution p(x) is
zero outside the Julesz ensemble and uniform inside. In
this manner, eq. (2) can be though as a uniform distribution over images that have a certain characteristic filter responses µ̄ = (µ̄1 , . . . , µ̄L ).
Note also that the texture is completely described by the
filter bank F = (F1 , . . . , FL ) and their characteristic responses µ̄. As discussed below, the filter bank is generally
fixed, so in this framework different textures are given by
different characteristics µ̄.
Generation-by-minimization. For any interesting choice
of the filter bank F , sampling from eq. (2) is rather challenging and classically addressed by Monte Carlo methods [20].
In order to make this framework more practical, Portilla
and Simoncelli [16] proposed instead to heuristically sample from the Julesz ensemble by the optimization process
x∗ = argmin L(x).
x∈X
(3)
If this optimization problem can be solved, the minimizer
x∗ is by definition a texture image. However, there is no
reason why this process should generate fair samples from
the distribution p(x). In fact, the only reason why eq. (3)
may not simply return always the same image is that the optimization algorithm is randomly initialized, the loss function is highly non-convex, and search is local. Only because
of this eq. (3) may land on different samples x∗ on different
runs.
Deep filter banks. Constructing a Julesz ensemble requires choosing a filter bank F . Originally, researchers considered the obvious candidates: Gaussian derivative filters,
Gabor filters, wavelets, histograms, and similar [20, 16, 21].
More recently, the work of Gatys et al. [4, 5] demonstrated
that much superior filters are automatically learned by deep
convolutional neural networks (CNNs) even when trained
for apparently unrelated problems, such as image classification. In this paper, in particular, we choose for L(x) the
style loss proposed by [4]. The latter is the distance between
the empirical correlation matrices of deep filter responses in
a CNN.1
Stylization. The texture generation method of Gatys et
al. [4] can be considered as a direct extension of the texture
generation-by-minimization technique (3) of Portilla and
Simoncelli [16]. Later, Gatys et al. [5] demonstrated that
the same technique can be used to generate an image that
mixes the statistics of two other images, one used as a texture template and one used as a content template. Content
is captured by introducing a second loss Lcont. (x, x0 ) that
compares the responses of deep CNN filters extracted from
the generated image x and a content image x0 . Minimizing
the combined loss L(x) + αLcont. (x, x0 ) yields impressive
artistic images, where a texture µ̄, defining the artistic style,
is fused with the content image x0 .
Feed-forward generator networks. For all its simplicity
and efficiency compared to Markov sampling techniques,
generation-by-optimization (3) is still relatively slow, and
certainly too slow for real-time applications. Therefore, in
the past few months several authors [8, 19] have proposed to
learn generator neural networks g(z) that can directly map
random noise samples z ∼ pz = N (0, I) to a local minimizer of eq. (3). Learning the neural network g amounts to
minimizing the objective
g ∗ = argmin E L(g(z)).
g
pz
(4)
1 Note that such matrices are obtained by averaging local non-linear
filters: these are the outer products of filters in a certain layer of the neural
network. Hence, the style loss of Gatys et al. is in the same form as eq. (1).
While this approach works well in practice, it shares the
same important limitation as the original work of Portilla
and Simoncelli: there is no guarantee that samples generated by g ∗ would be fair samples of the texture distribution (2). In practice, as we show in the paper, such samples
tend in fact to be not diverse enough.
Both [8, 19] have also shown that similar generator networks work also for stylization. In this case, the generator
g(x0 , z) is a function of the content image x0 and of the
random noise z. The network g is learned to minimize the
sum of texture loss and the content loss:
g ∗ = argmin E [L(g(x0 , z)) + αLcont. (g(x0 , z), x0 )].
g
px0 ,pz
(5)
Alternative neural generator methods. There are many
other techniques for image generation using deep neural
networks.
The Julesz distribution is closely related to the FRAME
maximum entropy model of [21], as well as to the concept
of Maximum Mean Discrepancy (MMD) introduced in [7].
Both FRAME and MMD make the observation that a probability distribution p(x) can be described by the expected
values µα = Ex∼p(x) [φα (x)] of a sufficiently rich set of
statistics φα (x). Building on these ideas, [14, 3] construct
generator neural networks g with the goal of minimizing the
discrepancy between
PNthe statistics averaged over a batch of
generated images i=1 φα (g(zi ))/N and the statistics avPM
eraged over a training set i=1 φα (xi )/M . The resulting
networks g are called Moment Matching Networks (MMN).
An important alternative methodology is based on the
concept of Generative Adversarial Networks (GAN; [6]).
This approach trains, together with the generator network
g(z), a second adversarial network f (·) that attempts to distinguish between generated samples g(z), z ∼ N (0, I) and
real samples x ∼ pdata (x). The adversarial model f can be
used as a measure of quality of the generated samples and
used to learn a better generator g. GAN are powerful but notoriously difficult to train. A lot of research is has recently
focused on improving GAN or extending it. For instance,
LAPGAN [2] combines GAN with a Laplacian pyramid and
DCGAN [17] optimizes GAN for large datasets.
3. Julesz generator networks
This section describes our first contribution, namely a
method to learn networks that draw samples from the Julesz
ensemble modeling a texture (section 2), which is an intractable problem usually addressed by slow Monte Carlo
methods [21, 20]. Generation-by-optimization, popularized
by Portilla and Simoncelli and Gatys et al., is faster, but
can only find one point in the ensemble, not sample from it,
with scarce sample diversity, particularly when used to train
feed-forward generator networks [8, 19].
Here, we propose a new formulation that allows to train
generator networks that sample the Julesz ensemble, generating images with high visual fidelity as well as high diversity.
A generator network [6] maps an i.i.d. noise vector
z ∼ N (0, I) to an image x = g(z) in such a way that
x is ideally a sample from the desired distribution p(x).
Such generators have been adopted for texture synthesis
in [19], but without guarantees that the learned generator
g(z) would indeed sample a particular distribution.
Here, we would like to sample from the Gibbs distribution (2) defined over the Julesz ensemble. This distribution can beR written compactly as p(x) = Z −1 e−L(x)/T ,
where Z = e−L(x)/T dx is an intractable normalization
constant.
Denote by q(x) the distribution induced by a generator
network g. The goal is to make the target distribution p and
the generator distribution q as close as possible by minimizing their Kullback-Leibler (KL) divergence:
Z
q(x)Z
dx
KL(q||p) = q(x) ln
p(x)
1
=
E L(x) + E ln q(x) + ln(Z) (6)
T x∼q(x)
x∼q(x)
1
E L(x) − H(q) + const.
=
T x∼q(x)
Hence, the KL divergence is the sum of the expected value
of the style loss L and the negative entropy of the generated
distribution q.
The first term can be estimated by taking the expectation
over generated samples:
E
x∼q(x)
L(x) =
E
z∼N (0,I)
L(g(z)).
(7)
This is similar to the reparametrization trick of [11] and is
also used in [8, 19] to construct their learning objectives.
The second term, the negative entropy, is harder to estimate accurately, but simple estimators exist. One which
is particularly appealing in our scenario is the KozachenkoLeonenko estimator [12]. This estimator considers a batch
of N samples x1 , . . . , xn ∼ q(x). Then, for each sample
xi , it computes the distance ρi to its nearest neighbour in
the batch:
ρi = min kxi − xj k.
(8)
j6=i
The distances ρi can be used to approximate the entropy as
follows:
N
DX
ln ρi + const.
(9)
H(q) ≈
N i=1
where D = 3W H is the number of components of the images x ∈ R3×W ×H .
An energy term similar to (6) was recently proposed
in [10] for improving the diversity of a generator network
in a adversarial learning scheme. While the idea is superficially similar, the application (sampling the Julesz ensemble) and instantiation (the way the entropy term is implemented) are very different.
Learning objective. We are now ready to define an objective function E(g) to learn the generator network g. This
is given by substituting the expected loss (7) and the entropy estimator (9), computed over a batch of N generated
images, in the KL divergence (6):
E(g) =
N
1 Xh 1
L(g(zi ))
N i=1 T
i
− λ ln min kg(zi ) − g(zj )k
j6=i
(10)
The batch itself is obtained by drawing N samples
z1 , . . . , zn ∼ N (0, I) from the noise distribution of the
generator. The first term in eq. (10) measures how closely
the generated images g(zi ) are to the Julesz ensemble. The
second term quantifies the lack of diversity in the batch by
mutually comparing the generated images.
Learning. The loss function (10) is in a form that allows optimization by means of Stochastic Gradient Descent
(SGD). The algorithm samples a batch z1 , . . . , zn at a time
and then descends the gradient:
N
1 X h dL dg(zi )
N i=1 dx> dθ>
λ
− (g(zi ) − g(zji∗ ))>
ρi
dg(zi ) dg(zji∗ )
−
dθ>
dθ>
i
(11)
where θ is the vector of parameters of the neural network g,
the tensor image x has been implicitly vectorized and ji∗ is
the index of the nearest neighbour of image i in the batch.
4. Stylization with instance normalization
The work of [19] showed that it is possible to learn
high-quality texture networks g(z) that generate images in
a Julesz ensemble. They also showed that it is possible to
learn good quality stylization networks g(x0 , z) that apply
the style of a fixed texture to an arbitrary content image x0 .
Nevertheless, the stylization problem was found to be
harder than the texture generation one. For the stylization
task, they found that learning the model from too many example content images x0 , say more than 16, yielded poorer
qualitative results than using a smaller number of such examples. Some of the most significant errors appeared along
Figure 2: Comparison of normalization techniques in image
stylization. From left to right: BN, cross-channel LRN at
the first layer, IN at the first layer, IN throughout.
16.0
Batch
Instance
15.5
15.0
14.5
14.0
13.5
13.0
16.5
16.0
15.5
15.0
14.5
14.0
13.5
13.0
Batch
Instance
Random
ln L(x)
ln L(x)
the border of the generated images, probably due to padding
and other boundary effects in the generator network. We
conjectured that these are symptoms of a learning problem
too difficult for their choice of neural network architecture.
A simple observation that may make learning simpler is
that the result of stylization should not, in general, depend
on the contrast of the content image but rather should match
the contrast of the texture that is being applied to it. Thus,
the generator network should discard contrast information
in the content image x0 . We argue that learning to discard
contrast information by using standard CNN building block
is unnecessarily difficult, and is best done by adding a suitable layer to the architecture.
To see why, let x ∈ RN ×C×W ×H be an input tensor
containing a batch of N images. Let xnijk denote its nijkth element, where k and j span spatial dimensions, i is the
feature channel (i.e. the color channel if the tensor is an
RGB image), and n is the index of the image in the batch.
Then, contrast normalization is given by:
0
3000
6000
9000
12000
Iteration
0
50
100 150 200 250 300
Iteration
(a) Feed-forward history.
(b) Finetuning history.
xnijk − µni
,
ynijk = p 2
σni +
µni =
W
H
1 XX
xnilm ,
HW
m=1
(c) Content.
(d) StyleNet IN.
(e) IN finetuned.
(f) Style.
(g) StyleNet BN.
(h) BN finetuned.
(12)
l=1
2
σni
=
W
H
1 XX
(xnilm − µni )2 .
HW
m=1
l=1
It is unclear how such as function could be implemented as
a sequence of standard operators such as ReLU and convolution.
On the other hand, the generator network of [19] does
contain a normalization layers, and precisely batch normalization (BN) ones. The key difference between eq. (12) and
batch normalization is that the latter applies the normalization to a whole batch of images instead of single ones:
xnijk − µi
ynijk = p 2
,
σi +
µi =
N X
W X
H
X
1
xnilm ,
HW N n=1
m=1
(13)
l=1
σi2 =
N X
W X
H
X
1
(xnilm − µi )2 .
HW N n=1
m=1
l=1
We argue that, for the purpose of stylization, the normalization operator of eq. (12) is preferable as it can normalize
each individual content image x0 .
While some authors call layer eq. (12) contrast normalization, here we refer to it as instance normalization (IN)
since we use it as a drop-in replacement for batch normalization operating on individual instances instead of the
Figure 3: (a) learning objective as a function of SGD iterations for StyleNet IN and BN. (b) Direct optimization of
the Gatys et al. for this example image starting from the result of StyleNet IN and BN. (d,g) Result of StyleNet with
instance (d) and batch normalization (g). (e,h) Result of
finetuing the Gatys et al. energy.
batch as a whole. Note in particular that this means that instance normalization is applied throughout the architecture,
not just at the input image—fig. 2 shows the benefit of doing
so.
Another similarity with BN is that each IN layer is followed by a scaling and bias operator s x + b. A difference
is that the IN layer is applied at test time as well, unchanged,
whereas BN is usually switched to use accumulated mean
and variance instead of computing them over the batch.
IN appears to be similar to the layer normalization
method introduced in [1] for recurrent networks, although
it is not clear how they handle spatial data. Like theirs, IN
is a generic layer, so we tested it in classification problems
as well. In such cases, it still work surprisingly well, but not
as well as batch normalization (e.g. AlexNet [13] IN has 2-
3% worse top-1 accuracy on ILSVRC [18] than AlexNet
BN).
5. Experiments
In this section, after discussing the technical details of
the method, we evaluate our new texture network architectures using instance normalization, and then investigate the
ability of the new formulation to learn diverse generators.
5.1. Technical details
Network architecture. Among two generator network
architectures, proposed previously in [19, 8], we choose the
residual architecture from [8] for all our style transfer experiments. We also experimented with architecture from [19]
and observed a similar improvement with our method, but
use the one from [8] for convenience. We call it StyleNet
with a postfix BN if it is equipped with batch normalization
or IN for instance normalization.
For texture synthesis we compare two architectures: the
multi-scale fully-convolutional architecture from [19] (TextureNetV1) and the one we design to have a very large receptive field (TextureNetV2). TextureNetV2 takes a noise
vector of size 256 and first transforms it with two fullyconnected layers. The output is then reshaped to a 4 × 4
image and repeatedly upsampled with fractionally-strided
convolutions similar to [17]. More details can be found in
the supplementary material.
Weight parameters. In practice, for the case of λ > 0,
entropy loss and texture loss in eq. (10) should be weighted
properly. As only the value of T λ is important for optimization we assume λ = 1 and choose T from the set of
three values (5, 10, 20) for texture synthesis (we pick the
higher value among those not leading to artifacts – see our
discussion below). We fix T = 10000 for style transfer experiments. For texture synthesis, similarly to [19], we found
useful to normalize gradient of the texture loss as it passes
back through the VGG-19 network. This allows rapid convergence for stochastic optimization but implicitly alters the
objective function and requires temperature to be adjusted.
We observe that for textures with flat lightning high entropy
weight results in brightness variations over the image fig. 7.
We hypothesize this issue can be solved if either more clever
distance for entropy estimation is used or an image prior introduced.
5.2. Effect of instance normalization
In order to evaluate the impact of replacing batch normalization with instance normalization, we consider first
the problem of stylization, where the goal is to learn a generator x = g(x0 , z) that applies a certain texture style to
the content image x0 using noise z as “random seed”. We
set λ = 0 for which generator is most likely to discard the
noise.
The StyleNet IN and StyleNet BN are compared in fig. 3.
Panel fig. 3.a shows the training objective (5) of the networks as a function of the SGD training iteration. The
objective function is the same, but StyleNet IN converges
much faster, suggesting that it can solve the stylization
problem more easily. This is confirmed by the stark difference in the qualitative results in panels (d) end (g). Since
the StyleNets are trained to minimize in one shot the same
objective as the iterative optimization of Gatys et al., they
can be used to initialize the latter algorithm. Panel (b) shows
the result of applying the Gatys et al. optimization starting
from their random initialization and the output of the two
StyleNets. Clearly both networks start much closer to an
optimum than random noise, and IN closer than BN. The
difference is qualitatively large: panels (e) and (h) show the
change in the StyleNets output after finetuning by iterative
optimization of the loss, which has a small effect for the IN
variant, and a much larger one for the BN one.
Similar results apply in general. Other examples are
shown in fig. 4, where the IN variant is far superior to BN
and much closer to the results obtained by the much slower
iterative method of Gatys et al. StyleNets are trained on images of a fixed sized, but since they are convolutional, they
can be applied to arbitrary sizes. In the figure, the top tree
images are processed at 512 × 512 resolution and the bottom two at 1024 × 1024. In general, we found that higher
resolution images yield visually better stylization results.
While instance normalization works much better than
batch normalization for stylization, for texture synthesis the
two normalization methods perform equally well. This is
consistent with our intuition that IN helps in normalizing
the information coming from content image x0 , which is
highly variable, whereas it is not important to normalize the
texture information, as each model learns only one texture
style.
5.3. Effect of the diversity term
Having validated the IN-based architecture, we evaluate
now the effect of the entropy-based diversity term in the
objective function (10).
The experiment in fig. 5 starts by considering the
problem of texture generation. We compare the new
high-capacity TextureNetV2 and the low-capacity TextureNetsV1 texture synthesis networks. The low-capacity
model is the same as [19]. This network was used there
in order to force the network to learn a non-trivial dependency on the input noise, thus generating diverse outputs
even though the learning objective of [19], which is the
same as eq. (10) with diversity coefficient λ = 0, tends to
suppress diversity. The results in fig. 5 are indeed diverse,
but sometimes of low quality. This should be contrasted
Content
StyleNet IN (ours)
StyleNet BN
Gatys et al.
Style
Figure 4: Stylization results obtained by applying different textures (rightmost column) to different content images (leftmost
column). Three methods are compared: StyleNet IN, StyleNet BN, and iterative optimization.
with TextureNetV2, the high-capacity model: its visual fidelity is much higher, but, by using the same objective function [19], the network learns to generate a single image, as
expected. TextureNetV2 with the new diversity-inducing
objective (λ > 0) is the best of both worlds, being both
high-quality and diverse.
The experiment in fig. 6 assesses the effect of the diversity term in the stylization problem. The results are similar
to the ones for texture synthesis and the diversity term effectively encourages the network to learn to produce different
results based on the input noise.
One difficultly with texture and stylization networks is
that the entropy loss weight λ must be tuned for each
learned texture model. Choosing λ too small may fail to
learn a diverse generator, and setting it too high may create
artifacts, as shown in fig. 7.
6. Summary
This paper advances feed-forward texture synthesis and
stylization networks in two significant ways. It introduces
instance normalization, an architectural change that makes
training stylization networks easier and allows the training
process to achieve much lower loss levels. It also introduces
a new learning formulation for training generator networks
Input
TextureNetV2 λ = 0
TextureNetV2 λ > 0 (ours)
TextureNetV1 λ = 0
Figure 5: The textures generated by the high capacity Texture Net V2 without diversity term (λ = 0 in eq. (10)) are nearly
identical. The low capacity TextureNet V1 of [19] achieves diversity, but has sometimes poor results. TextureNet V2 with
diversity is the best of both worlds.
Content
Style
StyleNet λ > 0
StyleNet λ = 0
Figure 6: The StyleNetV2 g(x0 , z), trained with diversity λ > 0, generates substantially different stylizations for different
values of the input noise z. In this case, the lack of stylization diversity is visible in uniform regions such as the sky.
to sample uniformly from the Julesz ensemble, thus explicitly encouraging diversity in the generated outputs. We
show that both improvements lead to noticeable improvements of the generated stylized images and textures, while
keeping the generation run-times intact.
References
[1] L. J. Ba, R. Kiros, and G. E. Hinton. Layer normalization.
CoRR, abs/1607.06450, 2016. 5
[2] E. L. Denton, S. Chintala, A. Szlam, and R. Fergus. Deep
generative image models using a laplacian pyramid of adversarial networks. In NIPS, pages 1486–1494, 2015. 3
Figure 7: Negative examples. If the diversity term λ is too
high for the learned style, the generator tends to generate
artifacts in which brightness is changed locally (spotting)
instead of (or as well as) changing the structure.
[3] G. K. Dziugaite, D. M. Roy, and Z. Ghahramani. Training
generative neural networks via maximum mean discrepancy
optimization. In UAI, pages 258–267. AUAI Press, 2015. 3
[4] L. Gatys, A. S. Ecker, and M. Bethge. Texture synthesis
using convolutional neural networks. In Advances in Neural Information Processing Systems, NIPS, pages 262–270,
2015. 1, 2, 3
[5] L. A. Gatys, A. S. Ecker, and M. Bethge. A neural algorithm
of artistic style. CoRR, abs/1508.06576, 2015. 1, 2, 3
[6] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
D. Warde-Farley, S. Ozair, A. C. Courville, and Y. Bengio.
Generative adversarial nets. In Advances in Neural Information Processing Systems (NIPS), pages 2672–2680, 2014. 3,
4
[7] A. Gretton, K. M. Borgwardt, M. Rasch, B. Schölkopf, and
A. J. Smola. A kernel method for the two-sample-problem.
In Advances in neural information processing systems,NIPS,
pages 513–520, 2006. 3
[8] J. Johnson, A. Alahi, and L. Fei-Fei. Perceptual losses for
real-time style transfer and super-resolution. In Computer
Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings,
Part II, pages 694–711, 2016. 1, 2, 3, 4, 6
[9] B. Julesz. Textons, the elements of texture perception, and
their interactions. Nature, 290(5802):91–97, 1981. 2
[10] T. Kim and Y. Bengio. Deep directed generative models
with energy-based probability estimation. arXiv preprint
arXiv:1606.03439, 2016. 4
[11] D. P. Kingma and M. Welling. Auto-encoding variational
bayes. CoRR, abs/1312.6114, 2013. 4
[12] L. F. Kozachenko and N. N. Leonenko. Sample estimate of
the entropy of a random vector. Probl. Inf. Transm., 23(12):95–101, 1987. 2, 4
[13] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
NIPS, pages 1106–1114, 2012. 5
[14] Y. Li, K. Swersky, and R. S. Zemel. Generative moment
matching networks. In Proc. International Conference on
Machine Learning, ICML, pages 1718–1727, 2015. 3
[15] J. Portilla and E. P. Simoncelli. A parametric texture model
based on joint statistics of complex wavelet coefficients.
IJCV, 40(1):49–70, 2000. 1
[16] J. Portilla and E. P. Simoncelli. A parametric texture model
based on joint statistics of complex wavelet coefficients.
IJCV, 2000. 2, 3
[17] A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR, abs/1511.06434, 2015. 3, 6
[18] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual
Recognition Challenge. International Journal of Computer
Vision (IJCV), 115(3):211–252, 2015. 5
[19] D. Ulyanov, V. Lebedev, A. Vedaldi, and V. S. Lempitsky.
Texture networks: Feed-forward synthesis of textures and
stylized images. In Proceedings of the 33nd International
Conference on Machine Learning, ICML 2016, New York
City, NY, USA, June 19-24, 2016, pages 1349–1357, 2016.
1, 2, 3, 4, 5, 6, 8
[20] S. C. Zhu, X. W. Liu, and Y. N. Wu. Exploring texture
ensembles by efficient markov chain monte carlotoward a
atrichromacyo theory of texture. PAMI, 2000. 2, 3
[21] S. C. Zhu, Y. Wu, and D. Mumford. Filters, random fields
and maximum entropy (FRAME): Towards a unified theory
for texture modeling. IJCV, 27(2), 1998. 3
| 1cs.CV
|
arXiv:1605.06909v3 [math.OA] 13 Sep 2017
UNITARIZABILITY, MAUREY-NIKISHIN FACTORIZATION,
AND POLISH GROUPS OF FINITE TYPE
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
Abstract. Let Γ be a countable discrete group, and π : Γ → GL(H) be a representation
of Γ by invertible operators on a separable Hilbert space H. We show that the semidirect
product group G = H ⋊π Γ is SIN (G admits a two-sided invariant metric compatible with
its topology) and unitarily representable (G embeds into the unitary group U(ℓ2 (N))), if and
only if π is uniformly bounded, and that π is unitarizable if and only if G is of finite type:
that is, G embeds into the unitary group of a II1 -factor. Consequently, we show that a
unitarily representable Polish SIN group need not be of finite type, answering a question of
Sorin Popa. The key point in our argument is an equivariant version of the Maurey–Nikishin
factorization theorem for continuous maps from a Hilbert space to the space L0 (X, m) of all
measurable maps on a probability space.
1. Introduction
In this paper, we continue our study [AM12-2] of Popa’s class Ufin of finite type Polish
groups. Here, a Polish group G is said to be of finite type if it is isomorphic as a topological
group to a closed subgroup of the unitary group U (M ) of some finite von Neumann algebra
M with separable predual, where U (M ) is endowed with the strong operator topology. The
class Ufin was introduced in [Po07], in which Popa proved the celebrated Cocycle Superrigidity
Theorem. This theorem asserts that if Γ is a w-rigid group and α : Γ y (X, µ) is a probability
measure preserving (p.m.p) action of Γ on a standard probability space (X, µ) with certain
properties, then any 1-cocycle w : X × Γ → G for α with values in a Polish group G ∈ Ufin
is cohomologous to a group homomorphism Γ → G, see [Po07] for details. Despite the
importance of Popa’s Theorem and its wide range of applications, little has been known
about the properties of the class Ufin . There are two necessary conditions for a Polish group
to be in Ufin . Namely, if G ֒→ U (M ) for some finite von Neumann algebra (M, τ ) with
separable predual, then clearly G ֒→ U (ℓ2 (N)). G is said to be unitarily representable if the
abovementioned condition is satisfied. Moreover, the restriction of the 2-norm metric d on
U (M ) given by
1
d(u, v) := ku − vk2 , kxk2 := τ (x∗ x) 2 (u, v ∈ U (M ))
defines a two-sided invariant metric on G that is compatible with the topology. It is wellknown that a metrizable topological group admits a two-sided invariant metric compatible
with the topology if and only if the group is SIN (see Definition 2.3). In [Po07, §6.5], Popa
questioned whether the abovementioned two conditions are sufficient:
Question (Popa). Let G be a unitarily representable SIN Polish group. Is G necessarily of
finite type?
Key words and phrases. uniformly bounded representation, unitarizability, Maurey-Nikishin factorization, Polish groups of finite type.
1
2
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
Motivated by the study [AM12-1] of the Lie structure of the class Ufin , two of us began to
research Popa’s Ufin -problem [AM12-2], and obtained some positive answers when additional
conditions were placed on G. Among other things, we showed that the question has a positive
answer if G is amenable. For example, the Hilbert-Schmidt group U (H)2 is a finite type Polish
group. However, the general case was left unaddressed.
Note that there is a similarity between the Ufin -problem and the problem of finding a
trace on a finite von Neumann algebra, which was established by Murray and von Neumann
in their classical papers on the subject. Yeadon [Ye71] found a relatively simpler proof of
Murray’s and von Neumann’s theorem as follows: Let M be a finite von Neumann algebra.
For simplicity, we assume that M has separable predual, and let ϕ be a faithful normal state
on M . We let the unitary group U (M ) act on the predual M∗ by conjugation:
(1)
αu (ψ) := uψu∗ = ψ(u∗ · u),
ψ ∈ M∗ , u ∈ U (M ).
Note that α is a linear isometric action on the Banach space M∗ . Using the finiteness assumption, Yeadon showed that the orbit Oϕ := {αu (ϕ); u ∈ U (M )} is relatively weakly compact.
Thus, its weak convex closure co(Oϕ ) ⊂ M∗ is a weakly compact convex subset of M∗ on
which U (M ) acts by affine isometries. By the Ryll-Nardzewski Theorem, this implies that
there exists an α-fixed point, which is easily seen to be a normal faithful tracial state τ .
In the first named author’s PhD thesis, we adopted the Yeadon argument up to a certain
point. This part of the thesis formed an unpublished joint work with the second named author.
The argument was as follows. Let G be a unitarily representable Polish SIN group. It is wellknown (see e.g. [Gao05]) that G is unitarily representable if and only if G admits a continuous
positive definite function ϕ that generates the topology. An analogous characterization for
Ufin is obtained in [AM12-2]. Let P(G) be the set of all normalized continuous positive
definite functions on G. Then G is of finite type if and only if there exists a character (i.e.,
a function ϕ ∈ P(G) which is conjugation-invariant) that generates a neighborhood basis of
the identity on G (see Theorem 2.4 for details). It is clear that P(G) is a convex subset of the
Banach space Cb (G) of all continuous bounded functions on G equipped with the sup norm.
Define an affine isometric action β : G y Cb (G) by:
βg (ψ) := ψ(g −1 · g),
ψ ∈ Cb (G), g ∈ G.
If the orbit Oϕ = {βg (ϕ); g ∈ G} is relatively weakly compact, then by the Ryll-Nardzewski
Theorem, there exists a fixed point ψ in the closed convex hull co(Oϕ ). Once we get a
fixed point ψ ∈ co(Oϕ ), one can show, using Theorem 2.4 below, that G is of finite type.
Unfortunately, we do not know if Oϕ is relatively weakly compact. What we could show is
the following result.
Theorem 1.1 (Ando–Matsuzawa, unpublished). Let G be a unitarily representable Polish
SIN group, and assume that ϕ ∈ P(G) generates the topology of G. Then the following
statements hold:
(i) Oϕ is relatively compact in the compact open topology.
(ii) If Oϕ satisfies one of the following conditions, then G is of finite type.
(iia) The closure of Oϕ with respect to the topology of pointwise convergence is normseparable.
(iib) Oϕ is relatively weakly compact.
Two of the current authors considered the abovementioned result as an indication that the
Ufin -problem may have a positive answer in general. However, there was no obvious means of
POLISH GROUPS OF FINITE TYPE
3
verifying the relation between the compact open topology and the weak topology in Cb (G)
or verifying either (iia) or (iib).
In this paper, we show that indeed the orbit Oϕ ⊂ Cb (G) is not always relatively weakly
compact. That is, we construct a family of unitarily representable SIN Polish groups that do
not fall into the class Ufin . This answers Popa’s abovementioned question. The construction
is based on the existence of uniformly bounded non-unitarizable representations of countable
discrete groups.
Theorem 1.2 (Main Theorem). Let Γ be a countable discrete group, and let π : Γ → GL(H)
be a group homomorphism from Γ to the group of bounded invertible operators on H. Let
G = H ⋊π Γ be the semidirect product.
(a) G is a unitarily representable Polish group. Moreover, the following two conditions
are equivalent.
(i) G is a SIN group.
(ii) π is uniformly bounded. That is, kπk := sups∈Γ kπ(s)k < ∞ holds.
(b) Assume that π is uniformly bounded. Then the following three conditions are equivalent.
(i) G is of finite type.
(ii) There exists a character f : H → C that generates a neighborhood basis of 0 ∈ H,
such that f (π(s)ξ) = f (ξ) for all ξ ∈ H and s ∈ Γ.
(iii) π is unitarizable. That is, there exists V ∈ GL(H) such that V −1 π(s)V is unitary
for all s ∈ Γ.
A group Γ is said to be unitarizable if all of its uniformly bounded representations are
unitarizable. The problem of unitarizability has a long history. Sz.-Nagy [SzN47] showed that
the group Z of integers is unitarizable. Later, this was independently generalized to amenable
groups by Dixmier [Di50], Nakamura–Takeda [NT51] and Day [Da50]. This led Dixmier to
pose the question whether every unitarizable group is amenable, a question that remains open
today. The first examples of non-unitarizable groups were found by Ehrenpreis and Mautner
[EM55], who showed that SL2 (R) is not unitarizable. For discrete groups, explicit families
of uniformly bounded non-unitarizable representations of free groups were later constructed
by Mantero–Zappa [MZ83-1, MZ83-2], Pitlyk–Szwarc [PS86], and Bozejko [Bo87]. We refer
the reader to [Pi01, Pi05, Oz06] for details. Because unitarizability passes to subgroups, it
follows that all discrete groups containing free subgroups are non-unitarizable. We also note
that examples of non-unitarizable groups not containing non-abelian free subgroups were
presented by Epstein–Monod [EM09], Osin [Os09] and Monod–Ozawa [MO10]. Therefore,
there exist many uniformly bounded representations that are not unitarizable. The present
article provides a new and unexpected connection between the study of unitarizability of
groups and the understanding of the class Ufin .
Finally, we explain the idea of the proof. Theorem 1.2 (a) is shown in Theorem 2.7. For (b),
(i)⇔(ii) is Theorem 2.8 and (iii)⇒(i) is Corollary 2.9. Finally, (i)⇒(iii) is shown in Theorem
3.1. The proofs of (a) and (b) (i)⇔(ii) and (iii)⇒(i) rely on our characterization [AM12-2]
of the class Ufin via positive definite functions. The most difficult step is to prove that (b)
(i)⇒(iii). Note that our argument is different from that of Vasilescu-Zsido [VZ74], who used
the Ryll-Nardzewski Theorem, and showed that if a uniformly bounded representation π
generates a finite von Neumann algebra, then π is unitarizable. From the assumption that
G = H ⋊π Γ is of finite type, there appears to be no obvious means to deduce the conditoin
that W ∗ (π(Γ)) is a finite von Neumann algebra (observe that when π is an irreducible unitary
4
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
representation, then H ⋊π Γ is of finite type, but π(Γ)′′ = B(H) is not a finite von Neumann
algebra if dim(H) = ∞). Assume that G = H ⋊π Γ is of finite type. Then there exists
f ∈ P(H) that generates the neighborhood basis of 0 and satisfies f (π(s)ξ) = f (ξ) (s ∈
Γ, ξ ∈ H). If dim(H) < ∞, then by Bochner’s Theorem, there exists a probability measure µ
on H (regarded as a real Hilbert space) such that
Z
eix·ξ dµ(x),
ξ ∈ H.
f (ξ) =
H
If dim(H) = ∞, Bochner’s Theorem does not hold (e.g. the positive definite function ξ 7→
2
e−kξk is not the Fourier transform of a probability measure on H). However, there exists a
cylinder set probability measure µ defined on the algebraic dual Ha′ of H, such that
Z
eihχ,ξi dµ(χ),
ξ ∈ H.
(2)
f (ξ) =
Ha′
It can, in fact, be shown that µ is supported on a Hilbert-Schmidt extension of H. See [Ya85]
for details. Intuitively, one may hope that π is unitarizable by the following argument: define
Z
hχ, ξihχ, ηi dµ(χ),
ξ, η ∈ H,
hξ, ηi′π :=
Ha′
which is (if well-defined) an R-bilinear form. Because f is π-invariant, its dual action π̃ of Γ on
Ha′ preserves µ, which then implies that h·, ·i′π is Γ-invariant: hπ(s)ξ, π(s)ηi′π = hξ, ηi′π (ξ, η ∈
H, s ∈ Γ). By complexification, we obtain a π(Γ)-invariant inner product h·, ·iπ . We may
then hope that the inner product provides the same topology as the original inner product,
which would imply that π is unitarizable (see Lemma 3.2). There are two difficulties in this
approach: (a) we must guarantee that second moments exist:
Z
?
hχ, ξi2 dµ(χ) < ∞
Ha′
and (b) we must prove that the new inner product is equivalent to the original one. It does
not seem likely that these conditions hold in general. We remedy this by noncommutative
integration theory: assume that α : G → U (M ) is an embedding into the unitary group of a
II1 -factor M . Then by setting α(s) := α(0, s) and α(ξ) := α(ξ, 1) (ξ ∈ H, s ∈ Γ), we obtain
f be the space of all unbounded closed operators affiliated
α(s)α(ξ)α(s)∗ = α(π(s)ξ). Let M
with M . Because t 7→ α(tξ) is a strongly continuous one-parameter unitary group for each
fsa such that α(tξ) = eitT (ξ) (t ∈ R). One
ξ ∈ H, there exists a self-adjoint operator T (ξ) ∈ M
fsa is an R-linear homeomorphism onto its range. If we can show
can show that T : H → M
2
that T (ξ) ∈ L (M, τ ) for every ξ ∈ H (we call it the L2 -condition), then
(3)
hξ, ηi′π := hT (ξ), T (η)iL2 (M,τ ) ,
(ξ, η ∈ H)
is a well-defined inner-product, which is Γ-invariant by α(s)T (ξ)α(s)∗ = T (π(s)ξ) (ξ ∈ H, s ∈
Γ). We can also show that the induced norm is equivalent to the original norm. By complexification, we are done by Lemma 3.2. Again, the real question is whether the L2 -condition
holds in general, a question essentially equivalent to postulating the existence of the second
moments of µ. However, because the von Neumann algebra A = {α(ξ); ξ ∈ H}′′ generated
by the image of α is commutative, we may represent it as A = L∞ (X, m) for some probability
space (X, m) equipped with an m-preserving Γ-action β such that the W∗ -dynamical system
e Ad(α( · ))) is equivariantly ∗-isomorphic to (L0 (X, m), β). Here, L0 (X, m) is the space of
(A,
POLISH GROUPS OF FINITE TYPE
5
all measurable maps on a probability space (X, m) equipped with the topology of convergence
in measure Because H is a Hilbert space (hence of Rademacher type 2, see [AK06, §6]), it
is known (by the so-called Maurey–Nikishin factorization [Ni70, Ma72]) that any continuous
e factors through a Hilbert space. That is, there exists
linear operator T : H → L0 (X, m) = A
2
a bounded operator T̃ : H → L (X, m) and an m-a.e. positive function ψ ∈ L0 (X, m) such
that the following diagram commutes:
T
H❍
❍
❍❍
❍❍
❍❍
❍##
T̃
// L0 (X, m)
q88
qqq
q
q
q
qqq Mψ
L2 (X, m)
Here, Mψ is the multiplication operator by Rthe function ψ. More precisely, there exists an
m-a.e. positive ϕ ∈ L∞ (X, m) such that X ϕ(x)[T ξ](x)2 dm(x) < ∞ for every ξ ∈ H.
1
Then T̃ ξ := M 21 · T ξ and ψ = ϕ− 2 gives the factorization. This result was first essentially
ϕ
proved by Nikishin [Ni70]. Nikishin’s proof was later simplified by Maurey [Ma72] using the
theory of Rademacher types. We refer the reader to [Pi86, AK06, GR85] for more details. In
§3, we show an equivariant version of the Maurey–Nikishin factorization, in which we use a
convexity argument to ensure (Theorem 3.5) that the operator T̃ can be made Γ-equivariant
if T is assumed to be Γ-equivariant. We can then define h·, ·i′π by the formula (3), with T
replaced by T̃ . Because T is a real linear homeomorphism onto its range, we can show that
T̃ is also a real linear homeomorphism onto its range too. The last step is to show that on
the image T̃ (H) of H under T̃ , the topology of the convergence in measure agrees with the
L2 -topology (Lemma 3.13). This is somewhat surprising: note that these two topologies do
not agree on the whole of L2 (X, m). The proof is done by using the fact [Be10, AM12-1]
f, the strong resolvent topology and the τ -measure topology of Nelson [Ne74] agree.
that on M
Once this is established, it is straightforward to show that h·, ·i′π generates the same topology
as that of the original inner product, and therefore that π is unitarizable.
Finally, we report that as a byproduct of the existence of an equivariant Maurey–Nikishin
factorization, we find a new, useful description of positive definite functions on a Hilbert space
(compare (2) with (4)). See also the comments following Corollary 3.16.
Theorem 1.3 (Corollary 3.16). Let H be a separable real Hilbert space, and let f ∈ P(H).
Then there exist
(i) a compact metrizable space X and a Borel probability measure m on X,
(ii) a bounded linear operator T : H → L2 (X, m; R), and
(iii) a positive measurable function ϕ ∈ L0 (X, m),
such that
(4)
f (ξ) =
Z
eiϕ(x)[T ξ](x) dm(x),
ξ ∈ H.
X
Moreover:
(iv) If f generates a neighborhood basis of 0, then T can be chosen to be a real-linear
homeomorphism onto its range.
(v) We may arrange T and ϕ such that if G is a uniformly bounded subgroup of the group
{u ∈ GL(H); f (uξ) = f (ξ) (ξ ∈ H)} of invertible operators on H fixing f pointwise,
6
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
then there exists a continuous homomorphism β from G to the group Aut(X, m) of all
m-preserving automorphisms of X with the weak topology, such that
[T (uξ)](x) = β(u)[T (ξ)](x),
ϕ(β(u)x) = ϕ(x),
u ∈ G, ξ ∈ H, m-a.e. x ∈ X.
2. The class Ufin and uniformly bounded representations
In this section we recall Popa’s class Ufin of finite type Polish groups and consider when a
specific semidirect group belongs to this class. The results of this section are extracted from
the first named author’s PhD thesis, jointly proved with the second named author.
Definition 2.1. Let G, G′ be topological groups.
(i) We say that G is embeddable into G′ , if there exists a topological group isomorphism
π from G onto a closed subgroup of G′ . Such π is called an embedding.
(ii) We denote by P(G) the set of all normalized continuous positive definite functions
on G, and by P(G)inv the set of all f ∈ P(G) which is conjugation-invariant, i.e.,
f (g−1 xg) = f (x) holds for all x, g ∈ G. An element of P(G)inv is called a character
on G.
Definition 2.2 ([Po07]). A Polish group G is of finite type if G is embeddable into the unitary
group U (M ) of some finite von Neumann algebra M equipped with strong operator topology.
The class of finite type Polish groups is denoted by Ufin .
If G is a finite type Polish group, then by [Po07, Lemma 2.6], the above M may chosen to
be a type II1 -factor with separable predual.
Definition 2.3. Let G be a topological group.
(i) G is called unitarily representable if G is embeddable into U (ℓ2 (N)) equipped with
strong operator topology.
(ii) G is called SIN (small invariant neighborhood) if for any neighborhood V of G at the
identity e, there exists a neighborhood U ⊂ V of e which is conjugation invariant, i.e.,
g−1 U g = U, ∀g ∈ G. Such U is called an invariant neighborhood of G.
If G is metrizable, then G is SIN if and only if G admits a bi-invariant metric d which is
compatible with its topology.
For more details about unitary representability, see e.g., [Gal09, Gao05, Me02]. We will
use the following characterization of the finite type Polish groups (cf. [Gao05]).
Theorem 2.4 ([AM12-2]). For a Polish group G, the following conditions are equivalent.
(i) G is of finite type.
(ii) There exists a family F ⊂ P(G)inv which generates a neighborhood basis of the identity eG of G. That is, for each neighborhood V at the identity, there are functions
f1 , . . . , fn ∈ F and open sets O1 , . . . , On in C such that
n
\
fi−1 (Oi ) ⊂ V.
eG ∈
i=1
(iii) There exists f ∈ P(G)inv that generates a neighborhood basis of the identity of G.
(iv) There exists a family F ⊂ P(G)inv which separates the identity of G and closed subsets
A with A 6∋ eG . That is, for each closed subset A with A 6∋ eG , there exists f ∈ F
such that
sup |f (x)| < |f (eG )|.
x∈A
POLISH GROUPS OF FINITE TYPE
7
(v) There exists f ∈ P(G)inv which separates the identity of G and closed subsets A with
A 6∋ eG .
We start from the following characterization for the semidirect product of a SIN group
by a SIN group to be SIN. Denote by N (G) the set of all neighborhoods of identity of a
topological group G.
Theorem 2.5. Let N, K be SIN-groups. Let α : K → Aut(N ) be a continuous action of K
on N . Then the following two conditions are necessary and sufficient for N ⋊α K to be a
SIN-group.
(SIN. 1) For any U ∈ N (N ), there exist V ∈ N (N ) with V ⊂ U such that
αk (V ) = V,
k ∈ K.
(SIN. 2) The action α is bounded. That is, for any V ∈ N (N ), there exists W ∈
N (K) such that
n−1 αk (n) ∈ V,
n ∈ N, k ∈ W.
Proof. We denote by eN (resp. eK ) the unit of N (resp. K).
We first prove sufficiency of the two conditions. Fix an arbitrary U ∈ N (N ⋊α K). By
definition of the product topology, there exist V ′′ ∈ N (N ) and W ′ ∈ N (K) such that
(eN , eK ) ∈ V ′′ × W ′ ⊂ U holds. Since N is SIN, there exists an invariant neighborhood
V ′ ∈ N (N ) which satisfies V ′ · V ′ ⊂ V ′′ . By (SIN. 1), we may take V ∈ N (N ) such that
V ⊂ V ′ and
αk (V ) = V,
k ∈ K.
By (SIN. 2) and the SIN property of K, there exists an invariant neighborhood W ∈ N (K)
satisfying W ⊂ W ′ and
n−1 αk (n) ∈ V, n ∈ N, k ∈ W.
Indeed, by (SIN. 2) there is W0 ∈ N (K) such that n−1 αk (n) ∈ V holds for all n ∈ N and
k ∈ W0 . By the SIN property of K, there is an invariant neighborhood W ∈ N (K) contained
in W0 ∩ W ′ . It is then clear that this W satisfies the above requirement.
Let (n1 , k1 ) ∈ V × W and (n2 , k2 ) ∈ N ⋊α K. We see that
−1
(n2 , k2 )(n1 , k1 )(n2 , k2 )−1 = (n2 αk2 (n1 ), k2 k1 )(αk−1 (n−1
2 ), k2 )
2
−1
= (n2 αk2 (n1 )αk2 k1 k−1 (n−1
2 ), k2 k1 k2 )
2
−1
−1
= {n2 αk2 (n1 )n−1
)},
k
k
k
}{n
α
−1 (n
2
2
1
2
2
2
k2 k1 k
2
∈ (n2 V n−1
2 ·V)×W
′
′
⊂ (n2 V ′ n−1
2 ·V )×W
= (V ′ · V ′ ) × W ′
⊂ U.
Hence, we have g−1 (V × W )g ⊂ U for all g ∈ N ⋊α (K). This shows
[
U ′ :=
g −1 (V × W )g ⊂ U
g∈N ⋊α K
is an invariant neighborhood of eN ⋊α H contained in U . Since U is arbitrary, this shows
N ⋊α K is a SIN-group.
We now prove necessity of the two conditions. Suppose N ⋊α K is a SIN-group. Given
U ∈ N (N ). Then U × K ∈ N (N ⋊α K). Thus by SIN-property of N ⋊α K we find an
8
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
invariant neighborhood A ∈ N (N ⋊α K) contained in U × K. Then we may find U0 ∈ N (N )
and W ∈ N (K) such that U0 × W ⊂ A. Since A is invariant, we have
g(U0 × {eK })g−1 ⊂ A,
g ∈ N ⋊α K.
Inparticular, for all k ∈ K and n ∈ U0 , we have
(eN , k)(n, eK )(eN , k)−1 = (αk (n), eK ) ∈ A ⊂ U × K.
This shows αk (n) ∈ U for all k ∈ K and n ∈ U0 ⊂ U . Then
[
V :=
αk (U0 ) ⊂ U
k∈K
is a neighborhood of eN satisfying αk (V ) = V for all k ∈ K. Hence we conclude that (SIN.
1) holds.
In order to prove (SIN. 2), suppose V ∈ N (N ) is given. Again by the same argument as
above, we may find an invariant neighborhood A ∈ N (N ⋊α K) contained in V × K and
V0 ∈ N (N ) and W ∈ N (K) with V0 × W ⊂ A ⊂ V × K. Since A is invariant, we have in
particular
g({eN } × W )g−1 ⊂ A ⊂ V × K,
g ∈ N ⋊α K.
Hence for n ∈ N and k ∈ W , we have
(n, eK )(eN , k)(n, eK )−1 = (nαk (n−1 ), k) ∈ A ⊂ V × K.
This shows nαk (n−1 ) ∈ V for all k ∈ W . This proves (SIN. 2).
Remark 2.6. A semidirect product G ⋊α Γ of a SIN group G by a discrete group Γ always
satisfies (SIN. 2) (take W = {eK }).
As an immediate corollary to Theorem 2.5, we obtain:
Theorem 2.7. Let Γ be a countable discrete group, H be a separable Hilbert space and
π : Γ → GL(H) be an invertible representation of Γ. Then the semidirect product group
G := H ⋊π Γ is unitarily representable. Moreover, the following conditions are equivalent.
(1) G is SIN.
(2) π is uniformly bounded, i.e., kπk := sups∈Γ kπ(s)k < ∞.
2
Proof. The unitarily representable group H has a positive definite function f (ξ) := e−kξk , ξ ∈
H and identifying H with H × {e} ⊂ H ⋊π Γ, we may extend f to be a continuous positive
definite function on H ⋊π Γ by taking value 0 outside H × {e}. It is immediate to see that
this extended function still generates the neighborhood basis of the identity. Then by [Gao05,
Theorem 2.1], G is unitarily representable.
(1)⇒(2) Assume that G is a unitarily representable and SIN group. Let Bδ (0) be the open
δ-neighborhood of H at 0 (δ > 0). Then by (SIN. 1), there exists a neighborhood V ⊂ Bδ (0)
of H at 0 such that π(s)V = V, (s ∈ Γ). Since V is a neighborhood, there exists ε > 0
such that Bε (0) ⊂ V holds. Thus π(s)Bε (0) ⊂ Bδ (0) holds for all s ∈ Γ. This shows that
kπ(s)k ≤ δε for every s ∈ Γ. Therefore π is uniformly bounded.
(2)⇒(1) Assume that π is uniformly bounded. To prove that G is SIN, by Theorem 2.5 we have
only to check (SIN.1) (see Remark 2.6). Consider Bε (0) (ε > 0), and C := 1+sups∈Γ kπ(s)k <
∞. Then for all ξ ∈ Bε/C (0), s ∈ Γ, we have
kπ(s)ξk ≤ Ckξk < ε.
POLISH GROUPS OF FINITE TYPE
9
This implies π(s)Bε/C (0) ⊂ Bε (0) for all s ∈ Γ. Then V :=
(SIN.1).
S
s∈Γ π(s)Bε/C (0)
satisfies
Theorem 2.8. Let Γ be a countable discrete group, H be a separable Hilbert space and
π : Γ → B(H) be a uniformly bounded representation of Γ. Then H ⋊π Γ is of finite type if
and only if there exists f ∈ P(H) such that
(F.1) f (π(s)ξ) = f (ξ),
∀s ∈ Γ, ξ ∈ H,
(F.2) f generates a neighborhood basis of 0 ∈ H.
Proof. We first prove “if” part. For this, we assume that there exists a positive, continuous
positive definite function f on H which satisfies (F.1) and (F.2). Then one can extend f on
H ⋊π Γ by
(
f (ξ) (s = 1),
f˜(ξ, s) :=
0
(s 6= 1).
It is easy to see that f˜ is continuous, positive definite and generates a neighborhood basis of
the identity. Take arbitrary ξ, η ∈ H, s, t ∈ Γ. Since
(η, t)(ξ, s)(η, t)−1 = η + π(t)ξ − π(tst−1 )η, tst−1 ,
we have
f˜ (η, t)(ξ, s)(η, t)−1 =
=
(
(
f˜ (π(t)ξ, 1)
0
(s = 1),
(s 6= 1).
f (ξ) (s = 1),
0
(s =
6 1).
= f˜(ξ, s).
Hence H ⋊π Γ is of finite type.
Conversely, let us assume that H ⋊π Γ is of finite type. Then there exists a positive,
continuous positive definite class function f˜ on H ⋊π Γ that generates a neighborhood basis
of the identity. We define a positive definite function f on H by
f (ξ) := f˜(ξ, 1),
ξ ∈ H.
It is easy to check that f is positive, continuous and satisfies (F.2). Since
(η, t)(ξ, 1)(η, t)−1 = (π(t)ξ, 1)
and f˜ is a class function, we obtain
f (ξ) = f˜(ξ, 1) = f˜ (η, t)(ξ, 1)(η, t)−1
= f˜ (π(t)ξ, 1) = f (π(t)ξ)
so that f satisfies (F.1).
Corollary 2.9. Let Γ be a countable discrete group, H be a separable Hilbert space and
π : Γ → B(H) be a uniformly bounded representation of Γ. If π is unitarizable, then H ⋊π Γ
is of finite type.
10
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
Proof. Since π is unitarizable, there exists an invertible operator T ∈ B(H) such that T π(s)T −1
is unitary for every s ∈ Γ. Define a positive definite function f on H by
2
f (ξ) := e−kT ξk ,
ξ ∈ H.
Then f is positive, continuous and satisfies (F.2). Moreover
f (π(s)ξ) = e−kT π(s)ξk = e−k{T π(s)T
2
−1
2
}T ξ k = e−kT ξk2 = f (ξ)
so that (F.1) holds.
Since amenable groups are unitarizable ([Da50, Di50, NT51]), we obtain:
Corollary 2.10. Let Γ be a countable discrete group, H be a separable Hilbert space and
π : Γ → B(H) be a uniformly bounded representation of Γ. Assume that Γ is amenable. Then
H ⋊π Γ is of finite type.
Remark 2.11. We remark that if Γ is amenable, then so is H ⋊π Γ, so that the conclusion of
Corollary 2.10 also follows from the fact [AM12-2, Theorem 2.2] that an amenable unitarily
representable SIN group is of finite type.
The purpose of the next section is to prove a converse to Corollary 2.9.
3. Equivariant Maurey–Nikishin Factorization
In this section, we prove the last and crucial step, Theorem 1.2 (b) (i)⇒(iii). Let π : Γ → H
be a uniformly bounded representation of a discrete countable group Γ on a separable Hilbert
space H, as before.
Theorem 3.1. If G = H ⋊π Γ is of finite type, then π is unitarizable.
Throughout this section, we assume that G = H ⋊π Γ is of finite type. Our aim is to find
a Γ-invariant inner product and apply the following classical result due to Dixmier:
Lemma 3.2 (Dixmier). Let π : Γ → B(H) be a uniformly bounded representation. Then π is
unitarizable if and only if there exists an inner product h·, ·i′ on H that generates the same
Hilbert space topology and which is π-invariant, i.e., hπ(s)ξ, π(s)ηi′ = hξ, ηi′ (ξ, η ∈ H, s ∈ Γ)
holds.
By assumption, there exists an embedding (as a topological group) α : G → U (M ), where
M is some II1 -factor with separable predual and with the normal faithtul tracial state τ . We
let M act on L2 (M, τ ). As usual, we identify H ⊂ G ⊂ U (M ) and Γ ⊂ U (M ). That is, ξ ∈ H
is identified with α(ξ) = α(ξ, 0), and s ∈ Γ is identified with α(s) = α(0, s). Therefore under
this identification, the uniformly bounded G-action π on H is implemented by unitaries in
M:
(5)
α(π(s)ξ) = α((0, s)(ξ, 1)(0, s−1 ) = α(s)α(ξ)α(s)∗ .
f be the space of all (possibly unbounded) closed and densely defined operators T
Now, let M
f,
on L2 (M, τ ) which are affiliated with M . There is a completely metrizable topology on M
called the τ -measure topology [Ne74]. A neighborhood basis of 0 in the τ -measure topology
is given by the family {N (ε, δ); ε, δ > 0}, where
f; ∃p ∈ Proj(M ), τ (p⊥ ) < δ, kAp⊥ k < ε},
N (ε, δ) := {A ∈ M
POLISH GROUPS OF FINITE TYPE
11
where Proj(M ) is the lattice of all projections of M . In [AM12-1], we showed that the
τ -measure topology coincides with the strong resolvent topology (SRT) (the same result is
f are selfobtained by Beltita [Be10] using the standard form). In particular, if An , A ∈ M
n→∞
n→∞
−1
adjoint, then An → A in the τ -measure topology, if and only if k(An −i) −(A−i)−1 k2 →
n→∞
0, if and only if (by the Kato-Trotter Theorem) keitAn − eitA k2 → 0 for every t ∈ R (in this
case the convergence is compact uniform in t). As was shown by Murray-von Neumann, the
f, although not locally convex, is actually a topological ∗-algebra with respect to the
space M
τ -measure topology. The addition (resp. the multiplication) is defined by (A, B) 7→ A + B
(resp. (A, B) 7→ AB), where the bar denotes the operator closure (it asserts that A + B, AB
are always closable and in fact A + B is self-adjoint whenever A, B are).
Now let ξ ∈ H. Then since α is continuous, R ∋ t 7→ α(tξ) ∈ U (M ) is a strongly continuous
one-parameter unitary group. Thus by Stone’s Theorem, there exists a unique self-adjoint
f such that
operator T (ξ) ∈ M
α(tξ) = eitT (ξ) ,
t ∈ R, ξ ∈ H.
fsa is a real-linear homeomorphism onto its range (with
Proposition 3.3. The map T : H → M
respect to the τ -measure topology or the SRT). Moreover, the following covariance property
holds.
T (π(s)ξ) = α(s)T (ξ)α(s)∗ ,
(6)
s ∈ Γ, ξ ∈ H.
Proof. Let s ∈ Γ and ξn , ξ, η ∈ H (n ∈ N). Since α is a topological isomorphism onto its
range,
n→∞
n→∞
ξn − ξ → 0 ⇔ α(tξn ) − α(tξ) = eitT (ξn ) − eitT (ξ) → 0 for all t ∈ R (SOT)
n→∞
⇔ T (ξn ) → T (ξ) (SRT).
Thus T is a homeomorphism from H onto its range. Also, for all s, t ∈ R, we have
α(tsξ) = eitsT (ξ) = α(t(sξ)) = eitT (sξ) .
Since t is arbitrary, the uniqueness part of Stone’s Theorem asserts that T (sξ) = sT (ξ).
Now since H is a commutative (additive) group, all {T (ξ); ξ ∈ H} are strongly commuting
self-adjoint operators. Thus by Kato-Trotter Theorem,
1
1
eitT (ξ)+T (η) = SOT − lim (eit n T (ξ) eit n T (η) )n = eitT (ξ) eitT (η)
n→∞
= α(tξ)α(tη) = α(t(ξ + η))
= eitT (ξ+η) .
Therefore we obtain T (ξ + η) = T (ξ) + T (η). Therefore T is a real-linear homeomorphism
onto its range. Finally, by the unitary covariance of the Borel functional calculus,
eitT (π(s)ξ) = α(tπ(s)ξ) = α(s)α(tξ)α(s)∗
= α(s)eitT (ξ) α(s)∗
∗
= eitα(s)T (ξ)α(s) .
Therefore T (π(s)ξ) = α(s)T (ξ)α(s)∗ .
12
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
Let A := {α(ξ); ξ ∈ H}′′ be the commutative von Neumann subalgebra of M generated
by the image of α. By RProposition 3.19, there exists a probability space (X, m) such that
(A, τ |A ) ∼
= (L∞ (X, m), X · dm). Then by (6), the formula
(7)
βs := Ad(α(s))|A ,
s ∈ Γ.
defines a trace-preserving action of Γ on A = L∞ (X, m). This action also defines a probability
measure preserving action on X. We denote this action by Γ × X ∋ (s, x) 7→ s · x ∈ X.
Therefore, βs (ϕ) = ϕ ◦ s−1 , ϕ ∈ L∞ (X, m). Below we keep this setting.
Definition 3.4. We denote by L0 (X, m) the space of all m-measurable C-valued functions
on X, and L0 (X, m)+ its positive cone. We also define the Γ-action on L0 (X, m) by βs =
Ad(α(s))|L0 (X,m) (s ∈ Γ) (we use the same symbol βs as the Γ-action on L∞ (X, m), but there
would be no danger of confusion).
Equipped with the convergence in measure, L0 (X, m) is a (non-locally convex) completely
metrizable topological vector space. Note that under the identification A = L∞ (X, m), we
fsa (cf. [AM12-1]). What we will do is to construct a Γ-equivariant
obtain Ae = L0 (X, m) ⊂ M
fsa . We almost exactly
Maurey–Nikishin factorization of the map T : H → L0 (X, m) ⊂ M
follow Maurey’s argument as in [GR85, Chapter VI.3], but we also need to take care of the
Γ-action, which makes it necessary to refine some of the arguments.
Theorem 3.5 (Equivariant Maurey-Nikishin Factorization). The map T : H → L0 (X, m)
factors Γ-equivariantly through a Hilbert space. More precisely, there exists a Γ-invariant
function ϕ ∈ L0 (X, m) which is m-a.e. positive and
Z
ϕ(x)[T ξ](x)2 dm(x) ≤ kξk2 ,
ξ ∈ H.
(8)
X
In particular, the following diagram commutes:
H❍
❍
T
❍❍
❍❍
❍❍
❍##
T̃
// L0 (X, m)
q88
qqq
q
q
q
qqq Mψ
L2 (X, m)
1
Here, Mψ is the multiplication operator by the Γ-invariant function ψ = ϕ− 2 and T̃ ξ :=
M 12 · T ξ is Γ-equivariant.
ϕ
Definition 3.6. For ϕ ∈ L0 (X, m), we define its non-increasing rearrangement ϕ⋆ : [0, 1] →
[0, ∞) by
ϕ⋆ (t) := inf{λ > 0; m({x ∈ X; |ϕ(x)| > λ}) ≤ t},
t ∈ [0, 1].
⋆
⋆
Then clearly (cϕ) (t) = cϕ (t) (c > 0, t ≥ 0) holds.
The next lemmata are elementary.
Lemma 3.7. Let S : H → L0 (X, m) be a linear operator. Then the following conditions are
equivalent.
(i) S is continuous.
(ii) There exists a non-increasing map C : (0, ∞) → [0, 1] such that limλ→∞ C(λ) = 0 and
(9)
m({x ∈ X; |Sξ(x)| ≥ λkξk}) ≤ C(λ)
(λ > 0, ξ ∈ H).
POLISH GROUPS OF FINITE TYPE
13
(iii) For every t > 0, there exists K(t) > 0 such that
(10)
(Sξ)⋆ (t) ≤ K(t)kξk
(ξ ∈ H).
Proof. (i)⇒(ii) For each ε, δ > 0, the set U (ε, δ) := {f ∈ L0 (X, m); m({x ∈ X; |f (x)| ≥ ε}) ≤
δ} is an neighborhood of 0 in the measure topology. For each λ > 0, define
C(λ) := inf{δ > 0; Sξ ∈ U (λ, δ) for all ξ ∈ H with kξk ≤ 1}.
Note that C(λ) ≤ 1. It is then clear that C(·) is nonincreasing and (9) holds. We claim
that limλ→∞ C(λ) = 0. Assume by contradiction that there exists δ > 0 such that C(λ) ≥ δ
holds for all λ > 0. Then for each n ∈ N, there exists ξn ∈ H with kξn k ≤ 1, such that
Sξn ∈
/ U (n, δ − n1 ), i.e., m({x ∈ X; |Sξn (x)| ≥ n}) > δ − n1 holds. Then ηn := n1 ξn (n ∈ N)
n→∞
satisfies kηn k → 0, but m({x ∈ X; |Sηn (x)| ≥ 1) > δ − n1 , so Sηn 6→ 0 in measure, a
contradiction.
(ii)⇒(iii) Let t > 0. Since limλ→∞ C(λ) = 0, there exists K(t) > 0 such that C(λ) < t (λ ≥
K(t)) holds. Then for ξ ∈ H with kξk ≤ 1, m({x ∈ X; |Sξ(x)| > λ}) ≤ C(λ), so that
(Sξ)⋆ (t) ≤ K(t). For general ξ 6= 0, we have (Skξk−1 ξ)⋆ (t) ≤ kξk−1 K(t), whence (10) holds.
(iii)⇒(i) Since H and L0 (X, m) are metrizable, it suffices to show the sequential continuity
n→∞
of S. Let ξn ∈ H be such that kξn k → 0. Let ε, δ > 0. By assumption, we have
n→∞
(Sξn )⋆ (δ) ≤ K(δ)kξn k → 0. Therefore there exists N ∈ N such that (Sξn )⋆ (δ) < ε holds for
all n ≥ N . In particular, we have m({x ∈ X; |Sξn (x)| ≥ ε}) ≤ δ. Since ε, δ > 0 are arbitrary,
Sξn converges to 0 in measure.
Lemma 3.8. Let ϕ ∈ L0 (X, m). Then for every δ > 0, m({x ∈ X; |ϕ(x)| > ϕ⋆ (δ)}) ≤ δ
holds.
Proof. Let λ := ϕ⋆ (δ). Then for each n ∈ N, there exists λn ∈ (λ, λ + n1 ) such that µ({x ∈
X; |ϕ(x)| > λn }) ≤ δ. Thus
m({x ∈ X; |ϕ(x)| > λ}) ≤ lim m({x ∈ X; |ϕ(x)| > λ + n1 })
n→∞
≤ lim sup m({x ∈ X; |ϕ(x)| > λn })
n→∞
≤ δ.
The crucial step in our proof is to establish the following proposition. This is the step
which allows to make the construction of the factorization canonical and hence equivariant
with respect to a group of symmetries.
Proposition 3.9. Let A be a convex subset of L0 (X, m)+ which is globally Γ-invariant. Then
for every ε > 0, there is a Γ-invariant measurable subset S(ε) ⊂ X with m(X \ S(ε)) ≤ ε and
Z
(11)
sup
ϕ(x) dm(x) ≤ 2 sup ϕ⋆ ( 2ε ).
ϕ∈A S(ε)
ϕ∈A
This is a Γ-invariant version of [GR85, Proposition VI.3.1]. In order to show Proposition
3.9, we use the following variant of the Minimax Lemma:
Lemma 3.10 (Minimax Lemma). Let X, Y be topological vector spaces, and let A (resp. B)
be a convex subset of X (resp. Y ). Assume that B is compact. If the function Φ : A × B →
R ∪ {+∞} satisfies :
14
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
(i) Φ(·, b) is a concave function on A for each b ∈ B,
(ii) Φ(a, ·) is a convex function on B for each a ∈ A,
(iii) Φ(a, ·) is lower-semicontinuous on B for each a ∈ A,
then the following identity holds:
min sup Φ(a, b) = sup min Φ(a, b).
b∈B a∈A
a∈A b∈B
Proof. See [GR85, Appendix A.2].
Proof of Proposition 3.9. Define
Z
ε
0
ϕ(x) dm(x) ≥ 1 −
B := ϕ ∈ L (X, m); 0 ≤ ϕ ≤ 1 (m-a.e.),
.
2
X
Then we first check that B is a globally Γ-invariant, weakly compact, convex subset of the
Hilbert space L2 (X, m): indeed, B ⊂ L∞ (X, m) ⊂RL2 (X, m) is clearly
a convex set contained
R
2
2
in the unit ball of L (X, m) (any ϕ ∈ B satisfies X |ϕ| dm ≤ X 1dm = 1). Since m ◦ βs =
m (s ∈ Γ), the Γ-invariance is clear. Suppose (ϕn )∞
n=1 is a sequence in B converging weakly
to ϕ ∈ L2 (X, m). For each k ∈ N, let Ek = {x ∈ X; ϕ(x) ≤ − k1 }. Then
Z
1
ϕ dm ≤ − m(Ek ) ≤ 0,
0 ≤ lim hϕn , 1Ek iL2 (X) = hϕ, 1Ek iL2 (X) =
n→∞
k
Ek
S
so that m(Ek ) = 0. Therefore E = {x ∈ X; ϕ(x) < 0} = k∈N Ek satisfies m(E) = 0, so
ϕ ≥ 0 a.e. Similarly, ϕ ≤ 1 a.e. holds, and
Z
ε
ϕ dm = hϕ, 1iL2 (X) = lim hϕn , 1i ≥ 1 − .
n→∞
2
X
Therefore ϕ ∈ B. This shows that B is weakly closed, so that B is a weakly compact convex
subset of L2 (X, m).
Next, define Φ : A × B → R ∪ {+∞} by
Z
ψϕ dm,
(ψ ∈ A, ϕ ∈ B).
Φ(ψ, ϕ) :=
X
It is clear that Φ satisfies conditions (i) and (ii) of Lemma 3.10. To verify the condition (iii)
of Lemma 3.10, define for ψ ∈ A, ϕ ∈ B and t > 0:
Et (ψ) : = {x ∈ X; ψ(x) ≤ t},
Z
ψϕ dm.
Φt (ψ, ϕ) :=
Et (ψ)
Then for ψ ∈ A and t > 0, the map Φt (ψ, ·) : B → R+ is weakly continuous, and for ϕ ∈ B,
we have
Φ(ψ, ϕ) = sup Φt (ψ, ϕ).
t>0
This shows that Φ(ψ, ·) is weakly lower-semicontinuous on B. Let M := supψ∈A ψ ⋆ ( 2ε ), and
assume that M < ∞ (if M = ∞, there is nothing to prove). Then for every ψ ∈ A,
Z
1EM (ψ) dm = m({x ∈ X; |ψ(x)| > M })
1−
X
≤ m({x ∈ X; |ψ(x)| > ψ ⋆ ( 2ε )}) ≤ 2ε .
POLISH GROUPS OF FINITE TYPE
15
Here we used Lemma 3.8. This shows that 1EM (ψ) ∈ B for every ψ ∈ A. Therefore by Lemma
3.10,
Z
Z
ψϕ dm
min sup
ψϕ dm = sup min
ϕ∈B ψ∈A X
ψ∈A ϕ∈B
≤ sup
(12)
X
Z
ψ dm ≤ M.
ψ∈A EM (ψ)
Therefore the minimum of the left hand side will be attained at some ϕ ∈ B. So the set B0 of
all such minimizers ϕ ∈ B is nonempty. We observe that B0 is a globally Γ-invariant weakly
closed, convex subset of B. Indeed, let 0 < λ < 1 and ϕ1 , ϕ2 ∈ B0 be minimizers and let
ϕ = λϕ1 + (1 − λ)ϕ2 . Then for ψ ∈ A,
Z
Z
Z
ψϕ2 dm
ψϕ1 dm + (1 − λ) sup
ψϕ dm ≤ λ sup
sup
ψ∈A X
ψ∈A X
ψ∈A X
Z
= min
sup
ψϕ′ dm.
′
ϕ ∈B ψ∈A X
This shows that ϕ is also a minimizer, hence in B0 . Next, let (ϕn )∞
n=1 be a sequence of
elements in B0 converging weakly to ϕ ∈ L2 (X, m). Then ϕ ∈ B, and Fatou’s Lemma shows
Z
Z
ψ1{x;ψ(x)≤k} ϕ dm
sup
ψϕ dm ≤ sup lim inf
{z
}
ψ∈A X
ψ∈A k→∞
X|
∈L2 (X,m)
Z
ψ1{x;ψ(x)≤k} ϕn dm
Z
ψϕn dm
≤ sup lim inf lim sup
ψ∈A k→∞ n→∞
X
Z
ψ ′ ϕn dm
≤ sup lim sup sup
= sup lim inf lim
ψ∈A k→∞ n→∞ X
ψ∈A n→∞ ψ′ ∈A X
ϕn ∈B0
Z
ψ ′ ϕ′ dm
lim sup min
sup
n→∞ ϕ′ ∈B ψ′ ∈A X
Z
ψ ′ ϕ′ dm.
= min
sup
′
=
ϕ ∈B ψ′ ∈A X
This shows that ϕ is a minimizer too, so ϕ ∈ B0 . Finally, let ϕ ∈ B0 and s ∈ Γ. Then since A
is globally Γ-invariant and m is Γ-invariant, we have (recall that the Γ action on X is given
so that βs (ϕ) = ϕ(s−1 · ))
Z
Z
βs−1 (ψ)ϕ dm(g·)
ψβs (ϕ) dm = sup
sup
ψ∈A X
ψ∈A X
Z
= sup
ψϕ dm
ψ∈βs−1 (A) X
= min
sup
′
Z
ϕ ∈B ψ∈A X
ψϕ′ dm.
This shows that βs (ϕ) ∈ B0 . Therefore B0 is a Γ-invariant, non-empty, weakly compact, and
convex subset of L2 (X, m). Let ϕ0 be an element in B0 of minimum L2 -norm. By the uniform
16
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
convexity, such an element is unique. Then since kβs (ϕ0 )kL2 (X) = kϕ0 kL2 (X) (s ∈ Γ), we have
βs (ϕ0 ) = ϕ0 (s ∈ Γ) by the uniqueness of ϕ0 . Then define
S(ε) := {x ∈ X; ϕ0 (x) ≥ 21 }.
Then S(ε) is a Γ-invariant measurable subset of X. Moreover,
m(X \ S(ε)) = m({x ∈ X; ϕ0 (x) < 12 })
Z
≤ 2 (1 − ϕ0 (x)) dm ≤ ε,
X
Z
Z
ψϕ0 dm ≤ 2M.
ψ dm ≤ 2 sup
sup
ψ∈A
ψ∈A S(ε)
X
Here in the last line we used (12).
In the sequel, we will mainly follow the arguments in [GR85, Theorem VI.3.3]. Since this
source is hardly accessible, we decided to repeat some of the arguments (and present them
in a streamlined form) in order to provide a complete proof. We set Z := {|(T ξ)(x)|2 ; ξ ∈
H, kξk ≤ 1} ⊂ L0 (X, m)+ and note that
n
n
X
X
kξj k2 ≤ 1 ⊂ L0 (X, m)+
(13)
A :=
|(T ξj )|2 ; n ∈ N, ξ1 , . . . , ξn ∈ H,
j=1
j=1
is the convex hull of Z. In order to apply Proposition 3.9 to (a subset of) A, we need to show
that the estimates from Lemma 3.7 extend to the convex hull of Z. This is the content of the
following proposition.
Proposition 3.11. There exists a function (0, ∞) ∋ λ 7→ C(λ) ∈ [0, 1] satisfying
lim C(λ) = 0,
λ→∞
such that for every finite family ξ1 , · · · , ξn ∈ H, one has
1
2
n
n
X
X
2
2
kξj k ≤ 1 ⇒ m x ∈ X;
|T ξj (x)|
≥ λ ≤ C(λ).
j=1
j=1
This is Condition (d) in [GR85, Theorem VI.3.3], and the proof can be found in [GR85,
Theorem VI.3.6]. We give an alternative and direct proof, following a strategy that we
learned from G. Lowther [Low12] on MathOverflow. First, we need a classical Khintchine
type inequality in L0 -space. We fix a Rademacher sequence (ǫn )∞
n=1 on a probability space
is
a
sequence
of
independent,
identically
distributed
random variables
(Ω, P). That is, (ǫn )∞
n=1
on (Ω, P) with the distribution P(ǫn = 1) = 21 = P(ǫn = −1) (n ∈ N). The symbol E denotes
the expectation value with respect to P.
Lemma 3.12. For all n ∈ N and for all x = (xj )nj=1 ∈ Rn , we have
2
n
n
X
X
1
1
ǫ j xj ≥ ·
x2j ≥
P
.
2
12
j=1
j=1
POLISH GROUPS OF FINITE TYPE
17
Pn
2
Proof. Without loss of generality, we may assume that
j=1 xj = 1. We consider the
P
2
n
random variable f (ǫ1 , . . . , ǫn ) =
. Then, using independence, we compute
j=1 ǫj xj
Pn
2
E(f (ǫ1 , . . . , ǫn )) = j=1 xj = 1 and
2
E(f (ǫ1 , . . . , ǫn ) ) =
n
X
x4j
+3
j=1
X
x2i x2j
i6=j
= 3
n
X
j=1
2
x2j
−2
n
X
x4j ≤ 3.
j=1
Using the Paley-Zygmund inequality, which asserts that if Y ≥ 0 is a random variable with
0 6= E(Y 2 ) < ∞, then
P(Y > θ · E(Y )) ≥ (1 − θ)2 ·
E(Y )2
E(Y 2 )
for 0 ≤ θ ≤ 1,
we obtain
1 E(f (ǫ1 , . . . , ǫn ))2
1
1
≥ .
P f (ǫ1 , . . . , ǫn ) ≥ · E(f (ǫ1 , . . . , ǫn )) ≥ ·
2
2
4 E(f (ǫ1 , . . . , ǫn ) )
12
This proves the claim.
P
Proof of Proposition 3.11. Fix ξ1 , . . . , ξn ∈ H with nj=1 kξj k2 ≤ 1. By Lemma 3.7, there
exists a non-increasing function C ′ : (0, ∞) → [0, 1], such that limλ→∞ C ′ (λ) = 0 satisfying
m ({x ∈ X; |(T ξ)(x)| ≥ λ}) ≤ C ′ (λ)
for all ξ ∈ H with kξk ≤ 1.
For any constant α > 0, we compute
Z
≥
(14)
≥
X
2
n
X
ǫj T (ξj )(x) ≥ α dm(x)
P
j=1
n
X
1
T (ξj )(x)2 dm(x)
P
ǫj T (ξj )(x) ≥ ·
1 Pn
2
2
{x∈X| 2 · j=1 T (ξj )(x) ≥α}
j=1
j=1
n
1 X
1
p(α)
(T ξj (x))2 ≥ α =:
· m x ∈ X; ·
.
12
2
12
Z
n
X
2
j=1
Here in the last inequality we used Lemma 3.12. For each ω ∈ Ω, define
A(ω) :=
x ∈ X;
n
X
j=1
ǫj (ω)(T ξj )(x) ≥ α ⊂ X,
2
18
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
p(α)
and set B := {ω ∈ Ω; m(A(ω)) ≥ p(α)
24 }. If P(B) < 24 , then by Fubini Theorem,
2
Z
Z
n
X
P (ω ∈ B c ; x ∈ A(ω)) dm(x)
ǫj (T ξj )(x) ≥ α dm(x) ≤ P(B) +
P
X
X
j=1
p(α)
<
+
24
<
p(α)
,
12
Z
Bc
Z
dm(x) dP(ω)
A(ω)
which contradicts (14). Therefore P(B) ≥ p(α)
24 holds. Since H is a Hilbert space, we have the
equality
2
n
n
X
X
E
ǫj ξ j =
kξj k2 ≤ 1.
j=1
j=1
Therefore by a similar argument, we can show that the set C := {ω ∈ Ω; k
p(α)
48 .
48
p(α) }
Pn
2
j=1 εj (ω)ξj k
≤
In particular, P(B ∩ C) > 0 holds (if p(α) = 0, then C = Ω, in
satisfies P(C) ≥ 1 −
P
which case the conclusion is still true). Let ω ∈ B ∩ C 6= ∅ and set ξ = nj=1 ǫj (ω)ξj . Then
48
kξk2 ≤ p(α)
, so that
n
X
p(α) = m x ∈ X;
(T ξj (x))2 ≥ 2α ≤ 24 · m({x ∈ X; (T ξ(x))2 ≥ α})
j=1
1 !
α · p(α) 2
′
≤ 24 · C
48
2
We apply this now with λ2 = 2α. If p(α) = p( λ2 ) > λ−1 , then αp(α) > λ2 and since C ′ (·) is
non-increasing, we have
n
X
1
2
2
′
λ 2
(T ξj (x)) ≥ λ
x ∈ X;
p(α) = m
≤ 24C
96
j=1
1
λ 2
. This finishes the proof.
We can now set C(λ) := max λ−1 , 24C ′
96
We are now ready to prove the existence of an equivariant Maurey–Nikishin factorization.
Proof of Theorem 3.5. This part of the proof is exactly the same as Maurey’s argument given
as in [GR85, Theorem 3.3]. We include the proof for the reader’s convenience. Define a subset
A′ of A (see (13)) by
n
n
X
X
A′ :=
kξj k2π ≤ 1 ⊂ L0 (X, m)+ ,
|(T ξj )|2 ; n ∈ N, ξ1 , . . . , ξn ∈ H,
j=1
j=1
where kξkπ := sups∈Γ kπ(s)ξk (ξ ∈ H). Then since kπ(s)ξkπ = kξkπ (s ∈ Γ, ξ ∈ H), by (6),
the set A′ is globally Γ-invariant and convex. Since T is continuous and using Proposition
3.11, a very similar argument as in the proof of Lemma 3.7 (ii)⇒(iii) implies the following:
POLISH GROUPS OF FINITE TYPE
19
for every t > 0, there exists K(t) > 0 such that for any finite sequence ξ1 , . . . , ξn ∈ H, one
has (using (ψ 2 )⋆ = (ψ ⋆ )2 ))
⋆
n
n
X
X
(T ξj )2 (t) ≤ K(t)2
kξj k2 .
j=1
j=1
By Proposition 3.9, for every ε > 0, there exists a Γ-invariant measurable subset S(ε) ⊂ X
with m(X \ S(ε)) < ε satisfying (11) with A replaced by A′ . In particular, this shows by (11)
that for every ϕ ∈ A′ ,
Z
2
ϕ(x) dm(x) ≤ 2K 2ε
S(ε)
holds. Therefore for every ξ ∈ H, kξk2π ≤ kπk2 kξk2 (where kπk := sups∈Γ kπ(s)k), so that we
have
Z
[T ξ](x)2 dm(x) ≤ 2K( 2ε )2 kπk2 kξk2 .
S(ε)
Set Cε := 2K
ε 2
2
kπk2 > 0. Then let
ϕ(x) :=
∞
X
x ∈ X,
kj 1S(1/j) (x),
j=1
P
j→∞
where kj > 0 and ∞
j=1 kj C1/j = 1. Since Cε is non-increasing in ε, kj → 0, which implies
that ϕ ∈ L∞ (X, m). Also, ϕ is m-a.e. positive and Γ-invariant. Moreover we have:
Z
Z
∞
X
2
[T ξ](x)2 dm(x)
kj
ϕ(x)[T ξ](x) dm(x) =
X
≤
j=1
∞
X
S1/j
kj C1/j kξk2 = kξk2 .
j=1
This finishes the proof.
f is
Recall from the non-commutative integration theory that a self-adjoint operator A ∈ M
square-integrable, if
Z
2
λ2 d(τ (EA (λ)) < ∞.
kAk2 :=
R
R
f is called squareHere, A = R λ dEA (λ) is the spectral decomposition of A. Then T ∈ M
1
integrable if |T | = (T ∗ T ) 2 is square-integrable. The set of all square-integrable operators can
naturally be identified with L2 (M ) (see [Ne74] for details). We need the following simple but
useful lemma.
f be an R-linear homeomorphism onto its range. If S(ξ) ∈
Lemma 3.13. Let S : H → M
2
L (M ) for every ξ ∈ H, then regarded as S : H → L2 (M ), S is an R-linear homeomorphism onto its range, where L2 (M ) is equipped with the L2 -topology (instead of the τ -measure
topology).
20
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
Proof. We first show that in L2 (M ), the L2 -topology is stronger than the τ -measure topology.
n→∞
Let An ∈ L2 (M ) (n ∈ N) be such that kAn k2 → 0. Then
Z
1 2
1
2
−
dτ (EAn (λ))
(An − i)−1 − (0 − i)−1 2 =
0−i
R λ−i
Z
λ2
dτ (EAn (λ))
=
2
R λ +1
Z
λ2 dτ (EAn (λ))
≤
R
n→∞
= kAn k22 → 0.
This shows the claim. Note also that this proves that S −1 : S(H) ∋ S(ξ) → ξ ∈ H is
continuous:
n→∞
n→∞
kS(ξn )k2 → 0 ⇒ S(ξn ) → 0 SRT
n→∞
⇔ kξn k → 0.
Next, we show that S : H → L2 (M ) is L2 -continuous. Since H, L2 (M ) are Banach spaces,
it suffices to show that S is closed, by the Closed Graph Theorem. So assume that ξn ∈ H
n→∞
converges to 0, and S(ξn ) converges in L2 to some A ∈ L2 (M ). Then kS(ξn ) − Ak2 → 0,
n→∞
whence by the first part, we also have S(ξn ) − A → 0 in SRT. Since SRT is a vector space
f (because M is finite), we obtain S(ξn ) n→∞
topology in M
→ A in SRT. However, S is continuous
n→∞
in SRT, whence S(ξn ) → 0 in SRT. Since SRT is Hausdorff, A = 0. Therefore S is closed,
whence L2 -continuous.
Remark 3.14. In the above lemma, we did not claim that in L2 (M ), SRT and the L2 topology are the same. In fact the two topologies are different (but they agree on the image
of S, i.e., S defines a strong embedding in the sense of [AK06, Definition 6.4.4]). To see
this in the commutative case, let (Ω, µ) be a diffuse probability space, and choose a sequence
n→∞
2
(Bn )∞
n=1 of measurable subsets of Ω such that 0 < µ(Bn ) → 0. Define fn ∈ L (Ω, µ) by
fn (x) =
1
1
µ(Bn ) 2
1Bn (x),
Then kfn k22 = 1 for n ∈ N. On the other hand,
−1
k(fn − i)
− (0 −
i)−1 k2L2 (Ω,µ)
=
≤
Z
ZBn
x ∈ Ω.
1
dµ(x)
1 + µ(Bn )
n→∞
1 dµ(x) = µ(Bn ) → 0
Bn
n→∞
This shows that fn → 0 (SRT) in L0 (Ω, µ), but kfn k2 = 1 6→ 0.
Proof of Theorem 3.1. By Theorem 3.5, there exists an m-a.e. positive ϕ ∈ L0 (X, m) such
that (8) holds. We define an R-bilinear form h·, ·i′π : H × H → R by
Z
′
ϕ(x)[T ξ](x)[T η](x) dm(x),
ξ, η ∈ H.
hξ, ηiπ :=
X
POLISH GROUPS OF FINITE TYPE
21
If hξ, ξi′π = 0, then ϕ|T ξ|2 = 0 a.e., and since ϕ > 0 a.e., T ξ = 0 a.e. But since T is injective,
we get ξ = 0. Thus this is an inner product. Let s ∈ Γ. Then for ξ, η ∈ H, the Γ-invariance
of ϕ and m show that
Z
′
ϕ(x)[T ξ](s−1 x)[T η](s−1 x) dm(x)
hπ(s)ξ, π(s)ηiπ =
X
Z
ϕ(sx)[T ξ](x)[T η](x) dm(sx)
=
X
Z
ϕ(x)[T ξ](x)[T η](x) dm(x).
=
X
h·, ·i′π
Therefore
is Γ-invariant. We show that this inner product generates the same topology
on H as the original norm. Define S : H → L2 (X, m) ⊂ L2 (M ) by S(ξ) := ϕ1/2 T (ξ) =
Mϕ1/2 T (ξ). Note that the multiplication operator Mϕ1/2 is a (possibly unbounded) operator
affiliated with M . Therefore, the function
f ∋ X 7→ ϕ1/2 X ∈ M
f
Mϕ1/2 : M
is continuous with continuous inverse Mϕ−1/2 , so it is a homeomorphism. Therefore we obtain
f is also an R-linear homeomorphism onto its range. Moreover, we have
that S : H → M
the inclusion S(H) ⊂ L2 (X, m) ⊂ L2 (M ). Thus by Lemma 3.13, S : H → L2 (M ) is a
homeomorphism onto its range where L2 (M ) is equipped with the L2 -topology. Let (ξn )∞
n=1
be a sequence in H. Then
n→∞
n→∞
kξn k → 0 ⇔ kS(ξn )kL2 (M ) = kξk′π → 0.
This shows that k · k′π generates the same topology as k · k. Finally, we complexify the inner
product. Define a new sesqui-linear form h·, ·iπ : H × H → C by
1
hξ, ηiπ := {hξ, ηi′π + hiξ, iηi′π + ihξ, iηi′π − ihiξ, ηi′π },
4
′
Since h·, ·iπ is R-bilinear, so is h·, ·iπ . But if ξ, η ∈ H, then
(15)
ξ, η ∈ H.
1
hiξ, ηiπ = {hiξ, ηi′π + h−ξ, iηi′π + ihiξ, iηi′π + ihξ, ηi′π }
4
1
= {ihξ, ηi′π + ihiξ, iηi′π − hξ, iηi′π + hiξ, ηi′π }
4
= ihξ, ηiπ ,
1
hξ, iηi′ = {hξ, iηi′π − hiξ, ηi′π − ihξ, ηi′π − ihiξ, iηi′π }
4
1
= {−ihξ, ηi′π − ihiξ, iηi′π + hξ, iηi′π − hiξ, ηi′π }
4
= −ihξ, ηiπ .
Thus h·, ·iπ is sesqui-linear. Note that
1
1
kξk2π = (kξk′π )2 + (kiξk′π )2 ,
4
4
so that h·, ·iπ also generates the same Hilbert space topology. Since π(s) (s ∈ Γ) is C-linear
and preserves h·, ·i′π , it is also clear that π(s) preserves h·, ·iπ as well. By Lemma 3.2, this
shows that π is unitarizable.
22
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
Note that in our construction G = H ⋊π Γ, it is essential that Γ is discrete. It is therefore
of interest to consider the following question:
Question 3.15. Is there a connected, unitarily representable SIN Polish group which is not
of finite type?
Note also that in the proof of Theorem 3.1, we have shown the following result, which is
of independent interest:
Corollary 3.16. Let H be a separable real Hilbert space, and let f ∈ P(H). Then there exist
(i) a compact metrizable space X and a Borel probability measure m on X,
(ii) a bounded operator T : H → L2 (X, m; R), and
(iii) a positive measurable function ϕ ∈ L0 (X, m),
such that
Z
eiϕ(x)[T ξ](x) dm(x),
ξ ∈ H.
f (ξ) =
X
Moreover:
(iv) If f generates a neighborhood basis of 0, then T can be chosen to be a real-linear
homeomorphism onto its range.
(v) We may arrange T and ϕ such that if G is a uniformly bounded subgroup of the group
{u ∈ GL(H); f (uξ) = f (ξ) (ξ ∈ H)} of invertible operators on H fixing f pointwise,
then there exists a continuous homomorphism β from G to the group Aut(X, m) of all
m-preserving automorphisms of X with the weak topology, such that
[T (uξ)](x) = β(u)[T (ξ)](x),
ϕ(β(u)x) = ϕ(x),
u ∈ G, ξ ∈ H, m-a.e. x ∈ X.
One may regard Corollary 3.16 as an equivariant version of the classical result (cf. [AMM85,
Lemma 4.2]) that any f ∈ P(H) is of the form
Z
eit[U ξ](x) dm(x),
ξ ∈ H, t ∈ R
f (tξ) =
X
for a probability space (X, m) and a continuous linear map U : H → L0 (X, m).
Remark 3.17. A typical example of the above corollary is the positive definite function
2
f (ξ) = e−kξk /4 . To see this, let us fix an orthogonal basis {ξi } for H. In this case, we can
choose X = (Ṙ)N the product of an infinite number of copies of the one-point compactification
of R, m = the infinite dimensional Gaussian measure, T (ξi ) : X ∋ x 7→ xi ∈ Ṙ and ϕ = 1.
The triple (X, m, T ) is called a Gaussian random process indexed by H. For more detatils,
see [Si74].
Remark 3.18. It is clear from the proof that it applies whenever H is a Banach space of
Rademacher type 2 (see [AK06, §6] for definitions). Thus we have shown as a corollary that a
Banach space B is isomorphic (as a Banach space) to a Hilbert space if and only if a positive
definite function generates the topology of B and is of type 2. Moreover, the isomorphism can
be made equivariant with respect to a group of symmetries that preserve the positive definite
function. One may compare this result with Kwapień’s Theorem [Kw72] that B is isomorphic
to a Hilbert space if and only if B is of type 2 and cotype 2. Note that the Kwapień Theorem
does not produce an equivariant isomorphism but requires less in the sense that any Banach
space isomorphic to a subspace of L0 (X, m) is of cotype 2, but not all cotype 2 Banach spaces
can be isomorphic to a subspace of L0 (X, m) (see the remark after [BL00, Corollary 8.17] and
[Pi78]).
POLISH GROUPS OF FINITE TYPE
23
Appendix
The following two results are well-known, but we include the proofs for the reader’s convenience.
Proposition 3.19. Let A be a commutative von Neumann algebra with separable predual, and
let τ be a normal faithful tracial state on A. Let G be a Polish group, and let β : G → Aut(A)
be a continuous τ -preserving action. Then there exists a separable compact metrizable space
X, a Borel probability measure m on X, an m-preserving continuous action G × X ∋ (g, x) 7→
gx ∈ X and a ∗-isomorphism Φ : A → L∞ (X, m) such that
Z
(16)
Φ(a)(x) dm(x),
a ∈ A,
τ (a) =
X
(17)
Φ(βg (a))(x) = Φ(a)(g−1 x),
g ∈ G, µ−a.e. a ∈ A.
Moreover, the G-action is continuous as a map G → Aut(X, m), where the group Aut(X, m)
of all m-preserving automorphism of X is equipped with the weak topology.
Lemma 3.20. Let M, N be von Neumann algebras, and let ψ, ϕ be normal faithful states on
M, N , respectively. Let M0 (resp. N0 ) be a ∗-strongly dense subalgebra of M (resp. N ). If
α : M0 → N0 is a ∗-isomorphism such that ϕ(α(x)) = ψ(x) for every x ∈ M0 , then α can be
extended to a ∗-isomorphism from M onto N such that ϕ(α(x)) = ψ(x) (x ∈ M ).
Proof. By the GNS construction, we may assume that M acts on L2 (M, ψ) and N acts
on L2 (N, ϕ) so that ψ = h · ξψ , ξψ i, ϕ = h · ξϕ , ξϕ i for unit vectors ξψ ∈ L2 (M, ψ) and
ξϕ ∈ L2 (N, ϕ). Define U0 : M0 ξψ → N0 ξϕ by
U0 xξψ := α(x)ξϕ ,
x ∈ M0 .
Since ψ is faithful, U0 is well-defined. Also, since M0 is ∗-strongly dense in M and ξψ is cyclic
for M , M0 ξψ is dense in L2 (M, ψ). Similarly, N0 ξϕ is dense in L2 (N, ϕ). By ϕ ◦ α = ψ on
M0 , we have
kU0 xξψ k2 = ϕ(α(x∗ x)) = ψ(x∗ x) = kxξψ k2 , x ∈ M0 .
Therefore, U0 extends to an isometry U from L2 (M, ψ) into L2 (N, ϕ) and the range of U
contains a dense subspace N0 ξϕ . Therefore U is onto. If x ∈ M0 , then for each y ∈ N0 ,
U xU ∗ yξϕ = U xα−1 (y)ξψ = α(x)yξϕ ,
so that by the density of N0 ξϕ , we have
U xU ∗ = α(x),
x ∈ M0 .
Let now x ∈ M . Then there exists a net (xi )i∈I in M0 converging ∗-strongly to x. Then
U xi U ∗ = α(xi ) (∈ N ) converges ∗-strongly to U xU ∗ . Therefore U xU ∗ ∈ N . Then it is clear
that M ∋ x 7→ U xU ∗ ∈ N defines a ∗-isomorphism with inverse N ∋ y 7→ U ∗ yU ∈ M .
The uniqueness of the extension is clear by the ∗-strong density of M0 and the fact that
isomorphisms between von Neumann algebras are normal. It is now clear that ϕ ◦ α = ψ on
M . This finishes the proof.
Proof of Proposition 3.19. Let A0 := {x ∈ A; limg→1 kαg (x) − xk = 0}, which is a normdense C∗ -subalgebra of A. Let {an }∞
n=1 be a ∗-strongly dense subset of A0 , and let C =
C ∗ ({βg (an ); g ∈ G, n ∈ N} ∪ {1}). Then since G is separable and g 7→ βg (an ) is normcontinuous for all n ∈ N, C is a norm-separable ∗-strongly dense C∗ -subalgebra of A, and let
X = Spec(C) (the Gelfand spectrum), which is a separable and compact metrizable space,
24
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
and let Φ : C → C(X) be the Gelfand transform. Since τ ◦ Φ−1 ∈ C(X)∗+ , there exists a Borel
probability measure m on X such that
Z
Φ(c)(x) dm(x),
c ∈ C.
τ (c) =
X
L2 (X, m)
We let C(X) act on
by multiplication, so C(X) ⊂RL∞ (X, m) is a ∗-strongly dense
∞
∗-subalgebra of L (X, m). The map τm : L∞ (X, m) ∋ f 7→ X f dm ∈ C is a normal faithful
tracial state on L∞ (X, m). Moreover, τm (Φ(c)) = τ (c) for every c ∈ C. Therefore by Lemma
3.20, Φ can be extended to a ∗-isomorphism Φ : C ′′ = A → C(X)′′ = L∞ (X, m) such that
τm (Φ(a)) = τ (a) for every a ∈ A. This shows (16). Now let g ∈ G and x ∈ X. Let
mx ∈ C(X)∗ be the delta probability measure at x ∈ X. Let β ∗ be the dual action of G on
C(X)∗ given by
(18)
hβg∗ (ν), f i := hν, Φ ◦ βg−1 ◦ Φ−1 (f )i,
ν ∈ C(X)∗ , f ∈ C(X), g ∈ G.
Since β is an action, βg∗ maps pure states (on C ∼
= C(X)) to pure states, and pure states on
C(X) are precisely the delta probability measures, there exists a unique point g · x ∈ X such
that
βg∗ (mx ) = mg·x .
We show that G × X ∋ (g, x) 7→ g · x ∈ X is a continuous m-preserving action. It is routine to
check that this is indeed an action of G on X. We show that the action is continuous. Since
G is a Polish group and X is a Polish space, it suffices to show that the action is separately
continuous (see e.g., [Kec96, Theorem 9.14]). Suppose that (xn )∞
n=1 ⊂ X converges to x ∈ X
n→∞
and let g ∈ G. Then by the definition of the Gelfand spectrum, mxn → mx (weak*). Since
βg induces an action on C (C is β-invariant by construction), we have
n→∞
βg∗ (mxn ) = mg·xn → βg∗ (mx ) = mg·x ,
n→∞
which implies g·xn → g·x (again by the definition of the topology on the Gelfand spectrum).
Let now x ∈ X and (gn )∞
n=1 ⊂ G be a sequence converging to g ∈ G. Then for each f ∈ C(X),
n→∞
hβg∗n (mx ), f i = hmx , Φ ◦ βgn−1 ◦ Φ−1 (f )i → hmx , Φ ◦ βg−1 ◦ Φ−1 (f )i
= hβg∗ (mx ), f i,
n→∞
whence gn · x → g · x. Therefore the G-action on X is continuous. Next, we show that for
each g ∈ G, the map X ∋ x 7→ g · x ∈ X preserves m. It suffices to show that
Z
Z
−1
f (x) dm(x),
f ∈ C(X).
f (g x) dm(x) =
Let ν =
Pn
i=1 λi mxi
X
X
∈ conv{mx ; x ∈ X} =: D. Then
βg∗ (ν) =
n
X
λi δg·xi = ν(g −1 · ).
i=1
By the Krein-Milman Theorem, D is weak*-dense in the state space of C(X). This shows
that
(19)
βg∗ (ν) = ν(g −1 · ),
ν ∈ C(X)∗+ .
POLISH GROUPS OF FINITE TYPE
25
In particular, by τ ◦ βg = τ ,
Z
Z
(19)
−1
f (x) dm(g(x)) = hβg∗−1 (m), f i
f (g x) dm(x) =
X
X
= hm, Φ ◦ βg ◦ Φ−1 (f )i
= τm (Φ ◦ βg ◦ Φ−1 (f )) = τ (βg ◦ Φ−1 (f ))
= τ (Φ−1 (f )) = τm (f )
Z
f (x) dm(x).
=
X
Therefore the G-action on X is m-preserving. Finally, we show (17). Since the map
βg′ : L∞ (X, m) ∋ f 7→ f (g −1 · ) ∈ L∞ (X, m)
defines a τm -preserving action and since C(X) is ∗-strongly dense in L∞ (X, m), it suffices to
show that (17) holds for every f = Φ(a) ∈ C(X) (a ∈ C), x ∈ X, and g ∈ G. But
f (g −1 x) = hµg−1 ·x , f i = hβg∗−1 (µx ), f i
= hµx , Φ ◦ βg ◦ Φ−1 (f )i
= Φ ◦ βg ◦ Φ−1 (f )(x).
n→∞
Finally, assume that gn → g in G. Then for every measurable set A ⊂ X, we have
Z
|Φ ◦ βgn−1 ◦ Φ−1 (1A )(x) − Φ ◦ βg−1 ◦ Φ−1 (1A )(x)| dm(x)
m(gn A △ gA) =
X
n→∞
≤ kβgn−1 ◦ Φ−1 (1A ) − βg−1 ◦ Φ−1 (1A )k2 → 0,
because the u-topology on Aut(A) is the same as the topology of pointwise k · k2 -convergence.
Thus the G-action defines a continuous homomorphism G → Aut(X, m). This finishes the
proof.
Acknowledgments
We thank Professors Lukasz Grabowski, Magdalena Musat, Yuri Neretin, Izumi Ojima and
Mikael Rørdam for helpful discussions on the Ufin -problem at early stages of the project. The
final part of the manuscript was completed while HA was visiting the Mittag-Leffler Institute
for the program “Classification of operator algebras: complexity, rigidity, and dynamics”. He
thanks the institute and the workshop organizers for the invitation and for their hospitality.
This research was supported by JSPS KAKENHI 16K17608 (HA), 6800055 and 26350231
(YM), ERC Starting Grant No. 277728 (AT) and the Danish Council for Independent Research through grant no. 10-082689/FNU. (AT), from which HA was also supported while
he was working at University of Copenhagen. Last but not least, we would like to thank the
anonymous referee for suggestions which improved the presentation of the paper.
References
[AMM85] I. Aharoni, B. Maurey and B. S. Mityagin, Uniform embeddings of metric spaces and of Banach
spaces into Hilbert spaces, Israel J. Math. 52 (1985), 251–265.
[AK06]
F. Albiac and N. Kalton, Topics in Banach space theory, Graduate Texts in Mathematics 233,
Springer (2006).
26
HIROSHI ANDO, YASUMICHI MATSUZAWA, ANDREAS THOM, AND ASGER TÖRNQUIST
[AM12-1] H. Ando and Y. Matsuzawa, Lie Group-Lie Algebra Correspondences of Unitary Groups in Finte
von Neumann Algebras, Hokkaido Math. J. 41 (2012), 31–99.
[AM12-2] H. Ando and Y. Matsuzawa, On Polish Groups of Finite Type, Publ. RIMS vol. 48 (2012), 389–408.
[Be10]
D. Beltita, Lie theoretic significance of the measure topologies associated with a finite trace, Forum
Math. 22 (2010), 241–253.
[BL00]
Y. Benyamini and J. Lindenstrauss, Geometric nonlinear functional analysis Vol. 1, American Mathematical Society Colloquium Publications, 48. American Mathematical Society, Providence, RI
(2000).
[Bo87]
M. Bozejko, Uniformly bounded representations of free groups, J. Reine Angew. Math. 377 (1987),
170–186.
[Da50]
M. Day, Means for the bounded functions and ergodicity of the bounded representations of semigroups, Trans. Amer. Math. Soc. 69 (1950), 276–291.
[Di50]
J. Dixmier, Les moyennes invariantes dans les semi-groupes et leurs applications, Acta Sci. Math.
Szeged 12 (1950) 213-227.
[EM55] L. Ehrenpreis and F. Mautner, Uniformly bounded representations of groups, Proc. Nat. Acad. Sci.
U. S. A. 41 (1955), 231–233.
[EM09] I. Epstein and N. Monod, Nonunitarizable representations and random forests, Int. Math. Res. Not.
IMRN 22 (2009), 4336–4353.
[Gal09]
J. Galindo, On unitary representability of topological groups, Math. Z. 263 (2009), 211–220.
[Gao05] S. Gao, Unitary group actions and Hilbertian Polish metric spaces, Contemp. Math. 380, Amer.
Math. Soc., Providence, RI, 2005.
[GR85]
J. Garcı́a-Cuerva and J. L. Rubio de Francia, Weighted norm inequalities and related topics, NorthHolland Mathematics Studies 116 (1985).
[Kec96] A. S. Kechris, Classical descriptive set theory, Springer (1996).
[Kw72]
S. Kwapień, Isomorphic characterizations of inner product spaces by orthogonal series with vector
valued coefficients, Studia Math. 44 (1972), 583–595.
[Low12] G. Lowther, a comment on Mathoverflow at http://mathoverflow.net/questions/98410
[MZ83-1] A. M. Mantero and A. Zappa, The Poisson transform and representations of a free group, J. Funct.
Anal. 51 (1983), 372–399.
[MZ83-2] A. M. Mantero and A. Zappa, Uniformly bounded representations and Lp -convolution operators on
a free group, Harmonic analysis (Cortona, 1982), 333–343, Lecture Notes in Math., 992, Springer
(1983).
[Ma72]
B. Maurey, Théorèmes de Nikishin: théorèmes de factorisation pour les applications lineaires a
valeurs dans un espace L0 (Ω, µ), Seminaire dfanalyse fonctionnelle (Polytechnique) 10 and 11 (19721973), 1–10.
[Me02]
M. Megrelishvili, Reflexively but not unitarily representable topological groups. Topology Proc. 25
(2002), 615–625.
[MO10] N. Monod and N. Ozawa, The Dixmier problem, lamplighters and Burnside groups, J. Funct. Anal.
258 no. 1 (2010), 255–259.
[Ne74]
E. Nelson, Notes on non-commutative integration, J. Funct. Anal. 15 (1974), 103–116.
[Ni70]
E. M. Nikishin, Resonance theorems and superlinear operators (Russian), Uspehi Mat. Nauk 25
(156) (1970), 129–191.
[NT51]
M. Nakamura and Z. Takeda, Group representation and Banach limit, Tohoku Math. J. 3 (1951),
132–135.
[Os09]
D. Osin, L2 -Betti numbers and non-unitarizable groups without free subgroups, Int. Math. Res.
Not. IMRN 22 (2009), 4220–4231.
[Oz06]
N. Ozawa, An Invitation to the Similarity Problems (after Pisier) (2006), available at
http://www.kurims.kyoto-u.ac.jp/∼narutaka/notes/similarity.pdf
[Po07]
S. Popa, Cocycle and orbit equivalence superrigidity for malleable actions of w-rigid groups, Invent.
Math. 170 (2007), 243–295.
[Pi78]
G. Pisier, Some results on Banach spaces without local unconditional structure, Compositio Math.
37 (1978), 3–19.
[Pi86]
G. Pisier, Factorization of linear operators and geometry of Banach spaces, CBMS Regional Conference Series in Mathematics, vol. 60, Published for the Conference Board of the Mathematical
Sciences, Washington, DC (1986).
POLISH GROUPS OF FINITE TYPE
[Pi01]
[Pi05]
[PS86]
[Si74]
[SzN47]
[Sz88]
[VZ74]
[Ya85]
[Ye71]
27
G. Pisier, Similarity problems and completely bounded maps. Second, Lecture Notes in Mathematics,
1618. Springer- Verlag, (2001).
G. Pisier, Are unitarizable groups amenable?, in: Infinite groups: geometric, combiatorial and
dynamical aspects, Progress in Mathematics 248, Birkhäuser (2005), 323–362.
T. Pitlyk and R. Szwarc, An analytic family of uniformly bounded representations of free groups,
Acta. Math. 157, 12 (1986), 287–309.
B. Simon, The P (Φ)2 Euclidean (Quantum) Field Theory, Princeton University Press (1974)
B. Sz.-Nagy, On uniformly bounded linear transformations in Hilbert space, Acta Sci. Math. (Szeged)
11 (1947), 152–157.
R. Szwarc, An analytic series of irreducible representations of the free group, Ann. Inst. Fourier
(Grenoble) 38 (1988), 87–110.
F. H. Vasilescu and L. Zsido, Uniformly bounded groups in finite W ∗ -algebras, Acta Sci. Math.
(Szeged), 36 (1974), 189–192.
Y. Yamasaki, Measures on infinite dimensional spaces, Series in Pure Mathematics vol. 5, World
Scientific (1985)
F. J. Yeadon, A new proof of the existence of a trace in a finite von Neumann algebra, Bull. Amer.
Math. Soc. 77 (1971), 257–260.
Hiroshi Ando, Department of Mathematics and Informatics, Chiba University, 1-33 Yayoi-cho,
Inage, Chiba, 263-0022, Japan
E-mail address: [email protected]
Yasumichi Matsuzawa, Department of Mathematics, Faculty of Education, Shinshu University,
6 -Ro, Nishinagano, Nagano City 380-8544, Japan
E-mail address: [email protected]
Andreas Thom, Fachbereich Mathematik, Institut für Geometrie, TU Dresden, 01062 Dresden,
Germany
E-mail address: [email protected]
Asger Törnquist, Department of Mathematical Sciences, University of Copenhagen, Universitetspark 5, 2100 Copenhagen, Denmark
E-mail address: [email protected]
| 4math.GR
|
Convergence and State Reconstruction of Time-varying
Multi-agent Systems from Complete Observability Theory
Brian D. O. Anderson, Guodong Shi, and Jochen Trumpf∗
arXiv:1512.06244v2 [cs.SY] 9 Aug 2016
Abstract
We study continuous-time consensus dynamics for multi-agent systems with undirected switching
interaction graphs. We establish a necessary and sufficient condition for exponential asymptotic consensus based on the classical theory of complete observability. The proof is remarkably simple compared
to similar results in the literature and the conditions for consensus are mild. This observability-based
method can also be applied to the case where negatively weighted edges are present. Additionally, as a
by-product of the observability based arguments, we show that the nodes’ initial value can be recovered
from the signals on the edges up to a shift of the network average.
1
Introduction
1.1
Motivation
Distributed consensus algorithms [1, 2, 3, 4] of multi-agent systems have become the foundation of many
distributed solutions to in-network control and estimation [5, 6], signal processing [7], and optimization
[8, 9]. Convergence conditions of consensus dynamics often play a key role in the performance of such
distributed designs, which are relatively well understood under discrete-time network dynamics when the
underlying interaction graph is either time-invariant [2, 3, 14] or time-varying [11, 13].
Continuous-time consensus dynamics, however, is difficult to analyze with switching interaction graphs.
For piece-wise constant graph signals with dwell time, we can treat the continuous-time flow by investigating the node state evolution along a selected sequence of time instants and then work on the discrete-time
evolution at the sampled times [1, 4]. For graphs that vary in structure and edge weights in a general
manner, the problem becomes much more challenging. Under such circumstances the maximum node state
difference serves as one major tool for convergence analysis [20, 12, 15].
In this paper, we study the convergence of continuous-time consensus dynamics with undirected and
switching interaction graphs from the classical theory on uniformly complete observability of linear timevarying systems [17, 19]. It turns out that the complexity of the convergence analysis can be drastically reduced in view of these classical results even under mild conditions. Interestingly enough, this
∗
Authors are listed in alphabetical order. B. D. O. Anderson is with the National ICT Australia (NICTA), and the
Research School of Engineering, The Australian National University, Canberra, Australia. G. Shi and J. Trumpf are with the
Research School of Engineering, The Australian National University, ACT 0200, Canberra, Australia. Email: brian.anderson,
guodong.shi, [email protected].
1
observability-based method can also be used to analyze consensus dynamics over networks with both positively and negatively weighted edges [21, 22], where methods based on maximum node state difference
function are no longer applicable.
1.2
The Model
Consider a multi-agent network with N nodes indexed in the set V = {1, . . . , N }. Each node i holds a
state xi (t) ∈ R for t ∈ R≥0 . The dynamics of xi (t) are described by
N
X
d
aij (t) xj (t) − xi (t) ,
xi (t) =
dt
(1)
j=1
where the aij (t) = aji (t) are, for the moment, assumed to be nonnegative numbers representing the weight
of the edge between node i and j. For the weight functions aij (t), we impose the following assumption as
a standing assumption throughout our paper.
Weights Assumption. (i) Each aij (t) is regulated in the sense that one-sided limits exist for all t ≥ 0;
(ii) The aij (t) are almost everywhere bounded, i.e., there is a constant A∗ > 0 such that |aij (t)| ≤ A∗ for
all i, j ∈ V and for almost all t ≥ 0.
The underlying interaction graph of (1) at time t is defined as an undirected (weighted) graph Gt =
(V, Et ), where {i, j} ∈ Et if and only if aij (t) > 0. We introduce the following definition on the connectivity
of the time-varying graph process {Gt }t≥0 .
Definition 1 The time-varying graph {Gt }t≥0 is termed jointly (δ, T )-connected if there are two real
numbers δ > 0 and T > 0 such that the edges
Z s+T
aij (t)dt ≥ δ, i, j ∈ V
{i, j} :
(2)
s
form a connected graph over the node set V for all s ≥ 0.
Along the system (1), it is obvious to see that the average of the node states
xave (t) :=
PN
j=1 xj (t)
N
is preserved over time. Therefore, if a consensus, i.e., all nodes asymptotically reaching a common state, is
achieved, then all the node states will converge to xave (t0 ) where t0 ≥ 0 is the system initial time. Denote by
1n the n-dimensional all one vector and x(t) = (x1 (t) · · · xN (t))T . We can therefore conveniently introduce
the following definition.
Definition 2 Exponential asymptotic consensus is achieved if there are two constants α > 0 and β > 0
such that
|xi (t) − xave (t0 )| ≤ βkx(t0 ) − xave (t0 )1N ke−α(t−t0 )
for all i ∈ V, all x(t0 ), and all t ≥ t0 .
2
1.3
Main Results
First of all we present a necessary and sufficient condition for exponential asymptotic consensus of the
system (1).
Theorem 1 Exponential asymptotic consensus is achieved for the system (1) if and only if there are δ > 0
and T > 0 such that the graph {Gt }t≥0 is jointly (δ, T )-connected.
Theorem 1 is established in view of the classical complete observability theory for linear time-varying
systems [18, 19]. A by-product of such observability-based analysis is a state reconstruction theorem, which
indicates that node states can be recovered from the signals on the edges up to a shift of the network
average. We recall the following definition.
Definition 3 ([17]) A linear time-varying system
ẋ = F (t)x
(3)
y = D T (t)x
(4)
is uniformly completely observable if there are some positive α1 , α2 and δ such that
Z s+δ
ΦT (t, s)D(t)D T (t)Φ(t, s)dt ≤ α2 I
α1 I ≤
(5)
s
for all s ≥ 0. Here Φ(·, ·) is the state transition matrix for system (3).
If the system (3) and (4) is uniformly completely observable, then
Z s+δ
−1
x(s) =
ΦT (t, s)D(t)D T (t)Φ(t, s)dt
s
Z s+δ
ΦT (t, s)D(t)y(t)dt.
(6)
s
In this way, x(s) is uniformly recovered from the signal y(t) : t ∈ [s, s + δ].
Theorem 2 Assume full knowledge of {Gt }t≥0 with the edge weights. The shifted node state xi (s) −
xave (0), i ∈ V can be uniformly recovered from the signals xj (t) − xi (t), {i, j} ∈ Et over the edges if there
are two constants δ > 0 and T > 0 such that the graph {Gt }t≥0 is jointly (δ, T )-connected.
1.4
Generalizations
We now present a few direct generalizations of the above established convergence results.
1.4.1
Robust Consensus
Further consider the system (1) subject to noises:
N
X
d
aij (t) xj (t) − xi (t) + wi (t), i ∈ V,
xi (t) =
dt
(7)
j=1
where wi (t) is a piecewise continuous function defined over R≥0 for each i ∈ V. We introduce the following
notion of robust consensus.
3
Definition 4 The system (7) achieves ∗-bounded robust consensus if for all w(t) := (w1 (t) . . . wN (t))T
satisfying
Z
s
s+ζ
wT (t)w(t)dt ≤ B0
(8)
for some ζ > 0, some B0 ≥ 0, and all s ≥ 0, there exists C(ζ, B0 ) > 0 such that
kxi (t) − xave (t)k ≤ C(ζ, B0 ), ∀i ∈ V, ∀t ≥ t0
(9)
when x1 (t0 ) = x2 (t0 ) = · · · = xN (t0 ).
The following robust consensus result holds.
Theorem 3 The system (7) achieves ∗-bounded robust consensus if and only if there are δ > 0 and T > 0
such that the graph {Gt }t≥0 is jointly (δ, T )-connected.
1.4.2
Signed Networks
Up to now we have assumed that all the aij (t) are non-negative numbers. In fact, the complete observability
analysis can be easily extended to the case when some of the aij (t) are negative. The motivation1 of
studying the effect of possibly negative aij (t) comes from the modelling of misbehaved links in engineering
network systems, as well as opinion dynamics over social networks with mistrustful interactions represented
by negative edges [22].
Although conventionally, the graph Laplacian is considered only for graphs with non-negative edges,
for any t ≥ 0, we continue to define the (weighted) Laplacian of the graph Gt as
LGt = DGt − AGt
PN
P
with A(Gt ) = [aij (t)] and D(Gt ) = diag( N
j=1 aN j (t)). Note that although the aij (t) can
j=1 a1j (t), . . . ,
now be negative, LGt remains symmetric with at least one zero eigenvalue with 1N being a corresponding
eigenvector. Introduce the following assumption.
Negative-Link Assumption. The Laplacian LGt is positive semi-definite for all t ≥ 0.
This negative-link assumption requires that the influence of the negative links can be reasonably overcome by the positive links for any t ≥ 0. The following result holds.
Theorem 4 Consider the system (1) with possibly negative aij (t). Assume the Negative-Link Assumption.
Then Theorems 1 and 3 continue to hold.
1.5
Some Remarks
Theorems 1 and 3 are based on a direct application of the classical complete observability theory for linear
time-varying systems [18, 19], after some useful properties of the Laplacian of the time-varying network
1
Another definition for a negative link between i and j is to replace xj by −xj in the system (1) [23]. The difference of
the two definitions of negative links was discussed in [22] from a social network point of view.
4
have been established. The theorems recover the results in [1, 20, 15] for undirected graphs under weaker
weight and noise assumptions. We would also like to emphasize that the applicability of the complete
observability theory relies heavily on undirected node interactions, and hence we believe that it will be
difficult to extend the analysis to the directed case.
Moreover, Theorem 4 illustrates that if the effect of the negative links is somewhat moderated by the
positive links, i.e., as in the Negative-Link Assumption, then the convergence property is not affected by
the negative links. This point, however, cannot be seen from the max state difference analysis used in
prior work [20, 15] because obviously the max state difference function is no longer non-increasing in the
presence of negative links, regardless of how small their weights are compared to the positive ones.
Theorem 2 discusses the possibility of reconstructing node initial states from signals over the edges.
Naturally one may wonder whether we can recover the nodes’ initial values by observing the state signals
at one or more nodes. By duality this is equivalent to the controllability of the network by imposing control
inputs over one or more anchor nodes. This problem, as shown in [24, 25], is challenging in general for
a full theoretical characterization on general graphs, and in fact, finding the minimum number of nodes
to control to ensure network controllability is a computationally intensive problem for general network
dynamics [26].
2
Proofs of Statements
2.1
Proof of Theorem 1
Consider the following linear time-varying system
ẋ = −V (t)V T (t)x,
(10)
where x ∈ Rn and V (·) is a piecewise continuous and almost everywhere bounded matrix function mapping
from R≥0 to Rn×r . Let Φ(·, ·) be the state transition matrix of the System (10). Our proof relies on the
following result which was proved using complete observability theory in [18].
Lemma 1 System (10) is exponentially asymptotically stable in the sense that there are γ1 , γ2 > 0 such
that kΦ(t, s)k ≤ γ1 e−γ2 (t−s) for all t ≥ s ≥ 0, if and only if for some positive α1 , α2 and δ, there holds
Z s+δ
V (t)V T (t)dt ≤ α2 I
(11)
α1 I ≤
s
for all s ∈ R≥0 .
Remark 1 To serve the technical purpose of the proof, we have made a slight variation in Lemma 1
compared to the original Theorem 1 in [18], where we have replaced the lower bound on exponentially
asymptotic stability required in [18] with the almost everywhere boundedness of V (·) adopted in the current
paper.
The system (1) can be written into the following compact form
ẋ = −LGt x.
5
(12)
Since LGt is positive semi-definite for all t ≥ 0 [16], we can easily construct V (t) so that the system (12)
can be re-written into the form of (10). Although any construction of V (t) would do, we use the incidence
matrix HGt . We make the following definition.
Definition 5 The N ×
N (N −1)
2
weighted incidence matrix of Gt , denoted as HGt , with rows indexed by
the nodes v ∈ V and columns indexed by the edges {i, j} (abbreviated as {ij}), i < j ∈ V, is defined as
HGt = [Hv−{ij} (t)] where2
We can now write (12) as
p
aij (t), if v = i;
−
p
Hv−{ij} (t) =
aij (t),
if v = j;
0,
otherwise.
T
ẋ = −LGt x = −HGt HG
x.
t
(13)
(14)
Define y(t) = (I − 1N 1T
N /N )x(t). Then we have
ẏ = (I − 1N 1T
N /N )ẋ(t)
= −(I − 1N 1T
N /N )LGt x(t)
a)
T
= −(LGt + 1N 1T
N /N )(I − 1N 1N /N )x(t)
= −(LGt + 1N 1T
N /N )y(t)
1N 1T T
1N 1T
b)
H Gt + √ r
y(t),
= − H Gt + √ r
rN
rN
(15)
T
where r = N (N − 1)/2, a) follows from the equations that (I − 1N 1T
N /N )LGt = LGt (I − 1N 1N /N ) and
2
T
T
(1N 1T
N /N ) = 1N 1N /N , and b) follows from direct computation in view of LGt = HGt HGt .
We can now conclude that the following statements are equivalent.
(i) System (12) achieves exponential consensus;
(ii) System (15) is exponentially asymptotically stable;
(iii) There are some positive α1 , α2 and T , such that
s+T
α1 I ≤
Z
s+T
=
Z
s
s
1N 1T
1N 1T T
H Gt + √ r
H Gt + √ r
dt
rN
rN
(LGt + 1N 1T
N /N )dt ≤ α2 I;
(16)
for all s ∈ R≥0 .
(iv) there are δ, T > 0 such that the graph {Gt }t≥0 is (δ, T )-connected.
2
In this definition we have used an orientation for the graph Gt in which i is the tail of an edge {i, j} if i < j. There
T
always holds LGt = HGt HG
independent of the choice of orientation in the definition of the incidence matrix HGt [16].
t
6
The equivalence of (i) and (ii) is straightforward. Lemma 1 gives the equivalence between (ii) and (iii).
We only need to show the equivalence between (iii) and (iv).
R s+T
T
(iii) ⇒ (iv): If (16) holds, then s (LGt + 1N 1T
N /N )dt is positive definite. Note that 1N 1N /N has
a unique positive eigenvalue equal to 1 and another eigenvalue equal to zero with multiplicity N − 1.
T /N L
Moreover, for any fixed t ≥ 0, LGt 1N 1T
/N
=
1
1
N
Gt = 0, and the zero eigenvalue of LGt and
N
N
the eigenvalue one of 1N 1T
N /N share a common eigenvector 1N . This implies that the second smallest
R s+T
R s+T
R s+T
eigenvalue of s
LGt dt, λ2 ( s
LGt dt), must satisfy λ2 ( s
LGt dt) ≥ α1 .
Consider λ2 (L), the second smallest eigenvalue of L, where L takes value from
L := L ∈ RN ×N : L is symmetric; [L]ij ≤ 0, i 6= j;
L1n = 0 .
Certainly λ2 (·) is a continuous function over the set L . In turn, there must exist δ > 0 (which depends
on α1 ) such that the edges
{i, j} :
Z
s+T
s
aij (t)dt ≥ δ, i, j ∈ V
(17)
form a connected graph over the node set V. In other words, the graph {Gt }t≥0 is (δ, T )-connected.
R s+T
(iv) ⇒ (iii): The matrices s
LGt dt, for all s ≥ 0 and all (δ, T )-connected graphs LGt , form a compact set
in L . Therefore, again noticing that λ2 (·) is a continuous function over the set L , we can find a constant
R s+T
α1 > 0 (which depends on δ) such that λ2 ( s
LGt dt) ≥ α1 for all s ≥ 0 and for all (δ, T )-connected
graph {Gt }t≥0 . Re-using the above properties of the two matrices LGt and 1N 1T
N /N we immediately get
the lower bound part of (16). The existence of α2 in (16) is always guaranteed by the Weights Assumption
on the almost everywhere boundedness of the aij (t).
We have now proved Theorem 1 by the equivalence of (i) and (iv).
2.2
Proof of Theorem 2
From the proof of Theorem 1, if there are δ, T > 0 such that the graph G(t) is (δ, T )-connected, then
there are some positive α1 , α2 and T , there holds
Z s+T
1N 1T
1N 1T T
H Gt + √ r
α1 I ≤
H Gt + √ r
dt ≤ α2 I;
rN
rN
s
for all s ≥ 0. This in turn implies that the following system (see [18])
ẏ = −(LGt + 1N 1T /N )y;
N
T
z = HG + 1√N 1Tr
y.
t
rN
(18)
(19)
is uniformly completely observable.
Note that
T
1N 1T
r
√
z(t) = HGt +
y(t)
rN
1r 1T
T
N
√
(I − 1N 1T
+
= HG
N /N )x(t)
t
rN
T
x(t).
= HG
t
(20)
This immediately translates to the desired theorem.
7
2.3
Proof of Theorem 3
Again we consider y(t) = (I − 1N 1T
N /N )x(t). Along the system (7) the evolution of y(t) obeys
d
T
y(t) = −(LGt + 1N 1T
N /N )y(t) + (I − 1N 1N /N )w(t).
dt
(21)
It is then obvious that system (7) achieves ∗-bounded robust consensus under Definition 4 if system (21)
is bounded*-input, bounded state stable as defined in [17].
If the graph {Gt }t≥0 is (δ, T )-connected, then (15) is exponentially asymptotically stable. This in
turn leads to the conclusion that the system (21) is bounded*-input, bounded state stable based on the
sufficiency part of Theorem 1 (p. 404) in [17]. Therefore, system (7) achieves ∗-bounded robust consensus.
On the other hand, note that there holds
T
T
(LGt + 1N 1T
N /N )(I − 1N 1N /N ) = (LGt + 1N 1N /N )
(22)
for all t. Consequently, we have
Φ(t, s)(I − 1N 1T
N /N ) = Φ(t, s)
(23)
for all t and s, where Φ(t, s) is the state-transition matrix of the system (21) with w(t) ≡ 0. Therefore,
applying the necessity proof of Theorem 1 in [17] we can directly conclude that if system (21) is bounded*input, bounded state stable, then (15) is exponentially asymptotically stable. This in turn leads to the
graph {Gt }t≥0 being (δ, T )-connected.
We have now completed the proof.
2.4
Proof of Theorem 4
When the Negative-Link Assumption holds, the construction of (14) can be replaced by
p
p
ẋ = −LGt x = − LGt LGt x.
(24)
It is then straightforward to see that the remaining arguments in the proofs of Theorems 1 and 3 will not
be affected, leading to the desired result.
3
Conclusions
We have studied continuous-time consensus dynamics with undirected switching interaction graphs in
view of the classical theory of uniform complete observability for linear time-varying systems. A necessary
and sufficient condition for exponential asymptotic consensus was established by much simplified analysis compared to related results in the literature. Interestingly and importantly, this observability-based
method can also be applied to signed networks with both positive and negative edges. We also established
a robust consensus result as well as a state reconstruction result indicating that the nodes’ initial values
can be recovered from signals on the edges up to a shift of the network average.
8
References
[1] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of groups of mobile autonomous agents using
nearest neighbor rules,” IEEE Trans. Autom. Control, 48(6): 988-1001, 2003.
[2] L. Xiao and S. Boyd, “Fast linear iterations for distributed averaging,” Syst. Contr. Lett., 53: 65-78,
2004.
[3] R. Olfati-Saber and R. Murray, “Consensus problems in the networks of agents with switching topology and time delays,” IEEE Trans. Autom. Control, 49(9): 1520-1533, 2004.
[4] W. Ren and R. Beard, “Consensus seeking in multi-agent systems under dynamically changing interaction topologies,” IEEE Trans. Autom. Control, 50(5): 655-661, 2005.
[5] S. Martinez, J. Cortés, and F. Bullo, “Motion coordination with distributed information,” IEEE
Control Syst. Mag., 27(4): 75-88, 2007.
[6] S. Kar, J. M. F. Moura and K. Ramanan, “Distributed parameter estimation in sensor networks: nonlinear observation models and imperfect communication,” IEEE Trans. Information Theory, 58(6):
3575-3605, 2012.
[7] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A Scaglione, “Gossip algorithms for
distributed signal processing,” Proceedings of IEEE, 98(11): 1847-1864, 2010.
[8] M. Rabbat and R. Nowak, “Distributed optimization in sensor networks,” in Proc. IPSN04, 20-27,
2004.
[9] A. Nedić and A. Ozdaglar, “Distributed subgradient methods for multi-agent optimization,” IEEE
Trans. Autom. Control, 54(1): 48-61, 2009.
[10] L. Moreau, “Stability of multiagent systems with time-dependent communication links,” IEEE Trans.
Automat. Control, 50: 169-182, 2005.
[11] M. Cao, A. S. Morse and B. D. O. Anderson, “Reaching a consensus in a dynamically changing
environment: convergence rates, measurement delays, and asynchronous events,” SIAM J. Control
Optim., 47: 601–623, 2008.
[12] Z. Lin, B. Francis, and M. Maggiore, “State agreement for continuous-time coupled nonlinear systems,” SIAM J. Control Optim., 46(1): 288-307, 2007.
[13] A. Nedić, A. Olshevsky, A. Ozdaglar, and J. N. Tsitsiklis, “On distributed averaging algorithms and
quantization effects,” IEEE Trans. Autom. Control, 54: 2506–2517, 2009.
[14] A. Olshevsky and J. N. Tsitsiklis, Convergence speed in distributed consensus and averaging, SIAM
J. Control Optim., 48: 33–55, 2009.
[15] G. Shi and K. H. Johansson, “Robust consusus for continuous-time multi-agent dynamics,” SIAM J.
Control Optim., 51(5): 3673-3691, 2013.
9
[16] C. Godsil and G. Royle. Algebraic Graph Theory. New York: Springer-Verlag, 2001.
[17] B. D. O. Anderson and J. B. Moore, “New results in linear system stability,” SIAM Journal on
Control, 7 (3): 398-414, 1969.
[18] B. D. O. Anderson, “Exponential stability of linear equations arising in adaptive identification,” IEEE
Trans. on Automatic Control, 22 (1): 83-88, 1977.
[19] A. P. Morgan and K. S. Narendra, “On the uniform asymptotic stability of certain linear nonautonomous differential equations,” SIAM J. Control Optim., 15(1): 5-24, 1977.
[20] L. Moreau, “Stability of continuous-time distributed consensus algorithms,” in Proceedings of the
IEEE Conference on Decision and Control, 3998-4003, 2004.
[21] G. Shi, M. Johansson, and K. H. Johansson, “How agreement and disagreement evolve over random
dynamic networks,” IEEE J. on Selected Areas in Communications, 31(6): 1061-1071, 2013.
[22] G. Shi, A. Proutiere, M. Johansson, J. S. Baras and K. H. Johansson, “The evolution of beliefs over
signed social networks,” Operations Research, in press, 2016, [arXiv 1307.0539].
[23] C. Altafini, “Consensus problems on networks with antagonistic interactions,” IEEE Trans. Automatic
Control, 58: 935-946, 2013.
[24] A. Rahmani, M. Ji, M. Mesbahi, and M. Egerstedt, “Controllability of multi-agent systems from a
graph-theoretic perspective,” SIAM J. Control Optim., 48(1): 162-186, 2009.
[25] M. Nabi-Abdolyousefi and M. Mesbahi, “On the controllability properties of circulant networks,”
IEEE Trans. Autom. Control, 58(12): 3179-3184, 2013.
[26] A. Olshevsky, “Minimal Controllability Problems,” IEEE Trans. Control of Network Systems, 1(3):
249–258, 2014.
10
| 3cs.SY
|
Submitted to IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018.
BUILDING COMPETITIVE DIRECT ACOUSTICS-TO-WORD MODELS
FOR ENGLISH CONVERSATIONAL SPEECH RECOGNITION
Kartik Audhkhasi, Brian Kingsbury, Bhuvana Ramabhadran, George Saon, Michael Picheny
IBM T. J. Watson Research Center, Yorktown Heights, New York
arXiv:1712.03133v1 [cs.CL] 8 Dec 2017
ABSTRACT
Direct acoustics-to-word (A2W) models in the end-to-end paradigm
have received increasing attention compared to conventional subword based automatic speech recognition models using phones,
characters, or context-dependent hidden Markov model states. This
is because A2W models recognize words from speech without
any decoder, pronunciation lexicon, or externally-trained language
model, making training and decoding with such models simple.
Prior work has shown that A2W models require orders of magnitude
more training data in order to perform comparably to conventional
models. Our work also showed this accuracy gap when using the
English Switchboard-Fisher data set. This paper describes a recipe
to train an A2W model that closes this gap and is at-par with stateof-the-art sub-word based models. We achieve a word error rate
of 8.8%/13.9% on the Hub5-2000 Switchboard/CallHome test sets
without any decoder or language model. We find that model initialization, training data order, and regularization have the most impact
on the A2W model performance. Next, we present a joint wordcharacter A2W model that learns to first spell the word and then
recognize it. This model provides a rich output to the user instead of
simple word hypotheses, making it especially useful in the case of
words unseen or rarely-seen during training.
Index Terms— End-to-end models, direct acoustics-to-word
models, automatic speech recognition, deep learning
1. INTRODUCTION
Conventional sub-word based automatic speech recognition (ASR)
typically involves three models - acoustic model (AM), pronunciation model and decision tree (PMT), and language model (LM) [1].
The AM computes the probability pAM (A|s) of the acoustics A
given the sub-word units s. The PMT models the probability
pPMT (w|s) of the word sequence w given the sub-word unit sequence s. The LM acts as a prior pLM (w) on the word sequence w.
Hence, finding the most likely sequence of words given the acoustics
A becomes a maximum aposteriori optimization problem over the
following probability density function:
pASR (w|A) ∝ pLM (w)p(A|w)
X
= pLM (w)
pAM (A|s, w)pPMT (s|w)
(1)
(2)
s
≈ pLM (w)
X
pAM (A|s)pPMT (s|w) .
(3)
s
While both the AM and LM in modern ASR systems use deep neural
networks (DNNs) and their variants [2], the PMT is usually based on
The authors thank Ewout Van Den Berg and Andrew Rosenberg of IBM,
and Florian Metze of Carnegie Mellon University for useful discussions.
decision trees and finite state transducers. Training the AM requires
alignments between the acoustics and sub-word units, and several
iterations of model training and re-alignment. Recent end-to-end
(E2E) models have obviated the need for aligning the sub-word units
to the acoustics. Popular E2E models include recurrent neural networks (RNNs) trained with the connectionist temporal classification
(CTC) loss function [3–10] and the attention-based encoder-decoder
RNNs [11–15].These approaches are not truly E2E because they still
use sub-word units, and hence require a decoder and separatelytrained LM to perform well.
In contrast, the recently-proposed direct acoustics-to-word
(A2W) models [16, 17] train a single RNN to directly optimize
pASR (w|A). This eliminates the need for sub-word units, pronunciation model, decision tree, decoder, and externally-trained LM,
which significantly simplifies the training and decoding process.
However, prior research on A2W models has shown that such models require several orders of magnitude more training data when
compared with conventional sub-word based models. This is because the A2W models need sufficient acoustic training examples
per word to train well. For example, [16] used more than 125,000
hours of speech to train an A2W model with a vocabulary of nearly
100,000 words that matched the performance of a state-of-the-art
CD state-based CTC model. Our prior work [17] explored A2W
models for the well-known English Switchboard task and presented
a few initialization techniques to effectively train such models with
only 2000 hours of data. However, we still observed a gap of around
3-4% absolute in WER between the Switchboard phone CTC and
the A2W models on the Hub5-2000 evaluation set.
This paper further improves the state-of-the-art in A2W models for English conversational speech recognition. We present a
training recipe that achieves WER of 8.8%/13.9% on the Switchboard/CallHome subsets of the Hub5-2000 evaluation set, compared
to our previous best result of 13.0%/18.8% [17]. These new results are at par with several state-of-the-art models that use sub-word
units, a decoder, and a LM. We quantify the gains made by each ingredient of our training recipe and conclude that model initialization,
training data order, and regularization are the most important factors.
Next, we turn our attention to the issue of data sparsity while
training A2W models. The conventional solution to this problem
uses a sub-word unit-based model that needs a decoder and LM during testing. As an alternative, we propose the spell and recognize
(SAR) CTC model that learns to first spell the word into its character sequence and then recognize it. Not only does this model retain
all advantages of a direct A2W model, it also provides rich hypotheses to the user which are readable especially in the case of unseen
or rarely-seen words. We illustrate the benefits of this model for
out-of-vocabulary (OOV) words.
The next section discusses the baseline A2W model [17]. Section 3 discusses the proposed training recipe, an analysis of the impact of individual ingredients, and the results. Section 4 presents
our SAR model that jointly models words and characters. The paper
concludes in Section 5 with a summary of findings.
2. BASELINE ACOUSTICS-TO-WORD MODEL
2.1. CTC Loss
Conventional losses used for training neural networks, such as crossentropy, require a one-to-one mapping (or alignment) between the
rows of the T ×D input feature vector matrix A and length-L output
label sequence y. The connectionist temporal classification (CTC)
loss relaxes this requirement by considering all possible alignments.
It introduces a special blank symbol φ that expands the length-L target label sequence y to multiple length-T label sequences ỹ containing φ, such that ỹ maps to y after removal of all repeating symbols
and φ. The CTC loss is then
p(y|A) =
X
p(ỹ|A) =
ỹ
T
X Y
p(ỹt |at ) ,
(4)
ỹ∈Ω(y) t=1
where the set Ω(y), is the set of CTC paths for y, and at /yt denote
the tth elements of the sequences. A forward-backward algorithm
efficiently computes the above loss function and its gradient, which
is then back-propagated through the neural network [3]. The next
section describes the baseline A2W model [17].
2.2. Baseline A2W Model
We used two standard training data sets for our experiments. The
“300-hour” set contained 262 hours of segmented speech from the
Switchboard-1 audio with transcripts provided by Mississippi State
University. The “2000-hour” set contained an additional 1698 hours
from the Fisher data collection and 15 hours from CallHome audio.
We extracted 40-dimensional logMel filterbank features over 25
ms frames every 10 ms from the input speech signal. We used
stacking+decimation [6], where we stacked two successive frames
and dropped every alternate frame during training. This resulted
in a stream of 80-dimensional acoustic feature vectors at half the
frame rate of the original stream. The baseline models also used
100-dimensional i-vectors [18] for each speaker, resulting in 180dimensional acoustic feature vectors.
The baseline A2W model consisted of a 5-layer bidirectional
LSTM (BLSTM) RNN with a 180-dimensional input and 320dimensional hidden layers in the forward and backward directions.
We picked words with at least 5 occurrences in the training data in
the vocabulary. This resulted in a 10,000-word output layer for the
300-hour A2W system and a 25,000-output layer for the 2000-hour
system. As noted in [17], initialization is crucial to training an A2W
model. Thus, we initialized the A2W BLSTM with the BLSTM
from a trained phone CTC model, and the final linear layer using
word embeddings trained using GloVe [19].
Table 1 gives the WERs of the baseline A2W and phone CTC
models reported in [17] on the Hub5-2000 Switchboard (SWB) and
CallHome (CH) test sets. We performed the decoding of the A2W
models via simple peak-picking over the output word posterior distribution, and removing repetitions and blank symbols. The phone
CTC model used a full decoding graph and a LM. We observe that
the 2000-hour A2W model lags behind the phone CTC model by
3.4%/2.8% absolute WER on SWB/CH, and the gap is much bigger
for the 300-hour models. We next discuss our new training recipe.
Table 1. This table shows the Switchboard/CallHome WERs for our
baseline 300/2000-hour A2W and phone CTC models from [17].
Data (hrs)
AM
LM
300
A2W
300
Phone CTC Small
2000
A2W
2000
Phone CTC Big
SWB
20.8
14.5
13.0
9.6
CH
30.4
25.1
18.8
16.0
3. UPDATED TRAINING RECIPE
Our prior experience with training A2W models led us to conclude
that model initialization and regularization are important aspects of
training such models. One key reason for this is the fact that A2W
models attempt to solve the difficult problem of directly recognizing words from acoustics with a single neural network. Hence, our
previously-proposed strategy of initializing the A2W BLSTM with
the phone CTC BLSTM and the final linear layer with word embeddings gave WER gains. In this work, we started by exploring
several other strategies in the same spirit. All our new experiments
were conducted in PyTorch [20] with the following changes compared to [17]:
• We included delta and delta-delta coefficients because they
slightly improved the WER. Hence, the total acoustic vector
was of size 340 after stacking+decimation and appending the
100-dimensional i-vectors.
• We initialized all matrices to samples of a uniform
distribu√
tion over (−, ) where is the inverse of the fan-in or the
size of the input vector [21]. This takes the dimensionality
of the input vectors at each layer in account, which improved
convergence compared to our old strategy of using = 0.01.
• In place of new-bob annealing, we kept √
a fixed learning rate
for the first 10 epochs and decayed it by 0.5 every epoch.
3.1. Training Data Order
Training data order is an important consideration for sequence-tosequence models such as E2E ASR systems because such models
operate on the entire input and output sequences. All training sequences have to be padded to the length of the longest sequence in
the batch in order to do GPU tensor operations. Random sequence
order during batch creation is not memory-efficient because batches
will contain a larger range of sequence lengths, which will lead to
more wasteful padding on average. Hence, sequences have to be
sorted before batch creation.
We compare the impact of sorting input acoustic sequences in order of ascending and descending length in Table 2. Our results show
that ascending order gives significantly better WER than sorting in
descending order. The intuition behind this result is that shorter sequences are easier to train on initially, which enables the network to
reach a better point in the parameter space. This can be regarded as
an instance of curriculum learning [22].
3.2. Momentum and Dropout
We also experimented with Nesterov momentum-based stochastic
gradient descent (SGD), which has been shown to give better convergence compared to simple SGD on several tasks. We use the
Table 2. This table shows the impact of various training strategies
on the 300-hour A2W model.
Model
Descending Order
Ascending Order
+Momentum
+Dropout
+Output Projection
+Phone BLSTM Init.
+Bigger Model
Previous best [17]
SWB
23.0
18.3
18.0
17.4
16.9
14.9
14.6
20.8
CH
30.7
28.1
27.4
26.9
26.3
23.8
23.6
30.4
Model
following parameter updates:
vn = ρvn−1 + λ∇f (Θn−1 + ρvn−1 )
Θn = Θn−1 − vn ,
Table 3. This table shows the WER of our current and previous-best
A2W model trained on the 2000-hour Switchboard+Fisher set. We
have also included several other published results for comparison.
Results with ∗ use data augmentation on the 2000-hour training set.
(5)
(6)
where vn is the velocity or a running weighted-sum of the gradient
of the loss function f . The constant ρ is usually set to 0.9 and λ is the
learning rate, set to 0.01 in our experiments. We also experimented
with a dropout of 0.25 in order to prevent over-fitting. Table 2 shows
that both momentum and dropout improve the WER.
Output Decoder/ SWB CH
Units
LM
BLSTM+LF MMI [24] CD state
Y/Y
8.5 15.3
LACE+LF MMI [25]
CD state
Y/Y
8.3 14.8
Dilated Conv. [26]
CD state
Y/Y
7.7 14.5
BLSTM [18]
CD state
Y/Y
7.7 13.9
Iterated CTC [27]
Char
Y/Y
11.3 18.7
CTC∗ [28]
Char
Y/Y
9.0 17.7
Gram-CTC∗ [28]
Char N-gm Y/Y
7.9 15.8
CTC+Gram-CTC∗ [28] Char N-gm Y/Y
7.3 14.7
RNN-T∗ [29]
Char
Y/N
8.5 16.4
Attention∗ [29]
Char
Y/N
8.6 17.8
CTC A2W [17]
Word
N/N
13.0 18.8
CTC A2W (current)
Word
N/N
8.8 13.9
We also note that our A2W model uses a vocabulary of 25,000
words which has an OOV rate of 0.5%/0.8% on the SWB/CH test
sets. All the other models used much bigger vocabularies and hence
did not suffer from OOV-induced errors.
3.3. Output Projection Layer
In contrast with phone or character-based CTC models, A2W models
have a large output size equal to the size of the vocabulary. Prior
research [23] has shown that decomposing the output linear layer of
size V × D into two layers of sizes V × d and d × D with d < D
speeds-up model training due to reduced number of parameters. We
experimented with a projection layer of size 256 and found that it
speeds-up training by a factor of 1.2x and also slightly improves the
WER, which we attribute to a reduction in over-fitting.
3.4. Phone BLSTM Initialization and Bigger Model
We finally initialized our model with the phone CTC BLSTM which
gave improvements in our previous work [17]. As expected, this
initialization lowered the WER, despite the presence of all the above
useful strategies. With dropout in place, we trained a bigger 6-layer
model with 512-dimensional BLSTM and saw slight gains in WER.
3.6. Ablation Study
In order to understand the impact of individual components of our
recipe, we conducted an ablation study on our best 300-hour A2W
model. We removed each component of the recipe while keeping
others fixed, trained a model, and decoded the test sets. Figure 1
shows the results of this experiment. We observed that changing the
training data order from ascending to descending order of length resulted in the biggest drop in performance. The second biggest factor
was dropout - excluding it leads to over-training because the heldout
loss rises after epoch 10. Choosing a smaller 5-layer model instead
of 6-layer led to the next largest drop in WER. Finally, as expected,
excluding the projection layer had the least impact on WER.
3.5. Final Model and Results
We initialized the 2000-hour A2W model with the best 300-hour
A2W model and used the same recipe for training. Table 3 shows
the WER of the resulting system along with our previous best A2W
WER and several other published results. We obtained a significant
improvement of 4.2%/4.9% absolute WER compared to our previous
result. We also see that our direct A2W is at par with most hybrid
CD state-based and E2E models, while utilizing no decoder or LM.
As noted in [18], the CallHome test set is more challenging than
Switchboard because 36 out of 40 speakers in the latter appear in
the training set. The results on the CallHome test set are especially
good, where our A2W model matches the best result obtained using
a hybrid BLSTM [18] that used exactly the same acoustic features1 .
1 Adding
additional FMLLR features gives a WER of 7.2%/12.7% [18].
Fig. 1. This figure shows the results of the ablation experiment on
our best 300-hour A2W model. The numbers in parentheses in the
legend give the SWB/CH WERs.
Despite strong results, the A2W model does not give any meaningful output to the user in case of OOV words but simply emits an
“UNK” tag. This is not a big problem for the Switchboard task because the OOV rates for the SWB/CH test sets are 0.5%/0.8% on
the 25,000 word vocabulary. But other tasks might be affected by
the limited vocabulary. As a solution, the next section discusses our
joint word-character model that aims to provide the user with a richer
output that is especially useful for unseen or rarely-seen words.
4. SPELL AND RECOGNIZE MODEL
The advantage of the A2W model is that it directly emits word hypotheses by forward-passing acoustic features through a RNN without needing a decoder or externally-trained LM. However, its vocabulary is fixed and OOV words cannot be recognized by the system.
Furthermore, words infrequently seen in the training data are not recognized well by the network due to insufficient training examples.
Prior approaches to dealing with the above limitations completely
rely on sub-word units. This includes work on character models [9,
27] and N-grams [28], RNN-Transducer [29], RNN-Aligner [30].
In contrast, our approach is to have the best of both worlds by
combining the ease of decoding a A2W model with the flexibility of
recognizing unseen/rarely-seen words with a character-based model.
One natural candidate is a multi-task learning (MTL) model containing a shared lower network, and two output networks corresponding
to the two tasks - recognizing words and characters. However, such
an MTL network is not suitable for our purpose because the recognized word and character sequences for an input speech utterance
are not guaranteed to be synchronized in time. This is because the
CTC loss does not impose time-alignment on the output sequence.
The proposed spell and recognize (SAR) model circumvents this
alignment problem by presenting training examples that contain both
words and characters. This allows us to continue to leverage an A2W
framework without resorting to more complex graph-based decoding
methodologies employed in, for example, word-fragment based systems [31–33]. Consider the output word sequence “THE CAT IS
BLACK” The SAR model uses the following target sequence:
b-t h e-e THE b-c a e-t CAT b-i e-s IS b-b l a c e-k BLACK
where lowercase alphabets are the character targets, and b-/edenote special prefixes for word beginning and end. Hence, the
model is trained to first spell the word and then recognize it. In
contrast with a MTL model, the SAR model has a single softmax
over words+characters in the output layer.
a higher OOV rate than the 2000-hour set and also contains several
rare words. We trained a 6-layer BLSTM with joint word and character targets after preparing the output training sequences as described
in the previous section. We initialized the SAR BLSTM using the
A2W BLSTM. The training recipe was the same as for the A2W
model presented previously. The SAR model permits three decodes:
• Word: Use only word predictions, similar to the A2W model.
• Characters: Use only character predictions, and combine
them into words using the word-begin characters.
• Switched: Use the character predictions up to the previous
word when the model predicts an “UNK” symbol, and use
the word prediction everywhere else.
Table 4. This table shows performance of the 300-hour SAR model
for three types of decodes.
Decode
Word
Characters
Switched
SWB
14.5
18.9
14.4
CH
23.9
30.9
24.0
We observe that the SAR model gives comparable performance
to the baseline A2W model, but additionally gives meaningful output
for OOVs, as illustrated by the following test set examples:
REF: SUCH AS LIKE (%HESITATION) THE MURDERING OF A COP OR
HYP: b-s u c e-h SUCH b-a e-s AS b-l i k e-e LIKE b-t h e-e THE b-u e-u
UH b-m u r d e r i n e-g UNK b-o e-f OF b-a A b-s o e-p COP b-o e-r OR
REF: THAT IS RIGHT WE ARE WE ARE FURTHERING HIGHER
HYP: b-t h a e-s THAT’S b-r i g h e-t RIGHT b-r h I RIGHT b-w e ’ r e-e
WE’RE b-w e ’ r e-e WE’RE b-f u r t h e r i n e-g UNK b-h i g h e e-r
HIGHER
REF: BUT SOMETIMES LIKE I JUST HAD TO DO THIS SUMMARY OF
THIS ONE YOU KNOW THESE SCHOLARLY JOURNALS AND STUFF
HYP: b-b u e-t BUT b-s o m e t i m e e-s SOMETIMES b-l i k e-e LIKE b-i I
b-j u s e-t JUST b-h a e-d HAD b-t e-o TO b-d e-o DO b-t h e-e THIS b-s
u mm e r e-y SUMMARY b-t h i e-s THIS b-o n e-e ONE b-y o e-u YOU
b-k n o e-w KNOW b-t h e s e-e THESE b-c o l a r l e-y UNK b-j o u r n a l
e-s UNK b-a n e-d AND b-s t u e-2f STUFF
The words in bold are OOVs. We observe that the SAR model
emits the UNK tag in these cases, but the characters preceding it contain the spelling of the word. In some cases, this spelling is incorrect,
e.g. ”SCHOLARLY → COLARLY”, but still is more meaningful to
the user than the UNK tag. Future research will try to fix these errors
using data-driven methods.
4.1. Choice of Character Set
We experimented with two character sets for the SAR model. The
first one is the simple character set consisting of a total of 41 symbols - alphabets a-z, digits 0-9, whitespace , and other punctuations.
The second character set is the one used in [5], and includes separate
character variants depending on position in a word - beginning, middle, and end. It also includes special symbols for repeated characters,
e.g. a separate symbol for ll. The intuition behind this character set is
that its symbols capture more context as compared to the simple set,
and also disambiguate legitimate character repetitions from double
peaks emitted by the CTC model. We observed that the performance
with the latter character set is slightly better than using simple characters. Hence, we present results only for this case.
4.2. Experiments and Results
We restricted ourselves to the 300-hour set for experiments on the
SAR model because it uses a 10,000 word vocabulary, leading to
5. CONCLUSION
Conventional wisdom and prior research suggests that direct acousticto-word (A2W) models require orders of magnitude more data than
sub-word unit-based models to perform competitively. This paper
presents a recipe to train a A2W model on the 2000-hour Switchboard+Fisher data set that performs at-par with several state-ofthe-art hybrid and end-to-end models using sub-word units. We
conclude that data order, model initialization, and regularization
are crucial to obtaining a competitive A2W model with a WER of
8.8%/13.9% on the Switchboard/CallHome subsets of the Hub52000 test set. Next, we present a spell and recognize (SAR) model
that learns to first spell a word and then recognize it. The proposed SAR model gives a rich and readable output to the user while
maintaining the training/decoding simplicity and performance of a
A2W model. We show some examples illustrating the SAR model’s
benefit for utterances containing OOV words.
6. REFERENCES
[1] F. Jelinek, Statistical methods for speech recognition, MIT
press, 1997.
[2] G. Hinton, L. Deng, D. Yu, G. E. Dahl, A. Mohamed, N. Jaitly,
A. Senior, V. Vanhoucke, P. Nguyen, T. N. Sainath, and
B. Kingsbury, “Deep neural networks for acoustic modeling in
speech recognition: The shared views of four research groups,”
IEEE Signal Processing Magazine, vol. 29, no. 6, pp. 82–97,
2012.
[3] A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber,
“Connectionist temporal classification: labelling unsegmented
sequence data with recurrent neural networks,” in Proc. ICML,
2006, pp. 369–376.
[4] A. Graves and N. Jaitly, “Towards end-to-end speech recognition with recurrent neural networks.,” in ICML, 2014, vol. 14,
pp. 1764–1772.
[5] G. Zweig, C. Yu, J. Droppo, and A. Stolcke, “Advances in allneural speech recognition,” in Proc. ICASSP, 2017, pp. 4805–
4809.
[6] H. Sak, A. Senior, K. Rao, and F. Beaufays, “Fast and accurate
recurrent neural network acoustic models for speech recognition,” in Proc. Interspeech, 2015.
[7] Y. Miao, M. Gowayyed, and F. Metze, “EESEN: End-to-end
speech recognition using deep RNN models and WFST-based
decoding,” in Proc. ASRU, 2015, pp. 167–174.
[8] Y. Miao, M. Gowayyed, X. Na, T. Ko, F. Metze, and A. Waibel,
“An empirical exploration of CTC acoustic models,” in Proc.
ICASSP, 2016, pp. 2623–2627.
[9] A. Hannun, C. Case, J. Casper, B. Catanzaro, G. Diamos,
E. Elsen, R. Prenger, S. Satheesh, S. Sengupta, A. Coates, and
A. Y. Ng, “Deep speech: Scaling up end-to-end speech recognition,” arXiv preprint arXiv:1412.5567, 2014.
[10] D. Amodei et al., “Deep speech 2: End-to-end speech recognition in English and Mandarin,” in International Conference
on Machine Learning, 2016, pp. 173–182.
[11] D. Bahdanau, J. Chorowski, D. Serdyuk, P. Brakel, and Y. Bengio, “End-to-end attention-based large vocabulary speech
recognition,” in Proc. ICASSP, 2016, pp. 4945–4949.
[12] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by jointly learning to align and translate,” in Proc. ICLR,
2015.
[13] A. L. Maas, Z. Xie, D. Jurafsky, and A. Y. Ng, “Lexicon-free
conversational speech recognition with neural networks,” in
Proc. HLT-NAACL, 2015, pp. 345–354.
[14] W. Chan, N. Jaitly, Q. Le, and O. Vinyals, “Listen, attend and
spell: A neural network for large vocabulary conversational
speech recognition,” in Proc. ICASSP, 2016, pp. 4960–4964.
[15] W. Chan, Y. Zhang, Q. Le, and N. Jaitly, “Latent sequence
decompositions,” arXiv preprint arXiv:1610.03035, 2016.
[16] H. Soltau, H. Liao, and H. Sak, “Neural Speech Recognizer:
Acoustic-to-Word LSTM Model for Large Vocabulary Speech
Recognition,” in Proc. Interspeech, 2017.
[17] K. Audhkhasi, B. Ramabhadran, G. Saon, M. Picheny, and
D. Nahamoo, “Direct acoustics-to-word models for English
conversational speech recognition,” in Proc. Interspeech, 2017,
pp. 959–963.
[18] G. Saon, G. Kurata, T. Sercu, K. Audhkhasi, S. Thomas,
D. Dimitriadis, X. Cui, B. Ramabhadran, M. Picheny, L. Lim,
B. Roomi, and P. Hall, “English conversational telephone
speech recognition by humans and machines,” in Proc. Interspeech, 2017.
[19] J. Pennington, R. Socher, and C. D. Manning, “Glove: Global
Vectors for Word Representation,” in Proc. EMNLP, 2014,
vol. 14, pp. 1532–1543.
[20] “PyTorch,”
pytorch.
https://github.com/pytorch/
[21] X. Glorot and Y. Bengio, “Understanding the difficulty of
training deep feedforward neural networks,” in Proc. AISTATS,
2010, pp. 249–256.
[22] Y. Bengio, J. Louradour, R. Collobert, and J. Weston, “Curriculum learning,” in Proc. ICML, 2009.
[23] T. N. Sainath, B. Kingsbury, V. Sindhwani, E. Arisoy, and
B. Ramabhadran, “Low-rank matrix factorization for deep neural network training with high-dimensional output targets,” in
Proc. ICASSP, 2013, pp. 6655–6659.
[24] D. Povey, V. Peddinti, D. Galvez, P. Ghahremani, V. Manohar,
X. Na, Y. Wang, and S. Khudanpur, “Purely Sequence-Trained
Neural Networks for ASR Based on Lattice-Free MMI,” in
Proc. Interspeech, 2016, pp. 2751–2755.
[25] W. Xiong, J. Droppo, X. Huang, F. Seide, M. Seltzer,
A. Stolcke, D. Yu, and G. Zweig, “Achieving human parity in conversational speech recognition,” arXiv preprint
arXiv:1610.05256, 2016.
[26] T. Sercu and V. Goel, “Dense prediction on sequences
with time-dilated convolutions for speech recognition,” arXiv
preprint arXiv:1611.09288, 2016.
[27] W. Xiong, J. Droppo, X. Huang, F. Seide, M. Seltzer, A. Stolcke, D. Yu, and G. Zweig, “The Microsoft 2016 conversational
speech recognition system,” in Proc. ICASSP, 2017, pp. 5255–
5259.
[28] H. Liu, Z. Zhu, X. Li, and S. Satheesh, “Gram-CTC: Automatic Unit Selection and Target Decomposition for Sequence
Labelling,” in Proc. ICML, 2017.
[29] E. Battenberg, J. Chen, R. Child, A. Coates, Y. Gaur, Y. Li,
H. Liu, S. Satheesh, D. Seetapun, A. Sriram, and Z. Zhu, “Exploring neural transducers for end-to-end speech recognition,”
arXiv preprint arXiv:1707.07413, 2017.
[30] H. Sak, M. Shannon, K. Rao, and F. Beaufays, “Recurrent
Neural Aligner: An Encoder-Decoder Neural Network Model
for Sequence to Sequence Mapping,” Proc. Interspeech 2017,
pp. 1298–1302, 2017.
[31] O. Siohan and M. Bacchiani, “Fast vocabulary-independent
audio search using path-based graph indexing,” in Proc. European Conference on Speech Communication and Technology,
2005.
[32] B. Ramabhadran, A. Sethy, J. Mamou, B. Kingsbury, and
U. Chaudhari, “Fast decoding for open vocabulary spoken term
detection,” in Proc. NAACL-HLT, 2009, pp. 277–280.
[33] A. Rastrow, A. Sethy, B. Ramabhadran, and F. Jelinek, “Towards using hybrid word and fragment units for vocabulary independent LVCSR systems,” in Interspeech, 2009, pp. 1931–
1934.
| 2cs.AI
|
The Dependent Doors Problem: An Investigation into Sequential
Decisions without Feedback∗
Amos Korman1 and Yoav Rodeh2
1
CNRS and University Paris Diderot, Paris, France, [email protected]
2
Weizmann Institute of Science, Rehovot, Israel, [email protected]
arXiv:1704.06096v1 [cs.AI] 20 Apr 2017
April 21, 2017
Abstract
We introduce the dependent doors problem as an abstraction for situations in which one must
perform a sequence of possibly dependent decisions, without receiving feedback information on the
effectiveness of previously made actions. Informally, the problem considers a set of d doors that are
initially closed, and the aim is to open all of them as fast as possible. To open a door, the algorithm
knocks on it and it might open or not according to some probability distribution. This distribution
may depend on which other doors are currently open, as well as on which other doors were open
during each of the previous knocks on that door. The algorithm aims to minimize the expected time
until all doors open. Crucially, it must act at any time without knowing whether or which other
doors have already opened. In this work, we focus on scenarios where dependencies between doors
are both positively correlated and acyclic.
The fundamental distribution of a door describes the probability it opens in the best of conditions
(with respect to other doors being open or closed). We show that if in two configurations of d
doors corresponding doors share the same fundamental distribution, then these configurations have
the same optimal running time up to a universal constant, no matter what are the dependencies
between doors and what are the distributions. We also identify algorithms that are optimal up to a
universal constant factor. For the case in which all doors share the same fundamental distribution we
additionally provide a simpler algorithm, and a formula to calculate its running time. We furthermore
analyse the price of lacking feedback for several configurations governed by standard fundamental
distributions. In particular, we show that the price is logarithmic in d for memoryless doors, but can
potentially grow to be linear in d for other distributions.
We then turn our attention to investigate precise bounds. Even for the case of two doors,
identifying the optimal sequence is an intriguing combinatorial question. Here, we study the case of
two cascading memoryless doors. That is, the first door opens on each knock independently with
probability p1 . The second door can only open if the first door is open, in which case it will open on
each knock independently with probability p2 . We solve this problem almost completely by identifying
algorithms that are optimal up to an additive term of 1.
1
Introduction
Often it is the case that one must accomplish multiple tasks whose success probabilities are dependent on
each other. In many cases, failure to achieve one task will tend to have a more negative affect on the
success probabilities of other tasks. In general, such dependencies may be quite complex, and balancing
the work load between different tasks becomes a computational challenge. The situation is further
complicated if the ability to detect whether a task has been accomplished is limited. For example, if task
B highly depends on task A then until A is accomplished, all efforts invested in B may be completely
wasted. How should one divide the effort between these tasks if feedback on the success of A is not
available?
In this preliminary work we propose a setting that captures some of the fundamental challenges that
are inherent to the process of decision making without feedback. We introduce the dependent doors
∗ This work has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020
research and innovation programme (grant agreement No 648032).
1
problem, informally described as follows. There are d ≥ 2 doors (representing tasks) which are initially
closed, and the aim is to open all of them as fast as possible. To open a door, the algorithm can “knock”
on it and it might open or not according to some governing probability distribution, that may depend
on other doors being open or closed1 . We focus on settings in which doors are positively correlated,
which informally means that the probability of opening a door is never decreased if another door is open.
The governing distributions and their dependencies are known to the algorithm in advance. Crucially,
however, during the execution, it gets no direct feedback on whether or not a door has opened unless all
d doors have opened, in which case the task is completed.
This research has actually originated from our research on heuristic search on trees [4]. Consider a
tree of depth d with a treasure placed at one of its leaves. At each step the algorithm can “check” a
vertex, which is child of an already checked vertex. Moreover, for each level of the tree, the algorithm has
a way to compare the previously checked vertices on that level. This comparison has the property that
if the ancestor of the treasure on that level was already checked, then it will necessarily be considered
as the “best” on that level. Note, however, that unless we checked all the vertices on a given level, we
can never be sure that the vertex considered as the best among checked vertices in the level is indeed
the correct one. With such a guarantee, and assuming that the algorithm gets no other feedback from
checked vertices, any reasonable algorithm that is about to check a vertex on a given level, will always
choose to check a child of the current best vertex on the level above it. Therefore, the algorithm can be
described as a sequence of levels to inspect. Moreover, if we know the different distributions involved,
then we are exactly at the situation of the dependent doors problem. See Appendix A for more details on
this example.
Another manifestation of d dependent doors can arise in the context of cryptography. Think about a
sequence of d cascading encryptions, and separate decryption protocols to attack each of the encryptions.
Investing more efforts in decrypting the i’th encryption would increase the chances of breaking it, but only
if previous encryptions where already broken. On the other hand, we get no feedback on an encryption
being broken unless all of them are.
The case of two doors can serve as an abstraction for exploration vs. exploitation problems, where it
is typically the case that deficient performances on the exploration part may result in much waste on the
exploitation part [10, 17]. It can also be seen as the question of balance between searching and verifying
in algorithms that can be partitioned thus [1, 15]. In both examples, there may be partial or even no
feedback in the sense that we don’t know that the first procedure succeeded unless the second one also
succeeds.
For simplicity, we concentrate on scenarios in which the dependencies are acyclic. That is, if we draw
the directed dependency graph between doors, then this graph does not contain any directed cycles.
The examples of searching and verifying and the heuristic search on trees can both be viewed as acyclic.
Moreover, despite the fact that many configurations are not purely acyclic, one can sometimes obtain a
useful approximation that is.
To illustrate the problem, consider the following presumably simple case of two dependent memoryless
doors. The first door opens on each knock independently with probability 1/2. The second door can only
open if the first door is open, in which case it opens on each knock independently, with probability 1/2.
What is the sequence of knocks that minimizes the expected time to open both doors, remembering that
we don’t know when door 1 opens? It is easy to see that the alternating sequence 1, 2, 1, 2, 1, 2, . . . results
in 6 knocks in expectation. Computer simulations indicate that the best sequence gives a little more than
5.8 and starts with 1, 1, 2, 2, 1, 2, 1, 2, 2, 1, 2, 2, 1, 2, 1, 2. Applied to this particular scenario, our theoretical
lower bound gives 5.747, and our upper bound gives a sequence with expected time 5.832.
1.1
Context and Related Work
This paper falls under the framework of decision making under uncertainty, a large research subject that
has received significant amount of attention from researchers in various disciplines, including computer
science, operational research, biology, sociology, economy, and even psychology and cognition, see, e.g.,
[2, 3, 5, 6, 7, 8, 9, 16].
Performing despite limited feedback would fit the framework of reinforced learning [17] and is inherent
to the study of exploration vs. exploitation type of problems, including Multi-Armed Bandit problems [10].
In this paper we study the impact of having no feedback whatsoever. Understanding this extreme scenario
may serve as an approximation for cases where feedback is highly restricted, or limited in its impact. For
1 Actually, the distribution associated with some door i may depend on the state of other doors (being open or closed)
not only at the current knock, but also at the time of each of the previous knocks on door i.
2
example, if it turns out that the price of lacking feedback is small, then it may well be worth to avoid
investing efforts in complex methods for utilizing the partial feedback.
Of particular interest is the case of two doors. As mentioned, difficulties resulting from the lack of
feedback can arise when one aims to find a solution by alternating between two subroutines: Producing
promising candidate solutions and verifying these candidates. Numerous strategies are based on this
interplay, including heuristics based on brute force or trail and error approaches [1, 15], sample and
predict approaches [11, 14, 17], iterative local algorithms [12, 13], and many others. Finding strategies
for efficiently balancing these two tasks can be therefore applicable.
1.2
Setting
There are d ≥ 2 doors and each door can be either open or closed. Doors start closed, and once a door
opens it never closes. To open a door, an algorithm can knock on it and it might open or not according
to some probability distribution. The goal is to minimize the expected number of knocks until all doors
open. Crucially, the algorithm has no feedback on whether or not a door has opened, unless all doors
have opened, in which case the task is completed.
The probability that a door opens may depend on the state of other doors (being open or closed) at
the time of the current knock as well as on their state during each of the previous knocks on the door. For
example, the probability that a certain knock at door i succeeds may depend on the number of previous
knocks on door i, but counting only those that were made while some other specific door j was open.
The idea behind this definition is that the more time we invest in opening a door the more likely it is to
open, and the quality of each knock depends on what is the state of the doors it depends on at the time
of the knock.
Below we provide a semi-formal description of the setting. The level of detail is sufficient to understand
the content of the main text, which is mainly concerned with independent and cascading configurations.
The reader interested in a more formal description of the model is deferred to Appendix B.
A specific setting of doors is called a configuration (normally denoted C). This includes a description
of all dependencies between doors and the resulting probability distributions. In this paper we assume
that the dependency graph of the doors is acyclic, and so we may assume that a configuration describes
an ordering of the doors, such that each door depends only on lower index doors. Furthermore, we assume
that the correlation between doors is positive, i.e., a door being open can only improve the chances of
other doors to open.
Perhaps the simplest configuration is when all doors are independent of each other. In this case, door
i can be associated with a function pi : N → [0, 1], where pi (n) is the probability that door i is not open
after knocking on it n times. Another family of acyclic configurations are cascading configurations. Here,
door i cannot open unless all doors of lower index are already open. In this case, the configuration can
again be described by a set of functions {pi }di=1 , where pi (n) describes the probability that door i is not
open after knocking on it n times, where the count starts only after door i − 1 is already open.
In general, given a configuration, each door i defines a non-decreasing function pi : N → [0, 1], called
the fundamental distribution of the door, where pi (n) is the probability that the door is not open after
knocking on it n times in the best of conditions, i.e., assuming all doors of lower index are open. In the
case of independent and cascading configurations, the fundamental distribution pi coincides with the
functions mentioned above. Two doors are similar if they have the same fundamental distribution. Two
configurations are similar if for every i, door i of the first configuration is similar to door i of the second.
When designing an algorithm, we will assume that the configuration it is going to run in is known.
As there is no feedback, a deterministic algorithm can be thought of as a possibly infinite sequence of
door knocks. A randomized algorithm is therefore a distribution over sequences, and as all of them will
have expected running time at least as large as that of an optimal sequence (if one exists), the expected
running time of a randomized algorithm cannot be any better. Denote by TC (π), the expected time until
all doors open when running sequence π in configuration C. We define TC = minπ TC (π). By Claim 23
in Appendix B.3, there exists a sequence achieving this minimum. Therefore, by the aforementioned
arguments, we can restrict our discussion to deterministic algorithms only.
If we had feedback
P∞ we would knock on each door until it opens, and then continue to the next.
Denoting by Ei = n=0 pi (n) the expected time to open door i on its own, the
Pexpected running time
then does not depend on the specific dependencies between doors at all, and is i Ei . Also, this value is
clearly optimal. To evaluate the impact of lacking feedback for a configuration C, we therefore define:
TC
Price(C) = P
i Ei
3
Obviously Price(C) ≥ 1, and for example, if all doors start closed and open after just 1 knock, it is in
fact equal to 1. Claim 22 in Appendix B.2 shows that Price(C) ≤ d.
1.3
Our Results
We have two main results. The first one, presented in Section 2, states that any two similar configurations
have the same optimal running time up to a constant factor. We stress that this constant factor is
universal in the sense that it does not depend on the specific distributions or on the number of doors d.
Furthermore, given a configuration, we identify an algorithm that is optimal for it up to a constant
factor. We then show that for configurations where all doors are similar, there is a much simpler algorithm
which is optimal up to a constant factor, and describe a formula that computes its approximate running
time. We conclude Section 2 by analysing the price of lacking feedback for several configurations governed
by standard fundamental distributions. In particular, we show that the price is logarithmic in d for
memoryless doors, but can potentially grow to be linear in d for other distributions.
We then turn our attention to identify exact optimal sequences. Perhaps the simplest case is the
case of two cascading memoryless doors. That is, the first door opens on each knock independently with
probability p1 . The second door can only open if the first door is open, in which case it opens on each
knock independently, with probability p2 . In Section 3 we present our second main result: Algorithms for
these configurations that achieve the precise optimal running time up to an additive term of 1.
On the technical side, to establish such an extremely competitive algorithm, we first consider a
semi-fractional variant of the problem and find a sequence that achieves the precise optimal bound. We
then approximate this semi-fractional sequence to obtain an integer solution losing only an additive term
of 1 in the running time. A nice anecdote is that in the case where p1 = p2 and are very small, the ratio
of 2-knocks over 1-knocks in the sequence we get approaches the golden ratio. Also, in this case, the
optimal running time approaches 3.58/p1 as p1 goes to zero. It follows that in this case, the price of
lacking feedback tends to 3.58/2 and the price of dependencies, i.e., the multiplicative gap between the
cascading and independent settings, tends to 3.58/3.
2
Near Optimal Algorithms
The following important lemma is proved in Appendix B.1 using a coupling argument:
Lemma 1. Consider similar configurations C, X and I, where X is cascading and I is independent. For
every sequence π, TI (π) ≤ TC (π) ≤ TX (π). This also implies that TI ≤ TC ≤ TX .
The next theorem presents a near optimal sequence of knocks for a given configuration. In fact, by
Lemma 1, this sequence is near optimal for any similar configuration, and so we get that the optimal
running time for any two similar configurations is the same up to a universal multiplicative factor.
Theorem 2. There is a polynomial algorithm2 , that given a configuration C generates a sequence π such
that TC (π) = Θ(TI ). In fact, TC (π) ≤ 2 + 4TI ≤ 2 + 4TC .
Proof. Denote by p1 , . . . , pd the fundamental distributions of the doors of C. For a finite sequence of
knocks α, denote by SCC (α) the probability that after running α in configuration C, some of the doors
are still closed. Note that if α is sorted, that is, if all knocks on door 1 are done first, followed by the
knocks on doors 2, etc., then SCX (α) = SCI (α).
We start by showing that for any T , we can construct in polynomial time a finite sequence αT of
length T that maximizes the probability that all doors will open, i.e., minimizes SCI (αT ). As noted
above, if we sort the sequence, this is equal to SCX (αT ).
The algorithm follows a dynamic programming approach, and calculates a matrix A, where A[i, t]
holds the maximal probability that a sequence of length t has of opening all of the doors 1, 2, . . . , i. All
the entries A[0, ·] are just 1, and the key point is that for each i and t, knowing all of the entries in A[i, ·],
it is easy to calculate A[i + 1, t]:
t
A[i + 1, t] = max A[i, t − k] · (1 − pi+1 (k))
k=0
2A
polynomial algorithm in our setting generates the next knock in the sequence in polynomial time in the index of the
knock and in d, assuming that reading any specific value of any of the fundamental distributions of a door takes constant
time.
4
Calculating the whole table takes O(dT 2 ) time, and A[d, T ] will give us the highest probability a sequence
of length T can have of opening all doors. Keeping tabs on the choices the max in the formula makes, we
can get an optimal sequence αT , and can take it to be sorted.
Consider the sequence π = α2 · α4 · · · α2n · · · . The complexity of generating this sequence up to place
T is O(dT 2 ), and so this algorithm is polynomial. Our goal will be to compare TX (π) with TI (π ? ), where
π ? is the optimal sequence for I.
The P
following observation stems from the fact that for any natural valued random variable X,
∞
E [X] = n=0 Pr [X > n] and Pr [X > n] is a non-increasing function of n.
∞
Observation 3. Let {an }n=1 be a strictly increasing sequence of natural numbers, and X be some
natural valued random variable. Then:
∞
∞
X
X
(an+1 − an )Pr [X > an+1 ] ≤ E [X] ≤ a1 +
(an+1 − an )Pr [X > an ]
n=1
n=1
of π of length n. In this terminology, TC (π) =
P∞For a sequence π, denote by π[n] the prefix
n
(π[n]).
Setting
a
=
2
+
4
+
.
.
.
+
2
in
the right side of Observation 3, and letting X
SC
n
C
n=0
be the number of rounds until all doors open when using π, we get:
TX (π) ≤ 2 +
=2+
∞
X
n=1
∞
X
2n+1 · SCX (π[2 + . . . + 2n ]) ≤ 2 +
2n+1 · SCX (α2n )
n=1
2n+1 · SCI (α2n ) ≤ 2 +
n=1
∞
X
2n+1 · SCI (π ? [2n ]) ≤ 2 + 4TI (π ? )
n=1
n−1
The last step is using Observation 3 with an = 2
2.1
∞
X
. Theorem 2 concludes.
Configurations where all Doors are Similar
In this section we focus on configurations where all doors have the same fundamental distribution p(n).
We provide simple algorithms that are optimal up to a universal constant, and establish the price of
lacking feedback with respect to a few natural distributions. Corresponding proofs appear in Appendix C.
2.1.1
Simple Algorithms
Let us consider the following very simple algorithm Asimp . It runs in phases, where in each phase it
knocks on each door once, in order. As a sequence, we can write Asimp = (1, 2, . . . , d)∞ . Let X1 , . . . , Xd
be i.i.d. random variables taking positive integer values, satisfying Pr [Xi > n] = p(n). The following is
straightforward:
Claim 4. TI (Asimp ) = Θ (d · E [max {X1 , . . . , Xd }])
This one is less trivial:
Claim 5. If all doors are similar then TI (Asimp ) = Θ(TI )
The claim above states that Asimp is optimal up to a multiplicative constant factor in the independent
case, where all doors are similar. As a result, we can also show:
n
n
Claim 6. Denote by αn the sequence 12 , . . . , d2 . If all doors are similar then for any configuration C,
TC (α0 · α1 · α2 · · ·) = Θ(TC ).
In plain words, the above claim states that the following algorithm is optimal up to a universal
constant factor for any configuration where all doors are similar: Run in phases where phase n consists of
knocking 2n consecutive times on each door, in order.
2.1.2
On the Price of Lacking Feedback
By Claims 4 and 5, investigating the price of lacking feedback when all doors are similar boils down to
understanding the expected maximum of i.i.d. random variables.
E [max {X1 , . . . , Xd }]
Price = Θ
(1)
E [X1 ]
Note that we omitted dependency on the configuration, as by Theorem 2, up to constant factors, it is the
same price as in the case where the doors are independent. Let us see a few examples of this value. First:
5
Lemma 7. If X1 , . . . , Xd are i.i.d. random variables taking natural number values, then:
!
∞
X
E [max(X1 , . . . , Xd )] = Θ κ + d
Pr [Xi > n]
n=κ
Where κ = min {n ∈ N | Pr [X1 > n] < 1/d}
Example 8. After the first knock on it, each door opens with probability 1 − 1/d and if it doesn’t, it
will open at its d + 1’st knock. The expected time to open each door on its own is 2. By Lemma 7, as
κ = d + 1, we get that Price = Ω(κ) = Ω(d). By Claim 22, Price = Θ(d).
Example 9. If p(n) = q n for some 1/2 < q < 1, then Price = Θ(log(d)).
1
Example 10. If for some c > 0 and a > 1, p(n) = min(1, c/na ), then Price = Θ(d a ).
Sometimes we know a bound on some moment of the distribution of opening a door. If E [X1 ] < M ,
then by Claim 22, T = O(d2 M ). Also,
1
1
Example 11. If E [X1a ] < M for some a > 1, then T = O d1+ a M 1/a (1 + a−1
) .
For example, if the second moment of the time to open a door on its own is bounded, we get an
O(d3/2 ) algorithm.
3
Two Memoryless Cascading Doors
One can say that by Theorem 2 we solved much of the dependent doors problem. There is an equivalence
of the independent and cascading models, and we give an up to constant factor optimal algorithm for any
situation. However, we still find the question of finding the true optimal sequences for cascading doors to
be an interesting one. What is the precise cost of having no feedback, in numbers? Even the simple case
of two doors, each opening with probability 1/2 on each knock, turns out to be quite challenging and has
a not so intuitive optimal sequence.
In this section, we focus on a very simple yet interesting case of the cascading door problem, and
solve it almost exactly. We have two doors. Door 1 opens with probability p1 each time we knock on
it, and door 2 opens with probability p2 . We further extend the setting to consider different durations.
Specifically, we assume that a knock on door 1 takes one time unit, and a knock on door 2 takes c time
units. Denote q1 = 1 − p1 and q2 = 1 − p2 . For brevity, we will call a knock on door 1 a 1-knock, and a
knock on door 2 a 2-knock.
The Semi-Fractional Model. As finding the optimal sequence directly proved to be difficult, we
introduce a relaxation of our original model, termed the semi-fractional model. In this model, we allow
1-knocks to be of any length. A knock of length t, where t is a non-negative real number, will have
probability of 1 − q1t of opening the door. In this case, a sequence consists of the alternating elements 1t
and 2, where 1t describes a knock of length t on door 1. We call sequences in the semi-fractional model
semi-fractional sequences, and to differentiate, we call sequences in the original model integer sequences.
As our configuration C will be clear from context, for a sequence π, we define E [π] = TC (π) to be
the expected running time of the sequence. Clearly, every integer sequence has a similar semi-fractional
sequence with the same expected running time. As we will see, the reverse is not far from being true.
That being so, finding the optimal semi-fractional sequence will give an almost optimal integer sequence.
3.1
Equivalence of Models
Theorem 12. Every semi-fractional sequence π has an integer sequence π 0 , s.t., E [π 0 ] ≤ E [π] + 1.
For this purpose, in this subsection only, we describe a semi-fractional sequence π as a sequence of
non-decreasing non-negative real numbers: π0 , π1 , π2 , . . ., where π0 = 0. This sequence describes the
following semi-fractional sequence (in our original terms):
1π1 −π0 · 2 · 1π2 −π1 · 2 · · ·
This representation simplifies our proofs considerably. Here are some observations:
6
• 1-knocks can be of length 0, yet we still consider them in our indexing.
• The sequence is an integer sequence iff for all i, πi ∈ N.
• The i-th 2-knock starts at time πi + c(i − 1) and ends at πi + ci.
• The probability of door 1 being closed after the completion of the i-th 1-knock is q1πi , and so the
π
probability it opens at 1-knock i is q1 i−1 − q1πi
Lemma 13. For two sequences π = (π0 , π1 , . . .) and π 0 = (π10 , π20 , . . .), if for all i, πi ≤ πi0 ≤ πi + 1 then
E [π 0 ] ≤ E [π] + 1.
Lemma 13 is the heart of our theorem. Indeed, once proven, Theorem 12 follows in a straightforward
manner. Given a semi-fractional sequence π, define πi0 = dπi e. Then, π 0 is an integer sequence, and it
satisfies the conditions of the lemma, so we are done. The lemma makes sense, as the sequence π 0 in
which for all i > 0, πi0 = πi + 1, can be thought of as adding a 1-knock of length one in the beginning
of the sequence. Even if this added 1-knock did nothing, the running time would increase by at most
1. However, the proof is more involved, since in the lemma, while some of the 2-knocks may have an
increased chance of succeeding, some may actually have a lesser chance.
Proof. Given a sequence π and an event X, we denote by E [π | X] the expected running time of π given
the event X. Let Xi denote the event that door 1 opens at its i-th 1-knock. As already said:
Z πi
πi−1
πi
Pr [Xi ] = q1
− q1 =
q1x ln(q1 ) dx
πi−1
Where the last equality comes as no surprise, as it can be seen as modelling door 1 in a continuous fashion,
having an exponential distribution fitting its geometrical one. Now:
Z ∞
∞ Z πi
∞
X
X
E [π] =
Pr [Xi ] E [π | Xi ] =
q1x ln(q1 ) dx · E [π | Xi ] =
q1x ln(q1 ) · E π Xi(x) dx
i=1
i=1
πi−1
0
Where i(x) = maxi {x ≥ πi−1 }, that is, the index of the 1-knock that x belongs to when considering only
time spent knocking on door 1. Defining Xi0 and i0 (x) in an analogous way for π 0 , we want to show that
for all x,
h
i
E π 0 Xi00 (x) ≤ 1 + E π Xi(x)
as using it with the last equality will prove the lemma. We need the following three claims:
1. If j ≤ i, then E [π | Xj ] ≤ E [π | Xi ]
2. For all x, i0 (x) ≤ i(x)
3. For all i, E [π 0 | Xi0 ] ≤ 1 + E [π | Xi ]
Together they give what we need:
h
i
E π 0 Xi00 (x) ≤ 1 + E π Xi0 (x) ≤ 1 + E π Xi(x)
The first is actually true trivially for all sequences, as the sooner the first door opens, the better the
expected time to finish. For the second, since for all i, πi0 ≥ πi , then x ≥ πi0 implies that x ≥ πi , and so:
0
i0 (x) = max x ≥ πi−1
≤ max {x ≥ πi−1 } = i(x)
i
i
For the third, denote by Yj the event that door 2 opens at the j’th 2-knock. Then:
E [π | Xi ] =
∞
X
(πj + cj)Pr [Yj | Xi ]
j=i
Let us consider this same expression as it occurs in π 0 . First note that Pr [Yj | Xi ] = Pr Yj0 Xi0 , as all
that matters for its evaluation is j − i. Therefore:
E [π
0
| Xi0 ]
∞
∞
X
0
X
0
0
=
(πj + cj)Pr Yj Xi ≤
(πj + 1 + cj)Pr [Yj | Xi ]
j=i
= E [π | Xi ] +
j=i
∞
X
Pr [Yj | Xi ] ≤ E [π | Xi ] + 1
j=i
7
3.2
The Optimal Semi-Fractional Sequence
A big advantage of the semi-fractional model is that we can find an optimal sequence for it. For that we
need some preparation:
Definition 14. For a semi-fractional sequence π, and some 0 ≤ x ≤ 1, denote by Ex [π] the expected
running time of π when started with door 1 being closed with probability x. In this notation, E [π] = E1 [π].
Lemma 15. Let y = x/(q2 + p2 x). Then:
Ex 1t · π = t + Eq1t x [π]
Ex [2 · π] = c +
x
Ey [π]
y
Proof. The first equation is clear, since starting with door 1 being closed with probability x, and then
knocking on it for t rounds, the probability that this door is closed is q1t x.
As for the second equation, if door 1 is closed with probability x, then knocking on door 2, we have a
probability of p2 (1 − x) of terminating, and so the probability we did not finish is:
1 − p2 (1 − x) = 1 − p2 + p2 x = q2 + p2 x =
x
y
It remains to show that conditioning on the fact that we indeed continue, the probability that door 1 is
closed is y. It is the following expression, evaluated after a 2-knock:
Pr [door 1 is closed]
x
=
=y
Pr [door 1 is closed] + Pr [door 1 is open but not door 2]
x + (1 − x)q2
Applying Lemma 15 iteratively on a finite sequence w, we get:
Ex [wπ] = a(x, w) + b(x, w)Eδ(x,w) [π]
(2)
Of specific interest is δ(x, w). It can be thought of as the state3 of our algorithm after running the
sequence w, when we started at state x. Lemma 15 and Equation (2) give us the behaviour of δ(x, w):
δ(x, 1t ) = q1t x
δ(x, 2) =
x
q2 + p2 x
δ(x, aw) = δ(δ(x, a), w)
We start with the state being 1, since we want to calculate E1 [π]. Except for this first moment, as we can
safely assume any reasonable algorithm will start with a 1-knock, the state will always be in the interval
(0, 1). A 1-knock will always decrease the state and a 2-knock will increase it.
Our point in all this, is that we wish to exploit the fact that our doors are memoryless, and if we
encounter a state we’ve already been at during the running of the sequence, then we should probably
make the same choice now as we did then. The following definition and lemma capture this point.
Definition 16. We say a non-empty finite sequence w is x-invariant, if δ(x, w) = x.
The following Lemma is proved in Appendix D.2, and formalizes our intuition about how an optimal
algorithm should behave.
Lemma 17. If w is x-invariant, and Ex [wπ] ≤ Ex [π] then Ex [w∞ ] ≤ Ex [wπ].
3.2.1
The Actual Semi-Fractional Sequence
Theorem 18. There is an optimal semi-fractional sequence π ? of the form 1s (21t )∞ , for some positive
real values s and t, and its running time is:
c + (1 − p2 z) logq1 (1 − p2 z)
E [π ? ] = min logq1 (1 − z) +
p2 z
z∈[0,1]
Proof. Claim 26 of Appendix D.1 says that there is an optimal semi-fractional sequence π. It clearly
starts with a non-zero 1-knock, and so we can write π = 1s 2π 0 . Intuitively, in terms of its state, this
sequence starts at 1, goes down for some time with a 1-knock, and then jumps back up with a 2-knock.
8
Figure 1: How the state evolves as a function of time. 1-knocks decrease the state, and 2-knocks increase it. Note
that r = logq1 (y) and s = logq1 (x).
The state it reaches now was already passed through on the first 1-knock, and so as this is an optimal
sequence we can assume it will choose the same as it did before, and keep zig-zaging up and down.
We next prove that indeed there is an optimal sequence following the zig-zaging form above. Again,
take some optimal π, and write π = 1s 2π 0 . Denote x = δ(1, 1s ) and y = δ(1, 1s 2) = δ(x, 2) > x (see
Figure 1). Taking r = logq1 (y) < s, we get δ(1, 1r ) = y. Denoting t = s − r, this means that 1t 2 is
y-invariant. Since π is optimal, then:
E [π] = E 1r (1t 2)π 0 ≤ E [1r π 0 ] which implies: Ey 1t 2π 0 ≤ Ey [π 0 ]
So by Lemma 17:
Ey (1t 2)∞ ≤ Ey 1t 2π 0
E 1r (1t 2)∞ ≤ E 1r 1t 2π 0 = E [π]
which implies:
Therefore, 1r (1t 2)∞ = 1s (21t )∞ is optimal. We denote this sequence π ? .
Now for the analysis of the running time of this optimal sequence. We will use Lemma 15 many times
in what follows.
E1 1s (21t )∞ = s + Ex (21t )∞
Denote α = (21t )∞ .
x
x
Ex [α] = Ex 21t α = c + Ey 1t α = c + (t + Ex [α])
y
y
Since t = s − r = logq1 (x/y):
Ex [α] =
c
1−
x
y
+
x
y
1−
x
y
logq1 (x/y)
By Lemma 15, as our y is the state resulting from a 2-knock starting at state x, it follows that
y = x/(q2 + p2 x). Since x/y = q2 + p2 x, then 1 − x/y = p2 (1 − x) and then we get:
c
q2 + p2 x
+
logq1 (q2 + p2 x)
p2 (1 − x) p2 (1 − x)
And in total:
c + (q2 + p2 x) logq1 (q2 + p2 x)
E1 1s (21t )∞ = logq1 (x) +
p2 (1 − x)
Changing variable to z = 1 − x, results in q2 + p2 x = 1 − p2 z, and we get the expression in the statement
of the theorem.
3 There is an intuitive meaning behind this. Going through Lemma 15, we can see that δ(1, w) is actually the probability
that after running w, door 1 is closed conditioned on door 2 being closed. Indeed, After running some finite sequence, the
only feedback we have is that the algorithm did not finish yet. We can therefore calculate from our previous moves what is
the probability that door 1 is closed, and that is the only information we need for our next steps.
9
3.3
Actual Numbers
Theorem 18 gives the optimal semi-fractional sequence and a formula to calculate its expected running
time. This formula can be approximated as accurately as we wish for any specific values of p1 , p2 and c, but
it is difficult to obtain a closed form formula from it. Lemma 27 in Appendix D.3 gives us a pretty good
result when p1 ≈ p2 , especially when they are small, as by Observation 25, we get log(1/(1 − p1 )) ≈ p1 ,
and so the additive mistake in the formula is something like 1.
In general, when p1 is small, then θ (see Lemma 27) is approximately cp1 /p2 , which is the expected
time to open door 2 on its own, divided by the time to open door 1 on its own - a natural measure of
the system. Then, ignoring the additive mistake, we get that the lower bound is approximately F(θ)/p1 ,
where F is some function not depending on the parameters of the system. For example F(1) = 3.58.
So opening two similar doors without feedback when p is small takes about 3.58 times more time than
opening one door as opposed to the case with feedback, where the factor is only 2.
We also note, that when the two doors are independent and similar, it is quite easy to see that the
optimal expected running time is at most 3/p (see Claim 28 in Appendix D.4). As a last interesting
point, in Appendix D.5 we show that if c = 1 and p = p1 = p2 approaches
√ zero, then the ratio between
the number of 2-knocks and the number of 1-knocks approaches 12 (1 + 5), which is the golden ratio.
3.4
Examples
For p1 = p2 = 1/2 and c = 1, the lower bound is 5.747. Simulations show that the best algorithm for this
case is slightly more than 5.8, so the lower bound is quite tight, but our upper bound is 6.747 which is
pretty far. However, the sequence we get from the upper bound proof starts with:
1, 1, 2, 1, 2, 2, 1, 2, 1, 2, 2, 1, 2, 2, 1, 2, 1, 2, 2, 1, 2, 1, 2, 2, 1, 2, 2, 1, 2, 1, 2, . . .
The value it gives is about 5.832, which is very close to optimal.
For p1 = p2 = 1/100 and c = 1, the sequence we get is:
197 , 2, 2, 1, 2, 2, 1, 2, 1, 2, 2, 1, 2, 1, 2, 2, 1, 2, 2, 1, 2, 1, 2, 2, 1, 2, 2, 1, 2, 1, 2, . . .
And the value it gives is about 356.756, while the lower bound can be calculated to be approximately
356.754. As we see this is much tighter than the +1 that our upper bound promises.
10
A
“Real Life” Example
The following scenario, much simplified, is the focus of [4]. A treasure τ is placed at a leaf of a ∆-regular
rooted tree of depth d. A mobile agent starting at the root wishes to find it as fast as possible and is
allowed to move along edges. At each node there is an advice pointer to one of its neighbours. With
probability p the advice is correct, i.e., directed towards the treasure, and with probability q = 1 − p it
is incorrect, i.e., points towards the one of the neighbours uniformly at random. The agent can move
between two neighbouring nodes in one unit of time and look at the advice at the current node hosting it.
Minimizing the expected running time until finding the treasure turns out to be not trivial, and the crux
of the problem is that
√ the advice is permanent, and so cannot be amplified by rechecking it. It is shown
in [4] that if q & 1/ ∆, then no algorithm has running time which is polynomial in d. Modeling this
problem as cascading doors gives a non-trivial solution for the cases where q is smaller.
Consider door i as open once the algorithm visits τi , the ancestor of the treasure that is at distance i
from the root. The purpose is then to open all doors. At any point in the algorithm, the candidates at
level i are those unvisited vertices at that level, whose parent is already visited. Also denote the score of
a vertex as the number of advice pointers that point towards it in the advice seen so far by the agent. A
knock on door i consists of visiting the highest scoring candidate on level i, where symmetry is broken
arbitrarily.
The difference in score between two candidates at the same level is affected by the advice on the
path between them only, and as the algorithm moves on edges, all of the advice on this path is known.
Consider a candidate at level i that is at distance l from τi . It will have a score that is at least as high
as the treasure if the number of advice pointers on the path connecting them that point towards it is
greater than the number of those pointing towards τi . The probability that this happens can be viewed
as the probability that a random walk of length l sums up to at least 0, where each step is (−1, 0, 1) with
respective probabilities (p
√ + q/∆, (1 − 2/∆)q, q/∆). Denote this probability by α(l). It is shown in [4]
that α(l) ≈ q l for q < 1/ ∆.
Denote by Ci the number of such candidates at level i that “beat” τi . Even assuming all of the vertices
at that level are now reachable:
E [Ci ] ≤
i
X
∆j α(2j − 1) ≈
j=1
i
√
1X
(∆q 2 )j = O( ∆)
q j=1
This is in fact an upper bound on the expected number of knocks until opening
door i, assuming door
√
i − 1 is already open. By Example 11, Asimp will need an expected O(d2 ∆) knocks to open all doors
and find the treasure. As moving
√ from one candidate to another takes
O(d) moves, the running time of
this algorithm is at most O(d3 ∆). If we were able to prove that E Ci2 is small, then by Example 11 we
could have dropped d’s exponent to 2.5, but it turns out that this second moment is actually exponential
in d and so this approach fails.
Of course, assuming there is no feedback at all in this situation is an over approximation, and while it
gives
√ a non-trivial result, using much more sophisticated arguments, it is shown in [4] that there is an
O( ∆d) algorithm, and that it is in fact optimal.
B
General Dependencies
In the main text of the paper we focus on two special cases, that of independent doors and that of
cascading doors. In what follows we introduce the possibility of much more general dependencies, and
show that in fact, the two cases above are the extreme ones and so proving their equivalence is enough to
prove it for all cases. For that we need to revisit our basic definition of doors and knocks.
Acyclic dependencies. We assume that the directed graph of dependencies between doors is
acyclic. In such cases, the doors can be ordered in a topological order such that a door may depend only
on lower index doors. In what follows, w.l.o.g., we shall always assume that doors are ordered in such an
order.
Configuration. A configuration C for d doors indexed 1, . . . , d describes the probabilities of each
door opening as a result of knocks on it. It relates a door i with the function:
φCi : {(X1 , . . . , Xn ) | n ≥ 1, ∀j.Xj ⊆ {1, . . . , i − 1}} → [0, 1]
11
Which given, for a sequence of n knocks on door i, the set of doors Xj that are open at the time of each
of those knocks, returns the probability that door i was opened by one of these knocks. We will omit the
superscript C when it is clear from context.
Monotonicity. The more we knock on a door the better our chances of opening it. More precisely,
the monotonicity property requires that if X = (X1 , . . . , Xn ), and X 0 is a sub-sequence of X (possibly a
non-consecutive one), then φi (X 0 ) ≤ φi (X).
Positive Correlation. We focus on the case where the doors are positively correlated, namely, a
door being open can never decrease the chances of other doors to open. Formally this means that for
every i, if for all j, Xj0 ⊆ Xj , then φi (X10 , . . . , Xn0 ) ≤ φi (X1 , . . . , Xn ).
Fundamental Distribution. The fundamental distribution4 of door i in configuration C is the
function pCi (again, we will omit the superscript) where pi (n) denotes the probability, in the best of
conditions, i.e., when all doors it depends on are open, that door i remains closed after being knocked on
n
n times. Formally, pi (n) = 1 − φi ({1, . . . , i − 1} ). So pi (0)
1 for every door i, and by the monotonicity
P=
∞
property pi is non-increasing. We also denote by Ei = n=0 pi (n) the expected time to open door i
assuming all the doors it depends on are already open. We will always assume that for all i, Ei < ∞.
Similarity. Two doors are similar if they have the same fundamental distribution. Two configurations are similar if for every i, door i of the first configuration is similar to door i of the second.
B.1
The Cascading and Independent Configurations
In light of the definitions above we define the two main configurations:
1. Independent doors. The distribution associated with a door is independent of whether or not other
k
doors are open. Formally, φi (X1 , . . . , Xk ) = φi ({1, . . . , i − 1} ).
2. Cascading doors. Door i > 1 cannot open unless door i − 1 is already open. Only after door i − 1
t
opens we start counting knocks on door i. Formally, φi (X1 , . . . , Xk ) = φi ({1, . . . , i − 1} ) where t is
the number of Xj ’s that are equal to {1, . . . , i − 1}.
Definition 19. For configurations A and B, we say that A dominates B, if for every i and every
B
X = (X1 , . . . , Xn ), we have: φA
i (X) ≥ φi (X).
First:
Claim 20. For configuration C, similar independent configuration I, and similar cascading configuration
X , I dominates C and C dominates X .
Proof. Denote n = |X|, and denote by k the number of elements of X that are equal to {1, . . . , i − 1}.
We get the following series of inequalities:
k
C
φX
i (X) = 1 − pi (k) = φi ({1, . . . , i − 1} )
n
≤ φCi (X) ≤ φCi ({1, . . . , i − 1} ) = 1 − pi (n) = φIi (X)
Where we used, in order: the definition of cascading configuration, the fact that pi is the fundamental
distribution of door i, monotonicity, positive correlation, the fact that p is the fundamental distribution
of door i, and the definition of independent configuration.
An important property of dominance is:
Claim 21. For any sequence π, if A dominates B then TA (π) ≤ TB (π).
Proof. A possible way to describe the random process governing the running of an algorithm in a particular
configuration, is as follows:
1. For each door i, choose uniformly at random a real number ai ∈ [0, 1]. Fix this number for the rest
of the run.
4 This
is actually not a distribution function, but rather the complement of an accumulative distribution function.
12
2. Denote by X the history of open doors as usual. Start it as the empty sequence.
3. Go over the knocks in the sequence in order, and when the knock is on door i, check if φi (X 0 ) > ai ,
where X 0 is part of X that is relevant to calculate φi (only the indices where there is a knock of
door i, and only the information about the doors of {1, . . . , i − 1}). If it is then consider door i as
open from this point on, and start marking it as such in X.
This way of describing the run is a little bizarre, but is in fact very natural, as our doors are described by
an accumulative distribution function.
For two histories X = (X1 , . . . Xn ) and X 0 = (X10 , . . . Xn0 ), we write X 0 X if for all j, Xj0 ⊆ Xj . We
note that Definition 19 combined with positive correlation, gives us that if X 0 X then φA (X) ≤ φB (X 0 ).
This fact together with a simple argument finishes the proof: Use the same random coins to run the
sequence π in both A and B. By induction and the fact above, the histories at any point in time satisfy
XB XA , and so the run on A will always be at least as fast as the run on B. Since this is true no
matter what ai ’s we got, it is true in expectation.
Together these two claims prove the lemma we need for the paper:
Lemma 1. Consider similar configurations C, X and I, where X is cascading and I is independent. For
every sequence π, TI (π) ≤ TC (π) ≤ TX (π). This also implies that TI ≤ TC ≤ TX .
B.2
A Simple Upper Bound on the Price of Lacking Feedback
Claim 22. For every configuration C, Price(C) ≤ d.
Proof. Denote by X the cascading configuration that is similar to C. Denote π = (1, 2, . . . , d)∞ . Using
Lemma 1,
TC ≤ TC (π) ≤ TX (π)
The behaviour of door i in the cascading case can be described in a simple manner: It doesn’t open
until all lower index doors are open, and from that time is behaves according to pi . Hence, the expected
number of knocks on door i until it opens when starting the count after all doors j < i are open, is
precisely Ei .
In sequence π, it takes dEi to guarantee that door i was knocked upon Ei times. Therefore,
Pd by
linearity of expectation, it follows that the expected time until we open all doors in X is at most d i=1 Ei .
Pd
Dividing by i=1 Ei , we get the result.
B.3
Existence of an Optimal Sequence
Claim 23. For any configuration C there is some sequence π such that for every π 0 , TC (π) ≤ TC (π 0 ).
Proof. Assume there is no optimal sequence. Recall we assume that the fundamental distribution of each
door allows it to be opened in finite expected time. It is then easy to see that the sequence (1, 2, . . . d)∞
will open all doors in finite time no matter what the configuration is as long as it is acyclic. Therefore,
I = inf π TC (π) exists.
Take a sequence of sequences π (1) , π (2) , . . . where limn→∞ TC (π (n) ) = I. W.l.o.g., we can assume that
(n+1)
π
agrees with π (n) on all the first n places. How so? there is at least one door number that appears
as the first knock in infinitely many of the sequences. Take one such number, and erase all sequences that
don’t have it as a first knock. Of the remaining sequences, take the first one, fix it as α(1) , and erase it.
Starting with the sequence of sequences that remains, find a number that appears infinitely often in the
second place. Erase all sequences not having it as the second knock, and then fix α(2) as the first of the
remaining sequences, erase it and continue thus. We get that the α(i) ’s are a sub-sequence of the π (i) ’s,
and satisfy the assumption.
(n)
Define πi = limn→∞ πi . It is clearly defined for such a sequence of sequences. This is our π. Now:
TC (π) = lim
n→∞
= lim
n→∞
n
X
i=0
n
X
Pr [π not finished by time i]
h
i
Pr π (n) not finished by time i
i=0
≤ lim TC (π (n) ) = I
n→∞
Where the second equality is because π = π (n) in the first n places.
13
C
C.1
Proofs Related to Section 2
Asimp is Optimal up to a Constant Factor for Identical Independent Doors
Claim 5. If all doors are similar then TI (Asimp ) = Θ(TI )
Proof. By Claim 23, there is some fixed sequence π such thatPTI (π) = TI . Denote by πi (t) the number
of times door i has been knocked on by time t in π. Clearly i πi (t) = t.
∞
X
TI = TI (π) =
Pr [some door is closed at time t]
t=0
∞
X
=
1 − Pr [all doors are open at time t] =
t=0
∞
X
1−
t=0
d
Y
!
(1 − p(πi (t)))
i=1
By time t, the number of doors that have been tried more than 2t/d is less than d/2. So at least half the
doors have been tried at most t0 = b2t/dc times. Therefore, each such door i satisfies p(πi (t)) ≥ p(t0 ).
d
Qd
We then have: i=1 (1 − p(πi (t))) ≤ (1 − p(t0 )) 2 . So TI is at least:
∞
X
d
1 − (1 − p (b2t/dc)) 2
t=0
In general, for any x ≤ 1, as t traverses all integers from 0 to infinity, btxc takes every natural value at
least b1/xc times. In our case we get:
X
∞
d
d
·
1 − (1 − p(t)) 2
(3)
TI ≥
2
t=0
We now turn to analyse the expected running time of Asimp . By Claim 4, TI (Asimp ) = O(d·E [max (X1 , . . . , Xd )]),
where Xi is the number of knocks on door i until it opens. Now:
E [max (X1 , . . . , Xd )] =
∞
X
d
1 − Pr [Xi ≤ t] =
t=0
d/2
Denote x(t) = (1 − p(t))
∞
X
, and then the sum
1 − x(t)2 =
t=0
∞
X
1 − (1 − p(t))d
t=0
P∞
t=0 (1
− (1 − p(t))d ) becomes:
∞
∞
∞
X
X
X
(1 − x(t))(1 + x(t)) ≤ 2
1 − x(t) = 2
1 − (1 − p(t))d/2
t=0
t=0
t=0
Applying this, and then using Equation (3) we get:
TI (Asimp ) = O d
∞
X
!
1 − (1 − p(t))
d/2
= O(TI )
t=0
C.2
A Simple Algorithm for General Configurations Where all Doors are
Similar
n
n
Claim 6. Denote by αn the sequence 12 , . . . , d2 . If all doors are similar then for any configuration C,
TC (α0 · α1 · α2 · · ·) = Θ(TC ).
Proof. Denote π = α0 · α1 · · · , and note that |αn | = 2n d. By Lemma 1 we need only consider TX (π).
Taking an = d + 2d + . . . + 2n d, and using the right side of Observation 3 (where indices are shifted to
account for the fact that a0 is the first element and not a1 ):
TX (π) ≤ d +
∞
X
2n+1 d · SCX (π[d + 2d + . . . + 2n d)]) ≤ d + 2
n=0
∞
X
=d+4
∞
X
n=0
2n−1 d · SCI (Asimp [2n d]) ≤ d + 4TI (Asimp )
n=0
14
2n d · SCX (αn )
Where for the last step we used the left side of Observation 3, taking an = 2n d. Seeing as all doors start
closed, TI (Asimp ) ≥ d, and we get that:
TC (π) = O(TI (Asimp )) = O(TI ) = O(TC )
Where for the last two steps, we used Claim 5, and then Theorem 1.
C.3
Expected Maximum of iid Random Variables
Lemma 7. If X1 , . . . , Xd are i.i.d. random variables taking natural number values, then:
!
∞
X
E [max(X1 , . . . , Xd )] = Θ κ + d
Pr [Xi > n]
n=κ
Where κ = min {n ∈ N | Pr [X1 > n] < 1/d}
Proof. Denote p(n) = Pr [Xi > n]. The expectation we are interested in is:
∞
X
d
1 − Pr [X ≤ t] =
t=0
The first term is at least:
∞
X
1 − (1 − p(t))d =
t=0
κ−1
X
1 − (1 − p(t))d +
1
1− 1−
d
t=0
d
1 − (1 − p(t))d
(4)
t=κ
t=0
κ−1
X
∞
X
1
≥κ 1−
e
and at most κ, and so is Θ(κ). For the second term, examine (1 − a)d when a ≤ 1/d. We use 1 + x ≤ ex
and Observation 24 (see below):
1
(1 − a)d ≤ e−ad ≤ 1 − ad
2
P∞
Hence the second term of (4) is Ω(d t=κ p(t)). On the other hand, by the same observation:
(1 − a)d ≥ e−2ad ≥ 1 − 2ad
P∞
And so the second term of (4) is O(d t=κ p(t)).
Observation 24. Every 0 ≤ x ≤ 1 satisfies e−x ≤ 1 − 12 x.
Proof. Define:
1
f (x) = 1 − x − e−x
2
Whenever f is positive the required inequality is satisfied. We note that f (0) = 0 and f (1) = 1 − 12 − 1e > 0.
Now,
1
f 0 (x) = − + e−x
2
It is positive for x < ln(2) < 1, zero at ln(2), and negative for larger values. So f starts as 0 at 0, climbs
up to reach its maximum at ln(2) and then decreases. Since f (1) > 0, it must be the case that for all
0 ≤ x ≤ 1, f is positive, which proves the lemma.
C.4
Proofs for the Examples of Subsection 2.1.2
Example 9. If p(n) = q n for some 1/2 < q < 1, then Price = Θ(log(d)).
Proof. In this case, κ = dlog1/q (d)e, and E [Xi ] = 1/(1 − q), so by (1):
∞
l
m
X
Price = Θ (1 − q) log1/q (d) + d(1 − q)
qi
!
i=κ
The second term inside the brackets is equal to dq κ ≤ 1. The first term is at least:
1−q
1
ln(d) ≥ q ln(d) ≥ ln(d)
ln(1/q)
2
15
Where we used Observation 25 below. On the other hand, it is at most:
ln(d)
ln(d)
(1 − q)
≤ (1 − q) 1 +
≤ 1 + ln(d) ≤ 3 ln(d)
1−q
1−q
Since d ≥ 2 and so 2 ln(d) > 1. So we get the result.
Observation 25. For 0 < q < 1,
1 − q ≤ ln
1
1−q
≤
q
q
Proof. The following is true for all x > −1:
x
≤ ln(1 + x) ≤ x
1+x
So for 0 < x < 1:
−x
≤ ln(1 − x) ≤ −x
1−x
Which is:
x ≤ ln
1
1−x
≤
x
1−x
Taking x = 1 − q we get the first result.
1
Example 10. If for some c > 0 and a > 1, p(n) = min(1, c/na ), then Price = Θ(d a ).
Proof. In this case κ = (dc)1/a . In this proof we have many approximations (such as dropping the
rounding above), and they all go into the constants.
1
The expected time to open just one door is (we assume c a is an integer, again this will only cost a
constant factor):
∞
X
1
p(n) = c a +
n=0
∞
X
n=c1/a
1
c
≈ ca + c
a
n
Z
∞
c1/a
1
1
c
1
a
dx ≈ c a +
1 = c
a
1−
x
(a − 1)c a
1+
1
a−1
On the other hand, in the terminology of Equation (1):
1
a
E [max {X1 , . . . , Xd }] = Θ (dc) + d
∞
X
!
p(i)
i=κ
We approximate the sum in second term in the brackets by an integral:
Z ∞
1
1
∞
X
c
c
dc
dc
da ca
≈
d
dx
=
≈
d
=
1
ia
xa
(a − 1)κa−1
a−1
(a − 1)(dc)1− a
κ
i=κ
So the expectation of the maximum is:
1
1
da ca
1+
1
a−1
And we get the result.
1
Example 11. If E [X1a ] < M for some a > 1, then T = O d1+ a M 1/a (1 +
1
a−1 )
.
Proof.
E [X1a ]
M
≤ a
na
n
So the current configuration dominates the independent door configuration where each door has fundamental distribution q(n) = M/na , and so by Claim 21 has algorithms with running
time at least
as good.
Folp(n) = Pr [X1 > n] = Pr [X1a > na ] <
1
1
lowing the proof of Example 10, there is such an algorithm with running time O d1+ a M a 1 +
16
1
a−1
.
D
D.1
Proofs Related to Section 3
The Existence of an Optimal Semi-Fractional Sequence
Claim 26. There is an optimal semi-fractional sequence π. That is, for every semi-fractional sequence
π 0 , E [π] ≤ E [π 0 ]
Proof. Assume there is not. But clearly, I = inf π (E [π]) exists. Take a series π 1 , π 2 , . . . where
limn→∞ E [π n ] = I.
We think of a sequence as its sequence of 1-knock lengths. That is, πin is the length of the i-th 1-knock
in π n . We first show that we can assume that for every i, the set {πin | n ≥ 1} is bounded.
For this purpose, we first note that if for some semi-fractional sequence α, E [α] < M , then for every
i, αi < M q2i−1 . That is because with probability at least q2i−1 the algorithm will actually run the i-th
1-knock, and if it’s longer than stated, then E [α] ≥ M , in contradiction. Since we can assume that for all
n, E [π n ] < 2I, then by the observation above, we get the boundedness property we were aiming for.
Now, we claim that we can assume that for every i, πin converges as n goes to infinity. For this, start
by taking a sub-series of the π n where π1n converges (it exists, because these values are bounded, as we
said). Erase all other π n . Take the first element of this series and put aside as the new first element.
From the rest, take a sub-series where π2n converges, and erase all others. Take the new first element, and
put it aside as the new second element. Continuing this, we get an infinite series as required.
Define πi = limn→∞ πin . We claim that π is optimal.
∞
i
X
X
(πj + c) Pr [π finishes at 2-knock i]
E [π] =
i=1
j=1
Denoting by Xi the event that π finishes at or after 2-knock i, this is equal to:
∞
X
(πi + c)Pr [Xi ] = lim
k→∞
i=1
k
X
(πi + c)Pr [Xi ]
i=1
Fix some k. And denote by Xin the event that π n finishes at or after 2-knock i. Since Pr [Xi ] is a
continuous function of π1 , . . . , πi , we get:
k
k
X
X
(πi + c)Pr [Xi ] = lim
(πin + c)Pr [Xin ] ≤ lim TI (π n ) = I
i=1
n→∞
n→∞
i=1
So E [π] ≤ I and we conclude.
D.2
Memoryless Doors Imply Memoryless Algorithms
Lemma 17. If w is x-invariant, and Ex [wπ] ≤ Ex [π] then Ex [w∞ ] ≤ Ex [wπ].
Proof. As in (2), for any sequence α:
Ex [wα] = a + bEx [α]
Where a and b are functions of x and w. Since w is not empty, and as 1-knocks decrease the state and
2-knocks increase it, there must be at least one 2-knock in w, and thus b < 1. So:
Ex [wπ] = a + bEx [π] ≥ a + bEx [wπ]
=⇒
Ex [wπ] ≥
a
1−b
Ex [w∞ ] = Ex [ww∞ ] = a + bEx [w∞ ]
=⇒
Ex [w∞ ] =
a
1−b
On the other hand:
And we conclude.
17
D.3
Approximating the Optimal Semi-Fractional Running Time
Theorem 18 gives a way to calculate the expectation of the best semi-fractional sequence π ? for our
configuration. Unfortunately, we were not able to obtain a close formula for this value. The following
lemma can be used to approximate it.
√
Lemma 27. Denoting θ = −c log(q1 )/p2 , and ψ = 12 ( θ2 + 4θ − θ), we have:
1
θ
p2
1
?
log
+ + 1 − 0,
E [π ] ∈
log(1/q1 )
1−ψ
ψ
log(1/q1 )
Proof. Recall the result of Theorem 18:
c + (1 − p2 z) logq1 (1 − p2 z)
E [π ? ] = min logq1 (1 − z) +
p2 z
z∈[0,1]
By the definition of θ in the statement of the lemma, and denoting:
Y =−
(1 − p2 z) log(1 − p2 z)
p2 z
We get:
1
θ
min log(1 − z) − − Y
log(q1 ) z∈[0,1]
z
1
θ
1
min log
+ +Y
=
log(1/q1 ) z∈[0,1]
1−z
z
E [π ? ] =
Next, since for x > −1,
x
≤ log(1 + x) ≤ x
1+x
Then for 0 < x < 1:
−x ≤ log(1 − x) ≤ −
x
1−x
Multiplying by −(1 − x)/x (a positive number):
(1 − x) log(1 − x)
≤1
x
Therefore, Y ∈ [1 − p2 z, 1] ⊆ [1 − p2 , 1]. It follows that:
1
1
θ
E [π ? ] ∈
min log
+
+ [1 − p2 , 1]
log(1/q1 ) z∈[0,1]
1−z
z
1−x≤−
(5)
For the minimization, we take the derivative and compare to 0
θ
z2
1
− 2 = 0 =⇒
+z−1=0
1−z
z
θ
p
√
1 + 4/θ − 1
θ2 + 4θ − θ
=⇒ z =
=
=ψ
2/θ
2
Where we took the root that is in [0, 1]. Assigning back in (5),
1
1
θ
p2
?
E [π ] ∈
log
+ + 1 − 0,
log(1/q1 )
1−ψ
ψ
log(1/q1 )
D.4
Similar Independent Memoryless Doors
The following simple claim implies that the expected time to open two similar memoryless doors is at
most 3 times the expected time to open one of them. A generalization to d doors can easily be established
based on the same idea.
Claim 28. Consider the configuration I of two similar doors that open on each knock independently with
probability p. Then TI (Asimp ) = p3 − 1.
Proof. Until the first door opens (either door 1 or door 2), each knock has probability p to open. Therefore,
the first door opens in expected time 1/p. From that time, every odd knock will be on the other door,
and will succeed with probability p. So the expected time to open the second door after the first one has
opened is 2/p − 1, and altogether we have expected time 3/p − 1.
18
D.5
The Golden Ratio
Returning to the case where
√ c = 1, and p1 = p2 are very small. As we said, θ of Lemma 27 tends to 1,
and so ψ there tends to ( 5 − 1)/2. This ψ is actually the value of z that minimizes the expression of
Theorem 18. Looking in the proof of the theorem, the length of 1-knocks (except the first), is
t = logq (x/y) = logq (q + px) = logq (1 − pz) =
log(1 − pz)
log(1 − p)
For small x, log(1 + x) ≈ x and so, as p goes to zero, the above ratio tends to z, and in our case to ψ. So
the length of 1-knocks is ψ, and that of the 2-knocks is 1. In the long run the length of the first 1-knock
is insignificant, and the transformation of Theorem 12 will make the ratio of between the number of
2-knocks and the number of 1-knocks approach 1/ψ, which is the golden ratio.
19
References
[1] Xiaohui Bei, Ning Chen, and Shengyu Zhang. On the complexity of trial and error. In Symposium
on Theory of Computing Conference, STOC’13, Palo Alto, CA, USA, June 1-4, 2013, pages 31–40,
2013. URL: http://doi.acm.org/10.1145/2488608.2488613, doi:10.1145/2488608.2488613.
[2] David E. Bell. Regret in decision making under uncertainty. Operations Research, 30(5):961–981,
1982. URL: http://dx.doi.org/10.1287/opre.30.5.961, doi:10.1287/opre.30.5.961.
[3] Michael Ben-Or and Avinatan Hassidim. The bayesian learner is optimal for noisy binary search (and
pretty good for quantum as well). In 49th Annual IEEE Symposium on Foundations of Computer
Science, FOCS, 2008, October 25-28, 2008, Philadelphia, PA, USA, pages 221–230, 2008. URL:
http://dx.doi.org/10.1109/FOCS.2008.58, doi:10.1109/FOCS.2008.58.
[4] Lucas Boczkowski, Amos Korman, and Yoav Rodeh. Searching on trees with noisy memory. CoRR,
abs/1611.01403, 2016. URL: http://arxiv.org/abs/1611.01403.
[5] Matthias Brand, Christian Laier, Mirko Pawlikowski, and Hans J. Markowitsch. Decision making
with and without feedback: The role of intelligence, strategies, executive functions, and cognitive styles. Journal of Clinical and Experimental Neuropsychology, 31(8):984–998, 2009. PMID:
19358007. URL: http://dx.doi.org/10.1080/13803390902776860, arXiv:http://dx.doi.org/
10.1080/13803390902776860, doi:10.1080/13803390902776860.
[6] Ehsan Emamjomeh-Zadeh, David Kempe, and Vikrant Singhal. Deterministic and probabilistic
binary search in graphs. In Proceedings of the 48th Annual ACM SIGACT Symposium on Theory
of Computing, STOC 2016, Cambridge, MA, USA, June 18-21, 2016, pages 519–532, 2016. URL:
http://doi.acm.org/10.1145/2897518.2897656, doi:10.1145/2897518.2897656.
[7] Uriel Feige, Prabhakar Raghavan, David Peleg, and Eli Upfal. Computing with noisy information. SIAM J. Comput., 23(5):1001–1018, October 1994. URL: http://dx.doi.org/10.1137/
S0097539791195877, doi:10.1137/S0097539791195877.
[8] L.A. Giraldeau and T. Caraco. Social Foraging Theory. Monographs in behavior and ecology.
Princeton University Press, 2000. URL: https://books.google.co.il/books?id=Q1HDQgAACAAJ.
[9] Richard M. Karp and Robert Kleinberg. Noisy binary search and its applications. In Proceedings
of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’07, pages 881–
890, Philadelphia, PA, USA, 2007. Society for Industrial and Applied Mathematics. URL: http:
//dl.acm.org/citation.cfm?id=1283383.1283478.
[10] Michael N. Katehakis and Arthur F. Veinott, Jr. The multi-armed bandit problem: Decomposition
and computation. Math. Oper. Res., 12(2):262–268, May 1987. URL: http://dx.doi.org/10.1287/
moor.12.2.262, doi:10.1287/moor.12.2.262.
[11] Michael J. Kearns and Umesh V. Vazirani. An Introduction to Computational Learning Theory. MIT
Press, Cambridge, MA, USA, 1994.
[12] Amos Korman, Jean-Sébastien Sereni, and Laurent Viennot. Toward more localized local algorithms:
removing assumptions concerning global knowledge. Distributed Computing, 26(5-6):289–308, 2013.
URL: http://dx.doi.org/10.1007/s00446-012-0174-8, doi:10.1007/s00446-012-0174-8.
[13] Michael Luby. A simple parallel algorithm for the maximal independent set problem. SIAM
J. Comput., 15(4):1036–1053, 1986. URL: http://dx.doi.org/10.1137/0215074, doi:10.1137/
0215074.
[14] Thomas M. Mitchell. Machine Learning. McGraw-Hill, Inc., New York, NY, USA, 1 edition, 1997.
[15] Douglas C. Montgomery. Design and Analysis of Experiments. John Wiley & Sons, 2006.
[16] Andrzej Pelc. Searching games with errors - fifty years of coping with liars. Theor. Comput.
Sci., 270(1-2):71–109, 2002. URL: http://dx.doi.org/10.1016/S0304-3975(01)00303-6, doi:
10.1016/S0304-3975(01)00303-6.
[17] Richard S. Sutton and Andrew G. Barto. Introduction to Reinforcement Learning. MIT Press,
Cambridge, MA, USA, 1st edition, 1998.
20
| 2cs.AI
|
arXiv:1611.05499v1 [math.CO] 16 Nov 2016
ENUMERATION OF COMMUTING PAIRS IN LIE
ALGEBRAS OVER FINITE FIELDS
JASON FULMAN AND ROBERT GURALNICK
Abstract. Feit and Fine derived a generating function for the number
of ordered pairs of commuting n × n matrices over the finite field Fq .
This has been reproved and studied by Bryan and Morrison from the
viewpoint of motivic Donaldson-Thomas theory. In this note we give a
new proof of the Feit-Fine result, and generalize it to the Lie algebra
of finite unitary groups and to the Lie algebra of odd characteristic
finite symplectic groups. We extract some asymptotic information from
these generating functions. Finally, we derive generating functions for
the number of commuting nilpotent elements for the Lie algebras of
the finite general linear and unitary groups, and of odd characteristic
symplectic groups.
1. Introduction
Let G0 = 1, and for n ≥ 1, let Gn be the number of ordered pairs of
commuting n × n matrices (not necessarily invertible) over Fq . Feit and
Fine [FF] proved the following theorem.
Theorem 1.1.
X
n≥0
YY
Gn un
1
=
.
| GL(n, q)|
1 − ui /q l−1
i≥1 l≥0
Theorem 1.1 was reproved and refined by Bryan and Morrison [BM] from
the viewpoint of motivic Donaldson-Thomas theory, using “power structures”. See also [BBS].
In this paper we give a new proof of the Feit-Fine theorem. Answering
a question of Matt Young (an expert in Donaldson-Thomas theory), we
obtain analogs of the Feit-Fine theorem for unitary groups and for odd
characteristic symplectic groups. More precisely, let U0 = 1 and for n ≥ 1,
let Un be the number of ordered pairs of commuting elements of the Lie
algebra of GU(n, q). We will show that
YY
X Un un
1
=
| GU(n, q)|
1 − (−1)l ui /q l−1
n≥0
i≥1 l≥0
Let S0 = 1 and for n ≥ 1, let Sn be the number of ordered pairs of commuting
elements of the Lie algebra of Sp(2n, q) in odd characteristic. We will prove
Date: October 17, 2016.
1
2
JASON FULMAN AND ROBERT GURALNICK
that
X
n≥0
Q
i
Sn un
i≥1 (1 + u )
=Q Q
.
i 2l−1 )
| Sp(2n, q)|
i≥1
l≥0 (1 − u /q
Our methods will also work for odd characteristic orthogonal groups, but
we do not treat them as this is quite tedious and the ideas are the same
as for odd characteristic symplectic groups. On the other hand, for various
reasons (e.g. nilpotent classes are no longer determined by their Jordan
form even in the algebraic group and the description of semisimple classes is
different as well), even characteristic symplectic and orthogonal groups are
much trickier, and are beyond the scope of this paper.
It is easy to extract asymptotic information from our generating functions.
For example from the original Feit-Fine result, we will conclude that for q
fixed and n tending to infinity,
Y
Gn
(1 − 1/q j )−j .
lim n2 +n =
n→∞ q
j≥1
This is consistent with the fact that the variety of n × n commuting matrices
over an algebraically closed field has dimension n2 + n.
Finally, we derive an analog of the Feit-Fine theorem for nilpotent commuting pairs. Let N G0 = 1, and for n ≥ 1, let N Gn be the number of
ordered pairs of commuting nilpotent n × n matrices over Fq . We will prove
that
YY
X N Gn un
1
=
.
| GL(n, q)|
1 − ui /q l
n≥0
i≥1 l≥1
We also obtain generating functions for N Un (the number of ordered pairs
of commuting nilpotent elements of the Lie algebra of GU(n, q)), and for
N Sn (the number of ordered pairs of commuting nilpotent elements of the
Lie algebra of the odd characteristic symplectic group Sp(2n, q)).
We note there has been quite a lot of work on commuting varieties. One
of the first results in this area is the result of Motzin and Taussky [MT]
that the variety of commuting pairs in Mat(n, k) with k algebraically closed
is irreducible of dimension n2 + n. This was extended by Richardson [R2]
to reductive Lie algebras g in characteristic 0 and by Levy [Le] in good
characteristic with an extra mild assumption. It follows easily that in this
case the dimension of the commuting variety is dim g + rank g. Recall that
good characteristic means any characteristic in type A, characteristic not 2
for groups of type B, C, D and characteristic not 2 or 3 for G2 , F4 , E6 and
E7 and characteristic not 2, 3 or 5 for E8 . Very good characteristic means
good characteristic for groups not of type A and characteristic not dividing
n + 1 for groups of type An .
There has also been quite a bit of study for the variety of commuting
nilpotent pairs in a simple Lie algebra g corresponding to a Dynkin diagram in good characteristic. The components are in bijection with the set
of distinguished orbits of nilpotent elements (i.e. the centralizer in the group
ENUMERATION OF COMMUTING PAIRS
3
contains no nontrivial torus) and all components have dimension dim g (under the assumption that the characteristic is very good). This was proved by
Premet [Pr] after some earlier results of Baranovosky [Ba] and Basili [Bs] for
gln (with some restrictions on the characteristic). In particular, this variety
is irreducible for Mat(n, k). It will have a large number of components for
some of the other simple Lie algebras. For example when g = sp(2n, k) and
the characteristic of k is not 2, the number of irreducible components is the
number of partitions of n with all parts of distinct sizes.
Our methods would give similar answers for the special orthogonal Lie
algebras in characteristic not 2. We do not work out the details of this case
and leave it to the interested reader.
The organization of this paper is as follows. Section 2 treats the general linear groups. It gives a new proof of the Feit-Fine theorem, obtains
asymptotic information from it, and enumerates commuting pairs of nilpotent matrices. Section 3 studies the Lie algebra of the finite unitary groups
GU(n, q), giving unitary analogs of all the results in Section 2. Section 4
studies the Lie algebra of the finite symplectic groups Sp(2n, q) in the case
of odd characteristic, giving symplectic analogs of all the results in Section
2.
2. General linear groups
We use the following notation about partitions. Let λ be a partition of
some non-negative integer |λ| into parts λ1 ≥ λ2 ≥ · · · . Let mi (λ) be the
number of parts of λ of size i, and let λ′ be the partition dual to λ in the
sense that λ′i = mi (λ)+ mi+1 (λ)+ · · · . It is also useful to define the diagram
associated to λ by placing λi boxes in the ith row. We use the convention
that the row index i increases as one goes downward. So the diagram of the
partition (5441) is
and λ′i can be interpreted as the size of the ith column.
The notation (1/q)r will denote (1 − 1/q)(1 − 1/q 2 ) · · · (1 − 1/q r ).
The next two lemmas are well known and will be useful. Lemma 2.1 is
due to Euler and is proved on page 19 of [A].
Lemma 2.1.
X
m≥0
Y
um
1
=
.
(1/q)m
1 − u/q l
l≥0
Lemma 2.2 is elementary. In its statement, and for the rest of the paper,
we let d(φ) denote the degree of a polynomial φ. For completeness, we
include a proof.
4
JASON FULMAN AND ROBERT GURALNICK
Lemma 2.2.
Y
(1 − ud(φ) ) = 1 − uq,
φ
where the product is over all monic, irreducible polynomials over the finite
field Fq .
Proof. By unique factorization in Fq [x], the coefficient of un in the reciprocal
of the left hand side is the number of degree n monic polynomials with
coefficients in Fq . This is equal to q n , which is the coefficient of un in
(1 − uq)−1 .
We let Mat(n, q) denote the set of n × n matrices (not necessarily invertible) over the field Fq . Recall the rational canonical form of an element of
Mat(n, q). This associates to each monic, irreducible polynomial φ over Fq
a partition λφ such that
X
d(φ)|λφ | = n.
φ
For further background on rational canonical forms, one can consult Chapter
6 of [H].
Lemma 2.3 is proved in Stong [S], and calculates the number of elements
of Mat(n, q) with given rational canonical form.
Lemma 2.3. Suppose that
X
d(φ)|λφ | = n,
φ
so that {λφ } is a possible rational canonical form of an element of Mat(n, q).
Then the number of elements of Mat(n, q) with rational canonical form {λφ }
is equal to
| GL(n, q)|
.
Q d(φ) P (λ′ )2 Q
d(φ) )
i φ,i
mi (λφ )
i (1/q
φq
(Recall that λ′φ,i is the size of the ith column of the partition λφ ).
Lemma 2.4 counts the number of elements of Mat(n, q) which commute
with a given element of Mat(n, q) which has rational canonical form {λφ }.
Lemma 2.4. Let x be an element of Mat(n, q) with rational canonical form
{λφ }. Then the number of elements of Mat(n, q) which commute with x is
equal to
Y d(φ) P (λ′ )2
i φ,i .
q
φ
Proof. Note that the exponent in the statement of the result is precisely the
dimension of the centralizer of x (either in Mat(n, q) or Mat(n, k) where k
is the algebraic closure of our finite field). Note that the centralizer breaks
up into the product of the centralizers corresponding to each irreducible
ENUMERATION OF COMMUTING PAIRS
5
polynomial φ. By passing to a splitting field for φ, it suffices to assume that
φ has degree 1 and indeed that φ = x.
Now we can decompose the underlying space V into a direct sum V1 ⊕
. . .⊕Vm where x acts via a single Jordan block on Vi for size λi . Then the dimension of the centralizer is the sum of the dimensions of of Homk[x](Vi , Vj ).
Since Vj and Vj are cyclic k[x]-modules, this dimension is the minimum of
λi , λj . Clearly
X
X
X
min(λi , λj ) = 2
kmk (λ)ml (λ) +
imi (λ)2 .
i,j≥1
This is equal to
i
k<l
P
′ 2
i (λi ) ,
λ′i
as can be seen by writing
= mi (λ) + mi+1 (λ) + · · · .
Now, we reprove the Feit-Fine theorem.
Proof. We let [un ]f (u) denote the coefficient of un in an expression f (u). It
follows from Lemmas 2.3 and 2.4 that Gn (the number of ordered pairs of
commuting elements of Mat(n, q)) is equal to
| GL(n, q)|[un ]
YX
φ
λ
Q
u|λ|d(φ)
,
d(φ) )
mi (λ)
i (1/q
where the product is over all monic irreducible polynomials φ over Fq . This
is equal to
Y Y X uim·d(φ)
| GL(n, q)|[un ]
.
d(φ) )
(1/q
m
φ i≥1 m≥0
By Lemma 2.1, this is equal to
| GL(n, q)|[un ]
YYY
φ i≥1 l≥0
1
1−
ui·d(φ) /q l·d(φ)
.
Switching the order of the products and applying Lemma 2.2 shows that
this is equal to
YYY
1
| GL(n, q)|[un ]
i·d(φ)
1−u
/q l·d(φ)
i≥1 l≥0 φ
YY
1
.
= | GL(n, q)|[un ]
1 − ui /q l−1
i≥1 l≥0
This completes the proof.
Next we derive asymptotic information from the Feit-Fine generating
function.
Given a power series f (u), let [un ]f (u) denote the coefficient of un in f (u).
The following lemma will be useful.
6
JASON FULMAN AND ROBERT GURALNICK
Lemma 2.5. If the Taylor series of f around 0 converges at u = 1, then
f (u)
= f (1).
n→∞
1−u
P
n
Proof. Write the Taylor expansion f (u) = ∞
n=0 an u . Then observe that
P
(u)
[un ] f1−u
= ni=0 ai .
lim [un ]
Theorem 2.6. For q fixed,
lim
n→∞
Gn
q n2 +n
=
Y
(1 − 1/q j )−j .
j≥1
Proof. We know that
Gn = | GL(n, q)|[un ]
YY
i≥1 l≥0
= | GL(n, q)|q n [un ]
1
1 − ui /q l−1
YY
i≥1 l≥0
= qn
2 +n
(1/q)n [un ]
YY
i≥1 l≥0
Thus
lim
n→∞
Gn
q
n2 +n
=
Y
(1 − 1/q j ) lim [un ]
j≥1
Since the i = 1, l = 0 case of
from Lemma 2.5 that
YY
lim [un ]
n→∞
i≥1 l≥0
which implies the theorem.
n→∞
1
1−
ui /q i+l−1
1
.
1 − ui /q i+l−1
YY
i≥1 l≥0
1/(1− ui /q i+l−1 )
1
1−
ui /q i+l−1
.
is equal to 1/(1− u), it follows
Y
1
=
(1 − 1/q j )−j−1 ,
1 − ui /q i+l−1
j≥1
Remark: Let Com(G) denote the number of ordered pairs of commuting
elements of a finite group G. It is easy to see that Com(G) is equal to
k(G)|G|, where k(G) is the number of conjugacy classes of G. From [FG],
limn→∞ k(GL(n,q))
= 1. Thus,
qn
| GL(n, q)| Y
Com(GL(n, q))
=
lim
=
(1 − 1/q j ).
n→∞
n→∞
q n2 +n
q n2
j≥1
lim
To close this section, we enumerate pairs of commuting nilpotent n × n
matrices. Let N G0 = 1, and for n ≥ 1, let N Gn be the number of ordered
pairs of commuting n × n nilpotent matrices over Fq .
Theorem 2.7.
X N Gn un
YY
1
=
.
| GL(n, q)|
1 − ui /q l
n≥0
i≥1 l≥1
ENUMERATION OF COMMUTING PAIRS
7
The next result gives the number of nilpotent elements commuting with
a given nilpotent element.
Lemma 2.8. The number of nilpotent n × n matrices that commute with a
fixed nilpotent matrix of type λ is equal to
q
P
′ 2
i (λi ) −
P
i
mi (λ)
.
Proof. It is well known and easy that the centralizer in Mat(n, q) is
Mat(m1 , q) × . . . × Mat(mr , q),
modulo its nilpotent radical. From [Lh], the number of nilpotent matrices
2
in Mat(m, q) is q m −m . The result follows.
Now we prove Theorem 2.7.
Proof. (Of Theorem 2.7) Lemma 2.3 gives that the number of n×n nilpotent
matrices of Jordan type λ is equal to
| GL(n, q)|
P ′ 2Q
.
(λ
q i i)
i (1/q)mi (λ)
Combining this with Lemma 2.8 gives that
X
1
Q
N Gn = | GL(n, q)|
.
mi (λ)
(1/q)
mi (λ) q
i
|λ|=n
Thus
N Gn = | GL(n, q)|[un ]
YX
i≥1 m≥0
Now replacing u by
ui /q
in Lemma 2.1 gives that
X
Y
uim
1
=
.
m
(1/q)m q
1 − ui /q l
m≥0
It follows that
l≥1
N Gn = | GL(n, q)|[un ]
YY
i≥1 l≥1
proving the theorem.
uim
.
(1/q)m q m
1
,
1 − ui /q l
3. Unitary groups
We first prove some analogs of the GL case.
Let F = Fq2 and let y q = ȳ denote the q-Frobenius map on F . Let
F0 = Fq . Let k denote the algebraic closure of F . We extend this map to
an F0 automorphism of Mat(n, q 2 ). Let A⊤ denote the transpose of A.
Since GU(n, q) are those matrices U (in Mat(n, q 2 )) satisfying U Ū ⊤ = I,
the Lie algebra of GU(n, q) are those matrices A ∈ Mat(n, q 2 ) such that
Ā⊤ = −A. We call these matrices skew Hermitian. We will occasionally
write GLǫ (n, q) for GL(n, q) with ǫ = + and GU(n, q) with ǫ = −.
8
JASON FULMAN AND ROBERT GURALNICK
Choose θ ∈ F so that θ θ̄ = −1 (if F has characteristic 2, then take θ = 1
and if F has odd characteristic, just take θ so that θ 2 ∈ F0 is a nonsquare).
Then A skew Hermitian implies that B := θA is Hermitian (i.e. B̄ ⊤ = B).
Let h denote the set of Hermitian matrices. Note that the orbits of GU(n, q)
on h are in bijection with those of θh and of course the set of commuting
pairs are in bijection as well. Thus, we will obtain counts for the set of
commuting pairs in h (note that h is a Jordan algebra).
We first observe that the orbits of GU(n, q) on h are in bijection (which
is not size preserving) with those of GL(n, q) on Mat(n, q).
Lemma 3.1. Let A ∈ Mat(n, q 2 ). Then A is similar to an element of h
if and only if A is similar to an element of Mat(n, q). Moreover, A, B ∈ h
are conjugate in GL(n, k) if and only if they are conjugate by an element of
GU(n, q).
Proof. First suppose that A ∈ h. Since Ā⊤ = A and A and A⊤ are similar, it
follows that A and Ā are similar. Thus all invariant factors of A are defined
over F0 , whence A is similar to an element of Mat(n, q).
Conversely, assume that A ∈ Mat(n, q). Since all nondegenerate Hermitian forms of dimension n over F are equivalent, it suffices to show that A is
self adjoint with respect to some Hermitian form – i.e. AH = H Ā⊤ for some
invertible Hermitian matrix H. By [TZ], AS = SA⊤ for some symmetric
invertible matrix S ∈ GL(n, q). Of course, S is Hermitian and A is self
adjoint with respect to the inner product determined by S, whence the first
statement holds.
We now prove the last statement. Of course, if A, B are conjugate via
an element of GU(n, q) they are conjugate in GL(n, k). The other direction
follows by the Lang-Steinberg theorem. If B = U AU −1 , then A, B ∈ h
implies that U AU −1 = Ū −⊤ AŪ ⊤ . Thus, Ū ⊤ U centralizes A. By the LangSteinberg theorem (since the centralizer of A in GL(n, k) is connected),
Ū ⊤ U = X̄ ⊤ X for some X centralizing A. Thus, U X −1 AXU −1 = B and
U X −1 ∈ GU(n, q).
Lemma 3.2. Let A ∈ h. Then the dimension of the centralizer (over F0 )
of A in h is the same as the dimension of the centralizer (over F ) of A
in Mat(n, q 2 ) (which is the same as the dimension of the centralizer in
Mat(n, k)).
Proof. We can write Mat(n, q 2 ) = h ⊕ αh (as F0 spaces) for any element
α ∈ F \ F0 . Note that the map X → [A, X] preserves this decomposition.
Thus, C the centralizer of A is a direct sum of its intersections with h and
αh. Clearly, dimF0 C ∩ h = dimF0 αh. Letting kC denote the centralizer of
A in Mat(n, k), we thus obtain dimF0 C ∩ h = dimF C = dimk kC.
Lemma 3.3. Let A ∈ h. Let f Q
(x) ∈ F0 [x] denote the characteristic polynomial of A and factor f (x) = i φi (x)ei where the φi are distinct monic
ENUMERATION OF COMMUTING PAIRS
9
irreducible polynomials in F0 [x]. Let V denote the column space of dimension n over F . Then V is an orthogonal direct sum of spaces Vi such that
the characteristic polynomial of A on Vi is a power of φi with φi 6= φj .
Proof. By the previous lemma, for any given similarity class, there is a
representative that is of the form given. Also by the previous lemma, the
similarity class determines the orbit under the unitary group and the result
follows.
We now want to determine the centralizer of A ∈ h. By the previous
lemma, it suffices to do this under the assumption that the characteristic
polynomial of A is φ(x)m where φ(x) ∈ F0 [x] is irreducible of degree d. The
similarity classPof such an A (given φ(x)) is determined by a partition of m.
So write m = i ei λi where λ1 > λ2 > . . . and the partition has ei parts of
length λi .
Lemma 3.4. Let A ∈ h. Let G = GU(n, q). Then CG (A) has unipotent
radical Q of order q ℓ where ℓ is the dimension of the unipotent radical of
CGL(n,k) (A) and
Y
GLǫ (ei , q d ),
CG (A)/Q ∼
=
i
where ǫ = + if d is even and − if d is odd.
Proof. Let X be the centralizer of A in GL(n, k). This has a unipotent
radical U and Q is just the fixed points of U under the endomorphism of
GL(n, k) defining GLǫ (n, q). For any Steinberg-Lang endomorphism on a
connected unipotent group, the size of the fixed point group is as given.
So we now focus on the reductive part of the centralizer. Let V denote
the natural module (over F ).
First suppose that d is even. Then φ(x) = φ1 (x)φ2 (x) where φi (x) has
degree d/2 and φ2 (x) is the Galois conjugate of φ1 (x) under the q power
map.
It follows that V = V1 ⊕ V2 where Vi is the kernel of φi (A)n . Then V1 and
V2 are totally singular spaces (with respect to the Hermitian form) and A
has characteristic polynomial a power of φi (x) on Vi .
Let Ai denote the restriction of A to Vi . Clearly the centralizer of A
preserves each Vi . Note that if g ∈ GU(n, q) fixes each Vi , then with respect
to an appropriate basis if gi denotes g restricted to Vi , then g2 = ḡ1−⊤ .
The subgroup stabilizing V1 and V2 is isomorphic to GL(n/2, q 2 ) and the
centralizer of A in U is isomorphic to the centralizer of A1 in GL(n, q 2 ).
The result then follows by the result for GL.
Now suppose that d is odd. Then φ(x) remains irreducible over F . By the
uniqueness of the canonical forms, we may assume that V is an orthogonal
direct sum of subspaces Vi such that on Vi the rational canonical form has all
blocks the same (i.e. the minimal and characteristic polynomials are both
powers of φ(x)).
10
JASON FULMAN AND ROBERT GURALNICK
By considering the GL case (over k) we see that the reductive part of the
centralizer (up to conjugacy) preserves each Vi . Thus, it suffices to assume
that V = Vi . In the notation given above, this means that λ1 = λ, e1 = e
and deλ = n.
First suppose that φ(x) is the minimal polynomial of A (i.e. A is a
semisimple element and λ = 1). Then we can decompose the space as
a direct sum of e orthogonal spaces each of dimension d so that A acts
irreducibly on each space. By the GL case, the centralizer in GL(n, F ) is just
GL(e, q 2d ). We can view the standard Hermitian form as inducing one on
the e dimensional vector space over Fq2d and clearly GL(e, q 2d )∩GU(de, q) =
GU(e, q) as claimed.
Finally suppose that λ > 1. Then we can decompose the space as V1 ⊕
. . . ⊕ Vλ (this is not an orthogonal decomposition and indeed V1 will be
totally singular) so that the matrix of A with respect to this decomposition
is
B I
0 ... 0
0 B I ... 0
.
···
0 0 0 ... B
Here B is semisimple and has minimal polynomial φ(x) and characteristic
polynomial φ(x)e . We can take the Hermitian form preserved by this matrix
to be in block diagonal form (as in the above equation) with the antidiagonal
blocks all being I and all other blocks being 0. So we need to compute the
centralizer with respect to the unitary group preserving this form and as
before we only need to compute the reductive part of this centralizer. Again,
considering the case of GL, we see that the reductive part of the centralizer
will be block diagonal. It is straightforward to see that if x is block diagonal
centralizing A, then all diagonal blocks are the same. By the previous case
this implies that x ∈ GU(e, q d ) and clearly any such element commutes with
A, whence the result.
Lemma 3.5 is a unitary analog of Lemma 2.3 and the proof is identical
given the previous results.
Lemma 3.5. The orbits of GU(n, q) on its Lie algebra correspond to {λφ }
satisfying
X
d(φ)|λφ | = n,
φ
that is to the possible rational canonical forms of an element of Mat(n, q).
The size of the orbit corresponding to data {λφ } is
| GU(n, q)|
Q
.
φ Cφ (λ)
ENUMERATION OF COMMUTING PAIRS
Here
Cφ (λ) = q d(φ)
P
2
′
i (λφ,i )
Y
11
(1/q d(φ) )mi (λφ )
i
if d(φ) is even, and
Cφ (λ) = q d(φ)
P
2
′
i (λφ,i )
Y
(1/q d(φ) )mi (λφ ),q7→−q
i
if d(φ) is odd. Here the notation (1/q d(φ) )mi (λφ ),q7→−q means that we replace
q by −q in the expression (1/q d(φ) )mi (λφ ) .
Lemma 3.6 is a unitary analog of Lemma 2.4, and the proof is identical
given the previous results.
Lemma 3.6. Let x be an element of the Lie algebra of GU(n, q), with data
{λφ }. Then the number of elements of the Lie algebra of GU(n, q) which
commute with x is equal to
Y d(φ) P (λ′ )2
i φ,i .
q
φ
Now we prove the unitary analog of the Feit-Fine theorem. Let U0 = 1
and for n ≥ 1, let Un be the number of ordered pairs of commuting elements
of the Lie algebra of GU(n, q).
Theorem 3.7.
X
n≥0
YY
1
Un un
=
| GU(n, q)|
1 − (−1)l ui /q l−1
i≥1 l≥0
Proof. Let [un ]f (u) denote the coefficient of un in an expression f (u). It
follows from Lemmas 3.5 and 3.6 that Un is equal to | GU(n, q)| multiplied
by the coefficient of un in
Y
φ
d(φ) even
X
λ
Q
u|λ|d(φ)
d(φ) )
mi (λ)
i (1/q
Y X
λ
φ
d(φ) odd
u|λ|d(φ)
.
d(φ) )
mi (λ),q7→−q
i (1/q
Q
This is equal to | GU(n, q)| multiplied by the coefficient of un in
Y
φ
d(φ) even
uim·d(φ)
(1/q d(φ) )m
i≥1 m≥0
YX
Y YX
φ
d(φ) odd
By Lemma 2.1, this is equal to
YYY
| GU(n, q)|[un ]
φ i≥1 l≥0
i m≥0
Q
uim·d(φ)
.
d(φ) )
m,q7→−q
i (1/q
1
1−
(ui (−1)l /q l )d(φ)
.
12
JASON FULMAN AND ROBERT GURALNICK
Switching the order of the products and applying Lemma 2.2 shows that
this is equal to
YYY
1
| GU(n, q)|[un ]
i
1 − (u (−1)l /q l )d(φ)
i≥1 l≥0 φ
YY
1
.
= | GU(n, q)|[un ]
1 − (−1)l ui /q l−1
i≥1 l≥0
This completes the proof.
Theorem 3.8 determines the asymptotics of the sequence Un .
Theorem 3.8. For q fixed,
Y
Y
Un
lim n2 +n =
(1 − 1/q i )−⌊i/4⌋ .
(1 − 1/q i )−1
n→∞ q
i even
i odd
Proof. We know that
Un = | GU(n, q)|[un ]
YY
i≥1 l≥0
= | GU(n, q)|q n [un ]
1
1 − (−1)l ui /q l−1
YY
i≥1 l≥0
Thus
Un
n
n→∞ q 2 +n
lim
=
Y
1
.
1 − (−1)l ui /q i+l−1
(1 − (−1)j /q j ) lim [un ]
n→∞
j≥1
YY
i≥1 l≥0
1
1−
(−1)l ui /q i+l−1
1/(1 − (−1)l ui /q i+l−1 )
.
Since the i = 1, l = 0 case of
is equal to 1/(1 − u), it
follows from Lemma 2.5 that
Un
1
1
lim
=
·
n→∞ q n2 +n
(1 − 1/q) (1 + 1/q 2 )(1 − 1/q 2 )
1
·
3
(1 − 1/q )(1 + 1/q 3 )(1 − 1/q 3 )
1
···
·
4
4
(1 + 1/q )(1 − 1/q )(1 + 1/q 4 )(1 − 1/q 4 )
Y
1
=
(1 − 1/q i )−1 ·
2
(1 + 1/q )(1 − 1/q 2 )
i odd
1
(1 + 1/q 3 )(1 − 1/q 3 )
1
··· .
·
4
4
(1 + 1/q )(1 − 1/q )(1 + 1/q 4 )(1 − 1/q 4 )
·
Writing (1 + 1/q k )(1 − 1/q k ) = (1 − 1/q 2k ) shows that this is equal to
Y
Y
(1 − 1/q i )−⌊i/4⌋ .
(1 − 1/q i )−1
i odd
i even
ENUMERATION OF COMMUTING PAIRS
13
Remark: Let Com(G) denote the number of ordered pairs of commuting
elements of a finite group G. As mentioned earlier, Com(G) is equal to
k(G)|G|, where k(G) is the number of conjugacy classes of G. From [FG],
k(GU(n, q)) Y (1 + 1/q i )
=
.
lim
n→∞
qn
(1 − 1/q i )
i≥1
Thus,
Com(GU(n, q))
n→∞
q n2 +n
lim
=
Y (1 + 1/q i )
i≥1
=
lim
(1 − 1/q i ) n→∞
| GU(n, q)|
q n2
Y (1 + 1/q i ) Y
(1 + 1/q i ).
(1 − 1/q i )
i≥1
i odd
Next we enumerate pairs of commuting nilpotent matrices in the Lie algebra of GU(n, q). Let N U0 = 1, and for n ≥ 1, let N Un be the number of
ordered pairs of commuting n × n nilpotent matrices in the Lie algebra of
GU(n, q).
Theorem 3.9.
YY
X N Un un
1
=
.
| GU(n, q)|
(1 + (−1)l ui /q l )
i≥1 l≥1
n≥0
We next state a lemma about counting nilpotent elements in a Lie subalgebra.
Lemma 3.10. Let G be a connected algebraic group with Lie algebra g over
an algebraically closed field of characteristic p. Let F be a Lang-Steinberg
endomorphism of G (so GF is a finite group). The number of nilpotent
elements in (Lie(G))F is q dim G−rank G (where q = qF is the size of the field
associated with F ).
Proof. Let Q be the unipotent radical of G. Of course Q is F -invariant.
Let h be the Lie algebra of Q. Then h consists of nilpotent elements and
hF will have cardinality q dim Q . Thus, we can pass to the case that G is
reductive. In that case, we just apply a result of Lehrer [Le, Cor. 1.15] (this
also follows by a result of Steinberg [St] in good characteristic – Steinberg
proves the corresponding result for unipotent elements).
Lemma 3.11. The number of nilpotent n × n matrices which are in the Lie
algebra of GU(n, q) and which commute with a fixed nilpotent matrix of type
λ in the Lie algebra of GU(n, q) is equal to
q
P
′ 2
i (λi ) −
P
i
mi (λ)
.
Proof. Let A ∈ h be nilpotent of type
P λ. As we have already noted, the
dimension of the centralizer of A is i (λ′i )2 . We have also seen that the
14
JASON FULMAN AND ROBERT GURALNICK
reductive part of the centralizer of A
Pis a direct product of GL(mi (λi )),
whence the rank of the centralizer is i mi (λ). Thus, the result follows by
Lemma 3.10.
Now we prove Theorem 3.9.
Proof. (Of Theorem 3.9) From Lemma 3.5 we know that the total number
of elements of the Lie algebra of GU(n, q) which are nilpotent of type λ is
equal to
| GU(n, q)|
P ′ 2Q
.
(λ
)
i
q
i (1/q)mi (λ),q7→−q
Here the notation (1/q)mi (λ),q7→−q means that we replace q by −q in the
expression (1/q)mi (λ) .
Combining this with Lemma 3.11 gives that
X
1
Q m (λ)
N Un = | GU(n, q)|
i
(1/q)mi (λ),q7→−q
iq
|λ|=n
= | GU(n, q)|[un ]
YX
= | GU(n, q)|[un ]
YX
i≥1 m≥0
i≥1 m≥0
It follows from Lemma 2.1 that
X [(−1)ui ]m
q m (1/q)m
m≥0
Replacing q by −q gives that
X
[(−1)ui ]m
[q m (1/q)m ]q7→−q
m≥0
Hence
N Un = | GU(n, q)|[un ]
=
[(−1)ui ]m
.
[q m (1/q)m ]q7→−q
Y
1
.
1 + ui /q l
Y
1
.
1 + (−1)l ui /q l
l≥1
=
uim
q m (1/q)m,q7→−q
l≥1
YY
i≥1 l≥1
1
.
(1 + (−1)l ui /q l )
4. Symplectic Groups
Here we consider the symplectic case. There are several extra difficulties
in characteristic 2. So let F = Fq with q an odd prime power. Let G =
Sp(2n, q). Let sp be the Lie algebra of G. This can be viewed in a number
of ways.
One way is as follows. Let J denote the skew symmetric matrix defining
G (i.e. G is the stabilizer of J). Then
sp := {A ∈ Mat(n, q)|AJ = (AJ)⊤ }.
ENUMERATION OF COMMUTING PAIRS
15
We can also view sp as the linear transformations on the natural module
V for G which are self adjoint with respect to the alternating form defining
G. Note that V is a self-dual G-module and so V ∼
= V ∗ (where V ∗ denotes
the dual module).
First let us make some remarks if we work over the algebraic closure k
of F . Note that if x ∈ sp, then if 0 6= α is an eigenvalue of x, so is −α
and the generalized α eigenspace of x is totally singular. The sum of the
generalized α and −α eigenspaces is nonsingular and the Jordan form of x
on the α space is the same as the −α space. Moreover the centralizer in
Sp(2m) is just the centralizer of x in GL(m) (on the α) space. In particular,
the centralizer is connected.
We first compute the dimension of the centralizer of an element x ∈ sp.
As usual, we reduce to the case where the characteristic polynomial of x is
a power of an irreducible polynomial φ(x) or the product of two irreducible
polynomials φi (x) where φ2 (x) = (−1)d φ1 (−x) where d is the degree of φi .
In the first case note that either φ(x) = x or φ(x) has even degree 2d (and
φ(x) = φ(−x)).
First consider the case that φ(x) = x (i.e. x is nilpotent). Then the orbit
of x under Sp(2n, k) is determined by the Jordan form of x which can be
described as a partition of n. Let λ1 ≥ λ2 ≥ . . . ≥ λr (such a partition
corresponds to an element of sp if and only if the multiplicity of the odd
pieces in the partition is even).
Recall that sp as a module either for G or for the Lie algebra is isomorphic
to Sym2 (V ) and this is a direct summand of V ⊗ V (which is isomorphic to
V ⊗ V ∗ as a module for x) since we are assuming that q is odd. Since the
dimension of the fixed point space on V ⊗ V depends only on the partition
and not on the characteristic, the same is true for the dimension of the
fixed point space on Sym2 (V ). This is well known (for example, this can
be computed in characteristic 0 using the decomposition of Sym2 (V ) as an
SL(2)-module). This proves the first expression in the following lemma.
Lemma 4.1. Let x ∈ sp be nilpotent with Jordan block sizes λ1 ≥ λ2 ≥
. . . ≥ λr on the natural module. Then the dimension of the centralizer of x
in sp is equal to both:
(1)
X
j<i
λi +
X
⌈λi /2⌉,
i
and
(2)
X
(λ′i )2 /2 + o(λ)/2,
i
where o(λ) is the number of odd parts of λ.
16
JASON FULMAN AND ROBERT GURALNICK
Proof. The expression in part 1 of the lemma is equal to
X
X
X
λi /2 + o(λ)/2 =
λi (i − 1) + |λ|/2 + o(λ)/2
λi +
j<i
i
i
=
X λ′
i
i
=
2
+ |λ|/2 + o(λ)/2
X
(λ′i )2 /2 + o(λ)/2.
i
Now consider the case that φ(x) 6= x.
Then the class of the element is determined by φ(x) (or the φi (x)) and a
partition of n/d.
To compute the dimension of the centralizer of x in sp, we can work
over the algebraic closure (or any extension field) and so reduce to the case
where the characteristic polynomial is a power of φ1 (x)φ2 (x). Then we can
write V = V1 ⊕ V2 where each Vi is a maximal totally singular subspace,
Vi is invariant under x and the characteristic polynomial of x on Vi is a
power of φi (x). Then we see that the centralizer of x in sp is isomorphic
to the centralizer of xi (the restriction of x to Vi ) in the algebra of linear
transformations of Vi and so the result for GL gives Theorem 4.2 below.
Recall that Jordan canonical form associates to each monic, irreducible
polynomial φ over Fq a partition λφ such that
X
d(φ)|λφ | = 2n.
φ
This data arises from an element of sp if and only if:
(1) The partition λx has all odd parts occur with even multiplicity
(2) For every φ, the partition λφ is equal to the partition λφ , where φ is
defined as (−1)d(φ) φ(−x).
Thus, we have proved:
Theorem 4.2. If {λφ } is the data corresponding to an element of sp, then
the dimension of its centralizer in sp is equal to
X
X
X
X
X
d(φ)
(λ′φ,i )2 .
d(φ)
(λ′φ,i )2 /2 +
(λ′x,i )2 /2 + o(λx )/2 +
i
φ=φ
φ6=x
i
{φ,φ}
φ6=φ
i
Finally, we need to count how many elements there are in sp with a given
canonical form. This amounts to computing the centralizers in Sp. As in the
discussion for centralizers in sp, we can reduce to the three cases as above.
It is a bit more involved as we now need the order of the centralizer rather
than just the dimension.
For nilpotent elements, the centralizer is not usually connected in the
algebraic group and so the orbits break up over the finite field. However,
ENUMERATION OF COMMUTING PAIRS
17
since the centralizer dimension in the Lie algebra depends only the orbit
over k, we only need to know the following.
Lemma 4.3. The number of nilpotent elements in sp corresponding to a
partition λ1 ≥ λ2 ≥ . . . ≥ λr is
q
P
′ 2
i (λi ) /2+o(λ)/2
Q
| Sp(2n, q)|
2
2⌊mi (λ)/2⌋ )
i (1 − 1/q ) · · · (1 − q
Proof. Since we are in good characteristic, the number of nilpotent elements
with a given Jordan form is the same as the number of unipotent elements
in Sp(2n, q) with the same distribution of Jordan blocks. This number is
determined in [F].
For the elements in sp that do not have 0 as an eigenvalue, we have
seen that the centralizers in the algebraic group are connected, whence the
Sp(2n, k)-orbits intersect sp(2n, q) are just the Sp(2n, q) orbits. So we just
need to compute the centralizers in Sp(2n, q).
The next result gives the number of elements in sp(2n, q) with a given
Jordan canonical form.
Theorem 4.4. If {λφ } is the data corresponding to an element x of sp,
then the number of elements of sp corresponding to this data is equal to
| Sp(2n, q)| multiplied by:
1
P ′ 2
Q
(λ
)
/2+o(λ
)/2
x
2
2⌊mi (λx )/2⌋ )
q i x,i
i (1 − 1/q ) · · · (1 − 1/q
Y
1
P
·
d(φ) i (λ′φ,i )2 /2 Q
d(φ)/2
) · · · (1 − (−1)mi (λφ ) /q mi (λφ )d(φ)/2 )
q
i (1 + 1/q
φ=φ
φ6=x
·
Y
{φ,φ}
φ6=φ
q
d(φ)
P
′
2
i (λφ,i )
1
Q
i (1
− 1/q d(φ) ) · · · (1 − 1/q mi (λφ )d(φ) )
Proof. By the discussion before Lemma 4.1 and Lemma 4.3 it suffices to
reduce to the case where φ 6= x and φ is either irreducible (and φ(x) =
φ(−x)) or φ = φ1 φ2 where φ1 is irreducible and φ2 is the dual polynomial
of φ1 (i.e. its roots are the negatives of the roots of φ1 ).
In the latter case, V decomposes as a direct sum of two totally isotropic
spaces V1 ⊕ V2 where the element has characteristic polynomial a power of
φi on Vi and the action on V2 is the dual of the action on V1 . Thus, the
centralizer of x is isomorphic to the centralizer in GL(V1 ) of x restricted to
V1 and the result follows by Lemma 2.4.
So it remains to consider the case that φ is irreducible. The unipotent
radical of the centralizer has order q dim Q where Q is the unipotent radical of
the centralizer in the algebraic group (and so can be computed as above). So
we only have to identify the reductive part. Modulo the unipotent radical,
we see that we can reduce to the case where all parts of the partition have
18
JASON FULMAN AND ROBERT GURALNICK
the same size e (this can be seen in the algebraic group). Let m denote the
this multiplicity and let 2d be the degree of φ.
If we pass to a quadratic extension of our finite field, we are in the previous
case and so the reductive part of the centralizer is GL(m, q 2d ). Arguing as
we did in the unitary case, we see that the reductive part of the centralizer
is GU(m, q d ).
Recall that φ is defined as (−1)d(φ) φ(−x). Note that if a monic irreducible
polynomial φ 6= x satisfies φ = φ, then φ has even degree. We let N (2d, q)
denote the number of monic irreducible polynomials φ of degree 2d such
that φ = φ. It is also helpful to let M (d, q) denote the number of unordered
pairs {φ, φ} of monic, irreducible, degree d polynomials such that φ 6= φ.
The following enumerative lemma will be helpful. It is a symplectic analog
of Lemma 2.2.
(1)
Lemma 4.5.
Y
(1 − ud )−N (2d,q)
Y
(1 − ud )−M (d,q) =
d≥1
d≥1
1−u
.
1 − qu
(2)
Y
(1 + ud )−N (2d,q)
d≥1
Y
(1 − ud )−M (d,q) = 1.
d≥1
Proof. To prove the first assertion, note that any monic polynomial (not
necessarily irreducible) which is invariant under the involution f (x) 7→
(−1)d(f ) f (−x) factors uniquely into irreducibles as a product of powers of
φ where φ = φ and of powers of φφ where φ 6= φ. Hence the coefficient of
un in
Y
Y
(1 − ud )−N (2d,q)
(1 − u)−1
(1 − ud )−M (d,q)
d≥1
d≥1
is equal to the number of monic degree 2n polynomials f such that f (x) =
1
, proving
f (−x). This is equal to q n , which is the coefficient of un in 1−qu
the first part of the theorem.
To prove the second part of the theorem, we claim that
Y
Y
1
(1 − u)−1
(1 − u2d )−N (2d,q)
.
(1 − ud )−2M (d,q) =
1 − qu
d≥1
d≥1
Indeed, the claimed equation follows since the coefficient of un in the left
hand side counts the total number of monic degree n polynomials, as does
the coefficient of un in 1/(1 − qu). Dividing the claimed equation by the
first assertion of the theorem immediately proves the second assertion of the
theorem.
Let S0 = 1, and for n ≥ 1, let Sn be the number of ordered pairs of
commuting elements of the Lie algebra of Sp(2n, q). Armed with the above
results, we can now derive a generating function for the sequence Sn .
ENUMERATION OF COMMUTING PAIRS
19
Theorem 4.6.
X
n≥0
Q
i
Sn un
i≥1 (1 + u )
=Q Q
.
i 2l−1 )
| Sp(2n, q)|
i≥1
l≥0 (1 − u /q
Proof. It is immediate from Theorems 4.2 and 4.4 that Sn is equal to
| Sp(2n, q)| multiplied by the coefficient of u2n in ABC where
X
u|λ|
Q
A=
2
4
2⌊mi (λ)/2⌋ )
i (1 − 1/q )(1 − 1/q ) · · · (1 − 1/q
λ
i odd =⇒ mi (λ) even
B=
YX
λ
φ=φ
φ6=x
C=
Q
ud(φ)|λ|
d(φ)/2 ) · · · (1 − (−1)mi (λ) /q mi (λ)d(φ)/2 )
i (1 + 1/q
Y X
{φ,φ}
φ6=φ
λ
Q
u2d(φ)|λ|
.
d(φ) ) · · · (1 − 1/q mi (λ)d(φ) )
i (1 − 1/q
Note that in A, the sum is over all partitions such that the odd parts occur
with even multiplicity, and that in B and C, the sum is over all partitions.
Note that A = (A1)(A2) where
A1 =
Y
i odd m
and A2 is equal to
"
Y
X
i even
i≥2
m
uim
(1 − 1/q 2 )(1 − 1/q 4 ) · · · (1 − 1/q m )
even
X
X
uim
uim
+
(1 − 1/q 2 ) · · · (1 − 1/q m )
(1 − 1/q 2 ) · · · (1 − 1/q m−1 )
even
m odd
Applying Lemma 2.1 gives that
A1 =
Y Y
i odd l≥0
A2 =
Y
1
(1 − u2i /q 2l )
(1 + ui )
i even
i≥2
Y
l≥0
1
.
(1 − u2i /q 2l )
Next observe that
B =
φ=φ
φ6=x
=
uimd(φ)
(1 + 1/q d(φ)/2 ) · · · (1 − (−1)m /q md(φ)/2 )
i≥1 m≥0
YYX
Y YY
φ=φ
φ6=x
i≥1 l≥0
1
(1 −
(−1)l uid(φ) /q ld(φ)/2 )
where the second equality used Lemma 2.1.
#
20
JASON FULMAN AND ROBERT GURALNICK
Similarly,
u2imd(φ)
(1 − 1/q d(φ) ) · · · (1 − 1/q md(φ) )
{φ,φ} i≥1 m≥0
Y YX
C =
φ6=φ
Y YY
=
{φ,φ}
φ6=φ
i≥1 l≥0
1
(1 −
u2id(φ) /q ld(φ) )
.
Switching the order of the products in B and C, and then applying both
parts of Lemma 4.5 gives that
Y
YY Y
(1 − (−1)l u2id /q ld )−N (2d,q)
(1 − u2id /q ld )−M (d,q)
BC =
l≥0 i≥1
=
d≥1
d≥1
Y Y (1 − u2i /q l )
.
(1 − u2i /q l−1 )
l even
l≥0
i≥1
Hence
Q
2i
i≥1 (1 + u )
Q
Q
,
ABC = (A1)(A2)(BC) =
2i 2l−1 )
i≥1
l≥0 (1 − u /q
which proves the theorem.
Next we use Theorem 4.6 to obtain asymptotic information about the
number of commuting pairs in the Lie algebra of Sp(2n, q).
Theorem 4.7. For q fixed,
Y
Sn
(1 + 1/q i )(1 − 1/q i )−⌊(i+1)/2⌋ .
lim 2n2 +2n =
n→∞ q
i≥1
Proof. We know from Theorem 4.6 that
Sn
Q
i
i≥1 (1 + u )
= | Sp(2n, q)|[u ] Q Q
i 2l−1 )
i≥1
l≥0 (1 − u /q
Q
i i
i≥1 (1 + u /q )
n n Q
Q
= | Sp(2n, q)|q [u ]
i 2l+i−1 )
i≥1
l≥0 (1 − u /q
n
= q 2n
(1 − 1/q 2 ) · · · (1 − 1/q 2n )
Q
i i
i≥1 (1 + u /q )
n Q
Q
.
·[u ]
i 2l+i−1 )
i≥1
l≥0 (1 − u /q
Thus
lim
n→∞
2 +2n
Sn
q 2n2 +2n
=
Y
Q
i i
i≥1 (1 + u /q )
(1 − 1/q ) lim [u ] Q Q
.
i 2l+i−1 )
n→∞
i≥1
l≥0 (1 − u /q
j≥1
2j
n
ENUMERATION OF COMMUTING PAIRS
21
Since the i = 1, l = 0 case of 1/(1 − ui /q 2l+i−1 ) is equal to 1/(1 − u), it
follows from Lemma 2.5 and elementary manipulations that
Q
Y
(1 + ui /q i )
2j
n Q
Qi≥1
(1 − 1/q ) lim [u ]
i 2l+i−1 )
n→∞
i≥1
l≥0 (1 − u /q
j≥1
is equal to
Y
(1 + 1/q i )(1 − 1/q i )−⌊(i+1)/2⌋ .
i≥1
Remark: Let Com(G) denote the number of ordered pairs of commuting
elements of a finite group G. As mentioned earlier, Com(G) is equal to
k(G)|G|, where k(G) is the number of conjugacy classes of G. From [FG],
since q is odd, one has that
k(Sp(2n, q)) Y (1 + 1/q i )4
.
lim
=
n→∞
qn
(1 − 1/q i )
i≥1
Thus,
Com(Sp(2n, q))
2
n→∞
q 2n +2n
lim
=
Y (1 + 1/q i )4
(1 −
Q
(1 + 1/q i )4
Q i≥1
.
i
i odd (1 − 1/q )
i≥1
=
lim
1/q i ) n→∞
| Sp(2n, q)|
2
q 2n +n
Our next goal is to count the number of nilpotent commuting pairs in sp.
We first require a result of Richardson [R1, Lemma 6.6] (which can be
proved quite easily in our special case). Note that for G = Sp(2n, k) all
characteristics other than 2 are very good.
Lemma 4.8. ([R1]) Let G be a simple algebraic group over an algebraically
closed field k. Assume that the characteristic is very good for G. Let g be
the Lie algebra of G. If x ∈ g, then {y ∈ g|[x, y] = 0} is the Lie algebra of
CG (x).
Lemma 4.9. If an element x of sp is nilpotent of type λ, then its centralizer
in sp has size
P ′ 2
P
q i (λi ) /2+o(λ)/2− i ⌊mi (λ)/2⌋ .
Proof. Note that we are in very good characteristic. As we have already
seen,
P the′ 2dimension of the centralizer in the group of a nilpotent element x
is
P i (λi ) /2 + o(λ)/2. It follows by [LS] that the rank of the centralizer is
i ⌊mi (λ)/2⌋. By the result of Richardson above, the same is true for the
centralizer of x in sp. The result now follows by Lemma 3.10,
This leads to the following theorem. In its statement, we let N S0 = 1,
and let N Sn denote the number of commuting ordered pairs of nilpotent
elements in the Lie algebra of Sp(2n, q).
22
JASON FULMAN AND ROBERT GURALNICK
Theorem 4.10.
X
n≥0
Q
i
N Sn un
i≥1 (1 + u )
=Q Q
.
i 2l+1 )
| Sp(2n, q)|
i≥1
l≥0 (1 − u /q
Proof. It follows from Lemmas 4.3 and 4.9 that N Sn is equal to
X
1
Q
.
| Sp(2n, q)|
2 ) · · · (1 − 1/q 2⌊mi (λ)/2⌋ )q ⌊mi (λ)/2⌋
(1
−
1/q
i
|λ|=2n
i odd =⇒ mi (λ) even
Letting [un ]f (u) denote the coefficient of un in a power series f (u), it follows
that
N Sn = | Sp(2n, q)|[u2n ]AB
where
Y X
uim
A=
(1 − 1/q 2 ) · · · (1 − 1/q m )q m/2
i odd m even
B =
Y X
i even
i≥2
=
Y
i even
i≥2
m
uim
(1 − 1/q 2 ) · · · (1 − 1/q 2⌊m/2⌋ )q ⌊m/2⌋
(1 + ui )
uim
.
(1 − 1/q 2 ) · · · (1 − 1/q m )q m/2
even
X
m
It follows from Lemma 2.1 that
Q
2i
i≥1 (1 + u )
Q
Q
AB =
,
2i 2l+1 )
i≥1
l≥0 (1 − u /q
proving the theorem.
5. Acknowledgements
Fulman was partially supported by Simons Foundation Grant 400528.
Guralnick was partially supported by NSF grant DMS-1600056. The authors
thank Matt Young for asking about analogs of the Feit-Fine theorem for
other Lie algebras.
References
[A]
[Ba]
Andrews, G., The theory of partitions, Cambridge University Press, 1984.
Baranovsky, V., The variety of pairs of commuting nilpotent matrices is irreducible,
Transform. Groups 6 (2001), 3-8.
[Bs]
Basili, R., On the irreducibility of commuting varieties of nilpotent matrices, J.
Algebra 268 (2003), 58-80.
[BBS] Behrend, K., Bryan, J., and Szendroi, B., Motivic degree zero Donaldson-Thomas
invariants, Invent. Math. 192 (2013), 111-160.
[BM] Bryan, J. and Morrison, A., Motivic classes of commuting varieties via power
structures, J. Algebraic Geom. 24 (2015), 183-199.
[FF] Feit, W. and Fine, N.J., Pairs of commuting matrices over a finite field, Duke
Math. J 27 (1960), 91-94.
ENUMERATION OF COMMUTING PAIRS
[F]
[FG]
[H]
[Lh]
[Le]
[LS]
[MT]
[Pr]
[R1]
[R2]
[St]
[S]
[TZ]
23
Fulman, J., A probabilistic approach to conjugacy classes in the finite symplectic
and orthogonal groups, J. Algebra 234 (2000), 207-224.
Fulman, J. and Guralnick, R., Bounds on the number and sizes of conjugacy
classes in finite Chevalley groups with applications to derangements, Transac.
Amer. Math. Soc. 364 (2012), 3023-3070.
Herstein, I.N., Topics in algebra, Second edition, Xerox College Publishing, 1975.
Lehrer, G., Rational tori, semisimple orbits and the topology of hyperplane complements, Comment. Math. Helvetici 67 (1992), 226-251.
Levy, P., Commuting varieties of Lie algebras over fields of prime characteristic,
J. Algebra 250 (2002), 473-484.
Liebeck, M. and Seitz, G., Unipotent and nilpotent classes in simple algebraic
groups and Lie algebras, Mathematical Surveys and Monographs, 180, American
Mathematical Society, Providence, RI, 2012.
Motzkin, T.S. and Taussky, O., Pairs of matrices with property L. II, Transac.
Amer. Math. Soc. 80 (1955), 387-401.
Premet, A., Nilpotent commuting varieties of reductive Lie algebras, Invent. Math.
154 (2003), 653-683.
Richardson, R.W., Conjugacy classes in Lie algebras and algebraic groups, Ann.
of Math. 86 (1967), 1-15.
Richardson, R.W., Commuting varieties of semisimple Lie algebras and algebraic
groups, Compositio Math. 38 (1979), 311-327.
Steinberg, R., Endomorphisms of linear algebraic groups, Memoirs Amer. Math.
Society 80 (1968).
Stong, R., Some asymptotic results on finite vector spaces, Adv. in Appl. Math. 9
(1988), 167-199.
Taussky, O. and Zassenhaus, H., On the similarity transformation between a matrix and its transpose, Pacific J. Math. 9 (1959), 893-896.
Department of Mathematics, University of Southern California, Los Angeles, CA, 90089
E-mail address: [email protected]
Department of Mathematics, University of Southern California, Los Angeles, CA, 90089
E-mail address: [email protected]
| 4math.GR
|
arXiv:1704.01685v6 [cs.IT] 10 Dec 2017
A lower bound on the 2-adic complexity of
modified Jacobi sequence
Yuhua Suna,b,c , Qiang Wangb, Tongjiang Yana,c
a College
of Sciences, China University of Petroleum, Qingdao 266555, Shandong,
China(e-mail:sunyuhua [email protected]), (e-mail:[email protected]),
(e-mail:[email protected]).
b School
c
of Mathematics and Statistics, Carleton University, Ottawa, Ontario, K1S
5B6, Canada.
Key Laboratory of Network Security and Cryptology, Fujian Normal University,
Fuzhou, Fujian 350117, China
Abstract
Let p, q be distinct primes satisfying gcd(p−1, q−1) = d and let Di , i = 0, 1, · · · , d−1,
∗ = ∪d−1 D . In this paper,
be Whiteman’s generalized cyclotomic classes with Zpq
i=0 i
we give the values of Gauss periods based on the generalized cyclotomic sets D0∗ =
P d2 −1
P d2 −1
∗ =
D
and
D
2i
1
i=0
i=0 D2i+1 . As an application, we determine a lower bound
on the 2-adic complexity of modified Jacobi sequence. Our result shows that the
2-adic complexity of modified Jacobi sequence is at least pq − p − q − 1 with period
N = pq. This indicates that the 2-adic complexity of modified Jacobi sequence is
large enough to resist the attack of the rational approximation algorithm (RAA)
for feedback with carry shift registers (FCSRs).
Key words: Gaussian period, generalized cyclotomic class, modified Jacobi
sequence, 2-adic complexity.
1
The work is supported by Shandong Provincial Natural Science Foundation of China (No. ZR2014FQ005), Fundamental Research Funds for the Central Universities (No. 15CX02065A, No. 15CX05060A, No. 15CX08011A, No.
15CX02056A, No. 16CX02013A, No. 16CX02009A), Fujian Provincial Key Laboratory of Network Security and Cryptology Research Fund (Fujian Normal University) (No.15002),Qingdao application research on special independent innovation
plan project (No. 16-5-1-5-jch)
Preprint submitted to Elsevier
22 March 2018
1
Introduction
Pseudo-random sequences with good statistical property are widely used as
basic blocks for constructing stream ciphers. Any key stream generators could
be implemented by both linear feedback shift registers (LFSRs) and feedback
with carry shift registers (FCSRs). However, after the Berlekamp-Massey algorithm (BMA) for LFSRs [12] and the rational approximation algorithm for
FCSRs [13] were presented, linear complexity and 2-adic complexity of the
key stream sequence have been regarded as critical security criteria and both
are required to be no less than one half of the period.
Sequences from cyclotomic and generalized cyclotomic classes are large and
important sequence families for constructing codebooks [10], [9], Frequency
Hopping Sequences [22], Optical Orthogonal Codes [4], [5],[2] and cyclic codes
[6]-[7] and most of cyclotomic sequences and generalized cyclotomic sequences
have been proved to be with large linear complexity [8], [1], [21],[11]. However,
there are a handful research papers that focus on 2-adic complexities of these
sequences. In fact, although the concept of 2-adic has been presented for more
than two decades, there are only a few kinds of sequences whose 2-adic complexities have been completely determined. For example, in 1997, Klapper has
pointed out that an m-sequence with prime period has maximal 2-adic complexity [13]. In 2010, Tian and Qi showed that the 2-adic complexity of all the
binary m-sequences is maximal [16]. Afterwards, Xiong et al. [18] presented
a new method using circulant matrices to compute the 2-adic complexities
of binary sequences. They showed that all the known sequences with ideal
2-level autocorrelation have maximum 2-adic complexity. Moreover, they also
proved that the 2-adic complexities of Legendre sequences and Ding-HellesethLam sequences with optimal autocorrelation are also maximal. Then, using
the same method as that in [18], Xiong et al. [19] pointed out that two other
classes of sequences based on interleaved structure have also maximal 2-adic
complexity. One of these two classes of sequences was constructed by Tang
and Ding [14], which has optimal autocorrelation, the other was constructed
by Zhou et al [24], which is optimal with respect to the Tang-Fan-Matsufuji
bound [15].
Modified Jacobi sequence is one of sequence families constructed by Whiteman generalized cyclotomic classes. Green and Choi have proved that these
sequences have large linear complexity and low autocorrelation in many cases.
But, as far as the authors known, among Jacobi sequence family, there is no
other result about the 2-adic complexities of these sequences other than twinprime sequences, which has ideal autocorrelation and has been proved to be
with maximal 2-adic complexity by Xiong et al [18], and another class of sequences constructed from Whiteman generalized cyclotomic classes of order
2 has been proved to be also with maximal 2-adic complexity by Zeng et al.
2
[22].
In this paper, we study the 2-adic complexity of modified Jacobi sequences.
And we give a general lower bound on the 2-adic complexity, i.e., we will
prove that the 2-adic complexity of all these sequences is lower bounded by
pq − p − q − 1 with period N = pq. As a special case, we can also confirm that
the twin-prime sequence has maximal 2-adic complexity.
The rest of this paper is organized as follows. Some necessary definitions,
notations, and previous results are introduced in Section 2. Gauss periods
based on two Whiteman generalized cyclic sets are given in section 3. And the
lower bound on the 2-adic of modified Jacobi sequence is given in Section 4.
Finally we summarize our results and give some remarks in Section 5.
2
Preliminaries
N −1
Let N be a positive integer, {si }i=0
a binary sequence of period N, and
S(x) =
NP
−1
i=0
si xi ∈ Z[x]. If we write
NP
−1
si 2 i
S(2)
m
i=0
= N
= , 0 ≤ m ≤ n, gcd(m, n) = 1,
N
2 −1
2 −1
n
(1)
N −1
then the 2-adic complexity Φ2 (s) of the sequence {si }i=0
is defined as the
integer ⌊log2 n⌋, i.e.,
2N − 1
Φ2 (s) = log2
,
gcd(2N − 1, S(2))
$
%
(2)
where ⌊x⌋ is the greatest integer that is less than or equal to x.
Let p = df +1 and q = df ′ +1 (p < q) be two odd primes with gcd(p−1, q−1) =
d. Define N = pq, L = (p − 1)(q − 1)/d. The Chinese Remainder Theorem
guarantees that there exists a common primitive root g of both p and q.
Then the order of g modulo N is L. Let x be an integer satisfying x ≡ g
(mod p), x ≡ 1 (mod q). The existence and uniqueness of x (mod N) are
also guaranteed by the Chinese Remainder Theorem. Whiteman [17] presented
the definition of the following generalized cyclotomic classes
Di = {g txi : t = 0, 1, · · · , L − 1}, i = 0, 1, · · · , d − 1
3
(3)
of order d. And he also has proved
ZN∗ = ∪d−1
i=0 Di , Di ∩ Dj = ∅ for i 6= j,
where ∅ denotes the empty set. The corresponding generalized cyclotomic
numbers of order d are defined by
(i, j) = |(Di + 1) ∩ Dj |, for all i, j = 0, 1, · · · , d − 1.
And the following properties of the generalized cyclotomic numbers have also
been given by Whiteman:
(i, j) = (d − i, j − i),
(i, j) =
d−1
X
(i, j) =
j=0
(j
(4)
+ d2 , i + d2 ), if f f ′ is even,
if f f ′ is odd,
(j, i),
(p − 2)(q − 2) − 1
+ δi ,
d
(5)
(6)
where
δi =
If we denote
1,
0,
if f f ′ is even and i = d2 , or if f f ′ is odd and i = 0,
otherwise.
P = {p, 2p, · · · , (q − 1)p}, Q = {q, 2q, · · · , (p − 1)q}, R = {0},
d
d
−1
−1
2
2
C0 = ∪i=0
D2i ∪ Q ∪ R, C1 = ∪i=0
D2i+1 ∪ P,
d
d
−1
−1
2
2
D0∗ = ∪i=0
D2i , D1∗ = ∪i=0
D2i+1 ,
then we have ZN = C0 ∪ C1 , Z∗N = D0∗ ∪ D1∗ and it is easy to see that D0∗ is a
N −1
subgroup of Z∗N . The modified Jacobi sequence {si }i=0
is defined by
si =
0,
1,
if i
(mod N) ∈ C0 ,
if i
(mod N) ∈ C1 .
(7)
Moreover, it is not difficult to verify that modified Jacobi sequence is also
equivalent to the following definition
si =
1,
1−(
0,
if i (mod N) ∈ P,
i
p
)( )
2
i
q
, if i ∈ ZN∗ ,
otherwise,
4
where (·) is the Legendre symbol. But, in this paper, we will discuss its 2-adic
complexity using properties of generalized cyclotomic classes.
From the above argument, we have known that D0∗ is a subgroup of Z∗N ,
then D0∗ and D1∗ are also generalized cyclotomic classes. In this paper, we will
determine the Gauss periods based on D0∗ and D1∗ . Then, using these gauss
periods, a lower bound on the 2-adic of modified Jacobi sequence will be given.
To this end, we will first list the following properties of the above sets (The
proofs of these properties can also be found in many literatures, for example,
see [20]).
Lemma 1 For any a ∈ ZN and B ⊆ ZN , we denote aB = {ab|b ∈ B}. Then
we have the following properties.
(1) For each fixed a ∈ Di , we have aDj = D(i+j) (mod d) , aP = P and aQ = Q,
where i, j = 0, 1, · · · , d − 1.
(2) For each fixed a ∈ P , if b runs through each element of Di , i = 0, 1, · · · , d−1,
then ab runs exactly each element of P p−1
times. Symmetrically, for each
2
fixed a ∈ Q, if b runs through each element of Di , i = 0, 1, · · · , d − 1, then
ab runs exactly each element of Q q−1
times.
2
(3) For each fixed a ∈ P , we have aP = P , aQ = R. Symmetrically, for each
fixed a ∈ Q, we have aQ = Q, aP = R.
∗
(4) For each fixed a ∈ Di∗ , we have aDj∗ = D(i+j)
(mod 2) , aP = P and aQ = Q,
where i, j = 0, 1.
√
Let ωN = e2π −1/N be a Nth complex primitive root of unity. Then the additive character χ of ZN is given by
x
χ(x) = ωN
, x ∈ ZN
(8)
and Gaussian periods of order d are defined by
ηi =
X
χ(x), i = 0, 1, · · · , d − 1.
x∈Di
It is well-known that
X
χ(x) = −1,
(9)
X
χ(x) = −1,
(10)
ηi = 1.
(11)
x∈P
x∈Q
and
d−1
X
i=0
5
Moreover, the following results, which have also been proved by Whiteman
[17], will be useful.
Lemma 2 The element −1 ∈ Z∗N satisfies
−1 ≡
d
gδ x 2
g L2
(mod N), if f f ′ is even,
if f f ′ is odd,
(mod N),
where δ is some fixed integer such that 0 ≤ δ ≤ L − 1
Lemma 3 For each u ∈ P ∪ Q,
|Di ∩ (Dj + u)| =
3
(p−1)(q−1)
,
d2
if i 6= j,
(p−1)(q−1−d)
,
d2
if i = j and u ∈ P,
(p−1−d)(q−1)
,
d2
if i = j and u ∈ Q.
Gaussian periods of Whiteman generalized cyclotomic classes
P d −1
2
Let Ω0 = x∈D0∗ χ(x) = i=0
η2i and Ω1 = x∈D1∗ χ(x) =
section, we will determine the values of Ω0 and Ω1 .
P
P
P d2 −1
i=0
η2i+1 . In this
Theorem 1 Let p = df + 1, q = df ′ be distinct primes satisfying gcd(p −1, q −
1) = d and let Di∗ be the generalized cyclotomic set be defined in section 2 and
Ω0 the gauss period based on Di∗ , where i = 0, 1. Then the Gauss periods are
given by
Ω0 =
X
χ(x) =
x∈D0∗
Ω1 = 1 − Ω0 .
√
1± pq
,
2
if f f ′ is odd, or if
f f ′ is even and d ≡ 0
√
1± −pq
, if
2
(mod 4),
f f ′ is even and d ≡ 2
(12)
(mod 4),
(13)
Proof. Above all, from the definition of generalized cyclotomic class, for any
τ ∈ Dk , it can be easy verify that τ −1 ∈ D(d−k) (mod d) . Then, by Lemma 1,
for any 0 ≤ i, j, k ≤ d − 1, we have
| (Di + τ ) ∩ Dj | = | (τ −1 Di + 1) ∩ τ −1 Dj |
= ((i + d − k) (mod d), (j + d − k) (mod d)) .
6
If f f ′ is odd, then, by Lemma 2, we have −1 ∈ D0 . Therefore, using Lemma
3, we get
2
d
−1
2
X
(Ω0 )2 =
=
i=0
d
−1
2
X
i=0
=
η2i
d
−1
2
X
=
d
−1
2
X
d−1
X
i=0
X
+
y−x
ωN
k=0
=
d−1
X
k=0
=
d−1
X
k=0
η2i η2j
+
d
−1
2
d
−1
2
X
X
i=0 j6=i,j=0
X
X
x∈D2i y∈D2j
y−x
ωN
(p − 1)(q − 1) (p − 1)(q − 1 − d)
−
d
d2
!
d−1
X
(p − 1)(q − 1) (p − 1)(q − 1)
−
(2i − k, 2j − k)ηk −
d2
d2
k=0
X
i=0 j6=i,j=0
=
X
(2i − k, 2i − k)ηk +
d
−1
2
d
−1
2
X
k=0
d
−1
2
d−1
X
d
−1
2
i=0 j6=i,j=0
(p − 1 − d)(q − 1)
−
d2
X
+
i=0
x∈D2i y∈D2i
X
2
η2i
d
−1 d2 −1
2
X X
(2i − k, 2j
i=0 j=0
d
−1 d2 −1
2
d
−1 d2 −1
2
− k) ηk +
X X
(d − (2i − k), 2(j
i=0 j=0
X X
(d − (2(j
j ′ =0 j=0
p−1 q−1
+
2
2
− i)) ηk +
p−1 q−1
+
2
2
− j ′ ) − k), 2j ′ ) ηk +
p−1 q−1
+
,
2
2
where Eq. (14) comes from Eq.(4). Similarly, we can get
(Ω1 )
2
2
d
−1
2
X
=
i=0
=
d−1
X
k=0
η2i+1
d
−1 d2 −1
2
X X
j ′ =0 j=0
=
d
−1
2
X
2
η2i+1
+
i=0
d
−1
2
X
d
−1
2
X
η2i+1 η2j+1
i=0 j6=i,j=0
(d − (2(j − j ′ ) + 1 − k), 2j ′ ) ηk +
Then,
2
2
(Ω0 ) + (Ω1 ) =
d−1
X
k=0
d
−1
2
d
−1
2
X X
j ′ =0
((d − (2(j − j ′ ) − k), 2j ′ )
j=0
+ (d − (2(j − j ′ ) + 1 − k), 2j ′ )))) ηk
7
!
p−1 q−1
+
.
2
2
(14)
+ (p − 1) + (q − 1)
=
d−1
X
k=0
=
d
−1
2
d−1
X
X
(j, 2j ′ )
ηk
X
j ′ =0
=
d−1
X
d
−1
2
j ′ =0
d−1
X
j=0
(j, 2j ′ ) + (p − 1) + (q − 1)
(j, 0) +
d
−1
2
X
j ′ =1
j=0
=
+ (p − 1) + (q − 1)
j=0
d−1
X
(0, j) +
d
−1
2
X
j ′ =1
j=0
d−1
X
j=0
′
(15)
(j, 2j ) + (p − 1) + (q − 1)
d−1
X
j=0
(2j ′ , j) + (p − 1) + (q − 1)
d (p − 2)(q − 2) − 1
×
+ (p − 1) + (q − 1)
2
d
pq − 1
,
=1+
2
(16)
(17)
=1+
Where Eq. (15) is by Eq. (11), Eq. (16) is by Eq. (5), and Eq. (17) hold because
of Eq. (6). Moreover, from Eq.(11), we know that Ω0 + Ω1 = 1 and
(Ω0 )2 + (Ω1 )2 = (Ω0 + Ω1 )2 − 2Ω0 Ω1
= 1 − 2Ω0 Ω1
Then we obtain Ω0 Ω1 =
1−pq
,
4
which implies Ω0 =
√
1± pq
2
and Ω1 =
√
1∓ pq
.
2
Suppose that f f ′ is even then, by Lemma 2, we have −1 ∈ D d . Let d ≡ 2
2
(mod 4), i.e., d2 is odd. Then, for any 0 ≤ i, j ≤ d2 − 1 with i 6= j, we know
that 2i + d2 6≡ 2j (mod d) and 2i + 1 + d2 6≡ 2j + 1 (mod d). Therefore,
(Ω0 )2 =
X X
d
−1 d2 −1
2
i=0 j=0
=
d
−1 d2 −1
2
X X
i=0 j=0
=
d−1
X
k=0
=
d−1
X
k=0
X
X
x∈D2i+ d y∈D2j
2
d−1
X
y−x
ωN
(p − 1)(q − 1)
d
(2i + − k, 2j − k)ηk − 2 ×
2
d2
k=0
d
−1 d2 −1
2
d
−1 d2 −1
2
d
(p − 1)(q − 1)
− (2i − k), 2(j − i) + ) ηk −
(18)
2
2
2
X X d
(
i=0 j=0
X X
(2(j
i=0 j=0
!
− i), k − 2i) ηk −
8
(p − 1)(q − 1)
2
(19)
=
d−1
X
k=0
d
−1 d2 −1
2
X X
(2j ′ , k
j ′ =0 j=0
− 2(j − j ′ )) ηk −
(p − 1)(q − 1)
,
2
where Eq. (18) comes again from Eq. (4) and Eq. (19) comes from Eq. (5).
Similarly, we can get
(Ω1 )2 =
d−1
X
k=0
d
−1 d2 −1
2
X X
(2j ′ , k
j ′ =0 j=0
− 2(j − j ′ ) − 1) ηk −
(p − 1)(q − 1)
.
2
Then we get
(Ω0 )2 + (Ω1 )2 =
d−1
X
k=0
d
−1
2
X
X
((2j ′ , k
d
−1
2
j ′ =0
j=0
− (p − 1)(q − 1)
=
d−1
X
k=0
=
d
−1
2
X
j ′ =0
d
−1
2
d−1
X
X
j ′ =0
d−1
X
j=0
j=0
− 2(j − j ′ )) + (2j ′ , k − 2(j − j ′ ) − 1)) ηk
(2j ′ , j) ηk
− (p − 1)(q − 1)
(2j ′ , j) − (p − 1)(q − 1)
d (p − 2)(q − 2) − 1
×
− (p − 1)(q − 1)
2
d
pq + 1
=1−
,
2
=
where Eq. (20) is by Eq. (11) and Eq. (21) is by Eq. (6). Similar
argument
√
1± −pq
1+pq
and Ω1 =
to √the above, we can get Ω0 Ω1 = 4 , which implies Ω0 =
2
1∓ −pq
. For the case of even f f ′ and d ≡ 0 (mod 4), the result can be
2
similarly obtained.
✷
4
A lower bound on the 2-adic complexity of Jacobi sequence
N −1
Let {si }i=0
be a binary sequence with period N and let A = (ai,j )N ×N be the
matrix defined by ai,j = si−j (mod N ) . In this section, using the periods which
have been determined in Section 3, we will give a lower bound on the 2-adic
complexity of modified Jacobi sequence. In order to derive the lower bound,
the following two results will be useful, which can be found in [18] and [3]
respectively.
9
(20)
(21)
Lemma 4 [18] Viewing A as a matrix over the rational fields Q, if det(A) 6=
0, then
gcd S(2), 2N − 1 |gcd det(A), 2N − 1 .
Lemma 5 [3] det(A) =
QN −1
a=0
(22)
a
S(ωN
), where ωN is defined as in Eq. (8).
N −1
Lemma 6 Let {si }i=0
be the modified Jacobi sequence with period N = pq.
Then we have
a
S(ωN
)=
(p+1)(q−1)
,
2
−Ω0 ,
−Ω1 ,
− p+1 ,
2
q−1
,
2
if a ∈ R,
if a ∈ D0∗ ,
(23)
if a ∈ D1∗ ,
if a ∈ P,
if a ∈ Q.
Proof. Recall that
a k
a k
a
a k
)
S(ωN
) = Σk∈C1 (ωN
) = Σk∈D1∗ (ωN
) + Σk∈P (ωN
k
k
= Σk∈aD1∗ ωN
+ Σk∈aP ωN
.
a
Firstly, if a = 0, then it is easy to see S(ωN
) = (p−1)(q−1)
+ (q − 1) =
2
∗
Secondly, if a ∈ D0 , by Lemma 1 and Eq. (9), then we have
k
k
a
k
k
+ Σk∈P ωN
S(ωN
) = Σk∈aD1∗ ωN
+ Σk∈aP ωN
= Σk∈D1∗ ωN
= Ω1 − 1 = −Ω0 (for Ω0 + Ω1 = 1 by Eq.(11)).
Similarly, if a ∈ D1∗ , we have
a
k
k
k
k
S(ωN
) = Σk∈aD1∗ ωN
+ Σk∈aP ωN
= Σk∈D0∗ ωN
+ Σk∈P ωN
= Ω0 − 1 = −Ω1 .
If a ∈ P , again by Lemma 1, we have
a
k
k
S(ωN
) = Σk∈aC1 ωN
+ Σk∈aP ωN
=
=−
p+1
p−1
−1=−
.
2
2
p−1
k
k
Σk∈P ωN
+ Σk∈P ωN
2
Similarly, if a ∈ Q we have
10
(p+1)(q−1)
.
2
q−1
k
Σk∈Q ωN
+ (q − 1)
2
q−1
q−1
+ (q − 1) =
.
=−
2
2
a
k
k
S(ωN
) = Σk∈aC1 ωN
+ Σk∈aP ωN
=
The result follows.
✷
Lemma 7 Let p and q be two distinct odd primes and N = pq. Then we have
N −1
N −1
gcd(2p − 1, 22p −1
) = gcd(2p − 1, q) and gcd(2q − 1, 22q −1
) = gcd(2q − 1, p).
N −1
Particularly, if q > p, we have gcd(2q − 1, p) = 1, i.e., gcd(2q − 1, 22q −1
) = 1.
Proof. Note that
2N − 1 = 2pq − 1 = (2p − 1)(2p(q−1) + 2p(q−2) + · · · + 2p + 1)
= 2pq − 1 = (2q − 1)(2q(p−1) + 2q(p−2) + · · · + 2q + 1).
pq
pq
−1
−1
≡ q (mod 2p − 1) and 22q −1
≡ p (mod 2q − 1), which
Then we get 22p −1
N −1
N −1
imply gcd(2p − 1, 22p −1
) = gcd(2p − 1, q) and gcd(2q − 1, 22q −1
) = gcd(2q − 1, p).
q
q
Particularly, for q > p, if gcd(2 −1, p) > 1, i.e., p|2 −1, then the multiplicative
order of 2 modular p, denoted as Ordp (2), is a divisor of q. But q is a prime,
then Ordp (2) = q. By Fermat Theorem, we know that p|2p−1 − 1. Therefore,
we have q ≤ p − 1, which contradicts to the fact p < q. The desired result
follows.
✷
Theorem 2 Let p = df + 1 and q = df ′ + 1 be two odd primes satisfying
N −1
gcd(p − 1, q − 1) = d and p < q. Suppose {si }i=0
is the modified Jacobi
N −1
sequence with period N = pq. Then the 2-adic complexity φ2 (s) of {si }i=0
is
bounded by
φ2 (s) ≥ pq − p − q − 1.
(24)
N −1
Specially, if q = p + 2, then the 2-adic complexity of {si }i=0
is maximal.
Proof. By Lemmas 5 and 6, we can get
det(A) =
NY
−1
a
S(ωN
)
a=0
=
Y
a
S(ωN
)
a∈R
=2
p+1
2
Y
a
S(ωN
)
q−1
2
a
S(ωN
)
a∈D1∗
a∈D0∗
q
Y
p
(Ω0 Ω1 )
Y
a
S(ωN
)
a∈P
(p−1)(q−1)
2
Y
a
S(ωN
)
a∈Q
.
if f f ′ is odd or if f f ′ is even and
From Theorem 1, we know that Ω0 Ω1 = 1−pq
4
d ≡ 0 (mod 4), and Ω0 Ω1 = 1+pq
if f f ′ is even and d ≡ 2 (mod 4). Then
4
11
we have
det(A) =
(p−1)(q−1)
2
p+1 q q−1 p pq−1
2
, if f f ′ is odd,
2
2
4
or if f f ′ is even and d ≡ 0 (mod
p+1 q q−1 p pq+1 (p−1)(q−1)
2
, if f f ′ is
2 2
2
4
even and d ≡ 2 (mod 4).
4)
(25)
Now, let r be an any prime factor of 2N − 1 and Ordr (2) the multiplicative
order of 2 modular r. Then we get Ordr (2)|N. Note that N = pq. Therefore,
Ordr (2) = pq, p or q. Next, we will prove gcd(det(A), 2N −1) ≤ (2p −1)(2q −1).
To this end, we need discuss the following three cases.
Case 1. Ordr (2) = pq. By Fermat Theorem, we know that r|2r−1 − 1, which implies
Ordr (2) ≤ r−1. Since Ordr (2) = pq, then we have pq ≤ r−1, i.e., r ≥ pq+1.
But we know that p+1
< pq + 1, q−1
< pq + 1, and pq±1
< pq + 1, then
2
2
4
gcd(det(A), 2N − 1) = 1 < (2p − 1)(2q − 1).
Case 2. Ordr (2) = p. Similar argument to that in Case 1, we can get r ≥ p + 1,
which implies gcd( p+1
, 2N − 1) = 1. It is obvious that r|2p − 1. If r| q−1
or
2
2
pq±1
pq∓1
′
′
r| 4 (Here 4 corresponds to the cases of odd f f and even f f ), then we
N −1
)=
have gcd(r, q) = 1. Furthermore, by Lemma 7, we know gcd(2p − 1, 22p −1
N −1
2
gcd(2p − 1, q). Thus we have gcd(r, 2p −1 ) = 1.
Case 3. Ordr (2) = q. It is obvious that r|2q − 1. From Lemma 7, we know that
N −1
N −1
gcd(2q − 1, 22q −1
) = 1, which implies that gcd(r, 22q −1
) = 1.
Combining Case 2 and Case 3, no matter Ordr (2) = p or Ordr (2) = q, we
2N −1
always have gcd(det(A), 2N −1)|(2p −1)(2q −1) and gcd(det(A), (2p −1)(2
q −1) ) =
N
p
q
1, which implies that gcd(det(A), 2 − 1) ≤ (2 − 1)(2 − 1). From Lemma 4,
we have
gcd S(2), 2N − 1 ≤ gcd det(A), 2N − 1 ≤ (2p − 1)(2q − 1).
Then, by Eq. (2), we get
2N − 1
2pq − 1
φ2 (s) = log2
≥
log
≥ pq − p − q − 1.
2
gcd(2N − 1, S(2))
(2p − 1)(2q − 1)
$
%
$
%
Specially, if q = p + 2, then we know that f f ′ is even, which implies that
det(A) = 2
p+1 p(p+2)+1
.
2
From the above argument, we know that
p + 1 p(p+2)
gcd
,2
− 1 = 1.
2
The desire result follows.
✷
12
5
Summary
In this paper, we derive gauss periods of a class of generalized cyclotomic
sets from Whiteman generalized cyclotomic classes. As an application, a lower
bound on the 2-adic complexity of modified Jacobi sequences is determined.
Our result shows that the 2-adic complexity is at least pq − p − q − 1 with
period N = pq, which is obviously large enough to resist against RAA for
FCSR.
Acknowledgement
Parts of this work were written during a very pleasant visit of the first author to
the University of Carleton University in School of Mathematics and Statistics.
She wishes to thank the hosts for their hospitality.
References
[1] Bai, E., Liu, X., Xiao, G.: Linear complexity of new generalized cyclotomic
sequences of order two of length pq. IEEE Trans. Inform. Theory 51, 1849-1853
(2005).
[2] Cai, H., Liang, H., Tang, X.: Constructions of optimal 2-D optical orthogonal
codes via generalized cyclotomic classes. IEEE Trans. Inform. Theory 61, 688-695
(2015).
[3] Davis, P J.: Circulant Matrices. New York, NY, USA: Chelsea, 1994.
[4] Ding, C., Xing, C.: Several classes of (2m − 1, w, 2) optical orthogonal codes.
Discrete Applied Mathematics 128, 103-120 (2003).
[5] Ding, C., Xing, C.: Cyclotomic optical orthogonal codes of composite lengths.
IEEE Trans. Inform. Theory 52, 263-268 (2004).
[6] Ding, C.: Cyclotomic constructions of cyclic codes with length being the product
of two primes. IEEE Trans. Inform. Theory 58, 2231-2236 (2012).
[7] Ding, C.: Cyclic codes from the two-prime sequences. IEEE Trans. Inform.
Theory 58, 3881-3891 (2012).
[8] Ding, C., Helleseth, T.: On the linear complexity of Legendre sequences. IEEE
Trans. Inform. Theory 44, 1693-1698 (1998).
[9] Fan, C., Ge, G.: A unified approach to Whiteman’s and Ding-Helleseth’s
generalized cyclotomy over residue classs rings. IEEE Trans. Information Theory
60, 1326-1336 (2014).
13
[10] Hu, L., Yue, Q.: Gauss periods and codebooks from generalized cyclotomic sets
of order four. Design, Codes and Cryptography 69, 233-246 (2013).
[11] Li, X., Ma, W., Yan, T., Zhao, X.: Linear complexity of a new generalized
cyclotomic sequence of order two of length pq. IEICE Transactions 96-A, 10011005 (2013).
[12] Massey,J. L.: Shift-register synthesis and BCH decoding. IEEE Trans. Inform.
Theory 15, 122-127 (1969).
[13] Klapper, A., Goresky, M.: Feedback shift registers, 2-adic span, and combiners
with memory. Journal of Cryptology 10, 111-147 (1997).
[14] Tang, X., Ding, C.: New classes of balanced quaternary and almost balanced
binary sequences with optimal autocorrelation Value. IEEE Trans. Inform. Theory
56, 6398-6405 (2010).
[15] Tang, X., Fan, P., Matsufuji,S.: Lower bounds on the maximum correlation of
sequences with low or zero correlation zone. Electron. Lett. 36, 551-552 (2000).
[16] Tian, T., Qi, W.: 2-Adic complexity of binary m-sequences. IEEE Trans. Inform.
Theory 56, 450-454 (2010).
[17] Whiteman, A L.: A family of difference sets. lllinois Journal of Mathematics 6,
107-121 (1962).
[18] Xiong, H., Qu, L., Li, C.: A new method to compute the 2-adic complexity of
binary sequences. IEEE Trans. Inform. Theory 60, 2399-2406 (2014).
[19] Xiong, H., Qu, L., Li, C.: 2-Adic complexity of binary sequences with interleaved
structure. Finite Fields and Their Applications 33, 14-28 (2015).
[20] Yan, T.: Study on Constructions and Properties of Pseudo-Random Sequence.
Ph. D Thesis, 2007.
[21] Yan, T., Du, X., Xiao, G., Huang, X.: Linear complexity of binary Whiteman
generalized cyclotomic sequences of order 2k . Information Sciences 179, 1019-1023
(2009).
[22] Zeng, X.,Cai, H., Tang, X., Yang, Y.: Optimal frequency sequences of odd
length. IEEE Trans. Inform. Theory 59, 3237-3248 (2013).
[23] Xiao, Z.,Zeng, X.: 2-Adic complexity of two classes of generalized cyclotomic
binary sequences. International Journal of Foundations of Computer Science 27,
879-893 (2016).
[24] Zhou, Z., Tang, X., Gong, G.: A new classes of sequences with zero or low
correlation zone based on interleaving technique. IEEE Trans.Inform. Theory 54,
4267-4273 (2008).
14
| 7cs.IT
|
arXiv:1610.08994v2 [math.GR] 30 Jan 2017
ON SELF-SIMILARITY OF WREATH PRODUCTS OF ABELIAN
GROUPS
ALEX C. DANTAS AND SAID N. SIDKI
Abstract. We prove that in a self-similar wreath product of abelian groups
G = BwrX, if X is torsion-free then B is torsion of finite exponent. Therefore,
in particular, the group ZwrZ cannot be self-similar Furthemore, we prove that
if L is a self-similar abelian group then Lω wrC2 is also self-similar.
1. Introduction
A group G is self-similar provided for some finite positive integer m, the group
has a faithful representation on an infinite regular one-rooted m-tree Tm such that
the representation is state-closed and is transitive on the tree’s first level. If a group
G does not admit such a representation for any m then we say G is not self-similar.
In determining that a group is not self-similar we will use the language of virtual
endomorphisms of groups. More precisely, a group G is not self-similar if and only
if for any subgroup H of G of finite index and any homomorphism f : H → G there
exists a non-trivial subgroup K of H which is normal in G and is f -invariant (in
the sense K f ≤ K).
Which groups admit faithful self-similar representation is an on going topic of
investigation. The first in depth study of this question was undertaken in [2] and
then in book form in [3]. Faithful self-similar representations are known for many
individual finitely generated groups ranging from the torsion groups of Grigorchuk
and Gupta-Sidki to free groups. Such representations have been also studied for
the family of abelian groups [4], of finitely generated nilpotent groups [7], as well
as for arithmetic groups [9]. See [8] for further references.
One class which has received attention in recent years is that of wreath products
of abelian groups G = BwrX, such as the classical lamplighter group [1] in which
B is cyclic of order 2 and X is infinite cyclic. The more general class G = Cp wrX
where Cp is cyclic of prime order p and X is free abelian of rank d ≥ 1 was the
subject of [8] where self-similar groups of this type are constructed for every finite
rank d.
We show in this paper that the properly of B being torsion is necessary to
guarantee self-similarity of G. More precisely, we prove
Theorem 1. Let G = BwrX be a self-similar wreath product of abelian groups. If
X is torsion free then B is a torsion group of finite exponent. In particular, ZwrZ
cannot be self-similar.
Date: January 18, 2017.
2000 Mathematics Subject Classification. Primary 20E08, 20F18.
Key words and phrases. Automorphisms of trees, state-closed groups, self-similar groups.
We thank the referee for suggestions which improved the original text.
1
2
ALEX C. DANTAS AND SAID N. SIDKI
Observe that though G = ZwrZ is not self-similar, it has a faithful finite-state
representation on the binary tree [5].
Next we produce a novel embedding of self-similar abelian groups into self-similar
wreath products having higher cardinality.
Theorem 2. Let L be a self-similar abelian group and Lω an infinite countable
direct sum of copies of L. Then Lω wrC2 is also self-similar.
2. Preliminaries
We recall a number of notions of groups acting on trees and of virtual endomorphisms of groups from [4].
2.1. State-closed groups. 1. Automorphisms of one-rooted regular trees T (Y )
indexed by finite sequences from a finite set Y of size m ≥ 2, have a natural
interpretation as automata on the alphabet Y , and with states which are again
automorphisms of the tree. A subgroup G of the group of automorphisms A (Y ) of
the tree is said to have degree m. Moreover, G is state-closed of degree m provided
the states of its elements are themselves elements of the same group.
2. Given an automorphism group G of the tree, v a vertex of the tree and l a
level of the tree, we let F ixG (v) denote the subgroup of G formed by its elements
which fix v and let StabG (l) denote the subgroup of G formed by elements which
fix all v of level l. Also, let P (G) denote the permutation group induced by G on
the first level of the tree. We say G is transitive provided P (G) is transitive.
3. A group G is said to be self-similar provided it is a state-closed and transitive
subgroup of A (Y ) for some finite set Y .
2.2. Virtual endomorphisms. 1. Let G be a group with a subgroup H of finite
index m. A homomorphism f : H → G is called a virtual endomorphism of G and
(G, H, f ) is called a similarity triple; if G is fixed then (H, f ) is called a similarity
pair.
2. Let G be a transitive state-closed subgroup of A (Y ) where Y = {1, 2, ..., m}.
Then the index [G : F ixG (1)] = m and the projection on the 1st coordinate of
F ixG (1) produces a subgroup of G; that is, π 1 : F ixG (1) → G is a virtual endomorphism of G.
3. Let G be a group with a subgroup H of finite index m and a homomorphism
f : H → G. If U ≤ H and U f ≤ U then U is called f -invariant. The largest
subgroup K of H, which is normal in G and is f -invariant is called the f -core(H).
If the f -core(H) is trivial then f and the triple (G, H, f ) are called simple.
4. Given a triple (G, H, f ) and a right transversal L = {x1 , x2 , ..., xm } of H in
G, the permutational representation π : G → P erm (1, 2, ..., m) is g π : i → j which
is induced from the right multiplication Hxi g = Hxj . Generalizing the KalujnineKrasner procedure [6] , we produce recursively a representation ϕ : G → A (Y ),
defined by
−1 f ϕ
ϕ
g =
xi g. x(i)gπ
gπ ,
1≤i≤m
seen as an element of an infinitely iterated wreath product of P erm (1, 2, ..., m).
The kernel of ϕ is precisely the f -core(H) and Gϕ is state-closed and transitive and
H ϕ = F ixGϕ (1).
SELF-SIMILAR WREATH PRODUCTS
3
Lemma 1. A group G is self-similar if and only if there exists a simple similarity
pair (H, f ) for G.
3. Proof of Theorem 1
We recall B, X are abelian groups, X is a torsion-free group and G = BwrX .
Denote the normal closure of B in G by A = B G . Let (H, f ) be the similarity pair
with respect to which G is self-similar and let [G : H] = m. Define
A0 = A ∩ H, L = (A0 )f ∩ A, Y = X ∩ (AH).
Note that if x ∈ X is nontrivial then the centralizer CA (x) is trivial. We develop
the proof in four lemmas.
f
Lemma 2. Either B m is trivial or (A0 ) ≤ A. In both cases A 6= A0 .
Proof. We have Am ≤ A0 and X m ≤ H. As A is normal abelian and X is abelian,
[Am , X m ] ⊳ G,
[Am , X m ] ≤
[A0 , X m ] ≤ A0 .
Also,
f
:
h
i
f
f
f
[Am , X m ] → (Am ) , (X m ) ≤ (A0 ) ∩ G′
≤
(A0 ) ∩ A = L.
f
(1) If L is trivial then [Am , X m ] ≤ ker (f ). Since f is simple, it follows that
ker (f ) = 1 and [Am , X m ] = 1 = [B m , X m ]. As X m 6= 1, we conclude Am = 1 =
B m . (2) If L is nontrivial then L is central in M = A (A0 )f = A (X ∩ M ) which
f
implies X ∩ M = 1 and (A0 ) ≤ A. (3) If B is a torsion group then tor(G) = A;
f
clearly, (A0 ) ≤ A and A 6= A0 .
Let G be a counterexample; that is, B has infinite exponent. By the previous
f
lemma (A0 ) ≤ A and so we may use Proposition
1 of [8] to replace the simple
˙
similarity pair (H, f ) by a simple pair Ḣ, f where Ḣ = A0 Y (Y ≤ X) and
f˙
f
(Y ) ≤ X. In other words, we may assume (Y ) ≤ X.
Lemma 3. If z ∈ X is nontrivial and x1 , ..., xt , z1 , ..., zl ∈ X, then there exists an
integer k such that
z k {z1 , ..., zl } ∩ {x1 , ..., xt } = ∅.
Proof. Note that the set {k ∈ Z | {z k zj } ∩ {x1 , ..., xt } =
6 ∅} is finite, for each j =
1, ..., l. Indeed, otherwise there exist k1 6= k2 such that z k1 −k2 = 1, a contradiction.
f
Lemma 4. If x ∈ X is nontrivial, then (xm ) is nontrivial.
Proof. Suppose that there exists a nontrivial x ∈ X such that xm ∈ ker (f ). Then
for each a ∈ A and each u ∈ X we have
(xm )f
m
m
(a−mu amux )f = (a−mu )f (amux )f = (a−mu )f (amu )f
= (a−mu )f (amu )f = 1,.
m(xm −1)
Since A
≤ ker (f ) and is normal in G, we have a contradiction.
4
ALEX C. DANTAS AND SAID N. SIDKI
Lemma 5. The subgroup Am is f -invariant.
Proof. Let a ∈ A. Consider T = {c1 , ..., cr } a transversal of A0 in A, where r
is a divisor of m. Since Am is a subgroup of A0 and A = ⊕x∈X B x , there exist
x1 , ..., xt such that
D
E
f
x1
⊕ ... ⊕ B xt
(cm
i ) | i = 1, ..., r ≤ B
and z1 , ..., zl ∈ X such that
(am )f ≤ B z1 ⊕ ... ⊕ B zl ..
Since [G : H] = m, it follows that X m ≤ Y . Fix a nontrivial x ∈ X and let
z = (xm )f .
For each integer k, define ik ∈ {1, ..., r} such that
ax
Then
But (ax
mk
mk
c−1
ik ∈ A0 .
f mk
m
mk
−1 f
x
m
(ax c−1
=
(a
c
∈ Am .
)
)
ik
ik
mk
m
mx
c−1
c−m
, thus
ik ) = a
ik
f
mk
zk −mf
f
mk f
m
= amf
cik .
c−m
(ax c−1
= amx
ik
ik )
By Lemma 4, z 6= 1. There exists by Lemma 3 an integer k ′ such that
′
′
{z k z1 , ..., z k zl } ∩ {x1 , ..., xt } = ∅,
and so,
(B z
k′
z1
⊕ ... ⊕ B z
k′
zl
) ∩ (B x1 ⊕ ... ⊕ B xt ) = 1.
It follows that
k′
(amf )z c−mf
∈ Am ∩ [(B z
ik
k′
z1
⊕ ... ⊕ B z
k′
zl
) ⊕ (B x1 ⊕ ... ⊕ B xt )].
But as
Am = ⊕x∈X B mx .,
we conclude, (amf )z
f
(Am ) ≤ Am .
k′
∈ B mz
k′
z1
⊕ ... ⊕ B mz
k′
zl
≤ Am and amf ∈ Am . Hence,
With this last lemma, the proof of Theorem 2 is finished.
4. Proof of Theorem 2
Let L be a self-similar abelian group with
P respect to a simple triple (L, M, φ);
then φ is a monomorphism. Define B = i≥1 Li , a direct sum of groups where
Li = L for each i . Let X be cyclic group of order 2 and G = BwrX, the wreath
product of B by X. Denote the normal closure of B in G by A; then,
X
A = B X = L 1 ⊕
Li × B
i≥2
G =
A • X.
SELF-SIMILAR WREATH PRODUCTS
5
Define the subgroup of G
H = M ⊕
X
i≥2
Li × B ;
an element of H has the form
β = (β 1 , β 2 )
where
=
βi
β ij
=
β1
β 1j
We note that [G : H] is finite; indeed,
j≥1
j≥1
, β ij ∈ L
, β 11 ∈ M .
[A : H] = [L : M ] and [G : H] = 2[L : M ].
Define the maps
X
φ′1 : M ⊕
Li → B, φ′2 : B → B
i≥2
where for β = (β 1 , β 2 ) = (β 1j ), (β 2j ) j≥1 , β 11 ∈ M,
φ′1 : β 1 7→ (β φ11 β 12 , β 13 , ...), φ′2 : β 2 7→ (β 22 , β 21 , β 23 , ...).
Since L is abelian, φ′1 is a homomorphism and clearly φ′2 is a homomorphism as
well.
Define the homomorphism
X
f : M ⊕
Li × B → A
i≥2
by
′
φ′
f : (β 1 , β 2 ) 7→ (β 1 ) 1 , (β 2 )φ2 .
Suppose by contradiction that K is a nontrivial subgroup of H, normal in G and
f -invariant and let κ = (κ1 , κ2 ) be a nontrivial element of K. Since X permutes
transitively the indices of κi , we conclude κi1 ∈ M for i = 1, 2. Let si (call it
degree) be the maximum index of the nontrivial entries of κi ; if κi = 0 then write
si = 0 . Choose κ with minimum s1 + s2 ; we may assume s1 be minimum among
those si 6= 0. Since
κ1
=
(κ1j )j≥1 , κ11 ∈ M,
φ′1
=
(κφ11 κ12 , κ13 , ...),
(κ1 )
φ′1
we conclude (κ1 )
has smaller degree and therefore
κ1 = (κ11 , e, e, e, ...) or κ11 , κ−φ
11 , e, e, ... .
6
ALEX C. DANTAS AND SAID N. SIDKI
Suppose κ1 = (κ11 , e, e, e, ...). As, κ = (κ1 , κ2 ) ∈ K, we have κ11 ∈ M and therefore
′
′
κf =
(κ1 )φ1 , (κ2 )φ2 ∈ K,
φ′
φ
(κ1 )
=
(κ11 ) , e, e, e, ... ,
φ
∈
(κ11 )
κf
2
2
φ′
(κ1 )( 1 )
φ2
(κ11 )
M;
2
φ′1 )
(
, κ2 ,
=
(κ1 )
φ2
=
(κ11 ) , e, e, e, ... ,
∈
M ; etc.
By simplicity of φ, this alternative is out. That is,
κ1 = κ11 , κ−φ
,
e,
e,
...
, κ11 ∈ M .
11
Therefore
κ
=
(κ1 , κ2 ) ,
κx
=
κxf x
=
φ′
(κ2 , κ1 ) , κxf = (κ2 ) 1 , κ−φ
,
κ
,
e,
e,
...
,
11
11
′
φ1
κ−φ
11 , κ11 , e, e, ... , (κ2 )
are elements of K and so, κφ11 ∈ M . Furthermore,
κxf xf
2
κ−φ
11 κ11
2
φ′1 φ′2
,
κ−φ
11 κ11 , e, e, ... , (κ2 )
=
∈
M;
E
D i
i
successive applications of f to κxf x produces κφ11 ∈ M . Therefore, κφ11 | i ≥ 0 is
a φ-invariant subgroup of M ; a contradiction is reached.
References
[1] Grigorchuk, R., Zuk, A., The Lamplighter group as a group generated by a 2-state automaton
and its spectrum. Geometriae Dedicata 87, (2001), 209-244.
[2] Nekrachevych,V., Sidki, S., Automorphisms of the binary tree: state-closed subgroups and
dynamics of 1/2 endomorphisms, In: Groups: London Mathematical Lecture Notes Series,
Vol. 311, (2004), 375-404. Topological, Combinatorial and Arithmetic Aspects, Muller, T. W.,
(Ed.)
[3] Nekrashevych, V., Self-similar groups, Math. Surveys and Monographs, 117, American Mathematical Society, Providence, RI, 2005.
[4] Brunner, A., Sidki, S., Abelian state-closed subgroups of automorphisms of m-ary trees,
Groups, Geometry and Dynamics, vol.3, (2010) 455-472.
[5] Sidki, S., Tree wreathing applied to the generation of groups by finite automata, International
Journal of Algebra and Computation, vol. 15. (2005), 1-8.
[6] Robinson, D. J. S., A course in the theory of groups, 2nd ed., Grad. Texts in Math. 80, Springer
Verlag, New York 1996.
[7] Berlatto, A., Sidki, S., Virtual endomorphisms of nilpotent groups, Groups, Geometry and
Dynamics, vol.1, (2007), 21-46.
[8] Dantas, A., Sidki, S., On state-closed representations of restricted wreath product of groups
of type Gp,d = Cp wrC d , arXiv:1505.05165, (2015).
SELF-SIMILAR WREATH PRODUCTS
7
[9] Kapovich, M., Arithmetic aspects of self-similar groups, Groups Geometry and Dynamics 6
(2012), 737–754.
Universidade Tecnológica Federal do Paraná, Guarapuava-PR 85053-525, Brasil
E-mail address: [email protected]
Departamento de Matematica, Universidade de Brası́lia, Brası́lia DF 70910-900, Brazil
E-mail address: [email protected]
| 4math.GR
|
Weak Memory Models: Balancing Definitional Simplicity and Implementation
Flexibility
arXiv:1707.05923v1 [cs.PL] 19 Jul 2017
Sizhuo Zhang, Muralidaran Vijayaraghavan, Arvind
Computer Science and Artificial Intelligence Laboratory
Massachusetts Institute of Technology
{szzhang, vmurali, arvind}@csail.mit.edu
Abstract—The memory model for RISC-V, a newly developed
open source ISA, has not been finalized yet and thus, offers an
opportunity to evaluate existing memory models. We believe
RISC-V should not adopt the memory models of POWER or
ARM, because their axiomatic and operational definitions are
too complicated. We propose two new weak memory models:
WMM and WMM-S, which balance definitional simplicity and
implementation flexibility differently. Both allow all instruction
reorderings except overtaking of loads by a store. We show
that this restriction has little impact on performance and it
considerably simplifies operational definitions. It also rules
out the out-of-thin-air problem that plagues many definitions.
WMM is simple (it is similar to the Alpha memory model),
but it disallows behaviors arising due to shared store buffers
and shared write-through caches (which are seen in POWER
processors). WMM-S, on the other hand, is more complex and
allows these behaviors. We give the operational definitions of
both models using Instantaneous Instruction Execution (I2 E),
which has been used in the definitions of SC and TSO. We also
show how both models can be implemented using conventional
cache-coherent memory systems and out-of-order processors,
and encompasses the behaviors of most known optimizations.
commercial processors and then constructing models to fit
these observations [7]–[10], [12]–[16].
The newly designed open-source RISC-V ISA [17] offers a
unique opportunity to reverse this trend by giving a clear definition with understandable implications for implementations.
The RISC-V ISA manual only states that its memory model
is weak in the sense that it allows a variety of instruction
reorderings [18]. However, so far no detailed definition has
been provided, and the memory model is not fixed yet.
In this paper we propose two weak memory models for
RISC-V: WMM and WMM-S, which balance definitional
simplicity and implementation flexibility differently. The difference between the two models is regarding store atomicity,
which is often classified into the following three types [19]:
•
•
Keywords-weak memory model
•
I. I NTRODUCTION
A memory model for an ISA is the specification of all
legal multithreaded program behaviors. If microarchitectural
changes conform to the memory model, software remains
compatible. Leaving the meanings of corner cases to be
implementation dependent makes the task of proving the
correctness of multithreaded programs, microarchitectures
and cache protocols untenable. While strong memory models
like SC and SPARC/Intel-TSO are well understood, weak
memory models of commercial ISAs like ARM and POWER
are driven too much by microarchitectural details, and
inadequately documented by manufacturers. For example, the
memory model in the POWER ISA manual [11] is “defined”
as reorderings of events, and an event refers to performing
an instruction with respect to a processor. While reorderings
capture some properties of memory models, it does not
specify the result of each load, which is the most important
information to understand program behaviors. This forces
the researchers to formalize these weak memory models by
empirically determining the allowed/disallowed behaviors of
Single-copy atomic: a store becomes visible to all processors at the same time, e.g., in SC.
Multi-copy atomic: a store becomes visible to the issuing
processor before it is advertised simultaneously to all other
processors, e.g., in TSO and Alpha [5].
Non-atomic (or non-multi-copy-atomic): a store becomes
visible to different processors at different times, e.g., in
POWER and ARM.
Multi-copy atomic stores are caused by the store buffer
or write-through cache that is private to each processor.
Non-atomic stores arise (mostly) because of the sharing
of a store buffer or a write-through cache by multiple
processors, and such stores considerably complicate the
formal definitions [7], [8]. WMM is an Alpha-like memory
model which permits only multi-copy atomic stores and
thus, prohibits shared store buffers or shared write-through
caches in implementations. WMM-S is an ARM/POWERlike memory model which admits non-atomic stores. We will
present the implementations of both models using out-oforder (OOO) processors and cache-coherent memory systems.
In particular, WMM and WMM-S allow the OOO processors
in multicore settings to use all speculative techniques which
are valid for uniprocessors, including even the load-value
speculation [20]–[24], without additional checks or logic.
We give operational definitions of both WMM and WMMS. An operational definition specifies an abstract machine,
and the legal behaviors of a program under the memory
Operational
model
SC
TSO
RMO
Alpha
RC
ARM and
POWER
WMM
WMM-S
Simple; I2 E [1]
Simple; I2 E [2]
Doesn’t exist
Doesn’t exist
Doesn’t exist
Complex; non
I2 E [7], [8]
Simple; I2 E
Medium; I2 E
Definition
Axiomatic model
Simple [1]
Simple [3]
Simple; needs fix [4]
Medium [5]
Medium [6]
Complex [9], [10]
Single-copy atomic
Multi-copy atomic
Multi-copy atomic
Multi-copy atomic
Unclear
Non-atomic
Model properties / Implementation flexibility
Allow shared write- Instruction reorderings
through cache/shared
store buffer
No
None
No
Only St-Ld reordering
No
All four
No
All four
No
All four
Yes
All four
Simple
Doesn’t exist
Multi-copy atomic
Non-atomic
No
Yes
Store atomicity
Figure 1.
All except Ld-St reordering
All except Ld-St reordering
Ordering
of
data-dependent
loads
Yes
Yes
Yes
No
Yes
Yes
No
No
Summary of different memory models
model are those that can result by running the program
on the abstract machine. We observe a growing interest
in operational definitions: memory models of x86, ARM
and POWER have all been formalized operationally [2], [7],
[8], [15], [25], and researchers are even seeking operational
definitions for high-level languages like C++ [26]. This is
perhaps because all possible program results can be derived
from operational definitions mechanically while axiomatic
definitions require guessing the whole program execution
at the beginning. For complex programs with dependencies,
loops and conditional branches, guessing the whole execution
may become prohibitive.
Unfortunately, the operational models of ARM and
POWER are too complicated because their abstract machines
involve microarchitectural details like reorder buffers (ROBs),
partial and speculative instruction execution, instruction
replay on speculation failure, etc. The operational definitions
of WMM and WMM-S are much simpler because they are
described in terms of Instantaneous Instruction Execution
(I2 E), which is the style used in the operational definitions of
SC [1] and TSO [2], [25]. An I2 E abstract machine consists
of n atomic processors and an n-ported atomic memory.
The atomic processor executes instructions instantaneously
and in order, so it always has the up-to-date architectural
(register) state. The atomic memory executes loads and
stores instantaneously. Instruction reorderings and store
atomicity/non-atomicity are captured by including different
types of buffers between the processors and the atomic
memory, like the store buffer in the definition of TSO. In
the background, data moves between these buffers and the
memory asynchronously, e.g., to drain a value from a store
buffer to the memory.
I2 E definitions free programmers from reasoning partially
executed instructions, which is unavoidable for ARM and
POWER operational definitions. One key tradeoff to achieve
I2 E is to forbid a store to overtake a load, i.e., disallow Ld-St
reordering. Allowing such reordering requires each processor
in the abstract machine to maintain multiple unexecuted
instructions in order to see the effects of future stores, and
the abstract machine has to contain the complicated ROBlike structures. Ld-St reordering also complicates axiomatic
definitions because it creates the possibility of “out-ofthin-air” behaviors [27], which are impossible in any real
implementation and consequently must be disallowed. We
also offer evidence, based on simulation experiments, that
disallowing Ld-St reordering has no discernible impact on
performance.
For a quick comparison, we summarize the properties of
common memory models in Figure 1. SC and TSO have
simple definitions but forbid Ld-Ld and St-St reorderings,
and consequently, are not candidates for RISC-V. WMM is
similar to RMO and Alpha but neither has an operational
definition. Also WMM has a simple axiomatic definition,
while Alpha requires a complicated axiom to forbid outof-thin-air behaviors (see Section V-B), and RMO has
an incorrect axiom about data-dependency ordering (see
Section X).
ARM, POWER, and WMM-S are similar models in the
sense that they all admit non-atomic stores. While the
operational models of ARM and POWER are complicated,
WMM-S has a simpler I2 E definition and allows competitive
implementations (see Section IX-B). The axiomatic models
of ARM and POWER are also complicated: four relations in
the POWER axiomatic model [10, Section 6] are defined in
a fixed point manner, i.e., their definitions mutually depend
on each other.
Release Consistency (RC) are often mixed with the concept
of “SC for data-race-free (DRF) programs” [28]. It should be
noted that “SC for DRF” is inadequate for an ISA memory
model, which must specify behaviors of all programs. The
original RC definition [6] attempts to specify all program
behaviors, and are more complex and subtle than the “SC for
DRF” concept. We show in Section X that the RC definition
fails a litmus test for non-atomic stores and forbids shared
write-through caches in implementation.
This paper makes the following contributions:
1) WMM, the first weak memory model that is defined
in I2 E and allows Ld-Ld reordering, and its axiomatic
definition;
2) WMM-S, an extension on WMM that admits non-atomic
stores and has an I2 E definition;
3) WMM and WMM-S implementations based on OOO pro-
cessors that admit all uniprocessor speculative techniques
(such as load-value prediction) without additional checks;
4) Introduction of invalidation buffers in the I2 E definitional
framework to model Ld-Ld and other reorderings.
Paper organization: Section II presents the related work.
Section III gives litmus tests for distinguishing memory
models. Section IV introduces I2 E. Section V defines
WMM. Section VI shows the WMM implementation using
OOO processors. Section VII evaluates the performance of
WMM and the influence of forbidding Ld-St reordering.
Section VIII defines WMM-S. Section IX presents the
WMM-S implementations with non-atomic stores. Section X
shows the problems of RC and RMO. Section XI offers the
conclusion.
dependency between I2 and I3 in Figure 2a. Thus the
resulting behaviors can arise only because of different store
atomicity properties. We use FENCELL for memory models
that can reorder data-dependent loads, e.g., I5 in Figure 2b
would be the MB fence for Alpha. For other memory models
that order data-dependent loads (e.g., ARM), FENCELL could
be replaced by a data dependency (like the data dependency
between I2 and I3 in Figure 2a). The Ld-Ld fences only
stop Ld-Ld reordering; they do not affect store atomicity in
these tests.
II. R ELATED W ORK
(a) SBE: test for multi-copy atomic stores
SC [1] is the simplest model, but naive implementations of
SC suffer from poor performance. Although researchers have
proposed aggressive techniques to preserve SC [29]–[38],
they are rarely adopted in commercial processors perhaps due
to their hardware complexity. Instead the manufactures and
researchers have chosen to present weaker memory models,
e.g., TSO [2], [3], [25], [39], PSO [4], RMO [4], Alpha [5],
Processor Consistency [40], Weak Consistency [41], RC [6],
CRF [42], Instruction Reordering + Store Atomicity [43],
POWER [11] and ARM [44]. The tutorials by Adve et al.
[45] and by Maranget et al. [46] provide relationships among
some of these models.
A large amount of research has also been devoted to
specifying the memory models of high-level languages:
C++ [26], [47]–[50], Java [51]–[53], etc. We will provide
compilation schemes from C++ to WMM and WMM-S.
Recently, Lustig et al. have used Memory Ordering
Specification Tables (MOSTs) to describe memory models,
and proposed a hardware scheme to dynamically convert
programs across memory models described in MOSTs
[19]. MOST specifies the ordering strength (e.g., locally
ordered, multi-copy atomic) of two instructions from the same
processor under different conditions (e.g., data dependency,
control dependency). Our work is orthogonal in that we
propose new memory models with operational definitions.
III. M EMORY M ODEL L ITMUS T ESTS
Here we offer two sets of litmus tests to highlight
the differences between memory models regarding store
atomicity and instruction reorderings, including enforcement
of dependency-ordering. All memory locations are initialized
to 0.
A. Store Atomicity Litmus Tests
Figure 2 shows four litmus tests to distinguish between
these three types of stores. We have deliberately added
data dependencies and Ld-Ld fences (FENCELL ) to these
litmus tests to prevent instruction reordering, e.g., the data
Proc. P1
I1 : St a 1
I2 : r1 = Ld a
I3 : r2 = Ld (b+r1 −1)
SC forbids but TSO allows:
Proc. P1
I1 : St a 2
Proc. P2
I4 : St b 1
I5 : r3 = Ld b
I6 : r4 = Ld (a+r3 −1)
r1 = 1, r2 = 0, r3 = 1, r4 = 0
Proc. P2
I2 : r1 = Ld a
I3 : St b (r1 − 1)
Proc. P3
I4 : r2 = Ld b
I5 : FENCELL
I6 : r3 = Ld a
TSO, RMO and Alpha forbid, but RC, ARM
and POWER allow: r1 = 2, r2 = 1, r3 = 0
(b) WRC: test for non-atomic stores [7]
Proc. P1
I1 : St a 2
Proc. P2
Proc. P3
I2 : r1 = Ld a
I4 : r2 = Ld b
I3 : St b (r1 − 1)
I5 : St a r2
TSO, RMO, Alpha and RC forbid, but ARM and
POWER allow: r1 = 2, r2 = 1, m[a] = 2
(c) WWC: test for non-atomic stores [46], [54]
Proc. P1
I1 : St a 1
Proc. P2
I2 : r1 = Ld a
I3 : FENCELL
I4 : r2 = Ld b
TSO, RMO and Alpha forbid,
POWER allow: r1 = 1, r2 = 0,
Proc. P3
I5 : St b 1
Proc. P4
I6 : r3 = Ld b
I7 : FENCELL
I8 : r4 = Ld a
but RC, ARM and
r3 = 1, r4 = 0
(d) IRIW: test for non-atomic stores [7]
Figure 2.
Litmus tests for store atomicity
SBE: In a machine with single-copy atomic stores (e.g., an
SC machine), when both I2 and I5 have returned value 1,
stores I1 and I4 must have been globally advertised. Thus
r2 and r4 cannot both be 0. However, a machine with store
buffers (e.g., a TSO machine) allows P1 to forward the value
of I1 to I2 locally without advertising I1 to other processors,
violating the single-copy atomicity of stores.
WRC: Assuming the store buffer is private to each processor
(i.e., multi-copy atomic stores), if one observes r1 = 2 and
r2 = 1 then r3 must be 2. However, if an architecture allows
a store buffer to be shared by P1 and P2 but not P3, then P2
can see the value of I1 from the shared store buffer before
I1 has updated the memory, allowing P3 to still see the old
value of a. A write-through cache shared by P1 and P2 but
not P3 can cause this non-atomic store behavior in a similar
way, e.g., I1 updates the shared write-through cache but has
not invalidated the copy in the private cache of P3 before I6
is executed.
WWC: This litmus test is similar to WRC but replaces the
load in I6 with a store. The behavior is possible if P1 and
P2 share a write-through cache or store buffer. However, RC
forbids this behavior (see Section X).
IRIW: This behavior is possible if P1 and P2 share a writethrough cache or a store buffer and so do P3 and P4.
B. Instruction Reordering Litmus Tests
Although processors fetch and commit instructions in order,
speculative and out-of-order execution causes behaviors as if
instructions were reordered. Figure 3 shows the litmus tests
on these reordering behaviors.
Proc. P1
Proc. P2
I1 : St a 1
I3 : St b 1
I2 : r1 = Ld b I4 : r2 = Ld a
SC forbids, but TSO allows:
r1 = 0, r2 = 0
Proc. P1
Proc. P2
I1 : St a 1 I3 : r1 = Ld b
I2 : St b 1 I4 : r2 = Ld a
TSO forbids, but Alpha and
RMO allow: r1 = 1, r2 = 0
(a) SB: test for St-Ld reordering [46] (b) MP: test for Ld-Ld and St-St
reorderings [7]
Proc. P1
Proc. P2
I1 : r1 = Ld b I3 : r2 = Ld a
I2 : St a 1
I4 : St b 1
TSO forbids, but Alpha,
RMO, RC, POWER and
ARM allow: r1 = r2 = 1
Proc. P1
Proc. P2
I1 : St a 1 I4 : r1 = Ld b
I2 : FENCE I5 : if(r1 6= 0) exit
I3 : St b 1 I6 : r2 = Ld a
Alpha, RMO, RC, ARM and
POWER allow: r1 = 1, r2 = 0
(c) LB: test for Ld-St reordering [7] (d) MP+Ctrl: test
dependency ordering
Proc. P1
Proc. P2
I1 : St a 1
I4 : r1 = Ld b
I2 : FENCE I5 : St (r1 + a) 42
I3 : St b 100 I6 : r2 = Ld a
Alpha, RMO, RC, ARM
and POWER allow: r1 =
100, r2 = 0
for
Proc. P1
Proc. P2
I1 : St a 1 I4 : r1 = Ld b
I2 : FENCE I5 : r2 = Ld r1
I3 : St b a
RMO, RC, ARM and
POWER forbid, but Alpha
allows: r1 = a, r2 = 0
(e) MP+Mem: test for memory- (f) MP+Data: test
dependency ordering
dependency ordering
Figure 3.
control-
for
data-
Litmus tests for instruction reorderings
SB: A TSO machine can execute I2 and I4 while I1 and I3
are buffered in the store buffers. The resulting behavior is as
if the store and the load were reordered on each processor.
MP: In an Alpha machine, I1 and I2 may be drained from
the store buffer of P1 out of order; I3 and I4 in the ROB of
P2 may be executed out of order. This is as if P1 reordered
the two stores and P2 reordered the two loads.
LB: Some machines may enter a store into the memory before
all older instructions have been committed. This results in the
Ld-St reordering shown in Figure 3c. Since instructions are
committed in order and stores are usually not on the critical
path, the benefit of the eager execution of stores is limited.
In fact we will show by simulation that Ld-St reordering
does not improve performance (Section VII).
MP+Ctrl: This test is a variant of MP. The two stores in
P1 must update memory in order due to the fence. Although
the execution of I6 is conditional on the result of I4 , P2 can
issue I6 speculatively by predicting branch I5 to be not taken.
The execution order I6 , I1 , I2 , I3 , I4 , I5 results in r1 = 1 and
r2 = 0.
MP+Mem: This test replaces the control dependency in
MP+Ctrl with a (potential) memory dependency, i.e., the
unresolved store address of I5 may be the same as the load
address of I6 before I4 is executed, However, P2 can execute
I6 speculatively by predicting the addresses are not the same.
This results in having I6 overtake I4 and I5 .
MP+Data: This test replaces the control dependency in
MP+Ctrl with a data dependency, i.e., the load address
of I5 depends on the result of I4 . A processor with loadvalue prediction [20]–[24] may guess the result of I4 before
executing it, and issue I5 speculatively. If the guess fails to
match the real execution result of I4 , then I5 would be killed.
But, if the guess is right, then essentially the execution of
the two data-dependent loads (I4 and I5 ) has been reordered.
C. Miscellaneous Tests
All programmers expect memory models to obey perlocation SC [55], i.e., all accesses to a single address appear
to execute in a sequential order which is consistent with the
program order of each thread (Figure 4).
Proc. P1
Proc. P2
I1 : r1 = Ld a I3 : St a 1
I2 : r2 = Ld a
Models with per-location
SC forbid: r1 = 1, r2 = 0
Figure 4.
Per-location SC
Proc. P1
Proc. P2
I1 : r1 = Ld b I3 : r2 = Ld a
I2 : St a r1
I4 : St b r2
All models forbid: r1 = r2 = 42
Figure 5.
Out-of-thin-air read
Out-of-thin-air behaviors (Figure 5) are impossible in real
implementations. Sometimes such behaviors are permitted
by axiomatic models due to incomplete axiomatization.
IV. D EFINING M EMORY M ODELS IN I 2 E
Figure 6 shows the I2 E abstract machines for SC,
TSO/PSO and WMM models. All abstract machines consist
of n atomic processors and an n-ported atomic memory m.
Each processor contains a register state s, which represents
all architectural registers, including both the general purpose
registers and special purpose registers, such as PC. The
abstract machines for TSO/PSO and WMM also contain a
store buffer sb for each processor, and the one for WMM
also contains an invalidation buffer ib for each processor as
shown in the figure. In the abstract machines all buffers are
unbounded. The operations of these buffers will be explained
shortly.
The operations of the SC abstract machine are the simplest:
in one step we can select any processor to execute the
next instruction on that processor atomically. That is, if
the instruction is a non-memory instruction (e.g., ALU or
branch), it just modifies the register states of the processor; if
it is a load, it reads from the atomic memory instantaneously
Processor ݅
Processor ݅
Reg State ݏ
Processor ݅
…
Reg State ݏ
…
…
Reg State ݏ
…
Store
Buffer ܾݏ
…
Inv
Buffer ܾ݅
Store
Buffer ܾݏ
Atomic Memory ݉
Atomic Memory ݉
Atomic Memory ݉
(a) SC
(b) TSO/PSO
(c) WMM
Figure 6.
…
I2 E abstract machines for different models
and updates the register state; and if it is a store, it updates
the atomic memory instantaneously and increments the PC.
A. TSO Model
The TSO abstract machine proposed in [2], [25] (Figure
6b) contains a store buffer sb for each processor. Just like SC,
any processor can execute an instruction atomically, and if
the instruction is a non-memory instruction, it just modifies
the local register state. A store is executed by inserting its
haddress, valuei pair into the local sb instead of writing the
data in memory. A load first looks for the load address in the
local sb and returns the value of the youngest store for that
address. If the address is not in the local sb, then the load
returns the value from the atomic memory. TSO can also
perform a background operation, which removes the oldest
store from a sb and writes it into the atomic memory. As
we discussed in Section III, store buffer allows TSO to do
St-Ld reordering, i.e., pass the SB litmus test (Figure 3a).
In order to enforce ordering in accessing the memory and
to rule out non-SC behaviors, TSO has a fence instruction,
which we refer to as Commit. When a processor executes
a Commit fence, it gets blocked unless its sb is empty.
Eventually, any sb will become empty as a consequence
of the background operations that move data from the sb to
the memory. For example, we need to insert a Commit fence
after each store in Figure 3a to forbid the non-SC behavior
in TSO.
We summarize the operations of the TSO abstract machine
in Figure 7. Each operation consists of a predicate and an
action. The operation can be performed by taking the action
only when the predicate is true. Each time we perform only
one operation (either instruction execution or sb dequeue)
atomically in the whole system (e.g., no two processors can
execute instructions simultaneously). The choice of which
operation to perform is nondeterministic.
Enabling St-St reordering: We can extend TSO to PSO by
changing the background operation to dequeue the oldest
store for any address in sb (see the PSO-DeqSb operation in
Figure 7). This extends TSO by permitting St-St reordering.
V. WMM M ODEL
WMM allows Ld-Ld reordering in addition to the reorderings allowed by PSO. Since a reordered load may read a stale
TSO-Nm (non-memory execution)
Predicate: The next instruction of a processor is a non-memory
instruction.
Action: Instruction is executed by local computation.
TSO-Ld (load execution)
Predicate: The next instruction of a processor is a load.
Action: Assume the load address is a. The load returns the value
of the youngest store for a in sb if a is present in the sb of
the processor, otherwise, the load returns m[a], i.e., the value of
address a in the atomic memory.
TSO-St (store execution)
Predicate: The next instruction of a processor is a store.
Action: Assume the store address is a and the store value is v.
The processor inserts the store ha, vi into its sb.
TSO-Com (Commit execution)
Predicate: The next instruction of a processor is a Commit and
the sb of the processor is empty.
Action: The Commit fence is executed simply as a NOP.
TSO-DeqSb (background store buffer dequeue)
Predicate: The sb of a processor is not empty.
Action: Assume the haddress, valuei pair of the oldest store in
the sb is ha, vi. Then this store is removed from sb, and the
atomic memory m[a] is updated to v.
PSO-DeqSb (background store buffer dequeue)
Predicate: The sb of a processor is not empty.
Action: Assume the value of the oldest store for some address a
in the sb is v. Then this store is removed from sb, and the atomic
memory m[a] is updated to v.
Figure 7.
Operations of the TSO/PSO abstract machine
value, we introduce a conceptual device called invalidation
buffer, ib, for each processor in the I2 E abstract machine (see
Figure 6c). ib is an unbounded buffer of haddress, valuei
pairs, each representing a stale memory value for an address
that can be observed by the processor. Multiple stale values
for an address in ib are kept ordered by their staleness.
The operations of the WMM abstract machine are similar
to those of PSO except for the background operation and
the load execution. When the background operation moves
a store from sb to the atomic memory, the original value
in the atomic memory, i.e., the stale value, enters the ib of
every other processor. A load first searches the local sb. If
the address is not found in sb, it either reads the value in
the atomic memory or any stale value for the address in the
local ib, the choice between the two being nondeterministic.
The abstract machine operations maintain the following
invariants: once a processor observes a store, it cannot observe
any staler store for that address. Therefore, (1) when a store
is executed, values for the store address in the local ib are
purged; (2) when a load is executed, values staler than
the load result are flushed from the local ib; and (3) the
background operation does not insert the stale value into
the ib of a processor if the sb of the processor contains the
address.
Just like introducing the Commit fence in TSO, to prevent
loads from reading the stale values in ib, we introduce the
Reconcile fence to clear the local ib. Figure 8 summarizes
the operations of the WMM abstract machine.
WMM-Nm (non-memory execution): Same as TSO-Nm.
WMM-Ld (load execution)
Predicate: The next instruction of a processor is a load.
Action: Assume the load address is a. If a is present in the sb
of the processor, then the load returns the value of the youngest
store for a in the local sb. Otherwise, the load is executed in
either of the following two ways (the choice is arbitrary):
1) The load returns the atomic memory value m[a], and all values
for a in the local ib are removed.
2) The load returns some value for a in the local ib, and all
values for a older than the load result are removed from the
local ib. (If there are multiple values for a in ib, the choice
of which one to read is arbitrary).
WMM-St (store execution)
Predicate: The next instruction of a processor is a store.
Action: Assume the store address is a and the store value is v.
The processor inserts the store ha, vi into its sb, and removes all
values for a from its ib.
WMM-Com (Commit execution): Same as TSO-Com.
WMM-Rec (execution of a Reconcile fence)
Predicate: The next instruction of a processor is a Reconcile.
Action: All values in the ib of the processor are removed.
WMM-DeqSb (background store buffer dequeue)
Predicate: The sb of a processor is not empty.
Action: Assume the value of the oldest store for some address
a in the sb is v. First, the stale haddress, valuei pair ha, m[a]i
is inserted to the ib of every other processor whose sb does not
contain a. Then this store is removed from sb, and m[a] is set
to v.
Figure 8.
Operations of the WMM abstract machine
A. Properties of WMM
Similar to TSO/PSO, WMM allows St-Ld and St-St
reorderings because of sb (Figures 3a and 3b). To forbid
the behavior in Figure 3a, we need to insert a Commit
followed by a Reconcile after the store in each processor.
Reconcile is needed to prevent loads from getting stale values
from ib. The I2 E definition of WMM automatically forbids
Ld-St reordering (Figure 3c) and out-of-thin-air behaviors
(Figure 5).
Ld-Ld reordering: WMM allows the behavior in Figure 3b
due to St-St reordering. Even if we insert a Commit between
the two stores in P1, the behavior is still allowed because
I4 can read the stale value 0 from ib. This is as if the two
loads in P2 were reordered. Thus, we also need a Reconcile
between the two loads in P2 to forbid this behavior in WMM.
No dependency ordering: WMM does not enforce any dependency ordering. For example, WMM allows the behaviors
of litmus tests in Figures 3d, 3e and 3f (I2 should be Commit
in case of WMM), because the last load in P2 can always
get the stale value 0 from ib in each test. Thus, it requires
Reconcile fences to enforce dependency ordering in WMM.
In particular, WMM can reorder the data-dependent loads
(i.e., I4 and I5 ) in Figure 3f.
Multi-copy atomic stores: Stores in WMM are multicopy atomic, and WMM allows the behavior in Figure 2a
even when Reconcile fences are inserted between Ld-Ld
pairs hI2 , I3 i and hI5 , I6 i. This is because a store can be
read by a load from the same processor while the store
is in sb. However, if the store is ever pushed from sb
to the atomic memory, it becomes visible to all other
processors simultaneously. Thus, WMM forbids the behaviors
in Figures 2b, 2c and 2d (FENCELL should be Reconcile in
these tests).
Per-location SC: WMM enforces per-location SC (Figure 4),
because both sb and ib enforce FIFO on same address entries.
B. Axiomatic Definition of WMM
Based on the above properties of WMM, we give a simple
axiomatic definition for WMM in Figure 9 in the style of
the axiomatic definitions of TSO and Alpha. A True entry
in the order-preserving table (Figure 9b) indicates that if
instruction X precedes instruction Y in the program order
(X <po Y ) then the order must be maintained in the global
memory order (<mo ). <mo is a total order of all the memory
and fence instructions from all processors. The notation
rf
S −→ L means a load L reads from a store S. The notation
max<mo {set of stores} means the youngest store in the
set according to <mo . The axioms are self-explanatory: the
program order must be maintained if the order-preserving
table says so, and a load must read from the youngest
store among all stores that precede the load in either the
memory order or the program order. (See Appendix A for
the equivalence proof of the axiomatic and I2 E definitions.)
These axioms also hold for Alpha with a slightly different
order-preserving table, which marks the (Ld,St) entry as
a = b. (Alpha also merges Commit and Reconcile into a
single fence). However, allowing Ld-St reordering creates
the possibility of out-of-thin-air behaviors, and Alpha uses an
additional complicated axiom to disallow such behaviors [5,
Chapter 5.6.1.7]. This axiom requires considering all possible
execution paths to determine if a store is ordered after a load
by dependency, while normal axiomatic models only examine
a single execution path at a time. Allowing Ld-St reordering
also makes it difficult to define Alpha operationally.
Axiom Inst-Order (preserved instruction ordering):
X <po Y ∧ order(X, Y ) ⇒ X <mo Y
Axiom Ld-Val (the value of a load):
rf
St a v −→ Ld a ⇒ St a v =
max<mo {St a v 0 | St a v 0 <mo Ld a ∨ St a v 0 <po Ld a}
(a) Axioms for WMM
HH Y
X H
H
Ld a
St a v
Reconcile
Commit
Ld b
St b v 0
Reconcile
Commit
a=b
False
True
False
True
a=b
True
True
True
False
True
True
True
True
True
True
(b) WMM order-preserving table, i.e. order(X, Y ) where X <po Y
Figure 9.
Axiomatic definition of WMM
C. Compiling C++11 to WMM
OOO Processor ܲ݅
…
Reorder Buffer (ROB)
C++ primitives [47] can be mapped to WMM instructions
in an efficient way as shown in Figure 10. For the purpose
of comparison, we also include a mapping to POWER [56].
C++ operations
Non-atomic Load
Load Relaxed
Load Consume
Load Acquire
Load SC
Non-atomic Store
Store Relaxed
Store Release
Store SC
Figure 10.
WMM instructions
Ld
Ld
Ld; Reconcile
Ld; Reconcile
Commit; Reconcile;
Ld; Reconcile
St
St
Commit; St
Commit; St
POWER instructions
Ld
Ld
Ld
Ld; cmp; bc; isync
sync; Ld; cmp;
bc; isync
St
St
lwsync; St
sync; St
Mapping C++ to WMM and POWER
The Commit; Reconcile sequence in WMM is the same as
a sync fence in POWER, and Commit is similar to lwsync.
The cmp; bc; isync sequence in POWER serves as a Ld-Ld
fence, so it is similar to a Reconcile fence in WMM. In case
of Store SC in C++, WMM uses a Commit while POWER
uses a sync, so WMM effectively saves one Reconcile. On
the other hand, POWER does not need any fence for Load
Consume in C++, while WMM requires a Reconcile.
Besides the C++ primitives, a common programming
paradigm is the well-synchronized program, in which all
critical sections are protected by locks. To maintain SC
behaviors for such programs in WMM, we can add a
Reconcile after acquiring the lock and a Commit before
releasing the lock.
For any program, if we insert a Commit before every store
and insert a Commit followed by a Reconcile before every
load, then the program behavior in WMM is guaranteed to
be sequentially consistent. This provides a conservative way
for inserting fences when performance is not an issue.
VI. WMM I MPLEMENTATION
WMM can be implemented using conventional OOO
multiprocessors, and even the most aggressive speculative
techniques cannot step beyond WMM. To demonstrate this,
we describe an OOO implementation of WMM, and show
simultaneously how the WMM model (i.e., the I2 E abstract
machine) captures the behaviors of the implementation. The
implementation is described abstractly to skip unrelated
details (e.g., ROB entry reuse). The implementation consists
of n OOO processors and a coherent write-back cache
hierarchy which we discuss next.
Ld resp
delay
Ld req
…
Store Buffer
Port ݅
St req
Mem Req Buffer ܾ݉ݎሾ݅ሿ
St resp
Write‐back
Cache Hierarchy
(CCM)
Atomic Memory ݉
Figure 11.
CCM+OOO: implementation of WMM
Consider a real n-ported write-back cache hierarchy with
each port i connected to processor P i. A request issued to
port i may be from a load instruction in the ROB of P i or
a store in the store buffer of P i. In conventional coherence
protocols, all memory requests can be serialized, i.e., each
request can be considered as taking effect at some time point
within its processing period [57]. For example, consider the
non-stalling MSI directory protocol in the Primer by Sorin
et al. [58, Chapter 8.7.2]. In this protocol, a load request
takes effect immediately if it hits in the cache; otherwise, it
takes effect when it gets the data at the directory or a remote
cache with M state. A store request always takes effect at
the time of writing the cache, i.e., either when it hits in the
cache, or when it has received the directory response and all
invalidation responses in case of miss. We also remove the
requesting store from the store buffer when a store request
takes effect. Since a cache cannot process multiple requests
to the same address simultaneously, we assume requests to
the same address from the same processor are processed in
the order that the requests are issued to the cache.
CCM (Figure 11) abstracts the above cache hierarchy
by operating as follows: every new request from port i is
inserted into a memory request buffer mrb[i], which keeps
requests to the same address in order; at any time we can
remove the oldest request for an address from a mrb, let
the request access the atomic memory m, and either send
the load result to ROB (which may experience a delay)
or immediately dequeue the store buffer. m represents the
coherent memory states. Removing a request from mrb and
accessing m captures the moment when the request takes
effect.
It is easy to see that the atomic memory in CCM
corresponds to the atomic memory in the WMM model,
because they both hold the coherent memory values. We will
show shortly that how WMM captures the combination of
CCM and OOO processors. Thus any coherent protocol that
can be abstracted as CCM can be used to implement WMM.
B. Out-of-Order Processor (OOO)
A. Write-Back Cache Hierarchy (CCM)
We describe CCM as an abstraction of a conventional
write-back cache hierarchy to avoid too much details. In the
following, we explain the function of such a cache hierarchy,
abstract it to CCM, and relate CCM to the WMM model.
The major components of an OOO processor are the ROB
and the store buffer (see Figure 11). Instructions are fetched
into and committed from ROB in order; loads can be issued
(i.e., search for data forwarding and possibly request CCM)
as soon as its address is known; a store is enqueued into the
store buffer only when the store commits (i.e., entries in a
store buffer cannot be killed). To maintain the per-location
SC property of WMM, when a load L is issued, it kills
younger loads which have been issued but do not read from
stores younger than L. Next we give the correspondence
between OOO and WMM.
Store buffer: The state of the store buffer in OOO is
represented by the sb in WMM. Entry into the store buffer
when a store commits in OOO corresponds to the WMM-St
operation. In OOO, the store buffer only issues the oldest
store for some address to CCM. The store is removed from
the store buffer when the store updates the atomic memory
in CCM. This corresponds to the WMM-DeqSb operation.
ROB and eager loads: Committing an instruction from
ROB corresponds to executing it in WMM, and thus the
architectural register state in both WMM and OOO must
match at the time of commit. Early execution of a load L to
address a with a return value v in OOO can be understood by
considering where ha, vi resides in OOO when L commits.
Reading from sb or atomic memory m in the WMM-Ld
operation covers the cases that ha, vi is, respectively, in the
store buffer or the atomic memory of CCM when L commits.
Otherwise ha, vi is no longer present in CCM+OOO at the
time of load commit and must have been overwritten in the
atomic memory of CCM. This case corresponds to having
performed the WMM-DeqSb operation to insert ha, vi into
ib previously, and now using the WMM-Ld operation to read
v from ib.
Speculations: OOO can issue a load speculatively by
aggressive predictions, such as branch prediction (Figure 3d),
memory dependency prediction (Figure 3e) and even loadvalue prediction (Figure 3f). As long as all predictions related
to the load eventually turn out to be correct, the load result
got from the speculative execution can be preserved. No
further check is needed. Speculations effectively reorder
dependent instructions, e.g., load-value speculation reorders
data-dependent loads. Since WMM does not require preserving any dependency ordering, speculations will neither break
WMM nor affect the above correspondence between OOO
and WMM.
Fences: Fences never go into store buffers or CCM in the
implementation. In OOO, a Commit can commit from ROB
only when the local store buffer is empty. Reconcile plays a
different role; at the time of commit it is a NOP, but while
it is in the ROB, it stalls all younger loads (unless the load
can bypass directly from a store which is younger than the
Reconcile). The stall prevents younger loads from reading
values that would become stale when the Reconcile commits.
This corresponds to clearing ib in WMM.
Summary: For any execution in the CCM+OOO implementation, we can operate the WMM model following the
above correspondence. Each time CCM+OOO commits an
instruction I from ROB or dequeues a store S from a store
buffer to memory, the atomic memory of CCM, store buffers,
and the results of committed instructions in CCM+OOO
are exactly the same as those in the WMM model when the
WMM model executes I or dequeues S from sb, respectively.
VII. P ERFORMANCE E VALUATION OF WMM
We evaluate the performance of implementations of WMM,
Alpha, SC and TSO. All implementations use OOO cores
and coherent write-back cache hierarchy. Since Alpha allows
Ld-St reordering, the comparison of WMM and Alpha will
show whether such reordering affects performance.
A. Evaluation Methodology
We ran SPLASH-2x benchmarks [59], [60] on an 8-core
multiprocessor using the ESESC simulator [61]. We ran all
benchmarks except ocean ncp, which allocates too much
memory and breaks the original simulator. We used simmedium inputs except for cholesky, fft and radix, where we
used sim-large inputs. We ran all benchmarks to completion
without sampling.
The configuration of the 8-core multiprocessor is shown
in Figures 12 and 13 . We do not use load-value speculation
in this evaluation. The Alpha implementation can mark a
younger store as committed when instruction commit is
stalled, as long as the store can never be squashed and
the early commit will not affect single-thread correctness. A
committed store can be issued to memory or merged with
another committed store in WMM and Alpha. SC and TSO
issue loads speculatively and monitor L1 cache evictions to
kill speculative loads that violate the consistency model. We
also implement store prefetch as an optional feature for SC
and TSO; We use SC-pf and TSO-pf to denote the respective
implementations with store prefetch.
Cores
L3 cache
Memory
8 cores (@2GHz) with private L1 and L2 caches
4MB shared, MESI coherence, 64-byte cache line
8 banks, 16-way, LRU replacement, max 32 req per bank
3-cycle tag, 10-cycle data (both pipelined)
5 cycles between cache bank and core (pipelined)
120-cycle latency, max 24 requests
Figure 12.
Frontend
ROB
Function
units
Ld queue
St queue
L1 D
cache
L2 cache
Multiprocessor system configuration
fetch + decode + rename, 7-cycle pipelined latency in all
2-way superscalar, hybrid branch predictor
128 entries, 2-way issue/commit
2 ALUs, 1 FPU, 1 branch unit, 1 load unit, 1 store unit
32-entry reservation station per unit
Max 32 loads
Max 24 stores, containing speculative and committed stores
32KB private, 1 bank, 4-way, 64-byte cache line
LRU replacement, 1-cycle tag, 2-cycle data (pipelined)
Max 32 upgrade and 8 downgrade requests
128KB private, 1 bank, 8-way, 64-byte cache line
LRU replacement, 2-cycle tag, 6-cycle data (both pipelined)
Max 32 upgrade and 8 downgrade requests
Figure 13.
Core configuration
empty
pendSt
flushRep
flushInv
flushLS
exe
active
1
0.5
(Bench.)
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
MM
W ha f
p
Al O-p
TSO
TS -pf
SC
SC
0
barnes
cholesky
fft
Figure 14.
fmm
lu_cb
lu_ncb
ocean_cp
radix
radiosity
raytrace
volrend
water_nsq
water_sp
Normalized execution time and its breakdown at the commit slot of ROB
B. Simulation Results
A common way to study the performance of memory
models is to monitor the commit of instructions at the commit
slot of ROB (i.e., the oldest ROB entry). Here are some
reasons why an instruction may not commit in a given cycle:
• empty: The ROB is empty.
• exe: The instruction at the commit slot is still executing.
• pendSt: The load (in SC) or Commit (in TSO, Alpha and
WMM) cannot commit due to pending older stores.
• flushLS: ROB is being flushed because a load is killed
by another older load (only in WMM and Alpha) or older
store (in all models) to the same address.
• flushInv: ROB is being flushed after cache invalidation
caused by a remote store (only in SC or TSO).
• flushRep: ROB is being flushed after cache replacement
(only in SC or TSO).
Figure 14 shows the execution time (normalized to WMM)
and its breakdown at the commit slot of ROB. The total
height of each bar represents the normalized execution time,
and stacks represent different types of stall times added to
the active committing time at the commit slot.
WMM versus SC: WMM is much faster than both SC and
SC-pf for most benchmarks, because a pending older store
in the store queue can block SC from committing loads.
WMM versus TSO: WMM never does worse than TSO or
TSO-pf, and in some cases it shows up to 1.45× speedup over
TSO (in radix) and 1.18× over TSO-pf (in lu ncb). There
are two disadvantages of TSO compared to WMM. First,
load speculation in TSO is subject to L1 cache eviction, e.g.,
in benchmark ocean cp. Second, TSO requires prefetch to
reduce store miss latency, e.g., a full store queue in TSO stalls
issue to ROB and makes ROB empty in benchmark radix.
However, prefetch may sometimes degrade performance due
to interference with load execution, e.g., TSO-pf has more
commit stalls due to unfinished loads in benchmark lu ncb.
WMM versus Alpha: Figure 15 shows the average number
of cycles that a store in Alpha can commit before it reaches
the commit slot. However, the early commit (i.e., Ld-St
reordering) does not make Alpha outperform WMM (see
Figure 14), because store buffers can already hide the store
miss latency. Note that ROB is typically implemented as
a FIFO (i.e., a circular buffer) for register renaming (e.g.,
freeing physical registers in order), precise exceptions, etc.
Thus, if the early committed store is in the middle of ROB, its
ROB entry cannot be reused by a newly fetched instruction,
i.e., the effective size of the ROB will not increase. In
summary, the Ld-St reordering in Alpha does not increase
performance but complicates the definition (Section V-B).
15
Cycles
Normalized time
1.5
10
5
0
bar cho fft
nes les
ky
Figure 15.
fmm lu_ lu_ oce rad rad ray vol wa wa
cb ncb an ix
ios trac ren ter_ ter_
_cp
ity
nsq sp
e d
Average cycles to commit stores early in Alpha
VIII. WMM-S M ODEL
Unlike the multi-copy atomic stores in WMM, stores in
some processors (e.g., POWER) are non-atomic due to shared
write-through caches or shared store buffers. If multiple
processors share a store buffer or write-through cache, a
store by any of these processors may be seen by all these
processors before other processors. Although we could tag
stores with processor IDs in the store buffer, it is infeasible
to separate values stored by different processors in a cache.
In this section, we introduce a new I2 E model, WMMS, which captures the non-atomic store behaviors in a way
independent from the sharing topology. WMM-S is derived
from WMM by adding a new background operation. We will
show later in Section IX why WMM-S can be implemented
using memory systems with non-atomic stores.
A. I2 E Definition of WMM-S
The structure of the abstract machine of WMM-S is the
same as that of WMM. To model non-atomicity of stores,
i.e., to make a store by one processor readable by another
processor before the store updates the atomic memory, WMMS introduces a new background operation that copies a store
from one store buffer into another. However, we need to
ensure that all stores for an address can still be put in a total
order (i.e., the coherence order), and the order seen by any
processor is consistent with this total order (i.e., per-location
SC).
To identify all the copies of a store in various store buffers,
we assign a unique tag t when a store is executed (by being
inserted into sb), and this tag is copied when a store is
copied from one store buffer to another. When a background
operation dequeues a store from a store buffer to the memory,
all its copies must be deleted from all the store buffers which
have them. This requires that all copies of the store are the
oldest for that address in their respective store buffers.
All the stores for an address in a store buffer can be strictly
ordered as a list, where the youngest store is the one that
entered the store buffer last. We make sure that all ordered
lists (of all store buffers) can be combined transitively to
form a partial order (i.e., no cycle), which has now to be
understood in terms of the tags on stores because of the
copies. We refer to this partial order as the partial coherence
order (<co ), because it is consistent with the coherence order.
Consider the states of store buffers shown in Figure 16
(primes are copies). A, B, C and D are different stores
to the same address, and their tags are tA , tB , tC and tD ,
respectively. A0 and B 0 are copies of A and B respectively
created by the background copy operation. Ignoring C 0 , the
partial coherence order contains: tD <co tB <co tA (D is
older than B, and B is older than A0 in P2), and tC <co tB
(C is older than B 0 in P3). Note that tD and tC are not
related here.
At this point, if we copied C in P3 as C 0 into P1, we
would add a new edge tA <co tC , breaking the partial order
by introducing the cycle tA <co tC <co tB <co tA . Thus
copying of C into P1 should be forbidden in this state.
Similarly, copying a store with tag tA into P1 or P2 should
be forbidden because it would immediately create a cycle:
tA <co tA . In general, the background copy operation must
be constrained so that the partial coherence order is still
acyclic after copying.
Enqueued later (younger)
↕
Enqueued earlier (Older)
Figure 16.
P1 sb
P2 sb
P3 sb
C’:
A’:
B:
D:
B’:
A:
C:
Example states of store buffers
Figure 17 shows the background operations of the WMMS abstract machine. The operations that execute instructions
in WMM-S are the same as those in WMM, so we do not
show them again. (The store execution operation in WMM-S
needs to also insert the tag of the store into sb).
Binding background copy with load execution: If the
WMM-S-Copy operation is restricted to always happen right
before a load execution operation that reads from the newly
created copy, it is not difficult to prove that the WMMS model remains the same, i.e., legal behaviors do not
change. In the rest of the paper, we will only consider this
“restricted” version of WMM-S. In particular, all WMM-SCopy operations in the following analysis of litmus tests fall
WMM-DeqSb (background store buffer dequeue)
Predicate: There is a store S in a store buffer, and all copies of
S are the oldest store for that address in their respective store
buffers.
Action: Assume the haddress, value, tagi tuple of store S is
ha, v, ti. First, the stale haddress, valuei pair ha, m[a]i is inserted
to the ib of every processor whose sb does not contain a. Then
all copies of S are removed from their respective store buffers,
and the atomic memory m[a] is updated to v.
WMM-S-Copy (background store copy)
Predicate: There is a store S that is in the sb of some processor
i but not in the sb of some other processor j. Additionally, the
partial coherence order will still be acyclic if we insert a copy of
S into the sb of processor j.
Action: Insert a copy of S into the sb of processor j, and remove
all values for the store address of S from the ib of processor j.
Figure 17.
Background operations of WMM-S
into this pattern.
B. Properties of WMM-S
WMM-S enforces per-location SC (Figure 4), because it
prevents cycles in the order of stores to the same address. It
also allows the same instruction reorderings as WMM does
(Figure 3). We focus on the store non-atomicity of WMM-S.
Non-atomic stores and cumulative fences: Consider the
litmus tests for non-atomic stores in Figures 2b, 2c and
2d (FENCELL should be Reconcile in these tests). WMM-S
allows the behavior in Figure 2b by copying I1 into the sb of
P2 and then executing I2 , I3 , I4 , I5 , I6 sequentially. I1 will
not be dequeued from sb until I6 returns value 0. To forbid
this behavior, a Commit is required between I2 and I3 in
P2 to push I1 into memory. Similarly, WMM-S allows the
behavior in Figure 2c (i.e., we copy I1 into the sb of P2
to satisfy I2 , and I1 is dequeued after I5 has updated the
atomic memory), and we need a Commit between I2 and
I3 to forbid the behavior. In both litmus tests, the inserted
fences have a cumulative global effect in ordering I1 before
I3 and the last instruction in P3.
WMM-S also allows the behavior in Figure 2d by copying
I1 into the sb of P2 to satisfy I2 , and copying I5 into the sb
of P4 to satisfy I6 . To forbid the behavior, we need to add a
Commit right after the first load in P2 and P4 (but before the
FENCELL /Reconcile that we added to stop Ld-Ld reordering).
As we can see, Commit and Reconcile are similar to release
and acquire respectively. Cumulation is achieved by globally
advertising observed stores (Commit) and preventing later
loads from reading stale values (Reconcile).
Programming properties: WMM-S is the same as WMM
in the properties described in Section V-C, including the
compilation of C++ primitives, maintaining SC for wellsynchronized programs, and the conservative way of inserting
fences.
IX. WMM-S I MPLEMENTATIONS
Since WMM-S is strictly more relaxed than WMM, any
WMM implementation is a valid WMM-S implementation.
However, we are more interested in implementations with
non-atomic memory systems. Instead of discussing each
specific system one by one, we explain how WMM-S can be
implemented using the ARMv8 flowing model, which is a
general abstraction of non-atomic memory systems [8]. We
first describe the adapted flowing model (FM) which uses
fences in WMM-S instead of ARM, and then explain how it
obeys WMM-S.
A. The Flowing Model (FM)
FM consists of a tree of segments s[i] rooted at the
atomic memory m. For example, Figure 19 shows four OOO
processors (P1. . .P4) connected to a 4-ported FM which has
six segments (s[1 . . . 6]). Each segment is a list of memory
requests, (e.g., the list of blue nodes in s[6], whose head is
at the bottom and the tail is at the top).
OOO interacts with FM in a slightly different way than
CCM. Every memory request from a processor is appended to
the tail of the list of the segment connected to the processor
(e.g., s[1] for P1). OOO no longer contains a store buffer;
after a store is committed from ROB, it is directly sent to
FM and there is no store response. When a Commit fence
reaches the commit slot of ROB, the processor sends a
Commit request to FM, and the ROB will not commit the
Commit fence until FM sends back the response for the
Commit request.
Inside FM, there are three background operations: (1) Two
requests in the same segment can be reordered in certain
cases; (2) A load can bypass from a store in the same segment;
(3) The request at the head of the list of a segment can flow
into the parent segment (e.g., flow from s[1] into s[5]) or the
atomic memory (in case the parent of the segment, e.g., s[6],
is m). Details of these operations are shown in Figure 18.
It is easy to see that FM abstracts non-atomic memory
systems, e.g., Figure 19 abstracts a system in which P1
and P2 share a write-through cache while P3 and P4 share
another.
Two properties of FM+OOO: First, FM+OOO enforces
per-location SC because the segments in FM never reorder
requests to the same address. Second, stores for the same
address, which lie on the path from a processor to m in
the tree structure of FM, are strictly ordered based on their
distance to the tree root m; and the combination of all
such orderings will not contain any cycle. For example, in
Figure 19, stores in segments s[3] and s[6] are on the path
from P3 to m; a store in s[6] is older than any store (for the
same address) in s[3], and stores (for the same address) in
the same segment are ordered from bottom to top (bottom
is older).
FM-Reorder (reorder memory requests)
Predicate: The list of segment s[i] contains two consecutive
requests rnew and rold (rnew is above rold in s[i]); and neither
of the following is true:
1) rnew and rold are memory accesses to the same address.
2) rnew is a Commit and rold is a store.
Action: Reorder rnew and rold in the list of s[i].
FM-Bypass (store forwarding)
Predicate: The list of segment s[i] contains two consecutive
requests rnew and rold (rnew is above rold in s[i]). rnew is a
load, rold is a store, and they are for the same address.
Action: we send the load response for rnew using the store value
of rold , and remove rnew from the segment.
FM-Flow (flow request)
Predicate: A segment s[i] is not empty.
Action: Remove the request r which is the head of the list of s[i].
If the parent of s[i] in the tree structure is another segment s[j],
we append r to the tail of the list of s[j]. Otherwise, the parent
of s[i] is m, and we take the following actions according to the
type of r:
• If r is a load, we send a load response using the value in m.
• If r is a store ha, vi, we update m[a] to v.
• If r is a Commit, we send a response to the requesting processor
and the Commit fence can then be committed from ROB.
Figure 18.
Background operations of FM
B. Relating FM+OOO to WMM-S
WMM-S can capture the behaviors of any program
execution in implementation FM+OOO in almost the same
way that WMM captures the behaviors of CCM+OOO. When
a store updates the atomic memory in FM+OOO, WMMS performs a WMM-S-DeqSb operation to dequeue that
store from store buffers to memory. When an instruction
is committed from a ROB in FM+OOO, WMM-S executes
that instruction. The following invariants hold after each
operation in FM+OOO and the corresponding operation in
WMM-S:
1) For each instruction committed in FM+OOO, the execution results in FM+OOO and WMM-S are the same.
2) The atomic memories in FM+OOO and WMM-S match.
3) The sb of each processor P i in WMM-S holds exactly all
the stores in FM+OOO that is observed by the commits
of P i but have not updated the atomic memory. (A store
is observed by the commits of P i if it has been either
committed by P i or returned by a load that has been
committed by P i).
4) The order of stores for the same address in the sb of any
processor in WMM-S is exactly the order of those stores
on the path from P i to m in FM+OOO.
It is easy to see how the invariants are maintained when
the atomic memory is updated or a non-load instruction is
committed in FM+OOO. To understand the commit of a load
L to address a with result v in processor P i in FM+OOO, we
still consider where ha, vi resides when L commits. Similar
to WMM, reading atomic memory m or local ib in the load
execution operation of WMM-S covers the cases that ha, vi
is still in the atomic memory of FM or has already been
overwritten by another store in the atomic memory of FM,
respectively. In case ha, vi is a store that has not yet updated
the atomic memory in FM, ha, vi must be on the path from
P i to m. In this case, if ha, vi has been observed by the
commits of P i before L is committed, then L can be executed
by reading the local sb in WMM-S. Otherwise, on the path
from P i to m, ha, vi must be younger than any other store
observed by the commits of P i. Thus, WMM-S can copy
ha, vi into the sb of P i without breaking any invariant. The
copy will not create any cycle in <co because of invariants 3
and 4 as well as the second property of FM+OOO mentioned
above. After the copy, WMM-S can have L read v from the
local sb.
Performance comparison with ARM and POWER: As
we have shown that WMM-S can be implemented using the
generalized memory system of ARM, we can turn an ARM
multicore into a WMM-S implementation by stopping Ld-St
reordering in the ROB. Since Section VII already shows
that Ld-St reordering does not affect performance, we can
conclude qualitatively that there is no discernible performance
difference between WMM-S and ARM implementations. The
same arguments apply to the comparison against POWER
and RC.
list tail req
list head req
FM
Atomic memory
Figure 19.
OOO+FM
Proc. P1
Proc P2
I1 : St a 1
I4 : r1 = Ld b
I2 : MEMBAR I5 : if(r1 6= 1) exit
I3 : St b 1
I6 : St c 1
I7 : r2 = Ld c
I8 : r3 = a+r2 −1
I9 : r4 = Ld r3
RMO forbids: r1 = 1, r2 = 1
r3 = a, r4 = 0
Figure 20. RMO dependency order
X. P ROBLEMS OF RC AND RMO
Here we elaborate the problems of RC (both RCsc and
RCpc ) and RMO, which have been pointed out in Section I.
RC: Although the RC definition [6] allows the behaviors of
WRC and IRIW (Figures 2b and 2d), it disallows the behavior
of WWC (Figure 2c). In WWC, when I2 reads the value of
store I1 , the RC definition says that I1 is performed with
respect to (w.r.t) P2. Since store I5 has not been issued due
to the data dependencies in P2 and P3, I1 must be performed
w.r.t P2 before I5 . The RC definition says that “all writes
to the same location are serialized in some order and are
performed in that order with respect to any processor” [6,
Section 2]. Thus, I1 is before I5 in the serialization order
of stores for address a, and the final memory value of a
cannot be 2 (the value of I1 ), i.e., RC forbids the behavior
of WWC and thus forbids shared write-through caches in
implementations.
RMO: The RMO definition [4, Section D] is incorrect in
enforcing dependency ordering. Consider the litmus test in
Figure 20 (MEMBAR is the fence in RMO). In P2, the
execution of I6 is conditional on the result of I4 , I7 loads
from the address that I6 stores to, and I9 uses the results
of I7 . According the definition of dependency ordering in
RMO [4, Section D.3.3], I9 depends on I4 transitively. Then
the RMO axioms [4, Section D.4] dictate that I9 must be
after I4 in the memory order, and thus forbid the behavior
in Figure 20. However, this behavior is possible in hardware
with speculative load execution and store forwarding, i.e., I7
first speculatively bypasses from I6 , and then I9 executes
speculatively to get 0. Since most architects will not be
willing to give up on these two optimizations, RISC-V should
not adopt RMO.
XI. C ONCLUSION
We have proposed two weak memory models, WMM and
WMM-S, for RISC-V with different tradeoffs between definitional simplicity and implementation flexibility. However
RISC-V can have only one memory model. Since there is
no obvious evidence that restricting to multi-copy atomic
stores affects performance or increases hardware complexity,
RISC-V should adopt WMM in favor of simplicity.
XII. ACKNOWLEDGMENT
We thank all the anonymous reviewers on the different
versions of this paper over the two years. We have also
benefited from the discussions with Andy Wright, Thomas
Bourgeat, Joonwon Choi, Xiangyao Yu, and Guowei Zhang.
This work was done as part of the Proteus project under the
DARPA BRASS Program (grant number 6933218).
A PPENDIX A.
P ROOF OF E QUIVALENCE BETWEEN WMM I2 E M ODEL
AND WMM A XIOMATIC M ODEL
Here we present the equivalence proof for the I2 E
definition and the axiomatic definition of WMM.
Theorem 1 (Soundness). WMM I2 E model ⊆ WMM axiomatic model.
Proof: The goal is that for any execution in the WMM
rf
I E model, we can construct relations h<po , <mo , −→i that
have the same program behavior and satisfy the WMM
axioms. To do this, we first introduce the following ghost
states to the I2 E model:
• Field source in the atomic memory: For each address a,
we add state m[a].source to record the store that writes
the current memory value.
• Fields source and overwrite in the invalidation buffer: For
each stale value ha, vi in an invalidation buffer, we add
state v.source to denote the store of this stale value, and
add state v.overwrite to denote the store that overwrites
v.source in the memory.
2
Per-processor list <po-i2e : For each processor, <po-i2e is
the list of all the instructions that has been executed by
the processor. The order in <po-i2e is the same as the
execution order in the processor. We also use <po-i2e to
represent the ordering relation in the list (the head of the
list is the oldest/minimum in <po-i2e ).
• Global list <mo-i2e : <mo-i2e is a list of all the executed
loads, executed fences, and stores that have been dequeued
from the store buffers. <mo-i2e contains instructions from
all processors. We also use <mo-i2e to represent the
ordering relation in the list (the head of the list is the
oldest/minimum in <mo-i2e ).
rf -i2e
rf -i2e
• Read-from relations −
−−−→: −−−−→ is a set of edges. Each
edge points from a store to a load, indicating that the load
had read from the store in the I2 E model.
m[a].source initially points to the initialization store, and
rf -i2e
<po-i2e , <mo-i2e , −−−−→ are all initially empty. We now show
how these states are updated in the operations of the WMM
I2 E model.
1) WMM-Nm, WMM-Com, WMM-Rec, WMM-St: Assume
the operation executes an instruction I in processor i. We
append I to the tail of list <po-i2e of processor i. If I is
a fence (i.e., the operation is WMM-Com or WMM-Rec),
then we also append I to the tail of list <mo-i2e .
2) WMM-DeqSb: Assume the operation dequeues a store S
for address a. In this case, we update m[a].source to be S.
Let S0 be the original m[a].source before this operation is
performed. Then for each new stale value ha, vi inserted
into any invalidation buffer, we set v.source = S0 and
v.overwrite = S. We also append S to the tail of list
<mo-i2e .
3) WMM-Ld: Assume the operation executes a load L for
address a in processor i. We append L to the tail of list
<po-i2e of processor i. The remaining actions depends
on how L gets its value in this operation:
• If L reads from a store S in the local store buffer, then
rf -i2e
we add edge S −−−−→ L, and append L to the tail of
list <mo-i2e .
• If L reads the atomic memory m[a], then we add edge
rf -i2e
m[a].source −−−−→ L, and append L to the tail of list
<mo-i2e .
• If L reads a stale value ha, vi in the local invalidation
rf -i2e
buffer, then we add edge v.source −−−−→ L, and we
insert L to be right before v.overwrite in list <mo-i2e
(i.e., L is older than v.overwrite, but is younger than
any other instruction which is older than v.overwrite).
As we will see later, at the end of the I2 E execution, <po-i2e ,
rf -i2e
rf
<mo-i2e and −−−−→ will become the h<po , <mo , −→i relations that satisfy the WMM axioms. Before getting there, we
show that the I2 E model has the following invariants after
each operation is performed:
1) For each address a, m[a].source in the I2 E model is the
•
youngest store for a in <mo-i2e .
2) All loads and fences that have been executed in the I2 E
model are in <mo-i2e .
3) An executed store is either in <mo-i2e or in store buffer,
i.e., for each processor i, the store buffer of processor i
contains exactly every store that has been executed in the
I2 E model but is not in <mo-i2e .
4) For any two stores S1 and S2 for the same address in the
store buffer of any processor i in the I2 E model, if S1 is
older than S2 in the store buffer, then S1 <po-i2e S2 .
5) For any processor i and any address a, address a cannot
be present in the store buffer and invalidation buffer of
processor i at the same time.
6) For any stale value v for any address a in the invalidation
buffer of any processor i in the I2 E model, the following
invariants hold:
a) v.source and v.overwrite are in <mo-i2e , and
v.source <mo-i2e v.overwrite, and there is no other
store for a between them in <mo-i2e .
b) For any Reconcile fence F that has been executed by
processor i in the I2 E model, F <mo-i2e v.overwrite.
c) For any store S for a that has been executed by
processor i in the I2 E model, S <mo-i2e v.overwrite.
d) For any load L for a that has been executed by
rf -i2e
processor i in the I2 E model, if store S −−−−→ L,
then S <mo-i2e v.overwrite.
7) For any two stale values v1 and v2 for the same address
in the invalidation buffer of any processor i in the I2 E
model, if v1 is older than v2 in the invalidation buffer,
then v1 .source <mo-i2e v2 .source.
8) For any instructions I1 and I2 , if I1 <po-i2e I2 and
order(I1 , I2 ) and I2 is in <mo-i2e , then I1 <mo-i2e I2 .
rf -i2e
9) For any load L and store S, if S −−−−→ L, then the
following invariants hold:
a) If S not in <mo-i2e , then S is in the store buffer of
the processor of L, and S <po-i2e L, and there is no
store S 0 for the same address in the same store buffer
such that S <po-i2e S 0 <po-i2e L.
b) If
S
is
in
<mo-i2e ,
then
S
=
maxmo-i2e {S 0 | S 0 .addr = L.addr ∧ (S 0 <po-i2e
L ∨ S 0 <mo-i2e L)}, and there is no other store
S 00 for the same address in the store buffer of the
processor of L such that S 00 <po-i2e L.
We now prove inductively that all invariants hold after each
operation R is performed in the I2 E model, i.e., we assume
all invariants hold before R is performed. In case performing
R changes some states (e.g., <mo-i2e ), we use superscript 0
to denote the state before R is performed (e.g., <0mo-i2e ) and
use superscript 1 to denote the state after R is performed
(e.g., <1mo-i2e ). Now we consider the type of R:
1) WMM-Nm: All invariants still hold.
2) WMM-St: Assume R executes a store S for address a
in processor i. R changes the states of the store buffer,
invalidation buffer, and <po-i2e of processor i. Now we
consider each invariant.
• Invariant 1, 2: These are not affected.
• Invariant 3: This invariant still holds for the newly
executed store S.
• Invariant 4: Since S becomes the youngest store in the
store buffer of processor i, this invariant still holds.
• Invariant 5: Since R will clear address a from the
invalidation buffer of processor i, this invariant still
holds.
• Invariant 6: Invariants 6a, 6b, 6d are not affected.
Invariant 6c still holds because there is no stale value
for a in the invalidation buffer of processor i after R
is performed.
• Invariant 7: This is not affected, because R can only
remove values from the invalidation buffer.
• Invariant 8: This is not affected because R is not in
<mo-i2e .
∗
∗
• Invariant 9: Consider load L and store S for address a
∗ rf -i2e
∗
∗
such that S −−−−→ L and L is from processor i. We
need to show that this invariant still holds for L∗ and
S ∗ . Since L∗ has been executed, we have L∗ <1po-i2e S.
Thus this invariant cannot be affected.
3) WMM-Com: Assume R executes a Commit fence F
in processor i. R adds F to the end of the <po-i2e of
processor i and adds F to the end of <mo-i2e . Now we
consider each invariant.
• Invariants 1, 3, 4, 5, 6, 7, 9: These are not affected.
• Invariant 2: This still holds because F is added to
<mo-i2e .
• Invariant 8: Consider instruction I in processor i such
that I <po-i2e F and order(I, F ). We need to show that
I <1mo-i2e F . Since order(I, F ), I can be a load, or
store, or fence. If I is a load or fence, since I has been
executed, invariant 2 says that I is in <0mo-i2e before R
is performed. Since F is added to the end of <mo-i2e ,
I <1mo-i2e F . If I is a store, the predicate of R says
that I is not in the store buffer. Then invariant 3 says
that I must be in <0mo-i2e , and we have I <1mo-i2e F .
4) WMM-Rec: Assume R executes a Reconcile fence F
in processor i. R adds F to the end of the <po-i2e of
processor i, adds F to the end of <mo-i2e , and clear the
invalidation buffer of processor i. Now we consider each
invariant.
• Invariants 1, 3, 4, 9: These are not affected.
• Invariant 2: This still holds because F is added to
<mo-i2e .
• Invariants 5, 6, 7: These invariant still hold because the
invalidation buffer of processor i is now empty.
• Invariant 8: Consider instruction I in processor i such
that I <po-i2e F and order(I, F ). We need to show that
I <1mo-i2e F . Since order(I, F ), I can be a load or
fence. Since I has been executed, I must be in <0mo-i2e
before R is performed according to invariant 2. Thus,
I <1mo-i2e F .
5) WMM-DeqSb: Assume R dequeues a store S for address
a from the store buffer of processor i. R changes the
store buffer of processor i, the atomic memory m[a], and
invalidation buffers of other processors. R also adds S
to the end of <mo-i2e . Now we consider each invariant.
• Invariant 1: This invariant still holds, because
m[a].source1 = S and S becomes the youngest store
for a in <1mo-i2e .
• Invariant 2: This is not affected.
• Invariant 3: This invariant still holds, because S is
removed from store buffer and added to <mo-i2e .
• Invariants 4: This is not affected because we only
remove stores from the store buffer.
• Invariant 5: The store buffer and invaliation buffer of
processor i cannot be affected. The store buffer and
invalidation buffer of processor j (6= i) may be affected,
because m[a]0 may be inserted into the invalidation
buffer of processor j. The predicate of R ensures that
the insertion will not happen if the store buffer of
processor j contains address a, so the invariant still
holds.
• Invariant 6: We need to consider the influence on both
existing stale values and the newly inserted stale values.
a) Consider stale value ha, vi which is in the invalidation buffer of processor j both before and after
operation R is performed. This implies j 6= i,
because the store buffer of processor i contains
address a before R is performed, and invariant 5
says that the invalidation buffer of processor i
cannot have address a before R is performed. Now
we show that each invariant still holds for ha, vi.
– Invariant 6a: This still holds because S is the
youngest in <mo-i2e .
– Invariant 6b: This is not affected.
– Invariant 6c: This is not affected because S is
not executed by processor j.
– Invariant 6d: Since S is not in <0mo-i2e , invariant 9a says that any load that has read S must
be from process i. Since i 6= j, this invariant
cannot be affected.
b) Consider the new stale value ha, vi inserted to
the invalidation buffer of process j (6= i). According to WMM-DeqSb, v = m[a]0 , v.source =
m[a].source0 , and v.overwrite = S. Now we check
each invariant.
– Invariant 6a: According to invariant 1,
v.source = m[a]0 .source is the youngest store
for a in <0mo-i2e . Since S (i.e., v.overwrite) is
appended to the tail of <0mo-i2e , this invariant
still holds.
•
•
– Invariant 6b: According to invariant 2, any
Reconcile fence F executed by processor j
must be in <0mo-i2e . Thus, F <1mo-i2e S =
v.overwrite, and the invariant still holds.
– Invariant 6c: The predicate of R says that the
store buffer of processor j cannot contain address
a. Therefore, according to invariant 3, any store
S 0 for a executed by processor j must be in
<0mo-i2e . Thus, S 0 <1mo-i2e S = v.overwrite, and
the invariant still holds.
– Invariant 6d: Consider load L for address a
that has been executed by processor j. Assume
rf -i2e
store S 0 −−−−→ L. The predicate of R says that
the store buffer of processor j cannot contain
address a. Thus, S 0 must be in <0mo-i2e according
to invariant 9a. Therefore, S 0 <1mo-i2e S =
v.overwrite, and the invariant still holds.
Invariant 8: Consider instruction I such that I <po-i2e S
and order(I, S). Since order(I, S), I can be a load,
fence, or store for a. If I is a load or fence, then
invariant 2 says that I is in <0mo-i2e , and thus I <1mo-i2e
S, i.e., the invariant holds. If I is a store for a, then
the predicate of R and invariant 4 imply that I is not
in the store buffer of processor i. Then invariant 3 says
that I must be in <0mo-i2e , and thus I <1mo-i2e S, i.e.,
the invariant holds.
Invariant 9: We need to consider the influence on both
loads that read S and loads that reads stores other than
S.
a) Consider load L for address a that reads from S,
rf -i2e
i.e., S −−−−→ L. Since S is not in <0mo-i2e before
R is performed, invariant 9a says that L must be
executed by processor i, S <po-i2e L, and there is
no store S 0 for a in the store buffer of processor
i such that S <po-i2e S 0 <po-i2e L. Now we show
rf -i2e
that both invariants still hold for S −−−−→ L.
– Invariant 9a: This is not affected because S is in
<1mo-i2e after R is performed.
– Invariant 9b: Since S <po-i2e L and S is the
youngest in <1mo-i2e , S satisfies the maxmo-i2e
formula. We prove the rest of this invariant by
contradiction, i.e., we assume there is store S 0
for a in the store buffer of processor i after R
is performed such that S 0 <po-i2e L. Note that
<po-i2e is not changed by R. The predicate of
R ensures that S is the oldest store for a in the
store buffer. Invariant 4 says that S <po-i2e S 0 .
Now we have S <po-i2e S 0 <po-i2e L (before R
is performed), contradicting with invariant 9a.
b) Consider load L for address a from processor j that
rf -i2e
reads from store S ∗ (6= S), i.e., S 6= S ∗ −−−−→ L.
Now we show that both invariants still hold for
rf -i2e
S ∗ −−−−→ L.
– Invariant 9a: This invariant cannot be affected,
because performing R can only remove a store
from a store buffer.
– Invariant 9b: This invariant can only be affected
when S ∗ is in <0mo-i2e . Since R can only remove
a store from a store buffer, the second half of
this invariant is not affected (i.e, no store S 00
in the store buffer and so on). We only need
to focus on the maxmo-i2e formula, i.e., S ∗ =
maxmo-i2e {S 0 | S 0 .addr = a ∧ (S 0 <po-i2e
L ∨ S 0 <mo-i2e L)}. Since L <1mo-i2e S, this
formula can only be affected when S <po-i2e L
and i = j. In this case, before R is performed, S
is in the store buffer of processor i, and S <po-i2e
L, and L reads from S ∗ 6= S. This contradicts
with invariant 9b which is assume to hold before
R is performed. Thus, the maxmo-i2e formula
cannot be affected either, i.e., the invariant holds.
6) WMM-Ld that reads from local store buffer: Assume R
executes a load L for address a in processor i, and L
reads from store S in the local store buffer. R appends
L to the <po-i2e of processor i, appends L to <mo-i2e ,
rf -i2e 1
and adds S −−−−→ L. Note that R does not change any
invalidation buffer or store buffer. Now we consider each
invariant.
• Invariants 1, 3, 4, 5, 7: These are not affected.
• Invariant 2: This still holds because L is added to
<mo-i2e .
• Invariant 6: We consider each invariant.
– Invariants 6a, 6b, 6c: These are not affected.
– Invariant 6d: L can only influence stale values for
a in the invalidation buffer of processor i. However,
since S is in the store buffer of processor i before
R is performed, invariant 5 says that the invalidation
buffer of processor i cannot contain address a.
Therefore this invariant still holds.
• Invariant 8: We consider instruction I such that
I <1po-i2e L and order(I, L). Since order(I, L), I can
only be a Reconcile fence or a load for a. In either
case, invariant 2 says that I is in <0mo-i2e . Since L is
appended to the end of <mo-i2e , I <1mo-i2e L, i.e., the
invariant still holds.
• Invariant 9: Since R does not change any store buffer
or any load/store already in <0mo-i2e , R cannot affect
this invariant for loads other than L. We only need to
rf -i2e 1
show that S −−−−→ L satisfies this invariant. Since
S is in the store buffer, invariant 3 says that S is
not in <mo-i2e . Therefore we only need to consider
invariant 9a. We prove by contradiction, i.e., we assume
there is store S 0 for a in the store buffer of processor i
and S <1po-i2e S 0 <1po-i2e L. Since R does not change
store buffer states, S and S 0 are both in the store buffer
before R is performed. We also have S <0po-i2e S 0
(because the only change in <po-i2e is to append L to
the end). According to the predicate of R, S should
be younger than S 0 , so S 0 <0po-i2e S (according to
invariant 4), contradicting with the previous conclusion.
Therefore, the invariant still holds.
7) WMM-Ld that reads from atomic memory: Assume R
executes a load L for address a in processor i, and
L reads from atomic memory m[a]. R appends L to
the <po-i2e of processor i, appends L to <mo-i2e , adds
rf -i2e 1
m[a].source −−−−→ L, and may remove stale values
from the invalidation buffer of processor i. Now we
consider each invariant.
• Invariants 1, 3, 4: These are not affected.
• Invariant 2: This still holds because L is added to
<mo-i2e .
• Invariants 5, 7: These are not affected because R only
remove values from an invalidation buffer.
• Invariant 6: We consider each invariant.
– Invariants 6a, 6b, 6c: These are not affected because
R only remove values from an invalidation buffer.
– Invariant 6d: L can only influence stale values for
a in the invalidation buffer of processor i. However,
R will remove address a from the the invalidation
buffer of processor i. Therefore this invariant still
holds.
• Invariant 8: We consider instruction I such that
I <1po-i2e L and order(I, L). Since order(I, L), I can
only be a Reconcile fence or a load for a. In either
case, invariant 2 says that I is in <0mo-i2e . Since L is
appended to the end of <mo-i2e , I <1mo-i2e L, i.e., the
invariant still holds.
• Invariant 9: Since R does not change any store buffer
or any load/store already in <0mo-i2e , R cannot affect
this invariant for loads other than L. We only need
rf -i2e 1
to show that m[a].source −−−−→ L satisfies this
invariant (m[a] is not changed before and after R is
performed). According to invariant 1, m[a].source is
the youngest store for a in <0mo-i2e . Therefore we
only need to consider invariant 9b. Since we also have
m[a].source <1po-i2e L, maxmo-i2e {S 0 | S 0 .addr =
a ∧ (S 0 <po-i2e L ∨ S 0 <mo-i2e L)} will return
m[a].source, i.e., the first half the invariant holds. The
predicate of R ensures that there is no store for a in
the store buffer of processor i, so the second half the
invariant also holds.
8) WMM-Ld that reads from the invalidation buffer: Assume
R executes a load L for address a in processor i, and L
reads from the stale value ha, vi in the local invalidation
buffer. R appends L to the <po-i2e of processor i, appends
rf -i2e 1
L to <mo-i2e , adds v.source −−−−→ L, and may remove
stale values from the invalidation buffer of processor i.
Now we consider each invariant.
• Invariants 1, 3, 4: These are not affected.
• Invariant 2: This still holds because L is added to
<mo-i2e .
• Invariants 5, 7: These are not affected because R can
only remove values from an invalidation buffer.
• Invariant 6: We consider each invariant.
– Invariants 6a, 6b, 6c: These are not affected because
R can only remove values from an invalidation buffer.
– Invariant 6d: Only stale values in the invalidation
buffer of processor i can be affected. Consider
stale value ha, v 0 i in the invalidation buffer of
processor i after R is performed. We need to
show that v.source <1mo-i2e v 0 .overwrite. Since
v 0 is not removed in R, v 0 must be either v or
younger than v in the invalidation buffer before
R is performed. According to invariant 7, either
v 0 .source = v.source or v.source <0mo-i2e v 0 .source.
Since v 0 .source <0mo-i2e v 0 .overwrite according to
invariant 6a, v.source <1mo-i2e v 0 .overwrite.
• Invariant 8: We consider instruction I such that
I <1po-i2e L and order(I, L), and we need to show
that I <1mo-i2e L. Since order(I, L), I can only be
a Reconcile fence or a load for a. If I is a Reconcile
fence, then invariant 6b says that I <0mo-i2e v.overwrite.
Since we insert L right before v.overwrite, we still have
I <1mo-i2e L. If I is a load for a, then invariant 6d
says that I <0mo-i2e v.overwrite, and thus we have
I <1mo-i2e L.
• Invariant 9: Since R does not change any store buffer
or any load/store already in <0mo-i2e , R cannot affect
this invariant for loads other than L. We only need to
rf -i2e 1
show that v.source −−−−→ L satisfies this invariant.
Since v.source is in <0mo-i2e , we only need to consider
invariant 9b. The predicate of R ensures that the store
buffer of processor i cannot contain address a, so the
second half of the invariant holds (i.e., there is no S 00
and so on).
Now we prove the first half of the invariant, i.e., consider maxmo-i2e {S 0 | S 0 .addr = a ∧ (S 0 <1po-i2e L ∨
S 0 <1mo-i2e L)}. First note that since v.source <0mo-i2e
v.overwrite, v.source <1mo-i2e L. Thus v.source is in
set {S 0 | S 0 .addr = a ∧ (S 0 <1po-i2e L ∨ S 0 <1mo-i2e
L)}. Consider any store S that is also in this set,
then S <1po-i2e L ∨ S <1mo-i2e L must be true.
If S <1po-i2e L, S is executed in processor i before
R is performed. Invariant 6c says that S <0mo-i2e
v.overwrite ⇒ S <1mo-i2e v.overwrite. If S <1mo-i2e L,
then S <1mo-i2e L <1mo-i2e v.overwrite. In either
case, we have S <mo-i2e v.overwrite. Since we have
proved invariant 6a holds after R is performed, either
S = v.source or S <mo-i2e v.overwrite. Therefore,
maxmo-i2e will return v.source.
It is easy to see that at the end of the I2 E execution
(of a program), there is no instruction to execute in each
processor and all store buffers are empty (i.e., all exected
loads stores and fences are in <mo-i2e ). At that time, if we
rf
define axiomatic relations <po , <mo , −→ as <po-i2e , <mo-i2e
rf -i2e
, −−−−→ respectively, then invariants 8 and 9b become
the Inst-Order and Ld-Val axioms respectively. That is,
rf -i2e
h<po-i2e , <mo-i2e , −−−−→i are the relations that satisfy the
WMM axioms and have the same program behavior as the
I2 E execution.
Theorem 2 (Completeness). WMM axiomatic model ⊆
WMM I2 E model.
Proof: The goal is that for any axiomatic relations
rf
h<po , <mo , −→i that satisfy the WMM axioms, we can
run the same program in the I2 E model and get the same
program behavior. We will devise an algorithm to operate
the I2 E model to get the same program behavior as in
rf
axiomatic relations h<po , <mo , −→i. In the algorithm, for
2
each instruction in the I E model, we need to find its
corresponding instruction in the <po in axiomatic relations.
Note that this mapping should be an one-to-one mapping, i.e.,
one instruction in the I2 E model will exactly correspond to
one instruction in the axiomatic relations and vice versa, so
we do not distinguish between the directions of the mapping.
The algorithm will create this mapping incrementally. Initially
(i.e., before the I2 E model performs any operation), for each
processor i, we only map the next instruction to execute
in processor i of the I2 E model to the oldest instruction in
the <po of processor i in the axiomatic relations. After the
algorithm starts to operate the I2 E model, whenever we have
executed an instruction in a processor in the I2 E model, we
map the next instruction to execute in that processor in the
I2 E model to the oldest unmapped instruction in the <po
of that processor in the axiomatic relations. The mapping
scheme obviously has the following two properties:
•
•
The k-th executed instruction in a processor in the I2 E
model is mapped to the k-th oldest instruction in the <po
of that processor in the axiomatic relations.
In the I2 E model, when a processor has executed x instructions, only the first x + 1 instructions (i.e., the executed
x instructions and the next instruction to execute) of that
processor are mapped to instructions in the axiomatic
relations.
Of course, later in the proof, we will show that the two corresponding instructions (one in the I2 E model and the other
in the axiomatic relations) have the same instruction types,
same load/store addresses (if they are memory accesses),
same store data (if they are stores), and same execution
results. In the following, we will assume the action of adding
new instruction mappings as an implicit procedure in the
algorithm, so we do not state it over and over again when
we explain the algorithm. When there is no ambiguity, we
do not distinguish an instruction in the I2 E model and an
instruction in the axiomatic relations if these two instructions
corresponds to each other (i.e., the algorithm has built the
mapping between them).
Now we give the details of the algorithm. The algorithm
begins with the I2 E model (in initial state), an empty set Z,
and a queue Q which contains all the memory and fence
instructions in <mo . The order of instructions in Q is the
same as <mo , i.e., the head of Q is the oldest instruction
in <mo . The instructions in Q and Z are all considered as
instructions in the axiomatic relations. In each step of the
algorithm, we perform one of the followings actions:
1) If the next instruction of some processor in the I2 E model
is a non-memory instruction, then we perform the WMMNm operation to execute it in the I2 E model.
2) Otherwise, if the next instruction of some processor in
the I2 E model is a store, then we perform the WMM-St
operation to execute that store in the I2 E model.
3) Otherwise, if the next instruction of some processor in
the I2 E model is mapped to a load L in set Z, then we
perform the WMM-Ld operation to execute L in the I2 E
model, and we remove L from Z.
4) Otherwise, we pop out instruction I from the head of Q
and process it in the following way:
a) If I is a store, then I must have been mapped to a
store in some store buffer (we will prove this), and
we perform the WMM-DeqSb operation to dequeue
I from the store buffer in the I2 E model.
b) If I is a Reconcile fence, then I must have been
mapped to the next instruction to execute in some
processor (we will prove this), and we perform the
WMM-Rec operation to execute I in the I2 E model.
c) If I is a Commit fence, then I must have been mapped
to the next instruction to execute in some processor
(we will prove this), and we perform the WMM-Com
operation to execute I in the I2 E model.
d) I must be a load in this case. If I has been mapped,
then it must be mapped to the next instruction to
execute in some processor in the I2 E model (we will
prove this), and we perform the WMM-Ld operation
to execute I in the I2 E model. Otherwise, we just
add I into set Z.
For proof purposes, we introduce the following ghost states
to the I2 E model:
•
•
Field source in atomic memory: For each address a, we
add state m[a].source to record the store that writes the
current memory value.
Fields source in invalidation buffer: For each stale value
ha, vi in an invalidation buffer, we add state v.source to
denote the store of this stale value.
These two fields are set when a WMM-DeqSb operation is
performed. Assume the WMM-DeqSb operation dequeues a
store S for address a. In this case, we update m[a].source
to be S. Let S0 be the original m[a].source before this
operation is performed. Then for each new stale value ha, vi
inserted into any invalidation buffer, we set v.source = S0 .
It is obvious that memory value m[a] is equal to the value
of m[a].source, and stale value v is equal to the value of
v.source.
For proof purposes, we define a function overwrite. For
each store S in <mo , overwrite(S) returns the store for the
same address such that
• S <mo overwrite(S), and
0
• there is no store S for the same address such that S <mo
0
S <mo overwrite(S).
In other words, overwrite(S) returns the store that overwrites
S in <mo . (overwrite(S) does not exist if S is the last store
for its address in <mo .)
Also for proof purposes, at each time in the algorithm,
we use Vi to represent the set of every store S in <mo that
satisfies all the following requirements:
1) The store buffer of processor i does not contain the
address of S.
2) overwrite(S) exists and overwrite(S) has been popped
from Q.
3) For each Reconcile fence F that has been executed by
processor i in the I2 E model, F <mo overwrite(S).
4) For each store S 0 for the same address that has been
executed by processor i in the I2 E model, S 0 <mo
overwrite(S).
5) For each load L for the same address that has been
rf
executed by processor i in the I2 E model, if store S 0 −→ L
0
in the axiomatic relations, then S <mo overwrite(S).
With the above definitions and new states, we introduce the
invariants of the algorithm. After each step of the algorithm,
we have the following invariants for the states of the I2 E
model, Z and Q:
1) For each processor i, the execution order of all executed
instructions in processor i in the I2 E model is a prefix
of the <po of processor i in the axiomatic relations.
2) The predicate of any operation performed in this step is
satisfied.
3) If we perform an operation to execute an instruction in
the I2 E model in this step, the operation is able to get
the same instruction result as that of the corresponding
instruction in the axiomatic relations.
4) The instruction type, load/stores address, and store data of
every mapped instruction in the I2 E model are the same
as those of the corresponding instruction in the axiomatic
relations.
5) All loads that have been executed in the I2 E model are
mapped exactly to all loads in <mo but not in Q or Z.
6) All fences that have been executed in processor i are
mapped exactly to all fences in <mo but not in Q.
7) All stores that have been executed and dequeued from
the store buffers in the I2 E model are mapped exactly to
all stores in <mo but not in Q.
8) For each address a, m[a].source in the I2 E model is
mapped to the youngest store for a, which has been
popped from Q, in <mo .
9) For each processor i, the store buffer of processor i
contains exactly every store that has been executed in the
I2 E model but still in Q.
10) For any two stores S1 and S2 for the same address in
the store buffer of any processor i in the I2 E model, if
S1 is older than S2 in the store buffer, then S1 <po S2 .
11) For any processor i and any address a, address a cannot
be present in the store buffer and invalidation buffer of
processor i at the same time.
12) For any processor i, for each store S in Vi , the invalidation
buffer of processor i contains an entry whose source field
is mapped to S.
13) For any stale value ha, vi in any invalidation buffer,
v.source has been mapped to a store in <mo , and
overwrite(v.source) exists, and overwrite(v.source) is not
in Q.
14) For any two stale values v1 and v2 for the same address
in the invalidation buffer of any processor i in the I2 E
model, if v1 is older than v2 in the invalidation buffer,
then v1 .source <mo v2 .source.
These invariants guarantee that the algorithm will operate
the I2 E model to produce the same program behavior as the
axiomatic model. We now prove inductively that all invariants
hold after each step of the algorithm, i.e., we assume all
invariants hold before the step. In case a state is changed in
this step, we use superscript 0 to denote the state before this
step (e.g., Q0 ) and use superscript 1 to denote the state after
this step (e.g., Q1 ). We consider which action is performed
in this step.
•
•
Action 1: We perform a WMM-Nm operation that executes
a non-memory instruction in the I2 E model. All the
invariants still hold after this step.
Action 2: We perform a WMM-St operation that executes
a store S for address a in processor i in the I2 E model.
We consider each invariant.
– Invariants 1, 2, 4: These invariants obviously hold.
– Invariants 3, 5, 6, 7, 8: These are not affected.
– Invariant 9: Note that S is mapped before this step. Since
S cannot be dequeued from the store buffer before this
step, invariant 7 says that S is still in Q. Thus, this
invariant holds.
– Invariant 10: Since S is the youngest store in store buffer
and invariant 1 holds after this step, this invariant also
holds.
– Invariant 11: Since the WMM-St operation removes
•
all stale values for a from the invalidation buffer of
processor i, this invariant still holds.
– Invariant 12: For any processor j (j 6= i), the action
in this step cannot change Vj or the invalidation buffer
of processor j. We only need to consider processor i.
The action in this step cannot introduce any new store
into Vi , i.e., Vi1 ⊆ Vi0 . Also notice that Vi1 does not
contain any store for a due to requirement 1. Since the
action in this step only removes values for address a
from the invalidation buffer of processor i, this invariant
still holds for i.
– Invariants 13, 14: These still hold, because we can only
remove values from the invalidation buffer in this step.
Action 3: We perform a WMM-Ld operation that executes
a load L in Z. (Note that L has been popped from Q
before.) We assume L is in processor i (both the axiomatic
relations and the I2 E model agree on this because of
the way we create mappings). We also assume that L has
rf
address a in the axiomatic relations, and that store S −→ L
in the axiomatic relations. According to invariant 4, L also
has load address a in the I2 E model. We first consider
several simple invariants:
– Invariants 1, 2, 5: These invariants obviously hold.
– Invariants 6, 7, 8, 9, 10: These are not affected.
– Invariant 11: Since the WMM-Ld operation does not
change store buffers and can only remove values from
the invalidation buffers, this invariant still holds.
– Invariants 13, 14: These still hold, because we can only
remove values from the invalidation buffer in this step.
We now consider the remaining invariants, i.e., 3, 4 and
12, according to the current state of Q (note that Q is not
changed in this step):
1) S is in Q: We show that the WMM-Ld operation can
read the value of S from the store buffer of processor
i in the I2 E model. We first show that S is in the store
buffer of processor i. Since L is not in Q, we have
L <mo S. According to the Ld-Val axiom, we know
S <po L, so S must have been executed. Since S is in
Q, invariant 7 says that S cannot be dequeued from the
store buffer, i.e., S is in the store buffer of processor i.
Now we prove that S is the youngest store for a in
the store buffer of processor i by contradiction, i.e.,
we assume there is another store S 0 for a which is in
the store buffer of processor i and is younger than S.
Invariant 10 says that S <po S 0 . Since S and S 0 are
stores for the same address, the Inst-Order axiom says
that S <mo S 0 . Since S 0 is in the store buffer, it is
executed before L. According to invariant 1, S 0 <po L.
rf
Then S −→ L contradicts with the Ld-Val axiom.
Now we can prove the invariants:
– Invariant 3: This holds because the WMM-Ld operation reads S from the store buffer.
– Invariant 4: This holds because invariant 3 holds after
this step.
– Invariant 12: The execution of L in this step cannot
introduce new stores into Vj for any j, i.e., Vj1 ⊆ Vj0 .
Since there is no change to any invalidation buffer
when WMM-Ld reads from the store buffer, this
invariant still holds.
2) S is not in Q but overwrite(S) is in Q: We show that
the WMM-Ld operation can read the value of S from
the atomic memory. Since S has been popped from Q
while overwrite(S) is not, S is the youngest store for
a in <mo that has been popped from Q. According to
invariant 8, the current m[a].source in the I2 E model is
S. To let WMM-Ld read m[a], we only need to show
that the store buffer of processor i does not contain any
store for a. We prove by contradiction, i.e., we assume
there is a store S 0 for a in the store buffer of processor
i. According to invariant 9, S 0 has been executed in
the I2 E model, and S 0 is still in Q. Thus, we have
S 0 <po L (according to invariant 1), and S <mo S 0 .
rf
Then S −→ L contradicts with the Ld-Val axiom.
Now we can prove the invariants:
– Invariant 3: This holds because the WMM-Ld operation reads S from the atomic memory m[a].
– Invariant 4: This holds because invariant 3 holds after
this step.
– Invariant 12: The execution of L in this step cannot
introduce new stores into Vj for any j, i.e., Vj1 ⊆ Vj0 .
Since there is no change to any invalidation buffer of
any processor other than i, we only need to consider
processor i. The WMM-Ld removes all values for
a from the invalidation buffer of processor i, so
the goal is to show that there is no store for a in
Vi1 . We prove by contradiction, i.e., assume there
is store S 0 for a in Vi1 . Requirement 5 for Vi says
that S <mo overwrite(S 0 ). Since overwrite(S) is in
Q, overwrite(S 0 ) is also in Q. This contradicts with
requirement 2. Therefore, there is no store for a in
Vi1 , and this invariant holds.
3) Both S and overwrite(S) are not in Q: We show that
the WMM-Ld operation can read the value of S from
the invalidation buffer of processor i. That is, we need
to show S ∈ Vi0 . We now prove that S satisfies all the
requirements for Vi0 :
– Requirement 1: We prove by contradiction, i.e., we
assume there is store S 0 for a in the store buffer of
processor i. Invariant 9 says that S 0 has been executed
but not in Q. Then we have S 0 <po L (invariant 1)
rf
and S <mo S 0 . Then S −→ L contradicts with the
Ld-Val axiom.
– Requirement 2: This satisfied because we assume
overwrite(S) is not in Q.
•
– Requirement 3: We prove by contradiction, i.e., we
assume that Reconcile fence F has been executed
by processor i, and overwrite(S) <mo F . Since F
is executed before L, invariant 1 says that F <po L.
Since order(F, L), the Inst-Order axiom says that
F <mo L. Now we have S <mo overwrite(S) <mo
rf
F <mo L. Thus, S −→ L contradicts with the Ld-Val
axiom.
– Requirement 4: We prove by contradiction, i.e.,
we assume that store S 0 for a has been executed
by processor i, and either S 0 = overwrite(S) or
overwrite(S) <mo S 0 . According to the definition
of overwrite, we have S <mo S 0 . Since S 0 has
been executed, invariant 1 says that S 0 <po L. Then
rf
S −→ L contradicts with the Ld-Val axiom.
– Requirement 5: We prove by contradiction, i.e., we
assume that store S 0 and load L0 are both for address
rf
a, L0 has been executed by processor i, S 0 −→ L0 ,
and either S 0 = overwrite(S) or overwrite(S) <mo
S 0 . According to the definition of overwrite, we have
S <mo S 0 . Since L0 has been executed, invariant 1
says that L0 <po L. Since order(L0 , L), the Instrf
Order axiom says that L0 <mo L. Since S 0 −→ L,
0
0
0
0
0
we have S <po L or S <mo L . Since L <po L
and L0 <mo L, we have S 0 <po L or S 0 <mo L.
rf
Since S <mo S 0 , S −→ L contradicts with the LdVal axiom.
Now we can prove the invariants:
– Invariant 3: This holds because the WMM-Ld operation reads S from the invalidation buffer of processor
i.
– Invariant 4: This holds because invariant 3 holds after
this step.
– Invariant 12: The execution of L in this step cannot
introduce new stores into Vj for any j, i.e., Vj1 ⊆ Vj0 .
Since there is no change to any invalidation buffer
of any processor other than i, we only need to
consider processor i. Assume the invalidation buffer
entry read by the WMM-Ld operation is ha, vi, and
v.source = S. The WMM-Ld rule removes any stale
value ha, v 0 i that is older than ha, vi from the invalidation buffer of processor i. The goal is to show that
v 0 .source cannot be in Vi1 . We prove by contradiction,
i.e., we assume that v 0 .source ∈ Vi1 . Since L has
been executed after this step, requirement 5 says that
S <mo overwrite(v 0 .source). Since v 0 is older than v
in the invalidation buffer before this step, invariant 14
says that v 0 .source <mo v.source = S. The above
two statements contradict with each other. Therefore,
this invariant still holds.
Action 4a: We pop a store S from the head of Q, and
we perform a WMM-DeqSb operation to dequeue S from
the store buffer. Assume that S is for address a, and
in processor i in the axiomatic relations. We first prove
that S has been mapped before this step. We prove by
contradiction, i.e., we assume S has not been mapped to
any instruction in the I2 E model before this step. Consider
the state right before this step. Let I be the next instruction
to execute in processor i in the I2 E model. We know I
is mapped and I <po S. The condition for performing
action 4a in this step says that I can only be a fence or
load, and we have order(I, S). According to the Inst-Order
axiom, I <mo S, so I has been popped from Q0 .
1) If I is a fence, since I is in <mo but not in Q0 ,
invariant 6 says that I must be executed, contradicting
our assumption that I is the next instruction to execute.
2) If I is a load, since I is not executed, and I is in <mo ,
and I is not in Q0 , invariant 5 says that I must be
in Z 0 . Then this algorithm step should use action 3
instead of action 4a.
Due to the contradictions, we know S must have been
mapped. Note that the next instruction to execute in
processor i cannot be store, because otherwise this step
will use action 2. According to invariant 4, S cannot be
mapped to the next instruction to execution in processor
i. Therefore S must have been executed in processor i in
the I2 E model before this step.
Also according to invariant 4, the address and data of S
in the I2 E model are the same as those in the axiomatic
relations. Now we consider each invariant.
– Invariants 1, 4, 7, 9, 10: These invariants obviously hold.
– Invariants 3, 5, 6: These are not affected.
– Invariant 2: We prove by contradiction, i.e., we assume
there is a store S 0 for a younger than S 0 in the store
buffer of processor i (before this step). According to
invariant 9, S 0 is in Q. Since S is the head of Q0 ,
S <mo S 0 . According to invariant 10, S 0 <po S. Since
order(S 0 , S), S 0 <mo S, contradicting with previous
statement. Thus, the predicate of the WMM-DeqSb
operation is satisfied, and the invariant holds.
– Invariant 8: S is the youngest instruction in <mo that
has been popped from Q, and m[a].source is updated
to S. Thus, this invariant still holds.
– Invariant 11: Since the WMM-DeqSb operation will not
insert the stale value into an invalidation buffer of a
processor if the store buffer of that processor contains
the same address, this invariant still holds.
– Invariant 12: For any processor j, the action this step
will not remove stores from Vj but may introduce new
stores to Vj , i.e., Vj0 ⊆ Vj1 . We consider the following
two types of processors.
1) Processor i: We show that Vi0 = Vi1 . We prove
by contradiction, i.e., we assume there is store S 0
such that S 0 ∈ Vi1 but S 0 ∈
/ Vi0 . Since S 0 satisfies
requirements 3, 4, 5 after this step, it also satisfies
these three requirements before this step. Then S 0
must fail to meet at least one of requirements 1 and
2 before this step.
a) If S 0 does not meet requirement 1 before this step,
then S 0 .addr is in the store buffer of processor i
before this step and S 0 .addr is not in this store
buffer after this step. Thus, S 0 .addr must be a.
Since S 0 meets requirement 4 before this step and
S has been executed by processor i before this
step, we know S <mo overwrite(S 0 ). Since S 0
meets requirement 2 after this step, overwrite(S 0 )
is not in Q1 . Since Q1 is derived by popping the
oldest store form Q0 , we know S is not in Q0 .
Since S is in the store buffer before this step,
this contradicts invariant 9. Therefore this case
is impossible.
b) If S 0 does not meet requirement 2 before this step,
then overwrite(S 0 ) is in Q0 but not in Q1 . Then
overwrite(S 0 ) = S. Since S has been executed
by processor i, S 0 will fail to meet requirement 4
after this step. This contradicts with S 0 ∈ Vi1 , so
this case is impossible either.
Now we have proved that Vi0 = Vi1 . Since the WMMDeqSb operation does not change the invalidation
buffer of processor i, this invariant holds for processor i.
2) Processor j (6= i): We consider any store S 0 such
that S 0 ∈ Vj1 but S 0 ∈
/ Vj0 . Since S 0 satisfies
requirements 1, 3, 4, 5 after this step, it also satisfies
these four requirements before this step. Then S 0
must fail to meet requirement requirement 2 before
this step, i.e., overwrite(S 0 ) is in Q0 but not in Q1 .
Then overwrite(S 0 ) = S. According to invariant 8,
we know S 0 is m[a].source. Consider the following
two cases.
a) The store buffer of processor j contains address
a: Since S 0 cannot meet requirement 1, Vj1 = Vj0 .
Since WMM-DeqSb cannot remove any value
from the invalidation buffer of processor j, this
invariant holds.
b) The store buffer of processor j does not contain
address a: In this case, the WMM-DeqSb operation will insert stale value ha, m[a]0 i into the
invalidation buffer of processor j, so the invariant
still holds.
– Invariant 13: The source field of all the newly inserted
stale values in this step are equal to m[a].source0 .
According to invariant 8, m[a].source0 is the youngest
store for a in <mo that is not in Q0 . Since S is the head
of Q0 , we know overwrite(m[a].source0 ) = S. Since S
is not in Q1 , this invariant still holds.
– Invariant 14: Assume ha, vi is the new stale value
inserted into the invalidation buffer of processor j in
•
•
this step. We need to show that for any stale value
ha, v 0 i that is in this invalidation buffer before this
step, v 0 .source <mo v.source. According to invariant 13,
overwrite(v 0 .source) is not in Q0 . Since v.source =
m[a].source0 , according to invariant 8, v.source is the
youngest store for a in <mo that is not in Q0 . Therefore
v 0 .source <mo v.source.
Action 4b: We pop a Reconcile fence F from Q, and
perform a WMM-Rec operation to execute it in the I2 E
model. Assume F is in processor i in the axiomatic
relations. We first prove that F has been mapped before
this step. We prove by contradiction, i.e., we assume F
is not mapped before this step. Consider the state right
before this step. Let I be the next instruction to execute in
processor i in the I2 E model. We know I is mapped and
I <po F . The condition for performing action 4b in this
step says that I can only be a fence or load, so we have
order(I, F ). According to the Inst-Order axiom, I <mo F ,
so I has been popped from Q0 .
1) If I is a fence, since I is in <mo but not in Q0 ,
invariant 6 says that I must be executed, contradicting
our assumption that I is the next instruction to execute.
2) If I is a load, since I is not executed, and I is in <mo ,
and Ik is not in Q0 , invariant 5 says that I must be
in Z 0 . Then this algorithm step should use action 3
instead of action 4a.
Due to the contradictions, we know F must have been
mapped before this step. According to invariant 6, since F
is in Q0 , F must have not been executed in the I2 E model.
Thus, F is mapped to the next instruction to execute in
processor i in the I2 E model.
Now we consider each invariant:
– Invariants 1, 2, 4, 6: These obviously hold.
– Invariants 3, 5, 7, 8, 9, 10: These are not affected.
– Invariants 11, 13, 14: These invariants hold, because
the invalidation buffer of any processor j (6= i) is not
changed, and the invalidation buffer of processor i is
empty after this step.
– Invariant 12: For any processor j (6= i), Vj1 = Vj0 and
the invalidation buffer of processor j is not changed in
this step. Thus, this invariant holds for any processor j
(6= i). We now consider processor i. The invalidation
buffer of processor i is empty after this step, so we need
to show that Vi1 is empty. We prove by contradiction, i.e.,
we assume there is a store S ∈ Vi1 . Since F has been
executed in processor i after this step, requirement 3
says that F <mo overwrite(S). Since F is the head of
Q0 , overwrite(S) must be in Q1 . Then S fails to meet
requirement 2 after this step, contradicting with S ∈ Vi1 .
Therefore Vi1 is empty, and this invariant also holds for
processor i.
Action 4c: We pop a Commit fence F from Q, and perform
a WMM-Com operation to execute it in the I2 E model.
•
Assume F is in processor i in the axiomatic relations.
Using the same argument as in the previous case, we can
prove that F is mapped to the next instruction to execute
in processor i in the I2 E model (before this step). Now
we consider each invariant:
– Invariants 1, 4, 6: These obviously hold.
– Invariants 3, 5, 7, 8, 9, 10, 11, 12, 13, 14: These are
not affected.
– Invariants 2: We prove by contradiction, i.e., we assume
there is store S in the store buffer of processor i before
this step. According to invariant 9, S has been executed
in processor i and is in Q0 . Thus, we have S <po F .
Since order(S, F ), the Ld-Val axiom says that S <mo F .
Then F is not the head of Q0 , contradicting with the
fact that we pop F from the head of Q0 .
Action 4d: We pop a load L from Q. Assume that L
is for address a and is in processor i in the axiomatic
relations. If we add L to Z, then all invariants obviously
hold. We only need to consider the case that we perform
a WMM-Ld operation to execute L in the I2 E model. In
this case, L is mapped before this step. Since L is in
Q0 , according to invariant 5, L must be mapped to an
unexecuted instruction in the I2 E model. That is, L is
mapped to the next instruction to execute in processor i
in the I2 E model. Invariant 4 ensures that L has the same
load address in the I2 E model. We first consider several
simple invariants:
– Invariants 1, 2, 5: These invariants obviously hold.
– Invariants 6, 7, 8, 9, 10: These are not affected.
– Invariant 11: Since the WMM-Ld operation does not
change store buffers and can only remove values from
the invalidation buffers, this invariant still holds.
– Invariants 13, 14: These still hold, because we can only
remove values from the invalidation buffer in this step.
rf
Assume store S −→ L in the axiomatic relations. We prove
the remaining invariants (i.e., 3, 4 and 12) according to
the position of S in <mo .
1) L <mo S: We show that the WMM-Ld can read S
from the store buffer of processor i in the I2 E model.
The Ld-Val axiom says that S <po L. Then S must
have been executed in processor i in the I2 E model
according to invariant 1. Since S is in Q0 , invariant 9
ensures that S is in the store buffer of processor i
before this step.
To let WMM-Ld read S from the store buffer, we
now only need to prove that S is the youngest store
for a in the store buffer of processor i. We prove by
contradiction, i.e., we assume there is another store S 0
for a which is in the store buffer of processor i and
is younger than S. Invariant 10 says that S <po S 0 .
Since S and S 0 are stores for the same address, the
Inst-Order axiom says that S <mo S 0 . Since S 0 is in
the store buffer, it is executed before L. According to
rf
invariant 1, S 0 <po L. Then S −→ L contradicts with
the Ld-Val axiom.
Now we can prove the invariants:
– Invariant 3: This holds because the WMM-Ld operation reads S from the store buffer.
– Invariant 4: This holds because invariant 3 holds after
this step.
– Invariant 12: The execution of L in this step cannot
introduce new stores into Vj for any j, i.e., Vj1 ⊆ Vj0 .
Since there is no change to any invalidation buffer
when WMM-Ld reads from the store buffer, this
invariant still holds.
2) S <mo L: We show that the WMM-Ld operation can
read the value of S from the atomic memory. Since L
is the head of Q0 , S is not in Q0 . According to the
Ld-Val axiom, there cannot be any store for a between
S and L in <mo . Thus, S is the youngest store for a
in <mo that has been popped from Q. According to
invariant 8, the current m[a].source in the I2 E model
is S.
To let WMM-Ld read m[a], we only need to show
that the store buffer of processor i does not contain
any store for a. We prove by contradiction, i.e., we
assume there is a store S 0 for a in the store buffer of
processor i. According to invariant 9, S 0 has executed
in the I2 E model, and S 0 is in Q0 . Thus, we have
S 0 <po L (according to invariant 1), and S <mo S 0 .
rf
Then S −→ L contradicts with the Ld-Val axiom.
Now we can prove the invariants:
– Invariant 3: This holds because the WMM-Ld operation reads S from the atomic memory m[a].
– Invariant 4: This holds because invariant 3 holds after
this step.
– Invariant 12: The execution of L in this step cannot
introduce new stores into Vj for any j, i.e., Vj1 ⊆ Vj0 .
Since there is no change to any invalidation buffer of
any processor other than i, we only need to consider
processor i. The WMM-Ld removes all values for
a from the invalidation buffer of processor i, so
the goal is to show that there is no store for a in
Vi1 . We prove by contradiction, i.e., assume there
is store S 0 for a in Vi1 . Requirement 5 for Vi says
that S <mo overwrite(S 0 ). Since S is the youngest
store for a that is in <mo but not Q0 , overwrite(S 0 )
must be in Q0 . This contradicts with requirement 2.
Therefore, there is no store for a in Vi1 , and this
invariant holds.
By combining Theorems 1 and 2, we prove the equivalence
between the I2 E model and the axiomatic model of WMM.
Theorem 3 (Equivalence). WMM I2 E model ≡ WMM
axiomatic model.
R EFERENCES
[1] L. Lamport, “How to make a multiprocessor computer that
correctly executes multiprocess programs,” Computers, IEEE
Transactions on, vol. 100, no. 9, pp. 690–691, 1979.
[2] P. Sewell, S. Sarkar, S. Owens, F. Z. Nardelli, and M. O.
Myreen, “x86-tso: a rigorous and usable programmer’s model
for x86 multiprocessors,” Communications of the ACM, vol. 53,
no. 7, pp. 89–97, 2010.
[3] SPARC International, Inc., The SPARC Architecture Manual:
Version 8. Prentice-Hall, Inc., 1992.
[4] D. L. Weaver and T. Gremond, The SPARC architecture
manual (Version 9). PTR Prentice Hall Englewood Cliffs,
NJ 07632, 1994.
[5] Alpha Architecture Handbook, Version 4. Compaq Computer
Corporation, 1998.
[6] K. Gharachorloo, D. Lenoski, J. Laudon, P. Gibbons, A. Gupta,
and J. Hennessy, “Memory consistency and event ordering in
scalable shared-memory multiprocessors,” in Proceedings of
the 17th International Symposium on Computer Architecture.
ACM, 1990, pp. 15–26.
[7] S. Sarkar, P. Sewell, J. Alglave, L. Maranget, and D. Williams,
“Understanding power multiprocessors,” in ACM SIGPLAN
Notices, vol. 46, no. 6. ACM, 2011, pp. 175–186.
[8] S. Flur, K. E. Gray, C. Pulte, S. Sarkar, A. Sezgin, L. Maranget,
W. Deacon, and P. Sewell, “Modelling the armv8 architecture,
operationally: Concurrency and isa,” in Proceedings of
the 43rd Annual ACM SIGPLAN-SIGACT Symposium on
Principles of Programming Languages, ser. POPL 2016.
New York, NY, USA: ACM, 2016, pp. 608–621. [Online].
Available: http://doi.acm.org/10.1145/2837614.2837615
[9] S. Mador-Haim, L. Maranget, S. Sarkar, K. Memarian,
J. Alglave, S. Owens, R. Alur, M. M. Martin, P. Sewell,
and D. Williams, “An axiomatic memory model for power
multiprocessors,” in Computer Aided Verification. Springer,
2012, pp. 495–512.
[10] J. Alglave, L. Maranget, and M. Tautschnig, “Herding cats:
Modelling, simulation, testing, and data mining for weak
memory,” ACM Transactions on Programming Languages and
Systems (TOPLAS), vol. 36, no. 2, p. 7, 2014.
[11] IBM, Power ISA, Version 2.07, 2013.
[12] J. Alglave, A. Fox, S. Ishtiaq, M. O. Myreen, S. Sarkar,
P. Sewell, and F. Z. Nardelli, “The semantics of power and
arm multiprocessor machine code,” in Proceedings of the 4th
workshop on Declarative aspects of multicore programming.
ACM, 2009, pp. 13–24.
[13] J. Alglave and L. Maranget, Computer Aided Verification:
23rd International Conference, CAV 2011, Snowbird, UT,
USA, July 14-20, 2011. Proceedings. Berlin, Heidelberg:
Springer Berlin Heidelberg, 2011, ch. Stability in Weak
Memory Models, pp. 50–66. [Online]. Available: http:
//dx.doi.org/10.1007/978-3-642-22110-1 6
[14] J. Alglave, “A formal hierarchy of weak memory models,”
Formal Methods in System Design, vol. 41, no. 2, pp. 178–
210, 2012.
[15] S. Sarkar, K. Memarian, S. Owens, M. Batty, P. Sewell,
L. Maranget, J. Alglave, and D. Williams, “Synchronising
c/c++ and power,” in ACM SIGPLAN Notices, vol. 47, no. 6.
ACM, 2012, pp. 311–322.
[16] J. Alglave, D. Kroening, V. Nimal, and M. Tautschnig,
“Software verification for weak memory via program transformation,” in Programming Languages and Systems. Springer,
2013, pp. 512–532.
[17] “The risc-v instruction set,” https://riscv.org/.
[18] A. Waterman, Y. Lee, D. A. Patterson, and K. Asanovi,
“The risc-v instruction set manual. volume 1: User-level isa,
version 2.1,” Technical Report UCB/EECS-2016-118, EECS
Department, University of California, Berkeley, May 2014.
[Online]. Available: https://people.eecs.berkeley.edu/∼krste/
papers/riscv-spec-v2.1.pdf
[19] D. Lustig, C. Trippel, M. Pellauer, and M. Martonosi, “Armor:
defending against memory consistency model mismatches
in heterogeneous architectures,” in Proceedings of the 42nd
Annual International Symposium on Computer Architecture.
ACM, 2015, pp. 388–400.
[20] M. H. Lipasti, C. B. Wilkerson, and J. P. Shen, “Value locality
and load value prediction,” ACM SIGOPS Operating Systems
Review, vol. 30, no. 5, pp. 138–147, 1996.
[21] M. M. K. Martin, D. J. Sorin, H. W. Cain, M. D.
Hill, and M. H. Lipasti, “Correctly implementing value
prediction in microprocessors that support multithreading
or multiprocessing,” in Proceedings of the 34th Annual
ACM/IEEE International Symposium on Microarchitecture,
ser. MICRO 34. Washington, DC, USA: IEEE Computer
Society, 2001, pp. 328–337. [Online]. Available: http:
//dl.acm.org/citation.cfm?id=563998.564039
[22] W. J. Ghandour, H. Akkary, and W. Masri, “The potential of
using dynamic information flow analysis in data value prediction,” in Proceedings of the 19th international conference
on Parallel architectures and compilation techniques. ACM,
2010, pp. 431–442.
[23] A. Perais and A. Seznec, “Eole: Paving the way for an effective
implementation of value prediction,” in Computer Architecture
(ISCA), 2014 ACM/IEEE 41st International Symposium on.
IEEE, 2014, pp. 481–492.
[24] A. Perais and A. Seznec, “Practical data value speculation
for future high-end processors,” in High Performance Computer Architecture (HPCA), 2014 IEEE 20th International
Symposium on. IEEE, 2014, pp. 428–439.
[25] S. Owens, S. Sarkar, and P. Sewell, “A better x86 memory
model: x86-tso,” in Theorem Proving in Higher Order Logics.
Springer, 2009, pp. 391–407.
[26] J. Kang, C.-K. Hur, O. Lahav, V. Vafeiadis, and D. Dreyer,
“A promising semantics for relaxed-memory concurrency,”
in Proceedings of the 44th ACM SIGPLAN Symposium on
Principles of Programming Languages, ser. POPL 2017.
New York, NY, USA: ACM, 2017, pp. 175–189. [Online].
Available: http://doi.acm.org/10.1145/3009837.3009850
[27] H.-J. Boehm and B. Demsky, “Outlawing ghosts: Avoiding
out-of-thin-air results,” in Proceedings of the Workshop on
Memory Systems Performance and Correctness, ser. MSPC
’14. New York, NY, USA: ACM, 2014, pp. 7:1–7:6. [Online].
Available: http://doi.acm.org/10.1145/2618128.2618134
[28] S. V. Adve and M. D. Hill, “Weak ordering a new definition,”
in ACM SIGARCH Computer Architecture News, vol. 18, no.
2SI. ACM, 1990, pp. 2–14.
[29] K. Gharachorloo, A. Gupta, and J. L. Hennessy, “Two
techniques to enhance the performance of memory consistency
models,” in Proceedings of the 1991 International Conference
on Parallel Processing, 1991, pp. 355–364.
[30] P. Ranganathan, V. S. Pai, and S. V. Adve, “Using speculative
retirement and larger instruction windows to narrow the
performance gap between memory consistency models,” in
Proceedings of the ninth annual ACM symposium on Parallel
algorithms and architectures. ACM, 1997, pp. 199–210.
[31] C. Guiady, B. Falsafi, and T. N. Vijaykumar, “Is sc+ ilp=
rc?” in Computer Architecture, 1999. Proceedings of the 26th
International Symposium on. IEEE, 1999, pp. 162–171.
[32] C. Gniady and B. Falsafi, “Speculative sequential consistency
with little custom storage,” in Parallel Architectures and Compilation Techniques, 2002. Proceedings. 2002 International
Conference on. IEEE, 2002, pp. 179–188.
[33] L. Ceze, J. Tuck, P. Montesinos, and J. Torrellas, “Bulksc: bulk
enforcement of sequential consistency,” in ACM SIGARCH
Computer Architecture News, vol. 35, no. 2. ACM, 2007, pp.
278–289.
[34] T. F. Wenisch, A. Ailamaki, B. Falsafi, and A. Moshovos,
“Mechanisms for store-wait-free multiprocessors,” in ACM
SIGARCH Computer Architecture News, vol. 35, no. 2. ACM,
2007, pp. 266–277.
[35] C. Blundell, M. M. Martin, and T. F. Wenisch, “Invisifence:
performance-transparent memory ordering in conventional
multiprocessors,” in ACM SIGARCH Computer Architecture
News, vol. 37, no. 3. ACM, 2009, pp. 233–244.
[36] A. Singh, S. Narayanasamy, D. Marino, T. Millstein, and
M. Musuvathi, “End-to-end sequential consistency,” in ACM
SIGARCH Computer Architecture News, vol. 40, no. 3. IEEE
Computer Society, 2012, pp. 524–535.
[39] S. Sarkar, P. Sewell, F. Z. Nardelli, S. Owens, T. Ridge,
T. Braibant, M. O. Myreen, and J. Alglave, “The semantics
of x86-cc multiprocessor machine code,” SIGPLAN Not.,
vol. 44, no. 1, pp. 379–391, Jan. 2009. [Online]. Available:
http://doi.acm.org/10.1145/1594834.1480929
[40] J. R. Goodman, Cache consistency and sequential consistency. University of Wisconsin-Madison, Computer Sciences
Department, 1991.
[41] M. Dubois, C. Scheurich, and F. Briggs, “Memory access
buffering in multiprocessors,” in ACM SIGARCH Computer
Architecture News, vol. 14, no. 2. IEEE Computer Society
Press, 1986, pp. 434–442.
[42] X. Shen, Arvind, and L. Rudolph, “Commit-reconcile and
fences (crf): A new memory model for architects and compiler
writers,” in Computer Architecture, 1999. Proceedings of the
26th International Symposium on. IEEE, 1999, pp. 150–161.
[43] Arvind and J.-W. Maessen, “Memory model = instruction
reordering + store atomicity,” in ACM SIGARCH Computer
Architecture News, vol. 34, no. 2. IEEE Computer Society,
2006, pp. 29–40.
[44] ARM, ARM Architecture Reference Manual, ARMv7-A and
ARMv7-R edition, 2013.
[45] S. V. Adve and K. Gharachorloo, “Shared memory consistency
models: A tutorial,” computer, vol. 29, no. 12, pp. 66–76, 1996.
[46] L. Maranget, S. Sarkar, and P. Sewell, “A tutorial introduction
to the arm and power relaxed memory models,” http://www.
cl.cam.ac.uk/∼pes20/ppc-supplemental/test7.pdf, 2012.
[47] R. Smith, Ed., Working Draft, Standard for Programming
Language C++. http://open-std.org/JTC1/SC22/WG21/docs/
papers/2015/n4527.pdf, May 2015.
[48] H.-J. Boehm and S. V. Adve, “Foundations of the c++
concurrency memory model,” in ACM SIGPLAN Notices,
vol. 43, no. 6. ACM, 2008, pp. 68–78.
[49] M. Batty, S. Owens, S. Sarkar, P. Sewell, and T. Weber,
“Mathematizing c++ concurrency,” in ACM SIGPLAN Notices,
vol. 46, no. 1. ACM, 2011, pp. 55–66.
[50] M. Batty, A. F. Donaldson, and J. Wickerson, “Overhauling
sc atomics in c11 and opencl,” SIGPLAN Not., vol. 51,
no. 1, pp. 634–648, Jan. 2016. [Online]. Available:
http://doi.acm.org/10.1145/2914770.2837637
[37] C. Lin, V. Nagarajan, R. Gupta, and B. Rajaram, “Efficient sequential consistency via conflict ordering,” in ACM SIGARCH
Computer Architecture News, vol. 40, no. 1. ACM, 2012, pp.
273–286.
[51] J. Manson, W. Pugh, and S. V. Adve, “The java
memory model,” in Proceedings of the 32Nd ACM
SIGPLAN-SIGACT Symposium on Principles of Programming
Languages, ser. POPL ’05. New York, NY, USA:
ACM, 2005, pp. 378–391. [Online]. Available: http:
//doi.acm.org/10.1145/1040305.1040336
[38] D. Gope and M. H. Lipasti, “Atomic sc for simple in-order
processors,” in High Performance Computer Architecture
(HPCA), 2014 IEEE 20th International Symposium on. IEEE,
2014, pp. 404–415.
[52] P. Cenciarelli, A. Knapp, and E. Sibilio, “The java memory model: Operationally, denotationally, axiomatically,” in
Programming Languages and Systems. Springer, 2007, pp.
331–346.
[53] J.-W. Maessen, Arvind, and X. Shen, “Improving the java
memory model using crf,” ACM SIGPLAN Notices, vol. 35,
no. 10, pp. 1–12, 2000.
[54] “Wwc+addrs test result in power processors,” http://www.cl.
cam.ac.uk/∼pes20/ppc-supplemental/ppc051.html#toc11.
[55] J. F. Cantin, M. H. Lipasti, and J. E. Smith, “The complexity
of verifying memory coherence,” in Proceedings of the
fifteenth annual ACM symposium on Parallel algorithms and
architectures. ACM, 2003, pp. 254–255.
[56] M. Batty, K. Memarian, S. Owens, S. Sarkar, and P. Sewell,
“Clarifying and compiling c/c++ concurrency: from c++ 11 to
power,” in ACM SIGPLAN Notices, vol. 47, no. 1. ACM,
2012, pp. 509–520.
[57] M. Vijayaraghavan, A. Chlipala, Arvind, and N. Dave,
Computer Aided Verification: 27th International Conference,
CAV 2015, San Francisco, CA, USA, July 18-24, 2015,
Proceedings, Part II. Cham: Springer International Publishing,
2015, ch. Modular Deductive Verification of Multiprocessor
Hardware Designs, pp. 109–127. [Online]. Available:
http://dx.doi.org/10.1007/978-3-319-21668-3 7
[58] D. J. Sorin, M. D. Hill, and D. A. Wood, “A primer on
memory consistency and cache coherence,” Synthesis Lectures
on Computer Architecture, vol. 6, no. 3, pp. 1–212, 2011.
[59] S. C. Woo, M. Ohara, E. Torrie, J. P. Singh, and A. Gupta,
“The splash-2 programs: Characterization and methodological
considerations,” in ACM SIGARCH Computer Architecture
News, vol. 23, no. 2. ACM, 1995, pp. 24–36.
[60] “Splash-2x benchmarks,”
parsec3-doc.htm#splash2x.
http://parsec.cs.princeton.edu/
[61] E. K. Ardestani and J. Renau, “Esesc: A fast multicore
simulator using time-based sampling,” in High Performance
Computer Architecture (HPCA2013), 2013 IEEE 19th International Symposium on. IEEE, 2013, pp. 448–459.
| 6cs.PL
|
An Energy Efficient Spectrum Sensing in Cognitive
Radio Wireless Sensor Networks
Atchutananda Surampudi and Krishnamoorthy Kalimuthu
arXiv:1711.09255v1 [cs.IT] 25 Nov 2017
Department of Electronics and Communication Engineering,
SRM University,
Chennai, India 603203.
[email protected], [email protected]
Abstract—The cognitive radio wireless sensor networks
have become an integral part of communicating spectrum
information to the fusion centre, in a cooperative spectrum
sensing environment. A group of battery operated sensors or
nodes, sensing information about spectrum availability in the
radio links, needs an energy efficient strategy to pass on the
binary information. A proper routing method through intelligent
decision making can be a good solution. In this paper, an energy
efficient routing protocol has been introduced and a performance
comparison is made with the existing system in terms of it’s
energy consumption. The proposed routing technique considers
both uniform and non-uniform clustering strategies and proves
to be an energy efficient method of communicaton.
Index Terms: Cluster head, clustering, minimum spanning tree,
prims algorithm, routing.
I. I NTRODUCTION
The cognitive radio wireless sensor network (CR-WSN) is
a group of sensors, performing the task of sensing spectrum
availability information in a cognitive radio (CR) environment
[1]. The spectrum sensing can happen in a cooperative or
a non-cooperative way. In a cooperative spectrum sensing
these sensor nodes communicate their respective information
to an information sink called the fusion centre (FC), by
forming an interconnected wireless network [2]. At the FC
the data is aggregated and further decision for utilization of
spectrum is taken. So, once a wireless network is formed,
the one bit information present with each node is routed
along a wireless route with the help of routing protocols and
associated algorithms defining the protocols. There have been
several routing protocols proposed for routing information
along these sensor networks [3], [4], [5]. In this paper, the
focus has been made on the routing strategy used by sensors
which sense spectrum availability information in cognitive
radio systems.
The existing system of sending the binary sensing
information is done by hierarchical clustering using the low
energy adaptive clustering hierarchy (LEACH) protocol and
there have been several developments on the same [6], [7]. In
such systems, energy efficiency has been the most important
key point of the network routing design. Here we introduce
a novel routing protocol which combines with hierarchical
clustering and inter cluster routing to communicate the
information to the FC in an energy efficient manner. This
research puts light on the energy demand for the sensors
to send the sensed information to the FC directly or along
a certain intelligently decided path. A greedy algorithm
called the prim’s algorithm [8] is used to form a low energy
minimum spanning tree at every round of cluster formation.
The rest depends on the elected cluster head to take an
energy efficient decision to send the sensed information. In
this analysis, both uniform and non uniform size of clustering
have been considered.
This paper describes the proposed methodology of the new
routing protocol and compares the energy parameter with
that of the existing protocol, the LEACH. This paper is
organized as follows: In section II, the clustering strategy
of existing system is described and the proposed method of
communication protocol is analyzed in section III. Section IV
discusses the results and comparisons made with the existing
system. Finally the paper concludes with section V.
II. E XISTING S YSTEM
A group of Wireless Sensors are placed for spectrum sensing
along with the FC, where the data is aggregated. These nodes
are further differentiated into advanced and normal nodes and
are powered by their residual energy. A probability of selection
of Cluster Head (CH) is pre-specified. The system undergoes
a certain number of rounds and selection of CH is based on
distance and energy parameters of each node from the FC.
From [9], the condition for electing the CH for each round is
given as
T (s) =
p
1 ;s
1−p(rmod p
)
∈ G,
(1)
0, otherwise,
where r is the current round of clustering, G is the set of nodes
that have not been CHs in the last 1/p rounds. Each sensor
node decides independently of other senor nodes whether it
will claim to be a CH or not, by picking a random value
between 0 and 1 and comparing it with a threshold T(s) based
on a user-specified probability p. After the CH is selected, it
aggregates information from its cluster of nodes and calculates
its distance from the fusion centre. This distance is used to
calculate the cost to send the one bit information to the fusion
centre soon after election and the sensing information is sent.
In this method, all elected CHs transmit the aggregated data to
the FC directly. If the CH is far away to the FC, then it needs
more energy to transmit. So, the overall lifetime becomes very
short and a considerable number of nodes die after certain
rounds.
III. P ROPOSED M ETHOD
In the proposed method an energy efficient routing protocol
has been introduced. As a further extension of the previously
described method, the proposed method introduces the
process of intelligent decision making and forwarding the
information along a minimum energy route by forming a
minimum spanning tree (MST) using a greedy algorithm
called the Prim’s algorithm. A greedy algorithm formulates
a certain path in a stepwise manner by greedily choosing
the next step with respect to the previous one, according to
the least value of the available metric. The Prim’s algorithm
finds an MST in a connected, weighted graph. It begins with
a vertex and no edges and then applies the greedy rule,
Add an edge of minimum weight that has one vertex in the
current tree and the other not in the current tree.
Using this Prims algorithm, the proposed communication is
made as follows
• Node is elected as a CH using (1). But it doesn’t directly
send the sensed information to the FC soon after election.
• At once all the required ’M’ number of CH’s are elected
for the ’rth’ round and each CH now calculates it’s
Cartesian distance from the fusion centre and distances
from every other node.
• Using this information an adjacency matrix A(i,j) is
formed, with distance as the metric.
• MST is formed using the Prims algorithm. This connects
all the elected ’M’ number of CH’s along a minimum
energy route in that round. This tree varies dynamically
with every round.
Now, as a development with respect to the earlier protocol,
each CH maintains a sensing table which provides a comprehensive information about the one bit information sensed by
itself and all other CH’s that are present. So, the CH sends
’M’ bit information rather than one bit along an intelligently
decided path. It calculates the respective cost (C) to the fusion
centre and to the nearest CH along the minimum spanning tree,
using the previously calculated distances as
(Et + Ed )M + Ef M d2 ; d ≤ do ,
C=
(2)
(Et + Ed )M + Em M d4 ; d > do .
where Et is energy spent by transmitting one bit information,
Ed is data aggregation energy, do is relative distance, Ef and
Em is energy spent for free space path loss and multipath
loss respectively. Now the CH takes a decision on which
route to select, either along the minimum spanning tree or
Table I
PARAMETERS : T HIS TABLE SHOWS THE PARAMETERS CONSIDERED IN
THIS STUDY.
PARAMETER
Initial energy Eo
Tranmission energy Et
Aggregation energy Ed
Reception energy Er
Free space energy Ef
Multipath energy Em
Path loss exponent α
Maximum number of rounds rmax
Probability of CH election p
Number of sensor nodes in the area N
VALUE
0.5J
50 × 10−9 J
5 × 10−9 J
50 × 10−9 J
10 × 10−12 J
0.0013×10−12 J
0.3
1500
0.1
100
directly to the fusion centre, based on the costs calculated.
The minimum cost path is selected. So, all the CHs perform
intelligent decision making and finally the fusion centre acquires complete information about the available spectrum in a
short period of time. Moreover network life time is extended
and comparatively very few nodes die after the same number
of rounds considered earlier. The energy expended during
transmission Et and reception Er for M bits to a distance
d between transmitter and receiver for the secondary user is
given by
Et = Ee M + Ea αd.
Er = Ee M.
where α is path loss component, Ea is energy constant
for propagation, Ee is the electronics energy. The following
assumptions are made while making the analysis,
• A secondary users’ network is stable and consists of one
FC, one primary transmitter and N number of cognitive
radio users or sensor nodes.
• The base station acts as the Fusion Centre and it has
the location information of all secondary users, possibly
determined using Global Positioning System (GPS).
• The instantaneous channel state information of the channel is available at the each secondary user.
• The channel between any two secondary users in the same
cluster is perfect since they are close to each other.
• In this research work, minimum tree cost determination
is made with the adjacency matrix of distances.
IV. S IMULATION R ESULTS
The performance of the proposed algorithm is carried with
the following parameters given in Table I. The performance
of the proposed routing protocol is compared with the
conventional protocol.
In the uniform clustering method, the number of clusters is
restricted to 10. From Fig. ??, the proposed communication
protocol proves to be energy efficient by 1.4 times (for 1500
rounds) and this increases the lifetime of the network. Moreover, in this research, when uniform clustering is restricted to
10 CHs, the network residual energy is still higher by 27% (for
1500 rounds) than non-uniform clustering (where the cluster
Energy remaining in all the nodes in Joules
60
50
40
30
LEACH
LEACH using Prims algorithm (non uniform clustering)
LEACH using Prims algorithm (uniform clustering)
20
10
R EFERENCES
0
200
400
600
800 1,000 1,200 1,400
Number of rounds (r)
Figure 1. This graph shows the total energy remaining in Joules, in all the
alive nodes N , at the end of each round r. Uniform and non-uniform clustering
strategies have also been compared for the proposed protocol.
Number of alive sensor nodes (N )
scales down a large size of network and proves to be an elegant
design paradigm of reducing energy consumption. Moreover,
all the elected CHs, along with the FC have complete information about the spectrum availability and as a result, network
convergence has taken place. When M number of bits are
sent only by those CHs who are nearer to FC, the problem of
channel congestion is also eliminated. As an extension work,
the sensing tables can be used to store control information
as well as to information storage and intelligence for sending
the sensed information. This may be useful in case of mission
critical situations like natural disasters, where network failure
is a common phenomenon.
100
80
60
40
LEACH
LEACH using Prims algorithm (non uniform clustering)
LEACH using Prims algorithm (uniform clustering)
20
0
200
400
600
800 1,000 1,200 1,400
Number of rounds (r)
Figure 2. This graph shows the number of alive sensor nodes N , at the end of
each round r. Uniform and non-uniform clustering strategies have also been
compared for the proposed protocol.
size is not uniform for every cluster). Thus, it is observed that
the overall network residual energy using proposed protocol
with uniform size clustering is much greater than the existing
system and non-uniform clustering as well. From Fig. 2, it is
observed that the number of active nodes after 1500 rounds of
clustering and spectrum sensing, in the proposed protocol with
uniform clustering is more than the number of active nodes in
the existing protocol by 7 times. So, with this it can be shown
that the Proposed method is an energy efficient strategy with
respect to the existing system of communication for CR-WSN.
V. C ONCLUSION
The analysis and simulation results show that the proposed
protocol with uniform or non-uniform clustering shows better
performance in terms of energy consumption. Moreover, this
[1] J. Mitola, ?Cognitive Radio?An Integrated Agent Architecture for Software Defined Radio,? 2000.
[2] A. S. Zahmati, S. Hussain, X. Fernando, and A. Grami, ?Cognitive
Wireless Sensor Networks: Emerging Topics and Recent Challenges,?
pp. 593?596, 2009.
[3] A. Ghasemi and E. S. Sousa, ?Spectrum Sensing in Cognitive Radio
Networks: Requirements, Challenges and Design Trade-offs,? IEEE Communications Magazine, vol. 46, no. 4, 2008.
[4] K.-L. A. Yau, P. Komisarczuk, and P. D. Teal, ?Cognitive Radio-Based
Wireless Sensor Networks: Conceptual Design and Open Issues,? in Local
Computer Networks, 2009. LCN 2009. IEEE 34th Conference on. IEEE,
2009, pp. 955?962.
[5] A. Jamal, C.-K. Tham, and W.-C. Wong, ?CR-WSN MAC: An Energy
Efficient and Spectrum Aware MAC Protocol for Cognitive Radio Sensor
Network,? in Cognitive Radio Oriented Wireless Networks and Communications (CROWNCOM), 2014 9th International Conference on. IEEE,
2014, pp. 67?72.
[6] J. Xu, N. Jin, X. Lou, T. Peng, Q. Zhou, and Y. Chen, ?Improvement of
LEACH Protocol for WSN,? in Fuzzy Systems and Knowledge Discovery
(FSKD), 2012 9th International Conference on. IEEE, 2012, pp. 2174?
2177.
[7] K. Akkaya and M. Younis, ?A Survey on Routing Protocols for Wireless
Sensor Networks,? Ad Hoc Networks, vol. 3, no. 3, pp. 325?349, 2005.
[8] C. J. Alpert, T. C. Hu, J. Huang, A. B. Kahng, and D. Karger, ?Primdijkstra tradeoffs for improved performance-driven routing tree design,?
IEEE Transactions on Computer-Aided Design of Integrated Circuits and
Systems, vol. 14, no. 7, pp. 890?896, 1995.
[9] M. Handy, M. Haase, and D. Timmermann, ?Low Energy Adaptive
Clus- tering Hierarchy with Deterministic Cluster-Head Selection,? in
Mobile and Wireless Communications Network, 2002. 4th International
Workshop on. IEEE, 2002, pp. 368?372.
| 7cs.IT
|
arXiv:1511.02554v3 [cs.NE] 17 Jan 2016
Deep Recurrent Neural Networks for Sequential
Phenotype Prediction in Genomics
Farhad Pouladi
Hojjat Salehinejad
Amir Mohammad Gilani
Faculty of Applied Sciences
of Communications,Tehran, Iran
UMTS Project Partners bv. (SPAA05),
Waalwijk, Netherlands
[email protected]
Farnoud Data Communication Company,
Kerman, Iran
[email protected]
Mobile Communication Company
of Iran (MCI), Tehran, Iran
[email protected]
Abstract—In analyzing of modern biological data, we are
often dealing with ill-posed problems and missing data, mostly
due to high dimensionality and multicollinearity of the dataset.
In this paper, we have proposed a system based on matrix
factorization (MF) and deep recurrent neural networks (DRNNs)
for genotype imputation and phenotype sequences prediction. In
order to model the long-term dependencies of phenotype data,
the new Recurrent Linear Units (ReLU) learning strategy is
utilized for the first time. The proposed model is implemented
for parallel processing on central processing units (CPUs) and
graphic processing units (GPUs). Performance of the proposed
model is compared with other training algorithms for learning
long-term dependencies as well as the sparse partial least square
(SPLS) method on a set of genotype and phenotype data with
604 samples, 1980 single-nucleotide polymorphisms (SNPs), and
two traits. The results demonstrate performance of the ReLU
training algorithm in learning long-term dependencies in RNNs.
Index Terms—genotype imputation; phenotype prediction; recurrent neural networks; sequence learning;
I. I NTRODUCTION
Machine learning is a practical approach to deal with real
world challenges such as to model neck pain and motor
training induced plasticity [10], [12]. Such methods have
been widely used to solve difficult optimization problems
[6], [7]. Opposition based learning is one example which
has achieved successful results in medical image processing
and optimization problems [25], [20], [11]. The genomewide association (GWA) studies have discovered many convincingly replicated associations for complex human diseases
using high-throughput single-nucleotide polymorphism (SNP)
genotypes [1], [2]. The genotype imputation has been used for
fine-map associations and facilitates the combination of results
across studies [1]. The issue of missing genotype data and its
imputation implies creating individualistic genotype data [2].
Impact of even small amounts of missing data on a multiSNP analysis is of great importance for the complex diseases
research [2]. There are several programs such as BEAGLE
[3], MaCH [4], and IMPUTE2 [5], which provide imputation
capability of untyped variants.
The sparse partial least squares (SPLS) and least absolute
shrinkage and selection operator (LASSO) methods are wellknown for simultaneous dimension reduction and variable
selection [8], [9]. The LASSO is a shrinkage and selection
method for linear regression, which attempts to minimize an
error function. This function is typically the sum of squared
errors with a bound on the sum of the absolute values of the coefficients [9]. The partial least squares (PLS) regression is used
as an alternative approach to the ordinary least squares (OLS)
regression method [8]. The SPLS method is the sparse version
of PLS method, which simultaneously works to achieve good
predictive performance and variable selection by producing
sparse linear combinations of the original predictors [8].
In general, matrix factorization is a technique to decompose
a matrix for multivariate data into two matrices with F latent
features [15]. Many matrix factorization techniques have been
proposed to increase its performance, such as non-negative
[16], sparse [17], non-linear [14], and kernel-based approaches
[13]. A kernel non-negative matrix factorization method is
proposed for feature extraction and classification of microarray data in [13]. Performance evaluation of this method for
eight different gene samples has showed better performance
over linear as well as other well-known kernel-based matrix
factorization approaches.
A sparse matrix factorization method has been proposed
for tumor classification using gene expression data, [17]. In
this approach, the gens are selected using a sparse matrix
factorization method and then the features are extracted to be
fed into a support vector machine (SVM) for tumor samples
classification. It is reported that the performance results have
been improved versus the non-sparse matrix factorization
techniques. The artificial neural network (ANN) is another
successful machine learning approach for prediction and classification applications [29]. As an example, a feed-forward
ANNs model and a Bayesian approach are utilized to impute
missing genotype data of SNPs in [2]. Sequence modelling
is one of the most areas in machine learning. This is due to
the fact that a large class of phenomenal and data around us
is made of sequences of data with particular patterns. Some
examples are retail data, speed recognition, natural language
modelling, music generation and genotype data for medical applications. With the great practical advances in deep learning,
this state-of-the-art machine learning technique is the key for
many problems in science and engineering. Recurrent neural
network (RNN), due to its recurrent connections is considered
Pre-processed Phenotype
Genotype and
Phenotype Data
Data Pre-processing
Matrix Factorization
Recurrent Neural
Network
Imputed Genotype
Predicted Phenotype
Fig. 1: The proposed system model.
1
v
0
gu,v
(
0
1
T
0
(
u
0
Update Feature Vectors
Optimizer
Objective Function
1 V
1U
U V
u
0
g'u,v
1
0
pu,1,
pu,F
u
1 U
P QT
UF
UV
v
0
1
0
1 V
qv,1,
qv,F v
V F
Fig. 2: The matrix factorization model.
as a subcategory of deep learning methods. This powerful
model is capable of learning temporal patters and in sequential
data. The power of RNN arises from its hidden state, which
works as the “memory” of system to remember past important
features for the future decision makings. The hidden state
is consisted of high-dimensional non-linear dynamics which
enables modelling any phenomena, if trained well [31], [30].
In this paper, we are proposing a new model for missing
genotype imputation and phenotype prediction using matrix
factorization and RNNs. In this model, a simple but efficient
matrix factorization method is used for missing genotype
imputation. Then, the imputed genotypes are used along the
sequence of available phenotype data to train our RNN with
the recently developed ReLU learning approach. In order to
evaluate performance of the ReLU approach in learning longterm dependencies in phenotype data, it is compared with the
LSTM-RNN and SRNN approaches.
In the next section, the data structure of the dataset used for
the experiments is described. In Section III, the methodologies
based on the matrix factorization technique and DRNN is
discussed in detail. The experimental results as well are
comparative analysis are provided in Section IV. Finally, the
paper is concluded in Section V and some guidelines are
further developments are provided.
II. DATA S TRUCTURE
For the experiments, we are using a set of data provided for
our research by Afzalipour research hospital. The genotype
data contains genotypes of 1980 SNPs for 604 observations
and the phenotype data provides measurements of two phenotype, called trait 1 and trait 2. Out of 1980 SNPs provided
in the genotype data, 5% contain missing genotypes. The
percentage of observations with missing genotypes for each
SNP varies from 1 to 25. For each trait, 30 randomly selected
observations have missing values.
III. M ETHODOLOGY
In order to deal with the missing genotype and phenotype
problem, we are utilizing the matrix factorization (MF) and
RNNs techniques to fit prediction models as in Figure 1. To
do so, after data pre-processing, the genotype dataset with
missing values is imported into the MF system to predict the
missing genotype values. By having the estimated genotype
dataset and corresponding phenotypes, the RNN is utilized in
a supervised manner to train a network model for prediction
of phenotypes, based on the known genotype-phenotype pairs.
Each stage is described in details in the following subsections.
A. Data Pre-processing
In general, the SNP genotypes (AA, BB, AB, or Null)
are denoted with integer numbers for computational purposes,
however, some programs may be able to work with this
AA/AB/BB format directly. The data pre-processing step is
an opportunity to clean data, remove noise, and translate the
genotype data and indicate/distinguish missing values from the
available data.
Output
Layer
y1 y1
yP
WHO
Hidden
Layer
ht+1WHH
y1 y1
WHO
hM
h1 h2
htWHH
WIH
Input
Layer
x1 x2
xN
h1 h2
t
hM
t
(a) Fodled SRNN.
h1 h2
yP
yP
y1 y1
W
ht+1WHH
HO
hM
hM
h1 h2
WIH
WIH
x1 x2
t+1
xN
x1 x2
xN
t+2
Time
(b) Unfolded SRNN through time.
Fig. 3: A simple recurrent neural network (SRNN) and its unfolded structure through time. To keep the figure simple, biases
are not shown.
Algorithm 1 Matrix factorization
B. Matrix Factorization Model
The proposed MF structure for genotype data imputation is
presented in Figure 2. In this model we consider U number of
samples and V number of SNPs. Therefore, the genotype data
is structured as a U ×V matrix, called GU×V . The objective of
MF technique is to estimate two matrices, PU×F and QV ×F
′
with F latent features, such that their product GU×V estimates
GU×V as:
′
G ≈ G = P × QT ,
(1)
′
where each element of the genotype matrix G is computed
by using the dot product such as:
′
gu,v =
F
X
pu,f qf,v .
1:
2:
3:
4:
5:
6:
7:
8:
9:
(2)
f =1
In order to find the best values for the matrices P and Q,
we need to minimize the objective function which describes
′
the difference between the G and G genotype matrices [15].
To do so, the gradient descent algorithm is utilized as the
optimizer in Figure 2 to update the feature matrices P and Q
iteratively.
The above procedure is illustrated in pseudocode as in
Algorithm 1. As it is demonstrated, the parameters are set
in the initialization step and random values in range [a, b]
are allocated to the feature matrices P and Q, [15]. Based
on the availability of each genotype such as Gu,v 6= ∅ for
all {u, v} ∈ {{1, ..., U }, {1, ..., V }}, the estimated genotype
′
matrix Gu,v is computed. The objective function is then
formulated with respect to P and Q as:
Initialization
P ← rand([a, b])
Q ← rand([a, b])
for each epoch do
if Gu,v 6= ∅ ∀ {u, v} ∈ {{1, ..., U }, {1, ..., V }} then
′
Gu,v ← P × QT
P
PV
′
β
2
e2 ← U
u=1
v=1 (Gu,v − Gu,v ) + 2 (kP kF +
kQkF )
2
P ← P − α ∂e
∂P2
Q ← Q − α ∂e
∂Q
P updated = P − α
∂e2
∂P
(4)
Qupdated = Q − α
∂e2
∂Q
(5)
and
respectively, where α represents the learning rate and is
practically set to 0.0001.
C. Recurrent Neural Network with Rectified Linear Unit
Model
The utilized RNN in the proposed model in Figure 1 is
consisted of input, hidden, and output layers, where each
layer is consisted of corresponding units. The input layer is
consisted of N input units, where its inputs are defined as a sequence of vectors through time t such as {..., xt−1 , xt , xt+1 ...}
where xt = (x1 , x2 , ..., xN ). In a fully connected RNN,
the
inputs units are connected to hidden units in the hidden
U X
V
X
′
β
layer,
where the connections are defined with a weight matrix
min e(G, P, Q)2 =
(Gu,v −Gu,v )2 + (kP kF +kQkF )
WIH . The hidden layer is consisted of M hidden units
2
u=1 v=1
(3) ht = (h1 , h2 , ..., hM ), which are connected to each other
where Forbenius norms of P and Q are used for regularization through time with recurrent connections. As it is demonstrated
under control of parameter β to prevent over-fitting of model in Figure 3b, the hidden units are initiated before feeding the
by penalizing it with extreme parameter values [15]. Normally inputs. The hidden layer structure defines the state space or
β is set to some values in the range of 0.02, such that P and Q “memory” of the system, defined as
can approximate G without having to contain large numbers.
ht = fH (WIH xt + WHH ht−1 + bh )
(6)
The feature matrices of P and Q are updated as:
Fig. 4: The “tanh” activation function.
Fig. 5: The “rectified linear unit (ReLU)” activation function.
where fH (.) is the hidden layer activation function and bh is
the bias vector of the hidden units. The hidden units are connected to the output layer with weighted connections WHO .
The output layer has P units such as yt = (y1 , y2 , ..., yP )
which are estimated as
yt = fO (WHO ht + b0 )
(7)
where fO (.) is the output layer activations functions and b0
is the bias vector.
Learning long term dependencies in RNNs is a difficult
task [30]. This is due to two major problems which are
vanishing gradients and exploding gradients. The long-shortterm memory (LSTM) method is one of the popular methods to
overcome the vanishing gradient problem. A recent proposed
method suggests that proper initialization of the RNN weights
with rectified linear units has good performance in modeling
long-range dependencies [30]. In this approach, the model is
trained by utilizing back-propagation through time (BPTT)
technique to compute the derivatives of error with respect to
the weights. The reported performance analysis show that this
method has comparable results in comparision to the LSTM
method, with much less complexity.
In this model, each new hidden state vector is inherited
from the previous hidden vector by copying its values, adding
the effects of inputs, and finally, replacing negative state
values by zero. In other words, this means that the recurrent
weight matrix is initialized to an identity matrix and the
biases are set to zero. This procedure is in fact replacing the
”tanh” activation function (Figure 4) with a rectified linear
unit (ReLU)(Figure 5). The ReLU in fact is modelling the
behaviour of LSTM. In LSTM, the gates are set in a way
that there is no decay to model long-term dependencies. In
ReLU, when the error derivatives for the hidden units are
back-propagated through time they remain constant provided
no extra error-derivatives are added [30].
IV. E XPERIMENTAL R ESULTS
The proposed model is implemented for parallel processing
using Theano in Python [21], [18], [19]. In this section,
we are presenting the performance result from comparison
between simple RNN, LSTM, and ReLU training methods
for phenotype sequence prediction. The ReLU method is then
TABLE I: Parameter setting for the experiments.
Parameter Name
α
β
NEP
U
V
F
eta1
eta2
K1
K2
[a, b]
Parameter Descritption
Learning Rate
Regularization Control
Number of Epochs
Number of Samples
Number of SNPs
Number of Features
eta for Trait 1
eta for Trait 2
K for Trait 1
K for Trait 2
Boundary values for P and Q
Parameter Value
0.001
0.02
5000
604
1980
400
0.2
0.2
3
4
[0,1]
compared with the well-known sparse partial least square
(SPLS) method.
A. Parameter Setting
As it is described in Section II, it is assumed that the
genotypes data is either missing or observed. Therefore, the
observed genotypes are represented as Gu,v ∈ {0, 1, 2} and
the missing (null) data is represented as Gu,v = 5 for the
experiments [1]. The genotype dataset GU,V is consisted of
V = 1980 SNPs for U = 604 observations. Parameter setting
for all the experiments are presented in Table I adapted from
the literature, [14], [19], [29], unless a change is mentioned. As
recommended in the literature, the 10-fold cross-validation is
used to tune the parameters eta and K for the SPLS algorithm
using the SPLS package in R programming language [?], [26],
[27], [8]. The parameter tuning is conducted separately for
each provided trait in the phenotype dataset for 1 ≤ K ≤ 10.
The optimal values are provided in Table I.
Due to small size of data samples, %80 of available
genotype and phenotype data is used for training the ANN,
%10 for validation, and %10 for testing. Regarding the SPLS
algorithm, %80 of the provided data is considered as training
data and %20 as test data.
B. Simulation Results Analysis
In this subsection, performance results of the proposed
method is presented and compared with the SPLS method for
the described dataset in Section II. Performance of the methods
is evaluated by measuring the correlation between the original
TABLE II: Performance results of the matrix factorization
technique success for different latent features F .
5
10
Average of Cost
50
100
300
500
700
1000
F
50
100
300
500
700
1000
4
10
3
10
2
10
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
Epoch
(a) Logarithmic view for average of costs.
10
50
100
300
500
700
1000
1
Average of Cost
4
2
10
Genotype Matrix Construction
%71.47
%76.11
%88.11
%95.72
%96.91
%97.56
TABLE III: Performance results of the ReLU-RNN for successful phenotype traits prediction.
Trait
5
Missing Genotype
%52.45
%65.16
%67.53
%68.80
%70.88
%72.48
Mode
Train
Validation
Test
Train
Validation
Test
Original Genotype
%82.75
%86.56
%84.66
%80.62
%85.25
%79.58
Missing Genotype
%80.53
%80.22
%80.25
%75.01
%80.54
%78.29
3
10
TABLE IV: Performance results in terms of correlation between the predicted and original phenotype data for the sparse
partial least squares (SPLS) method.
2
10
0
10
1
10
2
10
3
10
4
10
Epoch
Trait
1
(b) Logarithmic view for average of costs and epochs.
2
Fig. 6: Average of costs versus epochs for genotype imputation
using matrix factorization technique with different number of
latent features.
Mode
Train
Test
Train
Test
Original Phenotype
%75.10
%54.56
%80.53
%69.82
Missing Phenotype
%69.52
%51.53
%72.85
%56.42
better that before in each epoch, however, after epoch 110 the
progress is not very significant. The performance results for
the success percentile in missing genotype data and success
percentile in the whole genotype data construction is provided
in Table II. There is a trade-off between the number of features
and performance. This is due to the fact that increasing the
number of features increases the computational complexity.
Fig. 7: Training error comparison between SRNN, LSTMRNN, and ReLU-RNN.
genotype and phenotype values with the corresponding predicted values.
The average of cost for fitting the MF model for different
epochs is presented in Figure 6. In order to have a deeper look
in details, the figure is presented in linear and logarithmic
scales. The measure to compute the average of cost is the
mean-square error, which is the average of squared difference
between targets and the output of the MF model. As the
results show, the more number of latent features F results
in less average of cost. In addition, the model is trying to fit
In Table III, performance of the ReLU-RNN for the genotype prediction is presented. As it is showed, the ReLU
method has better training performance comparing to the
results of the SPLS method in Table IV. The missing genotype
represents the error for the test dataset in percentile while
the original genotype represents the training error values in
percentile. Since the training dataset has been seen by the
model during training, it is reasonable to see that performance
of the methods for the training dataset is better than the unseen
test data.
In Figure 7, the training error of the SRNN, LSTM-RNN,
and ReLU-RNN algorithms are compared. For better illustration, the first 100 training epochs are presented. As the results
show, the SRNN algorithm has more training error than the
LSTM-RNN method. This is while the ReLU-RNN approach
has the least training error comparing to LSTM-RNN. At the
early epochs, we see that the LSTM and ReLU are almost at
the same training loss, however, the ReLU achieves less error
in further epochs.
V. C ONCLUSION
AND
F UTURE W ORKS
In this paper, a novel model is proposed which utilizes matrix factorization and deep recurrent neural networks (DRNN)
for genotype imputation and phenotype sequences prediction.
Sine we are interested in keeping track of sequences with longterm dependencies in genomics, the state-of-the-art recited
linear unit learning method is used.
The performance results show the with the ReLU methods has a better performance in training comparing to the
LSTM-RNN and simple RNN methods. The ReLU learning
methods also has less computational complexity comparing
to the LSTM method. For future research, it is interesting to
analyze other recent advances in deep learning for genotypephenotype application; particularly that these algorithms are
moving toward more simple designs which is suitable for big
data application.
R EFERENCES
[1] J. Marchini and B. Howie, “Genotype imputation for genome-wide
association studies,” Nature Reviews Genetics, vol. 11, pp. 499-511,
2010.
[2] Y. V. Sun and S. LR. Kardia1, “Imputing missing genotypic data
of single-nucleotide polymorphisms using neural networks,” European
Journal of Human Genetics, vol. 16, pp. 487-495, 2008.
[3] SR. Browning and BL. Browning, ”Rapid and accurate haplotype
phasing and missing-data inference for whole-genome association
studies by use of localized haplotype clustering,” American journal
of human genetics, vol. 81, pp. 10841097, 2007.
[4] Y. Li, CJ. Willer, J. Ding, P. Scheet, and GR. Abecasis, “MaCH: using
sequence and genotype data to estimate haplotypes and unobserved
genotypes,” Genetic Epidemiology, vol. 34, pp. 816834 2010.
[5] BN. Howie, P. Donnelly, and J. Marchini, “A flexible and accurate
genotype imputation method for the next generation of genome-wide
association studies,“ PLoS genetics 5: e1000529, 2009.
[6] H. Salehinejad and S. Talebi. ”Dynamic fuzzy logic-ant colony systembased route selection system,” Applied Computational Intelligence and
Soft Computing, 2010.
[7] H. Salehinejad, et al. ”Micro-differential evolution with vectorized random mutation factor,” Evolutionary Computation, 2014 IEEE Congress
on (CEC), pp. 2055-2062, 2014.
[8] H. Chun and S. Kele, “Sparse partial least squares regression for
simultaneous dimension reduction and variable selection,” J R Stat Soc
Series B Stat Methodol, vol. 72, no. 1, pp. 3-25, 2010.
[9] R. Tibshirani, “Regression shrinkage and selection via the lasso,” J.
Royal. Statist. Soc B., Vol. 58, No. 1, pp. 267-288, 1996.
[10] J. Baarbé, et al. ”A novel protocol to investigate motor training-induced
plasticity and sensorimotor integration in the cerebellum and motor
cortex,” Journal of neurophysiology, 111.4, pp. 715-721, 2014.
[11] H. Salehinejad, S. Rahnamayan, and H. R. Tizhoosh, ”Type-II
opposition-based differential evolution,” Evolutionary Computation,
2014 IEEE Congress on (CEC), pp. 1768-1775, 2014.
[12] J. Daligadu, et al. ”Alterations in cortical and cerebellar motor processing in subclinical neck pain patients following spinal manipulation,”
Journal of manipulative and physiological therapeutics, 36.8, pp. 527537, 2013.
[13] Y. Li and A. Ngom, “A New Kernel Non-Negative Matrix Factorization and Its Application in Microarray Data Analysis,“ in Proc.
IEEE Symposium on Computational Intelligence in Bioinformatics and
Computational Biology, pp. 371-378, 2012.
[14] P. Binbin, L. Jianhuang, and C. Wen-Sheng, ”Nonlinear nonnegative
matrix factorization based on Mercer kernel construction,“ Pattern
Recognition, vol. 44, Iss. 1011, pp. 2800-2810, 2011.
[15] R. Gribonval and K. Schnass, “Dictionary IdentificationSparse MatrixFactorization via l1 -Minimization,” IEEE Transactions on Information
Theory, vol. 56, no. 7, pp. 3523-3539, 2010.
[16] D.D. Lee and S. Seung, “Learning the parts of objects by non-negative
matrix factorization,” Nature, vol. 401, pp. 788-791, 1999.
[17] C. Zheng and et. al., “Tumor Classification Based on Non-Negative
Matrix Factorization Using Gene Expression Data,” IEEE Transactions
on NanoBioscience, vol. 10, pp. 86-93, 2011.
[18] E. Jones, E. Oliphant, P. Peterson, and et. al. “SciPy: Open Source
Scientific Tools for Python,” 2001, http://www.scipy.org/ [Online; accessed 2015-04-07]
[19] J. Bergstra, O. Breuleux, F. Bastien, and et. al., “Theano: A CPU and
GPU Math Expression Compiler,“ in Proceedings of the Python for
Scientific Computing Conference (SciPy), Austin, TX, 2010.
[20] S. Rahnamayan, and H. R. Tizhoosh, ”Image thresholding using micro
opposition-based differential evolution (Micro-ODE).” In Evolutionary
Computation, IEEE Congress on, pp. 1409-1416, 2008.
[21] Python Software Foundation. Python Language Reference, version 2.7.
Available at http://www.python.org.
[22] MATLAB and Artificial Neural Networks Toolbox, Release 2012b, The
MathWorks, Inc., Natick, Massachusetts, United States.
[23] Marquardt, D., ”An Algorithm for Least-Squares Estimation of Nonlinear Parameters,“ SIAM Journal on Applied Mathematics, Vol. 11,
No. 2, pp. 431441, June 1963.
[24] A. Ranganathan, ”The Levenberg-Marquardt Algorithm,“ June 2004.
[25] S. Rahnamayan, H. R. Tizhoosh, and M. Salama, ”Opposition-based
differential evolution,” Evolutionary Computation, IEEE Transactions
on 12, no. 1, pp.64-79, 2008.
[26] D. Chung and et. al., “An Introduction to the ’spls’ Package, Version
1.0,” June 10, 2012.
[27] D. Chung and et. al., “Package spls,” R programming language,
February 20, 2015.
[28] F. Bastien, P. Lamblin, R. Pascanu, J. Bergstra, I. Goodfellow, A.
Bergeron, N. Bouchard, D. Warde-Farley and Y. Bengio. ?Theano: new
features and speed improvements?. NIPS 2012 deep learning workshop.
[29] S. Mahdavi-Jafari, H. Salehinejad, and S. Talebi. ”A Pistachio Nuts
Classification Technique: An ANN Based Signal Processing Scheme.”
In Computational Intelligence for Modelling Control and Automation,
2008 International Conference on, pp. 447-451. IEEE, 2008.
[30] Le, Q. V, Jaitly, N., & Hinton Google, G. E. ”A Simple Way to
Initialize Recurrent Networks of Rectified Linear Units,” pp. 1?9, 2015,
arXiv:1504.00941v2 [cs.NE] 7 Apr 2015.
[31] N. Boulanger-Lewandowski, Y. Bengio, and P. Vincent, ”Modeling
Temporal Dependencies in High-Dimensional Sequences: Application
to Polyphonic Music Generation and Transcription,” Appearing in Proc.
ICML, 2012.
| 5cs.CE
|
Weak convergence of quantile and expectile processes under
general assumptions
Tobias Zwingmann∗ and Hajo Holzmann†
Fachbereich Mathematik und Informatik
Philipps-Universität Marburg, Germany
arXiv:1706.04668v1 [math.ST] 14 Jun 2017
January 4, 2018
Abstract
We show weak convergence of quantile and expectile processes to Gaussian limit processes in the space of bounded functions endowed with an appropriate semimetric which is
based on the concepts of epi- and hypo convergence as introduced in Bücher et al. (2014).
We impose assumptions for which it is known that weak convergence with respect to the
supremum norm or the Skorodhod metric generally fails to hold. For expectiles, we only
require a distribution with finite second moment but no further smoothness properties of
distribution function, for quantiles, the distribution is assumed to be absolutely continuous
with a version of its Lebesgue density which is strictly positive and has left- and right-sided
limits. We also show consistency of the bootstrap for this mode of convergence.
Keywords. epi - and hypo convergence, expectile process, quantile process, weak convergence
1
Introduction
Quantiles and expectiles are fundamental parameters of a distribution which are of major interest
in statistics, econometrics and finance (Koenker, 2005; Bellini et al., 2014; Newey and Powell,
1987; Ziegel, 2016) .
The asymptotic properties of sample quantiles and expectiles have been addressed in detail
under suitable conditions. For quantiles, differentiability of the distribution function at the quantile with positive derivative implies asymptotic normality of the empirical quantile, and under a
continuity assumption on the density one obtains weak convergence of the quantile process to
a Gaussian limit process in the space of bounded functions with the supremum distance from
the functional delta-method (van der Vaart, 2000). However, without a positive derivative of the
distribution function at the quantile, the weak limit will be non-normal (Knight, 2002), and thus
process convergence to a Gaussian limit with respect to the supremum distance cannot hold true.
Similarly, for a distribution with finite second moment, the empirical expectile is asymptotically normally distributed if the distribution function is continuous at the expectile, but nonnormally distributed otherwise (Holzmann and Klar, 2016). For continuous distribution functions, process convergence of the empirical expectile process in the space of continuous functions
also holds true, but for discontinuous distribution functions this can no longer be valid.
∗
†
Email: [email protected]
Corresponding author, Email: [email protected]
1
In this note we discuss convergence of quantile and expectile processes from independent and
identically distributed observations under more general conditions. Indeed, we show that the
expectile process converges to a Gaussian limit in the semimetric space of bounded functions
endowed with the hypi-semimetric as recently introduced in Bücher et al. (2014) under the assumption of a finite second moment only. Since the Gaussian limit process is discontinuous in
general while the empirical expectile process is continuous, this convergence can hold neither
with respect to the supremum distance nor with a variant of the Skorodhod metric. Similarly,
we show weak convergence of the quantile process under the hypi-semimetric if the distribution
function is absolutely continuous with a version of its Lebesgue density that is strictly positive
and has left and right limits at any point. These results still imply weak convergence of important statistics such as Kolmogorov-Smirnov and Cramer-von Mises type statistics. We also show
consistency of the n out of n bootstrap in both situations.
A sequence of bounded functions (fn )n∈N on a compact metric space hypi-converges to
the bounded function f if it epi-converges to the lower-semicontinuous hull of f , and hypoconverges to its upper-semicontinuous hull, see the supplementary material for the definitions.
Bücher et al. (2014) show that this mode of convergence can be expressed in terms of a semimetric dhypi on the space of bounded functions, the hypi-semimetric.
We shall use the notation c−1 = 1/c for c ∈ R, c 6= 0, and denote the (pseudo-) inverse
L
of a function h with hInv . We write → for ordinary weak convergence of real-valued random
variables. Weak convergence in semimetric spaces will be understood in the sense of HofmannJørgensen, see van der Vaart and Wellner (2013) and Bücher et al. (2014).
2
Convergence of the expectile process
For a random variable Y with distribution function F and finite mean E [|Y |] < ∞, the τ expectile µτ = µτ (F ), τ ∈ (0, 1), can be defined as the unique solution of E [Iτ (x, Y )] = 0,
x ∈ R, where
Iτ (x, y) = τ (y − x)1 (y ≥ x) − (1 − τ ) (x − y)1 (y < x) ,
(1)
and 1 (·) is the indicator function. Given a sequence of independent and identically distributed
copies Y1 , Y2 , . . . of Y and a natural number n ∈ N, we let
n
µ̂τ,n = µτ Fn ,
1X
Fn (x) =
n
1 (Yk ≤ x) ,
k=1
be the empirical τ -expectile and the empirical distribution function, respectively.
Theorem 1. Suppose that E Y 2 < ∞. Given 0 < τl < τu < 1, the standardized
expectile
√
∞
process τ 7→ n µ̂τ,n − µτ , τ ∈ [τl , τu ], converges weakly in ` [τl , τu ], dhypi to the limit
−1
process ψ̇ Inv (Z)(τ ) τ ∈[τ ,τu ] . Here, ψ̇ Inv (ϕ)(τ ) = τ +(1−2 τ )F (µτ ) ϕ(τ ), ϕ ∈ `∞ [τl , τu ],
l
and (Zτ )τ ∈[τl ,τu ] is a centered
tight Gaussian process with continuous sample paths and covariance function cov Zτ , Zτ 0 = E [Iτ (µτ , Y ) Iτ 0 (µτ 0 , Y )] for τ, τ 0 ∈ [τl , τu ].
From Propositions 2.3 and 2.4 in Bücher et al. (2014), hypi-convergence of the expectile
process implies ordinary weak convergence of important statistics such as Kolmogorov-Smirnov
or Cramér-von Mises type statistics. We let
kϕk[τl ,τu ] =
sup |ϕ(τ )|,
τ ∈[τl ,τu ]
2
ϕ ∈ `∞ [τl , τu ],
denote the supremum norm on `∞ [τl , τu ].
Corollary 2. If E Y 2 < ∞, then we have as n → ∞ that
√
n µ̂·,n − µ·
L
[τl ,τu ]
→ ψ̇ Inv Z
[τl ,τu ]
.
Further, for p ≥ 1 and a bounded, non-negative weight function w on [τl , τu ],
Z τu
Z τu
p
L
p
p/2
µ̂τ,n − µτ
ψ̇ Inv (Z) (τ ) w(τ ) dτ.
n
w(τ ) dτ →
τl
τl
Remark 3. Evaluation at a given point x is only a continuous operation under the hypi-semimetric if the limit function is continuous at x, see Proposition 2.2 in Bücher et al. (2014). In
particular, this does not apply to the expectile process if the distribution function F is discontinuous at µτ . Indeed, Theorem 7 in Holzmann and Klar (2016) shows that the weak limit of the
empirical expectile is not normal in this case.
Next we turn to the validity of the bootstrap. Given n ∈ N let Y1∗ , . . . , Yn∗ denote a sample
drawn from Y1 , . . . , Yn with replacement, that is, having distribution function
Fn . Let Fn∗ denote
the empirical distribution function of Y1∗ , . . . , Yn∗ , and let µ∗τ,n = µτ Fn∗ denote the bootstrap
expectile at level τ ∈ (0, 1).
Theorem 4. Suppose that E Y 2 < ∞. Then, almost surely,conditionally on Y1 , Y2 , . . . the
√
standardized bootstrap expectile process τ 7→ n µ∗τ,n − µ̂τ,n , τ ∈ [τl , τu ], converges weakly
in L∞ [τl , τu ], dhypi to ψ̇ Inv (Z)(τ ) τ ∈[τ ,τu ] , where the map ψ̇ Inv and the process (Zτ )τ ∈[τl ,τu ]
l
are as in Theorem 1.
Remark. The simple n out of n bootstrap does not apply for the empirical expectile at level τ
if F is discontinuous at µτ , see Knight (1998) for a closely related result for the quantile. Thus,
Theorem 4 is somewhat surprising, but its conclusion is reasonable in view of Remark 3.
3
Convergence of the quantile process
Let qα = qα (F ) = inf{x ∈ R : F (x) ≥ α}, α ∈ (0, 1), denote the α-quantile of the distribution
function F , and let q̂α = qα (Fn ) be the empirical α-quantile of the sample Y1 , . . . , Yn .
Theorem 5. Suppose that the distribution function F is absolutely continuous and strictly increasing on [qαl , qαu ], 0 < αl < αu < 1. Assume that there is a version f of its density which is
bounded and bounded away from zero on [qαl , qαu ], and which admits right- and left-sided limits
in every point of [qαl , qαu ].
√
Then the standardized quantile process α 7→ n q̂α,n −qα , α ∈ [αl , αu ], converges weakly in
(L∞ [αl , αu ], dhypi ) to the process Vα/f (qα ) α∈[α ,αu ] , where (Vα )α∈[αl ,αu ] is a Brownian bridge
l
on [αl , αu ].
Furthermore, as n → ∞ we have that
√
L
n q̂·,n − q· [α ,αu ] → Ψ̇Inv V [α ,αu ] ,
l
l
as well as
n
p/2
Z
αu
p
L
Z
αu
q̂α,n − qα w(α) d α →
αl
p
Ψ̇Inv (V ) (α) w(α) dα.
αl
for p ≥ 1 and a bounded, non-negative weight function w on [αl , αu ].
3
∗ = q F ∗ denote the bootstrap quantile at level α ∈ (0, 1).
Let qα,n
α
n
Theorem 6. Let the assumptions of Theorem 5 be true. Then, almost surely, conditionally
on
√
∗
Y1 , Y2 , . . . the standardized bootstrap quantile
process
α
→
7
n
q
−
q̂
,
α
∈
[α
,
α
α,n
u ],
l
α,n
∞
V
α
/f (qα ) α∈[α ,αu ] .
converges weakly in L [αl , αu ], dhypi to
l
4
Appendix: Outlines of the proofs
In this section we present an outline of the proofs, technical details can be found in the supplementary material.
Proof of Theorem 1 (Outline). We give an outline of the proof of Theorem 1. For a distribution
function S with finite first moment
let [ψ(ϕ, S)](τ ) = −Iτ (ϕ(τ ), S), τ ∈ [τl , τu ] and ϕ ∈
R
`∞ [τl , τu ], where Iτ (x, S) = Iτ (x, y) dS(y), and set ψ0 = ψ( , F ) and ψn = ψ( , F̂n ). In the
following we simply write kϕk instead of kϕk[τl ,τu ] .
√
Step 1. Weak convergence of n ψ0 (µ̂·,n ) − ψ0 (µ· ) to Z in `∞ [τl , τu ], k · k
This step uses standard results from empirical process theory based on bracketing properties of
Lipschitz-continuous functions. The main issue in the proof of the lemma below is the Lipschitzcontinuity of τ 7→ µτ , τ ∈ [τl , τu ], for a general distribution function F .
Lemma 7. In (`∞ [τl , τu ], k · k) we have the weak convergence
√
n ψn (µ· ) − ψ0 (µ· ) (τ ) → Zτ , τ ∈ [τl , τu ].
(2)
Further, given δn & 0 we have as n → ∞ that
√
sup
sup
n ψn (µ· + ϕ)(τ ) − ψ0 (µ· + ϕ)(τ )
kϕk[τl ,τu ] ≤δn τ ∈[τl ,τu ]
(3)
− ψn (µ· )(τ ) − ψ0 (µ· )(τ ) = oP (1).
Since ψ0 (µ· ) = ψn (µ̂·,n ) = 0, (3) and the uniform consistency of µ̂·,n from Theorem 1 in
Holzmann and Klar (2016) give
√
√
n ψ0 (µ̂·,n ) − ψ0 (µ· ) = n ψ0 (µ̂·,n ) − ψn (µ̂·,n )
√
= − n ψn (µ· ) − ψ0 (µ· ) + oP (1),
where the remainder term is in (`∞ [τl , τu ], k · k), and (2) and the fact that Z and −Z have the
same law conclude the proof of
√
n ψ0 (µ̂·,n ) − ψ0 (µ· ) → Z weakly in `∞ [τl , τu ], k · k .
(4)
Step 2. Invertibility of ψ0 and semi-Hadamard differentiability of ψ0Inv with respect to dhypi .
The first part of Step 2. is observing the following lemma.
Lemma 8. The map ψ0 is invertible. Further ψ0Inv (ϕ) ∈ `∞ [τl , τu ] for any ϕ ∈ `∞ [τl , τu ].
The next result then is the key technical ingredient in the proof of Theorem 1.
4
Lemma 9. The map ψ0Inv is semi-Hadamard differentiable with respect to the hypi-semimetric in
0 ∈ C[τl , τu ] tangentially to C[τl , τu ] with semi-Hadamard derivative ψ̇ Inv (ϕ)(τ ) = τ + (1 −
−1
2 τ )F (µτ ) ϕ(τ ), ϕ ∈ `∞ [τl , τu ], that is, we have
Inv
Inv
Inv
t−1
n ψ0 (tn ϕn ) − ψ0 (0) → ψ̇ (ϕ)(τ )
for any sequence tn & 0, tn > 0 and ϕn ∈ `∞ [τl , τu ] with ϕn → ϕ ∈ C[τl , τu ] with respect to
dhypi .
The proof of the lemma is based on an explicit representation of increments of ψ0Inv , and novel
technical properties of convergence under the hypi-semimetric for products and quotients.
Step 3. Conclusion with the generalized functional delta-method.
From (4), Lemma 9 and the generalized functional delta-method, Theorem B.7 in Bücher et al.
(2014) we obtain
√
√
n µ̂·,n − µ· = n ψ0Inv ψ0 (µ̂·,n ) − ψ0Inv 0 → ψ̇ Inv (Z)
in (L∞ [τl , τu ], dhypi ).
Proof of Theorem 4 (Outline). The steps in the proof are similar to those of Theorem 1. In the
analogous result to Lemma 7 and (4), we require the uniform consistency of µ∗·,n as in Holzmann
and Klar (2016), Theorem 1. The weak convergence statements require the changing classes
central limit theorem, van der Vaart (2000), Theorem 19.28. In the second step we argue directly
with the extended continuous mapping theorem, Theorem B.3 in Bücher et al. (2014).
Proof of Theorem 5 (Outline). Let G be the distribution function of U(0, 1) and Gn the empirical
version coming from a sample U1 , . . . , Un . By the quantile transformation we can write
√
√
Inv
n q̂·,n − q· = n F Inv (GInv
(GInv (·)) .
n (·)) − F
√
Inv
∞
The process n(GInv
n − G ) converges in distribution in (l [αl , αu ], k · k[αl ,αu ] ) to V , see
Example 19.6, van der Vaart (2000).
To use a functional delta-method for the hypi-semimetric, we require
ϕ(α)
Inv
t−1
(α + tn ϕn (α)) − F Inv (α) →
n F
f (qα )
with respect to dhypi for tn & 0 and ϕn ∈ `∞ [αl , αu ], ϕ ∈ C[αl , αu ] such that dhypi (ϕn , ϕ) →
Inv
0, that is the semi-Hadamard-differentiability of the functional ΨInv
(α + ν(α)),
0 (ν)(α) = F
∞
ν ∈ ` [αl , αu ] as in Definition B.6, Bücher et al. (2014). To also be able to deal with the
bootstrap version, we directly show a slightly stronger version, the uniform semi-Hadamarddifferentiability.
Lemma 10. Let tn & 0 ∈ R, ϕn , νn ∈ `∞ [αl , αu ], ϕ ∈ C[αl , αu ] such that dhypi (ϕn , ϕ) → 0
and dhypi (νn , 0) → 0 holds. Then the convergence
h
i
ϕ(·)
Inv
Inv
t−1
F
id
(·)
+
ν
(·)
+
t
ϕ
(·)
−
F
id
(·)
+
ν
(·)
→
n
n n
n
[αl ,αu ]
[αl ,αu ]
n
f (q· )
with respect to dhypi is true.
5
√
Inv
Using the convergence of n(GInv
n − G ), Lemma 10 and the functional delta-method, Theorem B.7, Bücher et al. (2014), we conclude
√
i
√ h
Inv
Inv
n q̂·,n − q· = n F Inv GInv
G (·) →
n (·) − F
V
f (q· )
in (L∞ [αl , αu ], dhypi ).
Proof of Theorem 6. Given a sample U1 , . . . , Un of independent U[0, 1] distributed random variables and the corresponding empirical distribution function Gn , let U1∗ , . . . , Un∗ be a bootstrap
sample with empirical distribution function G∗n . Using the quantile transformation we obtain
i
√ h Inv ∗ Inv
Inv
√
∗
Inv
n q·,n − q̂·,n = n F
(Gn ) (·) − F
Gn (·) .
√
By Theorem 3.6.2, van der Vaart and Wellner (2013), the process n (G∗n )Inv (·) − GInv
n (·)
converges in distribution to V , conditionally, almost surely, with V as in Theorem 5. Now we
use Theorem 3.9.13, van der Vaart and Wellner (2013), extended to semi-metric spaces. To this
end we use Lemma 10, which covers (3.9.12) in van der Vaart and Wellner (2013). Moreover
Inv
we require that kGInv
n (·) − G (·)k → 0 holds almost surely, which is valid by the classical
Glivenko-Cantelli-Theorem, Theorem 19.4, van der Vaart (2000). Finally, condition (3.9.9) in
van der Vaart and Wellner (2013) is valid by their Theorem 3.6.2. Thus we can indeed apply
Theorem 3.9.13, van der Vaart and Wellner (2013), which concludes the proof.
Supplementary material
The supplementary material contains simulation results, a summary of convergence under the
hypi-semimetric and technical details for the proofs.
Acknowledgements
Tobias Zwingmann acknowledges financial support from of the Cusanuswerk for providing a
dissertation scholarship.
References
Bellini, F., B. Klar, A. Müller, and E. R. Gianin (2014). Generalized quantiles as risk measures.
Insur. Math. Econ. 54, 41–48.
Beyn, W.-J. and J. Rieger (2011). An implicit function theorem for one-sided lipschitz mappings.
Set-Valued and Variational Analysis 19(3), 343–359.
Bücher, A., J. Segers, and S. Volgushev (2014). When uniform weak convergence fails: empirical
processes for dependence functions and residuals via epi- and hypographs. Ann. Stat. 42(4),
1598–1634.
6
Holzmann, H. and B. Klar (2016). Expectile asymptotics. Electron. J. Statist. 10(2), 2355–2371.
Knight, K. (1998). Bootstrapping sample quantiles in non-regular cases. Stat. Probab. Lett. 37,
259–267.
Knight, K. (2002). What are the limiting distributions of quantile estimators? In Y. Dodge (Ed.),
Statistical Data Analysis Based on the L1-Norm and Related Methods, pp. 47–65. Birkhäuser
Basel.
Koenker, R. (2005). Quantile regression. Cambridge: Cambridge University Press.
Newey, W. and J. Powell (1987). Asymmetric least squares estimation and testing. Econometrica 55(4), 819–847.
Thomson, B., A. Bruckner,
www.classicalrealanalysis.com.
and
J.
Bruckner
(2008).
Real
Analysis.
van der Vaart, A. (2000). Asymptotic Statistics. Cambridge Series in Statistical and Probabilistic
Mathematics. Cambridge University Press.
van der Vaart, A. and J. Wellner (2013). Weak Convergence and Empirical Processes: With
Applications to Statistics. Springer Series in Statistics. Springer New York.
Ziegel, J. F. (2016). Coherence and elicitability. Math. Finance 26(4), 901–918.
5
Supplement: Simulation Study
In this section we illustrate the asymptotic results for the expectile process in a short simulation.
Let Y be a random variable with distribution function
Z x
y2
9
1
1
√
F (x) =
exp −
dy +
1 (x ≥ 1) ,
10 −∞ 4 2 π
32
10
which is a mixture of a N (0, 16) random variable and point-mass in 1, so that E [Y ] = 1/10
and E Y 2 = 14.5. We will concentrate on the weak convergence of the sup-norm of the
empirical expectile process. Using equation (2.7) in Newey and Powell (1987), we numerically
find µτ0 = 1 for τ0 ≈ 0.6529449, and investigate the expectile process on the interval [0.6, 0.7].
√
Figure 1 contains four paths of the expectile process n µ̂τ,n − µτ for samples of size
n = 104 . All plotted paths seem to evolve a jump around τ0 .
Now we investigate the distribution of the supremum norm of the expectile process on the
interval [0.6, 0.7]. To this end, we simulate M = 104 samples of sizes n ∈ {10, 102 , 104 }, compute the expectile process and its supremum norm. Plots of the resulting empirical distribution
function and density estimate of this statistic are contained in Figure 2. The distribution of the
supremum distance seems to converge quickly.
Finally, to illustrate performance of the bootstrap, Figure 3 displays the distribution of M =
√
4
2
3
4
10 bootstrap samples of k n µ∗·,n − µ̂·,n k based on a single sample of size n ∈
{10 , 10 , 10 }
√
from the n out of n bootstrap, together with the distribution of k n µ̂·,n − µ· k. The bootstrap
distribution for n = 104 is quite close to the empirical distribution.
7
(a) Path evolving a downward jump.
(b) Path evolving a downward jump.
(c) Path evolving an upward jump.
(d) Path evolving an upward jump.
Figure 1: The pictures show simulated paths of the empirical expectile process based on n = 104
observations of Y . If the path is negative (positive) around τ0 , a downward- (upward-) jump seems
to evolve. This is plausible when considering the form of the hulls of ψ̇ Inv in the limit process.
(a) Estimated cumulative distribution function.
(b) Estimated density.
√
Figure
the cumulative distribution function of the supremum norm of n µ̂τ,n −
2: Figure (a) shows
µτ , based on M = 104 samples of sizes n ∈ {10, 102 , 104 }, Figure (b) the corresponding density
estimate.
6
Supplement: weak-convergence under the hypi semimetric
Let us briefly discuss the concept of hypi-convergence as introduced by Bücher et al. (2014).
Let (T, d) be a compact, separable metric space, and denote by `∞ (T) the space of all bounded
functions h : T → R. The lower- and upper-semicontinuous hulls of h ∈ `∞ (T) are defined by
h∧ (t) = lim inf h(t0 ) | d(t, t0 ) < ε , h∨ (t) = lim sup h(t0 ) | d(t, t0 ) < ε
(5)
ε&0
ε%0
and satisfy h∧ , h∨ ∈ `∞ (T) as well as h∧ ≤ h ≤ h∨ . A sequence hn ∈ `∞ (T) hypi-converges
to a limit h ∈ `∞ (T), if it epi-converges to h∧ , that is,
for all t, tn ∈ T with tn → t :
for all t ∈ T there exist tn ∈ T, tn → t :
8
h∧ (t) ≤ lim inf hn (tn )
n→∞
h∧ (t) = lim hn (tn ),
n→∞
(6)
(a) Estimated bootstrap cumulative distribution function.
(b) Estimated bootstrap density.
√
Figure 3: Figure (a) shows the estimated cumulative bootstrap distribution function of k n µ̂∗·,n −
µ·,n k, figure (b) the estimated density thereof, obtained from M = 104 estimates of this statistic.
The red√line indicates the estimated empirical distribution and density function, respectively, taken
from k n µ·,n − µ· k for n = 104 .
and if it hypo-converges to h∨ , that is,
for all t, tn ∈ T with tn → t :
for all t ∈ T there exist tn ∈ T, tn → t :
lim sup hn (tn ) ≤ h∨ (t)
n→∞
lim hn (tn ) = h∨ (t).
(7)
n→∞
The limit function h is only determined in terms of its lower - and upper semicontinuous
hulls. Indeed, there is a semimetric, denoted by dhypi , so that the convergence in (5) and (7)
is equivalent to dhypi (hn , h) → 0, see Bücher et al. (2014) for further details. To transfer the
concept of weak convergence from metric to semimetric space, Bücher et al. (2014) consider the
space L∞ (T) of equivalence classes [h] = {g ∈ `∞ (T) | dhypi (h, g) = 0}. The convergence
of a sequence of random elements (Yn ) in g ∈ `∞ (T) to a Borel-measurable Y is defined in
terms of weak convergence of ([Yn ]) to [Y ] in the metric space (L∞ (T), dhypi ) in the sense of
Hofmann-Jørgensen, see van der Vaart and Wellner (2013)) and Bücher et al. (2014).
7
Supplement: Technical Details
P
We write En [h(Y )] = 1/n nk=1 h(Yk ), and use the abbreviation
√
√
k n Fn − F )kG = sup n En [h(Y )] − E [h(Y )]
h∈G
for a class of measurable functions G.
Recall from Holzmann and Klar (2016) the identity
Z ∞
Z
Iτ (x, F ) = τ
1 − F (y) dy − (1 − τ )
F (y) dy.
(8)
−∞
x
7.1
x
Details for the Proof of Theorem 1
We start with some technical preliminaries.
Lemma 11. We have that for x1 , x2 ∈ R,
Iτ (x1 , F ) − Iτ (x2 , F ) = (x2 − x1 ) τ + 1 − 2 τ
Z
0
9
1
F x2 + s(x1 − x2 ) ds .
(9)
Proof. Define g(s) = Iτ (x2 + s(x1 − x2 ), F ), s ∈ [0, 1], so that Iτ (x1 , F ) − Iτ (x2 , F ) =
g(1) − g(0). The map g is continuous, in addition it is decreasing, if x1 ≥ x2 , and increasing
otherwise. Hence it is of bounded variation and Theorem 7.23, Thomson et al. (2008), yields
Z 1
g(1) − g(0) =
g 0 (s) ds + µg ({s ∈ [0, 1] | g 0 (s) = ±∞}),
(10)
0
where µg is the Lebesgue-Stieltjes signed measure associated with g. From Holzmann and Klar
(2016), the right- and left-sided derivatives of Iτ (x, F ) are given by
∂+
∂−
Iτ (x, F ) = − τ + (1 − 2 τ )F (x) ,
Iτ (x, F ) = − τ + (1 − 2 τ )F (x−) . (11)
∂x
∂x
Both derivatives are bounded, such that {s ∈ [0, 1] | g 0 (s) = ±∞} = ∅ in (10), and we obtain
Z 1
Iτ (x1 , F ) − Iτ (x2 , F ) =
g 0 (s) ds
0
Z 1
− τ + 1 − 2 τ F x2 + s(x1 − x2 ) ds.
= (x1 − x2 )
0
Lemma 12. We have
min τl , 1 − τu ≤ τ + (1 − 2τ )s ≤ 3/2,
Proof. For the lower bound,
= 1/2,
τ + (1 − 2 τ )s ≥ τ,
≥ 1 − τ,
τ ∈ [τl , τu ], s ∈ [0, 1].
(12)
if τ = 1/2
if τ < 1/2 ≥ min 1/2, τl , 1 − τu = min τl , 1 − τu .
if τ > 1/2
The upper bound is proved similarly.
Next, we discuss Lipschitz-properties of relevant maps.
Lemma 13. For any x1 , x2 , y ∈ R and τ ∈ [τl , τu ],
Iτ (x1 , y) − Iτ (x2 , y) ≤ x2 − x1
(13)
Further, for any τ, τ 0 ∈ [τl , τu ] and x, y ∈ R,
Iτ (x, y) − Iτ 0 (x, y) ≤ τ − τ 0
|x| + |y|
(14)
Finally, the map τ 7→ µτ , τ ∈ [τl , τu ], is Lipschitz-continuous.
Proof. To show (13), for x1 ≤ x2 ,
Iτ (x1 , y) − Iτ (x2 , y) =
x2 − x1
τ 1 (y > x1 ) + (1 − τ )1 (y ≤ x2 )
≤ x2 − x1 .
As for (14),
Iτ (x, y) − Iτ 0 (x, y) = τ − τ 0 (y − x)1y≥x + (x − y)1y<x
≤ τ − τ 0 |x| + |y| .
For the Lipschitz-continuity of µ· , we use Corollary 1 of Beyn and Rieger (2011) for the function
x 7→ Iτ (x, F ), x ∈ BR (µτ ), for appropriately chosen R > 0, where BR (µτ ) is the open ball
around µτ with radius R. We observe that
10
(i) x 7→ Iτ (x, F ) is continuous for any τ ∈ [τl , τu ], which is immediate from (13),
(ii) x 7→ Iτ (x, F ) fulfils
Iτ (x1 , F ) − Iτ (x2 , F )
x1 − x2 ≤ −a (x1 − x2 )2
with a = min{τl , 1 − τu } > 0. This is clear from (9) and (12).
Let τ, τ 0 ∈ [τl , τu ], and set z = Iτ (µτ 0 , F ). Using (ii) above yields
1
1
z =
Iτ (µτ 0 , F ) − Iτ (µτ , F ) ≤ µτ − µτ 0 ≤ µτu − µτl .
a
a
Choosing R = µτu − µτl gives [µτl , µτu ] ⊂ BR (µτ ) ⊂ [µτl − R, µτu + R], τ ∈ [τl , τu ],
since µτ ∈ [µτl , µτu ], and Corollary 1, Beyn and Rieger (2011), now gives a x̄ ∈ BR (µτ ) with
Iτ (x̄, F ) = z and
1
µτ − x̄ ≤
z.
a
Since x 7→ Iτ (x, F ) is strictly decreasing and [µτl , µτu ] ⊂ BR (µτ ) as well as Iτ (x̄, F ) = z =
Iτ (µτ 0 , F ), we obtain x̄ = µτ 0 . We conclude that
1
1
Iτ (µτ 0 , F ) =
E [Iτ (µτ 0 , Y ) − Iτ 0 (µτ 0 , Y )]
a
a
|µτu | ∨ |µτl | + E [|Y |]
|µτ 0 | + E [|Y |]
≤ τ − τ0
,
≤ τ − τ0
a
a
µτ − µτ 0 ≤
(15)
where we used (14).
Details for Step 1.
Proof of Lemma 7. Proof of (2).
By Lemma 13 the function class
F = y 7→ −Iτ (µτ , y) | τ ∈ [τl , τu ]
is Lipschitz-continuous in the parameter τ for given y, and the Lipschitz constant (which depends
on y) is square-integrable under F . Indeed, the triangle inequality first gives
Iτ (µτ , y) − Iτ 0 (µτ 0 , y) ≤ Iτ (µτ , y) − Iτ 0 (µτ , y) + Iτ 0 (µτ , y) − Iτ 0 (µτ 0 , y) .
Using (14) the first summand on the right fulfils
Iτ (µτ , y) − Iτ 0 (µτ , y) ≤ |τ − τ 0 | (|µτl | ∨ |µτu | + |y|),
and the second is bounded by
Iτ 0 (µτ , y) − Iτ 0 (µτ 0 , y) ≤ |µτ − µτ 0 | ≤ |τ − τ 0 |
|µτu | ∨ |µτl | + E [|Y |]
,
a
utilizing (13) and (15). Thus
Iτ (µτ , y) − Iτ 0 (µτ 0 , y) ≤ |τ − τ 0 | (C + |y|)
(16)
for some constant C ≥ 1. By example 19.7 in combination with
Theorem 19.5 in van der Vaart
√
(2000), F is a Donsker class, so that n ψn (µ· ) − ψ0 (µ· ) converges to the process Z. The
11
same reasoning as in Theorem 8, Holzmann and Klar (2016), then shows continuity of the sample
paths of Z with respect to the Euclidean distance on [τl , τu ].
Proof of (3).
Setting
Fδn = y 7→ Iτ µτ + x, y − Iτ µτ , y | |x| ≤ δn , τ ∈ [τl , τu ]
we estimate that
sup
sup
√
n ψn (µ· + ϕ)(τ ) − ψ0 (µ· + ϕ)(τ ) − ψn (µ· )(τ ) − ψ0 (µ· )(τ )
kϕk[τl ,τu ] ≤δn τ ∈[τl ,τu ]
√
is smaller than k n Fn − F )kFδn . From the triangle inequality, for any τ, τ 0 ∈ [τl , τu ] and
x, x0 ∈ [−δ1 , δ1 ] we first obtain
Iτ µτ + x, y − Iτ µτ , y − Iτ 0 µτ 0 + x0 , y − Iτ 0 µτ 0 , y
≤ Iτ µτ + x, y − Iτ 0 µτ 0 + x0 , y + Iτ µτ , y − Iτ 0 µτ 0 , y ,
where the second term was discussed above and the first can be handled likewise to conclude
Iτ µτ + x, y − Iτ 0 µτ 0 + x0 , y ≤ |τ − τ 0 | + |x − x0 | C + δ1 + |y|
(17)
with the same C as above. Hence
Iτ µτ + x, y − Iτ µτ , y − Iτ 0 µτ 0 + x0 , y − Iτ 0 µτ 0 , y
≤ m(y) |τ − τ 0 | + |x − x0 |
with Lipschitz-constant
m(y) = 2 C + δ1 + 2 |y|,
which is square-integrable by assumption
on F . By example 19.7 in van der Vaart (2000) the
bracketing number N[ ] , Fδ1 , L2 (F ) of Fδ1 is of order −2 , so that for the bracketing integral
J[ ] n , Fδn , L2 (F ) ≤ J[ ] n , Fδ1 , L2 (F ) → 0 as n → 0.
From (13), the class Fδn has envelope δn , and hence using Corollary 19.35 in van der Vaart
(2000), we obtain
√
E k n Fn − F )kFδn ≤ J[ ] δn , Fδn , L2 (F ) → 0.
(18)
An application of the Markov inequality ends the proof of (3).
Details for Step 2.
Proof of Lemma 8. Given τ ∈ [τl , τu ], by (9) and the lower bound in (12), the function x 7→
Iτ (x, F ) is strictly decreasing, and its image is all of R. Hence, for any z ∈ R there is a unique
x satisfying Iτ (x, F ) = z, which shows that ψ0 is invertible.
Inv
Next for fixed ϕ ∈ `∞ [τl , τu ] the preimage (Iτ (·, F ) ([−kϕk, kϕk]) is by monotonicity an
interval [Lτ , Uτ ], |Lτ |, |Uτ | < ∞. By (8),
Z x
nZ ∞
o Z x
Iτ (x, F ) = τ
1 − F (y) dy +
F (y) dy −
F (y) dy,
−∞
x
−∞
thus the map τ 7→ Iτ (x, F ) is increasing, showing Lτ 0 ≤ Lτ and Uτ 0 ≤ Uτ for τ ≥ τ 0 . Hence
the solution of z = Iτ (x, F ) for z ∈ [−kϕk, kϕk] lies in [Lτl , Uτu ], which means that ψ0Inv (ϕ) is
bounded.
12
Before we turn to the proof of Lemma 9, we require the following technical assertions about
ψ0 .
Inv
Lemma 14. Given t > 0 and ν ∈ `∞ [τl , τu ], we have that
t−1 ψ0Inv (t ν) − ψ0Inv (0) (τ )
Z
n
= ν(τ ) τ + (1 − 2 τ )
1
F µτ + s ψ0Inv (t ν)(τ ) − µτ
ds
o−1
.
(19)
0
In particular, if νn ∈ `∞ [τl , τu ] with kνn k → 0, then kψ0Inv (νn )(·) − µ· k → 0, so that for any
τ ∈ [τl , τu ] and τn → τ ,
ψ0Inv (νn )(τn ) − µτn → 0.
(20)
Proof. For the first statement, given ρ ∈ `∞ [τl , τu ] it follows from (9) that
ψ0 ρ (τ ) = ψ0 µ· + ρ − µ· (τ ) − ψ0 µ· (τ )
= − Iτ µτ + ρ(τ ) − µτ , F − Iτ (µτ , F )
Z 1
o
n
F µτ + s(ρ(τ ) − µτ ) ds .
= ρ(τ ) − µτ τ + (1 − 2 τ )
0
The integral on the right hand side is bounded away from zero by (12) and thus choosing ρ =
ψ0Inv (t ν), observing µτ = ψ0Inv (0)(τ ) and reorganising the above equation leads to
t−1 ψ0Inv (0 + t ν) − ψ0Inv (0) (τ )
Z 1
n
o−1
−1
Inv
= t ψ0 ψ0 t ν (τ ) τ + (1 − 2 τ )
F µτ + s ψ0Inv (t ν)(τ ) − µτ ds
0
(
)−1
Z 1
Inv
= ν(τ ) τ + (1 − 2 τ )
F µτ + s ψ0 (t ν)(τ ) − µτ ds
.
(21)
0
In the sequel, given ϕ ∈ `∞ [τl , τu ] let us introduce the function
Z
cϕ (τ ) = τ + (1 − 2 τ )
1
F µτ + s ϕ(τ ) ds.
(22)
0
By (12), min τl , 1 − τu ≤ cϕ ≤ 3/2 holds uniformly for any ϕ.
Now, for the second part, set ϕn = ψ0Inv (νn )(·) − µ· , then (19) gives with t = 1
ψ0Inv (νn )(·) − µ· ≤ νn kcϕn k−1 ≤ νn
−1
min τl , 1 − τu
→ 0.
Then (20) follows by continuity of µ· , see Lemma 13.
Proof of Lemma 9. Let tn → 0, tn > 0, (ϕn )n ⊂ `∞ [τl , τu ] with ϕn → ϕ ∈ C[τl , τu ] with
respect to dhypi and thus uniformly by Proposition 2.1 in Bücher et al. (2014). From (19), and
using the notation (22) we can write
Inv
Inv
t−1
κn (τ ) = ψ0Inv (tn ϕn )(τ ) − µτ
n ψ0 (tn ϕn ) − ψ0 (0) = ϕn /cκn ,
13
and we need to show that
ϕn /cκn → ψ̇ Inv (ϕ) = ϕ/c0
(23)
with respect to dhypi , where c0 is as in (22) with ϕ = 0.
Now, since ϕn → ϕ uniformly and ϕ is continuous, to obtain (23) if suffices by Lemma 19,
i) and iii), to show that cκn → c0 under dhypi . To this end, by Lemma A.4, Bücher et al. (2014)
and Lemma 19, iii), it suffices to show that under dhypi
1
Z
F µτ + s ψ0Inv (tn ϕn )(τ ) − µτ
hn (τ ) =
ds → F (µτ ) = h(τ ),
(24)
0
for which we shall use Corollary A.7 in Bücher et al. (2014). Let
T = [τl , τu ],
S = T \ τ ∈ [τl , τu ] | F is not continuous in µτ ,
so that S is dense in T and h|S is continuous. Using the notation from Bücher et al. (2014),
Appendix A.2, we have that
h|S
S:T
∧
= h∧ = F (µ· −) and
h|S
S:T
∨
= h∨ = h,
(25)
where the first equalities follow from the discussion in Bücher et al. (2014), Appendix A.2, and
the second equalities from Lemma 19, ii) below, and where F (x−) = limt↑x F (t) denotes the
left-sided limit of F at x. If we show that
(i) for all τ ∈ [τl , τu ] with τn → τ it holds that lim inf n hn (τn ) ≥ F (µτ −) and
(ii) for all τ ∈ [τl , τu ] with τn → τ it holds that lim supn hn (τn ) ≤ F (µτ ),
Corollary A.7 in Bücher et al. (2014) implies (24), which concludes the proof of the convergence
in (23).
To this end, concerning (i), we compute that
Z 1
F (µτ −) ≤
lim inf F µτn + s ψ0Inv (tn ϕn )(τn ) − µτn ds
n
0
Z 1
≤ lim inf
F µτn + s ψ0Inv (tn ϕn )(τn ) − µτn ds = lim inf hn (τn ),
n
n
0
where the first inequality follows from (20) and the fact that F (µτ −) ≤ F (µτ ), and the second
inequality follows from Fatou’s lemma. For (ii) we argue analogously
Z 1
F (µτ ) ≥
lim sup F µτn + s ψ0Inv (tn ϕn )(τn ) − µτn ds
n
0
Z 1
≥ lim sup
F µτn + s ψ0Inv (tn ϕn )(τn ) − µτn ds = lim sup hn (τn ).
n
n
0
This concludes the proof of the lemma.
14
7.2
Details for the proof of Theorem 4
We let ψn∗ (ϕ)(τ ) = −Iτ ϕ(τ ), Fn∗ , ϕ ∈ `∞ [τl , τu ], and denote by P∗n the conditional law of
Y1∗ , . . . , Yn∗ given Y1 , . . . , Yn , and by E∗n expectation under this conditional law.
Lemma 15. We have, almost surely, conditionally on Y1 , Y2 , . . ., the following statements.
(i) If E [|Y |] < ∞, then
µ∗τ,n − µ̂τ,n = oP∗n (1).
(26)
(ii) Weakly in (`∞ [τl , τu ], k · k) it holds that
√
n ψn∗ (µ̂·,n ) − ψn (µ̂·,n ) → Z
(27)
sup
τ ∈[τl ,τu ]
Now assume E Y 2 < ∞.
with Z as in Theorem 1.
(iii) For every sequence δn → 0 it holds that
√
n ψn∗ (µ̂·,n + ϕ)(τ ) − ψn (µ̂·,n + ϕ)(τ )
sup
sup
kϕk≤δn τ ∈[τl ,τu ]
(28)
− ψn∗ (µ̂·,n )(τ ) − ψn (µ̂·,n +)(τ ) = oP∗n (1).
(iv) Weakly in (`∞ [τl , τu ], k · k) we have that
√
n ψn (µ∗·,n ) − ψn (µ̂·,n ) → Z.
(29)
Proof of Lemma 15. First consider (26). We start with individual consistency, the proof of which
is inspired by Lemma 5.10, van der Vaart (2000). Since Iτ (µ∗τ,n , Fn∗ ) = 0 and x 7→ Iτ (x, Fn∗ )
is strictly decreasing, for any ε, ηl , ηu > 0 the inequality Iτ (µ̂τ,n − ε, Fn∗ ) > ηl implies µ∗τ,n >
µ̂τ,n − ε and from Iτ (µ̂τ,n + ε, Fn∗ ) < −ηu it follows that µ∗τ,n < µ̂τ,n + ε. Thus
P∗n Iτ (µ̂τ,n − ε, Fn∗ ) > ηl , Iτ (µ̂τ,n + ε, Fn∗ ) < −ηu ≤ P∗n µ̂τ,n − ε < µ∗τ,n < µ̂τ,n + ε
and it suffices to show almost sure convergence of the left hand side to 1for appropriately chosen
ηl , ηu > 0, forwhich it is enough to deduce Pn∗ Iτ (µ̂τ,n − ε, Fn∗ ) > ηl → 1 and Pn∗ Iτ (µ̂τ,n +
ε, Fn∗ ) < −ηu → 1 almost surely. Choose 2 ηl = Iτ (µτ − ε, F ) 6= 0 to obtain
P∗n Iτ (µ̂τ,n − ε, Fn∗ ) > ηl ≥ P∗n Iτ (µτ − ε, F ) − |Iτ (µ̂τ,n − ε, Fn∗ ) − Iτ (µτ − ε, F )| > ηl
≥ P∗n |Iτ (µ̂τ,n − ε, Fn∗ ) − Iτ (µτ − ε, F )| < ηl
by the inverse triangle inequality. Similar for 2 ηu = Iτ (µτ + ε, F ) 6= 0 we get that
P∗n Iτ (µ∗τ,n + ε, Fn∗ ) < −ηu ≥ P∗n |Iτ (µ∗τ,n + ε, Fn∗ ) − Iτ (µτ + ε, F )| < ηu .
In both inequalities the right hand side converges to 1 almost surely, provided that almost surely,
Iτ (µ̂τ,n ± ε, Fn∗ ) → Iτ (µτ ± ε, F ) in probability conditionally on Y1 , Y2 , . . .. To this end, start
with
E∗n [Iτ (µ̂τ,n
±
ε, Fn∗ )]
=
n
X
P∗n (Y1∗ = Yi ) Iτ (µ̂τ,n ± ε, Yi ) = Iτ (µ̂τ,n ± ε, Fn ),
i=1
15
so that it remains to show convergence of the right hand side to Iτ (µτ ± ε, F ) for almost every
sequence Y1 , Y2 , . . .. To this end it holds that
Iτ (µ̂τ,n ± ε, Fn ) − Iτ (µτ ± ε, Fn ) ≤ µ̂τ,n − µτ
by Lemma 13, where the bound converges to 0 by the strong consistency of µ̂τ,n (Holzmann and
Klar, 2016, Theorem 2). Further Iτ (µτ ±ε, Fn ) → Iτ (µτ ±ε, F ) almost surely by the strong law
of large numbers, thus Iτ (µ̂τ,n ± ε, Fn ) converges to Iτ (µτ ± ε, F ) for almost every sequence
Y1 , Y2 , . . ., what concludes the proof of individual consistency of µ∗τ,n . To strengthen this to
uniform consistency, we use a Glivenko-Cantelli argument as in Holzmann and Klar (2016),
Theorem 2. Let dn = µ̂τu ,n − µ̂τl ,n and observe that dn → d = µτu − µτl almost surely by
consistency of µ̂τ,n . Let m ∈ N and choose τl = τ0 ≤ τ1 ≤ . . . ≤ τm = τu such that
µ̂τk ,n = µ̂τl ,n +
k dn
m
(k ∈ {1, . . . , m}),
which is possible because of the continuity of τ 7→ µ̂τ,n . Because the expectile functional is
strictly increasing in τ it follows that
µ∗τk ,n − µ̂τk+1 ,n ≤ µ∗τ,n − µ̂τ,n ≤ µ∗τk+1 ,n − µ̂τk ,n
(τk ≤ τ ≤ τk+1 ).
This implies
sup
τ ∈[τl ,τu ]
µ∗τ,n − µ̂τ,n ≤ max µ∗τk ,n − µ̂τk ,n +
1≤k≤m
dn
,
m
hence
lim sup kµ∗τ,n − µ̂τ,n k[τl ,τu ] ≤ lim sup max µ∗τk ,n − µ̂τk ,n + lim sup
n
n
1≤k≤m
n
dn
d
=
m
m
conditionally in probability for almost every sequence Y1 , Y2 , . . .. Letting m → ∞ completes
the proof.
Proof of (27). The idea is to use Theorem 19.28, van der Vaart (2000), for the random class
Fn = y 7→ −Iτ (µ̂τ,n , y) | τ ∈ [τl , τu ] ,
which is a subset of
Fηn = y 7→ −Iτ (µτ + x, y) | |x| ≤ ηn , τ ∈ [τl , τu ]
for the sequence ηn = supτ ∈[τl ,τu ] |µ̂τ,n − µτ |, hence, almost surely,
J[ ] (ε, Fn , L2 (Fn )) ≤ J[ ] (ε, Fηn , L2 (Fn )).
The class Fηn has envelope (|µτl | + |µτu | + ηn + y), which satisfies the Lindeberg condition.
By (17) the class Fηn is a class consisting of Lipschitz-functions with Lipschitz-constant
mn (y) = (C + ηn + |y|) for some C ≥ 1, such that the bracketing number fulfils
n
ηn + τu − τl o2
N[ ] (δ, Fηn , L2 (Fn )) ≤ K En mn (Y )2
δ
by Example 19.7, van der Vaart (2000),
where
K is some constant
not depending on n. By the
2
2
strong law of large numbers, En mn (Y ) → E (C + |Y |) holds almost surely, in addition
16
ηn → 0 almost surely by the strong consistency of µ̂τ,n , such that the above bracketing number
is of order 1/δ 2 . Thus the bracketing integral J[ ] (εn , Fηn , L2 (Fn )) converges to 0 almost surely
for every sequence εn & 0.
Next we show the almost sure convergence of the expectation En Iτ (µ̂τ,n , Y )Iτ 0 (µ̂τ 0 ,n , Y )
to E [Iτ (µτ , Y )Iτ 0 (µτ 0 , Y )]. First it holds that
En Iτ (µ̂τ,n , Y )Iτ 0 (µ̂τ 0 ,n , Y ) = En Iτ (µ̂τ,n , Y ) Iτ 0 (µ̂τ 0 ,n , Y ) − Iτ 0 (µτ 0 , Y )
+ En Iτ 0 (µτ 0 , Y ) Iτ (µ̂τ,n , Y ) − Iτ (µτ , Y )
+ En [Iτ (µτ , Y )Iτ 0 (µτ 0 , Y )] ,
where the last summand converges almost surely to E [Iτ (µτ , Y )Iτ 0 (µτ 0 , Y )] by the strong law
of large numbers. For the first term we estimate
En |Iτ (µ̂τ,n , Y )| Iτ 0 (µ̂τ 0 ,n , Y ) − Iτ 0 (µτ 0 , Y ) ≤ |µ̂τ,n | + En [|Y |] µ̂τ 0 ,n − µτ 0
with aid of Lemma 13, where the upper bound converges to 0 almost surely by the strong law of
large numbers, strong consistency of µ̂τ 0 ,n and boundedness of µ̂τ,n , which in fact also follows
from the strong consistency of the empirical expectile
and since τ ∈ [τl , τu ]. The remaining
summand above is treated likewise, hence En Iτ (µ̂τ,n , Y )Iτ 0 (µ̂τ 0 ,n , Y ) indeed converges almost surely to the asserted limit. The assertion now follows from Theorem 19.28, van der Vaart
(2000).
Proof of (28).
sup
sup
kϕk≤δn τ ∈[τl ,τu ]
= sup
sup
kϕk≤δn τ ∈[τl ,τu ]
≤
sup
Set ηn = supτ ∈[τl ,τu ] |µ̂τ,n − µτ | again, then as a first step we can estimate
√
√
n ψn∗ (µ̂·,n + ϕ)(τ ) − ψn (µ̂·,n + ϕ)(τ ) − ψn∗ (µ̂·,n )(τ ) − ψn (µ̂·,n )(τ )
n ψn∗ µ· + (µ̂·,n − µ· + ϕ) (τ ) − ψn µ· + (µ̂·,n − µ· + ϕ) (τ )
sup
|x1 |,|x2 |≤δn +ηn τ ∈[τl ,τu ]
√
− ψn∗ µ· + (µ̂·,n − µ· ) (τ ) − ψn µ· + (µ̂·,n − µ· ) (τ )
n ψn∗ (µ· + x1 )(τ ) − ψn (µ· + x1 )(τ )
− ψn∗ (µ· + x2 )(τ ) − ψn (µ· + x2 )(τ ) ,
√
so that for (28) it suffices to show the almost sure convergence E∗n k n(Fn∗ − Fn )kFρn → 0
for the class
Fρn = y 7→ Iτ µτ + x1 , y − Iτ µτ + x2 , y |x1 |, |x2 | ≤ ρn , τ ∈ [τl , τu ] , ρn = δn + ηn .
We use Corollary 19.35, van der Vaart (2000), which implies that
√
E∗n k n(Fn∗ − Fn )kFρn ≤ J[ ] En mn (Y )2 , Fρn , L2 (Fn )
almost surely, where mn (y) is an envelope function for Fρn . The proof consists of finding this
envelope and determining the order of the bracketing integral. By using Lemma 13 the class Fρn
consists of Lipschitz-functions, as for any τ, τ 0 ∈ [τl , τu ], x1 , x01 , x2 , x02 ∈ [−ρn , ρn ] the almost
sure inequality
Iτ (µτ + x1 , y) − Iτ (µτ + x2 , y) − Iτ 0 (µτ 0 + x01 , y) − Iτ 0 (µτ 0 + x02 , y)
≤ |x1 − x01 | + |x2 − x02 | + |τ − τ 0 | 2 C + ρn + |y|
(30)
17
is true for some constant C > 0, see also (17). Thus the bracketing integral J[ ] (εn , Fρn , L2 (Fn ))
converges to 0 almost surely for any sequence εn → 0 with the same arguments as above.
Finally, by (30) the function mn (y) = ηn 4 (C + |y|) is an envelope for Fρn . By the strong
lawof large numbers,
the square integrability of Y and since ηn → 0 almost surely it holds that
2
En mn (Y ) → 0 almost surely, so that
√
E∗n k n(Fn∗ − Fn )kFρn ≤ J[ ] En mn (Y )2 , Fρn , L2 (Fn ) → 0
is valid for almost every sequence Y1 , Y2 , . . ., which concludes the proof.
Proof of (29). By (26) and (28), and since ψn (µ̂·,n ) = ψn∗ (µ∗·,n ) = 0, we have that almost
surely,
√
√
n ψn (µ∗·,n ) − ψn (µ̂·,n ) = n ψn (µ∗·,n ) − ψn∗ (µ∗·,n )
√
= − n ψn∗ (µ̂·,n ) − ψn (µ̂·,n ) + oP∗n (1)
The right hand side converges to −Z conditionally in distribution, almost surely, by (27), which
equals Z in distribution.
Lemma 16. The map ψn is invertible. Further, if tn → 0, ϕ ∈ C[τl , τu ] and ϕn → ϕ with
respect to dhypi and hence uniformly, we have that almost surely, conditionally on Y1 , Y2 , . . .,
Inv
Inv
Inv
t−1
(31)
n ψn (tn ϕn ) − ψn (0) → ψ̇ (ϕ)
with respect to the hypi-semimetric.
Proof. The first part follows from Lemma 8 with F in ψ0 replaced by Fn in ψn as no specific
assumptions on F were used in that lemma. For (31), with the same calculations as for Lemma 14
we obtain the representation
Inv
Inv
t−1
n ψn (tn ϕn ) − ψn (0) (τ )
(
)Inv
Z 1
= ϕn (τ ) τ + (1 − 2 τ )
Fn µ̂τ,n + s ψnInv (tn ϕn )(τ ) − µ̂τ,n ds
,
0
and we have to prove hypi convergence to ψ̇ Inv (ϕ). By the same reductions as in the proof of
Theorem 9, it suffices to prove the hypi-convergence of
Z 1
hn (τ ) =
Fn µ̂τ,n + s ψnInv (tn ϕn )(τ ) − µ̂τ,n ds
0
to F (µτ ) for almost every sequence Y1 , Y2 , . . .. To this end, observe that for any s ∈ [0, 1]
the sequence µ̂τ,n + s ψnInv (tn ϕn )(τ ) − µ̂τ,n converges to µτ almost surely by the same arguments as in Lemma 14. Since µτ is continuous in τ , the almost sure convergence µ̂τn ,n +
s ψnInv (tn ϕn )(τn ) − µ̂τn ,n → µτ holds for any sequence τn → τ . By adding and subtracting
18
Fn (µ̂τ,n + s(ψnInv (tn ϕn )(τ ) − µ̂τ,n )) and using Lemma 17, we now can estimate
!
Z 1
Inv
lim inf Fn µ̂τ,n + s ψn (tn ϕn )(τ ) − µ̂τ,n
F (µτ −) ≤
n
0
!
Inv
+ lim sup F µ̂τ,n + s ψn (tn ϕn )(τ ) − µ̂τ,n
n
Z
Inv
− Fn µ̂τ,n + s ψn (tn ϕn )(τ ) − µ̂τ,n
ds
1
lim inf Fn µ̂τ,n + s ψnInv (tn ϕn )(τ ) − µ̂τ,n ds + lim sup kFn − F kR
n
n
0
Z 1
≤ lim inf
Fn µ̂τ,n + s ψnInv (tn ϕn )(τ ) − µ̂τ,n ds = lim inf hn (τn )
≤
n
n
0
almost surely, where the ’lim sup’ vanishes due to the Glivenko-Cantelli-Theorem for the empirical distribution function. Similarly we almost surely have
Z 1
lim sup Fn µ̂τ,n + s ψnInv (tn ϕn )(τ ) − µ̂τ,n ds + lim inf kFn − F kR
F (µτ ) ≥
n
n
0
Z 1
Fn µ̂τ,n + s ψnInv (tn ϕn )(τ ) − µ̂τ,n ds = lim sup hn (τn ).
≥ lim sup
n
n
0
The proof is concluded as that of Theorem 9 by using Corollary A.7, Bücher et al. (2014) .
√
Inv
Inv
Proof of Theorem 4. Set tn = 1/ n and define the function gn (ϕ) = t−1
n (ψn (tn ϕ)−ψn (0)).
Then from (31) the hypi-convergence gn (ϕn ) → ψ̇ Inv (ϕ) holds almost surely, whenever
ϕ ∈
√
C[τl , τu ] and ϕn → ϕ with respect to dhypi . In addition n ψn (µ∗·,n ) − ψn (µ̂·,n ) → Z conditional in distribution with respect to the sup-norm, almost surely, by (27), where Z is continuous
almost surely. Hence the convergence is also valid with respect to dhypi , such that
√
√
n µ∗·,n − µ̂·,n = gn n ψn (µ∗·,n ) − ψn (µ̂·,n ) → ψ̇ Inv (Z)
holds conditionally in distribution, almost surely, by using the extended continuous mapping
theorem, Theorem B.3, in Bücher et al. (2014).
7.3
Details for the proof of Theorem 5
Since F in Theorem 5 is assumed to be continuous and strictly increasing, it is differentiable
almost everywhere. If F is not differentiable in a point x, the assumptions at least guarantee
right- and left-sided limits of f in x. Since redefining a density on a set of measure zero is
possible without changing the density property, we can assume f (x) = f (x−) for such x. In
addition, the assumptions imply that F is invertible with continuous inverse, which we denote
by F Inv .
Proof of Lemma 10. Let tn & 0, αn → α ∈ [αl , αu ] and ϕn , νn ∈ `∞ [αl , αu ] with ϕn →
ϕ ∈ C[αl , αu ] and νn → 0 with respect to dhypi . Then ϕn (αn ) → ϕ(α) and νn (αn ) → 0 by
Proposition 2.1, Bücher et al. (2014). We have to deal with the limes inferior and superior of
Inv
t−1
(αn + νn (αn ) + tn ϕn (αn )) − F Inv (αn + νn (αn )) .
n F
Define the function g : [0, 1] → R with g(s) = F Inv (αn +νn (αn )+s tn ϕn (αn )). The difference
above then equals t−1
n (g(1) − g(0)). The function g is monotonically increasing (or decreasing,
19
depending on the sign of ϕn (αn )) and continuous in s. Hence with Theorem 7.23, Thomson
et al. (2008), we can write
Z 1
−1
tn g(1) − g(0) =
g 0 (s) ds + µg {g 0 = ±∞}
0
Z 1h
i−1
= ϕn (αn )
f F Inv αn + νn (αn ) + s tn ϕn (αn )
ds,
(32)
0
where µg is the Lebesgue-Stieltjes signed measure associated with g. The set {g 0 = ±∞} is
empty, as g 0 = ∞ if and only if f = 0, which is excluded by assumption, so that λg {g 0 =
±∞} = 0. Since ϕn → ϕ ∈ C[αl , αu ], using Lemma 19 i) we only need to deal with the
accumulation points of the integral in (32). To this end let
Z 1
−1
Hn (α) =
f F Inv αn + νn (αn ) + s tn ϕn (αn )
ds.
0
The assertion is that Hn (α) hypi-converges to f (qα )−1 . In the case for the expectile-process
we were able to exploit continuity properties of the asserted semi-derivative and thus utilize
Corollary A.7, Bücher et al. (2014)). In the present context we do not know whether the limit
f (q· )−1 can be attained by extending f (q· )−1 from the dense set where it is continuous. Thus,
we directly show the defining properties of hypi convergence in (6) and (7). For any s ∈ [0, 1] it
holds that F Inv αn + νn (αn ) + s tn ϕn (αn ) → qα by continuity of F Inv , hence we can estimate
Z 1
−1
lim inf
f F Inv αn + νn (αn ) + s tn ϕn (αn )
ds
n
0
Z 1
−1
1
≥
lim sup f F Inv αn + νn (αn ) + s tn ϕn (αn )
ds ≥
.
f∨ (qα )
n
0
with help of Fatou’s Lemma, Lemma 17 and the properties of the hulls. Similarly,
Z 1
−1
lim sup
f F Inv αn + νn (αn ) + s tn ϕn (αn )
ds
n
0
Z 1
−1
1
.
≤
lim inf f F Inv αn + νn (αn ) + s tn ϕn (αn )
ds ≤
n
f∧ (qα )
0
It remains to construct sequences αn such that limn Hn (αn ) equals the respective hull. By
Lemma 19 ii), it holds that
f∧ (qα ) = min f (qα −), f (qα +), f (qα )
and f (qα ) = max f (qα −), f (qα +), f (qα ) .
If F is differentiable in qα , all three values can differ; if not, we have set f (qα ) = f (qα −). We
shall choose a sequence converging from below, above or which is always equal to α, depending
on where the min / max is attained. We do this exemplary for f∨ (qα ) = f (qα −). Since ϕn → ϕ
uniformly and ϕ is bounded, there is a C > 0 for which kϕn k ≤ C. Then choose any sequence
εn & 0 and set αn = α − kνn k − tn C − εn , which is in [αl , αu ] for n big enough. For this
sequence it holds that αn + νn (αn ) + s tn ϕn (αn ) < α for any s ∈ [0, 1], so that the convergence
is from below. Bounded convergence yields
Z 1
−1
f F Inv αn + νn (αn ) + s tn ϕn (αn )
lim
ds
n
0
Z 1
−1
1
=
lim f F Inv αn + νn (αn ) + s tn ϕn (αn )
ds =
.
n
f (qα −)
0
20
7.4
Properties of hypi-convergence
An major technical issue in the above argument is to determine hypi-convergence of sums, products and quotients of hypi-convergent functions. We shall require the following basic relations
between lim sup and lim inf.
Lemma 17. Let (an )n , (bn )n be bounded sequences. Then
lim inf an + lim sup bn ≥ lim inf (an + bn ) ≥ lim inf an + lim inf bn
n→∞
n→∞
n→∞
n→∞
n→∞
and
lim sup an + lim inf bn ≤ lim sup(an + bn ) ≤ lim sup an + lim sup bn .
n→∞
n→∞
n→∞
n→∞
n→∞
If an 6= 0, then
lim inf
n
1
1
=
.
an
lim supn an
Proof. The first pair of inequalities follows from
ak + sup bl ≥ ak + bk ≥ ak + inf bl
l≥n
l≥n
for any n ∈ N, k ≥ n, by applying ’inf k≥n ’ and taking the limit n → ∞. The second pair of
inequalities follows similarly. For the last part note that
1
1
.
=
l≥n al
supl≥n al
inf
Taking n → ∞ yields the asserted equality.
Lemma 18. Let (bn )n be a bounded sequence and let (an )n be convergent with limit a ∈ R.
Then
lim inf an bn = lim inf a bn ,
n
lim sup an bn = lim sup a bn .
n
n
n
Proof. From Lemma 17,
lim inf (a0 bn − an bn ) + lim inf a0 bn ≤ lim inf an bn ≤ lim sup(a0 bn − an bn ) + lim inf a0 bn ,
n
n
n
n
n
lim inf (a0 bn − an bn ) + lim sup a0 bn ≤ lim sup an bn ≤ lim sup(a0 bn − an bn ) + lim sup a0 bn .
n
n
n
n
n
But a0 bn − an bn ≤ supm |bm | a0 − an → 0, which implies the stated result.
The next lemma was crucial in the proof of Theorem 9.
Lemma 19. Let c, cn , ϕn ∈ `∞ [l, u] and ϕ ∈ C[l, u].
i) If dhypi (ϕn , ϕ), dhypi (cn , c) → 0 holds, then cn ϕn hypi-converges to c ϕ. More precisely
cn ϕn epi-converges to (c ϕ)∧ and hypo-converges to (c ϕ)∨ , where
ϕ c ∧ = ϕ c∧ 1ϕ>0 + c∨ 1ϕ<0 ,
ϕ c ∨ = ϕ c∨ 1ϕ>0 + c∧ 1ϕ<0 .
(33)
21
ii) Assume c has left- and right-sided limits at every point in [l, u]. Then
c∧ (x) = min{c(x−), c(x), c(x+)}
and
c∨ (x) = max{c(x−), c(x), c(x+)}
holds. Especially, if c(x0 −) ≤ c(x0 ) ≤ c(x0 +) or c(x0 −) ≥ c(x0 ) ≥ c(x0 +) for some
x0 ∈ [l, u], then the equalities (c∧ )∨ (x0 ) = c∨ (x0 ) and (c∨ )∧ (x0 ) = c∧ (x0 ) are true.
iii) If cn , c > 0, the convergence dhypi (1/cn , 1/c) → 0 follows from dhypi (cn , c) → 0.
Proof of Lemma 19. From the definition in (5), for a function h ∈ `∞ [l, u] the lower semicontinuous hull h∧ at x ∈ [l, u] is characterized by the following conditions
For any sequence xn → x,
There is a sequence
x0n
→x
we have lim inf h(xn ) ≥ h∧ (x),
n→∞
for which lim h(x0n ) = h∧ (x),
(34)
n→∞
and similarly for h∨ .
Ad i): By continuity of ϕ we have ϕ(xn ) → ϕ(x) for any sequence xn → x. The statement
(33) now follows immediately using (34) and Lemma 18 and noting that for ϕ(x) < 0,
lim inf ϕ(x)c(xn ) = ϕ(x) lim sup c(xn ),
n→∞
lim sup ϕ(x)c(xn ) = ϕ(x) lim inf c(xn ).
n→∞
n→∞
n→∞
Further, by continuity of ϕ, the hypi-convergence of ϕn to ϕ actually implies the uniform convergence. Therefore, for any xn → x we have that ϕn (xn ) → ϕ(x). Using the pointwise criteria (6) and (7) for hypi-convergence, Lemma 18 and (33) we obtain the asserted convergence
ϕn cn → ϕ c with respect to the hypi semi-metric.
Ad ii): The proof of Lemma C.6, Bücher et al. (2014) show that for a function, which admits
right- and left-sided limits, the supremum over a shrinking neighbourhood around a point x
converges to the maximum of the three points c(x−), c(x) and c(x+). The analogues statement
holds for the infimum, which is the first part of ii). From Lemma C.5, Bücher et al. (2014), the
maps x 7→ c(x−) and x 7→ c(x+) both have a right-sided limit equal to c(x+) and a left-sided
limit equal to c(x−), hence this is also true for the functions c∨ (x) = max{c(x−), c(x), c(x+)}
and c∧ (x) = min{c(x−), c(x)c(x+)} . From the above argument, we obtain
(c∨ )∧ (x) = min c(x−), c(x+), max{c(x−), c(x), c(x+)} = min{c(x−), c(x+)}
and
(c∧ )∨ (x) = max c(x−), c(x+), min{c(x−), c(x), c(x+)} = max{c(x−), c(x+)}.
If c(x0 −) ≤ c(x0 ) ≤ c(x0 +) or c(x0 −) ≥ c(x0 ) ≥ c(x0 +), we obtain
c∨ (x0 ) = max{c(x−), c(x+)} = (c∧ )∨ (x0 ),
c∧ (x0 ) = min{c(x−), c(x+)} = (c∨ )∧ (x0 ).
Ad iii): From Lemma 17 (last statement) and (34) we obtain (1/c)∧ = 1/c∨ and (1/c)∨ = 1/c∧ .
The hypi-convergence of 1/cn to these hulls follows similarly from Lemma 17 (last statement)
and the pointwise criteria (6) and (7) for hypi-convergence.
22
| 10math.ST
|
Rethinking probabilistic prediction in the wake of the
2016 U.S. presidential election
Harry Crane∗ and
Ryan Martin†
arXiv:1704.01171v1 [stat.OT] 4 Apr 2017
April 6, 2017
Abstract
To many statisticians and citizens, the outcome of the most recent U.S. presidential election represents a failure of data-driven methods on the grandest scale.
This impression has led to much debate and discussion about how the election
predictions went awry—Were the polls inaccurate? Were the models wrong? Did
we misinterpret the probabilities?—and how they went right—Perhaps the analyses were correct even though the predictions were wrong, that’s just the nature of
probabilistic forecasting. With this in mind, we analyze the election outcome with
respect to a core set of effectiveness principles. Regardless of whether and how
the election predictions were right or wrong, we argue that they were ineffective
in conveying the extent to which the data was informative of the outcome and the
level of uncertainty in making these assessments. Among other things, our analysis sheds light on the shortcomings of the classical interpretations of probability
and its communication to consumers in the form of predictions. We present here
an alternative approach, based on a notion of validity, which offers two immediate
insights for predictive inference. First, the predictions are more conservative, arguably more realistic, and come with certain guarantees on the probability of an
erroneous prediction. Second, our approach easily and naturally reflects the (possibly substantial) uncertainty about the model by outputting plausibilities instead of
probabilities. Had these simple steps been taken by the popular prediction outlets,
the election outcome may not have been so shocking.
Keywords and phrases: Interpretation of probability; plausibility; prediction;
statistical modeling; validity.
1
Introduction
On the morning of November 8, 2016, the day of the United States presidential election,
Nate Silver’s FiveThirtyEight website estimated Hillary Clinton’s chance of being elected
President at 72%, The New York Times estimated Clinton at 91%, The Huffington Post
∗
Department of Statistics & Biostatistics, Rutgers, the State University of New Jersey,
[email protected]
†
Department of Statistics, North Carolina State University, [email protected]
1
98%, and the Princeton Election Consortium 99% [8, 10]. Bolstered and perhaps influenced by these probabilities, the analyses of many political pundits pointed to a likely, if
not inevitable, Clinton victory. Yet, at 3am the next morning, Donald Trump delivered
a victory speech as the President-Elect of the United States.
In the days and weeks that followed, discussion of the historic election quickly turned
into debate over what went wrong and who or what was to blame:
• some blamed the data, including Sam Wang, co-founder of the Princeton Election
Consortium, who told The New York Times that “state polls were off in a way that
has not been seen in previous presidential election years” [10];
• some changed their story, including FiveThirtyEight, who reframed the narrative
from “there’s a wide range of outcomes, and most of them come up Clinton” on the
day of the election [20] to the assertion that there were “plenty of paths to victory
for Trump” and “people should have been better prepared for it” afterwards [22];
• others were in denial, including Irizarry, who was bold enough to declare that
“statistically speaking, Nate Silver, once again, got it right” [7].
Based on this fallout, FiveThirtyEight, The New York Times, Princeton Election
Consortium, and many newcomers will likely put forth “new and improved” methods for
the 2020 election. Time will tell how improved those methods turn out to be. In the
meantime, however, the fields of statistics and data science will continue on much like
they did before the election. For the field at large, the specific methodologies undertaken
to obtain the predictions of 72%, 91%, 98%, 99% for Clinton are of little significance to
the many more salient and consequential applications of statistical methods in science,
medicine, finance, etc. With the interests of broader statistical practice in mind, we
dispense with any attempts to diagnose in what ways the election analyses were “right”
or “wrong.” We instead treat the election analysis meta-statistically, extracting three
main insights and highlighting pressing questions about how probabilistic predictions
ought to be conducted and communicated moving forward.
First, we address the tendency, even among statisticians and data scientists, to be
results-oriented in assessing the effectiveness of their analyses. The widely publicized
success of FiveThirtyEight’s 2008 and 2012 election predictions coupled with larger societal trends toward quantitative analysis fuel a kind of blind faith in data-driven methods.
Trends toward data-based methods across the spectrum of business, industry, and politics
have been accompanied by a shift away from principles in statistical circles. As Speed
stated plainly in a recent column of the IMS Bulletin:
If a statistical analysis is clearly shown to be effective [...] it gains nothing
from being described as principled [23].
But if the effectiveness of a certain method is primarily judged by the fact that it seems
to have worked in hindsight, then what are we left with when such a method fails? In
the wake of the 2016 election, we are left with outrage, confusion, questions, and not
many answers. Moreover, since the effectiveness of any given statistical analysis must
be assessed prior to observing the outcome, such a judgment cannot be entirely resultsdriven. It must be based on principles. In Section 2 we present three basic effectiveness
2
principles for any statistical analysis and discuss these principles in the context of the
2016 election.
Second, we discuss perhaps the most mysterious aspect of election analysis and probabilistic analyses in general: the probabilities themselves. Is the FiveThirtyEight probability of 72% for Clinton to be interpreted as a frequency, a degree of belief, a betting
quotient, or something else? Are we to assess the quality of these probabilities in terms of
coherence, calibration, proper scoring, or some other way? Do these differences matter?
Even among statisticians there is widespread disagreement on how to answer these questions. Worse is the impulse by some to dismiss questions about the meaning of probability
and the quality of predictions.
“[T]he key thing to understand is that [data science] is a tool that is not
necessarily going to give you answers, but probabilities,” said Erik Brynjolfsson, a professor at the Sloan School of Management at the Massachusetts
Institute of Technology.
Mr. Brynjolfsson said that people often do not understand that if the
chance that something will happen is 70 percent, that means there is a 30
percent chance it will not occur. The election performance, he said, is “not
really a shock to data science and statistics. It’s how it works” [10].
In high-impact and high-exposure applications like this election, when the reputation of
data science is on the line, the interpretation and communication of probabilities to the
general public ought to be taken more seriously than Brynjolfsson suggests. After discussing the merits and demerits of the classical interpretations of probability for election
predictions in Section 3, we propose a new approach that avoids the various shortcomings
of traditional perspectives by only requiring an interpretation of small probability values,
à la Cournot [16]. A key feature of this perspective is that the derived predictions are
valid, i.e., the probability of an erroneous prediction is no more than a specified threshold. This validity property is achieved by making more conservative predictions, thereby
avoiding the charge of “overselling precision” raised by some after the election [10].
Third, we make the important distinction, arising out of the discussion of Section 2,
between variability in statistical estimates and uncertainty in the model setup. While
it is common to report the former, it is uncommon to report the latter when making
predictions. More precisely, a statistical model encodes uncertainty in the data, which
can be used to derive measures of variability for estimators, confidence intervals, etc.
But these measures are only reliable insofar as the model is sound for a given application.
They do not account for variability in estimators due to uncertainty in the choice of
model. In Section 4, we describe how to account for this latter type of uncertainty in the
statistical analysis. Specifically, we propose an alternative output of plausibility rather
than probability, which accounts for uncertainty in a natural way that makes predictions
more conservative and assessments of predictive validity more stringent.
Finally, in Section 5, we conclude with some remarks about the statistical community’s
reaction to the election outcomes, and how the perceived failure of data science in the
election predictions might be turned into an opportunity to better communicate the
capabilities and limitations of statistical methods.
3
2
Were the analyses effective?
We begin with the premise that statistical analyses ought to be evaluated based on how
effective they are in providing insight about a scientific question of interest, not whether
they produce correct predictions. In election analysis, the question is simple: Which candidate will win the election? The analyses of several outlets, including FiveThirtyEight,
Princeton Election Consortium, etc., were conducted with this question in mind. Were
those analyses effective?
We first stress that in many statistical applications, the correct answer or outcome
may never be revealed, making a right/wrong classification of predictions impossible.
Even in cases like the election, where the outcome is eventually revealed, an assessment of
effectiveness is only useful if made prior to learning the outcome. Effectiveness, therefore,
cannot be judged by whether or not an approach “worked” in a particular instance, in
terms of giving a “right” or “wrong” prediction. Similarly, a determination of effectiveness
cannot be based on other statistical analyses, for then the effectiveness of the metaanalysis must be established, ad infinitum. Effectiveness must instead be established
by non-statistical means, which include principles, subject matter knowledge, qualitative
assessments, intuition, and logic. In this section, we evaluate the 2016 election analyses
based on a set of effectiveness principles [2], which are necessary to ensure that the three
main components of statistical practice—data, model, and inference—fit together in a
coherent way. In particular, data must be relevant for the question of interest; the model
must be sound for the given data; and the inference/prediction must be valid for the
given model.
Principle 1: The data must be relevant
For a statistical analysis to be effective, the data must be relevant for answering the
question under consideration. With regards to the question, who will win the presidential
election?, common sense suggests that polling data is relevant to the election outcome.
However, this judgment relies on an assumption that those polled are representative of
the general electorate and their responses are truthful (at least at the time the poll
is taken).1 In reality, both of these assumptions—and others from the classical survey
sampling theory—may have been violated to some extent in the 2016 election polling.
The challenge to statistical analysis, however, is not the binary question of relevance
or irrelevance, but rather the manner in which the data may be relevant for the given
question. Although polling data has been useful in the past and continues to be relevant
in the present, the way in which the polls convey this information is likely to change over
time. For example, the advent of social media and the decline of landline telephones has
changed the way in which polls are conducted, which in turn may affect (positively or
negatively) the information we are able to glean from the polls.
In the 2016 election, in particular, several statisticians, including Gelman, have admitted that the relevance of the polls was incorrectly assessed by many of the analyses.
1
For the sake of our discussion, we assume that the analyses carried out by FiveThirtyEight, Princeton
Election Consortium, The New York Times, etc., were based largely on polling data. FiveThirtyEight
offers a polls-only and a polls-plus prediction; the latter includes some additional information about
other current trends, but these estimates rarely differ by more than a few points.
4
So Nate [Silver] and I were both guilty of taking the easy way out and looking
at poll aggregates and not doing the work to get inside the polls. We’re doing
that now, in December, but I/we should’ve been doing it in October. Instead
of obsessing about details of poll aggregation, we should’ve been working more
closely with the raw data [5].
Such comments stand in stark contrast to the excuse by Sam Wang that the “polls were
off” and, more broadly, that
In this presidential election, analysts said, the other big problem was that some
state polls were wrong [10].
From a scientific point of view, the concept that the polls were wrong suggests a
perplexing juxtaposition of the experiment and the experimenter, the science and the
scientist. The polls are the data used by FiveThirtyEight, Princeton Election Consortium,
etc., to make predictions. And what is a data scientist’s job if not to understand data?
Just as it is a biologist’s job to understand the biology and a physicist’s job to understand
the physical world. Blaming a faulty statistical analysis on the data is tantamount to
a physicist blaming an electron or a biologist blaming a lab rat for a flawed scientific
hypothesis. Armed with relevant data, the data scientist is tasked with gleaning whatever
insights the data can provide. Attempts by Wang and others to blame faulty predictions
on erroneous polls, then, only highlight their failure to properly account for the key
sources of variation in the data. This is not an error in the data, but rather in the
modeling of that data.
Principle 2: The model must be sound
The statistical model must be sound in the sense that it accounts for sources of variation,
whether as a result of data collection or natural variation in the underlying system.
Claims, such as Wang’s, that the polls were off are based primarily on the fact that the
usual models did not fit the 2016 polls as well as they have fit past polls. Such reasoning is
backwards. The model assumed that the data was relevant in a way that it was not, likely
because the chosen model failed to adequately account for certain sources of variability
in the polling data, as Gelman points out above.
Leading up to election day, there was ample discussion of a possible “hidden” Trump
vote, akin to the so-called Bradley effect [25] from the 1982 California governor’s race.
The reasoning behind a possible hidden Trump vote was based on sound, albeit nonquantitative, reasoning: quite simply, it was socially unacceptable to openly support
Trump in many parts of the country and, as the theory goes, Trump voters in those regions
would either say nothing or claim to support Clinton but secretly vote for Trump on election day. This speculation was largely dismissed by data scientists because, unironically,
there was no evidence for it in the data. In the post-election analysis, FiveThirtyEight
holds to the claim that “shy” voters were not the reason why the outcome differed from
what the polls suggested [21], but still based on the assumption that the existence of the
“hidden” voter would have manifested itself in the data. Ultimately, it is unclear whether
the Trump vote was actually “hidden” or merely overlooked by the current statistical apparatus. What is clear, however, is that the predictors’ models were consistently unsound
in that they did not properly account for the variability in the polls.
5
The predictions also failed to reflect the uncertainty in what information the polls
conveyed about the eventual outcome. For example, rather than dismiss the possibility
of a hidden Trump vote, or any other explanation for why the polls may have been skewed,
the predictions would have been more insightful if they reflected the model uncertainty in
their reported predictions. While accounting for such uncertainty makes the predictions
less definitive, it makes them more reliable and less sensitive to arbitrary or unjustified
assumptions. We discuss the issue further in Section 4.
Principle 3: The prediction must be valid
For a statistical analysis to be effective, the inference drawn must be valid, meaning
that the conclusions can be justified to someone else who otherwise agrees with the
model setup. For this, it must be clear how the statistical output is to be interpreted.
In the present context, the output assigns a percentage to each candidate—e.g., 72%
for Clinton and 28% for Trump—and inference corresponds to a prediction based on
the output. Assessing the quality of these inferences requires an understanding of how
these probabilities are to be interpreted. Therefore, the probabilities themselves do not
constitute a prediction—a prediction is an interpretation of the probabilities.
For example, on October 24, 2016, Silver [19] estimated Clinton’s probability of victory
at 83–85%. He interpreted this figure to mean:
Hillary Clinton is probably going to become the next president.
Where does this conclusion come from? Perhaps Silver interpreted the estimated 83%/17%
Clinton/Trump probabilities (at that time) in terms of betting odds, making Clinton
roughly a 5-to-1 favorite to win. Or maybe he instead interprets these percentages vaguely
as “measures of risk,” as he said in a TV interview after the election [13]. Silver is free
to interpret these values as he pleases when arriving at his prediction that Clinton will
“probably” win, but a clear interpretation is necessary if the 83% probability is to convey any additional information in terms of how likely the prediction is to be correct.
In particular, Silver’s choice of the word probably is not insignificant; it suggests some
vague level of certainty, apparently stronger than “maybe” and weaker than “definitely,”
which can have a significant influence on the opinions—and even actions—of the average
FiveThirtyEight consumer. For example, contrast Silver’s prediction that Clinton will
“probably” win based on probability 83% on October 24th to his prediction that she will
“most likely” win based on probability 72% on November 8 [20]. Of course, 83% is bigger
than 17% so, if forced to choose between the two, it makes sense that Silver would choose
Clinton. But nothing forces Silver, or any one else, to make such a choice: one could,
quite reasonably, declare the race “too close to call.”
Since the various outlets, FiveThirtyEight included, offer many different numbers
without a clearly spelled out rationale for how to interpret them, it is hard to argue for
the validity of any one of these predictions. And, furthermore, it is unclear whether and
how to compare probabilities from different sources. The Princeton Election Consortium’s
99% probability conveys greater confidence in a Clinton win than FiveThirtyEight’s 72%,
but on what scale are we to make such comparisons?
6
3
Valid predictions
Consider the pair (X, Y ), where X represents polling data and any other information used
to arrive at a prediction and Y is the actual outcome of the election. In general, Y takes
values in the finite set Y consisting of all eligible presidential candidates, i.e., natural-born
U.S. citizens over the age of thirty-five who have resided within the United States for at
least fourteen years as of 2016. Much of our discussion below, however, specializes to the
case in which Y takes only two values, i.e., Y = {c, t} is binary, Clinton (c) or Trump
(t). The decision to disregard other candidates (Johnson, Stein, McMullen, etc.) is itself
an inference based on the fact that every conceivable model, as of early November 2016,
gave these candidates a negligible win probability; see Section 3.2.
Prediction probabilities are determined by a mapping x 7→ πx that takes data x to
a probability distribution πx on Y. Technically, πx could be any (possibly x-dependent)
distribution, but, following our discussion in Section 2, a natural approach is to start
with a model, i.e., a joint distribution P for (X, Y ), and to take πx as the conditional
distribution of Y , given X = x, under P, namely,
πx (y) = P(Y = y | X = x),
y ∈ Y.
(1)
For example, FiveThirtyEight’s final reported probabilities on election day were πx (c) =
0.72 and πx (t) = 0.28. It is not our intention here to question the modeling decisions
that produced these numbers. In fact, Silver’s analysis seems to be based on too many
ad hoc, inarticulable assumptions to evaluate in any scientific or systematic way, e.g.,
Silver’s weighting scheme for certain polls based on his own subjective grading of their
presumed accuracy. Instead, we ask some more basic questions: What do these numbers
mean? What do they tell us about the election outcome?
3.1
Probabilities and their interpretation
In the terminology of classical debates over probability, are these election probabilities to
be interpreted as frequencies, subjective degrees of belief, betting quotients, or something
else? It turns out that there are elements of each interpretation, at least in the predictions
from FiveThirtyEight. The discussion that follows highlights the difficulty in interpreting
these probabilities from these classical points of view, and we offer a first look at an
alternative approach that may be fruitful.
Objective probabilities and frequencies. In the case of FiveThirtyEight, this 72%
figure represents the frequency of times Clinton won in a large number of simulations
from their mathematical model. In this sense, the probabilities are frequencies, but how
do we interpret such a frequency in the context of the real world? The frequency of
72% derived from the FiveThirtyEight simulations is a synthetic frequency in that it
is calculated by repeatedly sampling from an idealized model for the election outcome.
It does not represent a frequency in the same sense as frequencies obtained by tossing
a physical coin a large number of times. And, of course, the concept of repeating the
election a large number of times is a fantasy.
Nevertheless, some of the post-election confusion stemmed from a widespread feeling
in the general public that prediction probabilities correspond to some objective reality.
7
For example, FiveThirtyEight’s final pre-election report hints at a real-world interpretation by stating Clinton’s “chance of winning” at 72% [20]; and The New York Times made
a similar statement, giving Clinton an 85% “chance to win” [9]. There is a tendency to
characterize Trump’s victory as “improbable” based on the low percentages he was given
under the models by FiveThirtyEight, Princeton Election Consortium, The New York
Times, and many others, but such an assessment affords those predictions an objective
interpretation that may not be warranted. For example, suppose that an urn contains
100 balls—50 blue and 50 red—and a ball is to be chosen completely at random. If we
know the composition of the urn, then neither color is improbable. On the other hand, if
we don’t know the composition of the urn, but we assume that 1 ball is blue and 99 are
red, then we would assign “low credence,” say, to drawing a blue ball even though such
an outcome is by no means improbable.
Subjective probabilities and calibration. Interpreting the probabilities as objective frequencies fails to account for the sensitivity of these probabilities to subjective
modeling decisions. In fact, the probabilities reported by FiveThirtyEight and others are
quite sensitive to subjective modeling assumptions; for example, on October 24th, Silver
estimated Clinton’s chance at 85%, but also discussed how the estimate would increase
to anywhere from 90% to 99.8% if certain assumptions were modified [19].
Instead of interpreting these percentages as the long run frequency that the specific
event “Clinton is elected president in 2016” occurs in a large number of repeated trials
of the election, we can possibly interpret the probabilities as frequencies relative to the
mechanism by which the percentages were derived. Specifically, we may interpret the
72% probability for Clinton in terms of how often we can expect any arbitrary event E
to occur when assigned probability 72% by the methods employed by FiveThirtyEight. If
the class of events E to which the FiveThirtyEight simulations assign a 72% probability
occur roughly 72% of the time, then the 72% probability may have a somewhat objective
interpretation as a frequency in terms of the mechanism by which it was calculated, but
not in terms of the event that Clinton will win.
With this interpretation, one can supplement the model specification step with checks
of how well the specified model fits the available data [14], with the goal of calibrating the
subjective probabilities to some kind of long-run frequencies. A necessary consequence of
this form of calibration is that 72% predictions have to be wrong about 28% of the time
[5]. Further shortcomings of calibration are highlighted in [18].
3.2
A new perspective—validity
Generally speaking, it seems safe to assume that consumers of the election predictions
don’t care about the various interpretations of probability, e.g., betting odds, frequencies,
etc. They just want the prediction to be correct. Since predictions come before the
election is decided, there is uncertainty in these predictions, and the only logical reason
to put stock in a given prediction is if the prediction is, in some sense, unlikely to be
wrong. With this in mind, we propose the condition of predictive validity, which we
describe informally as follows:
a prediction derived from a given method is valid if the probability, relative to
the posited model, that the method gives a wrong prediction is not too large.
8
This notion is closely related to Fisher’s logic of statistical inference and Occam’s razor:
given the model, if the prediction derived from a particular method is incorrect with low
probability, then there is no reason to doubt the prediction made in a particular case.
This goes beyond the idea of calibration discussed above in the sense that validity allows
us to control the accuracy of our prediction, to mitigate our damages so to speak. There
are also some connections with the guarantees advocated by [11, 12].
As we discuss further below, an important aspect of predictive validity is that it only
requires that we can interpret small probabilities. This feature aligns with the rationale of
Cournot’s principle [16], according to which small (and therefore large) probabilities are
the only interpretable or actionable probabilities. For example, a moderate probability of
40% for an event E is likely too large for us to comfortably predict that E will not occur,
that is, to “take action” in predicting that its complement, E c , will occur. But we may
well be comfortable to predict E c if the probability of E were, say, 0.001. Our proposed
approach also lends itself to accounting for uncertainty in the model, thereby avoiding
an exaggeration of the quality of predictions, the “overselling of precision” discussed by
some after the election [10]. We discuss this further in Section 4.
3.3
Prediction sets and their validity
By Cournot’s principle, the only meaningful or actionable probabilities are the small
ones. In keeping with this principle, our proposal is to avoid directly interpreting those
probabilities that are not small, and thus not easily interpretable. Towards making
this more formal, for a given prediction rule x 7→ πx and a suitably small α > 0 (see
Proposition 2), we define the α-level prediction set as
Πx (α) = y ∈ Y : πx (y) > α .
(2)
This set consists of all those election outcomes that are at least α-probable according
to the given prediction rule and the given polling data x. We propose to report the
prediction set Πx (α) instead of committing to choose a single candidate and then potentially struggling to interpret the magnitudes of the individual prediction probabilities
when making this choice. By working with the set in (2), we only commit ourselves to
interpreting the probabilities with small magnitudes, i.e., we only rule out a candidate if
his/her prediction probability is less than the specified threshold α.
For the 2016 election, the reported probabilities for Clinton were 0.72, 0.91, 0.98, and
0.99 for FiveThirtyEight, The New York Times, The Huffington Post, and the Princeton
Election Consortium, respectively. Taking these probabilities as the values for πx (c), if
α = 0.05, then the resulting prediction sets are {c, t} for FiveThirtyEight and The New
York Times, and {c} for the two others.
Our proposed approach that predicts with a set of candidates is more conservative, in
general, than an approach that is obligated to choose a single candidate. But if the polls
are not informative enough to distinguish two or more candidates, then the analysis is
prudent to avoid such a potentially misleading distinction. Indeed, an important point
that came up in the post-election discussions is that the predictions that heavily favored
Clinton may have unintentionally caused some Clinton supporters not to vote. Our
proposal to make specific predictions only when the data makes it a clear decision helps
to avoid the undesirable situation where the prediction itself affects the outcome. We
9
note, however, that even under the suggested approach, the analysis by the Princeton
Election Consortium and others that assigned probabilities of 98% or higher to Clinton
would still have given the wrong predictions, recalling the importance of a sound model
in arriving at the probabilities πx .
As we demonstrate below, under mild conditions, the α-level prediction set Πx (α)
controls the probability of making a wrong prediction at α. More formally, we say that
prediction based on a α-level prediction set Πx (α) defined in (2) is valid at level α, relative
to the specified model P for (X, Y ), if
P{ΠX (α) 63 Y } ≤ α.
(3)
In words, the prediction is valid at level α if the P-probability of making a wrong prediction with the set ΠX (α) is no more than the specified α. Compare this to the informal
notion of validity introduced in Section 3.2 above. Proposition 1 gives a sufficient condition for (3) that is easy to arrange.
Proposition 1. A prediction set ΠX (α) as defined in (2) is valid in the sense of (3) if
the corresponding x 7→ πx satisfies
P{πX (Y ) ≤ α} ≤ α.
(4)
Proof. By definition of ΠX (α),
P{ΠX (α) 63 Y } = P{πX (Y ) ≤ α},
so the validity result (3) follows immediately from (4).
We note that the setup is very general—any choice of x 7→ πx that satisfies (4) would
lead to valid prediction—but (4) automatically holds for an interval of α values when πx
is the conditional probability (1) based on the model P.
Proposition 2. Let πx be the P-conditional probability in (1). Then there exists A > 0
such that (4) holds for all α ∈ (0, A). In particular, A can be taken as
A = min πx(2) ,
x
(5)
(2)
where πx is the second smallest among the strictly positive πx (y) values, for the given x,
as y ranges over the (finite) set Y.
Proof. Using iterated expectation and the definition of πx , we get
P{πX (Y ) ≤ α} = E{1πX (Y )≤α }
nX
o
=E
1πX (y)≤α P(Y = y | X)
y∈Y
=E
nX
o
1πX (y)≤α πX (y)
y∈Y
If α < A, then the events {πX (y) ≤ α}, y ∈ Y, are mutually exclusive, so at most one
term in the above summation is positive. Since each term is no more than α, this implies
that the summation itself is no more than α, hence (4).
10
Note that if N = 2 in Proposition 2, say, Y = {c, t}, then A in (5) reduces to
A = min max{πx (c), πx (t)}.
x
Unless there are some counter-intuitive constraints on the Clinton/Trump probabilities,
e.g., πx (c) ≤ 31 for all values of x, we expect A = 12 . So, the interval (0, A) would certainly
be wide enough to accommodate traditionally “small” values, such as α = 0.05.
The above result is conservative in that the conclusion (4) may hold for a wider range
of α than that given by A in (5). In fact, in some cases, (4) holds for all α ∈ (0, 1),
which amounts to πX (Y ) being stochastically no smaller than a random variable having
a uniform distribution on (0, 1); see Figure 2. There are cases, however, where (4) does
not hold for all α ∈ (0, 1). For an extreme example, in the case Y = {c, t}, suppose that
the polling data is completely uninformative, so that πx (c) = πx (t) = 12 for all x. Then
we have P{πX (Y ) ≤ α} = 0 for all α < 12 and P{πX (Y ) ≤ α} = 1 for all α ≥ 21 , so that
(4) holds for α ∈ (0, 12 ) but in a trivial way.
3.4
Illustrative example
Here we consider an example that is simple enough that the necessary calculations can be
done without much trouble but also relevant enough that it illustrates our main points.We
revisit this example in Section 4.3 below.
Assume that there are only two viable candidates, Clinton and Trump, so that Y =
{c, t} and N = 2. Start with a fixed sample size n, poll n voters, and record the number
X that indicate support for Trump. (For now, we assume that there is no non-response,
so that n−X of the polled voters indicate support for Clinton; we relax this assumption in
Section 4.3.) A simple choice is a binomial model, i.e., X ∼ Bin(n, θ), with a default flat
prior for θ, resulting in a model for X that is uniform on {0, 1, . . . , n}. To complete our
specification of P, the joint distribution for the polling data X and the election outcome
Y , it remains to describe the conditional distribution of Y , given X = x, in (1). Here we
propose a logistic-type model
πx (t) =
exp{λ(θ̂ − 21 )}
1 + exp{λ(θ̂ − 21 )}
and πx (c) = 1 − πx (t),
(6)
where θ̂ = x/n and λ > 0 is a specified constant. Assuming that the poll is reasonably
informative, if θ̂ is much larger than 0.5—even, say, 0.6—then the probability assigned
to a Trump victory ought to be close to 1, so a relatively large value of λ makes sense;
for this illustration, we consider λ = 10. A plot of this πx (t) curve, as a function of
θ̂, is shown in Figure 1. The distribution function G of πX (Y ), derived from the joint
distribution of (X, Y ) under this model, is displayed in Figure 2; note the stochastic
ordering from (4). Validity of the corresponding prediction set ΠX (α), for any α ∈ (0, 1),
then follows immediately from Propositions 1 and 2.
The prediction set Πx (α), for small α, would contain both c and t unless
θ̂ ≤
1 1
α
+ log
2 λ
1−α
or θ̂ ≥
11
1 1
1−α
+ log
.
2 λ
α
1.0
0.8
0.6
0.4
0.0
0.2
πx(T)
0.0
0.2
0.4
0.6
0.8
1.0
θ^
Figure 1: Plot of the probability πx (t) in Section 3.4, as a function of θ̂ = x/n.
For a numerical example, suppose that X = 470 and we take the model in (6) with
λ = 10. According to the classical distribution theory, there is roughly a 3% margin of
error, so this seems to give an impression that Clinton has a significant lead. But under
the model in (6), the prediction probabilities are
πx (t) =
exp{10(0.47 − 0.50)}
= 0.426 and πx (c) = 1 − 0.426 = 0.574,
1 + exp{10(0.47 − 50)}
giving both candidates an appreciable probability of winning. For sure, no validity guarantees can be given for a prediction that Clinton wins based only on this available information. Therefore, we argue that, based on this data and this model, the analyst should
acknowledge that the race is too close to make a prediction, and report the set {c, t}
based on this poll data.
4
4.1
Predictions accounting for uncertainty
Probabilities to plausibilities
A serious concern about the use of (conditional) probabilities πx (y) in (1) for prediction
is that these only describe the variability of the outcome Y , given X = x, with respect
to a given model P. They do not reflect the uncertainty in the choice of P itself. Since
the outcome Y is usually binary in the election context, variability is difficult to interpret
and arguably less important than model uncertainty.
For example, recall Silver’s October 24th acknowledgment [19] that his estimates for
Clinton could change from 85% up to 99.8% depending on the choice of model. But
then he ultimately reports an estimate that does not actually reflect this uncertainty. An
12
1.0
0.8
0.6
0.4
0.0
0.2
G(π)
0.0
0.2
0.4
0.6
0.8
1.0
π
Figure 2: Distribution function G(π) = P{πX (Y ) ≤ π}, for π ∈ (0, 1), derived from the
joint distribution of (X, Y ) described in Section 3.4.
important lesson from this past election is that model uncertainty can be substantial, so
it should somehow be incorporated into the predictions.
A realistic situation is one where subject-area knowledge is used to narrow down to a
collection P of reasonable models P, but unable to identify a single best model. That is,
we “don’t know” which model within this collection is most appropriate. For example,
in this past election, there was speculation of a “hidden” Trump vote, and among the
collection of models that differ in the nature and extent to which the “hidden” vote
influences the outcome, we don’t know which is most appropriate; see Section 4.3. Silver
and others chose models that ignored the possible “hidden” Trump vote, and it turns out
that this may have been a poor choice. As discussed above, model choice is a subjective
decision, so the only thing wrong with the choices made by Silver and others is in not
accounting for their uncertainty in this choice.
To account for uncertainty, we propose an adjustment to the output reported for
prediction. Specifically, given a collection P of candidate models, we recommend reporting
features of the set function
π x (B) = sup P(Y ∈ B | X = x),
B ⊆ Y,
(7)
P∈P
in particular,
π x (y) = sup P(Y = y | X = x),
y ∈ Y.
P∈P
In Section 4.2, we discuss a number of important consequences in using the alternative
prediction output in (7). Here we simply point out that, since the supremum in π x (y) is
taken separately for each y ∈ Y, the resulting π x need not be a probability distribution.
For example, in the case where Y = {c, t}, it may be that π x (c) + π x (t) > 1. Therefore,
13
since π x (y) may not be probabilities, we instead call them plausibilities, following Shafer’s
terminology [15]. In this terminology, an outcome Y = y is plausible given X = x as long
as Y = y has a large enough probability under some candidate model P ∈ P. Conversely,
an outcome is implausible if it has a small probability under all candidate models P ∈ P.
4.2
Interpreting plausibilities
Our motivation for shifting from probabilities in Section 3.1 to plausibilities in Section 4.1
is to account for model uncertainty in the prediction output. Our proposal in (7) satisfies
this objective for two main reasons:
• Evaluating π x (y) requires the data analyst to be explicit about the class of models
being entertained. Compare this to standard practice where several models are
considered but the results of only one cherry-picked analysis are reported [6].
• Maximizing over P is consistent with our general understanding (via Cournot’s
principle) that only small probability values are actionable. In particular, when the
model itself is uncertain—meaning that different candidate models in P assign vastly
different probabilities to a particular outcome—then we should be less inclined to
take action based on the assessment of the more favorable model. Taking the
supremum in (7) has the effect of pushing some borderline-actionable probabilities
into the unactionable category.
Plausibilities are surely less familiar and arguably more complicated than probabilities, so some additional care is needed to interpret them. First, any set-function p on
the power set of Y that satisfies p(∅) = 0, p(B) ≥ 0, and p(B1 ) + p(B2 ) ≥ p(B1 ∪ B2 )
is a plausibility function. It is easy to check that the function π x in (7) meets these
requirements. Towards an interpretation, it helps to regard π x it as an upper probability
[3]; that is, π x (B) is the maximal probability assigned to B as P varies over the collection
P. Conversely, there is a corresponding lower probability
π x (B) = 1 − π x (B c ),
B ⊆ Y,
or belief function [15], obtained by replacing supremum in (7) with infimum.
The lower probabilities π x and, therefore, also upper probabilities π x admit a clear
interpretation in terms of bets [17, 26]. If our lower probability on an event B ⊆ Y is given
by π x (B) and we are offered the opportunity to bet on B, then we would accept any price
up to π x (B) since such a bet is coherent with respect to every conceivable model P in P.
This interpretation of π x (B) in terms of betting is more conservative than the ordinary
Bayesian arguments for coherence in terms of two-sided bets. In the conventional setting,
we are offered a bet at price pB on B and 1 − pB on B c . Note that by the requirement
π x (B) + π x (B c ) ≥ 1, we have π x (B) + π x (B c ) ≤ 1, so that if pB ∈ (π x (B), π x (B)), then
we would not accept a bet on either B or B c .
Another interesting feature of the upper and lower probabilities is that the difference
π x (B) − π x (B)
measures model uncertainty about B, what Dempster [4] refers to as a “don’t know”
probability, which measures the amount of uncertainty about B in the model P. The
14
greater this difference, the more uncertainty there is about how to assign probabilities
to the event B, and hence the more we “don’t know” about the proper assessment of B.
In the election context, where we are interested primarily in the probabilities assigned to
particular candidates, we have
π x (y) = sup P(Y = y | X = x) and π x (y) = inf P(Y = y | X = x),
P∈P
P∈P
and the uncertainty about the model is reflected in the fact that the difference π x (y) −
π x (y) is positive at least for some y ∈ Y. Under the view that only small probabilities
are interpretable, the most meaningful/useful output is π x (y), y ∈ Y, since having π x (y)
small means that the event Y = y has small probability under all candidate models P ∈ P,
and is thus actionable even in the presence of model uncertainty.
Finally, our choice to work with the plausibility or upper probability π x is consistent
with the proposed notion of validity and the corresponding prediction sets. Indeed,
the following proposition says that, for suitable α, the α-prediction set Πx (α), using
the plausibility π x (y) in place of the probability πx (y), satisfies the prediction validity
condition (3) with respect to any P in P.
Proposition 3. Suppose that there exists α > 0 such that
P{πX (Y ) ≤ α} ≤ α
for all P ∈ P, where πx is the P-conditional probability. Then, for that α value, the
plausibility-based α-prediction set
Πx (α) = {y ∈ Y : π x (y) > α}
satisfies
sup P{ΠX (α) 63 Y } ≤ α.
P∈P
4.3
Illustrative example, cont.
To follow up on the example from Section 3.4, suppose now that, among the n = 1000
polled individuals, 475 pledge support to Clinton, 425 pledged support to Trump, and 100
did not respond. A naive approach to handling non-response is to assume that those who
did not respond are no different, on average, than those who did respond; in statistics
jargon, this corresponds to a missing-at-random assumption. But based on the available
information, there is no justification for such a model assumption. In fact, all available
information indicates that these 100 individuals, in choosing not to respond to the poll,
acted drastically differently from the other 900 who did respond, suggesting that their
preferences may also exhibit different behavior.
Here, the added uncertainty is in which category—c or t—to put the 100 nonresponses. The missing-at-random model would split the 100 proportionally into the
two groups, which boils down to ignoring them altogether. Another model, extreme but
consistent with the hypothesis of a “hidden” Trump vote would put all 100 non-responses
with Trump; a similarly extreme model puts the 100 non-responses with Clinton, consistent with a “hidden Clinton” vote that was discussed by some [1]. There may be a
15
whole spectrum of models in P, differing in how they handle the non-responses, but we
have identified the two extremes, so we can evaluate the plausibilities by plugging the θ̂
corresponding to these extremes into the formula in Section 3.4. That is, set
θ̂lo =
425 + 0
425 + 100
= 0.425 and θ̂hi =
= 0.525,
1000
1000
and evaluate the plausibilities (for λ = 10)
π x (t) =
exp{λ(θ̂hi − 12 )}
1 + exp{λ(θ̂hi − 12 )}
π x (c) = 1 −
= 0.562,
exp{λ(θ̂lo − 12 )}
1 + exp{λ(θ̂lo − 12 )}
= 0.679.
This is equivalent to a pair of lower probabilities, 0.321 for t and 0.438 for c, along with
an extra 0.241 of “don’t know.” These results indicate that the information in the polling
data is too weak to make a clear prediction about the winner. Compare this to a naive
analysis that ignores the non-response in assuming θ̂ = 425/900 = 0.472 and reports the
pair of probabilities
πx (t) =
exp{λ(0.472 − 0.50)}
= 0.430 and πx (c) = 1 − πx (t) = 0.570.
1 + exp{λ(0.472 − 0.50)}
This latter analysis gives an indication that Clinton has a semi-healthy lead, while ignoring that a critical and arguably unjustifiable missing-at-random assumption was made.
5
Concluding remarks
The field of data science and statistics has grown rapidly over the past decade in no small
part because of the perceived success of data-driven methods in predicting the 2008 and
2012 presidential elections. Writing for a general audience, Kass et al (2016) pay explicit
acknowledgment to “FiveThirtyEight for bringing statistics to the world (or at least to
the media).” But the same publicity which gained statistics and data science great fanfare
after the 2008 and 2012 elections led to the public outcry after its perceived failure in the
2016 election. Such is the risk that comes with being relevant, and statisticians ought to
avoid the mistakes of the past when responding to these criticisms.
Statisticians are inclined to defend against these criticisms. For example, the brazenness in the declaration on the SimplyStatistics blog that Statistically speaking, Nate Silver,
once again, got it right, leaves many ordinary citizens scratching their heads: Silver was
praised for correctly predicting the 2008 and 2012 elections; and now he’s being praised
for incorrectly predicting the 2016 election. If Silver was correct in 2008 and 2012, when
his predictions were right, and also correct in 2016, when his predictions were wrong,
then is it possible for the analysis to ever be incorrect? Though professional statisticians
may dismiss such reasoning as naive, defending against these common sense refutations
with jargon heavy arguments only arouses more suspicion.
To be clear, we are not claiming that those analyses were wrong, only that they were
ineffective. And by all accounts, the analyses failed. They failed because the 2016 election
16
was one of the most heavily covered news events in recent history, offering a major platform to publicize the power of data-driven techniques in addressing real world problems.
But that’s OK. Statisticians should not feel obligated to defend these failed predictions.
We should, instead—as urged by Jessica Utts [24], 2016 President of the American Statistical Association—take the opportunity to clarify the purpose of statistical analyses
in the first place. Statistics is not a tool for predicting the future as much as it is of
making sense of what insights can be reasonably gleaned from data. Statistical analysis
is conservative by its nature, and the limelight enjoyed by Nate Silver and other data
scientists shouldn’t distract from that.
As Yogi Berra famously said, “It’s tough to make predictions, especially about the
future,” so even a perfectly sound analysis is not guaranteed to be right. One of the
things we can learn from the 2016 election is just how embarrassing incorrect predictions
can be. Since there is sure to be uncertainty in the 2020 election, we must be more careful
about how we justify and communicate our conclusions to consumers. In this paper, we
have put forward some alternative ways to convert probabilities to valid predictions in a
way that accounts for model uncertainties. These are, by design, more conservative than
traditional prediction strategies, especially when there is uncertainty about the model
itself, but we show that it is straightforward to arrange some bounds on the probability
of making an erroneous prediction.
References
[1] P. Bump. Is there a ‘hidden’ Hillary Clinton vote among women? The Washington Post, November 6 2016.
https://www.washingtonpost.com/news/thefix/wp/2016/11/06/is-there-a-hidden-hillary-clinton-vote-amongwomen/?utm_term=.5249e8e57113.
[2] H. Crane and R. Martin. Is statistics meeting the needs of science? Unpublished
manuscript, 2017.
[3] A. P. Dempster. Upper and lower probabilities induced by a multivalued mapping.
Ann. Math. Statist., 38:325–339, 1967.
[4] A. P. Dempster. The Dempster–Shafer calculus for statisticians. Internat. J. Approx.
Reason., 48(2):365–377, 2008.
[5] A. Gelman. Survey weighting and that 2% swing. http://andrewgelman.com/
2016/12/01/survey-weighting-2-swing/, December 1 2016.
[6] A. Gelman and E. Lokin. Data-dependent analysis—a “garden of forking paths”—
explains why statistically significance comparisons don’t hold up. American Scientist,
102(6):460, 2014.
[7] R. Irizarry.
Not all forecasters got it wrong: Nate Silver does it again
(again).
http://simplystatistics.org/2016/11/09/not-all-forecastersgot-it-wrong, November 9 2016.
17
[8] N. Jackson and A. Hooper. Forecast: President. The Huffington Post, October
10 (updated November 8) 2016. http://electons.huffingtonpost.com/2016/
forecast/president.
[9] J. Katz. Who will be president? https://www.nytimes.com/interactive/2016/
upshot/presidential-polls-forecast.html, November 8 2016.
[10] S. Lohr and N. Singer. How data failed us in calling an election. The New York Times,
November 10 2016.
http://www.nytimes.com/2016/11/10/technology/thedata-said-clinton-would-win-why-you-shouldnt-have-believed-it.html.
[11] R. Martin and R. Lingham. Prior-free probabilistic prediction of future observations.
Technometrics, 58(2):225–235, 2016.
[12] R. Martin and C. Liu. Validity and the foundations of statistical inference. Unpublished manuscript, arXiv:1607.05051, 2016.
[13] T. Noah (Host). The Daily Show with Trevor Noah. http://www.cc.com/fullepisodes/vnv022/the-daily-show-with-trevor-noah-november-14--2016--nate-silver---adrian-grenier-season-22-ep-22024, November 14 2016.
[14] D. B. Rubin. Bayesianly justifiable and relevant frequency calculations for the applied statistician. Ann. Statist., 12(4):1151–1172, 1984.
[15] G. Shafer. A Mathematical Theory of Evidence. Princeton University Press, Princeton, N.J., 1976.
[16] G. Shafer. From Cournot’s principle to market efficiency. In J.-P. Touffut, editor,
Augustin Cournot: Modelling Economics, pages 55–95. Edward Elgar, 2007.
[17] G. Shafer. A betting interpretation for probabilities and Dempster–Shafer degrees
of belief. Internat. J. Approx. Reason., 52(2):127–136, 2011.
[18] G. Shafer, V. Vovk, and A. Takemura. Defensive forecasting. In R. Cowell and
Z. Ghahramani, editors, AISTATS 2005, pages 365–372, 2005.
[19] N. Silver. Election update: Why our model is more bullish than others on
Trump. http://fivethirtyeight.com/features/election-update-why-ourmodel-is-more-bullish-than-others-on-trump/, October 24 2016.
[20] N. Silver. Final election update: There’s a wide range of outcomes, and most
of them come up Clinton.
http://fivethirtyeight.com/features/finalelection-update-theres-a-wide-range-of-outcomes-and-most-of-themcome-up-clinton/, November 8 2016.
[21] N. Silver.
“Shy” voters probably arent why the polls missed Trump.
http://fivethirtyeight.com/features/shy-voters-probably-arent-whythe-polls-missed-trump/, November 16 2016.
18
[22] N. Silver. Why FiveThirtyEight gave Trump a better chance than almost anyone else. http://fivethirtyeight.com/features/why-fivethirtyeight-gavetrump-a-better-chance-than-almost-anyone-else/, November 11 2016.
[23] T. Speed. Terence’s Stuff: Principles. IMS Bulletin, page 17, September 2016.
[24] J. Utts. Presidential transitions. AmStat News, pages 3–5, December 2016.
[25] Wikipedia. Bradley effect — Wikipedia, The Free Encyclopedia. https://en.
wikipedia.org/w/index.php?title=Bradley_effect&oldid=763056842.
[Accessed 4-February-2017].
[26] P. M. Williams. On a new theory of epistemic probability. Brit. J. Philos. Sci.,
29(4):375–387, 1978.
19
| 10math.ST
|
Optimization of Trading Physics Models of Markets
Lester Ingber
Lester Ingber Research
POB 06440 Sears Tower, Chicago, IL 60606
arXiv:physics/0007075 22 Jul 2000
and
DRW Investments LLC
311 S Wacker Dr, Ste 900, Chicago, IL 60606
[email protected], [email protected]
and
Radu Paul Mondescu
DRW Investments LLC
311 S Wacker Dr, Ste 900, Chicago, IL 60606
[email protected]
ABSTRACT
We describe an end-to-end real-time S&P futures trading system. Inner-shell stochastic nonlinear
dynamic models are developed, and Canonical Momenta Indicators (CMI) are derived from a fitted
Lagrangian used by outer-shell trading models dependent on these indicators. Recursive and adaptive
optimization using Adaptive Simulated Annealing (ASA) is used for fitting parameters shared across
these shells of dynamic and trading models.
Keywords: Simulated Annealing; Statistical Mechanics; Trading Financial Markets
© 2000 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this
material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be
obtained from the IEEE.
This work has been submitted to the IEEE for possible publication. Copyright may be transferred without
notice, after which this version may no longer be accessible.
This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and
all rights therein are retained by authors or by other copyright holders. All persons copying this information
are expected to adhere to the terms and constraints invoked by each author’s copyright. In most cases, these
works may not be reposted without the explicit permission of the copyright holder.
Optimization of Trading
-2-
Ingber & Mondescu
1. INTRODUCTION
1.1. Approaches
Real-world problems are almost intractable analytically, yet methods must be devised to deal with
this complexity to extract practical information in finite time. This is indeed true in the field of financial
engineering, where time series of various financial instruments reflect nonequilibrium, highly non-linear,
possibly even chaotic [1] underlying processes. A further difficulty is the huge amount of data necessary
to be processed. Under these circumstances, to develop models and schemes for automated, profitable
trading is a non-trivial task.
In the context of this paper, it is important to stress that dealing with such complex systems
invariably requires modeling of dynamics, modeling of actions on these dynamics, and algorithms to fit
parameters in these models to real data. We have elected to use methods of mathematical physics for our
models of the dynamics, artificial intelligence (AI) heuristics for our models of trading rules acting on
indicators derived from our dynamics, and methods of sampling global optimization for fitting our
parameters. Too often there is confusion about how these three elements are being used for a complete
system. For example, in the literature there often is discussion of neural net trading systems or genetic
algorithm trading systems. However, neural net models (used for either or both models discussed here)
also require some method of fitting their parameters, and genetic algorithms must have some kind of cost
function or process specified to sample a parameter space, etc.
Some powerful methods have emerged during years, appearing from at least two directions: One
direction is based on inferring rules from past and current behavior of market data leading to learningbased, inductive techniques, such as neural networks, or fuzzy logic. Another direction starts from the
bottom-up, trying to build physical and mathematical models based on different economic prototypes. In
many ways, these two directions are complementary and a proper understanding of their main strengths
and weaknesses should lead to synergetic effects beneficial to their common goals.
Among approaches in the first direction, neural networks already have won a prominent role in the
financial community, due to their ability to handle large quantities of data, and to uncover and model
nonlinear functional relationships between various combinations of fundamental indicators and price
data [2,3].
Optimization of Trading
-3-
Ingber & Mondescu
In the second direction we can include models based on non-equilibrium statistical mechanics [4]
fractal geometry [5], turbulence [6], spin glasses and random matrix theory [7], renormalization group [8],
and gauge theory [9]. Although the very complex nonlinear multivariate character of financial markets is
recognized [10], these approaches seem to have had a lesser impact on current quantitative finance
practice, although it is becoming increasing clear that this direction can lead to practical trading strategies
and models.
To bridge the gap between theory and practice, as well as to afford a comparison with neural
networks techniques, here we focus on presenting an effective trading system of S&P futures, anchored in
the physical principles of nonequilibrium statistical mechanics applied to financial markets [4,11].
Starting with nonlinear, multivariate, nonlinear stochastic differential equation descriptions of the
price evolution of cash and futures indices, we build an algebraic cost function in terms of a Lagrangian.
Then, a maximum likelihood fit to the data is performed using a global optimization algorithm, Adaptive
Simulated Annealing (ASA) [12]. As firmly rooted in field theoretical concepts, we derive market
canonical momenta indicators, and we use these as technical signals in a recursive ASA optimization that
tunes the outer-shell of trading rules. We do not employ metaphors for these physical indicators, but
rather derive them directly from models fit to data.
The outline of the paper is as follows: Just below we briefly discuss the optimization method and
momenta indicators. In the next three sections we establish the theoretical framework supporting our
model, the statistical mechanics approach, and the optimization method. In Section 5 we detail the
trading system, and in Section 6 we describe our results. Our conclusions are presented in Section 7.
1.2. Optimization
Large-scale, non-linear fits of stochastic nonlinear forms to financial data require methods robust
enough across data sets. (Just one day, tick data for regular trading hours could reach 10,000−30,000 data
points.) Simple regression techniques exhibit deficiencies with respect to obtaining reasonable fits. They
too often get trapped in local minima typically found in nonlinear stochastic models of such data. ASA is
a global optimization algorithm that has the advantage — with respect to other global optimization
methods as genetic algorithms, combinatorial optimization, etc. — not only to be efficient in its
importance-sampling search strategy, but to have the statistical guarantee of finding the best
Optimization of Trading
-4-
Ingber & Mondescu
optima [13,14]. This gives some confidence that a global minimum can be found, of course provided care
is taken as necessary to tune the algorithm [15].
It should be noted that such powerful sampling algorithms also are often required by other models
of complex systems than those we use here [16]. For example, neural network models have taken
advantage of ASA [17-19], as have other financial and economic studies [20,21].
1.3. Indicators
In general, neural network approaches attempt classification and identification of patterns, or try
forecasting patterns and future evolution of financial time series. Statistical mechanical methods attempt
to find dynamic indicators derived from physical models based on general principles of non-equilibrium
stochastic processes that reflect certain market factors. These indicators are used subsequently to generate
trading signals or to try forecasting upcoming data.
In this paper, the main indicators are called Canonical Momenta Indicators (CMI), as they faithfully
mathematically carry the significance of market momentum, where the “mass” is inversely proportional to
the price volatility (the “masses” are just the elements of the metric tensor in this Lagrangian formalism)
and the “velocity” is the rate of price changes.
2. MODELS
2.1. Langevin Equations for Random Walks
The use of Brownian motion as a model for financial systems is generally attributed to
Bachelier [22], though he incorrectly intuited that the noise scaled linearly instead of as the square root
relative to the random log-price variable. Einstein is generally credited with using the correct
mathematical description in a larger physical context of statistical systems. However, several studies
imply that changing prices of many markets do not follow a random walk, that they may have long-term
dependences in price correlations, and that they may not be efficient in quickly arbitraging new
information [23-25]. A random walk for returns, rate of change of prices over prices, is described by a
Langevin equation with simple additive noise η , typically representing the continual random influx of
information into the market.
Optimization of Trading
-5-
Ingber & Mondescu
Ṁ = − f + gη ,
Ṁ = dM/dt ,
< η (t) >η = 0 , < η (t), η (t′) >η = δ (t − t′) ,
(1)
where f and g are constants, and M is the logarithm of (scaled) price, M(t) = log(( P(t)/P(t − dt))). Price,
although the most dramatic observable, may not be the only appropriate dependent variable or order
parameter for the system of markets [26]. This possibility has also been called the “semistrong form of
the efficient market hypothesis” [23].
The generalization of this approach to include multivariate nonlinear nonequilibrium markets led to
a model of statistical mechanics of financial markets (SMFM) [11].
2.2. Adaptive Optimization of F x Models
Our S&P model for the futures F is
dF = µ dt + σ F x dz ,
< dz > = 0 ,
< dz(t) dz(t′) > = dt δ (t − t′)
We have used this model in several ways to fit the distribution’s volatility defined in terms of a scale
and an exponent of the independent variable [4].
A major component of our trading system is the use of adaptive optimization, essentially constantly
retuning the parameters of our dynamic model each time new data is encountered in our training, testing
and real-time applications. The parameters { µ , σ } are constantly tuned using a quasi-local simplex
code [27,28] included with the ASA (Adaptive Simulated Annealing) code [12].
We have tested several quasi-local codes for this kind of trading problem, versus using robust ASA
adaptive optimizations, and the faster quasi-local codes seem to work quite well for adaptive updates after
a zeroth order parameters set is found by ASA [29,30].
Optimization of Trading
-6-
Ingber & Mondescu
3. STATISTICAL MECHANICS OF FINANCIAL MARKETS (SMFM)
3.1. Statistical Mechanics of Large Systems
Aggregation problems in nonlinear nonequilibrium systems typically are “solved” (accommodated)
by having new entities/languages developed at these disparate scales in order to efficiently pass
information back and forth between scales. This is quite different from the nature of quasi-equilibrium
quasi-linear systems, where thermodynamic or cybernetic approaches are possible. These thermodynamic
approaches typically fail for nonequilibrium nonlinear systems.
Many systems are aptly modeled in terms of multivariate differential rate-equations, known as
Langevin equations [31],
G
Ṁ = f G + ĝGj η j , (G = 1, . . . , Λ) , ( j = 1, . . . , N ) ,
G
Ṁ = dM G /dt ,
< η j (t) >η = 0 , < η j (t), η j′ (t′) >η = δ jj′δ (t − t′) ,
(2)
where f G and ĝGj are generally nonlinear functions of mesoscopic order parameters M G , j is a
microscopic index indicating the source of fluctuations, and N ≥ Λ. The Einstein convention of summing
over repeated indices is used. Vertical bars on an index, e.g., |j|, imply no sum is to be taken on repeated
indices.
Via a somewhat lengthy, albeit instructive calculation, outlined in several other papers [11,32,33],
involving an intermediate derivation of a corresponding Fokker-Planck or Schr̈odinger-type equation for
the conditional probability distribution P[M(t)|M(t 0 )], the Langevin rate Eq. (2) is developed into the
more useful probability distribution for M G at long-time macroscopic time event t = (u + 1)θ + t 0 , in
terms of a Stratonovich path-integral over mesoscopic Gaussian conditional probabilities [34-38]. Here,
macroscopic variables are defined as the long-time limit of the evolving mesoscopic system.
The corresponding Schr̈odinger-type equation is [36,37]
∂P/∂t =
1
2
(gGG′ P),GG′ − (gG P),G + V ,
gGG′ = k T δ jk ĝGj ĝG′
k ,
Optimization of Trading
1
gG = f G +
2
δ
-7-
jk G′ G
ĝ j ĝ k,G′
Ingber & Mondescu
,
[. . .],G = ∂[. . .]/∂M G .
(3)
This is properly referred to as a Fokker-Planck equation when V ≡ 0. Note that although the partial
differential Eq. (3) contains information regarding M G as in the stochastic differential Eq. (2), all
references to j have been properly averaged over. I.e., ĝGj in Eq. (2) is an entity with parameters in both
microscopic and mesoscopic spaces, but M is a purely mesoscopic variable, and this is more clearly
reflected in Eq. (3).
The path integral representation is given in terms of the “Feynman” Lagrangian L.
P[M t |M t 0 ]dM(t) =
∫ . . . ∫ DM exp(−S)δ [M(t0) = M0]δ [M(t) = M t ] ,
t
S = k T−1 min
∫t dt′L ,
0
DM = lim
u+1
(2π θ )−1/2 dM vG ,
Π g1/2 Π
G
u→∞ v=1
G
L( Ṁ , M G , t) =
hG = gG −
1
2
1
2
G
( Ṁ − hG )gGG′ ( Ṁ
G′
− hG′ ) +
1
2
hG ;G + R/6 − V ,
g−1/2 (g1/2 gGG′ ),G′ ,
gGG′ = (gGG′ )−1 ,
g = det(gGG′ ) ,
F
G
−1/2 1/2 G
hG ;G = hG
(g h ),G ,
,G + ΓGF h = g
Γ FJK ≡ g LF [JK , L] = g LF (g JL,K + g KL,J − g JK ,L ) ,
R = g JL R JL = g JL g JK R FJKL ,
R FJKL =
1
2
M N
M N
(g FK ,JL − g JK ,FL − g FL,JK + g JL,FK ) + g MN (Γ FK
Γ JL − Γ FL
Γ JK ) .
(4)
Optimization of Trading
-8-
Ingber & Mondescu
Mesoscopic variables have been defined as M G in the Langevin and Fokker-Planck representations, in
terms of their development from the microscopic system labeled by j. The Riemannian curvature term R
arises from nonlinear gGG′ , which is a bona fide metric of this space [36]. Even if a stationary solution,
G
G
i.e., Ṁ = 0, is ultimately sought, a necessarily prior stochastic treatment of Ṁ terms gives rise to these
Riemannian “corrections.” Even for a constant metric, the term hG ;G contributes to L for a nonlinear
mean hG . V may include terms such as
J T ′G M G , where the Lagrange multipliers J T ′G are constraints
Σ
T′
on M G , which are advantageously modeled as extrinsic sources in this representation; they too may be
time-dependent.
For our purposes, the above Feynman Lagrangian defines a kernel of the short-time conditional
probability distribution, in the curved space defined by the metric, in the limit of continuous time, whose
iteration yields the solution of the previous partial differential equation Eq. (3). This differs from the
Lagrangian which satisfies the requirement that the action is stationary to the first order in dt — the
WKBJ approximation, but which does not include the first-order correction to the WKBJ approximation
as does the Feynman Lagrangian. This latter Lagrangian differs from the Feynman Lagrangian,
essentially by replacing R/6 above by R/12 [39]. In this sense, the WKBJ Lagrangian is more useful for
some theoretical discussions [40]. However, the use of the Feynman Lagrangian coincides with the
numerical method we use for long-time development of our distributions using our PATHINT code for
other financial products, e.g., options [4]. This also is consistent with our use of relatively short-time
“forecast” of data points using the most probable path [41]
dM G /dt = gG − g1/2 (g−1/2 gGG′ ), G′ .
(5)
Using the variational principle, J TG may also be used to constrain M G to regions where they are
empirically bound. More complicated constraints may be affixed to L using methods of optimal control
theory [42]. With respect to a steady state P, when it exists, the information gain in state P is defined by
ϒ[P] =
∫ . . . ∫ DM′ P ln (P/P) ,
DM′ = DM/dM u+1 .
(6)
In the economics literature, there appears to be sentiment to define Eq. (2) by the Itô, rather than the
Stratonovich prescription. It is true that Itô integrals have Martingale properties not possessed by
Optimization of Trading
-9-
Ingber & Mondescu
Stratonovich integrals [43] which leads to risk-neural theorems for markets [44,45], but the nature of the
proper mathematics — actually a simple transformation between these two discretizations — should
eventually be determined by proper aggregation of relatively microscopic models of markets. It should be
noted that virtually all investigations of other physical systems, which are also continuous time models of
discrete processes, conclude that the Stratonovich interpretation coincides with reality, when
multiplicative noise with zero correlation time, modeled in terms of white noise η j , is properly considered
as the limit of real noise with finite correlation time [46]. The path integral succinctly demonstrates the
difference between the two: The Itô prescription corresponds to the prepoint discretization of L, wherein
θ Ṁ(t) → M v+1 − M v and M(t) → M v . The Stratonovich prescription corresponds to the midpoint
discretization of L, wherein θ Ṁ(t) → M v+1 − M v and M(t) →
1
2
(M v+1 + M v ). In terms of the functions
appearing in the Fokker-Planck Eq. (3), the Itô prescription of the prepoint discretized Lagrangian, L I , is
relatively simple, albeit deceptively so because of its nonstandard calculus.
1
G
L I ( Ṁ , M G , t) =
2
G
( Ṁ − gG )gGG′ ( Ṁ
G′
− gG′ ) − V .
(7)
In the absence of a nonphenomenological microscopic theory, the difference between a Itô prescription
and a Stratonovich prescription is simply a transformed drift [39].
There are several other advantages to Eq. (4) over Eq. (2). Extrema and most probable states of
M G , << M G >>, are simply derived by a variational principle, similar to conditions sought in previous
studies [47]. In the Stratonovich prescription, necessary, albeit not sufficient, conditions are given by
δ G L = L ,G − L ,Ġ:t = 0 ,
L ,Ġ:t = L ,ĠG′ Ṁ
G′
+ L ,Ġ Ġ′ M̈
G
G′
.
(8)
G
G
For stationary states, Ṁ = 0, and ∂L/∂M = 0 defines << M >>, where the bars identify stationary
variables; in this case, the macroscopic variables are equal to their mesoscopic counterparts. Note that L
is not the stationary solution of the system, e.g., to Eq. (3) with ∂P/∂t = 0. However, in some cases [48],
L is a definite aid to finding such stationary states. Many times only properties of stationary states are
G
examined, but here a temporal dependence is included. E.g., the Ṁ terms in L permit steady states and
their fluctuations to be investigated in a nonequilibrium context. Note that Eq. (8) must be derived from
the path integral, Eq. (4), which is at least one reason to justify its development.
Optimization of Trading
- 10 -
Ingber & Mondescu
3.2. Algebraic Complexity Yields Simple Intuitive Results
It must be emphasized that the output of this formalism is not confined to complex algebraic forms
or tables of numbers. Because L possesses a variational principle, sets of contour graphs, at different
long-time epochs of the path-integral of P over its variables at all intermediate times, give a visually
intuitive and accurate decision-aid to view the dynamic evolution of the scenario. For example, this
Lagrangian approach permits a quantitative assessment of concepts usually only loosely defined.
“Momentum” = ΠG =
“Mass” = gGG′ =
“Force” =
∂L
,
∂(∂M G /∂t)
∂2 L
,
∂(∂M G /∂t)∂(∂M G′ /∂t)
∂L
,
∂M G
“F = ma”: δ L = 0 =
∂L
∂L
∂
,
−
G
∂M
∂t ∂(∂M G /∂t)
(9)
where M G are the variables and L is the Lagrangian. These physical entities provide another form of
intuitive, but quantitatively precise, presentation of these analyses. For example, daily newspapers use
some of this terminology to discuss the movement of security prices. In this paper, the ΠG serve as
canonical momenta indicators (CMI) for these systems.
3.2.1. Derived Canonical Momenta Indicators (CMI)
The extreme sensitivity of the CMI gives rapid feedback on changes in trends as well as the
volatility of markets, and therefore are good indicators to use for trading rules [29]. A time-locked
moving average provides manageable indicators for trading signals. This current project uses such CMI
developed as a byproduct of the ASA fits described below.
3.3. Correlations
In this paper we report results of our one-variable trading model. However, it is straightforward to
include multi-variable trading models in our approach, and we have done this, for example, with coupled
cash and futures S&P markets.
Optimization of Trading
- 11 -
Ingber & Mondescu
Correlations between variables are modeled explicitly in the Lagrangian as a parameter usually
designated ρ . This section uses a simple two-factor model to develop the correspondence between the
correlation ρ in the Lagrangian and that among the commonly written Wiener distribution dz.
Consider coupled stochastic differential equations for futures F and cash C:
dF = f F (F, C)dt + ĝ F (F, C)σ F dz F ,
dC = f C (F, C)dt + ĝC (F, C)σ C dzC ,
< dz i >= 0 , i = {F, C} ,
< dz i (t)dz j (t′) >= dt δ (t − t′) , i = j ,
< dz i (t)dz j (t′) >= ρ dt δ (t − t′) , i ≠ j ,
0 , , t ≠ t′ ,
1 , t = t′ ,
δ (t − t′) =
(10)
where < . > denotes expectations with respect to the multivariate distribution.
These can be rewritten as Langevin equations (in the Itô prepoint discretization)
dF/dt = f F + ĝ F σ F (γ +η 1 + sgn ρ γ −η 2 ) ,
dC/dt = gC + ĝC σ C (sgn ρ γ −η 1 + γ +η 2 ) ,
γ± =
1
2
√
[1 ± (1 − ρ 2 )1/2 ]1/2 ,
ni = (dt)1/2 pi ,
where p1 and p2 are independent [0,1] Gaussian distributions.
The equivalent short-time probability distribution, P, for the above set of equations is
P = g1/2 (2π dt)−1/2 exp(−Ldt) ,
L=
1
2
M † gM ,
(11)
Optimization of Trading
M=
- 12 -
Ingber & Mondescu
F
dF/dt − f
,
dC/dt − f C
g = det(g) .
(12)
g, the metric in {F, C}-space, is the inverse of the covariance matrix,
( ĝ F σ )2
g−1 = F FC
ρ ĝ ĝ σ F σ C
ρ ĝ F ĝC σ F σ C
.
( ĝC σ C )2
(13)
The CMI indicators are given by the formulas
ΠF =
(dF/dt − f F )
ρ (dC/dt − f C )
−
,
( ĝ F σ F )2 (1 − ρ 2 ) ĝ F ĝC σ F σ C (1 − ρ 2 )
ΠC =
ρ (dF/dt − f F )
(dC/dt − f C )
−
.
( ĝC σ C )2 (1 − ρ 2 ) ĝC ĝ F σ C σ F (1 − ρ 2 )
(14)
3.4. ASA Outline
The algorithm Adaptive Simulated Annealing (ASA) fits short-time probability distributions to
observed data, using a maximum likelihood technique on the Lagrangian. This algorithm has been
developed to fit observed data to a theoretical cost function over a D-dimensional parameter space [13],
adapting for varying sensitivities of parameters during the fit. The ASA code can be obtained at no
charge, via WWW from http://www.ingber.com/ or via FTP from ftp.ingber.com [12].
3.4.1. General Description
Simulated annealing (SA) was developed in 1983 to deal with highly nonlinear problems [49], as an
extension of a Monte-Carlo importance-sampling technique developed in 1953 for chemical physics
problems. It helps to visualize the problems presented by such complex systems as a geographical terrain.
For example, consider a mountain range, with two “parameters,” e.g., along the North−South and
East−West directions. We wish to find the lowest valley in this terrain. SA approaches this problem
similar to using a bouncing ball that can bounce over mountains from valley to valley. We start at a high
“temperature,” where the temperature is an SA parameter that mimics the effect of a fast moving particle
in a hot object like a hot molten metal, thereby permitting the ball to make very high bounces and being
Optimization of Trading
- 13 -
Ingber & Mondescu
able to bounce over any mountain to access any valley, given enough bounces. As the temperature is
made relatively colder, the ball cannot bounce so high, and it also can settle to become trapped in
relatively smaller ranges of valleys.
We imagine that our mountain range is aptly described by a “cost function.” We define probability
distributions of the two directional parameters, called generating distributions since they generate possible
valleys or states we are to explore. We define another distribution, called the acceptance distribution,
which depends on the difference of cost functions of the present generated valley we are to explore and
the last saved lowest valley. The acceptance distribution decides probabilistically whether to stay in a new
lower valley or to bounce out of it. All the generating and acceptance distributions depend on
temperatures.
In 1984 [50], it was established that SA possessed a proof that, by carefully controlling the rates of
cooling of temperatures, it could statistically find the best minimum, e.g., the lowest valley of our
example above. This was good news for people trying to solve hard problems which could not be solved
by other algorithms. The bad news was that the guarantee was only good if they were willing to run SA
forever. In 1987, a method of fast annealing (FA) was developed [51], which permitted lowering the
temperature exponentially faster, thereby statistically guaranteeing that the minimum could be found in
some finite time. However, that time still could be quite long. Shortly thereafter, Very Fast Simulated
Reannealing (VFSR) was developed in 1987 [13], now called Adaptive Simulated Annealing (ASA),
which is exponentially faster than FA.
ASA has been applied to many problems by many people in many disciplines [15,16,52]. The
feedback of many users regularly scrutinizing the source code ensures its soundness as it becomes more
flexible and powerful.
3.4.2. Mathematical Outline
ASA considers a parameter α ki in dimension i generated at annealing-time k with the range
α ki ∈[ Ai , Bi ] ,
calculated with the random variable y i ,
i
α k+1
= α ki + y i (Bi − Ai ) ,
(15)
Optimization of Trading
- 14 -
Ingber & Mondescu
y i ∈[−1, 1] .
(16)
The generating function gT (y) is defined,
D
gT (y) = Π
i=1
D
1
≡
gTi (y i ) ,
2(|y i | + T i ) ln(1 + 1/T i ) Π
i=1
(17)
where the subscript i on T i specifies the parameter index, and the k-dependence in T i (k) for the annealing
schedule has been dropped for brevity. Its cumulative probability distribution is
yD
y1
∫ −1∫
G T (y) =
...
−1
G Ti (y i ) =
D
dy′1 . . . dy′ D gT (y′) ≡ Π G Ti (y i ) ,
i=1
1 sgn (y i ) ln(1 + |y i |/T i )
+
.
2
2
ln(1 + 1/T i )
(18)
y i is generated from a ui from the uniform distribution
ui ∈U[0, 1] ,
1
y i = sgn (ui − )T i [(1 + 1/T i )|2u −1| − 1] .
i
2
(19)
It is straightforward to calculate that for an annealing schedule for T i
T i (k) = T 0i exp(−c i k 1/D ) ,
(20)
a global minima statistically can be obtained. I.e.,
∞
∞
0
0
D
1
1
] =∞.
gk ≈ Σ [ Π
Σ
i
i=1 2|y |c i k
k
k
(21)
Control can be taken over c i , such that
T fi = T 0i exp(−m i ) when k f = exp ni ,
c i = m i exp(−ni /D) ,
(22)
where m i and ni can be considered “free” parameters to help tune ASA for specific problems.
ASA has over 100 OPTIONS available for tuning. A few important ones were used in this project.
Optimization of Trading
- 15 -
Ingber & Mondescu
3.4.3. Reannealing
Whenever doing a multi-dimensional search in the course of a complex nonlinear physical problem,
inevitably one must deal with different changing sensitivities of the α i in the search. At any given
annealing-time, the range over which the relatively insensitive parameters are being searched can be
“stretched out” relative to the ranges of the more sensitive parameters. This can be accomplished by
periodically rescaling the annealing-time k, essentially reannealing, every hundred or so acceptanceevents (or at some user-defined modulus of the number of accepted or generated states), in terms of the
sensitivities si calculated at the most current minimum value of the cost function, C,
si = ∂C/∂α i .
(23)
In terms of the largest si = smax , a default rescaling is performed for each k i of each parameter dimension,
whereby a new index k′i is calculated from each k i ,
k i → k′i ,
T ′ik′ = T ik (smax /si ) ,
k′i = ( ln(T i0 /T ik′ )/c i ) D .
(24)
T i0 is set to unity to begin the search, which is ample to span each parameter dimension.
3.4.4. Quenching
Another adaptive feature of ASA is its ability to perform quenching in a methodical fashion. This
is applied by noting that the temperature schedule above can be redefined as
T i (k i ) = T 0i exp(−c i k i
Qi /D
),
c i = m i exp(−ni Qi /D) ,
(25)
in terms of the “quenching factor” Qi . The sampling proof fails if Qi > 1 as
D
Σk Π 1/k Q /D = Σk 1/k Q < ∞ .
i
i
(26)
This simple calculation shows how the “curse of dimensionality” arises, and also gives a possible
way of living with this disease. In ASA, the influence of large dimensions becomes clearly focussed on
Optimization of Trading
- 16 -
Ingber & Mondescu
the exponential of the power of k being 1/D, as the annealing required to properly sample the space
becomes prohibitively slow. So, if resources cannot be committed to properly sample the space, then for
some systems perhaps the next best procedure may be to turn on quenching, whereby Qi can become on
the order of the size of number of dimensions.
The scale of the power of 1/D temperature schedule used for the acceptance function can be altered
in a similar fashion. However, this does not affect the annealing proof of ASA, and so this may used
without damaging the sampling property.
3.4.5. Avoiding Repeating Cost Functions
Doing a recursive optimization is very CPU expensive, as essentially the cross-product of parameter
spaces among the various levels of optimization is required.
Therefore, we have used an ASA OPTION for some of the parameters in the outer-shell trading
model optimization of training sets, ASA_QUEUE, which sets up a first-in first-out (FIFO) queue, of
user-defined size Queue_Size to collect generated states. When a new state is generated, its parameters
are tested, within specified resolutions of a user-defined array Queue_Resolution[]. When parameters sets
are repeated within this queue, the saved value of the cost function is returned without having to repeat
the calculation.
3.4.6. Multiple Local Minima
Our criteria for the global minimum of our cost function is minus the largest profit over a selected
training data set (or in some cases, this value divided by the maximum drawdown). However, in many
cases this may not give us the best set of parameters to find profitable trading in test sets or in real-time
trading. Other considerations such as the total number of trades developed by the global minimum versus
other close local minima may be relevant. For example, if the global minimum has just a few trades,
while some nearby local minima (in terms of the value of the cost function) have many trades and was
profitable in spite of our slippage factors, then the scenario with more trades might be more statistically
dependable to deliver profits across testing and real-time data sets.
Therefore, for the outer-shell global optimization of training sets, we have used an ASA OPTION,
MULTI_MIN, which saves a user-defined number of closest local minima within a user-defined resolution
Optimization of Trading
- 17 -
Ingber & Mondescu
of the parameters. We then examine these results under several testing sets.
4. TRADING SYSTEM
4.1. Use of CMI
As the CMI formalism carries the relevant information regarding the prices dynamics, we have used
it as a signal generator for an automated trading system for S&P futures.
Based on a previous work [30] applied to daily closing data, the overall structure of the trading
system consists in 2 layers, as follows: We first construct the “short-time” Lagrangian function in the Itô
representation (with the notation introduced in Section 3.3)
2
1
dF i
F
L(i|i − 1) =
− f
2x dt
2σ 2 F i−1
(27)
with i the post-point index, corresponding to the one factor price model
dF = f F dt + σ F x dz(t) ,
(28)
where f F and σ > 0 are taken to be constants, F(t) is the S&P future price, and dz is the standard
Gaussian noise with zero mean and unit standard deviation. We perform a global, maximum likelihood fit
to the whole set of price data using ASA. This procedure produces the optimization parameters {x, f F }
that are used to generate the CMI. One computational approach was to fix the diffusion multiplier σ to 1
during training for convenience, but used as free parameters in the adaptive testing and real-time fits.
Another approach was to fix the scale of the volatility, using an improved model,
dF = f F dt + σ
x
F
dz(t) ,
< F >
(29)
where σ now is calculated as the standard deviation of the price increments ∆F/dt 1/2 , and < F > is just
the average of the prices.
As already remarked, to enhance the CMI sensitivity and response time to local variations (across a
certain window size) in the distribution of price increments, the momenta are generated applying an
adaptive procedure, i.e., after each new data reading another set of { f F , σ } parameters are calculated for
the last window of data, with the exponent x — a contextual indicator of the noise statistics — fixed to the
Optimization of Trading
- 18 -
Ingber & Mondescu
value obtained from the global fit.
The CMI computed in this manner are fed into the outer shell of the trading system, where an AItype optimization of the trading rules is executed, using ASA once again.
The trading rules are a collection of logical conditions among the CMI, prices and optimization
parameters that could be window sizes, time resolutions, or trigger thresholds. Based on the relationships
between CMI and optimization parameters, a trading decision is made. The cost function in the outer
shell is either the overall equity or the risk-adjusted profit (essentially the return). The inner and outer
shell optimizations are coupled through some of the optimization parameters (e.g., time resolution of the
data, window sizes), which justifies the recursive nature of the optimization.
Next, we describe in more details the concrete implementation of this system.
4.2. Data Processing
The CMI formalism is general and by construction permits us to treat multivariate coupled markets.
In certain conditions (e.g., shorter time scales of data), and also due to superior scalability across different
markets, it is desirable to have a trading system for a single instrument, in our case the S&P futures
contracts that are traded electronically on Chicago Mercantile Exchange (CME). The focus of our system
was intra-day trading, at time scales of data used in generating the buy/sell signals from 10 to 60 secs. In
particular, we here give some results obtained when using data having a time resolution ∆t of 55 secs (the
time between consecutive data elements is 55 secs). This particular choice of time resolution reflects the
set of optimization parameters that have been applied in actual trading.
It is important to remark that a data point in our model does not necessarily mean an actual tick
datum. For some trading time scales and for noise reduction purposes, data is pre-processed into
sampling bins of length ∆t using either a standard averaging procedure or spectral filtering (e.g., wavelets,
Fourier) of the tick data. Alternatively, the data can be defined in block bins that contain disjoint sets of
averaged tick data, or in overlapping bins of widths ∆t that update at every ∆t′ < ∆t, such that an effective
resolution ∆t′ shorter than the width of the sampling bin is obtained. We present here work in which we
have used disjoint block bins and a standard average of the tick data with time stamps falling within the
bin width.
Optimization of Trading
- 19 -
Ingber & Mondescu
In Figs. 1 and 2 we present examples of S&P futures data sampled with 55 secs resolution. We
remark that there are several time scales — from mins to one hour — at which an automated trading
system might extract profits. Fig. 2 illustrates the sustained short trading region of 1.5 hours and several
shorter long and short trading regions of about 10-20 mins. Fig. 1 illustrates that the profitable regions are
prominent even for data representing a relatively flat market period. I.e., June 20 shows an uptrend region
of about 1 hour 20 mins and several short and long trading domains between 10 mins and 20 mins. In
both situations, there are a larger number of opportunities at time resolutions smaller than 5 mins.
The time scale at which we sample the data for trading is itself a parameter that is extracted from
the optimization of the trading rules and of the Lagrangian cost function Eq. (27). This is one of the
coupling parameters between the inner- and the outer-shell optimizations.
4.3. Inner-Shell Optimization
A cycle of optimization runs has three parts, training and testing, and finally real-time use — a
variant of testing. Training consists in choosing a data set and performing the recursive optimization,
which produces optimization parameters for trading. In our case there are six parameters: the time
resolution ∆t of price data, the length of window W used in the local fitting procedures and in
computation of moving averages of trading signals, the drift f F , volatility coefficient σ and exponent x
from Eq. (28), and a multiplicative factor M necessary for the trading rules module, as discussed below.
The optimization parameters computed from the training set are applied then to various test sets and
final profit/loss analysis are produced. Based on these, the best set of optimization parameters are chosen
to be applied in real-time trading runs. We remark once again that a single training data set could support
more than one profitable sets of parameters and can be a function of the trader’s interest and the specific
market dynamics targeted (e.g., short/long time scales). The optimization parameters corresponding to
the global minimum in the training session may not necessarily represent the parameters that led to robust
profits across real-time data.
The training optimization occurs in two inter-related stages. An inner-shell maximum likelihood
optimization over all training data is performed. The cost function that is fitted to data is the effective
action constructed from the Lagrangian Eq. (27) including the pre-factors coming from the measure
element in the expression of the short-time probability distribution Eq. (12). This is based on the fact [39]
Optimization of Trading
- 20 -
Ingber & Mondescu
that in the context of Gaussian multiplicative stochastic noise, the macroscopic transition probability
P(F, t|F′, t′) to start with the price F′ at t′ and reach the price F at t is determined by the short-time
Lagrangian Eq. (27),
P(F, t|F′, t′) =
1
N
exp
− Σ L(i|i − 1)dt i ,
2x
2
1/2
i=1
(2π σ F i−1 dt i )
(30)
with dt i = t i − t i−1 . Recall that the main assumption of our model is that price increments (or the
logarithm of price ratios, depending on which variables are considered independent) could be described
by a system of coupled stochastic, non-linear equations as in Eq. (10). These equations are deceptively
simple in structure, yet depending on the functional form of the drift coefficients and the multiplicative
noise, they could describe a variety of interactions between financial instruments in various market
conditions (e.g., constant elasticity of variance model [53], stochastic volatility models, etc.). In
particular, this type of models include the case of Black-Scholes price dynamics (x = 1).
In the system presented here, we have applied the model from Eq. (28). The fitted parameters were
the drift coefficient f F and the exponent x. In the case of a coupled futures and cash system, besides the
corresponding values of f F and x for the cash index, another parameter, the correlation coefficient ρ as
introduced in Eq. (10), must be considered.
4.4. Trading Rules (Outer-Shell) Recursive Optimization
In the second part of the training optimization, we calculate the CMI and execute trades as required
by a selected set of trading rules based on CMI values, price data or combinations of both indicators.
Recall that three external shell optimization parameters are defined: the time resolution ∆t of the
data expressed as the time interval between consecutive data points, the window length W (in number of
time epochs or data points) used in the adaptive calculation of CMI, and a numerical coefficient M that
scales the momentum uncertainty discussed below.
At each moment a local refit of f F and σ over data in the local window W is executed, moving the
window M across the training data set and using the zeroth order optimization parameters f F and x
resulting from the inner-shell optimization as a first guess. It was found that a faster quasi-local code is
sufficient for computational purposes for these adaptive updates. In more complicated models, ASA can
be successfully applied recursively, although in real-time trading the response time of the system is a
Optimization of Trading
- 21 -
Ingber & Mondescu
major factor that requires attention.
All expressions that follow can be generalized to coupled systems in the manner described in
Section 3. Here we use the one factor nonlinear model given by Eq. (28). At each time epoch we
calculate the following momentum related quantities:
ΠF =
1
σ 2 F 2x
Π0F = −
dF
− fF ,
dt
fF
,
σ 2 F 2x
∆Π F = < (Π F − < Π F >)2 >1/2 =
1
σ F x√
dt
,
(31)
where we have used < Π F >= 0 as implied by Eqs. (28) and (27). In the previous expressions, Π F is the
CMI, Π0F is the neutral line or the momentum of a zero change in prices, and ∆Π F is the uncertainty of
momentum. The last quantity reflects the Heisenberg principle, as derived from Eq. (28) by calculating
∆F ≡ < (dF− < dF >)2 >1/2 = σ F x √
dt ,
∆Π F ∆F ≥ 1 ,
(32)
where all expectations are in terms of the exact noise distribution, and the calculation implies the Itô
approximation (equivalent to considering non-anticipative functions). Various moving averages of these
momentum signals are also constructed. Other dynamical quantities, as the Hamiltonian, could be used as
well. (By analogy to the energy concept, we found that the Hamiltonian carries information regarding the
overall trend of the market, giving another useful measure of price volatility.)
Regarding the practical implementation of the previous relations for trading, some comments are
necessary. In terms of discretization, if the CMI are calculated at epoch i, then dF i = F i − F i−1 ,
dt i = t i − t i−1 = ∆t, and all prefactors are computed at moment i − 1 by the Itô prescription (e.g.,
x
σ F x = σ F i−1
). The momentum uncertainty band ∆Π F can be calculated from the discretized theoretical
value Eq. (31), or by computing the estimator of the standard deviation from the actual time series of Π F .
There are also two ways of calculating averages over CMI values: One way is to use the set of local
optimization parameters { f F , σ } obtained from the local fit procedure in the current window W for all
CMI data within that window (local-model average). The second way is to calculate each CMI in the
Optimization of Trading
- 22 -
Ingber & Mondescu
current local window W with another set { f F , σ } obtained from a previous local fit window measured
from the CMI data backwards W points (multiple-models averaged, as each CMI corresponds to a
different model in terms of the fitting parameters { f F , σ }).
The last observation is that the neutral line divides all CMI in two classes: long signals, when
Π F > Π0F , as any CMI satisfying this condition indicates a positive price change, and short signals when
Π F < Π0F , which reflects a negative price change.
After the CMI are calculated, based on their meaning as statistical momentum indicators, trades are
executed following a relatively simple model: Entry in and exit from a long (short) trade points are
defined as points where the value of CMIs is greater (smaller) than a certain fraction of the uncertainty
band M ∆Π F (−M ∆Π F ), where M is the multiplicative factor mentioned in the beginning of this
subsection. This is a choice of a symmetric trading rule, as M is the same for long and short trading
signals, which is suitable for volatile markets without a sustained trend, yet without diminishing too
severely profits in a strictly bull or bear region.
Inside the momentum uncertainty band, one could define rules to stay in a previously open trade, or
exit immediately, because by its nature the momentum uncertainty band implies that the probabilities of
price movements in either direction (up or down) are balanced. From another perspective, this type of
trading rule exploits the relaxation time of a strong market advance or decline, until a trend reversal
occurs or it becomes more probable.
Other sets of trading rules are certainly possible, by utilizing not only the current values of the
momenta indicators, but also their local-model or multiple-models averages. A trading rule based on the
maximum distance between the current CMI data ΠiF and the neutral line Π0F shows faster response to
markets evolution and may be more suitable to automatic trading in certain conditions.
Stepping through the trading decisions each trading day of the training set determined the
profit/loss of the training set as a single value of the outer-sell cost function. As ASA importancesampled the outer-shell parameter space {∆t, W , M}, these parameters are fed into the inner shell, and a
new inner-shell recursive optimization cycle begins. The final values for the optimization parameters in
the training set are fixed when the largest net profit (calculated from the total equity by subtracting the
transactions costs defined by the slippage factor) is realized. In practice, we have collected optimization
parameters from multiple local minima that are near the global minimum (the outer-shell cost function is
Optimization of Trading
- 23 -
Ingber & Mondescu
defined with the sign reversed) of the training set.
The values of the optimization parameters {∆t, W , M, f F , σ , x} resulting from a training cycle are
then applied to out-of-sample test sets. During the test run, the drift coefficient f F and the volatility
coefficient σ are refitted adaptively as described previously. All other parameters are fixed. We have
mentioned that the optimization parameters corresponding to the highest profit in the training set may not
be the sufficiently robust across test sets. Then, for all test sets, we have tested optimization parameters
related to the multiple minima (i.e., the global maximum profit, the second best profit, etc.) resulting from
the training set.
We performed a bootstrap-type reversal of the training-test sets (repeating the training runs
procedures using one of the test sets, including the previous training set in the new batch of test sets),
followed by a selection of the best parameters across all data sets. This is necessary to increase the
chances of successful trading sessions in real-time.
5. RESULTS
5.1. Alternative Algorithms
In the previous sections we noted that there are different combinations of methods of processing
data, methods of computing the CMI and various sets of trading rules that need to be tested — at least in a
sampling manner — before launching trading runs in real-time:
1. Data can be preprocessed in block or overlapping bins, or forecasted data derived from the most
probable transition path [41] could be used as in one of our most recent models.
2. Exponential smoothing, wavelets or Fourier decomposition can be applied for statistical
processing. We presently favor exponential moving averages.
3. The CMI can be calculated using averaged data or directly with tick data, although the
optimization parameters were fitted from preprocessed (averaged) price data.
4. The trading rules can be based on current signals (no average is performed over the signal
themselves), on various averages of the CMI trading signals, on various combination of CMI data
(momenta, neutral line, uncertainty band), on symmetric or asymmetric trading rules, or on mixed priceCMI trading signals.
Optimization of Trading
- 24 -
Ingber & Mondescu
5. Different models (one and two-factors coupled) can be applied to the same market instrument,
e.g., to define complementary indicators.
The selection process evidently must consider many specific economic factors (e.g., liquidity of a
given market), besides all other physical, mathematical and technical considerations. In the work
presented here, as we tested our system and using previous experience, we focused toward S&P500
futures electronic trading, using block processed data, and symmetric, local-model and multiple-models
trading rules. In Table 1 we show results obtained for several training and testing sets in the mentioned
context.
5.2. Trading System Design
The design of a successful electronic trading system is complex as it must incorporate several
aspects of a trader’s actions that sometimes are difficult to translate into computer code. Three important
features that must be implemented are factoring in the transactions costs, devising money management
techniques, and coping with execution deficiencies.
Generally, most trading costs can be included under the “slippage factor,” although this could easily
lead to poor estimates. Given that the margin of profits from exploiting market inefficiencies are thin, a
high slippage factor can easily result in a non-profitable trading system. In our situation, for testing
purposes we used a $35 slippage factor per buy & sell order, a value we believe is rather high for an
electronic trading environment, although it represents less than three ticks of a mini-S&P futures contract.
(The mini-S&P is the S&P futures contract that is traded electronically on CME.) This higher value was
chosen to protect ourselves against the bid-ask spread, as our trigger price (at what price the CMI was
generated) and execution price (at what price a trade signaled by a CMI was executed) were taken to be
equal to the trading price. (We have changed this aspect of our algorithm in later models.) The slippage
is also strongly influenced by the time resolution of the data. Although the slippage is linked to bid-ask
spreads and markets volatility in various formulas [54], the best estimate is obtained from experience and
actual trading.
Money management was introduced in terms of a trailing stop condition that is a function of the
price volatility, and a stop-loss threshold that we fixed by experiment to a multiple of the mini-S&P
contract value ($200). It is tempting to tighten the trailing stop or to work with a small stop-loss value,
Optimization of Trading
- 25 -
Ingber & Mondescu
yet we found — as otherwise expected — that higher losses occurred as the signals generated by our
stochastic model were bypassed.
Regarding the execution process, we have to account for the response of the system to various
execution conditions in the interaction with the electronic exchange: partial fills, rejections, uptick rule
(for equity trading), etc. Except for some special conditions, all these steps must be automated.
5.3. Some Explicit Results
Typical CMI data in Figs. 3 and 4 (obtained from real-time trading after a full cycle of trainingtesting was performed) are related to the price data in Figs. 1 and 2. We have plotted the fastest (55 secs
apart) CMI values Π F , the neutral line Π0F and the uncertainty band ∆Π F . All CMI data were produced
using the optimization parameters set {55 secs, 88 epochs, 0. 15} of the second-best net profit obtained
with the training set “4D ESM0 0321-0324” (Table 1).
Although the CMIs exhibit an inherently ragged nature and oscillate around a zero mean value
within the uncertainty band — the width of which is decreasing with increasing price volatility, as the
uncertainty principle would also indicate — time scales at which the CMI average or some persistence
time are not balanced about the neutral line.
These characteristics, which we try to exploit in our system, are better depicted in Figs. 5 and 6.
One set of trading signals, the local-model average of the neutral line < Π0F > and the uncertainty band
multiplied by the optimization factor M = 0. 15, and centered around the theoretical zero mean of the
CMI, is represented versus time. Note entry points in a short trading position (< Π0F > > M ∆Π F ) at
around 10:41 (Fig. 5 in conjunction with S&P data in Fig. 1) with a possible exit at 11:21 (or later), and a
first long entry (< Π0F > < − M ∆Π F ) at 12:15. After 14:35, a stay long region appears (< Π0F > < 0),
which indicates correctly the price movement in Fig.1.
In Fig. 6 corresponding to June 22 price data from Fig. 2, a first long signal is generated at around
12:56 and a first short signal is generated at 14:16 that reflects the long downtrend region in Fig. 2. Due
to the averaging process, a time lag is introduced, reflected by the long signal at 12:56 in Fig. 4, related to
a past upward trend seen in Fig. 2; yet the neutral line relaxes rather rapidly (given the 55 sec time
resolution and the window of 88 ≈ 1.5 hour) toward the uncertainty band. A judicious choice of trading
rules, or avoiding standard averaging methods, helps in controlling this lag problem.
Optimization of Trading
- 26 -
Ingber & Mondescu
In Tables 1 and 2 we show some results obtained for several training and testing sets following the
procedures described at the end of the previous section. In both tables, under the heading “Training” or
“Testing Set” we specify the data set used (e.g., “4D ESM0 0321-0324” represents four days of data from
the mini-S&P futures contract that expired in June). The type of trading rules used is identified by
“LOCAL MODEL” or “MULTIPLE MODELS” tags. These tags refer to how we calculate the averages
of the trading signals: either by using a single pair of optimization parameters { f F , σ } for all CMI data
within the current adaptive fit window, or a different pair { f F , σ } for each CMI data. In the “Statistics”
column we report the net (subtracting the slippage) profit or loss (in parenthesis) across the whole data
set, the total number of trades (“trades”), the number of days with positive balance (“days +”), and the
percentage of winning trades (“winners”). The “Parameters” are the optimization parameters resulting
from the first three best profit maxima of each listed training set. The parameters are listed in the order
{∆t, W , M}, with the data time resolution ∆t measured in seconds, the length of the local fit window W
measured in time epochs, and M the numerical coefficient of the momentum uncertainty band ∆Π F .
Recall that the trading rules presented are symmetric (the long and short entry/exit signals are
controlled by the same M factor), and we apply a stay-long condition if the neutral-line is below the
average momentum < Π F >= 0 and stay-short if Π0F > 0. The drift f F and volatility coefficient σ are
refitted adaptively and the exponent x is fixed to the value obtained in the training set. Typical values are
f F ∈ ± [0. 003: 0. 05], x ∈ ± [0. 01: 0. 03]. During the local fit, due to the shorter time scale involved, the
drift may increase by a factor of ten, and σ ∈[0. 01: 1. 2].
Comparing the data in the training and testing tables, we note that the most robust optimization
factors — in terms of maximum cumulative profit resulted for all test sets — do not correspond to the
maximum profit in the training sets: For the local-model rules, the optimum parameters are
{55, 88, 0. 15}, and for the multiple models rules the optimum set is {45, 72, 0. 2}, both realized by the
training set “4D ESM0 0321-0324.”
Other observations are that, for the data presented here, the multiple-models averages trading rules
consistently performed better and are more robust than the local-model averages trading rules. The
number of trades is similar, varying between 15 and 35 (eliminating cumulative values smaller than 10
trades), and the time scale of the local fit is rather long in the 30 mins to 1.5 hour range. In the current
set-up, this extended time scale implies that is advisable to deploy this system as a trader-assisted tool.
Optimization of Trading
- 27 -
Ingber & Mondescu
An important factor is the average length of the trades. For the type of rules presented in this work,
this length is of several minutes, up to one hour, as the time scale of the local fit window mentioned above
suggested.
Related to the length of a trade is the length of a winning long/short trade in comparison to a losing
long/short trade. Our experience indicates that a ratio of 2:1 between the length of a winning trade and
the length of a losing trade is desirable for a reliable trading system. Here, using the local-model trading
rules seems to offer an advantage, although this is not as clear as one would expect.
Finally, the training sets data (Table 1) show that the percentage of winners is markedly higher in
the case of multiple-models average than local-average trading rules. In the testing sets (Table 2) the
situation is almost reversed, albeit the overall profits (losses) are higher (smaller) in the multiple-model
case. Apparently, the multiple-model trading rules can stay in winning trades longer to increase profits,
relative to losses incurred with these rules in losing trades. (In the testing sets, this correlates with the
higher number of trades executed using local-model trading rules.)
6. CONCLUSIONS
6.1. Main Features
The main stages of building and testing this system were:
1. We developed a multivariate, nonlinear statistical mechanics model of S&P futures and cash
markets, based on a system of coupled stochastic differential equations.
2. We constructed a two-stage, recursive optimization procedure using methods of ASA global
optimization: An inner-shell extracts the characteristics of the stochastic price distribution and an outershell generates the technical indicators and optimize the trading rules.
3. We trained the system on different sets of data and retained the multiple minima generated
(corresponding to the global maximum net profit realized and the neighboring profit maxima).
4. We tested the system on out-of-sample data sets, searching for most robust optimization
parameters to be used in real-time trading. Robustness was estimated by the cumulative profit/loss across
diverse test sets, and by testing the system against a bootstrap-type reversal of training-testing sets in the
optimization cycle.
Optimization of Trading
- 28 -
Ingber & Mondescu
Modeling the market as a dynamical physical system makes possible a direct representation of
empirical notions as market momentum in terms of CMI derived naturally from our theoretical model.
We have shown that other physical concepts as the uncertainty principle may lead to quantitative signals
(the momentum uncertainty band ∆Π F ) that captures other aspects of market dynamics and which can be
used in real-time trading.
6.2. Summary
We have presented a description of a trading system composed of an outer-shell trading-rule model
and an inner-shell nonlinear stochastic dynamic model of the market of interest, S&P500. The inner-shell
is developed adhering to the mathematical physics of multivariate nonlinear statistical mechanics, from
which we develop indicators for the trading-rule model, i.e., canonical momenta indicators (CMI). We
have found that keeping our model faithful to the underlying mathematical physics is not a limiting
constraint on profitability of our system; quite the contrary.
An important result of our work is that the ideas for our algorithms, and the proper use of the
mathematical physics faithful to these algorithms, must be supplemented by many practical
considerations en route to developing a profitable trading system. For example, since there is a subset of
parameters, e.g., time resolution parameters, shared by the inner- and outer-shell models, recursive
optimization is used to get the best fits to data, as well as developing multiple minima with approximate
similar profitability. The multiple minima often have additional features requiring consideration for realtime trading, e.g., more trades per day increasing robustness of the system, etc. The nonlinear stochastic
nature of our data required a robust global optimization algorithm. The output of these parameters from
these training sets were then applied to testing sets on out-of-sample data. The best models and
parameters were then used in real-time by traders, further testing the models as a precursor to eventual
deployment in automated electronic trading.
We have used methods of statistical mechanics to develop our inner-shell model of market
dynamics and a heuristic AI type model for our outer-shell trading-rule model, but there are many other
candidate (quasi-)global algorithms for developing a cost function that can be used to fit parameters to
data, e.g., neural nets, fractal scaling models, etc. To perform our fits to data, we selected an algorithm,
Adaptive Simulated Annealing (ASA), that we were familiar with, but there are several other candidate
Optimization of Trading
- 29 -
Ingber & Mondescu
algorithms that likely would suffice, e.g., genetic algorithms, tabu search, etc.
We have shown that a minimal set of trading signals (the CMI, the neutral line representing the
momentum of the trend of a given time window of data, and the momentum uncertainty band) can
generate a rich and robust set of trading rules that identify profitable domains of trading at various time
scales. This is a confirmation of the hypothesis that markets are not efficient, as noted in other
studies [11,30,55].
6.3. Future Directions
Although this paper focused on trading of a single instrument, the futures S&P 500, the code we
have developed can accommodate trading on multiple markets. For example, in the case of tickresolution coupled cash and futures markets, which was previously prototyped for inter-day
trading [29,30], the utility of CMI stems from three directions:
(a) The inner-shell fitting process requires a global optimization of all parameters in both futures
and cash markets.
(b) The CMI for futures contain, by our Lagrangian construction, the coupling with the cash market
through the off-diagonal correlation terms of the metric tensor. The correlation between the futures and
cash markets is explicitly present in all futures variables.
(c) The CMI of both markets can be used as complimentary technical indicators for trading in
futures market.
Several near term future directions are of interest: orienting the system toward shorter trading time
scales (10-30 secs) more suitable for electronic trading, introducing fast response “averaging” methods
and time scale identifiers (exponential smoothing, wavelets decomposition), identifying mini-crashes
points using renormalization group techniques, investigating the use of CMI in pattern-recognition based
trading rules, and exploring the use of forecasted data evaluated from most probable transition path
formalism.
Our efforts indicate the invaluable utility of a joint approach (AI-based and quantitative) in
developing automated trading systems.
Optimization of Trading
- 30 -
Ingber & Mondescu
6.4. Standard Disclaimer
We must emphasize that there are no claims that all results are positive or that the present system is
a safe source of riskless profits. There as many negative results as positive, and a lot of work is necessary
to extract meaningful information. Our purpose here is to describe an approach to developing an
electronic trading system complementary to those based on neural-networks type technical analysis and
pattern recognition methods. The system discussed in this paper is rooted in the physical principles of
nonequilibrium statistical mechanics, and we have shown that there are conditions under which such a
model can be successful.
ACKNOWLEDGMENTS
We thank Donald Wilson for his financial support. We thank K.S. Balasubramaniam and Colleen
Chen for their programming support and participation in formulating parts of our trading system. Data
was extracted from the DRW Reuters feed.
Optimization of Trading
- 31 -
Ingber & Mondescu
REFERENCES
[1]
E. Peters, Chaos and Order in the Capital Markets, Wiley & Sons, New York,NY, 1991.
[2]
E.M. Azoff, Neural Network Time Series Forecasting of Financial Markets, Wiley & Sons, New
York,NY, 1994.
[3]
E. Gately, Neural Networks for financial Forecasting, Wiley & Sons, New York,NY, 1996.
[4]
L. Ingber, ‘‘High-resolution path-integral development of financial options,’’ Physica A, vol. 283,
pp. 529-558, 2000.
[5]
B.B. Mandelbrot, Fractals and Scaling in Finance, Springer-Verlag, New York, NY, 1997.
[6]
R.N. Mantegna and H.E. Stanley, ‘‘Turbulence and financial markets,’’ Nature, vol. 383, pp.
587-588, 1996.
[7]
L. Laloux, P. Cizeau, J.-P. Bouchaud, and M. Potters, ‘‘Noise dressing of financial correlation
matrices,’’ Phys. Rev. Lett., vol. 83, pp. 1467-1470, 1999.
[8]
A. Johansen, D. Sornette, and O. Ledoit, Predicting financial crashes using discrete scale
invariance, http://xxx.lanl.gov/cond-mat, 2000.
[9]
K. Ilinsky and G. Kalinin, Black-Scholes equation from gauge theory of arbitrage,
http://xxx.lanl.gov/hep-th/9712034, 1997.
[10] J.C. Hull, Options, Futures, and Other Derivatives, 4th Edition, Prentice Hall, Upper Saddle River,
NJ, 2000.
[11] L. Ingber, ‘‘Statistical mechanics of nonlinear nonequilibrium financial markets,’’ Math. Modelling,
vol. 5, pp. 343-361, 1984.
[12] L. Ingber, ‘‘Adaptive Simulated Annealing (ASA),’’ Global optimization C-code, Caltech Alumni
Association, Pasadena, CA, 1993.
[13] L. Ingber, ‘‘Very fast simulated re-annealing,’’ Mathl. Comput. Modelling, vol. 12, pp. 967-973,
1989.
[14] L. Ingber and B. Rosen, ‘‘Genetic algorithms and very fast simulated reannealing: A comparison,’’
Oper. Res. Management Sci., vol. 33, pp. 523, 1993.
Optimization of Trading
- 32 -
Ingber & Mondescu
[15] L. Ingber, ‘‘Adaptive simulated annealing (ASA): Lessons learned,’’ Control and Cybernetics, vol.
25, pp. 33-54, 1996.
[16] L. Ingber, ‘‘Simulated annealing: Practice versus theory,’’ Mathl. Comput. Modelling, vol. 18, pp.
29-57, 1993.
[17] G. Indiveri, G. Nateri, L. Raffo, and D. Caviglia, ‘‘A neural network architecture for defect
detection through magnetic inspection,’’ Report, University of Genova, Genova, Italy, 1993.
[18] B. Cohen, Training synaptic delays in a recurrent neural network, M.S. Thesis, Tel-Aviv University,
Tel-Aviv, Israel, 1994.
[19] R.A. Cozzio-Büeler, ‘‘The design of neural networks using a priori knowledge,’’ Ph.D. Thesis,
Swiss Fed. Inst. Tech., Zurich, Switzerland, 1995.
[20] D.G. Mayer, P.M. Pepper, J.A. Belward, K. Burrage, and A.J. Swain, ‘‘Simulated annealing - A
robust optimization technique for fitting nonlinear regression models,’’ in Proceedings ’Modelling,
Simulation and Optimization’ Conference, International Association of Science and Technology for
Development (IASTED), 6-9 May 1996 Gold Coast, 1996.
[21] S. Sakata and H. White, ‘‘High breakdown point conditional dispersion estimation with application
to S&P 500 daily returns volatility,’’ Econometrica, vol. 66, pp. 529-567, 1998.
[22] L. Bachelier, ‘‘Théorie de la Spéculation,’’ Annales de l’Ecole Normale Supérieure, vol. 17, pp.
21-86, 1900.
[23] M. C. Jensen, ‘‘Some anomalous evidence regarding market efficiency, an editorial introduction,’’ J.
Finan. Econ., vol. 6, pp. 95-101, 1978.
[24] B. B. Mandelbrot, ‘‘When can price be arbitraged efficiently? A limit to the validity of the random
walk and martingale models,’’ Rev. Econ. Statist., vol. 53, pp. 225-236, 1971.
[25] S. J. Taylor, ‘‘Tests of the random walk hypothesis against a price-trend hypothesis,’’ J. Finan.
Quant. Anal., vol. 17, pp. 37-61, 1982.
[26] P. Brown, A. W. Kleidon, and T. A. Marsh, ‘‘New evidence on the nature of size-related anomalies
in stock prices,’’ J. Fin. Econ., vol. 12, pp. 33-56, 1983.
[27] J.A. Nelder and R. Mead, ‘‘A simplex method for function minimization,’’ Computer J. (UK), vol.
7, pp. 308-313, 1964.
Optimization of Trading
- 33 -
Ingber & Mondescu
[28] G.P. Barabino, G.S. Barabino, B. Bianco, and M. Marchesi, ‘‘A study on the performances of
simplex methods for function minimization,’’ in Proc. IEEE Int. Conf. Circuits and Computers,
1980, pp. 1150-1153.
[29] L. Ingber, ‘‘Canonical momenta indicators of financial markets and neocortical EEG,’’ in Progress
in Neural Information Processing, , ed. by S.-I. Amari, L. Xu, I. King, and K.-S. Leung, Springer,
New York, 1996, pp. 777-784.
[30] L. Ingber, ‘‘Statistical mechanics of nonlinear nonequilibrium financial markets: Applications to
optimized trading,’’ Mathl. Computer Modelling, vol. 23, pp. 101-121, 1996.
[31] H. Haken, Synergetics, 3rd ed., Springer, New York, 1983.
[32] L. Ingber, M.F. Wehner, G.M. Jabbour, and T.M. Barnhill, ‘‘Application of statistical mechanics
methodology to term-structure bond-pricing models,’’ Mathl. Comput. Modelling, vol. 15, pp.
77-98, 1991.
[33] L. Ingber, ‘‘Statistical mechanics of neocortical interactions: A scaling paradigm applied to
electroencephalography,’’ Phys. Rev. A, vol. 44, pp. 4017-4060, 1991.
[34] K.S. Cheng, ‘‘Quantization of a general dynamical system by Feynman’s path integration
formulation,’’ J. Math. Phys., vol. 13, pp. 1723-1726, 1972.
[35] H. Dekker, ‘‘Functional integration and the Onsager-Machlup Lagrangian for continuous Markov
processes in Riemannian geometries,’’ Phys. Rev. A, vol. 19, pp. 2102-2111, 1979.
[36] R. Graham, ‘‘Path-integral methods in nonequilibrium thermodynamics and statistics,’’ in
Stochastic Processes in Nonequilibrium Systems, , ed. by L. Garrido, P. Seglar, and P.J. Shepherd,
Springer, New York, NY, 1978, pp. 82-138.
[37] F. Langouche, D. Roekaerts, and E. Tirapegui, ‘‘Discretization problems of functional integrals in
phase space,’’ Phys. Rev. D, vol. 20, pp. 419-432, 1979.
[38] F. Langouche, D. Roekaerts, and E. Tirapegui, ‘‘Short derivation of Feynman Lagrangian for
general diffusion process,’’ J. Phys. A, vol. 113, pp. 449-452, 1980.
[39] F. Langouche, D. Roekaerts, and E. Tirapegui, Functional Integration and Semiclassical
Expansions, Reidel, Dordrecht, The Netherlands, 1982.
Optimization of Trading
- 34 -
Ingber & Mondescu
[40] M. Rosa-Clot and S. Taddei, A path integral approach to derivative security pricing: I. Formalism
and analytic results, INFN, Firenze, Italy, 1999.
[41] H. Dekker, ‘‘On the most probable transition path of a general diffusion process,’’ Phys. Lett. A, vol.
80, pp. 99-101, 1980.
[42] P. Hagedorn, Non-Linear Oscillations, Oxford Univ., New York, NY, 1981.
[43] B. Oksendal, Stochastic Differential Equations, Springer, New York, NY, 1998.
[44] J.M. Harrison and D. Kreps, ‘‘Martingales and arbitrage in multiperiod securities markets,’’ J.
Econ. Theory, vol. 20, pp. 381-408, 1979.
[45] S.R. Pliska, Introduction to Mathematical Finance, Blackwell, Oxford, UK, 1997.
[46] C.W. Gardiner, Handbook of Stochastic Methods for Physics, Chemistry and the Natural Sciences,
Springer-Verlag, Berlin, Germany, 1983.
[47] R. C. Merton, ‘‘An intertemporal capital asset pricing model,’’ Econometrica, vol. 41, pp. 867-887,
1973.
[48] L. Ingber, ‘‘Statistical mechanics of neocortical interactions: Stability and duration of the 7+−2 rule
of short-term-memory capacity,’’ Phys. Rev. A, vol. 31, pp. 1183-1186, 1985.
[49] S. Kirkpatrick, C.D. Gelatt, Jr., and M.P. Vecchi, ‘‘Optimization by simulated annealing,’’ Science,
vol. 220, pp. 671-680, 1983.
[50] S. Geman and D. Geman, ‘‘Stochastic relaxation, Gibbs distribution and the Bayesian restoration in
images,’’ IEEE Trans. Patt. Anal. Mac. Int., vol. 6, pp. 721-741, 1984.
[51] H. Szu and R. Hartley, ‘‘Fast simulated annealing,’’ Phys. Lett. A, vol. 122, pp. 157-162, 1987.
[52] M. Wofsey, ‘‘Technology: Shortcut tests validity of complicated formulas,’’ The Wall Street
Journal, vol. 222, pp. B1, 1993.
[53] J. C. Cox and S. A. Ross, ‘‘The valuation of options for alternative stochastic processes,’’ J. Fin.
Econ., vol. 3, pp. 145-166, 1976.
[54] P.J. Kaufman, Trading systems and Methods, 3rd edition, John Wiley & Sons, New York, NY, 1998.
[55] W. Brock, J. Lakonishok, and B. LeBaron, ‘‘Simple technical trading rules and the stochastic
properties of stock returns,’’ J. Finance, vol. 47, pp. 1731-1763, 1992.
Optimization of Trading
- 35 -
Ingber & Mondescu
FIGURE CAPTIONS
Figure 1. Futures and cash data, contract ESU0 June 20: solid line — futures; dashed line — cash.
Figure 2. Futures and cash data, contract ESU0 June 22: solid line — futures; dashed line — cash.
Figure 3. CMI data, real-time trading June 20: solid line — CMI; dashed line — neutral line;
dotted line — uncertainty band.
Figure 4. CMI data, real-time trading, June 22: solid line — CMI; dashed line — neutral line;
dotted line — uncertainty band.
Figure 5. CMI trading signals, real-time trading June 20: dashed line — local-model average of the
neutral line; dotted line — uncertainty band multiplied by the optimization parameter M = 0. 15.
Figure 6. CMI trading signals, real-time trading June 22: dashed line — local-model average of the
neutral line; dotted line — uncertainty band multiplied by the optimization parameter M = 0. 15.
Optimization of Trading
TABLE CAPTIONS
Table 1. Matrix of Training Runs.
Table 2. Matrix of Testing Runs.
- 36 -
Ingber & Mondescu
Optimization of Trading
- Figure 1 -
Ingber & Mondescu
ESU0 data June 20
time resolution = 55 secs
1505
Futures
Cash
1500
1495
S&P
1490
1485
1480
1475
1470
1465
06-20 10:46:16
06-20 11:45:53
06-20 12:45:30
06-20 13:45:07
TIME (mm-dd hh-mm-ss)
06-20 14:44:44
Optimization of Trading
- Figure 2 -
Ingber & Mondescu
ESU0 data June 22
time resolution = 55 secs
1485
Futures
Cash
1480
1475
S&P
1470
1465
1460
1455
1450
06-22 12:56:53
06-22 13:56:30
TIME (mm-dd hh-mm-ss)
06-22 14:56:07
Optimization of Trading
- Figure 3 -
Ingber & Mondescu
Canonical Momenta Indicators (CMI)
time resolution = 55 secs
F
Π (CMI Futures)
8
F
Π 0 (neutral CMI)
F
∆Π (theory)
CMI
4
0
-4
-8
06-20 10:46:16
06-20 11:45:53
06-20 12:45:30
06-20 13:45:07
TIME (mm-dd hh-mm-ss)
06-20 14:44:44
Optimization of Trading
- Figure 4 -
Ingber & Mondescu
Canonical Momenta Indicators (CMI)
time resolution = 55 secs
F
8
Π (CMI Futures)
Π
F
(neutral CMI)
0
F
∆Π (theory)
CMI
4
0
-4
-8
06-22 12:56:53
06-22 13:56:30
TIME (mm-dd hh-mm-ss)
06-22 14:56:07
Optimization of Trading
- Figure 5 -
Ingber & Mondescu
Canonical Momenta Indicators (CMI)
time resolution = 55 secs
1
F
<Π 0> (local)
M ∆Π
F
CMI
0.5
0
-0.5
-1
06-20 10:46:16
06-20 11:45:53
06-20 12:45:30
06-20 13:45:07
TIME (mm-dd hh-mm-ss)
06-20 14:44:44
Optimization of Trading
- Figure 6 -
Ingber & Mondescu
Canonical Momenta Indicators
time resolution = 55 secs
1
F
<Π 0>
M ∆Π
F
CMI
0.5
0
-0.5
-1
06-22 12:56:53
06-22 13:56:30
TIME (mm-dd hh-mm-ss)
06-22 14:56:07
Optimization of Trading
- Table 1 -
TRAINING SET
TRADING RULES
STATISTICS
4D ESM0 0321-0324
LOCAL MODEL
Parameters
$ profit (loss)
# trades
# days +
% winners
Parameters
$ profit (loss)
# trades
# days +
% winners
Parameters
$ profit (loss)
# trades
# days +
% winners
Parameters
$ profit (loss)
# trades
# days +
% winners
Parameters
$ profit (loss)
# trades
# days +
% winners
Parameters
$ profit (loss)
# trades
# days +
% winners
MULTIPLE MODELS
5D ESM0 0327-0331
LOCAL MODEL
MULTIPLE MODELS
5D ESM0 0410-0414
LOCAL MODEL
MULTIPLE MODELS
Ingber & Mondescu
PARAMETERS (∆ t W M)
55 90 0.125
1390
16
3
75
45 76 0.175
2270
18
4
83
20 22 0.60
437
15
3
67
45 74 0.25
657.5
3
5
100
50 102 0.10
1875
35
3
60
45 46 0.25
2285
39
3
72
55 88 0.15
1215
16
3
75
45 72 0.20
2167.5
17
4
88
20 24 0.55
352
16
3
63
40 84 0.175
635
19
3
68
50 142 0.10
1847
19
3
58
40 48 0.30
2145
23
3
87
60 40 0.275
1167
17
3
76
60 59 0.215
1117.5
17
3
76
10 54 0.5
(35)
1
0
0
30 110 0.15
227.5
26
2
65
35 142 0.10
1485
34
4
62
60 34 0.30
1922.5
29
3
72
Optimization of Trading
- Table 2 -
TESTING SETS
STATISTICS
5D ESM0 0327-0331
$ profit (loss)
# trades
# days +
% winners
$ profit (loss)
# trades
# days +
% winners
$ profit (loss)
# trades
# days +
% winners
4D ESM0 0403-0407
5D ESM0 0410-0414
Ingber & Mondescu
PARAMETERS (∆ t W M)
LOCAL MODEL
MULTIPLE MODELS
55 90 0.125
55 88 0.15
60 40 0.275
45 76 0.175
45 72 0.20
60 59 0.215
(712)
20
2
50
(30)
18
3
56
750
30
3
60
(857)
17
2
47
258
13
3
54
1227
21
3
62
(1472)
16
1
44
602
16
2
56
(117)
23
3
48
(605)
18
3
67
1340
16
1
50
(530)
23
2
48
(220)
12
1
67
2130
17
1
53
(1125)
20
2
50
(185)
11
1
54
932
13
1
38
(380)
18
3
50
| 5cs.CE
|
arXiv:1802.00918v1 [cs.IT] 3 Feb 2018
Typicality Matching for Pairs of Correlated Graphs
Farhad Shirani
Siddharth Garg
Elza Erkip
Department of Electrical
and Computer Engineering
New York University
New York, New York, 11201
Email: [email protected]
Department of Electrical
and Computer Engineering
New York University
New York, New York, 11201
Email: [email protected]
Department of Electrical
and Computer Engineering
New York University
New York, New York, 11201
Email: [email protected]
Abstract—In this paper, the problem of matching pairs of
correlated random graphs with multi-valued edge attributes is
considered. Graph matching problems of this nature arise in
several settings of practical interest including social network deanonymization, study of biological data, web graphs, etc. An
achievable region for successful matching is derived by analyzing
a new matching algorithm that we refer to as typicality matching.
The algorithm operates by investigating the joint typicality of
the adjacency matrices of the two correlated graphs. Our main
result shows that the achievable region depends on the mutual
information between the variables corresponding to the edge
probabilities of the two graphs. The result is based on bounds on
the typicality of permutations of sequences of random variables
that might be of independent interest.
I. Introduction
Graphical models emerge naturally in a wide range of
phenomena including social interactions, database systems,
and biological systems. In many applications such as DNA
sequencing, pattern recognition, and image processing, it is
desirable to find algorithms to match correlated graphs. In
other applications, such as social networks and database systems, privacy considerations require the network operators to
preclude de-anonymization using graph matching by enforcing
security safeguards. As a result, there is a large body of
work dedicated to characterizing the fundamental limits of
graph matching (i.e. to determine the necessary and sufficient
conditions for reliable matching), as well as the design of
efficient algorithms to achieve these limits.
In the graph matching problem, an agent is given a pair of
correlated graphs: i) an ‘anonymized’ unlabeled graph, and ii)
a ‘de-anonymized’ labeled graph. The agent’s objective is to
recover the correct labeling of the vertices in the anonymized
graph by matching its vertex set to that of the de-anonymized
graph. This is shown in Figure 1. This problem has been considered under varying assumptions on the joint graph statistics.
Graph isomorphism studied in [1]–[3] is an instance of the
matching problem where the two graphs are identical copies
of one another. Under the Erdös- Rényi graph model tight
necessary and sufficient conditions for graph isomorphism
have been derived [4], [5] and polynomial time algorithms
have been proposed [1]–[3]. The problem of matching nonidentical pairs of correlated Erdös-Rényi graphs have been
studied in [6]–[12]. Furthermore, graphs with community
structure have been considered in [13]–[16]. Seeded versions
?
?
?
?
?
?
Fig. 1: An instance of the graph matching problem where the
anonymized graph on the right is to be matched to the deanonymized graph on the left. The edges take values in the
set [0, 3]. The edges with value 0 are represented by vertex
pairs which are not connected and the edges taking values
{1, 2, 3} are represented by the colored edges.
of the graph matching problem, where the agent has access
to side information in the form of partial labelings of the
unlabeled graph have also been studied in [12], [17]–[20].
While great progress has been made in characterizing the
fundamental limits of graph matching, many of the methods
in the literature are designed for specific graph models such
as pairs of Erdős-Rényi graphs with binary valued edges and
are not extendable to general scenarios.
In this work, we propose a new approach for analyzing
graph matching problems based on the concept of typicality
in information theory [21]. The proposed approach finds a
labeling for the vertices in the anonymized graph which results
in a pair of jointly typical adjacency matrices for the two
graphs, where typicality is defined with respect to the induced
joint statistics of the adjacency matrices. It is shown that if
!
log n
I(X1 ; X2 ) = ω
,
n
then it is possible to label the vertices in the anonymized graph
such that almost all of the vertices are labeled correctly, where
I(X1 ; X2 ) represents the mutual information between the edge
distributions in the two graphs and n is the number of vertices.
The proposed approach is general and leads to a matching
strategy which is applicable under a wide range of statistical models. In addition to yielding sufficient conditions
for matching correlated random graphs with multi-valued
edges, our analysis also includes investigating the typicality
of permutations of sequences of random variables which is of
independent interest.
The rest of the paper is organized as follows: Section II
contains the problem formulation. Section III develops the
mathematical machinery necessary to analyze the new matching algorithm. Section IV introduces the typicality matching
algorithm and evaluates its performance. Section V concludes
the paper.
II. Problem Formulation
In this section, we provide our formulation of the graph
matching problem. There are two aspects of our formulation
that differ from the ones considered in [6]–[12]. First, we
consider graphs with multi-valued (i.e. not necessarily binaryvalued) edges. Second, we consider a relaxed criteria for
successful matching. In prior works, a matching algorithm
is said to succeed if every vertex in the anonymized graph
is matched correctly to the corresponding vertex in the deanonymized graph with vanishing probability of error. In our
formulation, a matching algorithm is successful if a randomly
and uniformly chosen vertex in the graph is matched correctly
with vanishing probability of error. Loosely speaking, this
requires that almost all of the vertices be matched correctly.
We consider graphs whose edges take multiple values.
Graphs with multi-valued edges appear naturally in various
applications where relationships among entities have attributes
such as social network de-anonymization, study of biological
data, web graphs, etc. An edge which has an attribute assignment is called a marked edge. The following defines an
unlabeled graph whose edges take l different values where
l ≥ 2.
Definition 1. An (n, l)-unlabeled graph g is a structure
(Vn , En ), where n ∈ N and l ≥ 2. The set Vn =
{vn,1 , vn,2 , · · · , vn,n } is called the vertex set, and the set En ⊂
{(x, vn,i , vn, j )|x ∈ [0, l − 1], i ∈ [1, n], j ∈ [1, n]} is called the
marked edge set of the graph. For the marked edge (x, vn,i , vn, j )
the variable ‘x’ represents the value assigned to the edge
(vn,i , vn, j ) between vertices vn,i and vn, j .
Without loss of generality, we assume that for any arbitrary
pair of vertices (vn,i , vn, j ), there exists a unique x ∈ [0, l − 1]
such that (x, vn,i , vn, j ) ∈ En . As an example, for graphs with
binary valued edges if the pair vn,i and vn,i are not connected,
we write (0, vn,i , vn, j ) ∈ En , otherwise (1, vn,i , vn, j ) ∈ En .
Definition 2. For an (n, l)-unlabeled graph g = (Vn , En ), a
labeling is defined as a bijective function σ : Vn → [1, n].
The structure g̃ = (g, σ) is called an (n, l)-labeled graph.
For the labeled graph g̃ the adjacency matrix is defined as
eσ = [G
eσ,i, j ]i, j∈[1,n] where G
eσ,i, j is the unique value such
G
eσ,i, j , σ−1 (i), σ−1 ( j)) ∈ En . The upper triangle (UT)
that (G
eUT
e
corresponding to g̃ is the structure G
σ = [G σ,i, j ]i< j .
Any pair of labelings are related through a permutation as
described below.
Definition 3. For two labelings σ and σ0 , the (σ, σ0 )permutation is defined as the bijection π(σ,σ0 ) , where:
π(σ,σ0 ) (i) = j,
if
σ0 −1 ( j) = σ−1 (i), ∀i, j ∈ [1, n].
Proposition 1 given bellow follows from Definition 3.
Proposition 1. Given an (n, l)-unlabeled graph g = (Vn , En )
and a pair of arbitrary permutations σ, σ0 ∈ S n , the adjacency
matrices corresponding to g̃ = (g, σ) and g̃0 = (g, σ0 ) satisfy
the following equality:
eσ,i, j = G
eσ0 ,π(σ,σ0 ) (i),π(σ,σ0 ) ( j) .
G
eσ0 = Π(σ,σ0 ) (G
eσ ), where Π(σ,σ0 ) is an n2 -length
We write G
UT
eUT
f0 UT
permutation. Similarly, we write G
σ = Π(σ,σ0 ) (G σ ).
Definition 4. Let the random variable X be defined on the
probability space (X, PX ), where X = [0, l − 1]. A marked
Erdös-Rényi (MER) graph gn,PX is a randomly generated (n, l)unlabeled graph with vertex set Vn and edge set En , such that
Pr((x, vn,i , vn, j ) ∈ En ) = PX (x), ∀x ∈ [1, l − 1], vn,i , vn, j ∈ Vn ,
and edges between different vertices are mutually independent.
We consider families of correlated pairs of marked labeled
= (g̃1n,PX , g̃2n,PX ).
Erdös-Rényi graphs g
n,Pn,X1 ,X2
2
1
Definition 5. Let the pair of random variables (X1 , X2 ) be
defined on the probability space (X × X, Pn,X1 ,X2 ), where
X = [0, l−1]. A correlated pair of marked labeled Erdös-Renyi
graphs (CMER) g̃
= (g̃1n,PX , g̃2n,pX ) is characterized by: i)
n,PX1 ,X2
1
2
the pair of marked Erdös-Renyi graphs gin,PX , i ∈ {1, 2}, ii) the
i
pair of labelings σi for the unlabeled graphs gin,PX , i ∈ {1, 2},
i
and iii) the probability distribution PX1 ,X2 (x1 , x2 ), (x1 , x2 ) ∈
X × X, such that:
1)The pair gin,PX , i ∈ {1, 2} have the same set of vertices
i
Vn = Vn1 = Vn2 .
2) For any two marked edges ei = (xi , vin,s1 , vin,s2 ), i ∈
{1, 2}, x1 , x2 ∈ [0, l − 1], we have
Pr e1 ∈ E1n , e2 ∈ E2n =
if σ1 (v1n,s j ) = σ2 (v2n,s j ), j ∈ {1, 2}
PX1 ,X2 (x1 , x2 ),
.
P (x )P (x ), Otherwise
X1
1
X2
2
Definition 6. For a given joint distribution PX1 ,X2 , a correlated
pair of marked partially labeled Erdös-Renyi graphs (CMPER)
g
= (g̃1n,PX , g2n,PX ) is characterized by: i) the pair of
n,PX1 ,X2
1
2
marked Erdös-Renyi graphs gin,PX , i ∈ {1, 2}, ii) a labeling σ1
i
for the unlabeled graph g1n,pX , and iii) a probability distribu1
tion PX1 ,X2 (x1 , x2 ), (x1 , x2 ) ∈ X × X, such that there exists a
2
2
labeling σ for the graph gn,PX for which (g̃1n,PX , g̃2n,PX ) is a
2
1
2
CMER, where g̃2n,PX , (g2n,PX , σ2 ).
2
2
The following defines a matching algorithm:
Definition 7. A matching algorithm for the family of CMPERs
= (g̃1n,Pn,X , g2n,Pn,X ), n ∈ N is a sequence of functions
g
n,Pn,X1 ,X2
1
2
7→ σ̂2n such that P σ2n (v2n,J ) = σ̂2n (v2n,J ) → 1
fn : g
n,Pn,X1 ,X2
as n → ∞, where the random variable J is uniformly
distributed over [1, n] and σ2n is the labeling for the graph
g2n,Pn,X for which (g̃1n,Pn,X , g̃2n,Pn,X ) is a CMER, where g̃2n,Pn,X ,
2
1
2
2
(g2n,Pn,X , σ2n ).
2
Note that in the above definition, for fn to be a matching
algorithm, the fraction of vertices whose labels are matched
incorrectly must vanish as n approaches infinity. This is a
relaxation of the criteria considered in [6]–[12] where all of the
vertices are required to be matched correctly simultaneously
with vanishing probability of error as n → ∞.
The following defines an achievable region for the graph
matching problem.
Definition 8. For the graph matching problem, a family of
e = (Pn )n∈N is said to be in the achievable
sets of distributions P
region if for every sequence of distributions Pn,X1 ,X2 ∈ Pn , n ∈
N, there exists a matching algorithm.
III. Permutations of Typical Sequences
In this section, we develop the mathematical tools necessary
to analyze the performance of the typicality matching strategy.
In summary, the typicality matching strategy operates as
the strategy finds a labeling
follows. Given a CMPER g
n,Pn,X1 ,X2
e1 1 , G
e2 2 ) are
σ̂2 such that the pair of adjacency matrices (G
σ
σ̂
jointly typical with respect to Pn,X1 ,X2 , where joint typicality is
defined in Section IV. Each labeling σ̂2 gives a permutation of
e2 2 . Hence, analyzing the performance
the adjacency matrix G
σ
of the typicality matching strategy requires bounds on the
probability of typicality of permutations of correlated pairs
of sequences of random variables. The necessary bounds are
derived in this section. The details of typicality matching and
its performance are described in Section IV.
Definition 9. Let the pair of random variables (X, Y) be
defined on the probability space (X×Y, PX,Y ), where X and Y
are finite alphabets. The -typical set of sequences of length
n with respect to PX,Y is defined as:
An (X, Y) =
o
n
1
(xn , yn ) : N(α, β|xn , yn ) − PX,Y (α, β) ≤ , ∀(α, β) ∈ X × Y ,
n
where
> 0, n ∈ N, and N(α, β|xn , yn ) =
Pn
i=1 1 ((xi , yi ) = (α, β)).
We follow the notation used in [22] in our study of permutation groups.
Definition 10. A permutation on the set of numbers [1, n] is
a bijection π : [1, n] → [1, n]. The set of all permutations on
the set of numbers [1, n] is denoted by S n .
Definition 11. A permutation π ∈ S n , n ∈ N is called a cycle
if there exists m ∈ [1, n] and α1 , α2 , · · · , αm ∈ [1, n] such that
i) π(αi ) = αi+1 , i ∈ [1, m − 1], ii) π(αn ) = α1 , and iii) π(β) = β
if β , αi , ∀i ∈ [1, m]. The variable m is called the length of the
cycle. The element α is called a fixed point of the permutation
if π(α) = α. We write π = (α1 , α2 , · · · , αm ). The permutation π
is called a non-trivial cycle if m ≥ 2.
Lemma 1. [22] Every permutation π ∈ S n , n ∈ N has
a unique representation as a product of disjoint non-trivial
cycles.
Definition 12. For a given sequence yn ∈ Rn and permutation
π ∈ S n , the sequence zn = π(yn ) is defined as zn = (yπ(i) )i∈[1,n] .1
For a correlated pair of independent and identically distributed (i.i.d) sequences (X n , Y n ) and an arbitrary permutation π ∈ S n , we are interested in bounding the probability
P((X n , π(Y n )) ∈ An (X, Y)). As an intermediate step, we first
find a suitable permutation π0 for which P((X n , π(Y n )) ∈
An (X, Y)) ≤ P((X n , π0 (Y n )) ∈ An (X, Y)). In our analysis, we
make extensive use of the standard permutations defined
below.
Definition 13. For a given n, m, c ∈ N, and 1 ≤ i1 ≤ i2 ≤
P
· · · ≤ ic ≤ n such that n = cj=1 i j + m, an (m, c, i1 , i2 , · · · , ic )permutation is a permutation in S n which has m fixed points
and c disjoint cycles with lengths i1 , i2 , · · · , ic , respectively.
The (m, c, i1 , i2 , · · · , ic )-standard permutation is defined as
the (m, c, i1 , i2 , · · · , ic )-permutation consisting of the cycles
P
Pk−1
Pk
( k−1
j=1 i j + 1,
j=1 i j + 2, · · · ,
j=1 i j ), k ∈ [1, c]. Alternatively,
the (m, c, i1 , i2 , · · · , ic )-standard permutation is defined as:
π = (1, 2, · · · , i1 )(i1 + 1, i1 + 2, · · · , i1 + i2 ) · · ·
c
c−1
c−1
X
X
X
i j )(n − m + 1)(n − m + 2) · · · (n).
i j + 2, · · · ,
( i j + 1,
j=1
j=1
j=1
Example 1. The (2, 2, 3, 2)-standard permutation is a permutation which has m = 2 fixed points and c = 2 cycles.
The first cycle has length i1 = 3 and the second cycle has
length i2 = 2. It is a permutation on sequences of length
P
n = cj=1 i j + m = 3 + 2 + 2 = 7. The permutation is
given by π = (123)(45)(6)(7). For an arbitrary sequence
α = (α1 , α2 , · · · , αn ), we have:
π(α) = (α3 , α1 , α2 , α5 , α4 , α6 , α7 ).
Proposition 2. Let (X n , Y n ) be a pair of i.i.d sequences defined
on finite alphabets. We have:
i) For an arbitrary permutation π ∈ S n ,
P((π(X n ), π(Y n )) ∈ An (X, Y)) = P((X n , Y n ) ∈ An (X, Y)).
ii) let n, m, c, i1 , i2 , · · · , ic ∈ N be numbers as described
in Definition 13. Let π1 be an arbitrary (m, c, i1 , i2 , · · · , ic )permutation and let π2 be the (m, c, i1 , i2 , · · · , ic )-standard
permutation. Then,
P((X n , π1 (Y n )) ∈ An (X, Y)) = P((X n , π2 (Y n )) ∈ An (X, Y)).
1 Note that in Definitions 10 and 12 we have used π to denote both a scalar
function which operates on the set [1, n] as well as a function which operates
on the vector space Rn .
Proof. The proof of part i) follows from the fact that permuting both X n and Y n by the same permutation does not change
their joint type. For part ii), it is straightforward to show that
there exists a permutation π such that π(π1 ) = π2 (π) [22]. Then
the statement follows from part i):
P (X n , π1 (Y n )) ∈ An (X, Y)
= P (π (X n ) , π (π1 (Y n ))) ∈ An (X, Y)
= P (π (X n ) , π2 (π (Y n ))) ∈ An (X, Y)
(a)
en , π2 Y
en ∈ An (X, Y)
=P X
(b)
= P (X n , π2 (Y n )) ∈ An (X, Y) ,
en , Y
en ) = (π(X n ), π(Y n )). and (b)
where in (a) we have defined (X
n
n
e ,Y
e ) has the same distribution as (X n , Y n ).
holds since (X
For a given permutation π ∈ S n , and sequences (X n , Y n ),
n
A
define U(π)
= π(Y n ). Furthermore, define Z(π),i
= 1((Xi , U(π),i ) ∈
P
A), A ⊆ X × Y. Define PX PY (A) = (x,y)∈A PX (x)PY (y) and
P
PX,Y (A) = (x,y)∈A PX,Y (x, y).
Theorem 1. Let (X n , Y n ) be a pair of i.i.d sequences defined
on finite alphabets X and Y, respectively. There exists ζ > 0
such that for any (m, c, i1 , i2 , · · · , ic )-permutation π, and 0 <
< I(X;Y)
|X||Y| :
P((X n , π(Y n )) ∈ An (X, Y)) ≤ 2−ζn(I(X;Y)−|X||Y|) ,
where√n, m, c, i1 , i2 , · · · , ic ∈ N such that i1 ≥ i2 ≥ · · · ≥ ic , and
m < n.
The proof is provided in the Appendix.
IV. The Typicality Matching Strategy
In this section, we describe the typicality matching algorithm and characterize its achievable region. Given a CMPER
g
= (g̃1n,Pn,X , g2n,Pn,X ), the typicality matching algorithm
n,Pn,X1 ,X2
1
The proof is provided in the Appendix.
Remark 1. For graphs with binary valued edges, Theorem
2 provides bounds on the condition for successful matching
which improve upon the bound given in ( [10] Theorem 1).
It should be noted that a stronger definition for successful
matching is used in [10].
V. Conclusion
We have introduced the typicality matching algorithm for
matching pairs of correlated graphs. The probability of typicality of permutations of sequences of random variables has
been investigated. An achievable region for the typicality
matching algorithm has been derived. The region characterizes
the conditions for successful matching both for graphs with
binary valued edges as well graphs with finite-valued edges.
Appendix
A. Proof of Theorem 1
The proof builds upon some of the results in [25]. Fix an
integer t ≥ 2. We provide an outline of the proof when m = 0
and i1 ≤ t. Let A = {(x, y) ∈ X × X PX PY (x, y) < PX,Y (x, y)}
and ∈ [0, min(x,y)∈X×X (|PX,Y (x, y) − PX PY (x, y)|)]. Note that
Pr((X n , π(Y n )) ∈ An (X, Y)) ≤
n
\
\ 1 X
{(x,y)}
Z(π),i
> PX,Y (x, y) −
Pr
n i=1
(x,y)∈A
n
\ 1 X
{(x,y)}
Z(π),i
< PX,Y (x, y) −
n i=1
(x,y)∈Ac
2
operates as follows. The algorithm finds a labeling σ̂2 , for
e1,UT
e2,UT
which the pair of UT’s G
and G
are jointly typical
σ1
σ̂2
with respect to Pn,X1 ,X2 when viewed as vectors of length n(n−1)
2 .
Specifically, it returns a randomly picked element σ̂2 from the
set:
n(n−1)
2
b
e1,UT
e2,UT
Σ = {σ̂2 |(G
,G
) ∈ A
σ1
σ̂2
ω( 1n ),
provided that Pn,X1 ,X2 is bounded away from 0 as n → ∞.
},
where =
and declares σ̂ as the correct labeling. Note
b
that the set Σ may have more than one element. We will show
that under certain conditions on the joint graph statistics, all of
the elements of b
Σ satisfy the criteria for successful matching
given in Definition 7. In other words, for all of the elements
of b
Σ the probability of incorrect labeling for any given vertex
is arbitrarily small for large n.
For brevity let α x,y =
1
n
Pn
i=1
{(x,y)}
Z(π),i
, x, y ∈ X and let
ij
T (x,y)
j
Also, define c̄ =
2
n
t
1 X {(x,y)}
Z
, j ∈ [1, c].
=
n k=1 (π),k
and t x,y = loge
PX,Y (x,y)
PX (x)PY (y) .
Then,
\
\
P
c̄α x,y > c̄PX,Y (x, y) −
(x,y)∈A
\
c̄α x,y < c̄PX,Y (x, y) −
(x,y)∈Ac
=P
\
ec̄tx,y αx,y > ec̄tx,y PX,Y (x,y)− ),
(x,y)∈X×Y
Theorem 2. For the typicality matching algorithm, a given
e = (Pn )n∈N is achievable, if for
family of sets of distributions P
every sequence of distributions Pn,X1 ,X2 ∈ Pn , n ∈ N:
I(X1 ; X2 ) = ω(
log n
),
n
(1)
where we have used the fact that by construction:
(x, y) ∈ A
t x,y > 0 if
t x,y < 0 if
(x, y) ∈ Ac .
(2)
So,
we have:
P
\
\
P
c̄α x,y > c̄PX,Y (x, y) −
c̄α x,y < c̄PX,Y (x, y) −
\
Y
≤P
c̄t x,y α x,y
e
(x,y)∈X×Y
P
− x,y c̄t x,y PX,Y (x,y)−
(b)
≤e
=e
−
(c)
=e
P
−
x,y
P
c̄t x,y PX,Y (x,y)−
x,y
Y
>
e
c̄t x,y PX,Y (x,y)−
=e
(3)
Y
ec̄tx,y αx,y )
E(
E(e n
c
Y
c̄t x,y PX,Y (x,y)−
P
j=1
E(e
1
t
(4)
(x,y)
x,y t x,y T j
P
)
{(x,y)}
x,y t x,y T j
),
PX,Y (x,y)
PX (x)PY (y) −)
n
= e− t (I(X;Y)−|X||Y| ).
(x,y)∈X×Y
x,y
c̄ Pc
\
α x,y < nPX,Y (x, y) −
(x,y)∈Ac
P
− nt x,y (PX,Y (x,y) loge
(x,y)∈Ac
(a)
α x,y > nPX,Y (x, y) −
(x,y)∈A
(x,y)∈A
\
\
In the next step, we prove the theorem when i1 > t and m = 0.
The proof is similar to the previous case. Following the steps
above, we get:
Pr((X n , π(Y n )) ∈ An (X, Y)) = Pr(V(PX,Y , P̂) ≤ )
c
Y
{(x,y)}
1 P
(c) − P c̄t x,y PX,Y (x,y)−
E(e t x,y tx,y T j )
= e x,y
(5)
j=1
j=1
where in (a) we have used the fact that the exponential
function is increasing and positive, (b) follows from the
Markov inequality and (c) follows from the independence
of T {(x,y)}
and T i{(x,y)} when i , j for arbitrary (x, y) and
j
1
P
{(x,y)}
(x0 , y0 ). Next, we investigate the term E(e t x,y tx,y T j ). Note
P
that by Definition, x,y∈X T {(x,y)}
= i j , ∀ j ∈ [1, c]. Define
j
1 {(x,y)}
S {(x,y)}
=
T
,
j
∈
[1,
c].
Let
B = {(s{(x,y)}
) j∈[1,c],x,y∈X :
j
t j
Pj
ij
{(x,y)}
= t , ∀ j ∈ [1, c]} be the set of possible values
x,y∈X s j
{(x,y)}
taken by (S j
) j∈[1,c],x,y∈X . Note that:
P
E(e
{(x,y)}
x,y t x,y S j
{(x,y)}
(s j
P((s{(x,y)}
) j∈[1,c],x,y∈X )e
j
P
{(x,y)}
x,y t x,y s j
.
(6)
)i∈[1,n],x,y∈X ∈β
For a fixed vector (s{(x,y)}
) j∈[1,c],x,y∈X ∈ β, let V(x,y) be defined as
j
the random variable for which P(V = t(x,y) ) = s{(x,y)}
, x, y ∈ X
j
ij
and P(V = 0) = 1 − t (note that PV is a valid probability
distribution). From Equation (6), we have:
P
E(e
{(x,y)}
x,y t x,y S j
{(x,y)}
(s j
X
≤
P((s{(x,y)}
) j∈[1,c],x,y∈X )eE(Vx,y )
j
=
{(x,y)}
(s j
{(x,y)}
x,y t x,y T 1
EXt (E(e
1
t
P
= EXt (E(e
{(x,y)}
x,y t x,y T 1
1
t
0
where, T 1{(x,y)} =
P((s{(x,y)}
) j∈[1,c],x,y∈X )(1
j
) j∈[1,c],x,y∈X ∈β
1
n
1
t
)|Xt )
0
{(x,y)}
x,y t x,y T 1
1
|Xt )E(e t
Pt
A
k=1 Z(π),k
0 {(x,y)}
P
x,y t x,y T 1
P
00 {(x,y)}
x,y t x,y T 1
00
and T 1 {(x,y)} =
|Xt ). Define,
)|Xt ),
1 Pi1
A
k=t+1 Z(π),k .
n
0
S 1{(x,y)} = 1t T 1{(x,y)} .
0
0 {(x,y)}
x,y t x,y S j
|Xt )
X
ij
ij − 1
1
= (1 − +
(
PX (x)PY (y) + PX (x)PY|X (y|Xt ))etx,y )
n
t
t
x,y
E(e
= 1.
The theorem is proved by the repetitive application of the
above arguments.
n(n−1)
2
i j X {(x,y)} tx,y
sj
e )
− +
t
x,y
ij X
{(x,y)} t x,y
+
E(S (π),i
)e )
t
x,y
ij X ij
= (1 − +
PX (x)PY (y)etx,y ).
n
t
x,y
= (1 −
{(x,y)}
x,y t x,y S j
)→1
as n → ∞.
So, P(b
Σ = φ) → 0 as n → ∞ since the correct labeling is a
member of the set b
Σ. Let (λn )n∈N be an arbitrary sequence of
numbers such that λn = Θ(n). We will show that the probability
that a labeling in b
Σ labels λn vertices incorrectly goes to 0 as
n → ∞. Define the following:
E = {σ0 2 ||σ2 − σ0 2 ||1 ≥ λn },
(7)
We replace t x,y , x, y ∈ X×Y with loge
(7), we conclude that E(e
P
e1,UT
e2,UT
P((G
,G
) ∈ A
σ1
σ2
P((s{(x,y)}
) j∈[1,c],x,y∈X )E(eVx,y )
j
P
P
)=
First, note that
) j∈[1,c],x,y∈X ∈β
(s{(x,y)}
) j∈[1,c],x,y∈X ∈β
j
X
P
B. Proof of Theorem 2
X
)=
1
E(e t
We investigate E(e
Then,
)
X
=
Assume that i1 > t, then we ‘break’ the cycle into smaller
cycles as follows:
PX,Y (x,y)
PX (x)PY (y) .
From Equation
) = 1. From Equation (5),
where || · ||1 is the L1 -norm. The set E is the set of all labelings
which match more than λn vertices incorrectly.
We show the following:
P(E ∩ b
Σ , φ) → 0,
as
n → ∞.
Note that:
[
P(E ∩ b
Σ , φ) = P(
{σ0 2 ∈ b
Σ})
σ02 :||σ2 −σ0 2 ||1 ≥λn
(a)
≤
n
X
X
P(σ0 2 ∈ b
Σ)
i=λn σ0 2 :||σ2 −σ0 2 ||1 =i
n
X
(b)
=
(c)
≤
X
i=λn σ0 2 :||σ2 −σ0 2 ||1 =i
n
X
X
n(n−1)
2
e1,UT
e2,UT
P((G
, Πσ2 ,σ0 2 (G
)) ∈ A
σ1
σ2
2−(
)
n(n−1) λn (λn −1)
)(I(X1 ;X2 )−|X||Y|)
2 −
2
i=λn σ0 2 :||σ2 −σ02 ||1 =i
!
n
X
n
(d)
−Θ(n2 )(I(X1 ;X2 )− 2 )
=
i=λn
i
≤ nn e−Θ(n
2
(!i)2
2
)(I(X1 ;X2 )−|X||Y|)
≤ 2−Θ(n )(I(X1 ;X2 )−|X||Y|)−Θ(
log n
n ))
,
where (a) follows from the union bound, (b) follows from
the definition of b
Σ and Proposition 1, in (c) we have used
Theorem 1 and the fact that ||σ2 − σ0 2 ||1 means that Πσ2 ,σ02
has λn (λ2n −1) fixed points, in (d) we have denoted the number
of derangement of sequences of length i by !i. Note that the
right hand side in the last inequality approaches 0 as n → ∞
as long as (1) holds since = O( logn n ).
References
[1] L. Babai, P. Erdos, and S. M Selkow. Random graph isomorphism.
SIaM Journal on computing, 9(3):628–635, 1980.
[2] B. Bollobás. Random graphs. 2001. Cambridge Stud. Adv. Math, 2001.
[3] T. Czajka and G. Pandurangan. Improved random graph isomorphism.
Journal of Discrete Algorithms, 6(1):85–92, 2008.
[4] P. Erdos and A. Rényi. On the evolution of random graphs. Publ. Math.
Inst. Hung. Acad. Sci, 5(1):17–60, 1960.
[5] E. M Wright. Graphs on unlabelled nodes with a given number of edges.
Acta Mathematica, 126(1):1–9, 1971.
[6] E. Kazemi. Network alignment: Theory, algorithms, and applications.
2016.
[7] L. Yartseva and M. Grossglauser. On the performance of percolation
graph matching. In Proceedings of the first ACM conference on Online
social networks, pages 119–130. ACM, 2013.
[8] P. Pedarsani, D. R Figueiredo, and M. Grossglauser. A bayesian method
for matching two similar graphs without seeds. In Communication, Control, and Computing (Allerton), 2013 51st Annual Allerton Conference
on, pages 1598–1607. IEEE, 2013.
[9] S. Ji, W. Li, M. Srivatsa, and R. Beyah.
Structural data deanonymization: Quantification, practice, and implications. In Proceedings of the 2014 ACM SIGSAC Conference on Computer and
Communications Security, pages 1040–1053. ACM, 2014.
[10] D. Cullina and N. Kiyavash. Exact alignment recovery for correlated
erdos renyi graphs. arXiv preprint arXiv:1711.06783, 2017.
[11] V. Lyzinski. Information recovery in shuffled graphs via graph matching.
arXiv preprint arXiv:1605.02315, 2016.
[12] F. Shirani, S. Garg, and E. Erkip. Seeded graph matching: Efficient
algorithms and theoretical guarantees. In 2017 51st Asilomar Conference
on Signals, Systems and Computers - Pacific Grove, CA, United States,,
Oct. 2017, https : //wp.nyu.edu/ f arhad_shirani/asilomar/.
[13] X. Fu, Z. Hu, Z. Xu, L. Fu, and X. Wang. De-anonymization of social
networks with communities: When quantifications meet algorithms.
arXiv preprint arXiv:1703.09028, 2017.
[14] D. Cullina and N. Kiyavash. Improved achievability and converse
bounds for erdos-renyi graph matching. SIGMETRICS Perform. Eval.
Rev., 44(1):63–72, June 2016.
[15] E. Onaran, S. Garg, and E. Erkip. Optimal de-anonymization in random
graphs with community structure. In Signals, Systems and Computers,
2016 50th Asilomar Conference on,, pages 709–713. IEEE, 2016.
[16] E. Kazemi, L. Yartseva, and M. Grossglauser. When can two unlabeled
networks be aligned under partial overlap? In 2015 53rd Annual Allerton
Conference on Communication, Control, and Computing (Allerton),
pages 33–42, Sept 2015.
[17] E. Kazemi, H. Hassani, and M. Grossglauser. Growing a graph matching
from a handful of seeds. Proceedings of the VLDB Endowment,
8(10):1010–1021, 2015.
[18] C. Chiasserini, M. Garetto, and E. Leonardi. Social network deanonymization under scale-free user relations. IEEE/ACM Transactions
on Networking, 24(6):3756–3769, 2016.
[19] V. Lyzinski, D. E Fishkind, and C. E Priebe. Seeded graph matching for
correlated erdös-rényi graphs. Journal of Machine Learning Research,
15(1):3513–3540, 2014.
[20] M. Fiori, P. Sprechmann, J. Vogelstein, P. Musé, and Guillermo Sapiro.
Robust multimodal graph matching: Sparse coding meets graph matching. In Advances in Neural Information Processing Systems, pages 127–
135, 2013.
[21] I. Csiszár and J. Korner. Information Theory: Coding Theorems for
Discrete Memoryless Systems. Academic Press Inc. Ltd., 1981.
[22] I. Martin Isaacs. Algebra: a graduate course, volume 100. American
Mathematical Soc., 1994.
[23] I. Csiszár. Large-scale typicality of Markov sample paths and consistency of mdl order estimators. IEEE Transactions on Information
Theory, 48(6):1616–1628, 2002.
[24] W. Hoeffding. Probability inequalities for sums of bounded random
variables. Journal of the American statistical association, 58(301):13–
30, 1963.
[25] Xinjia Chen. Concentration inequalities for bounded random vectors.
arXiv preprint arXiv:1309.0003, 2013.
| 7cs.IT
|
NOETHER RESOLUTIONS IN DIMENSION 2
arXiv:1704.01777v1 [math.AC] 6 Apr 2017
ISABEL BERMEJO, EVA GARCÍA-LLORENTE, IGNACIO GARCÍA-MARCO, AND MARCEL MORALES
A BSTRACT. Let R := K[x1 , . . . , xn ] be a polynomial ring over an infinite field K, and let I ⊂ R
be a homogeneous ideal with respect to a weight vector ω = (ω1 , . . . , ωn ) ∈ (Z+ )n such that
dim (R/I) = d. In this paper we study the minimal graded free resolution of R/I as A-module,
that we call the Noether resolution of R/I, whenever A := K[xn−d+1 , . . . , xn ] is a Noether normalization of R/I. When d = 2 and I is saturated, we give an algorithm for obtaining this resolution that involves the computation of a minimal Gröbner basis of I with respect to the weighted
degree reverse lexicographic order. In the particular case when R/I is a 2-dimensional semigroup
ring, we also describe the multigraded version of this resolution in terms of the underlying semigroup. Whenever we have the Noether resolution of R/I or its multigraded version, we obtain
formulas for the corresponding Hilbert series of R/I, and when I is homogeneous, we obtain a
formula for the Castelnuovo-Mumford regularity of R/I. Moreover, in the more general setting
that R/I is a simplicial semigroup ring of any dimension, we provide its Macaulayfication.
As an application of the results for 2-dimensional semigroup rings, we provide a new upper
bound for the Castelnuovo-Mumford regularity of the coordinate ring of a projective monomial
curve. Finally, we describe the multigraded Noether resolution and the Macaulayfication of either
the coordinate ring of a projective monomial curve C ⊆ PnK associated to an arithmetic sequence or
n−1
the coordinate ring of any canonical projection πr (C) of C to PK
.
Keywords: Graded algebra, Noether normalization, semigroup ring, minimal graded free resolution, CohenMacaulay ring, Castelnuovo-Mumford regularity.
1. I NTRODUCTION
Let R := K[x1 , . . . , xn ] be a polynomial ring over an infinite field K, and let I ⊂ R be a
weighted homogeneous ideal with respect to the vector ω = (ω1 , . . . , ωn ) ∈ (Z+ )n , i.e., I is
homogeneous for the grading degω (xi ) = ωi . We denote by d the Krull dimension of R/I and
we assume that d ≥ 1. Suppose A := K[xn−d+1 , . . . , xn ] is a Noether normalization of R/I,
i.e., A ֒→ R/I is an integral ring extension. Under this assumption R/I is a finitely generated
A-module, so to study the minimal graded free resolution of R/I as A-module is an interesting
problem. Set
ψp
ψ1
ψ0
F : 0 −→ ⊕v∈Bp A(−sp,v ) −→ · · · −→ ⊕v∈B0 A(−s0,v ) −→ R/I −→ 0
this resolution, where for all i ∈ {0, . . . , p} Bi denotes some finite set, and si,v are nonnegative
integers. This work concerns the study of this resolution of R/I, which will be called the Noether
resolution of R/I. More precisely, we aim at determining the sets Bi , the shifts si,v and the
morphisms ψi .
One of the characteristics of Noether resolutions is that they have shorter length than the minimal graded free resolution of R/I as R-module. Indeed, the projective dimension of R/I as
A-module is p = d − depth(R/I), meanwhile its projective dimension of R/I as R-module is
n − depth(R/I). Studying Noether resolutions is interesting since they contain valuable information about R/I. For instance, since the Hilbert series is an additive function, we get the Hilbert
series of R/I from its Noether resolution. Moreover, whenever I is a homogeneous ideal, i.e.,
homogeneous for the weight vector ω = (1, . . . , 1), one can obtain the Castelnuovo-Mumford
regularity of R/I in terms of the Noether resolution as reg(R/I) = max{si,v − i | 0 ≤ i ≤ p,
v ∈ Bi }.
1
2
I. BERMEJO, E. GARCÍA-LLORENTE, I. GARCÍA-MARCO, AND M. MORALES
In Section 2 we start by describing in Proposition 1 the first step of the Noether resolution of
R/I. By Auslander-Buchsbaum formula, the depth of R/I equals d − p. Hence, R/I is CohenMacaulay if and only if p = 0 or, equivalently, if R/I is a free A-module. This observation
together with Proposition 1, lead to Proposition 2 which is an effective criterion for determining
whether R/I is Cohen-Macaulay or not. This criterion generalizes [Bermejo & Gimenez (2001),
Proposition 2.1]. If R/I is Cohen-Macaulay, Proposition 1 provides the whole Noether resolution
of R/I. When d = 1 and R/I is not Cohen-Macaulay, we describe the Noether resolution of
R/I by means of Proposition 1 together with Proposition 3. Moreover, when d = 2 and xn is a
nonzero divisor of R/I, we are able to provide in Theorem 1 a complete description of the Noether
resolution of R/I. All these results rely in the computation of a minimal Gröbner basis of I with
respect to the weighted degree reverse lexicographic order. As a consequence of this, we provide
in Corollary 1 a description of the weighted Hilbert series in terms of the same Gröbner basis.
Whenever I is a homogeneous ideal, as a consequence of Theorem 1, we obtain in Corollary 2 a
formula for the Castelnuovo-Mumford regularity of R/I which is equivalent to the one provided
in [Bermejo & Gimenez (2000), Theorem 2.7].
In section 3 we study Noether resolutions when R/I is a simplicial semigroup ring, i.e., whenever I is a toric ideal and A = K[xn−d+1 , . . . , xn ] is a Noether normalization of R/I. We recall
that I is a toric ideal if I = IA with A = {a1 , . . . , an } ⊂ Nd and ai = (ai1 , . . . , aid ) ∈ Nd ; where
IA denotes the kernel of the homomorphism of K-algebras ϕ : R → K[t1 , . . . , td ]; xi 7→ tai =
t1ai1 · · · tadid for all i ∈ {1, . . . , n}. If we denote by S ⊂ Nd the semigroup generated by a1 , . . . , an ,
then the image of ϕ is K[S] := K[ts | s ∈ S] ≃ R/IA . By [Sturmfels (1996), Corollary 4.3],
IA is multigraded with respect to the grading induced by S which assigns degS (xi ) = ai for all
i ∈ {1, . . . , n}. Moreover, whenever A is a Noether normalization of K[S] we may assume without loss of generality that an−d+i = wn−d+i ei for all i ∈ {1, . . . , d}, where ωn−d+i ∈ Z+ and
{e1 , . . . , ed } is the canonical basis of Nd . In this setting we may consider a multigraded Noether
resolution of K[S], i.e., a minimal multigraded free resolution of K[S] as A-module:
ψp
ψ1
ψ0
0 −→ ⊕s∈Sp A · s −→ · · · −→ ⊕s∈S0 A · s −→ K[S] −→ 0,
where Si are finite subsets of S for all i ∈ {0, . . . , p} and A · s denotes the shifting of A by s ∈ S.
We observe that this multigrading is a refinement of the grading given by ω = (ω1 , . . . , ωn ) with
P
ωi := dj=1 aij ∈ Z+ ; thus, IA is weighted homogeneous with respect to ω. As a consequence,
whenever we get the multigraded Noether resolution or the multigraded Hilbert series of K[S],
we also obtain its Noether resolution and its Hilbert series with respect to the weight vector ω.
A natural and interesting problem is to describe combinatorially the multigraded Noether resolution of K[S] in terms of the semigroup S. This approach would lead us to results for simplicial semigroup rings K[S] which do not depend on the characteristic of the field K. In general, for any toric ideal, it is well known that the minimal number of binomial generators of IA
does not depend on the characteristic of K (see, e.g., [Sturmfels (1996), Theorem 5.3]), but the
Gorenstein, Cohen-Macaulay and Buchsbaum properties of K[S] depend on the characteristic
of K (see [Hoa (1991)], [Trung & Hoa (1986)] and [Hoa (1988)], respectively). However, in the
context of simplicial semigroup rings, these properties do not depend on the characteristic of K
(see [Goto et al. (1976)], [Stanley (1978)] and [García-Sánchez & Rosales (2002)], respectively).
These facts give support to our aim of describing the whole multigraded Noether resolution of
K[S] in terms of the underlying semigroup S for simplicial semigroup rings.
The results in section 3 are the following. In Proposition 5 we describe the first step of the
multigraded Noether resolution of a simplicial semigroup ring K[S]. As a byproduct we recover
in Proposition 6 a well-known criterion for K[S] to be Cohen-Macaulay in terms of the semigroup.
When d = 2, i.e., IA is the ideal of an affine toric surface, Theorem 2 describes the second step of
the multigraded Noether resolution in terms of the semigroup S. When d = 2, from Proposition
5 and Theorem 2, we derive the whole multigraded Noether resolution of K[S] by means of S
and, as a byproduct, we also get in Corollary 3 its multigraded Hilbert series. Whenever IA a is
NOETHER RESOLUTIONS IN DIMENSION 2
3
homogeneous ideal, we get a formula for the Castelnuovo-Mumford regularity of K[S] in terms
of S, see Remark 1.
Given an algebraic variety, the set of points where X is not Cohen-Macaulay is the non CohenMacaulay locus. Macaulayfication is an analogous operation to resolution of singularities and was
considered in Kawasaki [Kawasaki (2000)], where he provides certain sufficient conditions for X
to admit a Macaulayfication. For semigroup rings Goto et al. [Goto et al. (1976)] and Trung and
Hoa [Trung & Hoa (1986)] proved the existence of a semigroup S ′ satisfying S ⊂ S ′ ⊂ S̄, where
¯ is the normalization of K[S], such that we have an
S¯ denotes the saturation of S and thus K[S]
exact sequence:
0 −→ K[S] −→ K[S ′ ] −→ K[S ′ \ S] −→ 0
with dim(K[S ′ \S]) ≤ dim(K[S])−2. In this setting, K[S ′ ] satisfies the condition S2 of Serre, and
is called the S2 -fication of K[S]. Moreover, when S is a simplicial semigroup, [Morales (2007),
Theorem 5] proves that this semigroup ring K[S ′ ] is exactly the Macaulayfication of K[S]; indeed,
he proved that K[S ′ ] is Cohen-Macaulay and the support of K[S ′ \ S] coincides with the non
Cohen-Macaulay locus of K[S]. In [Morales (2007)], the author provides an explicit description
of the Macaulayfication of K[S] in terms of the system of generators of IA provided K[S] is a
codimension 2 simplicial semigroup ring. Section 4 is devoted to study the Macaulayfication of
any simplicial semigroup ring. The main result of this section is Theorem 4, where we entirely
describe the Macaulayfication of any simplicial semigroup ring K[S] in terms of the set S0 , the
subset of S that provides the first step of the multigraded Noether resolution of K[S].
In sections 5 and 6 we apply the methods and results obtained in the previous ones to certain dimension 2 semigroup rings. More precisely, a sequence m1 < · · · < mn determines
the projective monomial curve C ⊂ PnK parametrically defined by xi := smi tmn −mi for all
i ∈ {1, . . . , n − 1}, xn = smn , xn+1 := tmn . If we set A = {a1 , . . . , an+1 } ⊂ N2 where
ai := (mi , mn − mi ), an := (mn , 0) and an+1 := (0, mn ), it turns out that the homogeneous coordinate ring of C is K[C] := K[x1 , . . . , xn+1 ]/IA and A = K[xn , xn+1 ] is a Noether normalization
of R/IA .
The main result in Section 5 is Theorem 5, where we provide an upper bound on the CastelnuovoMumford regularity of K[C], where C is a projective monomial curve. The proof of this bound
is elementary and builds on the results of the previous sections together with some classical
results on numerical semigroups. It is known that reg(K[C]) ≤ mn − n + 1 after the work
[Gruson et al. (1983)]. In our case, [L’vovsky (1996)] obtained a better upper bound, indeed if
we set m0 := 0 he proved that reg(K[C]) ≤ max1≤i<j≤n {mi − mi−1 + mj − mj−1 } − 1. The
proof provided by L’vovsky is quite involved and uses advanced cohomological tools, it would be
interesting to know if our results could yield a combinatorial alternative proof of this result. Even
if L’vovsky’s bound usually gives a better estimate than the bound we provide here, we easily
construct families such that our bound outperforms the one by L’vovsky.
Also in the context of projective monomial curves, whenever m1 < · · · < mn is an arithmetic
sequence of relatively prime integers, the simplicial semigroup ring R/IA has been extensively
studied (see, e.g., [Molinelli & Tamone (1995), Li et. al (2012), Bermejo et al. (2017)]) and the
multigraded Noether resolution is easy to obtain. In Section 6, we study the coordinate ring of
the canonical projections of projective monomial curves associated to arithmetic sequences, i.e.,
the curves Cr whose homogeneous coordinate rings are K[Sr ] = R/IAr , where Ar := A \ {ar }
and Sr ⊂ N2 the semigroup generated by Ar for all r ∈ {1, . . . , n − 1}. In Corollary 5 we give
a criterion for determining when the semigroup ring K[Sr ] is Cohen-Macaulay; whenever it is
not Cohen-Macaulay, we get its Macaulayfication in Corollary 6. Furthermore, in Theorem 7 we
provide an explicit description of their multigraded Noether resolutions. Finally, in Theorem 8 we
get a formula for their Castelnuovo-Mumford regularity.
4
I. BERMEJO, E. GARCÍA-LLORENTE, I. GARCÍA-MARCO, AND M. MORALES
2. N OETHER
RESOLUTION .
G ENERAL
CASE
Let R := K[x1 , . . . , xn ] be a polynomial ring over an infinite field K, and let I ⊂ R be
a ω-homogeneous ideal, i.e., a weighted homogeneous ideal with respect to the vector ω =
(ω1 , . . . , ωn ) ∈ (Z+ )n . We assume that A := K[xn−d+1 , . . . , xn ] is a Noether normalization
of R/I, where d := dim(R/I). In this section we study the Noether resolution of R/I, i.e., the
minimal graded free resolution of R/I as A-module:
(1)
ψp
ψ1
ψ0
F : 0 −→ ⊕v∈Bp A(−sp,v ) −→ · · · −→ ⊕v∈B0 A(−s0,v ) −→ R/I −→ 0,
where for all i ∈ {0, . . . , p} Bi is a finite set of monomials, and si,v are nonnegative integers.
In order to obtain the first step of the resolution, we will deal with the initial ideal of I +
(xn−d+1 , . . . , xn ) with respect to the weighted degree reverse lexicographic order >ω .
We recall that >ω is defined as follows: xα >ω xβ if and only if
• degω (xα ) > degω (xβ ), or
• degω (xα ) = degω (xβ ) and the last nonzero entry of α − β ∈ Zn is negative.
For every polynomial f ∈ R we denote by in (f ) the initial term of f with respect to >ω .
Analogously, for every ideal J ⊂ R, in (J) denotes its initial ideal with respect to >ω .
Proposition 1. Let B0 be the set of monomials that do not belong to in (I + (xn−d+1 , . . . , xn ))
Then,
{xα + I | xα ∈ B0 }
is a minimal set of generators of R/I as A-module and the shifts of the first step of the Noether
resolution (1) are given by degω (xα ) with xα ∈ B0 .
Proof. Since A is a Noether normalization of R/I we have that B0 is a finite set. Let B0 =
{xα1 , . . . , xαk }. To prove that B := {xα1 + I, . . . , xαk + I} is a set of generators of R/I as
βn−d
A-module it suffices to show that for every monomial xβ := xβ1 1 · · · xn−d
∈
/ in (I), one has that
β
α1
αk
x + I ∈ R/I can be written as a linear combination of {x + I, . . . , x + I}. Since {xα1 + (I +
(xn−d+1 , . . . , xn )), . . . , xαk + (I + (xn−d+1 , . . . , xn ))} is a K-basis of R/(I + (xn−d+1 , . . . , xn )),
P
we have that g := xβ − ki=1 λi xαi ∈ I + (xn−d+1 , . . . , xn ) for some λ1 , . . . , λk ∈ K. Then we
deduce that in (g) ∈ in (I + (xn−d+1 , . . . , xn )) which is equal to in (I) + (xn−d+1 , . . . , xn ), and
thus in (g) ∈ in (I). Since xβ ∈
/ in (I) and xαi ∈
/ in (I) for all i ∈ {1, . . . , k}, we conclude that
P
k
αi
β
g = 0 and x + I = ( i=1 λi x ) + I. The minimality of B can be easily proved.
When R/I is a free A-module or, equivalently, when the projective dimension of R/I as Amodule is 0 and hence R/I is Cohen-Macaulay, Proposition 1 provides the whole Noether resolution of R/I. In Proposition 2 we characterize the Cohen-Macaulay property for R/I in terms
of the initial ideal in (I) previously defined. This result generalizes [Bermejo & Gimenez (2001),
Theorem 2.1], which applies for I a homogeneous ideal.
Proposition 2. Let A = K[xn−d+1 , . . . , xn ] be a Noether normalization of R/I. Then, R/I is
Cohen-Macaulay if and only if xn−d+1 , . . . , xn do not divide any minimal generator of in (I).
Proof. We denote by {ev | v in B0 } the canonical basis of ⊕v∈B0 A(−degω (v)). By Proposition 1
we know that ψ0 : ⊕v∈B0 A(−degω (v)) −→ R/I is the morphism induced by ev 7→ v + I ∈ R/I.
By Auslander-Buchsbaum formula, R/I is Cohen-Macaulay if and only if ψ0 is injective.
(⇒) By contradiction, we assume that there exists α = (α1 , . . . , αn ) ∈ Nn such that xα =
α1
x1 · · · xαnn is a minimal generator of in (I) and that αi > 0 for some i ∈ {n − d + 1, . . . , n}.
αn−d
, since in(I + (xn−d+1 , . . . , xn )) = in(I) + (xn−d+1 , . . . , xn ), we have that
Set u := xα1 1 · · · xn−d
′
αn−d+1
· · · xαnn ∈ A and f the remainder of xα modulo the reduced
u ∈ B0 . We also set xα := xn−d+1
Gröbner basis of I with respect to >ω . P
Then xα − f ∈ I and every monomial in f does not
′
belong to in (I). As a consequence, f = ti=1 ci xβi , where ci ∈ K and xβi = vi xβi with vi ∈ B0
NOETHER RESOLUTIONS IN DIMENSION 2
5
P
′
′
′
and xβi ∈ A for all i ∈ {1, . . . , t}. Hence, xα eu − ti=1 ci xβi evi ∈ Ker(ψ0 ) and R/I is not
Cohen-Macaulay.
P
(⇐) Assume that there exists a P
nonzero g ∈ Ker(ψ0 ), namely, g = v∈B0 gv ev ∈ Ker(ψ0 )
with gv ∈ A for all v ∈ B0 . Then, v∈B0 gv v ∈ I. We write in (g) = cxα u with c ∈ K, xα ∈ A
and u ∈ B0 . Since xn−d+1 , . . . , xn do not divide any minimal generator of in (I), we have that
u ∈ in (I), a contradiction.
When R/I has dimension 1, its depth can be either 0 or 1. When depth(R/I) = 1, then R/I
is Cohen-Macaulay and the whole Noether resolution is given by Proposition 1. When R/I is not
Cohen-Macaulay, then its depth is 0 and its projective dimension as A-module is 1. In this setting,
to describe the whole Noether resolution it remains to determine B1 , ψ1 and the shifts s1,v ∈ N for
all v ∈ B1 . In Proposition 3 we explain how to obtain B1 and ψ1 by means of a Gröbner basis of
I with respect to >ω .
Consider χ1 : R −→ R the evaluation morphism induced by xi 7→ xi for i ∈ {1, . . . , n − 1},
xn 7→ 1.
Proposition 3. Let R/I be 1-dimensional ring of depth 0. Let L be the ideal χ1 (in(I)) · R. Then,
B1 = B0 ∩ L
in the Noether resolution (1) of R/I and the shifts of the second step of this resolution are given
by degω (uxδnu ), where u ∈ B1 and δu := min{δ | uxδn ∈ in(I)}.
α
n−1
Proof. For every u = xα1 1 · · · xn−1
∈ B0 ∩ L, there exists δ ∈ N such that uxδn ∈ in(I); let δu be
the minimum of all such δ. Consider pu ∈ R the remainder of uxδnu modulo the reduced Gröbner
basis of I with respect to >ω . Thus uxδnu − pu ∈ I is ω-homogeneous and every monomial xβ
appearing in pu does not belong to in(I), then by Proposition 1 it can be expressed as xβ = vxβnn ,
where βn ≥ 0 and v ∈ B0 . Moreover, since uxδnu >ω xβ , then βn ≥ δu and u >ω v. Thus, we can
write
X
pu =
xδnu muv v,
v∈B0
u>ω v
with muv = cxαuv ∈ A = K[xn ] a monomial (possibly 0) for all v ∈ B0 , u >ω v.
Now we denote by {ev | v in B0 } the canonical basis of ⊕v∈B0 A(− degω (v)) and consider the
graded morphism ψ0 : ⊕v∈B0 A(− degω (v)) −→ R/I induced by ev 7→ v + I ∈ R/I. The above
construction yields that
X
hu := xδnu (eu −
muv ev ) ∈ Ker(ψ0 )
v∈B0
u>ω v
for all u ∈ B0 ∩ L. We will prove that Ker(ψ0 ) is a free A-module with basis
C := {hu | u ∈ B0 ∩ L}.
Firstly, we observe that the A-module generated by the elements of C is free
P due to the triangular
form of the matrix formed by the elements of C. Let us now take g = v∈B0 gv ev ∈ Ker(ψ0 )
with gv ∈ A, we assume that g ∈ ⊕v∈B0 A(− degω (v)) is ω-homogeneous and, thus, gv is either
0 or a monomial of the form cxβnv with c ∈ K and βv ∈ N for all v ∈ B0 . We consider ψ¯0 :
of A-modules induced by ev 7→ v. Since ψ0 (g) =
⊕v∈B0 A(− degω (v)) −→ R the monomorphism
P
′
γ
0, then the polynomial g ′ := ψ¯0 (g) =
g
u∈B0 u u ∈ I and in(g ) = cxn w for some w ∈ B0 ,
γ ∈ N and c ∈ K. Since in(g ′ ) ∈ in(I), we get that w ∈ B0 ∩ L and γ ≥ δw . Hence,
w
g1 := g − cxγ−δ
n−1 hw ∈ Ker(ψ0 ). If g1 is identically zero, then g ∈ ({hu | u ∈ B0 ∩ L}). If g1 is
not zero, we have that 0 6= in(ψ¯0 (g1 )) < in(ψ̄0 (g)) and we iterate this process with g1 to derive
that {hu | u ∈ B0 ∩ L} generates Ker(ψ0 ).
6
I. BERMEJO, E. GARCÍA-LLORENTE, I. GARCÍA-MARCO, AND M. MORALES
The rest of this section concerns I a saturated ideal such that R/I is 2-dimensional and it is not
Cohen-Macaulay (and, in particular, depth(R/I) = 1). We assume that A = K[xn−1 , xn ] is a
Noether normalization of R/I and we aim at describing the whole Noether resolution of R/I. To
achieve this it only remains to describe B1 , ψ1 and the shifts s1,v ∈ N for all v ∈ B1 . In Proposition
4 we explain how to obtain B1 and ψ1 by means of a Gröbner basis of I with respect to >ω . Since
K is an infinite field, I is a saturated ideal and A is a Noether normalization of R/I, one has that
xn + τ xn−1 is a nonzero divisor on R/I for all τ ∈ K but a finite set. Thus, by performing a mild
change of coordinates if necessary, we may assume that xn is a nonzero divisor on R/I.
Now consider χ : R −→ R the evaluation morphism induced by xi 7→ xi for i ∈ {1, . . . , n−2},
xi 7→ 1 for i ∈ {n − 1, n}.
Proposition 4. Let R/I be 2-dimensional, non Cohen-Macaulay ring such that xn is a nonzero
divisor. Let J be the ideal χ(in(I)) · R. Then,
B1 = B0 ∩ J
in the Noether resolution (1) of R/I and the shifts of the second step of this resolution are given
u
), where u ∈ B1 and δu := min{δ | uxδn−1 ∈ in(I)}.
by degω (uxδn−1
Proof. Since xn is a nonzero divisor of R/I and I is a ω-homogeneous ideal, then xn does not
αn−2
divide any minimal generator of in (I). As a consequence, for every u = xα1 1 · · · xn−2
∈ B0 ∩ J,
there exists δ ∈ N such that uxδn−1 ∈ in(I); by definition, δu is the minimum of all such δ.
u
Consider pu ∈ R the remainder of uxδn−1
modulo the reduced Gröbner basis of I with respect to
δu
>ω . Then uxn−1 − pu ∈ I is ω-homogeneous and every monomial xβ appearing in pu does not
βn−1 βn
belong to in(I), then by Proposition 1 it can be expressed as xβ = vxn−1
xn , where βn−1 , βn ≥ 0
δu
β
and v ∈ B0 . Moreover, we have that uxn−1 >ω x which implies that either βn ≥ 1, or βn = 0,
βn−1 ≥ δu and u >ω v. Thus, we can write
X
X
u
pu =
xn guv v,
fuv v +
xδn−1
v∈B0
u>ω v
v∈B0
with fuv ∈ K[xn−1 ] for all v ∈ B0 , u >ω v and guv ∈ A for all v ∈ B0 .
Now we denote by {ev | v in B0 } the canonical basis of ⊕v∈B0 A(− degω (v)) and consider the
graded morphism ψ0 : ⊕v∈B0 A(− degω (v)) −→ R/I induced by ev 7→ v + I ∈ R/I. The above
construction yields that
X
X
u
u
xn guv ev ∈ Ker(ψ0 )
eu −
hu := xδn−1
fuv ev −
xδn−1
v∈B0
u>ω v
v∈B0
for all u ∈ B0 ∩ J. We will prove that Ker(ψ0 ) is a free A-module with basis
C := {hu | u ∈ B0 ∩ J}.
PFirstly, we prove that the A-module generated by the elements of C is free. Assume that
u∈B0 ∩J qu hu = 0 where qu ∈ A for all u ∈ B0 ∩ J and we may also assume that xn does
not divide qv forPsome v ∈ B0 ∩ J. We P
consider the evaluation morphism
τ induced by xn 7→ 0
P
δu
u
and we get that u∈B0 ∩J τ (qu ) τ (hu ) = u∈B0 ∩J τ (qu ) (xn−1 eu + v∈B0 xδn−1
fuv ev ) = 0, which
u>ω v
implies that τ (qu ) =
P0 for all u ∈ B0 ∩ J and, hence, xn | qu for all u ∈ B0 ∩ J, a contradiction.
Let us take g = v∈B0 gv ev ∈ Ker(ψ0 ) with gv ∈ A, we assume that g ∈ ⊕v∈B0 A(− degω (v))
is ω-homogeneous and, thus, gv is either 0 or a ω-homogeneous polynomial for all v ∈ B0 .
We may also suppose that there exists v ∈ B0 such that xn does not divide gv . We consider
by ev 7→ v. Since
ψ¯0 : ⊕v∈B0 A(− degω (v)) −→ R the monomorphism
P of A-modules induced
′
′
¯
ψ0 (g) = 0, then the polynomial g := ψ0 (g) = u∈B0 gu u ∈ I and in(g ) = cxγn−1 w for some
w ∈ B0 and some c ∈ K, which implies that w ∈ B0 ∩ J. By definition of δw we get that γ ≥ δw ,
w
hence g1 := g − cxγ−δ
n−1 hw ∈ Ker(ψ0 ). If g1 is identically zero, then g ∈ ({hu | u ∈ B0 ∩ J}).
NOETHER RESOLUTIONS IN DIMENSION 2
7
If g1 is not zero, we have that 0 6= in(ψ¯0 (g1 )) < in(ψ¯0 (g)) and we iterate this process with g1 to
derive that {hu | u ∈ B0 ∩ J} generates Ker(ψ0 ).
From Propositions 1 and 4 and their proofs, we can obtain the Noether resolution F of R/I by
means of a Gröbner basis of I with respect to >ω . We also observe that for obtaining the shifts
of the resolution it suffices to know a set of generators of in (I). The following theorem gives the
resolution.
Theorem 1. Let R/I be a 2-dimensional ring such that xn is a nonzero divisor. We denote by G
be a Gröbner basis of I with respect to >ω . If δu := min{δ | uxδn−1 ∈ in (I)} for all u ∈ B1 , then
ψ1
ψ0
F : 0 −→ ⊕u∈B1 A(− degω (u) − δu ωn−1 ) −→ ⊕v∈B0 A(− degω (v)) −→ R/I −→ 0,
is the Noether resolution of R/I, where
ψ0 : ⊕v∈B0 A(− degω (v)) → R/I,
ev 7→ v + I
and
whenever
P
ψ1 : ⊕u∈B1 A(− degω (u) − δu ωn−1 ) −→ ⊕v∈B0 A(−P
degω (v)),
δu
eu 7→ xn−1 eu − v∈B0 fuv ev
v∈B0
u
by G.
fuv v with fuv ∈ A is the remainder of the division of uxδn−1
From this resolution, we can easily describe the weighted Hilbert series of R/I.
Corollary 1. Let R/I be a 2-dimensional ring such that xn is a nonzero divisor, then its Hilbert
series is given by:
P
P
degω (v)
− u∈B1 tdegω (u)+δu wn−1
v∈B0 t
HSR/I (t) =
(1 − tωn−1 )(1 − tωn )
In the following example we show how to compute the Noether resolution and the weighted
Hilbert series of the graded coordinate ring of a surface in A4K .
Example 1. Let I be the defining ideal of the surface of A4K parametrically defined by f1 := s3 +
s2 t, f2 := t4 +st3 , f3 := s2 , f4 := t2 ∈ K[s, t]. Using S INGULAR [Decker et al. (2015)], C O C OA
[Abbott et al. (2015)] or M ACAULAY 2 [Grayson & Stillman (2015)] we obtain that whenever K
is a characteristic 0 field, the polynomials {g1 , g2 , g3 , g4} constitute a minimal Gröbner basis of
its defining ideal with respect to >ω with ω = (3, 4, 2, 2), where g1 := 2x2 x23 − x21 x4 + x33 x4 −
x23 x24 , g2 := x41 − 2x21 x33 + x63 − 2x21 x23 x4 − 2x53 x4 + x43 x24 , g3 := x22 − 2x2 x24 − x3 x34 + x44 and
g4 := 2x21 x2 − x21 x3 x4 + x43 x4 − 3x21 x24 − 2x33 x24 + x23 x34 . In particular,
in (I) = (x2 x23 , x41 , x22 , x21 x2 ).
Then, we obtain that
• B0 = {u1, . . . , u6 } with u1 := 1, u2 := x1 , u3 := x2 , u4 := x21 , u5 := x1 x2 , u6 := x31 ,
• J = (x2 , x41 ) ⊂ K[x1 , x2 , x3 , x4 ], and
• B1 = {u3}.
Since x3 divides a minimal generator of in (I), by Proposition 2 we deduce that R/I is not CohenMacaulay. We compute δ3 = min{δ | u3xδ3 ∈ in (I)} and get that δ3 = 2 and that r3 = −x4 u4 +
(x33 x4 − x23 x24 )u1 is the remainder of the division of u3 x23 by G. Hence, following Theorem 1, we
obtain the Noether resolution or R/I:
ψ
F :0−
→ A(−8) −
→
A ⊕ A(−3) ⊕ A(−4)⊕
→ R/I −
−
→ 0,
⊕A(−6) ⊕ A(−7) ⊕ A(−9)
8
I. BERMEJO, E. GARCÍA-LLORENTE, I. GARCÍA-MARCO, AND M. MORALES
where ψ is given by the matrix
−x33 x4 + x23 x24
0
2
x3
x4
0
0
Moreover, by Corollary 1, we obtain that the weighted Hilbert series of R/I is
1 + t3 + t4 + t6 + t7 − t8 + t9
.
(1 − t2 )2
If we consider the same parametric surface over an infinite field of characteristic 2, we obtain
that {x21 + x33 + x23 x4 , x22 + x3 x34 + x44 } is a minimal Gröbner basis of I with respect to >ω , the
weighted degree reverse lexicographic order with ω = (3, 4, 2, 2). Then we have that
HSR/I (t) =
B0 = {v1 := 1, v2 := x1 , v3 := x2 , v4 := x1 x2 },
and B1 = ∅, so R/I is Cohen-Macaulay. Moreover, we also obtain the Noether resolution of R/I
F′ : 0 −
→ A ⊕ A(−3) ⊕ A(−4) ⊕ A(−7) −
→ R/I −
→0
and the weighted Hilbert series of R/I is
HSR/I (t) =
1 + t3 + t4 + t7
.
(1 − t2 )2
To end this section, we consider the particular case where I is standard graded homogeneous,
i.e., ω = (1, . . . , 1). In this setting, we obtain a formula for the Castelnuovo-Mumford regularity
of R/I in terms of in (I) or, more precisely, in terms of B0 and B1 . This formula is equivalent to
that of [Bermejo & Gimenez (2000), Theorem 2.7] provided xn is a nonzero divisor of R/I.
Corollary 2. Let R/I be a 2-dimensional standard graded ring such that xn is a nonzero divisor.
Then,
reg (R/I) = max{deg(v), deg(u) + δu − 1 | v ∈ B0 , u ∈ B1 }
In the following example we apply all the results of this section.
Example 2. Let K be a characteristic zero field and let us consider the projective curve C of P4K
parametrically defined by:
x1 = s3 t5 − st7 , x2 = s7 t, x3 = s4 t4 , x4 = s8 , x5 = t8 .
A direct computation with S INGULAR, C O C OA or M ACAULAY 2 yields that a minimal Gröbner
basis G of the defining ideal I ⊂ R = K[x1 , . . . , x5 ] of C with respect to the degree reverse
lexicographic order consists of 10 elements and that
in (I) = (x41 , x42 , x31 x3 , x1 x3 x24 , x21 x2 , x1 x22 , x1 x2 x3 , x22 x3 , x21 x4 , x23 ).
Then, we obtain that the set B0 is the following
B0 = {u1 := 1, u2 := x1 , u3 := x2 , u4 := x3 , u5 := x21 , u6 := x1 x2 , u7 := x22 ,
u8 := x1 x3 , u9 := x2 x3 , u10 := x31 , u11 := x32 , u12 := x21 x3 }
and the ideal J is
J = (x21 , x1 x3 , x23 , x22 x3 , x42 ) ⊂ R.
Thus, B1 = {u5 , u8, u10 , u12 }. For i ∈ {5, 8, 10, 12} we compute δi , the minimum integer such
that ui xδ4i ∈ in (I) and get that δ4 = δ10 = δ12 = 1 and δ8 = 2. If we set ri the remainder of the
division of ui xδ4i for all i ∈ {4, 8, 10, 12}, we get that
• r4 = −x4 x25 b1 + 2x4 x5 b4 + x5 b6 + x5 b7 ,
NOETHER RESOLUTIONS IN DIMENSION 2
9
• r8 = x24 x5 b3 + x5 b11 ,
• r10 = x24 x5 b2 + 3x4 x5 b8 + (x25 − x4 x5 )b9 , and
• r12 = x24 x25 b1 + x4 x5 b6 + x25 b7 .
Hence, we obtain the following minimal graded free resolution of R/I
ψ
F :0−
→ A(−3) ⊕ A3 (−4) −
→ A ⊕ A3 (−1)A5 (−2) ⊕ A3 (−3) −
→ R/I −
→ 0,
where ψ is given by the matrix
x4 x25
0
0
−x24 x25
0
0
−x24 x5
0
0
0
0
0
−2x4 x5 −x24 x5
0
0
x4
0
0
0
−x5
0
0
−x4 x5
−x5
0
0
−x25
0
x24
−3x4 x5
0
2
0
0
x4 x5 − x5
0
0
0
x4
0
0
−x5
0
0
0
0
0
x4
Moreover, the Hilbert series of R/I is
HSR/I (t) =
and reg(R/I) = max{3, 4 − 1} = 3.
3. N OETHER
1 + 3t + 5t2 + 2t3 − 3t4
.
(1 − t)2
RESOLUTION .
S IMPLICIAL
SEMIGROUP RINGS
This section concerns the study of Noether resolutions in simplicial semigroup rings R/I, i.e.,
whenever I = IA with A = {a1 , . . . , an } ⊂ Nd and an−d+i = wn−d+i ei for all i ∈ {1, . . . , d},
where {e1 , . . . , ed } is the canonical basis of Nd . In this setting, R/IA is isomorphic to the semigroup ring K[S], where S is the simplicial semigroup generated by A. When K is infinite, IA
is the vanishing ideal of the variety given parametrically by xi := tai for all i ∈ {1, . . . , n} (see,
e.g., [Villarreal (2015)]) and, hence, K[S] is the coordinate ring of a parametric variety. In this
section we study the multigraded Noether resolution of K[S] with respect to the multigrading
degS (xi ) = ai ∈ S; namely,
ψp
ψ1
ψ0
F : 0 −→ ⊕s∈Sp A · s −→ · · · −→ ⊕s∈S0 A · s −→ K[S] −→ 0.
where Si ⊂ S for all i ∈ {0, . . . , p}. We observe that this multigrading is a refinement of the
P
grading given by ω = (ω1 , . . . , ωn ) with ωi := dj=1 aij ∈ Z+ ; thus, IA is ω-homogeneous and
the results of the previous section also apply here.
Our objective is to provide a description of this resolution in terms of the semigroup S. We
completely achieve this goal when K[S] is Cohen-Macaulay (which includes the case d = 1) and
also when d = 2.
For any value of d ≥ 1, the first step of the resolution corresponds to a minimal set of generators
of K[S] as A-module and is given by the following well known result.
Proposition 5. Let K[S] be a simplicial semigroup ring. Then,
S0 = {s ∈ S | s − ai ∈
/ S for all i ∈ {n − d + 1, . . . , n}} .
Moreover, ψ0 : ⊕s∈S0 A · s −→ K[S] is the homomorphism of A-modules induced by es 7→ ts ,
where {es | s ∈ S0 } is the canonical basis of ⊕s∈S0 A · s.
10
I. BERMEJO, E. GARCÍA-LLORENTE, I. GARCÍA-MARCO, AND M. MORALES
Proposition 5 gives us the whole multigraded Noether resolution of K[S] when K[S] is CohenMacaulay.
In [Goto et al. (1976), Theorem 1] (see also [Stanley (1978), Theorem 6.4]), the authors provide
a characterization of the Cohen-Macaulay property of K[S]. In the following result we are proving
an equivalent result that characterizes this property in terms of the size of S0 . The proof shows
how to obtain certain elements of Ker(ψ0 ) and this idea will be later exploited to describe the
whole resolution when d = 2 and K[S] is not Cohen-Macaulay.
Q
d
ω
/[Zd : ZS],
Proposition 6. Let S be a simplicial semigroup as above. Set D :=
n−d+i
i=1
where [Zd : ZS] denotes the index of the group generated by S in Zd . Then, K[S] is CohenMacaulay ⇐⇒ |S0 | = D.
Proof. By Auslander-Buchsbaum formula we deduce that K[S] is Cohen-Macaulay if and only
if ψ0 is injective, where ψ0 is the morphism given in Proposition 5. We are proving that ψ0 is
injective if and only if |S0 | = D. We define an equivalence relation on Zd , u ∼ v ⇐⇒ u − v ∈
Z{ωn−d+1 e1 , . . . , ωn ed }. This relation partitions ZS into D = [ZS : Z{ωn−d+1 e1 , . . . , ωn ed }]
equivalence classes. Since
Zd /Z S ≃ Zd /Z{ωn−d+1 e1 , . . . , ωn ed } / (Z S/Z{ωn−d+1 e1 , . . . , ωn ed }) ,
Q
d
d
ω
we get that D =
i=1 n−d+i /[Z : ZS]. Moreover, the following two facts are easy to check:
for every equivalence class there exists an element b ∈ S0 , and S = S0 + N{ωn−d+1e1 , . . . , ωn ed }.
This proves that |S0 | ≥ D.
Assume that |S0 | > D, then there exist u, v ∈ S0 such that u ∼ v or, equivalently, u +
Pd
Pd
i=1 λi ωn−d+i ei = v +
i=1 µi ωn−d+i ei for some λi , µi ∈ N for all i ∈ {1, . . . , d}. Thus
1
1
· · · xλnd eu − xµn−d+1
· · · xµnd ev ∈ Ker(ψ0 ) and ψ0 is not injective.
xλn−d+1
Assume now that |S0 | = D, then for every s1 , s2 ∈ S0 , s1 6= s2 , we have that s1 6∼ s2 . As a
consequence, an element ρ ∈ ⊕s∈S0 A · s is homogeneous if and only if it is a monomial, i.e.,
ρ = cxα es for some c ∈ K, xα ∈ A and s ∈ S0 . Since the image by ψ0 of a monomial is
another monomial, then there are no homogeneous elements in Ker(ψ0 ) different from 0, so ψ0 is
injective.
From now on suppose that K[S] is a 2-dimensional non Cohen-Macaulay semigroup ring. In
this setting, we consider the set
∆ := {s ∈ S | s − an−1 , s − an ∈ S and s − an − an−1 ∈
/ S} .
The set ∆ or slight variants of it has been considered by other authors (see, e.g., [Goto et al. (1976),
Stanley (1978), Trung & Hoa (1986)]). We claim that ∆ has exactly |S0 | − D elements. Indeed,
if we consider the equivalence relation ∼ of Proposition 6, then ∼ partitions ZS in D classes
C1 , . . . , CD and it is straightforward to check that |∆ ∩ Ci | = |S0 ∩ Ci | − 1 for all i ∈ {1, . . . , D}.
From here, we easily deduce that |∆| = |S0 | − D. Hence, a direct consequence of Proposition 6
is that ∆ is nonempty because K[S] is not Cohen-Macaulay. Furthermore, as Theorem 2 shows,
the set ∆ is not only useful to characterize the Cohen-Macaulay property but also provides the set
of shifts in the second step of the multigraded Noether resolution of K[S].
Theorem 2. Let K[S] be a 2-dimensional semigroup ring and let
∆ = {s ∈ S | s − an−1 , s − an ∈ S and s − an − an−1 ∈
/ S} ,
as above. Then, S1 = ∆.
Proof. Set B0 the monomial basis of R/(in(IA ), xn−1 , xn ), where in (IA ) is the initial ideal of IA
with respect to >ω . For every u = xα1 1 · · · xαnn ∈ B1 we set δu ≥ 1 the minimum integer such that
u
u
uxδn−1
∈ in(IA ). Consider pu ∈ R the remainder of uxδn−1
modulo the reduced Gröbner basis of
NOETHER RESOLUTIONS IN DIMENSION 2
11
u
− pu ∈ IA . Since IA is a binomial ideal, we get that pu = xγ
IA with respect to >ω , then uxδn−1
for some (γ1, . . . , γn ) ∈ Nn . Moreover, the condition xα > xγ and the minimality of δu imply that
γ
γn > 0 and γn−1 = 0, so xγ = vu xnvu with vu ∈ B0 . As we proved in Proposition 4, if we denote
γ
u
by {ev | v ∈ B0 } the canonical basis of ⊕v∈B0 A(−degS (v)) and hu := xδn−1
eu − xnvu evu for all
u ∈ B1 , then Ker(ψ0 ) is the A-module minimally generated by C := {hu | u ∈ B1 }. Let us prove
that
{degS (hu ) | u ∈ B1 } = {s ∈ S | s − an−1 , s − an ∈ S and s − an−1 − an ∈
/ S}.
Take s = degS (hu ) for some u ∈ B1 , then s = degS (hu ) = degS (u)+δu an−1 = degS (v
Pu )+γvu an .
Since δu , γvu ≥ 1, we get that both s−an−1 , s−an ∈ S. Moreover, if s−an−1 −an = ni=1 δi ai ∈
δu −1
S, then xn−1
u − xλ xn+1 ∈ IA , which contradicts the minimality of δu .
Take now s ∈ S such that s−an−1 , s−an ∈ S and s−an−1 −an ∈
/ S. Since s−an−1 , s−an ∈ S,
′ ′′
′
there exists s , s ∈ S0 and γ1 , γ2 , λ1 , λ2 ∈ N such that s−an = s +γ1 an−1 +γ2 an and s−an+1 =
s′′ +λ1 an−1 +λ2 an . Observe that γ2 = 0, otherwise s−an−1 −an = s′ +γ1 an−1 +(γ2 −1)an ∈ S,
a contradiction. Analogously λ1 = 0. Take u, v ∈ B0 such that degS (u) = s′ and degS (v) = s′′ .
γ1
1
We claim that u ∈ J and that δu = γ1 . Indeed, f := uxn−1
− vxλn2 ∈ IA and in(f ) = uxγn−1
, so
′
u ∈ B1 . Moreover, if there exists γ < δu , then s − an−1 − an ∈ S, a contradiction.
One of the interests of Proposition 6 and Theorem 2 is that they describe multigraded Noether
resolutions of dimension 2 semigroup rings in terms of the semigroup S and, in particular, they
do not depend on the characteristic of the field K.
Now we consider the multigraded Hilbert Series of K[S], which is defined by
X
X
HSK[S](t) =
ts =
ts11 · · · tsdd ,
s∈S
s=(s1 ,...,sd )∈S
When d = 2, from the description of the multigraded Noether resolution of K[S] we derive an
expression of its multigraded Hilbert series in terms of S0 and S1 .
Corollary 3. Let K[S] be a dimension 2 semigroup ring. The multigraded Hilbert series of K[S]
is:
P
P
s
s
s∈S0 t −
s∈S1 t
HSK[S](t) =
.
ω
(1 − t1 n−1 )(1 − tω2 n )
Remark 1. When K[S] is a two dimensional semigroup ring and S is generated by the set A =
{a1 , . . . , an } ⊂ N2 , if we set ω = (ω1 , . . . , ωn ) ∈ Nn with ωi := ai,1 + ai,2 for all i ∈ {1, . . . , n},
then IA is ω-homogeneous, as observed at the beginning of this section. The Noether resolution of
K[S] with respect to this grading is easily obtained from the multigraded one. Indeed, it is given
by the following expression:
ψ1
ψ0
F : 0 −→ ⊕(b1 ,b2 )∈S1 A(−(b1 + b2 )) −→ ⊕(b1 ,b2 )∈S0 A(−(b1 + b2 )) −→ K[S] −→ 0.
In addition, the weighted Hilbert series of K[S] is obtained from the multigraded one by just
considering the transformation tα1 1 tα2 2 7→ tα1 +α2 .
When ω1 = · · · = ωn , then IA is a homogeneous ideal. In this setting, the Noether resolution
with respect to the standard grading is
ψ1
ψ0
F : 0 −→ ⊕(b1 ,b2 )∈S1 A(−(b1 + b2 )/ω1 ) −→ ⊕(b1 ,b2 )∈S0 A(−(b1 + b2 )/ω1 ) −→ K[S] −→ 0.
Thus, the Castelnuovo-Mumford regularity of K[S] is
b1 + b2
b1 + b2
| (b1 , b2 ) ∈ S0 ∪
− 1 | (b1 , b2 ) ∈ S1
.
(2)
reg (K[S]) = max
ω1
ω1
Moreover, the Hilbert series of K[S] is obtained from the multigraded Hilbert series by just considering the transformation tα1 1 tα2 2 7→ t(α1 +α2 )/ω1 .
12
I. BERMEJO, E. GARCÍA-LLORENTE, I. GARCÍA-MARCO, AND M. MORALES
4. M ACAULAYFICATION
OF SIMPLICIAL SEMIGROUP RINGS
Given K[S] a simplicial semigroup ring, the semigroup ring K[S ′ ] is the Macaulayfication of
K[S] if the three following conditions are satisfied:
(1) S ⊂ S ′ ,
(2) K[S ′ ] is Cohen-Macaulay, and
(3) the Krull dimension of K[S ′ \ S] is ≤ d − 2, where d is the Krull dimension of K[S].
The existence and uniqueness of a K[S ′ ] fulfilling the previous properties for simplicial semigroup
rings is guaranteed by [Morales (2007), Theorem 5]. In this section we describe explicitly the
Macaulayfication of any simplicial semigroup ring in terms of the set S0 . For this purpose we
consider the same equivalence relation in Zd as in proof of Proposition 6, namely, for s1 , s2 ∈ Zd
s1 ∼ s2 ⇐⇒ s1 − s2 ∈ Z{ωn−d+1 e1 , . . . , ωn ed }.
As we have seen, S0 ⊂ Zd is partitioned into D := ωn−d+1 · · · ωn /[Zd : ZS] equivalence classes
S 1 , . . . , S D . For every equivalence class S i we define a vector bi in the following way. We
take S i = {s1 , . . . , st }, where sj = (sj1 , . . . , sjd) ∈ Nd for all j ∈ {1, . . . , t} and define
bi = (bi1 , . . . , bid ) ∈ Nd as the vector whose k-th coordinate bik equals the minimum of the kth coordinates of s1 , . . . , st , this is, bik := min{s1k , . . . , stk }. We denote B := {b1 , . . . , bD }
and
(3)
S ′ := B + N{ωn−d+1 e1 , . . . , ωn ed }.
The objective of this section is to prove that K[S ′ ] is the Macaulayfication of K[S]. The main
issue in the proof is to show that dim(K[S ′ \ S]) ≤ d − 2. For this purpose we use a technique
developed in [Morales & Nhan (2003)] which consists of determining the dimension of a graded
ring by studying its Hilbert function. More precisely, for L an ω-homogeneous ideal, if we denote
by h(i) the Hilbert function of R/L, by [Morales (1985), Lemma 1.4], there exist some polynomials Q1 , ..., Qs ∈ Z[t] with s ∈ Z+ such that h(ls + i) = Qi (l) for all i ∈ {1, . . . , s} and l ∈ Z+
large enough. Moreover, in [Morales (2016)], the author proves the following.
Theorem 3. Let L be a ω-homogeneous
ideal and denote by h : N → N the Hilbert function
P
of R/L. If we set h0 (n) = ni=0 h(i), then there exist s polynomials f1 , ..., fs ∈ Z[t] such that
h0 (ls+i) = fi (l) for all i ∈ {1, . . . , s} and l ∈ Z+ large enough. Moreover, all these polynomials
f1 , . . . , fs have the same leading term c tdim(R/L) /(dim(R/L))! with c ∈ Z+ .
In the proof of Theorem 4, we relate the Hilbert function of K[S ′ \ S] with that of several
monomial ideals and use of the following technical lemma.
Lemma 1. Let M ⊂ K[y1 , . . . , yd] be a monomial ideal. If for all i ∈ {1, . . . , d} there exist
xα ∈ M such that xi ∤ xα , then dim(K[y1 , . . . , yd ]/M) ≤ d − 2.
Proof. Let us prove that M has height ≥ 2. By contradiction, assume that M has an associated
prime P of height one. Since M is monomial,
then so is P. Therefore, P = (xi ) for some
√
i ∈ {1 . . . , d}. Hence we get that M ⊂ M ⊂ P = (xi ), a contradiction.
Now we can proceed with the proof of the main result of this section.
Theorem 4. Let K[S] be a simplicial semigroup ring and let S ′ be the semigroup described in
(3). Then, K[S ′ ] is the Macaulayfication of K[S].
Proof. Is is clear that S ⊂ S ′ . In order to obtain the result it suffices to prove that S ′ is a semigroup, that K[S ′ ] is Cohen-Macaulay and that dim(K[S ′ \ S]) ≤ dim(K[S]) − 2 (see, e.g.,
[Morales (2007)]).
Let us first prove that S ′ is a semigroup. Take s1 , s2 ∈ S ′ , then there exists i, j ∈ {1, . . . , D}
such that s1 = bi + c1 and s2 = bj + c2 for some c1 , c2 ∈ N{ωn−d+1 e1 , . . . , ωn ed }. Then s1 + s2 =
bi + bj + c1 + c2 . We take k ∈ {1, . . . , D} such that bk ∼ bi + bj . By construction of B we have
that bk = bi + bj + c3 for some c3 ∈ {ωn−d+1 e1 , . . . , ωn ed } and, hence, s1 + s2 ∈ S ′ .
NOETHER RESOLUTIONS IN DIMENSION 2
13
To prove that S ′ is Cohen-Macaulay it suffices to observe that B = {b ∈ S ′ | b − ai ∈
/ S ′ for all
i ∈ {1, . . . , d}} and that |B| = D, so by Proposition 6 it follows that S ′ is Cohen-Macaulay.
Let us prove that dim (K[S ′ \S]) ≤ d−2. For all s = (s1 , . . . , sm ) ∈ Nm we consider the gradP
′
b
of K[S], K[S ′ ] and K[S ′ \
ing deg(ts ) = m
i=1 si and we denote h, h and h the Hilbert functions
P
PD
′
S] respectively, then b
h = h′ − h. Moreover, we have that h′ = D
i=1 hi and h =
i=1 hi where
′
′
s
s
hi (d) := |{s ∈ S | deg t = d and s ∼ bi }| and hi (d) := |{s ∈ S | deg t = d and s ∼ bi }|. For
each i ∈ {1, . . . , D} we define a monomial ideal Mi ⊂ k[y1 , . . . , yd ] as follows: for every b ∈ S
P
such that b ∼ bi we define the monomial mb := y1β1 · · · ydβd if b = bi + di=1 βi ωn−d+i ei and
Mi := ({mb | b ∈ S, b ∼ bi }). We consider in K[y1 , . . . , yd ] the grading degω (yi ) = ωn−d+i and
denote by hωi the corresponding ω-homogeneous Hilbert function of K[y1 , . . . , yd ]/Mi . We have
P
P
the following equality hωi (λ) = h′i ( dj=1 bij + λ) − hi ( dj=1 bij + λ) because y β ∈
/ Mi ⇐⇒
Pd
′
bi + i=1 βi ωn−d+i ei ∈ S \ S. Hence, we have expressed the Hilbert function b
h of K[S \ S ′ ] as
a sum of D Hilbert functions of K[y1 , . . . , yd]/Mi , for some monomial ideals M1 , . . . , MD and,
by Lemma 1, dim(K[y1 , . . . , yd ]/Mi ) ≤ d − 2. Thus, by Theorem 3, we can conclude that the
dimension of K[S ′ \ S] equals the maximum of dim(K[y1 , . . . , yd ]/Mi ) ≤ d − 2 and we get the
result.
We finish this section with an example showing how to compute the Macaulayfication by means
of the set S0 . Moreover, this example illustrates that even if K[S] = R/IA with IA a homogeneous
ideal, it might happen that the ideal associated to K[S ′ ] is not standard homogeneous.
Example 3. We consider the semigroup ring K[S], where S ⊂ N2 is the semigroup generated by
A := {(1, 9), (4, 6), (5, 5), (10, 0), (0, 10)} ⊂ N2 . Then, K[S] = R/IA and IA is homogeneous.
If we compute the set S0 we get that
S0 = {(0, 0), (1, 9), (2, 18), (3, 27), (13, 17), (4, 6), (5, 5), (6, 14), (7, 23), (8, 12), (9, 11)} .
Moreover we compute D = 100/[Zd : ZS] = 10 and get S 1 = {(0, 0)}, S 2 = {(1, 9)},
S 3 = {(2, 18)}, S 4 = {(3, 27), (13, 17)}, S 5 = {(4, 6)}, S 6 = {(5, 5)}, S 7 = {(6, 14)},
S 8 = {(7, 23)}, S 9 = {(8, 12)} and S 10 = {(9, 11)}. So, the Macaulayfication K[S ′ ] of K[S] is
given by S ′ = B + N{(10, 0), (0, 10)}, where
B = {(0, 0), (1, 9), (2, 18), (3, 17), (4, 6), (5, 5), (6, 14), (7, 23), (8, 12), (9, 11)}.
Or equivalently, S ′ is the semigroup generated by
A′ = {(1, 9), (3, 17), (4, 6), (5, 5), (10, 0), (0, 10)}.
We observe that K[S ′ ] ≃ K[x1 , . . . , x6 ]/IA′ and that IA′ is ω-homogeneous with respect to ω =
(1, 2, 1, 1, 1, 1) but not standard homogeneous.
5. A N
UPPER BOUND FOR THE
C ASTELNUOVO -M UMFORD
REGULARITY OF PROJECTIVE
MONOMIAL CURVES
Every sequence m1 < . . . < mn of relatively prime positive integers with n ≥ 2 has associated
the projective monomial curve C ⊂ PnK given parametrically by xi := smi tmn −mi for all i ∈
{1, . . . , n − 1}, xn = smn , xn+1 := tmn . If we set A := {a1 , . . . , an+1 } ⊂ N2 where ai :=
(mi , mn − mi ), an := (mn , 0) and an+1 := (0, mn ), it turns out that IA ⊂ K[x1 , . . . , xn+1 ] is
the defining ideal of C. If we denote by S the semigroup generated by A, then the 2-dimensional
semigroup ring K[S] is isomorphic to K[x1 , . . . , xn+1 ]/IA , the homogeneous coordinate ring of
C. Hence, the methods of the previous sections apply here to describe its multigraded Noether
resolution, and the formula (2) in Remark 1 for the Castelnuovo-Mumford regularity holds in this
context (with ω1 = mn ). The goal of this section is to use this formula to prove Theorem 5,
which provides an upper bound for the Castelnuovo-Mumford regularity of K[S]. The proof we
are presenting is elementary and uses some classical results on numerical semigroups. We will
14
I. BERMEJO, E. GARCÍA-LLORENTE, I. GARCÍA-MARCO, AND M. MORALES
introduce now the results on numerical semigroups that we need for our proof (for more on this
topic we refer to [Rosales & García-Sánchez (2009)] and [Ramírez Alfonsín (2005)]).
Given m1 , . . . , mn a set of relatively prime integers, we denote by R the numerical subsemigroup of N spanned by m1 , . . . , mn . The largest integer that does not belong to R is called the
Frobenius number of R and will be denoted by g(R). We consider the Apery set of R with respect
to mn , i.e., the set
Ap(R, mn ) := {a ∈ R | a − mn ∈
/ R}.
It is a well known and easy to check that Ap(R, mn ) constitutes a full set of residues modulo
mn (and, in particular, has mn elements) and that max(Ap(R, mn )) = g(R) + mn . We will also
use an upper bound on g(R) which is a slight variant of the one given in [Selmer (1977)] (which
was deduced from a result of [Erdös & Graham (1972)]). The reason why we do not use Selmer’s
bound itself is that it is only valid under the additional hypothesis that n ≤ m1 . This is not a
restrictive hypothesis when studying numerical semigroups, because whenever m1 < · · · < mn
is a minimal set of generators of R, then n ≤ m1 . In our current setting of projective monomial
curves, the case where m1 < · · · < mn is not a minimal set of generators of R is interesting by
itself (even the case m1 = 1 is interesting); hence, a direct adaptation of the proof of Selmer yields
that
jm k
τ
(4)
g(R) ≤ 2mn
− mτ ,
n
for every mτ ≥ n. Note that mn ≥ n and then, such a value τ always exists..
We first include a result providing an upper bound for reg(K[S]) when K[S] is Cohen-Macaulay.
Proposition 7. Let m1 < . . . < mn be a sequence of relatively prime positive integers with n ≥ 2
and let τ ∈ {1, . . . , n} such that mτ ≥ n. If K[S] is Cohen-Macaulay, then
k
j
jm k
τ
− mτ + mn )/m1 .
reg(K[S]) ≤ (2mn
n
j
k
In particular, if m1 ≥ n, we have that reg(K[S]) ≤ mn n2 + m11 − 1 .
Proof. We consider the equivalence relation ∼ of Section 4. Indeed, since now ZS = {(x, y) | x +
y ≡ 0 (mod mn )}, then we have that ∼ partitions the set S0 in exactly mn equivalence classes.
Moreover, since K[S] is Cohen-Macaulay, we have that
• each of these classes has a unique element,
• S1 = ∅, and
n
o
b1 +b2
• reg(K[S]) = max mn | (b1 , b2 ) ∈ S0 (see Remark 1).
Pn−1
Pn−1
αi . Moreover,
αi ai and (b1 + b2 )/mn = i=1
Let us take (b1 , b2 ) ∈ S0 , then (b1 , b2 ) = i=1
we claim that b1 ∈ Ap(R, mn ). Otherwise, b1 − mn ∈ R and there would be another element
(c1 , c2 ) ∈ S0 such that (c1 , c2 ) ∼ (b1 , b2 ), a contradiction. Hence, by (4),
!
n−1
n−1
jm k
X
X
τ
− mτ + mn .
αi m1 ≤
αi mi = b1 ≤ g(R) + mn ≤ 2mn
n
i=1
i=1
And, from this expression we conclude that
jm k
b1 + b2 X
τ
− mτ + mn )/m1 .
=
αi ≤ (2mn
mn
n
i=1
n−1
When m1 ≥ n, then it suffices to take τ = 1 to get the result.
And now, we can prove the main result of the section.
NOETHER RESOLUTIONS IN DIMENSION 2
15
Theorem 5. Let m1 < . . . < mn be a sequence of relatively prime positive integers with n ≥ 2.
If we take τ, λ such that mτ ≥ n and mn − mλ ≥ n. Then,
$
%
λ
+
m
2mn mn −m
2mn mnτ − mτ + mn
λ
n
− 2.
+
reg(K[S]) ≤
m1
(mn − mn−1 )
k
j
1
−4.
In particular, if m1 ≥ n and mn −mn−1 ≥ n, then reg(K[S]) ≤ mn n4 + m11 + mn −m
n−1
Proof. We consider E one of the equivalence classes of ZS induced by the equivalence relation ∼.
First, assume that S0 ∩ E has a unique element which we call (b1 , b2 ). Then,
m S1 ∩ E = ∅, and the
b1 +b2
same argument as in the proof of Proposition 7 proves that mn ≤ (2mn nτ − mτ + mn )/m1 .
Assume now that S0 ∩ E = {(x1 , y1 ), . . . , (xr , yr )} with r ≥ 2 and x1 < x2 < · · · < xr . We
claim that the following properties hold:
(a) x1 ≡ x2 ≡ · · · ≡ xr (mod mn ),
(b) y1 > · · · > yr and y1 ≡ y2 ≡ · · · ≡ yr (mod mn ),
(c) x1 ∈ Ap(R, mn ),
(d) yr ∈ Ap(R′ , mn ), where R′ is the numerical semigroup generated by mn − mn−1 <
mn − mn−2 < · · · < mn − m1 < mn ,
(e) S1 ∩ n
E = {(x2 , y1), (x3 , y2 ), . .o. , (xr , yr−1
n)}, and
o
b1 +b2
(f) max mn | (b1 , b2 ) ∈ S0 ∩ E ≤ max b1m+bn 2 | (b1 , b2 ) ∈ S1 ∩ E − 1.
Properties (a) and (b) are evident. To prove (c) and (d) it suffices to take into account the following
facts: S ⊂ R×R′ , and for every b1 ∈ R, b2 ∈ R′ there exist c1 , c2 ∈ N such that (b1 , c2 ), (c1 , b2 ) ∈
S. To prove (e) we first observe that
S ∩ E = {b + λ(mn , 0) + µ(0, mn ) |b ∈ S0 ∩ E, λ, µ ∈ N}.
Take now (x, y) ∈ S1 ∩ E and we take the minimum value i ∈ {1, . . . , r} such that (x, y) =
(xi , yi ) + λ(mn , 0) + µ(0, mn ) with λ, µ ∈ N; we observe that
• λ > 0; otherwise (x, y) − (mn , 0) ∈
/ S,
• µ = 0; otherwise (x, y) − (mn , mn ) = (xi , yi ) + (λ − 1)(mn , 0) + (µ − 1)(0, mn ) ∈ S, a
contradiction,
• y ≥ yr−1 ; otherwise i = r and, since (x, y) − (0, mn ) ∈ S ∩ E, we get that µ ≥ 1,
• x ≤ xi+1 ; otherwise (x, y) = (xi+1 , yi+1 ) + λ′ (mn , 0) + µ′ (0, mn ) with λ′ , µ′ ≥ 1, a
contradiction, and
• x ≥ xi+1 ; otherwise (x, y) − (0, mn ) ∈
/ S.
Hence, (x, y) = (xi+1 , yi ) and S1 ∩ E ⊆ {(x2 , y1 ), (x3 , y2 ), . . . , (xr , yr−1)}. Take now i ∈
{1, . . . , r − 1}, and consider (xi+1 , yi ) ∈ S. Since (xi , yi ), (xi+1 , yi+1) ∈ E, xi ≡ xi+1 (mod mn )
and yi ≡ yi+1 (mod mn ), then (xi+1 , yi ) ∈ E. We also have that there exist γ, δ ∈ N such that
(xi+1 , yi )−(mn , 0) = (xi , yi )+γ(mn , 0) ∈ S and (xi+1 , yi )−(0, mn ) = (xi+1 , yi+1)+δ(0, mn ) ∈
S. We claim that (xi+1 , yi ) − (mn , mn ) ∈
/ S. Otherwise there exists j ∈ {1, . . . , r} such that
(xi+1 −mn , yi −mn ) = (xj , yj )+λ′(mn , 0)+µ′(0, mn ); this is not possible since xi+1 −mn < xi+1
implies that j ≤ i, and yi − mn < yi implies that j ≥ i + 1. Thus, (xi+1 , yi ) ∈ S1 and (e) is
proved. Property (f) follows from (e).
Moreover, since x1 ∈ Ap(R, mn ), the same argument as in Proposition 7 proves that
jm k
x1 + y1
τ
(5)
≤ 2mn
− mτ + mn /m1 ,
mn
n
and a similar argument with yr ∈ Ap(R′ , mn ) proves that
mn − mλ
xr + yr
+ mλ /(mn − mn−1 ).
≤ 2mn
(6)
mn
n
16
I. BERMEJO, E. GARCÍA-LLORENTE, I. GARCÍA-MARCO, AND M. MORALES
And, since,
xi+1 + yi
xr + y1
x1 + y1 xr + yr
−1≤
−1≤
+
− 2,
mn
mn
mn
mn
(7)
putting together (5), (6) and (7) we get the result. If m1 ≥ n and mn − mn−1 ≥ n, it suffices to
take τ = 1 and λ = n − 1 to prove the result.
It is not difficult to build examples such that the bound provided by Theorem 5 outperforms the
bound of L’vovsky’s. Let us see an example.
Example 4. Set n ≥ 6 and consider mi = n + i for all i ∈ {1, . . . , n − 1} and mn = 3n, then we
can take τ = 1 and λ = n − 1 and apply Theorem 5 to prove that
1
1
4
− 4 = 13,
+
+
reg(K[S]) ≤ 3n
n n+1 n+1
meanwhile the result of L’vovsky provides an upper bound of 2n + 1.
6. N OETHER
RESOLUTION AND M ACAULAYFICATION OF PROJECTIVE MONOMIAL CURVES
ASSOCIATED TO ARITHMETIC SEQUENCES AND THEIR CANONICAL PROJECTIONS .
Every sequence m1 < . . . < mn of positive integers with n ≥ 2 has associated the projective
monomial curve C ⊂ PnK given parametrically by xi := smi tmn −mi for all i ∈ {1, . . . , n−1}, xn =
smn , xn+1 := tmn . If we set A := {a1 , . . . , an+1 } ⊂ N2 where ai := (mi , mn −mi ), an := (mn , 0)
and an+1 := (0, mn ), it turns out that IA ⊂ K[x1 , . . . , xn+1 ] is the defining ideal of C. Moreover,
if we denote by S the semigroup generated by A, then K[S] ≃ K[x1 , . . . , xn+1 ]/IA is a dimension
2 semigroup ring and the methods of the previous sections apply here to describe its multigraded
Noether resolution.
In [Li et. al (2012)], the authors studied the set S0 whenever m1 < · · · < mn is an arithmetic
sequence of relatively prime integers, i.e., there exist d, m1 ∈ Z+ such that mi = m1 + (i − 1) d
for all i ∈ {1, . . . , n} and gcd{m1 , d} = 1. In particular, they obtained the following result.
Theorem
6.
[Li
et. al (2012), Theorem 3.4]
j
S0 =
mn − jd, jd | j ∈ {0, . . . , mn − 1}
n−1
From the previous result and Proposition 6 we deduce that K[S] is Cohen-Macaulay (see also
[Bermejo et al. (2017), Corollary 2.3]), we obtain the shifts of the only step of the multigraded
Noether resolution and, by Corollary 2, we also derive that reg(K[S]) = ⌈(mn − 1)/(n − 1)⌉ (see
also [Bermejo et al. (2017), Theorem 2.7]). In the rest of this section we are using the tools developed in the previous sections to study the canonical projections of C, i.e., for all r ∈ {1, . . . , n−1}
and n ≥ 3 we aim at studying the curve Cr := πr (C) obtained as the image of C under the projecn−1
tion πr from PnK to PK
defined by (p1 : · · · : pn+1 ) 7→ (p1 : · · · : pr−1 : pr+1 : · · · : pn+1 ). We
know that the vanishing ideal of Cr is IAr , where Ar = A \ {ar } for all r ∈ {1, . . . , n − 1}. Note
that C1 is the projective monomial curve associated to the arithmetic sequence m2 < · · · < mn
and, thus, its Noether resolution can also be obtained by means of Theorem 6. Also when n = 3,
C2 is the curve associated to the arithmetic sequence m1 < m3 . For this reason, the rest of this
section only concerns the study of the multigraded Noether resolution of Cr for r ∈ {2, . . . , n − 1}
and n ≥ 4.
Remark 2. Denote by Cn and Cn+1 the Zariski closure of πn (C) and πn+1 (C) respectively. Then,
both Cn and Cn+1 are projective monomial curves associated to arithmetic sequences and, thus,
their Noether resolutions can also be obtained by means of Theorem 6. More precisely, the corresponding arithmetic sequences are m1 < · · · < mn−1 for Cn and 1 < 2 < · · · < n − 1 for Cn+1 ,
i.e., Cn+1 is the rational normal curve of degree n − 1.
NOETHER RESOLUTIONS IN DIMENSION 2
17
We denote by Pr the semigroup generated by Ar for r ∈ {2, . . . , n − 1} and n ≥ 4. Proposition
8 shows how to get the semigroups Pr from S. In the proof of this result we will use the following
two lemmas, both of them can be directly deduced from [Bermejo et al. (2017), Lemma 2.1].
Lemma 2. Set q := ⌊(m1 − 1)/(n − 1)⌋ ∈ N; then,
P
(a) q + d + 1 = min{b ∈ Z+ | bm1 ∈ ni=2 Nmi }
Pn−1
Nmi }
(b) q + 1 = min{b ∈ Z+ | bmn ∈ i=1
(c) (q + d)a1 + ai = al+i + qan + dan+1 for all i ∈ {1, . . . , n − l}, where l := m1 − q(n − 1) ∈
{1, . . . , n − 1}.
P
Lemma 3. For all r ∈ {2, . . . , n − 1}, we have that mr ∈
i∈{1,...,n}\{r} Nmi if and only if
r > m1 .
Proposition 8. Set q := ⌊(m1 − 1)/(n − 1)⌋ and l := m1 − q(n − 1). If r ≤ m1 , then
(a.1) for r = 2,
{µa1 + a2 + λ an+1 | 0 ≤ µ ≤ q + d − 1, λ ∈ N} , if l 6= n − 1,
S \ P2 =
{µa1 + a2 + λ an+1 | 0 ≤ µ ≤ q + d, λ ∈ N} ,
if l = n − 1,
(a.2) for r ∈ {3, . . . , n − 2}, S \ Pr = {ar + λ an+1 | λ ∈ N} , and
(a.3) for r = n −1,
{an−1 + µan + λ an+1 | 0 ≤ µ ≤ q − 1 or 0 ≤ λ ≤ d − 1} , if l 6= n − 1,
S\Pn−1 =
{an−1 + µan + λ an+1 | 0 ≤ µ ≤ q or 0 ≤ λ ≤ d − 1} ,
if l = n − 1.
If r > m1 , then
(b.1) for r = 2, S \ P2 = {µa1 + a2 + λan+1 | 0 ≤ µ, λ ≤ d − 1} ,
(b.2) for r ∈ {3, . . . , n − 2}, S \ Pr = {ar + λan+1 | 0 ≤ λ ≤ d − 1}, and
(b.3) for r = n − 1, S \ Pn−1 = {an−1 + µan + λan+1 | µ ∈ N, 0 ≤ λ ≤ d − 1}.
Proof. We express every s ∈ S as s = α1 a1 + ǫi ai + αn an + αn+1 an+1 , with α1 , αn , αn+1 ∈ N,
i ∈ {2, . . . , n − 1} and ǫi ∈ {0, 1}. Whenever ǫi = 0 or i 6= r, it is clear that s ∈ Pr . Hence,
we assume that s = α1 a1 + ar + αn an + αn+1 an+1 and the idea of the proof is to characterize the
values of α1 , αn , αn+1 so that s ∈ Pr in each case.
Assume first that r ∈ {3, . . . , n − 2} and let us prove (a.2) and (b.2). If α1 > 0 or αn > 0, the
equalities a1 + ar = a2 + ar−1 and ar + an = ar+1 + an−1 yield that s ∈ Pr , so it suffices to
consider when s = ar + αn+1 an+1 . If r ≤ m1 , then by Lemma 3 we get that s ∈
/ Pr because the
first coordinate of s is precisely mr . This proves (a.2). If r > m1 and αn+1 ≥ d, then the equality
ar + dan+1 = da1 + ar−m1 yields that s ∈ Pr . However, if αn+1 < d we are proving that s ∈
/ Pr .
Suppose by contradiction that s ∈ Pr and αn+1 < d, then
X
(8)
s = ar + αn+1 an+1 =
βj aj
P
j∈{1,...,n+1}\{r}
for some βj ∈ N, then d ≥ 1 + αn+1 = j∈{1,...,n+1}\{r} βj . Moreover, observing the first coordiP
P
nates in (8) we get that mr = j∈{1,...,n}\{r} βj mj . Hence, m1 +(r −1)d = j{1,...,n}\{r} βj (m1 +
P
(j − 1)d) and, since gcd{m1 , d} = 1, this implies that d divides ( j{1,...,n}\{r} βj ) − 1, but
P
0 < ( j∈{1,...,n}\{r} βj ) − 1 < d, a contradiction. Thus (b.2) is proved.
Since the proof of (a.1) is similar to the proof of (a.3) we are not including it here. So let
us prove (b.1). Assume that r = 2. If αn > 0 the equality a2 + an = a3 + an−1 yields that
s ∈ P2 , so it suffices to consider when s = α1 a1 + ar + αn+1 an+1 . If α1 ≥ d, then the identity
da1 +a2 = a3 +dan+1 yields that s ∈ P2 . For α1 < d, if αn+1 ≥ d, the equality α1 a1 +a2 +dan+1 =
(α1 + d + 1)a1 also yields that s ∈ P2 . Thus, to conclude (b.1) it only remains
to proof that
P
s∈
/ P2 when α1 , αn+1 < d. Indeed, assume that α1 a1 + a2 + αn+1 an+1 = j∈{1,3,...,n+1} βj aj .
P
Observing the first coordinate of the equality we get that α1 + m2 =
j∈{1,3,...,n} βj mj , but
α1 + m2 < m3 < · · · < mn , so β3 = · · · = βn+1 = 0. But this implies that β1 = α1 + d + 1 and,
hence, βn+1 < 0, a contradiction.
18
I. BERMEJO, E. GARCÍA-LLORENTE, I. GARCÍA-MARCO, AND M. MORALES
Assume now that r = n−1. If α1 > 0, the equality a1 + an−1 = a2 + an−2 yields that s ∈ Pn−1 ,
so it suffices to consider
when s = an−1 + αn an + αn+1 an+1 . Whenever s ∈ Pn−1 , then s can be
P
expressed as s = j∈{1,...,n−2,n,n+1} βj aj , if we consider both expressions of s, we get that
P
(i) j∈{1,...,n−2,n,n+1} βi = 1 + αn + αn+1 , and
P
(ii) j∈{1,...,n−2,n} βj mj = mn−1 + αn mn .
If αn+1 < d we are proving that s ∈
/ Pn−1 . Assume by contradiction that s ∈ Pn−1 . From (ii) and
Lemma 2 we deduce that βn < αn . Moreover, if we expand (ii) considering
P that mi = m1 +(i−1)d
for all i ∈ {1, . . . , n} and that gcd{m1 , d} = 1, we get that d divides j∈{1,...,n−2,n} βj −αn −1 =
αn+1 − βn+1 , a contradiction to 0 < αn+1 − βn+1 < d.
Case 1: m1 ≥ n − 1. Assume that s ∈ Pn−1 . By (ii) and Lemma 3 we have that βn < αn , so there
exists j0 ∈ {1, . . . , n − 2} such that βj0 >
βn mn in both sides
P0. As a consequence, if we add d − P
of (ii) we get that (αn + 1 − βn )mn = j∈{1...,n−2} βj mj − mj0 + mj0 +1 ∈ j∈{1,...,n−1} Nmj .
Hence, by Lemma 2 we have that αn ≥ αn − βn ≥ q. If l < n − 1, for αn ≥ q, αn+1 ≥ d
the equality of Lemma 2 (q + d)a1 + an−l−1 = an−1 + qan + dan+1 shows that s ∈ Pn−1 . This
proves (a.3) whenever l ≤ n − 1. If l = n − 1, for αn ≥ q + 1, αn ≥ d, again the equality
(q + d + 1)a1 = (q + 1)an + dan+1 shows that s ∈ Pn−1 . It only remainsP
to prove that if αn = q;
then s ∈
/ P2 . Assume by contradiction that an−1 + qan + αn+1 an+1 = j∈{1,...,n−2,n,n+1} βj aj .
P
Then, the first coordinates of this equality yield that mn−1 + qmn =
j∈{1,...,n−2,n} βj mj and
we deduce by Lemma 3 that βn < q and, hence, there exists j0 ∈ {1, . . . , n − 2} such that
βj0 > 0. We denote βn−1 := 0, λj := βj for all j ∈ {1, . . . , n} \ {j0 , j0 − 1}, λj0 = βj0 − 1,
λj0 +1 = βj0 +1 + 1, then adding d in
and using Lemma 2, we get that
Pboth sides of the equality
Pn
(q + 1)mn = (q + d + 1)m1 =
Nm
λ
m
∈
i . However, λ1 6= q + d + 1,
i=1
j∈{1,...,n} j j
λn 6= q + 1, P
so applying iteratively the equalities ai + aj = ai−1 + aj+1 for all 2 ≤ i ≤ j ≤ n − 1
we express j∈{1,...,n} λj mj as µ1 m1 + ǫk mk + µn mn with µ1 , µm ∈ N, k ∈ {2, . . . , n − 1},
ǫk ∈ {0, 1}. It is clear that µ1 6= q + d + 1 and that µn 6= q + 1 and one of those is nonzero, so
this contradicts the minimality of q + d + 1 or q + 1.
To prove (b.3) it only remains to prove that if αn+1 ≥ d, then s ∈ Pn−1 , but this easily follows
from the relation an−1 + dan+1 = da1 + an−1−m1 .
From the previous result and Proposition 5 it is not difficult to obtain the following corollary,
which provides the shifts of the first step of a multigraded Noether resolution of K[Pr ] for all
r ∈ {2, . . . , n − 1}, namely (Pr )0 := {s ∈ Pr | s − an , s − an+1 ∈
/ Pr }. Indeed, Corollary 4
describes (Pr )0 from the set S0 given by Theorem 6.
Corollary 4. We denote tµ := µa1 + a2 for all µ ∈ N. If r ≤ m1 , then
(a.1) for r = 2,
(S0 \ {tµ | 0 ≤ µ ≤ q + d − 1}) ∪ {tµ + an | 0 ≤ µ ≤ q + d − 1} , if l 6= n − 1,
(P2 )0 =
(S0 \ {tµ | 0 ≤ µ ≤ q + d}) ∪ {tµ + an | 0 ≤ µ ≤ q + d} ,
if l = n − 1,
(a.2) for r ∈ {3, . . . , n − 2}, (Pr )0 = (S0 \ {ar }) ∪ {ar + an },
(a.3) for r = n −1,
(S0 \ {an−1 }) ∪ {an−1 + qan + dan+1 },
if l 6= n − 1,
(Pn−1 )0 =
(S0 \ {an−1 }) ∪ {an−1 + (q + 1)an + dan+1 }, if l = n − 1.
If r > m1 , then
(b.1) for r = 2, (P2 )0 = (S0 \ {tµ | 0 ≤ µ ≤ d − 1}) ∪ {tµ + an , tµ + dan+1 | 0 ≤ µ ≤ d − 1},
(b.2) for r ∈ {3, . . . , n − 2}, (Pr )0 = (S0 \ {ar }) ∪ {ar + an , ar + dan+1 } , and
(b.3) for r = n − 1, (Pn−1 )0 = (S0 \ {an−1 }) ∪ {an−1 + dan+1 }.
From Corollary 4 and Proposition 6, we get the following characterization of the CohenMacaulay property for this family of semigroup rings taking into account that D in Proposition 6
equals mn in these cases.
NOETHER RESOLUTIONS IN DIMENSION 2
19
Corollary 5. K[Pr ] is Cohen-Macaulay ⇐⇒ r ≤ m1 or r = n − 1.
Moreover, as a consequence of Theorem 4 and Corollary 4, we get the following result.
Corollary 6. For all r ∈ {2, . . . , n − 2} and r > m1 , the Macaulayfication of K[Pr ] is K[S].
In order to get the whole multigraded Noether resolution of K[Pr ] for all r ∈ {2, . . . , n − 2}
and r > m1 , it remains to study its second step. By Theorem 2, its shifts are given by the set
(Pr )1 := {s ∈ Pr | s − an , s − an+1 ∈ Pr and s − an − an+1 ∈
/ Pr }.
Corollary 7.
{µa1 + a2 + an + dan+1 |µ ∈ {0, . . . , d − 1}} , if r = 2 and m1 = 1,
(Pr )1 =
{ar + an + dan+1 },
if r ∈ {3, . . . , n − 2} and m1 < r.
As a consequence of the above results, we are able to provide the multigraded Noether resolution of K[Pr ] for all r ∈ {2, . . . , n − 1}.
Theorem 7.
q := ⌈(m1 − 1)/(n − 1)⌉ and l := m1 − q(n − 1). If we
q, l ∈ N be the integers
Let
λ
mn − λd, λd ∈ N2 for all λ ∈ {0, . . . , mn − 1}, then the multigraded Noether
set sλ := n−1
resolution of K[Pr ] is given by the following expressions:
• For m1 ≥ 2, then
n −1
0−
→ ⊕m
A
·
s
⊕
⊕
A
·
(s
+
a
)
→ K[P2 ] −
−
→ 0,
λ
λ∈Λ1
λ
n
λ=0, λ∈Λ
/ 1
where Λ1 := {µ(n − 1) − 1 | 1 ≤ µ ≤ q + d + ǫ}, and ǫ = 1 if l = n − 1, or ǫ = 0
otherwise.
• For r ∈ {3, . . . , n − 2} and r ≤ m1 , then
n −1
0−
→ ⊕m
→ K[Pr ] −
→0
λ=0, λ6=n−r A · sλ ⊕ A · (ar + an ) −
• For r = n − 1 ≤ m1 , then
n −1
0−
→ ⊕m
→ K[Pn−1 ] −
→ 0,
λ=0, λ6=1 A · sλ ⊕ A · (an−1 + (q + ǫ)an + dan+1 ) −
where ǫ = 1 if l = n − 1, or ǫ = 0 otherwise.
• For m1 = 1, then
n −1
⊕m
λ=0,λ∈Λ
/ 2 A · sλ
⊕
− K[P2 ] −
→ 0,
0−
→ ⊕λ∈Λ2 A · (sλ + an + dan+1 ) →
−
⊕λ∈Λ2 A · (sλ + an ) →
⊕
⊕λ∈Λ2 A · (sλ + dan+1 )
where Λ2 := {µ(n − 1) − 1 | 1 ≤ µ ≤ d}.
• For r ∈ {3, . . . , n − 2} and r > m1 , then
n −1
⊕m
λ=0, λ6=n−r A · sλ
0−
→ A · (ar + an + dan+1 ) →
−
− K[Pr ] −
→
→ 0.
⊕
A · (ar + an ) ⊕ A · (ar + dan+1 )
• For r = n − 1 > m1 , then
n −1
0−
→ ⊕m
→ K[Pn−1 ] −
→ 0.
λ=0, λ6=1 A · sλ ⊕ A · (an−1 + dan+1 ) −
It is worth pointing out that from Theorem 7 and Remark 1, one can obtain the Noether resolution of K[Pr ] with respect to the standard grading. In addition, the description of (Pr )i for
all r ∈ {2, . . . , n − 1}, i ∈ {0, 1}, allows us to use Remark 1 to provide a formula for the
Castelnuovo-Mumford regularity of K[Pr ].
Theorem 8. The Castelnuovo-Mumford regularity of K[Pr ] equals:
20
I. BERMEJO, E. GARCÍA-LLORENTE, I. GARCÍA-MARCO, AND M. MORALES
mn −1
⌈ n−1 ⌉ + 1,
2d,
reg(K[Pr ]) =
mn −1
⌈ n−1 ⌉,
if r ∈ {2, n − 1} and r ≤ m1 ,
if r = 2 and m1 = 1, and
if r ∈ {3, . . . , n − 2}, or r = n − 1 and m1 < r
Let us illustrate the results of this section with an example.
Example 5. Consider the projective monomial curve given parametrically by:
x1 = st6 , x2 = s5 t2 , x4 = s7 , x5 = t7 .
We observe that the curve corresponds to C2 , where C is the curve associated to the arithmetic
sequence m1 < · · · < mn with m1 = 1, d = 2 and n = 4. Hence, by Theorem 7, we get that the
multigraded Noether resolution of K[P2 ] is
A ⊕ A · (1, 6) ⊕ A · (5, 2)⊕
− K[P2 ] −
→ 0.
0−
→ A · (10, 18) ⊕ A · (11, 24) →
− A · (2, 12) ⊕ A · (6, 8) ⊕ A · (10, 4)⊕ →
A · (3, 18) ⊕ A · (11, 10) ⊕ A · (4, 24)
By Corollary 3, we get that the multigraded Hilbert series of K[P2 ] is
4 24
3 18
5 2
6 8
10 4
10 18
11 10
11 24
1 + t1 t62 + t21 t12
2 + t1 t2 + t1 t2 + t1 t2 + t1 t2 + t1 t2 − t1 t2 + t1 t2 − t1 t2
.
(1 − t71 )(1 − t72 )
Following Remark 1, if we consider the standard grading on R, we get the following Noether
resolution of K[P2 ]:
HSK[P2 ] (t1 , t2 ) =
0−
→ A(−4) ⊕ A(−5) −
→
A ⊕ A(−1)2 ⊕ A(−2)3
→ K[P2 ] −
−
→ 0,
A(−3)2 ⊕ A(−4)
and the following expression for the Hilbert series of K[P2 ]:
HSK[P2 ] (t) =
We also have that reg(K[P2 ]) = 4.
1 + 2t + 3t2 + 2t3 − t5
.
(1 − t)2
ACKNOWLEDGEMENTS
The authors want to thank the anonymous referees for their comments and suggestions that we
believe have helped to improve this manuscript. In particular, Section 5 was included to answer a
question made by the referees.
The first three authors were supported by the Ministerio de Economía y Competitividad, Spain
(MTM2013-40775-P and MTM2016-78881-P).
R EFERENCES
[Abbott et al. (2015)] J. Abbott, A. M. Bigatti, G. Lagorio. CoCoA-5: a system for doing Computations in Commutative Algebra. Available at http://cocoa.dima.unige.it
[Bermejo et al. (2017)] I. Bermejo, E. García-Llorente, I. García-Marco, On projective monomial curves associated
to generalized arithmetic sequences. J. Symbolic Comput. 81 (2017), 1–19.
[Bermejo & Gimenez (2000)] I. Bermejo, Ph. Gimenez, On Castelnuovo-Mumford regularity of projective curves.
Proc. Amer. Math. Soc. 128 (2000), no. 5, 1293–1299.
[Bermejo & Gimenez (2001)] I. Bermejo, Ph. Gimenez, Computing the Castelnuovo-Mumford regularity of some
subschemes of PnK using quotients of monomial ideals, Effective methods in algebraic geometry (Bath, 2000). J.
Pure Appl. Algebra bf 164 (2001), no. 1-2, 23–33.
[Decker et al. (2015)] W. Decker, G. M. Greuel, G. Pfister and H. Schöenemann, S INGULAR 4-0-2 A computer
algebra system for polynomial computations. http://www.singular.uni-kl.de (2015).
[Erdös & Graham (1972)] P. Erdös and R. L. Graham, On a linear diophantine problem of Frobenius, Acta Arithm.
21 (1972), 399–408.
[García-Sánchez & Rosales (2002)] P. A. García-Sánchez, J. C. Rosales, On Buchsbaum simplicial affine semigroups. Pacific J. Math. 202 (2002), no. 2, 329–393.
NOETHER RESOLUTIONS IN DIMENSION 2
21
[Gruson et al. (1983)] L. Gruson, R. Lazarsfeld, C. Peskine, On a theorem of Castelnuovo, and the equations defining
space curves, Invent. Math. 72(3) (1983), 491–506.
[Goto et al. (1976)] S. Goto, N. Suzuki, K. Watanabe, On affine semigroups. Japan J. Math. 2 (1976), 1–12.
[Grayson & Stillman (2015)] D. R. Grayson, M. E. Stillman. Macaulay2, a software system for research in algebraic
geometry. Available at http://www.math.uiuc.edu/Macaulay2/
[Hoa (1988)] L.T. Hoa, On Segre products of semigroup rings. Nagoya Math. J., Vol.110 (1988), 113–128.
[Hoa (1991)] L.T. Hoa, The Gorenstein property depends upon characteristic for affine semigroup rings. Arch. Math.
56 (1991), 228–235.
[Kawasaki (2000)] T. Kawasaki, On Macaulayfication of Noetherian schemes. Trans. AMS 352(6), 2517–2552.
[L’vovsky (1996)] S. L’vovsky, On inflection points, monomial curves and hypersurfaces containing projective
curves, Math. Ann. 306(4)(1996), 719–735.
[Li et. al (2012)] P. Li, D. P. Patil, L. G. Roberts, Bases and ideal generators for projective monomial curves. Comm.
Algebra 40 (2012), no. 1, 173–191.
[Molinelli & Tamone (1995)] S. Molinelli, G. Tamone, On the Hilbert function of certain rings of monomial curves.
J. Pure Appl. Algebra 101 (1995), no. 2, 191–206
[Morales (1985)] M. Morales, Fonctions de Hilbert, genre géométrique d’une singularité quasi homogène CohenMacaulay [Hilbert functions, geometric genus of a Cohen-Macaulay quasihomogeneous singularity]. C. R. Acad.
Sci. Paris Sér. I Math. 301 (1985), no. 14, 699–702.
[Morales & Nhan (2003)] M. Morales, L. T. Nhan, On length of generalized fractions 1/(xn1 , . . . , xnd , 1). Vietnam J.
Math. 31 (2003), no. 4, 359–365 .
[Morales (2007)] M. Morales, On the S2-fication of some toric varieties. Comm. Algebra 35 (2007), no. 1, 2409–
2430.
[Morales (2016)] M. Morales, On Hilbert function and Poincaré series of multigraded rings, work in progress
[Ramírez Alfonsín (2005)] J. L. Ramírez Alfonsín, The Diophantine Frobenius Problem, Oxford Lecture Series in
Mathematics and its Applications 30, Oxford University Press, 2005.
[Rosales & García-Sánchez (2009)] J. C. Rosales y P. A. García-Sánchez, Numerical semigroups, Developments in
Mathematics, 20, Springer, New York, 2009.
[Selmer (1977)] E. S. Selmer, On the linear diophantine Problem of Frobenius, J. Reine Angewandte Math.
293/294(1) (1977), 1–17.
[Stanley (1978)] R. P. Stanley, Hilbert Functions of Graded Algebras. Adv. Math. 28:1 (1978), 57–83.
[Sturmfels (1996)] B. Sturmfels, Gröbner Bases and Convex Polytopes. University Lecture Series 8, American Mathematical Society, Providence, RI, 1996.
[Trung & Hoa (1986)] N. V. Trung, L. T. Hoa, Affine semigroups and Cohen-Macaulay rings generated by monomials. Trans. AMS 298(1) (1986), 145–167.
[Villarreal (2015)] R. H. Villarreal, Monomial Algebras, Second Edition, Monographs and Research Notes in Mathematics. Chapman and Hall/CRC, 2015.
U NIVERSIDAD DE L A L AGUNA . FACULTAD DE C IENCIAS . S ECCIÓN DE M ATEMÁTICAS . AVDA . A S F RANCISCO S ÁNCHEZ , S / N . A PARTADO DE CORREOS 456. 38200-L A L AGUNA . T ENERIFE . S PAIN .
E-mail address: [email protected]
TROFÍSICO
U NIVERSIDAD DE L A L AGUNA . FACULTAD DE C IENCIAS .
TROFÍSICO F RANCISCO S ÁNCHEZ , S / N . A PARTADO DE CORREOS
S ECCIÓN DE M ATEMÁTICAS . AVDA . A S 456. 38200-L A L AGUNA . T ENERIFE . S PAIN .
E-mail address: [email protected]
A IX -M ARSEILLE U NIVERSITÉ , CNRS, LIF UMR 7279, M ARSEILLE , F RANCE .
E-mail address: [email protected], [email protected]
U NIVERSITÉ DE G RENOBLE I, I NSTITUT F OURIER , UMR 5582, B.P.74, 38402 S AINT-M ARTIN D’H ERES
C EDEX , G RENOBLE AND ESPE DE LYON , U NIVERSITÉ DE LYON 1, LYON , F RANCE .
E-mail address: [email protected]
| 0math.AC
|
1
Achievable Rates of Buffer-Aided Full-Duplex
Gaussian Relay Channels
arXiv:1704.02596v2 [cs.IT] 2 Aug 2017
Ahmed El Shafie†, Ahmed Sultan⋆ , Ioannis Krikidis∗ , Naofal Al-Dhahir† , Ridha Hamila††
† University of Texas at Dallas, USA.
⋆ King Abdullah University of Science and Technology (KAUST), Saudi Arabia.
∗ University of Cyprus, Cyprus.
†† Qatar University.
Abstract—We derive closed-form expressions for the achievable
rates of a buffer-aided full-duplex (FD) multiple-input multipleoutput (MIMO) Gaussian relay channel. The FD relay still suffers
from residual self-interference (RSI) after the application of
self-interference mitigation techniques. We investigate both cases
of a slow-RSI channel where the RSI is fixed over the entire
codeword, and a fast-RSI channel where the RSI changes from
one symbol duration to another within the codeword. We show
that the RSI can be completely eliminated in the slow-RSI case
when the FD relay is equipped with a buffer while the fast
RSI cannot be eliminated. For the fixed-rate data transmission
scenario, we derive the optimal transmission strategy that should
be adopted by the source node and relay node to maximize the
system throughput. We verify our analytical findings through
simulations.
Index Terms—Buffer, full-duplex, relay, MIMO, achievable
rate, precoding.
I. I NTRODUCTION
Relay nodes play an important role in wireless communications due to their ability to increase the data rate between
a pair of communicating nodes [1]. Relays can operate in
three different modes, namely, half-duplex (HD) mode [2]–
[4], full-duplex (FD) mode [5]–[11], or hybrid HD/FD mode
[12]–[14]. In the FD mode, data transmission and reception
at the FD relay node occur simultaneously and over the same
frequency band. However, due to the simultaneous reception
and transmission, FD relays are impaired by loopback selfinterference (LSI), which occurs due to energy leakage from
the transmitter radio-frequency (RF) chain into the receiver
RF chain [15]–[18]. LSI can be suppressed by up to 120
dB in certain scenarios, as discussed in [19]. However, the
LSI cancellation process is never perfect, thereby leaving
some non-negligible residual self-interference (RSI). In many
modern communication systems such as WiFi, Bluetooth,
and Femtocells, the nodes’ transmit power levels and the
distances between communicating nodes have been decreasing.
In such scenarios, the high computation capabilities of modern
terminals enable efficient implementation of the FD radio
technology [20]–[22]. In the HD mode, transmission and
reception occur over orthogonal time slots or frequency bands.
As a result, HD relays do not suffer from RSI, but at the cost of
This paper was made possible by NPRP grant number 8-627-2-260 from
the Qatar National Research Fund (a member of Qatar Foundation). The
statements made herein are solely the responsibility of the authors.
wasting time and frequency resources. Hence, the achievable
data rates of an FD relaying system might be significantly
higher than that of an HD relaying system when the RSI
has low power. In the hybrid HD/FD mode [14], the relay
can operate in either HD mode or FD mode to maximize the
achievable rate. The key idea is to dynamically switch between
the two modes based on the RSI level. When the RSI level
is high, the HD mode can achieve higher rates. On the other
hand, when the RSI level is low, the FD mode can result in
much higher data rates.
Integrating multiple-input multiple-output (MIMO) techniques with relaying further improves the communication
performance and data rates [23], [24]. Although most previous
research efforts have focused on MIMO-HD relaying, recent
research has also investigated the performance of MIMO-FD
relaying [25]–[27]. MIMO techniques provide an effective
means to mitigate the RSI effects in the spatial domain [25]–
[27]. With multiple transmit or receive antennas at the FD
relay node, data precoding at the transmit side and filtering at
the receive side can be jointly optimized to mitigate the RSI
effects. Minimum mean square error (MMSE) and zero forcing
(ZF) are two widely adopted criteria in the literature for the
precoding and decoding designs [28]. ZF aims to completely
cancel out the undesired self-interference signals and results
in an interference-free channel at the relay node’s receive side.
Although ZF normally results in a sub-optimal solution to the
achievable performance (i.e., data rate and bit error rate), its
performance is asymptotically optimal in the high signal-tonoise ratio (SNR) regime. On the other hand, MMSE improves
the performance of the precoder/decoder design compared to
ZF since it takes into account the noise impact at the cost
of a higher complexity. However, due to the implementation
simplicity and optimality in the high-SNR regime, ZF has been
proposed as a useful design criterion to completely cancel
the RSI and separate the source-relay and relay-destination
channels.
Assuming there is no processing delay at the relay, the
optimal precoding matrix for a Gaussian FD amplify-andforward (AF) relay that maximizes the achievable rate under
an average power constraint is studied in [29]. In this case,
the design approach and the resulting precoding solution are
similar to the HD case. The joint precoding and decoding
design for an FD relay is studied in [17], [30], where both
ZF and MMSE solutions are discussed. The ZF solution used
2
in [17], [30] and most early works use a conventional approach
based on the singular value decomposition (SVD) of the RSI
channel. The main drawback of this approach is that the
ZF solution only exists given that the numbers of antennas
at the source, FD relay and the receiver satisfy a certain
dimensionality condition. To overcome this limitation, [26]
adopts an alternative criterion and proposes to maximize the
signal-to-interference ratios between the power of the useful
signal to the power of RSI at the relay input and output,
respectively. Conventional ZF precoding and decoding are
designed based on the singular vectors of the RSI channels. In
[31], a joint design of ZF precoding and decoding is proposed
to fully cancel the RSI at the relay, taking into account the
source-relay and relay-destination channels. In [8] and [32],
the precoding and decoding vectors are jointly optimized to
maximize the end-to-end performance.
Buffer-aided schemes for decode-and-forward (DF)-FD
Gaussian relay channels were proposed in [14], [33], [34].
In [33], the authors assumed that the RSI at the FD relay is
negligible which is not realistic. The authors in [34] assumed
that the RSI is fixed and does not vary with time. This may or
may not be the case depending on system parameters and the
employed self-interference cancellation techniques [35], [36].
In addition, the authors in [34] do not investigate the case
when both the source and the relay transmit with a fixed rate
in all time slots; a scenario which is investigated in this paper.
The authors of [12], [14] proposed a hybrid HD/FD scheme to
maximize the throughput of a relaying system for fixed-rate
data transmissions. However, the authors neglected the fact
that the relay knows its transmitted data signal and can do
better in mitigating its impact as will be fully investigated in
this paper.
Most of the aforementioned research assumed that the
RSI is known but the data symbols are unknown. The first
assumption is impractical since, by definition, the RSI is the
remaining interference after applying all kinds of practically
feasible interference mitigation techniques. Plausibility of the
assumption that the data symbols are unknown depends on the
operating scenario. For instance, in DF-FD relaying, the relay
needs to know the entire codeword to know the transmitted
sequence. Hence, it makes sense to assume that the symbols
are unknown until the entire codeword is decoded. But if we
assume that the relay has a buffer to store the data received
from the source node, the relay will have its own data which
are possibly different from the data that are currently received
from the source. Hence, an FD mode can be applied and the
entire transmitted sequence is known a priori by the relay.
In this paper, we consider a buffer-aided MIMO-FD Gaussian relay channel. Since the relay has a buffer, it knows
the codewords that it transmits. Given this information, the
contributions of this paper are summarized as follows
•
•
We derive closed-form expressions for the achievable
rates of the source-relay and relay-destination links when
the RSI is changing slowly or quickly.
We show that the buffer can help in completely canceling
the impact of RSI for the case of slow RSI, when the
buffer is non-empty. The maximum achievable rate of
the source-relay link under the FD mode is that of the
source-relay link without interference.
• For fast RSI, we show that the achievable rate of the
source-relay link is degraded due to RSI and the degradation is quantified analytically. When the optimal precoder
that maximizes the achievable rate of the relay-destination
link is used, we derive a closed-form expression for the
achievable rate of the source-relay link.
Notation: Unless otherwise stated, lower- and upper-case
bold letters denote vectors and matrices, respectively. IN
denotes the identity matrix whose size is N × N . CM×N
denotes the set of all complex matrices of size M ×N . 0M×N
denotes the all-zero matrix with size M × N . (·)⊤ , (·)∗ , and
(·)H denote transpose, complex conjugate, and Hermitian (i.e.,
complex-conjugate transpose) operations, respectively. | · | denotes the absolute of the value in brackets. CN (x, y) denotes a
complex circularly-symmetric Gaussian random variable with
mean x and variance y. E{·} denotes statistical expectation. ⊗
is the Kronecker product. diag = {·} denotes a diagonal matrix
with the enclosed elements as its diagonal elements. Trace{·}
denotes the sum of the diagonal entries of the matrix enclosed
in braces. vec{·} converts the input M × N matrix into a
column vector of size M N × 1.
II. S YSTEM M ODEL AND M AIN A SSUMPTIONS
We consider a dual-hop DF-FD MIMO Gaussian relay
channel, where a multi-antenna source node communicates
with its multi-antenna destination node through an FD multiantenna relay node, as shown in Fig. 1. Each node is equipped
with M antennas.1 A direct link between the source and its
destination (i.e., source-destination link) does not exist due
to shadowing and large distances between them [14], [33],
[34]. We assume that the relay node is equipped with a finitesize buffer/queue to store the incoming data traffic from the
source node. We denote the buffer at the relay node as QR
and its maximum size as Qmax . The source node is always
backlogged with data to transmit. It is assumed that the time
is partitioned into discrete equal-size time slots of T seconds,
where the duration of one time slot is equal to the channel
coherence time and the channel bandwidth is W . We use
subscripts S, R, and D to denote the source node, relay node,
and destination node, respectively.
Each wireless link exhibits a quasi-static fading where a
channel matrix between two nodes remains unchanged within
the duration of one time slot and changes independently from
one time slot to another. We consider slow-RSI and fast-RSI
scenarios, where the RSI is fixed over the entire codeword
or changes from one symbol duration to another within
the codeword, respectively.2 We denote the RSI coefficient
matrix by HRR in the slow-RSI case and by H̃RR in the
fast-RSI case. The elements of the RSI coefficient matrices
1 For simplicity of presentation, we assume equal number of antennas at all
nodes. However, the same analysis can be easily extended to the scenario of
different number of antennas at all nodes.
2 Typically, slow-RSI is assumed in the literature (e.g. [14], [33]), which
represents an optimistic assumption and the best-case scenario in system
design. However, in this paper, we investigate both scenarios of slow-/fastRSI.
3
۶ோோ : slow-RSI case
ܯ-Antenna
Source
۶ௌோ
෩ ோோ : fast-RSI case
۶
ܯ-Antenna
FD Relay
۶ோ
ܯ-Antenna
Destination
Figure 1. The considered dual-hop network. In the figure, we denote the RSI
coefficient matrix by HRR ∈ CM ×M in the slow-RSI case and by H̃RR ∈
CnM ×nM in the fast-RSI case. The two matrices are different, since in the
slow-RSI case, the RSI matrix remains constant over the entire codeword,
while in the fast-RSI case, the RSI matrix changes from one symbol duration
to another within the codeword. HSR ∈ CM ×M and HRD ∈ CM ×M are
the channel matrices of the source-relay link and the relay-destination link,
respectively.
RSI case). In addition, the slow-RSI model is suitable for
the cases of fixed-rate transmission and when analyzing the
system based on average performance [41]. Hence, it will be
adopted when we study the fixed-rate transmission scenario in
Section V.
In the following sections, we derive the closed-form expressions for the achievable rate of the source-relay link for
slow-RSI and fast-RSI cases.
III. S LOW-RSI C ASE
In this case, the RSI varies across time slots (i.e., from one
coherence time duration to another), but remains fixed within
each time slot.
A. Achievable Rates Derivations
are independent and identically distributed (i.i.d.) zero-mean
circularly-symmetric Gaussian-distributed random variables
2
with variance σRR
[35]–[38]. Each link is also corrupted by
an additive white Gaussian noise (AWGN) process with zero
mean and variance κm . It is assumed that the average transmit
power at Node m ∈ {S, R} is Pm . For notation simplicity,
unless otherwise stated, we drop the time slot (coherence time)
index from the equations and use only the symbol duration
index. However, it is worth mentioning that the channels
between the source and the relay and between the relay and
the destination are constant within each coherence time and
changes independently from one coherence time to another.
The RSI channel is time-varying even when the communication links do not exhibit fading [17], [37], [39]–[41]. The
RSI variations are due to the cumulative effects of various
distortion sources including noise, carrier frequency offset,
oscillator phase noise, analog-to-digital/digital-to-analog conversion (ADC/DAC) imperfections, in-phase/quadrature (I/Q)
imbalance, imperfect channel estimation, etc [17], [37], [39]–
[41]. These impairments and distortions have a significant
impact on the RSI channel due to the very small distance
between the transmitter-end and the receiver-end of the LSI
channel. Moreover, the variations of the RSI channel are
random and thereby cannot be accurately estimated at the FD
node [17], [37], [39]–[41]. The statistical properties of the
RSI variations are dependent on the hardware configuration
and the adopted LSI suppression techniques. In [37], the RSI
is assumed to be fixed/constant during the transmission of
a codeword comprised of many symbols. Hence, the RSI
model proposed in [37], and most of the papers in the
literature, captures only the long-term, i.e., codeword-bycodeword, statistical properties of the RSI channel. However,
the symbol-by-symbol RSI variations are not captured by the
model proposed in [37] since these variations are averaged
out. Nevertheless, for a meaningful information-theoretical
analysis, the symbol-by-symbol variations of the RSI should
be taken into consideration. The statistics of the RSI variations
affect the achievable rates of the considered FD Gaussian relay
channels. In this paper, we derive the achievable rates of the
considered Gaussian relay channels for both the best-case RSI
model (slow-RSI case) and the worst-case RSI model (fast-
The achievable rates of the communication links in bufferaided relay networks change based on the relaying queue state
(i.e., empty or non-empty). That is, if the relay’s queue is
empty, FD mode operation is not possible since, as mentioned
in [42], the practicality of DF-FD relaying is questionable,
when the relay does not have the entire codeword prior to
data transmission. Hence, we simply assume that, when the
relay is empty, it operates in an HD mode and it receives data
of the source node. Assume that the source node transmits M
independent codewords of length n, n > M . The data matrix
transmitted by the source node, denoted by XS ∈ Cn×M , is
given by
X (1)
S,1
XS,1 (2)
XS =
..
.
XS,1 (n)
XS,2 (1)
XS,2 (2)
..
.
XS,2 (n)
...
...
..
.
...
XS,M (1)
XS,M (2)
..
.
XS,M (n)
,
(1)
where the elements of XS are assumed to be i.i.d. Gaussian
circularly-symmetric random variables with zero mean and
variance P̃S = PS /M (i.e., variance per data stream). We
assume Gaussian data signals at the source and relay nodes as
in, e.g., [14], [33], [35]–[38] and the references therein. The
received signal at the relay is given by
YR = XS H⊤
SR + ǫ R ,
(2)
where ǫ R ∈ Cn×M is the AWGN noise matrix at the relay,
and HSR ∈ CM×M is the channel matrix between the source
node and the relay node with element (v, ℓ) in HSR being the
channel coefficient between the source’s v-th antenna and the
relay’s ℓ-th antenna. Hence, the achievable rate of the sourcerelay link is given by
!
M
X
P̃S
HD
(3)
log2 1 +
ISR =
ηv ,
κR
v=1
where ηv is the v-th eigenvalue of HSR HH
SR . The expression
in (3) can be deduced from Appendix A by setting the relay’s
data precoding matrix to zero (i.e., Ψ = 0M×M ).
If the relay’s queue is non-empty, the relay transmits a
codeword that is different from the source’s codeword and
4
operation in the FD mode is possible. The received signal at
the relay’s receiver is given by
⊤
YR = XS H⊤
SR + XR Ψ HRR + ǫ R ,
(4)
where XR ∈ Cn×M is the data matrix transmitted by the
relay node and has the same structure as XS in (1) but with
the codewords transmitted by the relay which are independent
from those transmitted by the source node. The elements of
XR are i.i.d. with n1 E{vec{XR }H vec{XR }} = PR and,
hence, a realization of XR has a rank equal to M with probability one. The relay node applies the data precoding matrix
ΨΨ H } = M .
Ψ ∈ CM×M to its data vectors, where Trace{Ψ
M×M
Matrix HRR ∈ C
is the RSI coefficient matrix. Element
(v, ℓ) in H⊤
RR represents how transmission from the relay’s
ℓ-th antenna impacts the relay’s received signal on its v-th
antenna. We collect the elements of the matrix YR in a vector
to compute the information rate under Gaussian signaling as
follows [43]
n
o
1h
H
FD
ISR
=
log2 det E vec {YR } vec {YR }
n
(5)
ǫR
−log2 det E vec XRΨ H⊤
RR +ǫ
oi
H
,
× vec XR ΨH⊤
RR + ǫ R
where
vec{YR } = (HSR ⊗ In ) vec{XS } + (IM ⊗ XRΨ ) vec{H⊤
RR }
+ vec{ǫǫ R },
(6)
by exploiting the following property of vec{.}
vec{AB} = B⊤ ⊗ In vec{A} = (IM ⊗ A) vec{B},
(7)
where A ∈ Cn×l and B ∈ Cl×M .
Proposition 1. The information rate of the source-relay link
under Gaussian signaling is given by
M
1X
FD
2
ISR
=
log2 det P̃S ηv In +σRR
XRΨΨ H XH
R +κR In
n v=1
!
(8)
2
H
H
−log2 det σRR XRΨΨ XR +κR In ,
where ηv is the v-th eigenvalue of HSR HH
SR .
Proof. See Appendix A.
Proposition 2. The precoder that maximizes the information
rate of the source-relay link under Gaussian relay channel,
FD
which is referred to as the ISR
-maximizing precoder, is rank1.
Proof. See Appendix B.
In the next two propositions, we present a closed-form
expression for the achievable rate of the source-relay link
FD
when the relay uses the rank-1 precoder (i.e., ISR
-maximizing
FD
precoder) in Proposition 2 and the IRD -maximizing precoder
derived in Appendix C, respectively.
√
where q ∈ CM×1 with
Proposition 3. Letting Ψ = M qqH , √
H
q q = 1, and substituting with Ψ = M qqH into (8), the
information rate of the source-relay link under the slow-RSI
scenario is given by
P̃S
ηv
log2 1 +
κR
v=1
M
2
σRR
1X
+
log2 1 +
qH XH
R XR q
n v=1
(P̃S ηv + κR )
!
2
σRR
H H
q XR XR q
− log 2 1 +
,
κR
FD
ISR
=
M
X
(9)
where q is the normalized eigenvector corresponding to the
minimum eigenvalue of XH
R XR .
Proof. See Appendix D.
Proposition 4. When the optimal precoder that maximizes the
information rate of the relay-destination channel derived in
Appendix C is used by the relay node, the achievable rate
expression of the source-relay link can be rewritten as
M
1X
n log2 (1 + P̃S ηv )
n v=1
2
σRR
+ log2 det IM +
EEH XH
X
R
R
P̃S ηv + κR
!
σ2
,
− log2 det IM + RR EEH XH
R XR
κR
FD
ISR
=
(10)
where Ψ = EQH
RD is full rank with E denoting a diagonal
matrix such that EEH contains the power fractions assigned
to each data stream and Trace{ΨΨH } = M .
Proof. See Appendix E.
Proposition 5. When n goes to infinity, the achievable rate in
the slow-RSI case is given by
FD
ISR
=
P̃S
log2 1 +
ηv ,
κR
v=1
M
X
(11)
which is the achievable rate of the source-relay channel with
no interference.
Proof. When n → ∞, the diagonal elements of n1 XH
R XR in
(10) converge to PR and the off-diagonal elements scaled by
1/n converge to zero almost surely [44]. Thus,
M
1X
P̃S
n log2 1 +
ηv
n→∞ n
κR
v=1
M
2
X
σRR
2
log2 1 +
+
nPR |Eℓ |
P̃S ηv + κR
ℓ=1
!
M
2
X
σRR
2
,
nPR |Eℓ |
log2 1 +
−
κR
ℓ=1
FD
ISR
= lim
(12)
where Eℓ is the ℓ-th element on the main diagonal of E. The
last two terms go to zero for finite M (n ≫ M ). Thus, we
get the expression in (11).
The result in Proposition 5 is promising since it implies
that, regardless of the precoder employed at the relay, the
achievable rate of the source-relay channel under FD operation
5
equals to the rate of the source-relay channel with no interference. Hence, if the relay uses the precoder that maximizes
FD
the achievable rate of the relay-destination link (i.e., IRD
maximizing precoder derived in Appendix C), the achievable
rates of the two links (i.e., source-relay and relay-destination
links) will be simultaneously maximized. Accordingly, the
channel capacities of the two links (source-relay and relaydestination links) can be achieved.
IV. FAST-RSI C ASE
In the case of fast-RSI, the RSI changes independently from
one symbol duration to another. That is, each symbol within
the codeword experiences a different RSI realization.
A. Achievable Rates Derivations
Assuming M independent codewords transmitted by the
relay node, the data matrix, denoted by X̃R ∈ Cn×nM , is
given by
X (1)
R
0
X̃R =
..
.
0
0
XR (2)
..
.
0
...
...
..
.
...
0
0
..
.
XR (n)
,
(13)
where XR (j) = [XR,1 (j) XR,2 (j) . . . XR,M (j)] ∈ C1×M
is the data symbols vector transmitted by the relay at the
jth symbol duration. The RSI coefficient matrix, denoted by
H̃RR ∈ CnM×M , is given by
H̃RR =
H̃⊤
RR (1)
H̃⊤
RR (2)
.
.
.
H̃⊤
RR (n)
,
(14)
where each block H̃RR (j) is M × M .
Hence, the received signal vector at the relay’s receiver
YR =
XS H⊤
SR
+ X̃R Ψ̃R H̃RR + ǫ R ,
(15)
where ǫ R ∈ CN ×M is the noise matrix at the relay and
Ψ R ∈ CnM×nM is the data precoding matrix used at the
Ψ̃
relay node, H̃RR ∈ CnM×nM is the RSI channel matrix.
ΨR ∈ CnM×nM has the block diagonal structure
Matrix Ψ̃
Ψ R = diag{Φ
Φ, Φ , . . . , Φ }, where Φ is an M × M matrix with
Ψ̃
ΦΦ H } = M .
Trace{Φ
Proposition 6. The achievable rate of the source-relay link
for the fast-RSI case is given by
M
H
1X
2
log2 det P̃S ηv In +σRR X̃R Ψ̃ X̃R Ψ̃ +κR In
n v=1
!
H
2
+ κR In .
−log2 det σRR X̃R Ψ̃ X̃R Ψ̃
FD
ISR
=
(16)
Proof. See Appendix F.
Using the same approach as in the slow-RSI case to
derive the precoder that maximizes the information rate
FD
of source-relay link (i.e., ISR
-maximizing precoder), if
Ψ̃
=
diag{Φ
,
Φ
,
.
.
.
,
Φ
},
then choosing Φk to be
1 2
N
R
⊤
X∗
R (j)XR (j)
IM − kXR (j)k2
nulls the relay’s transmission. In other
words, although this precoder cancels the relay’s transmission
from the relay’s receive side, but it also cancels the transmissions from everywhere else. Thus, this precoder reduces
the achievable rate of the relay-destination link to zero, and
effectively makes the relay operate as an HD terminal. In
the sequel, we study the achievable rates of the source-relay
and relay-destination links, respectively, when the precoder Φ
FD
FD
follows the ISR
-maximizing and IRD
-maximizing precoder
designs employed in the slow-RSI case. Moreover, we study
the asymptotic case as n → ∞.
FD
Note that the ISR
-maximizing precoder is not necessarily
the precoder that also maximizes the achievable rate of the
relay-destination link. The optimal precoder at the relay should
be designed based on a selected performance criterion (e.g.,
maximum rate between the two rates of the communications
hops, minimum end-to-end bit-error-rate probability, maximum sum-rate of the two communication hops, etc). For example, if the goal is to maximize the minimum between the inforFD
FD
mation rates of the two hops (i.e. maximize min{ISR
, IRD
}),
we need to derive the optimal precoder based on that. That is,
FD
we need to find the optimal precoder that maximizes IRD
.
However, this precoder is difficult to obtain analytically even
for M = 2. To gain some insights, we provide a heuristic
solution which is realized as follows. The relay uses the two
FD
FD
precoders: ISR
-maximizing and IRD
-maximizing precoders.
Then, it computes the minimum achievable rates of the two
hops under each case. After that, the relay selects the precoder
with the highest minimum achievable rate.
Proposition 7. The achievable rate of the source-relay link for
FD
the fast-RSI case, when the relay uses the ISR
-maximizing
pre√
coder of the slow-RSI case, which has the form Φ = M qqH ,
is given by
FD
ISR
=
2
σRR
X
M
P̃S
κR
2
log2 1+
|XR (j)q|
E log2 1 +
ηv +
S
κR
1 + κP̃R
ηv
v=1
v=1
σ2
.
− log2 1 + RR |XR (j)q|2
κR
(17)
M
X
Proof. See Appendix G.
The next proposition considers the case where the relay uses
the precoder that maximizes the achievable rate of the relayFD
destination link (i.e., the IRD
-maximizing precoder).
Proposition 8. The achievable rate of the source-relay link
FD
for the fast-RSI case, when the relay uses the IRD
-maximizing
precoder of the slow-RSI, which has the form Φ = EQ∗RD , is
given by
P̃S
ηv
log 2 1 +
κR
v=1
2
PM
σRR
2
M
X
i=1 |XR,i (j)|
M κR
E log2 1 +
+
S
1 + κP̃R
ηv
v=1
!)!
(
M
σ2 X
|XR,i (j)|2
.
− E log2 1 + RR
M κR i=1
FD
ISR
=
M
X
(18)
6
Proof. See Appendix H.
For the case of equal power allocation toP
data streams, when
2
M is large, we can approximate X (j) ≈ M
i=1 |XR,i (j)| =
M PR from the strong law of large numbers. Hence, the
achievable rate of the source-relay link is
FD
ISR
=
M
X
v=1
− log2
log2
σ 2 PR
P̃S
ηv + RR
+1
κR
κR
!
2
σRR
PR
+1
κR
(19)
.
A special case- single-input single-output (SISO):
Let hSR and hRD denote the channel coefficients of the
source-relay and relay-destination links, respectively. Since
nP
is very large, from the strong law of large numbers,
n
1
2
i=1
n
log2 1 + γq |xR (i)|
will almost surely converge
2
to E log2 1 + γq |xR (i)|
where γq ∈ {γ1 , γ2 } with
γ1 = κ +P̃ 1|h |2 and γ2 = κ1R . Since |xR (i)|2 is
R
S
SR
an exponentially-distributed
random variable, the average of
log2 1 + γq |xR (i)|2 is given by
E log2 1+γq |xR (i)|2 =
Z
∞
log2 1 + γq |xR (i)|2 d|xR (i)|2
exp γ σ 21 P
1
q RR R
,
Ei
=
2
ln(2)
γq σRR
PR
(20)
0
R∞
where Ei(x) = x exp(−u)
du is the exponential integral.
u
Substituting in the information rate expression of the sourcerelay link, we have
P̃S
FD
ISR
= log2 1 + |hSR |2
κR
1
1
1
+
Ei
exp
2
2
ln(2)
γ1 σRR
PR
γ1 σRR
PR
1
1
1
−
Ei
.
exp
2
2
ln(2)
γ2 σRR PR
γ2 σRR PR
(21)
If the relaying queue is non-empty, the optimal transmission
scheme is that both the source and the relay transmit data
simultaneously. This is because the two links (i.e., sourcerelay and relay-destination links) are completely independent
and separable because, when RSI is slow and n → ∞, the
self-interference at the relay is removed and the source-relay
channel is not affected by relay transmissions. The probability
that the queue transits from state ℓ > 0 to state ℓ + 1, denoted
by aℓ , is equal to the probability that the source-relay link is
not in outage and that of the relay-destination is in outage.
Hence, aℓ is given by
FD
FD
(24)
aℓ = a = Pr ISR
≥ R Pr IRD
<R ,
where ℓ > 0 and it denotes the state of the relaying queue
FD
(i.e., number of packets at the relaying queue) and IRD
is
the achievable rate of the source-destination link which is
derived in Appendix C. Similarly, the probability that the
queue transits from state 0 < ℓ < Qmax to state ℓ − 1, denoted
by bℓ , is equal to the probability that the source-relay link is
in outage, whereas the relay-destination link is not. Hence, bℓ
is given by
FD
FD
(25)
bℓ = b = Pr IRD
≥ R Pr ISR
<R .
When the relaying buffer is full, the transition probability,
denoted by bQmax , is given by
FD
(26)
≥R ,
bQmax = Pr IRD
since the relay cannot accept any new packets before delivering
the ones stored in its buffer.
Analyzing the relaying queue Markov chain as in [4], the
local balance equations are given by
βν aν = βν+1 bν+1 , 0 ≤ ν ≤ Qmax −1,
where βν denotes the probability of having ν packets in the
relaying queue. Using the balance equations recursively, the
stationary distribution of βν is given by
The achievable rate of the relay-destination link is given by
FD
IRD
=
HD
IRD
2 PR
= log2 1 + |hRD |
κD
.
(22)
V. A C ASE S TUDY: F IXED -R ATE T RANSMISSION
In this section, we study the fixed-rate transmission case
where the source and relay transmit with a fixed rate of R
bits/sec/Hz. Since we assume fixed-rate transmissions under
queueing constraints, the RSI channel is assumed to be slow
to capture only the long-term, i.e., codeword-by-codeword,
statistical properties [41]. The relaying queue can be modeled
as a birth-death process since only one packet is decoded at
the relay, one packet is transmitted by the relay, or one packet
is decoded and one packet is transmitted by the relay at the
same time.
When the relaying queue is empty, the probability that the
source packet is correctly decoded and stored at the relay (i.e.,
the queue state transits from state 0 to state 1) is given by
HD
(23)
a0 = Pr ISR
≥R .
(27)
βν = β0
ν−1
Y
̺=0
a̺
,
b̺+1
−1
(28)
P
Qmax Qν−1 a̺
is obtained using the
where β0 = 1+ ν=1
̺=0 b
PQmax ̺+1
normalization condition ν=0 βν = 1.
By using the normalization condition, we get
βν = β0
ν−1
Y
̺=0
a̺
b̺+1
=
(
ν
β0 aa0 abν , ν < Qmax
.
a0 b
aQmax
, ν = Qmax
β0 abQ
bQmax
(29)
max
The probability of the queue being empty is given by
β0 = 1 +
a0
a
Qmax −1
X
ν=1
a ν
b a Qmax
+
b
bQmax b
!!−1
!−1
a Qmax −1
a0 b a Qmax
a0 1 − b
.
+
= 1+
b
abQmax b
1 − ab
(30)
If the queue is unlimited in size (i.e., Qmax → ∞), a < b
is a necessary condition for the queue stability and for the
steady-state solution to exist. Simplifying Eqn. (30), we get
−1
a0
1
b−a
β0 = 1 +
.
=
(31)
b 1 − ab
b − a + a0
7
Slow-RSI: block size n=50
12
Slow-RSI: block size n=2000
12
10
10
8
8
6
6
4
4
2
2
0
0
0
5
10
15
20
25
30
0
5
10
Time slot number
Figure 2. Achievable rate of the source-relay channel for the slow-RSI
scenario when block size is n = 50.
The system throughput in packets/slot, which is the number
of correctly decoded packets at the destination per time slot,
is given by
FD
µd = (1 − β0 ) Pr{IRD
≥ R},
15
20
25
30
Time slot number
(32)
which represents the probability that the queue is non-empty
and that the relay-destination link is not in outage.
VI. N UMERICAL R ESULTS AND S IMULATIONS
In this section, we verify the analytical findings in each of
the investigated scenarios. We start with the slow-RSI case
followed by the fast-RSI case. Then, we show numerical
results for the case of fixed-rate transmission. Unless otherwise
stated, we use the following system’s parameters to generate
the results: the fading channels are assumed to be complex
circularly-symmetric Gaussian random variables with zero
mean and unit variance, κR = κD = κ, PS /κ = 10 dB,
2
PR /κ = 10 dB, and σRR
= 0 dB.
A. Slow-RSI Case
To verify our derivations, we provide some numerical results
for the achievable rate in the case of slow RSI. Our main
message from the numerical results in this subsection is to
verify that the optimal precoder that maximizes the achievable
rate of the source-relay link in case of finite block size n is
FD
the rank-1 precoder (which we refer to as the IRD
-maximizing
precoder). Moreover, we want to verify that when the block
size is sufficiently large, any precoder can be used, including
the one that maximizes the achievable rate of the relaydestination link, with no rate loss (i.e., the information rate
under slow RSI converges to the information rate of the no
interference case). Figs. 2 and 3 show the achievable rate of
FD
the source-relay link for both cases of ISR
-maximizing and
FD
IRD
-maximizing precoders when the block size is finite and
equal to n = 50 and n = 2000 symbols, respectively. We also
show the maximum achievable rate for the source-relay link
when the RSI is zero. Figs. 2 and 3 are generated using unitvariance channels, M = 2, and the instantaneous randomlygenerated channel matrices in Table I for three time slots.
Figure 3. Achievable rate of the source-relay channel for the slow-RSI
scenario when block size is n = 2000.
Slow-RSI
18
16
14
12
10
8
6
4
2
1
Figure 4.
scenario.
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
Average achievable rate of source-relay link for the slow-RSI
FD
As shown in Fig. 2, the ISR
-maximizing precoder achieves a
FD
rate which is closer to the upper bound than that of the IRD
maximizing precoder. In Fig. 3, all curves overlap thereby
implying that for slow-RSI regardless of the used precoder at
the relay, the RSI is completely canceled when n is sufficiently
high; which verifies our theoretical findings. For the case of
M > 2 and due to the significant increase in the number of
system’s parameters and channel matrices, we plot the average
achievable rate versus M in Fig. 4. As it can be seen from
FD
the figure, the ISR
-maximizing precoder achieves almost the
no-interference achievable rate when the block size is finite,
i.e., n = 50. Increasing the number of antennas increases the
achievable rate of the source-relay link.
B. Fast-RSI Case
We evaluate the achievable rate expressions that we obtained
for the fast-RSI scenario. First, we present some numerical
results for the instantaneous achievable rate expressions by
using Table I for the case of M = 2. Then, for the case
of M > 2 and since the size of the channel matrices and
the system’s parameters increase significantly, we present the
8
Table I
C HANNEL MATRICES USED TO GENERATE THE FIRST THREE TIME SLOTS IN THE FIGURES .
Slot number
1
1
1
2
2
2
3
3
3
Channel
HSR
HRR
HRD
HSR
HRR
HRD
HSR
HRR
HRD
Value
[0.013 + 0.0025i, 0.8374 − 0.8441i; 0.1166 − 0.3759i, 0.7537 + 0.2233i]
[1.6356 − 0.8668i, 0.1591 − 2.6461i; 0.7404 − 0.3748i, −0.7763 + 0.2951i]
[−0.2688 − 1.1046i, 1.0703 + 0.2583i; 0.8433 + 1.1624i, −0.3841 + 0.1363i]
[−0.3025 − 0.4487i, 0.6548 − 0.3400i; −0.4097 + 0.6069i, 0.0039 + 1.0534i]
[−0.445 + 0.9228i, −0.4446 + 0.5459i; −0.42 + 0.2586i, 0.2519 + 0.8876i]
[0.3088 − 1.7069i, 0.0019 + 0.2925i; −1.2754 + 0.2317i, −0.1195 − 0.4767i]
[0.184 − 1.0777i, 0.071 + 0.1647i; −0.3857 + 0.2473i, −0.5182 + 0.4624i]
[−0.5975 + 1.9031i, −0.347 − 0.4618i; −0.5693 + 0.2627i, −0.7111 + 0.3i]
[1.3800 + 1.5198i, 0.9294 + 1.6803i; −0.3835 + 0.5156i, 0.0726 + 0.5129i]
Fast-RSI
12
Fast-RSI
18
16
10
14
8
12
6
10
8
4
6
2
4
0
0
5
10
15
20
25
30
2
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
Time slot number
Figure 5. Achievable rate of source-relay link for the fast-RSI scenario.
average of the achievable rate expressions, averaged across
channel realizations, versus M . In Fig. 5, we show the
FD
achievable rate of the fast-RSI scenario when both the ISR
FD
maximizing and IRD -maximizing precoders of the slow-RSI
FD
are used by the relay. As expected, the ISR
-maximizing
precoder achieves a higher source-relay link achievable rate
FD
FD
than the IRD
-maximizing precoder. This is because the ISR
maximizing precoder decreases the interference caused by the
data transmissions at the relay. Fig. 6 shows the average
achievable rate of the source-relay link versus M for the
FD
FD
cases of IRD
-maximizing and ISR
-maximizing precoders.
FD
The ISR -maximizing precoder achieves a higher rate than
FD
the IRD
-maximizing precoder since the latter increases the
interference at the FD relay’s receiver due to the increased
number of data streams transmitted by the relay.
In Fig. 7, we show the minimum between the achievable
rates of the source-relay and the relay-destination links. When
FD
M = 2 and for the given channel realizations, the ISR
FD
maximizing precoder outperforms the IRD -maximizing preFD
coder. However, this is not true in general since the ISR
maximizing precoder degrades the achievable rate of the relaydestination link significantly, especially at high M . This is
clear from the values of achievable rate evaluated for the other
channel realizations as shown in Fig. 7 and in the average
achievable rate curves presented in Fig. 8. It is noteworthy
that when M = 2 as shown in Fig. 7, the relay might switch
FD
FD
between IRD
-maximizing precoder and the ISR
-maximizing
precoder to maximize the minimum achievable rate of the two
Figure 6. Average achievable rate of the source-relay link for the fast-RSI
FD -maximizing and I FD -maximizing precoders.
case under IRD
SR
FD
FD
hops, i.e., maximize min{ISR
, IRD
}. As shown in Fig. 8, the
expected value of the minimum between the achievable rate
of the source-relay link and the achievable rate of the relayFD
destination link when the ISR
-maximizing precoder is slightly
FD
better than the IRD -maximizing precoder when M = 2.
FD
Starting from M = 3, the IRD
-maximizing precoder is
superior and can achieve very high rates. On the other hand,
FD
the ISR
-maximizing precoder remains fixed regardless of M
since the total achievable rate is determined by the minimum
rate between the two communications links which is degraded
by the use of a single data stream at the relay’s transmit side.
C. Fixed-Rate Transmission
In Fig. 9, we plot the throughput of our proposed scheme
and the conventional FD scheme for R = 1 bits/sec/Hz. In
the conventional FD scheme, the source node and the relay
cooperatively transmit the data in each time slot using the DF
relaying scheme and the RSI is treated as a noise signal with a
known variance. As shown in Fig. 9, the throughput increases
by increasing the buffer size at the relay. This is expected since
increasing the buffer size allows more data transfer to and from
the relay; however, the increase is insignificant. Moreover, the
throughput is fixed for all queue sizes that are greater than or
equal to 3 packets. This implies that a data buffer with size 3
packets can be used without any throughput loss. Moreover,
our proposed scheme achieves a throughput higher than that
achieved by the conventional FD relaying. The throughput
9
Fast-RSI
7
1
0.9
6
Throughput [packets/slot]
0.8
5
4
3
2
0.7
Outer bound
0.6
Buffer-aided FD
0.5
Conventional FD
0.4
0.3
0.2
1
0.1
0
0
5
10
15
20
25
30
Time slot number
Fast-RSI
16
14
12
10
8
6
4
2
1
1.5
2
2.5
3
3.5
4
2
4
6
8
10
12
14
Maximum buffer size [packets]
Figure 7. The minimum between the achievable rates of the source-relay link
FD and the relay-destination link for the fast-RSI scenario. The case of IRD
FD -maximizing precoders are considered.
maximizing and ISR
18
0
4.5
5
5.5
6
Figure 8. Average achievable rate of the source-relay link for the fast-RSI
FD -maximizing and I FD -maximizing precoders. The average
case under IRD
SR
of the minimum between the achievable rate of the source-relay and relayFD -maximizing and I FD destination links is plotted for both cases of IRD
SR
maximizing precoders.
gain is more than 2866% when the buffer’s maximum size is
Qmax ≥ 3 packets. We also plot an upper bound which is the
case when the relay always has data packets and sends them
to the destination. As shown in Fig. 9, the buffer-aided scheme
outperforms the conventional FD scheme and it is closer to the
upper bound. The throughput gap between the upper bound
and the buffer-aided FD scheme is 6% for Qmax ≥ 2 packets.
In Fig. 10, we plot the throughput in bits/sec/Hz versus the
transmission rate R. The throughput in bits/sec/Hz is given
by µd × R. The throughput in bits/sec/Hz increases with R
until a peak is reached. This is expected since the throughput
in packets/slot, given by µd , is monotonically non-increasing.
Thus, multiplying µd by R results in a peak at some R. After
that, the throughput decreases until it reaches zero. The value
of R that maximizes the throughput for the buffer-aided FD
case is 2.5 bits/sec/Hz. The figure also shows the significant
gain of our scheme relative to the conventional FD case. To
Figure 9. Throughput versus the maximum buffer size at the relay, Qmax .
show the impact of the RSI variance, we plotted the cases
2
2
of σRR
= 0 dB and σRR
= −10 dB. The buffer-aided FD
scheme does not depend on the RSI since it can be completely
canceled as it was shown in the analytical proof in Appendix
A and verified here through simulations. On the other hand,
the conventional FD scheme suffers from self-interference and
the throughput increases with decreasing RSI variance.
Finally, we demonstrate the impact of the number of antennas M on the system’s throughput in Fig. 11 for two different
values of R, i.e., R = 1 and R = 6 bits/sec/Hz. It can be seen
that the throughput is monotonically nondecreasing with M .
When R = 1 bits/sec/Hz and R = 6 bits/sec/Hz, the throughput is almost equal to 1 packet/slot, which is the maximum
value for the system’s throughput, for M ≥ 2 and M ≥ 3,
respectively. Increasing R increases the outage probabilities of
the communications’ links and, hence, degrades the throughput
measured in packets/slot.
VII. C ONCLUSIONS
AND
F UTURE W ORK
We derived closed-form expressions for the achievable rates
of the communications links in a buffered FD wireless relay
network under the two scenarios of slow and fast RSI. We
showed that, when the relay is equipped with a buffer, the
impact of slow RSI can be completely eliminated in the time
slots when the buffer is non-empty since the relay transmits
a known codeword that is different from the source. That
is, when the buffer is non-empty, the achievable rate of the
source-relay link in the FD mode is equal to the achievable rate
of the source-relay link without RSI. For fast RSI, we showed
that the achievable rate of the source-relay link is degraded
due to RSI and the degradation was quantified analytically.
We designed two precoders that can be used at the relay,
FD
FD
namely, the ISR
-maximizing and IRD
-maximizing precoders.
We derived the closed-form expressions for the achievable
rate of the source-relay and relay-destination links under each
precoder. For the fixed-rate transmission scenarios, when the
RSI is slow and the block size is large, we proposed an optimal
scheme that maximizes the throughput, which is the number of
packets received at the destination per time slot. Our numerical
10
1
1.6
Buffer-aided FD
Conventional FD
2
=-10
R-R
1.2
Throughput [packets/slot]
Throughput [bits/sec/Hz]
1.4
0.9
dB
Conventional FD
1
2
=0
R-R
0.8
dB
0.6
0.4
0.2
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0
1
2
3
4
5
6
7
8
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
Figure 10. Throughput in bits/sec/Hz versus the transmission rate, R.
Figure 11. Throughput in packets/slot versus the number of antennas, M .
results showed that the throughput gain of our proposed bufferaided FD scheme is substantial relative to the conventional FD
scenario.
A possible future extension of this work is to consider the
case of relay selection and study the gain of the buffers in such
cases under FD constraints. Moreover, the multi-user scenario
where multiple source nodes are communicating with a single
or a set of relay nodes can be investigated.
Consider the eigendecomposition for HSR HH
=
SR
QSRΛ SR QH
where
QSR
is
unitary
and
SR ,
ΛSR
is
diagonal.
We
can
write
the
term
HSR HH
as
QSRΛSR QH
=
SR ⊗ In
SR ⊗ In
ΛSR ⊗ In ) QH
(QSR ⊗ In ) (Λ
.
Hence,
the
achievable
⊗
I
n
SR
rate expression can be rewritten as in (37) at the top of next
page where ηv denotes the v-th eigenvalue of HSR HH
SR .
P ROOF
A PPENDIX A
P ROOF OF P ROPOSITION 1
Starting from Eqn. (6), the Hermitian of the vec{YR } is
given by
vec{YR }H = vec{XS }H HH
SR ⊗ In
H
ǫR }H .
+ vec{H⊤
IM ⊗ Ψ H XH
RR }
R +vec{ǫ
THAT THE
A PPENDIX B
O PTIMAL P RECODER
R ANKq
-1
FOR SLOW-RSI IS
σ2
RR
Let γv = 1 + P̃κSRηv and Γ =
κR XR . Note that γv > 0.
The achievable rate expression in (8) becomes
M
1X
ΨΨ H ΓH
log2 det γv In + ΓΨ
n v=1
ΨΨ H ΓH .
− log2 det In + ΓΨ
(33)
FD
ISR
=
The expectation of vec{YR }vec{YR }H over XS and HRR
is given by
(38)
ΨΨ H ΓH = QΛ
ΛQH = Qdiag{λ1 , λ2 , . . . , λn }QH ,
Let ΓΨ
ΨΨ H ΓH
where λ1 ≥ λ2 ≥ · · · ≥ λn are the eigenvalues of ΓΨ
2
Pn
σ
P
R
ΨΨ H ΓH } = RR
with k=1 λk = Trace{ΓΨ
κR n = λtot . The
+(IM ⊗XRΨ ) Ω (IM ⊗XRΨ )H +κR InM ,
achievable
rate
is
thus
given
by
(34)
o
n
E vec{YR }vec{YR }H = P̃S (HSR ⊗In ) (HSR ⊗In )H
where E vec{XS }vec{XS }H
= P̃S InM and Ω =
⊤
H
E{vec{H⊤
}(vec{H
})
}.
Using
the Kronecker product
RR
RR
properties,
n
o
E vec{YR }vec{YR }H = P̃S HSR HH
SR ⊗In
+(IM ⊗XRΨ ) Ω (IM ⊗XRΨ )H+κR InM .
(35)
2
If Ω = σRR
IM 2 , we can use the achievable rate expression
for Gaussian vectors to obtain
1
log2 det P̃S HSR HH
SR ⊗ In
n
2
+σRR
(IM ⊗ XRΨ ) (IM ⊗ XRΨ )H + κR InM
1
2
(IM ⊗ XRΨ ) (IM ⊗ XRΨ )H +κR InM .
− log2 det σRR
n
(36)
FD
ISR
=
FD
ISR
"
#
n
M
1 XX
log2 (γv + λk ) − log2 (1 + λk ) .
=
n v=1
k=1
By using λ1 = λtot −
rewritten as
Pn
k=2
(39)
λk , the achievable rate can be
!
!
"
n
n
M
X
X
1X
λk
λk −log2 1 + λtot −
log 2 γv +λtot −
n v=1
k=2
k=2
#
n
X
1
log2 (γv + λk ) − log2 (1 + λk ) .
+
n
k=2
(40)
FD
ISR
=
FD
The derivative of ISR
with respect to λj (j ∈ {2, 3, . . . , n}),
is given by (41) at the top of next page. If λ1 > λj , then
FD
∂ISR
∂λj < 0. Hence, λj = 0 maximizes the information rate.
That is, λ1 = λtot and λj = 0 for j ∈ {2, 3, . . . , n}. If
11
1
ΛSR ⊗ In )
log2 det P̃S (Λ
n
2
H
+σRR QSR ⊗ In IM ⊗ XRΨΨ H XH
R (QSR ⊗ In ) + κR InM
1
2
− log2 det σRR
IM ⊗ XRΨΨ H XH
R + κR InM
n
M
1
2
2
ΛSR ⊗ In ) + σRR
= log2 det P̃S (Λ
log2 det σRR
X RΨ Ψ H X H
IM ⊗ XRΨΨ H XH
R + κR In
R + κR InM −
n
n
!
n
M
2
2
P̃S
σ
σRR
1 XX
H
H
H
H
RR
log 2 1 +
ηv +
XR (j)ΦΦ XR (j) −log2 1 +
XR (j)ΦΦ XR (j)
=
n v=1 j=1
κR
κR
κR
!
n
M
P̃S
σ2
σ2
1 XX
log 2 1 +
ηv + RR |XR (j)q|2 − log 2 1 + RR |XR (j)q|2
.
=
n v=1 j=1
κR
κR
κR
(37)
M
FD
X
1
1
1
1
1
∂ISR
Pn
Pn
−
=
−
+
+
∂λj
n ln(2) v=1
γ + λtot − k=2 λk
1 + λtot − k=2 λk
γ + λj
1 + λj
"
#
M
X
1
γv − 1
1
−
=
Pn
Pn
n ln(2)
(γv + λj ) (1 + λj )
γv + λtot − k=2 λk 1 + λtot − k=2 λk
v=1
2
Pn
Pn
Pn
2
M
X
γv − 1 γv + γv λj + λj + λj − γv + γv λtot − k=2 λk + λtot − k=2 λk + λtot − k=2 λk
P
P
=
n ln(2)
γv + λtot − n
1 + λtot − n
k=2 λk
k=2 λk (γv + λj ) (1 + λj )
v=1
"
#
M
2
2
X
γ v − 1 γ v + γ v λj + λj + λj − γ v + γ v λ1 + λ1 + λ1
=
.
n ln(2)
(γv + λ1 ) (1 + λ1 ) (γv + λj ) (1 + λj )
v=1
(41)
FD
ISR
=
∂I FD
λ1 < λj , ∂λSR
> 0. In this case, λj = λtot and λi = 0 for
j
FD
all i 6= j maximizes the information rate. Therefore, ISR
is
H
H
ΨΨ Γ is λtot
maximized when one of the eigenvalues of ΓΨ
ΨΨH ΓH
and the rest are zeros. This implies that the matrix ΓΨ
should be a rank-1 matrix. Since Γ has a rank of M , ΨΨ H
and, consequently, Ψ is a rank-1 matrix.
By using the achievable rate expression, we get
A PPENDIX C
ACHIEVABLE R ATE OF THE R ELAY-D ESTINATION
C HANNEL
The received signal matrix at the destination node is given
by
By using Sylvester’s determinant identity, we have
YD = XRΨ H⊤
RD + ǫ D ,
(42)
where XR ∈ Cn×M is the data matrix transmitted by the
relay node, HRD ∈ CM×M is the channel matrix between
the relay node and the destination node, Ψ ∈ CM×M is the
data precoding matrix used at the relay node, and ǫ D ∈ CN ×M
is the noise matrix at the destination node. Writing the matrix
YD in a vector form, we have
ǫ D }.
vec{YD } = (IM ⊗ XRΨ ) vec{H⊤
RD } + vec{ǫ
(43)
The expectation of vec{YD }vec{YD }H over XR is given by
n
o
H
E vec{YD }vec{YD }H = PR HRD Ψ⊤ ⊗In HRD Ψ⊤ ⊗In
+κD InM
= PR HRD Ψ⊤ Ψ∗ HH
RD ⊗ In
+ κD InM .
(44)
PR
1
log2 det IM n +
HRD Ψ⊤ Ψ∗ HH
RD ⊗ In
n
κD
(45)
PR
HRD Ψ⊤ Ψ∗ HH
.
= log2 det IM +
RD
κD
FD
IRD
=
PR H
FD
IRD
= log 2 det IM +
HRD HRD Ψ⊤ Ψ∗ .
κD
Consider
the
eigendecomposition HH
RD HRD
QH
Λ
Q
.
Thus,
the achievable rate is given by
RD RD RD
PR
FD
.
Λ RD QRD Ψ⊤ Ψ∗ QH
IRD
= log 2 det IM +
RD
κD
(46)
=
(47)
According to Hadamard’s inequality for Hermitian posFD
itive semidefinite matrices, IRD
is maximized when
QRD Ψ⊤ Ψ∗ QH
is
diagonal.
Hence,
Ψ⊤ = QH
RD
RD E where
H
E is a diagonal matrix such that EE contains the power
fractions assigned to each data stream and its trace is equal to
Trace{ΨΨH } = M . Accordingly, the achievable rate of the
relay-destination link in (47) is rewritten as
FD
IRD
=
PR
Λ RD (v)|Ev |2 .
log2 1 +
κD
v=1
M
X
(48)
12
A PPENDIX D
P ROOF OF P ROPOSITION 3
σ2
σ2
, bv = P̃ ηRR
< a, and α =
Setting a = κRR
R
S v +κR
qH XH
X
q,
we
maximize
the
following
function
over α
R
R
J(α) =
M
X
v=1
!
log2 (1 + bv α) − log2 (1 + aα) .
(49)
∂J(α)
1
=
∂α
ln(2)
v=1
bv
a
−
1 + bv α
1 + aα
!
A PPENDIX F
OF P ROPOSITION 6
We vectorize the elements of the matrix YR in (15) to obtain
vec{YR } = (HSR ⊗ In ) vec{XS }
+ IM ⊗ X̃R Ψ̃ vec{H̃RR } + vec{ǫǫ R }.
(50)
bv − a
1
.
=
ln(2) (1 + bv α)(1 + aα)
o
n
E vec{YR }vec{YR }H = P̃S (HSR ⊗ In ) (HSR ⊗ In )H
H
2
+σRR
IM ⊗ X̃R Ψ̃ IM ⊗ X̃R Ψ̃
+κR InM ,
(55)
Since bv < a, the derivative is always negative. Hence, J(α) by assuming vec{H̃RR }vec{H̃H }H
RR
is maximized when α is minimized. Now, α = qH XH
achievable rate is thus given by
R XR q is
minimized when q is the normalized eigenvector correspond
1
FD
H
I
=
log
det
P̃
H
H
⊗
I
S
SR
n
ing to the minimum eigenvalue of XH
X
.
Substituting
the
SR
SR
2
R R
n
H
precoder matrix, we get
2
M
1X
2
log2 det P̃S ηv In + σRR
XR qqH XH
R + κR In
n v=1
!
2
− log2 det σRR
XR qqH XH
+
κ
I
R n
R
FD
=
ISR
=
M
X
log2 1 +
v=1
P̃S
ηv
κR
M
2
σRR
1X
log2 1 +
+
qH XH
X
q
R
R
n v=1
P̃S ηv + κR
!
σ2
.
− log2 1 + RR qH XH
R XR q
κR
(51)
M
1X
2
log2 det (P̃S ηv + κR )In + σRR
XR EEH XH
R
n v=1
!
2
H
H
− log2 det σRR XR EE XR + κR In .
(52)
Taking the data signal component as a common factor, we get
M
1X
n log2 P̃S ηv + κR
n v=1
2
σRR
XR EEH XH
+ log2 det In +
R
P̃S ηv + κR
!
2
− log2 det σRR
XR EEH XH
.
R + κR In
FD
ISR
=
(53)
By using Sylvester’s determinant identity, we get the expression in (10).
=
2
σRR
InM . The
+σRR IM ⊗ X̃R Ψ̃ IM ⊗ X̃R Ψ̃
+ κR InM
H
1
2
IM ⊗ X̃R Ψ̃ IM ⊗ X̃R Ψ̃ +κR InM .
− log2 det σRR
n
(56)
Using the same matrix eigendecomposition for HSR HH
SR as
in the slow-RSI case, the achievable rate expression in (56)
can be rewritten as:
FD
ISR
=
A PPENDIX E
P ROOF OF P ROPOSITION 4
When the optimal precoder that maximizes the achievable
rate of the relay-destination channel derived in Appendix C is
used, Ψ is the product of a unitary and a diagonal matrix.
Assuming that the matrix Ψ = EQH
RD is full rank, the
achievable rate expression of the source-relay link can be
rewritten as
FD
ISR
=
(54)
The expected value of vec{YR }vec{YR }H is
Taking the first derivative with respect to α, we get
M
X
P ROOF
1
ΛSR ⊗ In )
log2 det P̃S (Λ
n
H
2
+σRR
QH
IM ⊗ X̃R Ψ̃ X̃R Ψ̃
(QSR ⊗ In )
SR ⊗ In
+κR InM )
H
1
IM ⊗ X̃R Ψ̃ X̃R Ψ̃
+ κR InM
− log 2 det
n
H
1
2
ΛSR ⊗In )+σRR
IM ⊗ X̃R Ψ̃ X̃R Ψ̃
= log2 det P̃S (Λ
n
+κR InM )
H
M
2
−
log2 det σRR
X̃R Ψ̃ X̃R Ψ̃ +κR In .
n
(57)
After simplifications, we get the expression in (16).
A PPENDIX G
P ROOF OF P ROPOSITION 7
FD
By assuming the ISR
-maximizing
precoder of the slow-RSI
√
case, which has the form Φ = M qqH , the information rate
of the source-relay link is given by
n
M
2
P̃S
σRR
1 XX
H
H
log2 1+
ηv +
XR (j)ΦΦ XR (j)
n v=1 j=1
κR
κR
!
2
σRR
H
H
− log2 1 +
XR (j)ΦΦ XR (j)
κR
n
M
P̃S
σ2
1 XX
log2 1 +
ηv + RR |XR (j)q|2
=
n v=1 j=1
κR
κR
!
σ2
− log2 1 + RR |XR (j)q|2
.
κR
(58)
FD
ISR
=
13
This can be simplified as follows
P̃S
log 2 1 +
ηv
κR
v=1
2
σRR
n
M
1 XX
κR
|XR (j)q|2
log 2 1 +
+
S
n v=1 j=1
1 + κP̃R
ηv
!
2
σRR
2
.
|XR (j)q|
− log2 1 +
κR
FD
ISR
=
M
X
(59)
Since q ∈ CM×1 is unit norm and XR (j) ∈ C1×M
is a complex Gaussian random vector with i.i.d. elements, XR (j)q is a Gaussian random variable with
the same statistics as any element in XR (j). From
the strong law of large numbers, when n is large,
!
2
σRR
PM Pn
κR
1
2
the term n v=1 j=1 log2 1 + P̃S |XR (j)q|
−
1+ κ ηv
R
!
σ2
log2 1 + κRR
converges to its statistical mean.
|XR (j)q|2
R
Accordingly, when n → ∞, the achievable rate is given by
(17).
A PPENDIX H
P ROOF OF P ROPOSITION 8
FD
Assuming that the IRD
-maximizing precoder of the slowRSI case, which has the form Φ = EQ∗RD , is used by the
relay, the information rate of the source-relay link in (61) for
the fast-RSI case is rewritten as
n
M
P̃S
σ2
1 XX
log2 1 +
ηv + RR XR (j)EEH XH
R (j)
n v=1 j=1
κR
κR
!
2
σRR
H
H
.
XR (j)EE XR (j)
− log 2 1 +
κR
(60)
FD
ISR
=
Assuming that Ei is the i-th element on the main diagonal
of E, the information rate is thus given by
FD
ISR
=
n
M
1 XX
log2
n v=1 j=1
− log2
2
σRR
M
X
i=1
2
P̃S ηv +σRR
M
X
i=1
2
2
2
|XR,i (j)| |Ei | +κR
2
|XR,i (j)| |Ei | + κR
!!
P̃S
ηv +
log2 1 +
κR
v=1
2
PM
σRR
2
2
M
n
1 XX
i=1 |XR,i (j)| |Ei |
κR
log2 1 +
+
S
n v=1 j=1
1 + κP̃R
ηv
!!
M
σ2 X
|XR,i (j)|2 |Ei |2
.
− log2 1 + RR
κR i=1
=
M
X
!
(61)
From
strong law of large numbers,
we can approximate
2 theP
σRR
M
2
2
log2 κR
i=1 |XR,i (j)| |Ei | + 1 as follows
M lim
n→∞
n log
X
2
j=1
→ M E log2
2
σRR
κR
2
σRR
κR
2
|X
(j)|
+1
R,i
i=1
PM
n
X (j)+1
,
(62)
PM
2
2
where X (j) =
i=1 |XR,i (j)| |Ei | is the weighted-sum
of exponentially-distributed random variables and is equal
in distribution to a scaled Chi-squared with 2M degrees of
freedom when |Ei |2 = 1/M (i.e., the case of equal power
allocation to the data streams). Note that E is assumed fixed
over the entire codeword for a given channel realization and,
hence, {X (j)}nj=1 are i.i.d. random variables, which is true
since all X (j) are identical.
R EFERENCES
[1] T. Cover and A. E. Gamal, “Capacity theorems for the relay channel,”
IEEE Trans. inf. theory, vol. 25, no. 5, pp. 572–584, 1979.
[2] B. Rankov and A. Wittneben, “Spectral efficient protocols for halfduplex fading relay channels,” IEEE J. Sel. Areas Commun., vol. 25,
no. 2, 2007.
[3] P. Xu, Z. Ding, I. Krikidis, and X. Dai, “Achieving optimal diversity
gain in buffer-aided relay networks with small buffer size,” IEEE Trans.
Veh. Technol., vol. 65, no. 10, pp. 8788–8794, Oct 2016.
[4] A. E. Shafie, M. G. Khafagy, and A. Sultan, “Optimization of a relayassisted link with buffer state information at the source,” IEEE Commun.
Lett., vol. 18, no. 12, pp. 2149–2152, Dec 2014.
[5] D. Qin and Z. Ding, “Transport capacity analysis of wireless in-band
full duplex ad hoc networks,” IEEE Trans. Commun., vol. 65, no. 3, pp.
1303–1318, March 2017.
[6] M. Mohammadi, B. K. Chalise, H. A. Suraweera, C. Zhong, G. Zheng,
and I. Krikidis, “Throughput analysis and optimization of wirelesspowered multiple antenna full-duplex relay systems,” IEEE Trans.
Commun., vol. 64, no. 4, pp. 1769–1785, April 2016.
[7] C. Zhong, H. A. Suraweera, G. Zheng, I. Krikidis, and Z. Zhang,
“Wireless information and power transfer with full duplex relaying,”
IEEE Trans. Commun., vol. 62, no. 10, pp. 3447–3461, Oct 2014.
[8] H. A. Suraweera, I. Krikidis, G. Zheng, C. Yuen, and P. J. Smith, “Lowcomplexity end-to-end performance optimization in MIMO full-duplex
relay systems,” IEEE Trans. Wireless Commun., vol. 13, no. 2, pp. 913–
927, February 2014.
[9] I. Krikidis and H. A. Suraweera, “Full-duplex cooperative diversity with
Alamouti space-time code,” IEEE Wireless Commun. Lett., vol. 2, no. 5,
pp. 519–522, October 2013.
[10] L. Samara, M. Mokhtar, O. Ozdemir, R. Hamila, and T. Khattab,
“Residual self-interference analysis for full-duplex OFDM transceivers
under phase noise and I/Q imbalance,” IEEE Commun. Lett., vol. 21,
no. 2, pp. 314–317, Feb 2017.
[11] M. Mokhtar, N. Al-Dhahir, and R. Hamila, “OFDM full-duplex DF
relaying under I/Q imbalance and loopback self-interference,” IEEE
Trans. Veh. Tech., vol. 65, no. 8, pp. 6737–6741, Aug 2016.
[12] T. Riihonen, S. Werner, and R. Wichman, “Hybrid full-duplex/halfduplex relaying with transmit power adaptation,” IEEE Trans. Wireless
Commun., vol. 10, no. 9, pp. 3074–3085, 2011.
[13] I. Krikidis, G. Zheng, and B. Ottersten, “Harvest-use cooperative networks with half/full-duplex relaying,” in IEEE Wireless Communications
and Networking Conference, April 2013, pp. 4256–4260.
[14] M. G. Khafagy, A. El Shafie, A. Sultan, and M.-S. Alouini, “Throughput
maximization for buffer-aided hybrid half-/full-duplex relaying with selfinterference,” in IEEE International Conference on Communications,
2015, pp. 1926–1931.
[15] D. Bliss, T. Hancock, and P. Schniter, “Hardware and environmental
phenomenological limits on full-duplex MIMO relay performance,” in
IEEE the Asilomar Conference on Signals, Systems, and Computers,
2012.
[16] T. Riihonen, S. Werner, and R. Wichman, “Hybrid full-duplex/halfduplex relaying with transmit power adaptation,” IEEE Trans. Wireless
Commun., vol. 10, no. 9, pp. 3074–3085, 2011.
[17] ——, “Mitigation of loopback self-interference in full-duplex MIMO
relays,” IEEE Trans. Signal Process., vol. 59, no. 12, pp. 5983–5993,
2011.
[18] B. P. Day, A. R. Margetts, D. W. Bliss, and P. Schniter, “Full-duplex
MIMO relaying: Achievable rates under limited dynamic range,” IEEE
J. Sel. Areas Commun., vol. 30, no. 8, pp. 1541–1553, 2012.
[19] G. Liu, F. R. Yu, H. Ji, V. C. Leung, and X. Li, “In-band full-duplex
relaying: A survey, research issues and challenges,” IEEE Commun.
Surveys Tuts., vol. 17, no. 2, pp. 500–524.
[20] M. Duarte, “Full-duplex wireless: Design, implementation and characterization,” Ph.D. dissertation, Rice University, 2012.
14
[21] M. Duarte, A. Sabharwal, V. Aggarwal, R. Jana, K. Ramakrishnan,
C. W. Rice, and N. Shankaranarayanan, “Design and characterization
of a full-duplex multiantenna system for WiFi networks,” IEEE Trans.
Veh. Technol., vol. 63, no. 3, pp. 1160–1177, 2014.
[22] M. Jain, J. I. Choi, T. Kim, D. Bharadia, S. Seth, K. Srinivasan, P. Levis,
S. Katti, and P. Sinha, “Practical, real-time, full duplex wireless,” in
Annual International Conference on Mobile Computing and Networking.
ACM, 2011, pp. 301–312.
[23] Y. Fan and J. Thompson, “MIMO configurations for relay channels:
Theory and practice,” IEEE Trans. Wireless Commun., vol. 6, no. 5,
2007.
[24] R. Mo, Y. H. Chew, and C. Yuen, “Information rate and relay precoder
design for amplify-and-forward MIMO relay networks with imperfect
channel state information,” IEEE Trans. Veh. Tech., vol. 61, no. 9, pp.
3958–3968, 2012.
[25] D. Bliss, P. Parker, and A. Margetts, “Simultaneous transmission and
reception for improved wireless network performance,” in IEEE/SP
Workshop on Statistical Signal Processing, 2007, pp. 478–482.
[26] P. Lioliou, M. Viberg, M. Coldrey, and F. Athley, “Self-interference suppression in full-duplex MIMO relays,” in IEEE the Asilomar Conference
on Signals, Systems, and Computers, 2010, pp. 658–662.
[27] Y. Sung, J. Ahn, B. Van Nguyen, and K. Kim, “Loop-interference suppression strategies using antenna selection in full-duplex MIMO relays,”
in IEEE International Symposium on Intelligent Signal Processing and
Communication Systems, 2011, pp. 1–4.
[28] W. Zhang, X. Ma, B. Gestner, and D. V. Anderson, “Designing lowcomplexity equalizers for wireless systems,” IEEE Commun. Mag.,
vol. 47, no. 1, pp. 56–62, 2009.
[29] Y. Y. Kang and J. H. Cho, “Capacity of MIMO wireless channel with
full-duplex amplify-and-forward relay,” in IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, 2009,
pp. 117–121.
[30] T. Riihonen, S. Werner, and R. Wichman, “Spatial loop interference suppression in full-duplex MIMO relays,” in IEEE the Asilomar Conference
on Signals, Systems, and Computers, 2009, pp. 1508–1512.
[31] D. Choi and D. Park, “Effective self interference cancellation in full
duplex relay systems,” IET Elec. lett., vol. 48, no. 2, pp. 129–130, 2012.
[32] B. Chun and H. Park, “A spatial-domain joint-nulling method of self-
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
interference in full-duplex relays,” IEEE Commun. Lett., vol. 16, no. 4,
pp. 436–438, 2012.
N. Zlatanov, D. Hranilovic, and J. S. Evans, “Buffer-aided relaying
improves throughput of full-duplex relay networks with fixed-rate transmissions,” IEEE Commun. Lett., vol. 20, no. 12, pp. 2446–2449, 2016.
K. T. Phan and T. Le-Ngoc, “Power allocation for buffer-aided fullduplex relaying with imperfect self-interference cancelation and statistical delay constraint,” IEEE Access, vol. 4, pp. 3961–3974, 2016.
T. Riihonen, S. Werner, and R. Wichman, “Mitigation of loopback selfinterference in full-duplex MIMO relays,” IEEE Trans. Signal Process.,
vol. 59, no. 12, pp. 5983–5993, 2011.
B. P. Day, A. R. Margetts, D. W. Bliss, and P. Schniter, “Full-duplex
MIMO relaying: Achievable rates under limited dynamic range,” IEEE
J. Sel. Areas Commun., vol. 30, no. 8, pp. 1541–1553, 2012.
M. Duarte, C. Dick, and A. Sabharwal, “Experiment-driven characterization of full-duplex wireless systems,” IEEE Trans. Wireless Commun.,
vol. 11, no. 12, pp. 4296–4307, 2012.
D. Bharadia, E. McMilin, and S. Katti, “Full duplex radios,” in ACM
Special Interest Group on Data Communication Computer Communication Review, vol. 43, no. 4. ACM, 2013, pp. 375–386.
D. Korpi, T. Riihonen, V. Syrjälä, L. Anttila, M. Valkama, and R. Wichman, “Full-duplex transceiver system calculations: Analysis of ADC and
linearity challenges,” IEEE Trans. Wireless Commun., vol. 13, no. 7, pp.
3821–3836, 2014.
D. Korpi, L. Anttila, V. Syrjälä, and M. Valkama, “Widely linear
digital self-interference cancellation in direct-conversion full-duplex
transceiver,” IEEE J. Sel. Areas Commun., vol. 32, no. 9, pp. 1674–
1687, 2014.
N. Zlatanov, E. Sippel, V. Jamali, and R. Schober, “Capacity of the Gaussian two-hop full-duplex relay channel with residual self-interference,”
IEEE Trans. Commun., vol. 65, no. 3, pp. 1005–1021, March 2017.
I. Krikidis, H. A. Suraweera, S. Yang, and K. Berberidis, “Full-duplex
relaying over block fading channel: A diversity perspective,” IEEE
Trans. Wireless Commun., vol. 11, no. 12, pp. 4524–4535, 2012.
T. M. Cover and J. A. Thomas, Elements of information theory. John
Wiley & Sons, 2012.
L. Liang, W. Xu, and X. Dong, “Low-complexity hybrid precoding
in massive multiuser MIMO systems,” IEEE Wireless Commun. Lett.,
vol. 3, no. 6, pp. 653–656, 2014.
| 7cs.IT
|
1
Discrete Power control stochastic game on the
Multiple access channel.
arXiv:1710.08800v1 [cs.SY] 24 Oct 2017
Prashant N, Lakshmi Narasimhan Theagarajan∗
Indian Institute of Information Technology Sri City, India
∗ Indian Institute of Technology Palakkad, India
Email:[email protected],∗ [email protected]
Abstract—We consider the scenario where multiple users send
information to a single receiver over a common channel. Each
user is aware of only his channel fading state and has statistical
knowledge of the fading process of the other users. The fading
process of each user is modeled as ergodic finite state Markov
chains. We consider two problems. The first problem is the noncooperative game where the receiver decodes the the message of
each transmitter by considering the signals of the other users as
noise. In the second problem, the receiver employs a successive
interference decoder to decode the messages of the users. The
two problems are posed as stochastic games, the first one as a
non-cooperative game where the aim is to maximize the expected
time average throughput of each user in a selfish manner. In the
second problem, the aim is maximize the expected time average
sum throughput of the system. We show the existence of Nash
equilibrium policies for both the games and give a best response
algorithm to find them for both the games. Our analysis begins
first by showing that the second game is the potential game of
the first one, hence the equilibria policies for both the games are
same. We then analyze the first game at large but finite number
of users. We show there exist a natural number such that when
the number of users exceed this number, then the equilibrium
policies are unique. Furthermore the equilibrium policies of the
users are invariant as long as the number of users remain above
the latter threshold. These policies are then shown to achieve the
global maximum of the expected time average sum throughput of
the second game. We give a algorithm which is linear in number
of users to compute these threshold equilibrium policies.
Index Terms—Multiple Access Channel, Stochastic Games,
Nash equilibrium, Markov Decision process, Potential Games,
Dominating strategies, Best response sets, Best response algorithm.
I. I NTRODUCTION
The Multiple access channel (MAC) is a abstract model of
the scenario where more than one transmitter simultaneously
sends information to a single receiver. In the wireless setup, the
time varying fading process and knowledge of channel state
information affects the throughput of the users. Resources such
as power, time and rate have to be dynamically allocated to
improve the throughput of the users. In general, this problem
can be approached in two ways, a centralized scheme or a
distributed scheme. The centralized scheme provides the best
possible throughputs for each transmitter while requiring a
huge overhead. On the other hand a distributed implementation
provides the flexibility required for a practical system, but
gives less throughput in comparison to the former. In this
work we adopt a game theoretic approach to analyze the above
problem. This approach is a distributed scheme and here the
users are modeled as rational and selfish players interested in
maximizing their utilities.
Another problem considered in the distributed approach
is the availability of each users channel statistics across
other users. A centralized approach assumes that all such
information is available to every user at each time instant.
On the other hand, a distributed scheme may or may not
include such assumptions and hence vary from users having
complete channel state information of other users to knowing
only channel statistics of other users. These assumptions
significantly affect system modeling as well as the throughputs
achieved by various users in the system. We now provide a
brief survey on the related works.
The centralized problem where each user’s transmitter and
receiver knows every other users channel state was considered
in [1], here the authors considered the problem of finding resource allocation schemes which are information theoretically
optimal. The authors of [1] considered the problem finding
all possible achievable rates of all users under fixed average
power constraints. This set of achievable rates was shown to
be a polymatroid and using results in [2], the authors showed
that the solution of any rate and power allocation schemes
over this set can be computed in polynomial time by using
a greedy algorithm. In particular they showed the scheme of
assigning the channel to the user with the highest channel
gain maximizes the ergodic sum capacity. [4] considered the
same problem in a game theoretic setup. They assumed full
channel state information of each user is available to every
user and showed that when each user maximizes his own rate
in a selfish manner, while treating the other user signal as
interference, assigning the channel to the user with the highest
channel gain is a Nash equilibrium. The work in [5], [6]
considered the problem where each user maximized his/her
mutual information in a selfish manner and provided sufficient
conditions for the existence of unique Nash equilibrium. Here
also the authors assumed that each user had access of channel
gain of every other user.
The work in [7] formulated the above problem in a Bayesian
scenario, that is each user is only aware of his/her channel gain
and has statistical knowledge of the channel gains of the other
user. They showed there exist a unique Nash equilibrium. In
this model, the power policies was modeled as a continuous
variable while the channel could only take values from a finite
set. A similar setup in which the power was considered discrete
was considered in [9], [22]. In fact our model is similar to
2
these papers. In [9] the authors showed the existence of Nash
equilibria for the game where each user maximizes his/her own
Shannon capacity while treating the other users as noise. An
algorithm was provided for computing the equilibrium policies
for two users while a best response algorithm is provided to
compute any such equilibrium policy for any number of users.
A proof was provided for the correctness of the two user
algorithm while the best response algorithm was shown to
converge only numerically. We note that a method of proving
that the algorithm in [9] converges was provided in [22]. We
provide the whole proof for convergence of the algorithm in
our work.
We note that in all of these models the number of users was
fixed and algorithms were provided to compute or converge to
point of interest. However as the number of users increases, the
algorithms provided above become computationally infeasible.
To overcome this problem new models where the number
of users can be very large have become quite popular in
literature [12], [13]. In these models the numbers of agents
are significantly large, such that, each user can be modeled as
a rational agent playing against a continuum of agents. The
infinite interaction between the users is now captured by means
of a parameter called the mean field. Each user is now modeled
to interact with other users only through the mean field. This
modeling assumption significantly reduces the complexity by
converting the large user problem into a simpler optimization
problem. However in many cases the equilibria policies arising
from this type of modeling becomes solution to difficult
forward - backward partial differential equations. In general,
only solutions from numerical algorithms can be given for
such problems. It is to be noted, that in all mean field models
an assumption of symmetric agents is assumed, that is the
players in these type of games are interchangeable.[12], [13].
In the wireless context such modeling has been considered in
[14], [15], [16], [17], [18], [19]. However we focus specifically
on [19], as this model is similar to the one we consider.
In this model, each user maximizes his/her own signal to
interference and noise ratio where the signal from the other
users is treated as noise. The authors show that as number
of users tend towards infinity, the problem can be analyzed
via a mean field setup. Interestingly they show via numerical
simulations that the optimal equilibria policy at large number
of users outperforms any well known equilibrium policy. We
prove a similar result in our work.
In this paper, we study the model where the receiver knows
the channel states of all the users, while each transmitter only
knows her/his channel state information. We assume that the
each user is aware of channel statistics of the other users.
We consider two scenarios, one where the receiver uses a
standard matched filter decoder and in the other one, the
receiver employs a successive interference decoder. In the first
scenario while decoding the signal of a user, the receiver
treats the message of other users as noise. Here each user
selfishly maximizes his/her own Shannon capacity. In the
second scenario a prespecified ordering of the users is made
available to each transmitter and here each user maximizes the
multiple access channel sum rate. We assume that the channel
gain as well as the feasible power levels are finite. Our results
are as follows.
1) In both scenarios we show the existence of Nash equilibria and prove the convergence of the best response
algorithm given in [9], [11]. We also show that the
utility function of the second problem is the potential
function[21] corresponding to the utilities of the first
one. Hence we need to analyze only one of the problem.
Henceforth we analyze the first scenario only.
2) We then consider the case as the numbers of users
becomes large under the assumption of finite sets of
symmetric users. This assumption implies that all users
belonging to a particular set are interchangeable and
there are only finite number of such sets. We prove that
there exist a finite threshold number of users beyond
which the Nash equilibrium of each user in each set
is unique and invariant. We refer to this policy as a
infinitely invariant policy. Hence as long as the number
of users remains greater than this threshold, the Nash
equilibrium policy of each user in the system does not
change. This implies we need to compute the equilibria
policies only for a finite number of users using the
best response algorithm. We also show that under this
assumption the best response algorithm has linear complexity in number of users.
3) We then show at sufficiently large number of users, the
equilibrium policy is a strictly dominating policy among
the set of all rational policies of each user. from this
result, we derive as a corollary that at sufficiently large
number of users, in the scenario where each user knows
only his, her channel gain , the infinitely invariant policy
is the optimal policy which maximizes the ergodic sum
rate capacity of the MAC channel.
4) We then provide a sequence of optimization problems
which computes the infinitely invariant policy. We show
that this policy maximizes time average snr of the user
and then minimizes the variance of the time average snr.
By using structural results from Markov decision theory,
we show that this policy arises from a greedy algorithm
which has linear complexity in number of channel states
as well as number of power levels.
Our paper is structured as follows. The system model and
problem formulation along with the various required definitions are provided in section II and section III respectively. We
then provide the existence of Nash equilibria and best response
algorithm used to compute them in section IV. Section V
considers the case when interference is very large. This is
an intermediate step in the analysis of the games at large
number of users. Here we provide the algorithm used in the
computation of the infinitely invariant policy. In section VI
we consider the scenario where the numbers of users is very
large and show the existence of a infinitely invariant policy
in section VII. In section VIII we provide numerical results
verifying the same.
II. S YSTEM
MODEL
We consider a discrete time scenario involving the multiaccess fading channel where a set N = {1, 2, 3, · · · , N } of
3
users access a base-station simultaneously over each time-slot.
The fading channel observed by user i at time n is modeled as
an finite state ergodic Markov chain hi [n] which takes values
from a finite set Hi . The results in this work remain the same
for any stationary and ergodic fading process. Any element
of this set is denoted as Q
hi . The set of channel states of all
users is denoted as H = i∈N Hi and h := (hi , h2 , · · · , hN )
denotes any element of this set.
′
The transition probability of user i going from channel hi
to hi is equal to λh′ ,hi . The Markov chains hi [n] of users are
i
assumed to be processes independent of each other. We assume
that the fading process of user i is known perfectly by both
the user as well as the base-station but is unknown to the other
users. We denote the row vector πi := (πi (hi ) : hi ∈ Hi ) as
the vector of steady state probabilities of user i corresponding
to the Markov chain hi [n]. At each time n user i invests power
pi [n] from the finite set Pi . Any element of this set is denoted
as pi . It is assumed that sometimes the user will not transmit,
hence the action pi = 0 belongs to the set of power Pi of
user Q
i. The set of power action of all users is denoted as
P = i∈N Pi and p := (pi , p2 , · · · , pn ) denotes any element
of this set.
We assume that the base station has complete information
about the power level pi [n] utilized by each user i for every
time-slot n. On the other hand, users other than user i only
know the set of channel states Hi , power levels Pi of user i.
We also assume that users other than user i know perfectly the
channel statistics of user i. is to be noted that, this allows users
other than user i to compute the steady state probabilities πi
of user i.
We associate with each user i, a real valued function ti
which signifies the reward ti (h, p) of user i when the the
channel states and power actions of all the users are h and
p respectively. It is assumed that this utility/reward function
is known to all users as well as the base-station. In this work
we consider the following reward functions.
1) The base-station employs a regular matched filter to
decode the message of a user by considering the signals
from other user as noise. In such a case a upper-bound
for the maximum rate, user i can achieve when the the
channel states and power actions of all the users are h
and p is,
hi p i
n
.
(1)
ti (h, p) := log2 1 +
PN
N0 + j=1,j6=i hj pj
2) The base-station uses a successive interference cancellation filter. Given an ordering scheme on the the set
of users N , the received packet of a user i is decoded
after canceling out the decoded transmission of other
users lying below user i in the predefined order from the
received transmission. We assume perfect cancellation
of the decoded signal from the received transmission
known to all users at each time slot. If the users want to
maximize the aggregated throughput in a decentralized
manner, then the utility function of each user is,
PN
j=1 hj pj
s
.
(2)
t (h, p) = log2 1 +
N0
It is to be noted that the objective function here is the
same for all users.
III. P ROBLEM F ORMULATION
In this section, we formulate the game theoretic problem
arising from the system model and the two reward functions.
We do so by first considering the class of stationary policies.
It is to be noted that, this problem is going to be analyzed a
constrained Markov game with independent state information.
The interested reader can refer to [11] for motivation about
the forthcoming definitions.
A. Stationary policies:
Let M(G) denote the set of probability measures over a set
G. A stationary policy for user i is a function ui : Hi 7−→
M(Pi ), such that ui (·/hi ) ∈ M(Pi ). The value ui (pi /hi )
represents the probability of user i using power pi when her/his
channel state is hi . We denote the set of stationary policies
for user i as Ui .
B. User and Order Notations:
Given a set Gi associated with user i we refer to an element
of the set as gi . The global set of all such elements of all
users is defined as
QNthe Cartesian product of these sets and is
denoted as G = i=1 Gi . Any element of this set is denoted
as g = (gi , · · · , gN ). The set of all elements other than the
element of user i is the Cartesian product of these sets other
than
QN the set Gi associated with user i and is denoted as G−i =
j=1,j6=i Gj . Any element of this set is denoted as g−i =
(gi , · · · , gi−1 , gi+1 , · · · , gN ). Given any set G we denote |G|
as the cardinality of set G. We use capital letters to denote any
random variable or matrix occurring in this work. It will be
clear from context, whether the occurring notation in capital
letter is a matrix or a random variable. We denote R, N and
Z as the set of real numbers, natural numbers and integers
respectively. The set Z+ denotes the set of strictly positive
natural numbers. Let f (·) and g(·) denote two functions on
the set of natural numbers. If the function f is asymptotically
bounded above (below) the function, we denote it as,
f (n) = O(g(n)) (f (n) = Ω(g(n)))
C. Expected time-average reward and cost:
Let βi represent a initial probability distribution on the set
of states Hi of user i and β := (βi , β2 , · · · , βN ) represent the
initial joint probability distribution of all users on the global
set of states H. Given a multi-policy u for all players, we
can define a probability measure λβu which determines the
distribution of the stochastic process (h[n], p[n]) of states and
actions. The expectation due to this distribution is denoted as
Eβu . We define the time-average expected rate Ti (u) of user i
as a function of the multi-policy u of all users. It is calculated
as
Ti (u) := lim inf
T →∞
T
1 X β
E (ti (h[n], p[n])).
T n=1 u
(3)
4
Similarly let λβui denote the probability measure with respect
(w.r.t) to the policy ui of user i and Eui denotes the expectation
due to this distribution. The time average expected power
Pi (ui ) used by user i is now calculated as
T
1 X βi
Pi (ui ) := lim sup
Eui (pi [n]).
T →∞ T n=1
(4)
D. Game Theoretic formulation
Let P i denote a constraint on the time average expected
power of user i. A policy ui is called i−feasible if it satisfies
Pi (ui ) ≤ P i . A multi-policy u is called feasible if it is ifeasible for all users i ∈ N. User i chooses a i-feasible policy
so as to maximize his reward Ti (ui , u−i ). However Ti (u)
depends on the stationary policy of other users also leading to
a non-cooperative game.
Let [u−i , vi ] denote the multi-policy where, users k 6= i
use stationary policy uk while user i uses policy vi . We now
define the Constrained Nash Equilibrium (CNE).
Definition 1. A multi-policy u ∈ U is called a CNE if for
each player i ∈ N and for any vi ∈ Ui such that [u−i , vi ] is
feasible,
Ti (u) ≥ Ti (u−i , vi ).
(5)
We denote the above formulation as a constrained Markov
game (CMG) [11],
"
#
Γ = N , (Hi )i∈N , (Pi )i∈N , (Λi )i∈N , (ti )i∈N , (P i , )i∈N ,
A CMG is called an identical interest CMG if the immediate
reward function ti is same for all user. The games Γn and
Γs represent the Markov games when the reward function for
user i are tni and tsi respectively. In the next two subsections
we illustrate how to compute the best response sets associated
with this game.
E. Calculation of Optimal response
Given the stationary policy ui , we define the occupation
measure zi (hi , pi ) of user i for each channel-power pair
(hi , pi ) as
zi (hi , pi ) = π(hi ) · ui (pi |hi ).
(6)
The occupation measure zi (hi , pi ) for user i is the steadystate probability of the user being in state hi and using power
pi . Define the immediate reward for user i, when user i has
channel hi and employs power pi and other users use multipolicy u−i as,
X
X Y
z
Ri −i (hi , pi ) =
[ zl (pl , hl )]ti (h, p),
(7)
h−i ∈Hi p−i ∈P−i l6=i
Then the time-average expected reward of user i under the
multi-policy u are:
X X z
Ti (u) =
(8)
Ri −i (hi , pi )zi (hi , pi ).
hi ∈Hi pi ∈Pi
Similarly the time-average expected power of user i is
X X
Pi (ui ) =
pi zi (xi , ai ).
(9)
hi ∈Hi pi ∈Pi
As we can compute occupation measures from stationary
policies and conversely using (6), we restrict ourselves, in
this work to finding occupation measures. Hence from now
on, without loss of generality, we denote time-average expected reward and time-average expected power of user i as
Ti (u) := Ti (z) and Pi (ui ) := Pi (zi ). We also use the term
occupation measure and stationary policy interchangeably,
even though from now on, we will only be computing the
occupation measures zi of each user i.
F. Best response of player i
Let all users other than user i use the multi-policy u−i . Then
user i has an optimal stationary best response policy which is
independent of the initial state x0 [11]. Let the set of optimal
stationary policies of user i be denoted as Bi (z−i ). We can
compute the elements of this set from the following Linear
program. Find zi∗ ∈ Zi that satisfies,
{Ti (zi∗ , z−i ) − Ti (zi , z−i ) ≥ 0 :, ∀zi ∈ Zi },
(10)
where the polyhedron Zi associated with user i is represented
as,
Zi =
X
(
zi ∈ R|Hi |×|Pi | : Pi (zi ) ≤ P i ,
zi (pi , hi ) = πi (hi ), ∀ hi ∈ Hi ,
pi ∈Pi
)
zi (pi , hi ) ≥ 0, ∀ (hi , pi ) ∈ Hi × Pi .
We note that Zi is the set of all occupation measures of
user i corresponding to i-feasible stationary policies. The
set of allQoccupation measures of all users is represented
as Z := i∈N Zi . We denote E(Zi ) as the set of extreme
points/vertices of the polyhedron Zi . It is to noted that, the
set of occupation measures Zi of each user i is nonempty
as the policy of never transmitting satisfies all the constraints
describing the polyhedron Zi .
IV. E XISTENCE AND C OMPUTATION
E QUILIBRIUM
OF
NASH
In this section we show how to compute a constrained
Nash equilibrium for these games. The existence of equilibria
for these games was shown in theorem 2.1, [11] under the
assumption that the Markov process resulting from policy of
each user is irreducible and has a single ergodic class. As
in our models the policy does not affect the Markov chain
transition probability, under the assumption of ergodicity of
the channel state process, the result follows. The existence of
equilibria for the game Γc was also shown in theorem 2, [9].
5
We state and prove the convergence of the best response
algorithm proposed in [9] used to compute the equilibrium
policies for game Γc . To show this result, we first extend the
idea of potential games developed in [21] for normal form
games to the constrained Markov games in this paper. We first
state the definition of potential function and potential games
for the games considered in this work.
A. Potential games
Definition 2. A function T̂ : Z 7−→ R is called a potential
function for the constrained Markov game Γ if for all users
i ∈ N , for all multi-policy z−i of all users other than user i
and for any pair of policies (zi , ẑi ) of user i,
Ti (zi , z−i ) − Ti (ẑi , z−i ) = T̂ (zi , z−i ) − T̂ (ẑi , z−i ).
(11)
The Markov game where each user use the expected time
average reward T̂ is considered as the Markov potential game
corresponding to the Markov game Γ.
It is to be noted that the expected time average reward T̂
for a Markov potential game is the same for all users, that is, a
Markov potential game is an identical interest game. One can
also observe from the above definition that the equilibria for
the Markov potential game are the same as the equilibria for
the corresponding Markov game. We now give a sufficient
condition for for verifying whether there exist a potential
function these Markov games.
Theorem 1. Suppose for a Markov game Γ there exist a
function t̂ : H × P 7−→ R such that for all users i ∈ N , for all
channel power (h−i , p−i ) tuples of all users other than user
i and any pair of channel power tuples (hi , pi ) and (ĥi , p̂i )
of user i the function ti (h, p) satisfies,
ti (hi , pi , h−i , p−i ) − ti (ĥi , p̂i , h−i , p−i ) =
(12)
t̂(hi , pi , h−i , p−i ) − t̂(ĥi , p̂i , h−i , p−i ).
Then there exist a potential function for the Markov game
Γ. Furthermore, the constrained Markov game where the
immediate reward of each user is t̂ is the Markov potential
game corresponding to the Markov game Γ.
Proof: The proof of this theorem is a consequence of
lemma 2.10 in [21]. Let us consider the Markov game Γ̂, where
user i has the immediate reward t̂. In this case we the denote
the immediate reward of user i, when when user i has channel
hi and employs power pi and other users use multi-policy z−i
z
as R̂i −i . Also denote T̂ (zi , z−i ) as the time-average expected
reward of each user for this game. The result (11) follows
from expressions (12), (7) and (8).
The next result shows that the Markov game Γs is the
Markov potential game corresponding to the Markov game
Γn .
Theorem 2. Γs is the Markov potential game of the Markov
game Γn .
Proof: The reader can verify the following identity,
tni (hi , pi , h−i , p−i ) − tni (ĥi , p̂i , h−i , p−i ) =
(13)
ts (hi , pi , h−i , p−i ) − ts (ĥi , p̂i , h−i , p−i ).
The proof then follows from theorem 1.
As the equilibria policies of the Markov potential game
and the corresponding Markov game are the same, following
theorem 2, we can restrict our analysis to the Markov game
Γn . We now state the algorithm stated in [9] used to compute
the equilibrium policies for a Markov game when it has a
potential function. The algorithm is used to compute Epsilon
constrained Nash equilibrium(ǫ-CNE) which we define below.
Definition 3. A multi-policy u ∈ U is called a ǫ-CNE if for
each player i ∈ N and for any vi ∈ Ui such that [u−i , vi ] is
feasible,
Ti (u) − Ti (u−i , vi ) ≥ −ǫ.
(14)
Algorithm 1 Best Response Algorithm
1: Set k = 0.
2: Find a z(0) ∈ Z.
3: Set ǫ > 0.
4: while z(k + 1) 6= z(k) do
5:
for i = 1 : |N |, do
6:
ẑ−i (k)
←
z1 (k + 1), .., zi−1 (k + 1)
, zi+1 (k), .., z|N | (k) .
7:
Set zi ∈ Bi (ẑ−i (k)).
8:
if |Ti (zi , ẑ−i (k))− Ti (zi (k), ẑ−i (k))| ≤ ǫ, then
9:
Set zi (k + 1) ← zi (k).
10:
else
11:
zi (k + 1) ← zi .
12:
k ←k+1
We analyze the Best response algorithm and prove its
convergence in two cases. With the error approximation ǫ = 0,
the algorithm is shown to terminate in finite time only if
the best response in the algorithm is computed using the
simplex algorithm. We assume here that the simplex algorithm
computes a vertex solution of a linear program, that is the
solution of the linear program is a extreme point of the
corresponding polyhedron. In the case the error ǫ is strictly
positive we show that in finite number of iterations the best
response algorithm computes an ǫ− CNE.
Theorem 3. 1) Let the error approximation ǫ = 0. If the
best response policy for each user in the best response
algorithm is a vertex of the polyhedron Zi , then the
algorithm terminates in finite number of iterations in a
CNE.
2) Let the error approximation ǫ > 0. Then the Best
response algorithm terminates in finite number of iterations in an ǫ− CNE.
The proof of this theorem is provided in Appendix A. It
is noted that the above theorem shows the existence of a
CNE which are the vertices of the set of feasible occupation
measures Zi for each user i. In particular, this implies the
existence of CNE policies for games Γs and Γn which are
vertices of the set of occupation measures. We also note that
the ǫ− best response will find a CNE which are close to the
vertices of feasible occupation measures.
6
It is to be noted that the best response algorithm only
guarantees convergence to a CNE, however depending upon
the ordering of users, the initial policy chosen for the algorithm
may converge to different CNE if there exist multiple CNE’s.
In the next subsection we analyze global maximizers of
identical interest games.
B. Global Maximizer of Identical Interest Markov games
Definition 4. A multi-policy z ∗ ∈ Z is the global maximizer
of an identical interest Markov games Γ with time average
reward T , if for any multi-policy z ∈ Z
T (z ∗ ) ≥ T (z).
(15)
We now state the result which shows there exist a global
maximum, whose each single policy is a vertex of the polyhedron formed by the set of feasible occupation measures.
Theorem 4. There exist a multi-policy z ∗ ∈ Z which is the
global maximizer of the identical interest Markov game Γ.
Each policy zi∗ corresponding to the multi-policy z ∗ is a vertex
of set of feasible occupation measure Zi for each user i.
Proof: As the function T (z) is continuous over a compact
set Z, there exist a global maximum point z ∗ ∈ Z. Now we
prove there exist a pure policy ẑi ∈ Zi for each user i, such
that T (ẑ) = T (z). We give the proof for two users which can
be easily extended to N users. Let ẑi denote the best response
policy of user i, which is a vertex of the polyhedron Zi , when
user 2 plays policy z2∗ . Then we have T (ẑi , z2∗ ) ≥ T( z ∗ ),
but as z ∗ is global maximizer we have T (ẑi , z2∗ ) = T (z ∗ ).
This implies (ẑi , z2∗ ) is also a global maximizer. Similarly we
can find ẑ2 which is a vertex of the polyhedron Z2 such that
T (ẑi , ẑ2 ) = T (ẑi , z2∗ ), hence we have T (ẑ) = T (z ∗ ) and each
policy zi∗ corresponding to the multipolicy z ∗ is a vertex of
set of feasible occupation measure Zi for each user i.
We can easily show that the global maximum of an identical
interest Markov game is a CNE. The above results shows that
the game Γs has a global maximizer and a CNE. In the next
subsection, we define dominating policies for Markov games.
V. T IME
AVERAGE E XPECTED REWARD AT LARGE
INTERFERENCE
We consider in this section only one user. At each time slot
n, the user i gets a one-step reward of
ti (hi [n], pi [n], I) = log2 (1 +
hi [n]pi [n]
),
N0 + I
(17)
where I is the interference faced by the user and (hi [n], pi [n])
is her/his channel state-power action pair. User i wants to
maximize her/his time average expected reward, i.e maximize,
ti (I, zi ) = lim inf
T →∞
subject to
T
1 X βi
E (ti (hi [n], pi [n], I)),
T n=1 u
pi (zi ) := lim sup
T →∞
T
1 X βi
E (pi [n]) ≤ P i .
T n=1 u1
where βi is the initial distribution of player i and zi is the
occupation measure corresponding to the stationary policy ui
computed using (6).
The optimal policy zi to the above optimization problem
considered by user i is independent of the initial distribution
βi and can be computed by the corresponding linear program
LP (i, I). Find a policy zi∗ which maximizes
o
n
(18)
ti (I, zi ) : zi ∈ Zi ,
where
ti (zi , I) :=
X
ti (hi , pi , I)zi (hi , pi ).
(hi ,pi )
C. Dominating Policies
Definition 5. Let Ẑi ⊆ Zi denote a subset of policies of all
users other than user i. A policy zi∗ for user i is called a
dominating policy over the set Ẑi if for any multi-policy z−i
in the set Ẑi of all users other than user i, we have for all
policies zi of user i,
Ti (zi∗ , z−i ) ≥ Ti (zi , z−i ).
used to analyze games Γn and Γs at large number of users.
To motivate the analysis of the games Γn and Γs at large
number of users, we first understand how the expected time
average maximizing policy of user i behaves in presence of
a very large deterministic interference. This analysis in itself
is interesting as well as provides intuition on the behavior of
the previously mentioned Markov games at large number of
users. The next section addresses this analysis.
(16)
The policy zi∗ is called strictly dominating over the set Ẑi if
the above inequalities are strict.
If the set Ẑi = Zi in the above definition, the definition
becomes the standard definition of a dominating policy. It also
follows that if there exist a multi-policy z ∗ of all users such
that the policy zi∗ of user i is a dominating policy over the set
of all feasible occupation measures Zi , then the multi-policy
z ∗ is a CNE. In case the policy zi∗ is a strictly dominating
policy, the CNE is unique. This simple observation will be
We denote Si (I) denote the set of solutions to the linear
program LP (i, I). is to be noted that the set of solutions Si (I)
is equal to the set of best response Bi (z−i ) of user i in the
Markov game Γn when users other than user i use the policy
where they do not transmit any data and the noise variance N0
in the system is equal to 1 + I. In this section, we characterize
the set of solutions Si (I) of the linear program LP (i, 1) at
large interference. To do so, we first define the notion of the
outer-limit of a sequence of sets.
Definition 6. We define the outer limit limI→∞ S i (I) of the
sequence of sets Si (I) as the interference I tends towards
infinity as,
∞ [
∞
n
\
limI→∞ S i (I) := zi ∈ Zi : zi ∈
Si (Ik ),
n=1 k=n
o
where Ik → ∞ as k → ∞ for some sequence {Ik } .
7
The outer limit contains all policies of user i which belong
to each set Si (Ik ) of solutions for some sequence Ik of
interference which tends towards infinity as the index k tends
towards infinity. The outer limit can be thought upon as the
set of all infinitely often occurring optimal solutions of the
linear programs LP (i, I) as the interference I tends toward to
infinity. We show the outer limit limI→∞ S i (I) is a nonempty
set in the next subsection.
A. Analysis as the interference I tends towards infinity.
At very large interference I, the function ti (hi , pi , I) can
be approximated as hi pi [25], [26]. By considering this
approximation the objective function ti (zi , I) of the linear
program can be approximated as,
X
(hi pi )zi (hi , pi ).
li1 (zi ) :=
respectively. Then consider stationary policies zi of user i such
that,
P
l
if rj=l π(hji )pqi ≤ P i , a = q
π(hi ),
Pr
zi (hli , pai ) := 0,
if j=l π(hji )pqi ≤ P i , a 6= q
Pr
j q
π(hli )α(pai ), if
j=l π(hi )pi > P i ,
where α(·) is a probability distribution on the set Pi which
satisfies,
π(hli )
r
X
pai α(pai )
= pi −
a=1
q
r
X
X
zi (hji , pâi )pâi .
j=l+1 â=1
Let the set Qi denote the above class of policies. To understand, why the policies in this set are structured, we observe
the following properties of policies in the set Qi .
1) The function ui (·/hki ) : Hi → Pi defined by
(hi ,pi )
We now consider the linear program (LP (1, 1)) induced by
this approximation. Find a policy zi∗ which maximizes,
o
n
(19)
li1 (zi ) : zi ∈ Zi .
The set of solutions of this linear program is denoted by the
set Si1 and any policy in this set is called a 1-sensitive optimal
solution. We now show that the outer limit limI→∞ S i (I) of
the sequence of solutions Si (I) of the linear program LP (i, I)
is a subset of the solutions of the above linear program.
Theorem 5. The outer limit limI→∞ S i (I) of the sequence of
sets Si (I) as the interference I tends towards infinity is a non
empty subset of the set Si1 of all 1-sensitive optimal solutions.
Proof: We note that each value of the interference I, the
linear program LP (I, i) always gives a solution which is a
vertex of the polyhedron Zi . As the number of vertices of
the polyhedron Zi is finite, we can find a subsequence {Im }
such that a single vertex zi ∈ Zi is a subset of the set of
solutions Si (Im ) for each integer m. Clearly zi lies in the
set limI→∞ S i (I). This proves that the set limI→∞ S i (I) is
nonempty.Let the policy zi lie in the set limI→∞ S i (I). Hence
there exists, a subsequence {Ik } such that the policy zi lies in
the set of solutions Si (Ik ) for each integer k and limk→∞ Ik =
∞. Hence we have,
Ik · ti (Ik , zi ) ≥ Ik · ti (Ik , ẑi ), ∀ẑi ∈ Zi .
Si1 .
Hence, in the limit we have zi ∈
We note that in most cases the outer limit limI→∞ S i (I)
is a strict subset of the set Si1 of all 1− sensitive optimal
solutions. In the next theorem we describe all the 1− sensitive
optimal policies. We do so by first defining the following set
of structural policies.
Let us represent the set Hi and Pi as sets of strictly
increasing positive real numbers, i.e.,
n
o
n
o
Hi = 0, h1i , · · · , hri and Pi = 0, p1i , · · · , pqi
ui (pki /hli ) =
zi (hli , pki )
,
πi (hli )
is monotonically increasing function
of the channel gain
Pr
hli for all index l satisfying j=l π(hji )pqi ≤ P i .
2) TheP
average power expenditure in each channel state hli ,
i.e a pai zi (hli , pai ) is increasing in the channel index l.
The next result shows that the all the policies in the above set
are optimal solutions of the linear program LP (i, 1).
Theorem 6. The set Si1 of solutions of linear program
LP (i, 1) is equal to the set Qi of structured policies.
This proof is also provided in appendix B. In general there
exist infinite 1− sensitive optimal solutions. Only in the case
when the power set Pi = {0, p1i } is the solution unique. In this
case the outer limit limI→∞ Si (I) is equal to the set Si1 af all
1− sensitive optimal solutions. We next provide a algorithm
which computes one of the policies of the set Si1 .
Algorithm 2 Infinitely invariant Optimal Policy
forPl = r : 0 do
r
if j=l π(hji )pqi ≤ P i then
Set zi (hli , pqi ) = π(hli ).
SetP
zi (hli , pki ) = 0, ∀k ∈ {0,
P1,r · · · , q − 1}.
r
else if j=l π(hji )pqi > P i ≥ j=l+1 π(hji )pqi then
Find a ∈ {0,
P1, · · · , q − 1} with
.
π(hli )pai ≤ rj=l π(hji )pqi − P i ≤ π(hli )pa+1
i
Set zi (hli , pai ) =
Set
Set
else
Set
Set
Pr
q
j
j=l+1 π(hi )pi −P i
a+1
a
pi −pi
π(hli )pa+1
+
i
zi (hli , pa+1
) = π(hli ) − zi (hli , pai )
i
zi (hli , pki ) = 0 ∀k 6= {a, a + 1}
zi (hli , p0i ) = π(hli ).
zi (hli , pki ) = 0, ∀k ∈ {1, 2, · · · , q}.
We will show in the next subsection that the policy generated
by the algorithm 2 is the unique policy belonging to the outer
limit limI→∞ Si (I).
8
B. Infinitely invariant Optimal Solution
In the previous subsection, we considered policies that
belonged to the set Si (I) of solutions of the linear program
LP (i, I) infinitely often as the interference tends to infinity.
In this subsection, we consider policies that belong to every
set Si (I) always after some finite value of interference I0 .
Definition 7. We define the inner-limit of the sets Si (I) as the
interference I tends towards infinity as
\
Si (I),
limI→∞ Si (I) =
I≥I0
where I0 is some finite interference. We refer to any point
zi belonging to the set limI→∞ Si (I) as a infinitely invariant
optimal solution. W The interference I0 is called the threshold
interference of user i.
If there exist a infinitely invariant optimal solution, then it
can be shown that the inner limit limI→∞ Si (I) is a subset of
the outer limit limI→∞ Si (I). If both the sets are equal, then
we say that the limI→∞ Si (I) exists and define
lim Si (I) = limI→∞ Si (I) = limI→∞ Si (I).
I→∞
We show the existence of an infinitely invariant optimal
solution for the hypothetical case when the set of feasible
solutions Zi has two vertices zi1 and zi2 . This proof motivates
the definitions later defined in this subsection. The proof for
the existence of an infinitely invariant optimal solution for a
general set of feasible solutions Zi is presented later.
Theorem 7. Assume that the polyhedron Zi is equal to the
convex closure of two policies zi1 and zi2 . Then there exist an
infinitely invariant optimal solution.
Proof: We define the set Ki (ki ) as the set of channelpower pair (hi , pi ) whose product is equal to the signal to
unit noise variance (snr) ki , that is,
o
n
(20)
Ki (ki ) = (hi , pi ) ∈ Hi × Pi : hi pi = ki .
Let Ki denote the set of possible distinct snr possible values.
Suppose the policies zi1 and zi2 satisfy for each distinct snr ki
X
X
zi2 (hi , pi ).
(21)
zi1 (hi , pi ) =
(hi ,pi )∈Ki (ki )
(hi ,pi )∈Ki (ki )
We can now see in such a case, the expected time average
reward ti (zi1 , I) and ti (zi2 , I) under the two policies are
the same for any interference I. Therefore, we have the
interference I = 0 as the interference threshold and the inner
limit limI→∞ Si (I) is equal to the set of feasible policies Zi
for all interference I greater than zero. On the other hand
suppose condition (20) does not hold true, then we will show
by contradiction the existence of an interference threshold.
So let us assume there does not exist any finite interference
threshold I0 , then the inner limit limI→∞ Si (I) is empty. We
know that there exist a vertex solution of the linear program
LP (i, I) for each interference I. Hence, either of the vertices
zi1 or zi2 belong to the set Si (I) of linear program LP (i, I)
for each interference I. As the inner limit limI→∞ Si (I) is
empty, we can find a sequence {Ik }k≥1 of distict interference
values, such that the set Si (Ik ) contains only the vertex zi1 for
odd values of index k and only the vertex zi2 for even values
of index k. Consider the real valued continuous function f (I)
defined as
f (I) := ti (zi1 , I) − ti (zi2 , I),
then the function f (Ik ) is strictly positive at odd values of
indices k and strictly negative at even indices. This implies,
we can find another sequence {Iˆk } such that f (Iˆk ) = 0 and
Ik ≤ Iˆk ≤ Ik+1 . By mean-value theorem, we can find another
sequence {Ik∗ } such that the derivative f 1 (I) of the function
f (I) evaluates to zero at the values {Ik∗ } and Iˆk ≤ Ik∗ ≤ Iˆk+1 .
However, the derivative f 1 (I) =
X
X
−ki
zi1 (hi , pi ) − zi2 (hi , pi ),
(1 + I + ki )(1 + I)
ki
(hi ,pi )∈Ki (ki )
is a rational function of the variable I and can have at most
|Ki | real numbers I such that f 1 (I) = 0. Since the derivative
f 1 (I) of the function f (I) evaluates to zero more than |Ki |
number of times, we have f 1 (I) = 0 for all I, and hence
the function f (I) is a constant function for all interference
value I. However, it can be easily seen that the function f (I)
decreases to 0 as the interference I tends towards infinity,
which contradicts our claim that f (I) is a constant function.
Hence, there exist an in infinitely invariant optimal solution.
In the proof of theorem 7, we differentiated between the
policies zi1 and zi2 , only if the condition (21) did not hold
true. We generalize this concept now. Given any occupation
measure zi of user i, we define a random variable Zi which
takes snr value ki with probability zi (ki )
X
zi (hi , pi ).
zi (ki ) = Pr(Zi = ki ) =
(hi ,pi )∈Ki (ki )
We refer to the distribution of this random variable as the snr
distribution corresponding to the policy zi . Note that for any
policy zi of user i, the time average expected reward of user
i when she/he faces an interference I is equal to
X
ti (ki , I)zi (ki ),
ti (zi , I) =
ki ∈Ki
where the one step reward ti (ki , I) of user i at snr ki and
interference I is,
ki
ti (ki , I) = log2 1 +
.
1+I
We restate condition (21) in the next theorem. The second
condition of the next theorem gives a sufficient condition for
verifying when two policies have the same snr distribution.
Theorem 8. Suppose zi1 and zi2 are two distinct policies of
user i.
1) If the snr random variables Zi1 and Zi2 corresponding
to distinct policies zi1 and zi2 have the same distribution,
then the policies give rise to the same the time average
expected reward of user ifor any interference I, i.e.
ti (zi1 , I) = ti (zi2 , I), ∀I ≥ 0.
9
2) Suppose time average expected reward of user i evaluated at the two policies zi1 and zi2 are the same at |Ki |−1
distinct interference values {Ik }, 1 ≤ k ≤ |Ki |−1 , then
the snr random variables Zi1 and Zi2 corresponding to
distinct policies zi1 and zi2 have the same distribution.
The proof is provided in Appendix D. The above theorem
ensures that we need only to consider the snr distribution
corresponding to a policy, as two distinct policies with the
same snr distribution give the same time average expected
reward. We state the generalization of theorem 7 for a general
polyhedron Zi of user i in the next theorem.
Theorem 9. There exist an infinitely invariant optimal solution
for user i. Further more, if policies zi1 and zi2 are infinitely
invariant optimal solutions, then the snr distributions corresponding to these policies are same.
The proof is provided in Appendix D. Using theorem 9 and
theorem 8, we note that for any interference I larger than the
threshold interference I0 , the infinitely invariant policies give
the same time average expected reward. In the next theorem,
we show that the outer limit is precisely the set of all infinitely
invariant optimal policies.
Theorem 10. The outer limit limI→∞ Si (I) is equal to the
inner limit limI→∞ Si (I) for any user i.
Proof: We have already established in theorem 9
that the inner limit limI→∞ Si (I) is nonempty. We shall
prove the outer limit limI→∞ Si (I) contains the inner limit
limI→∞ Si (I) thus proving theorem 10. Let policy zi1 be an
infinitely invariant optimal solution and policy zi2 belong to
the outer limit limI→∞ Si (I). Then there exist a sequence
{Ik } of interference terms which tends towards to infinity as
k tends towards to infinity and the policy zi2 belongs to the
set Si (Ik ) of solutions of the linear program LP (i, Ik ) for
each index k. Hence we can find at least |Ki | − 1 distinct
interference {Il } terms such that the sets Si (Il ) contain both
policies zi1 and zi2 . Thus by theorem 8, part 2, we have the
snr distributions corresponding to these policies are the same.
Hence from theorem 8, part 1 both policies give rise to the
same time average expected reward. Thus the policy zi2 also
belongs to inner limit limI→∞ Si (I).
As a corollary of theorem 5 and theorem 10, the set Si1
of all 1−sensitive optimal solutions contains all the infinitely
sensitive solutions of user i. In the case the set Pi of power
of user i is of the form {0, p1i }, the set Si1 is unique, thus one
can solve the approximate linear program LP (i, 1) to compute
the infinitely invariant solution. However, in general the set Si1
contains policies other than the infinitely invariant policies. In
the next subsection we shall show that algorithm 2 computes
an infinitely invariant optimal solution.
C. Computation of an infinitely invariant optimal solution.
In the previous subsection, we defined and showed the
existence of an infinitely invariant optimal solution and also
proved that the outerlimit contains only these policies. In
this subsection we shall show how to compute an infinitely
invariant optimal solution. We shall also show that there exist
only one such policy. We do so by refining the set Si1 of 1−
sensitive optimal solutions by considering the corresponding
linear program LP (i, 2). Find a policy zi∗ of user i which
minimizes,
n
o
li2 (zi ) : zi ∈ Zi , li1 (zi ) = li,1 ,
(22)
where
li2 (zi ) :=
X
(hi pi )2 zi (hi , pi ),
(hi ,pi )
and
li,1 := max{li1 (zi ) : zi ∈ Zi }
We denote the set of optimal solutions of the linear program
LP (i, 2) as Si2 and any policy belonging to this set is called
a 2− sensitive optimal policy. We note that by definition, any
2− sensitive optimal policy is also 1− sensitive optimal policy.
The next theorem shows that the previously defined set Si2
contains all the infinitely invariant optimal solutions.
Theorem 11. Let zi be an infinitely invariant optimal solution
of user i. Then the policy zi is 2− sensitive optimal and
belongs to the set Si2 . Furthermore this set only contains one
policy and is computed by algorithm 2.
The proof is provided in Appendix E. It is to be noted
that the above theorem establishes that it is sufficient to solve
LP (i, 2) to compute the unique infinitely invariant optimal
solution. Theorem 11 also shows that rather than solving
the linear program LP (i, 2), we can use algorithm 2 to
compute this policy. We note that the objective functions
li1 (zi ) and li2 (zi ) can be formulated in terms of snr distribution
corresponding to the policy zi as,
X
ki zi (ki ) = E(Zi ),
li1 (zi ) =
ki ∈Ki
and
li2 (zi ) =
X
ki2 zi (ki ) = E(Zi2 )
ki ∈Ki
respectively. One now notices that the infinitely sensitive
optimal policy minimizes the variance of the snr random
variable Zi over all policies which maximizes the mean of
the snr random variable Zi . In the next section we shall prove
that under simple sufficient conditions an infinitely invariant
optimal solution is the unique Nash equilibrium of the Markov
game Γn at sufficiently but finite large number of users.
VI. M ARKOV
GAMES AT LARGE NUMBER OF USERS
In this section, we shall prove that the infinitely invariant
optimal solution is a strictly dominating policy over the set of
all rational policies of the other users for any Markov game,
once the number of users crosses a particular threshold. As
we assume that all the users are rational, at sufficiently large
number of users, the infinitely invariant optimal solution for
each user is the Nash equilibrium policy of that user.
To achieve this result, we shall first characterize the set of all
rational polices of each user. We shall show that the set of all
10
policies where each user’s corresponding snr random variables
has a strictly positive mean contains his/her rational policies.
Then we shall show that at large number of users, when all
users employ rational policies, the best response of each user is
to employ her/his infinitely invariant optimal solution. Recall
the Markov game where each user treats interference as noise
was denoted as Γn . We note that the Markov game Γn when
the set of users is N is denoted from now on as ΓN . The set
of all possible users from now on is the set of strictly positive
integers Z+ and any set of users N is a strict finite subset
of the set Z+ . Since we shall analyze any game with the i′ th
user in perspective, we always consider that user i belongs to
the set of users N unless otherwise mentioned. We first define
the idea of an Infinitely invariant Nash equilibrium policy for
each user.
Definition 8. A sequence of policies {zi∗ }i≥1 are called the
Infinitely invariant Nash equilibrium policies of all possible
users, if there exists a natural number N ∗ such that for all
the finite subsets of users N ⊆ Z+ satisfying |N | ≥ N ∗ , we
have for all users i in the set N , the policy zi∗ is a Nash
equilibrium policy of each user i for the Markov game ΓN .
It is to be noted that the existence of this sequence of policies ensures that we need to use the best response algorithm
only if the cardinality of set N of users does not exceed the
threshold N ∗ . As we shall show that infinitely invariant Nash
equilibrium policy of any user is his/her infinitely invariant
optimal solution, once the number of users exceed N ∗ , we
can simply use algorithm 2 to compute the infinitely invariant
optimal policy and find all the Nash equilibrium policies. As
algorithm 2computes this policy in O(|Hi ||Pi |) steps, we can
compute the Nash equilibrium policies of the Markov game in
O(|N ||Hi ||Pi |) steps. Suppose the set N of users is strictly
greater than the number N ∗ , then with any addition of a new
user, we would no longer need to recompute the equilibria
policies using the best response algorithm, but rather just find
the infinitely invariant Nash equilibrium policy of the new user
and adjoin this policy to the older set of equilibrium policies.
This new set of policies will constitute a Nash equilibrium
for the new Markov game obtained by addition of this new
user. Similarly if any user leaves the system and yet the
residual number of users in the system still remains greater
than the threshold, we would not need to recompute the Nash
equilibria of this new system, but continue using the previously
computed policies. We also recall that the best response
algorithm required each user to know the channel statistics
of the other users. However, if an infinitely invariant Nash
equilibrium exist, then this condition is no longer required
and each user only needs to know his/her channel gain and
channel statistics. In the next subsection we analyze the set of
rational policies of each user.
A. Set of all best responses
Previously, we defined the best response set Bi (z−i ) of user
i as the set of all his/her policies, which maximizes his time
average expected reward Ti (zi , z−i ) when users other than
user i use multi-policy z−i . We use this set to define the set
(Bi ) of all best response of user i.
Definition 9. The set (Bi ) of all best response of user i is the
set of all policies zi of user i, such that each policy belonging
to this set belongs to the best response set Bi (z−i ) for some
multi-policy z−i of users other than user i, for some set N of
users containing user i.
It can be observed that the all the possible rational policies
for any game in which that user is involved is contained in the
N
set of all his/her best response policies. If we denote Z−i
as
the set of all feasible multi policies of users other than user i,
when the set N containing the number i denotes all the users
in the game, then the set Bi of all best responses of user i can
be obtained as,
[
[
Bi =
Bi (z−i ).
N
N ⊆Z+ z−i ∈Z−i
As all the users are rational, policies not in the set Bi (z−i )
are not used by user i. This implies that rational user need
not consider polcies in the set Zi /Bi (z−i ). We summarize the
result below.
Theorem 12. Let ΓN be a Markov game involving the set N
of users. If z−i denotes the policies of users other than user
i, then for any policy zi not in the set Bi of all his/her best
responses, there exist another policy zi∗ in the set Bi such that,
Ti (zi∗ , z−i ) > Ti (zi , z−i ).
With the previous theorem in mind, we can restrict our
attention to simply the policies belonging to the set Bi of
all best responses. Hence, we provide a important property of
policies in this set. We denote Mi (γ) as the set of all policies
of user i, for which the corresponding snr random variable Zi
has a mean which is bounded below by a strictly positive real
number γ, i.e,
n
o
X
Mi (γ) = zi ∈ Zi : E(Zi ) =
ki zi (ki ) ≥ γ .
ki ∈Ki
In the next theorem, we show that the set Bi of all best
responses of user i is a subset of the set Mi (γ) of all policies
of user i for some strictly positive real number γ. Furthermore
we provide a lower bound on this real number γ. Recall that,
we defined the set Pi of power levels of user i as the ordered
set of strictly increasing positive real numbers,
o
n
Pi = 0, p1i , p2i , · · · , pqi .
We also represent the stationary probability of user i being in
channel state 0 as πi (0).
Theorem 13. The set Bi of all best responses of user i is
equal to the set Mi (γ0,i ) of all policies of user i for some
strictly positive constant γ0,i . Also there exist a policy zi of
user i such that corresponding snr random variable Zi has
a mean equal to γ0,i . Furthermore the positive constant γi,0
satisfies,
γi,0 > min(h1i p1i (1 − π(0), h1i P i ).
The proof of this theorem is provided in Appendix F . We
call the constant γ0,i as the interference multiple of user i.
One can observe that the interference multiple γ0,i is strictly
11
positive if the stationary probability of user i being in channel
state 0 as πi (0) is strictly less than 1. We note that this
condition always holds true, as the contrapositive of this
condition would imply that user i always has a channel state
where he/she cannot get any strictly positive reward. In such a
case user i does not need to be considered as one of the user
playing the Markov game and can be neglected. Similarly the
power constraint P i of user i needs to be positive which we
assume is always true. We also note that from theorem 13 the
average interference caused by user i when he/she employs
any policy belonging to his set of all best responses is at least
the interference multiple γ0,i . In the next subsection, we show
the existence of infinitely invariant Nash equilibrium policies.
VII. E XISTENCE
OF THE INFINITELY INVARIANT
NASH
EQUILIBRIUM
In this section at large number of users, we will show that
the infinitely invariant optimal policy for each user strictly
dominates every other policy whenever the other users play
policies from their set of all best responses. We shall then later
show that the infinitely invariant optimal solution becomes the
unique Nash equilibrium at large number of players. We state
the next result.
Theorem 14. Suppose the interference multiple γi,0 for each
user i satisfies γi,0 ≥ γ0 and for some strict positive constant
γ0 . Then there exist a natural number N0,i for each user i
such that for all sets |N | ≥ N0,i + 1, the infinitely invariant
optimal policy of user i strictly dominates all her/his policies
in the game ΓN , when all users other than user i use policies
in the set of all best responses.
The proof is provided in Appendix G. One can observe in
the proof of this theorem that the number Ni,0 is a function of
only the set of channel and power Hi and Pi , the interference
threshold I0,i of user i and the lower bound γ0 . We note
that the above theorem does not guarantee that the infinitely
invariant optimal policy strictly dominates all policy when all
the other players play arbitrarily. Indeed, if all the users other
than users i do not transmit at all, then the best response of user
i belongs to the set Si (0) of solutions of the linear program
LP (i, 0). In general, the infinitely invariant optimal solution
need not belong to the latter set and will give a strictly lower
time average reward as compared to any policy in the set Si (0).
However all the users being rational always transmit from their
best response set and hence the previously mentioned scenario
does not occur.
We note that if we can ensure that the threshold N0,i is
bounded above for all users i by some constant, then there
exists an infinitely invariant Nash equilibrium. This result is
in general not easy to show. However in some cases as stated
next, the result is easily provable.
Proposition 1. [Finite sets of symmetric users] Suppose the
set Z+ of strictly positive integers can be partitioned into finite
sets N1 , N2 , · · · , Nk such that if user i and user j belong to
some set Nl , 1 ≤ l ≤ k, then both the users satisfy P i = P j ,
Hi = Hj , Pi = Pj and πi = πj .
TABLE I
PARAMETERS
Model
1
2
3
4
N
7
7
7
7
r
2
3
4
4
q
5
4
3
3
Power Constraints
2.1 ×1(6)
1.9 ×1(5)
1.35 ×1(6)
(1.35,1.25,1.25,1.35,1.35,1.35,1.35)
N∗
4
4
5
6
It is to be noted that the above assumptions ensures that the
users belonging in the same partition, then they are symmetric
and every optimization problem for each of the user in that
partition is the same. Under the above assumption we show
the uniform boundedness of the interference thresholds.
Theorem 15. Suppose the assumption "Finite sets of symmetric users " holds true. Then the numbers Ni,0 defined in the
statement of theorem 14 is uniformly bounded, that is for each
user i, we have Ni,0 ≤ maxj∈Z+ Nj,0 < ∞.
Proof: First we note that under the assumptions of symmetric users, the lower bound γ0 is strictly positive. It can
also be verified by that the interference threshold I0,i is upper
bounded by maxj∈Z+ Ij,0 . Hence the number N0,i is the same
for all users i belonging to the same partition. Hence, we have
Ni,0 ≤ maxj∈Z+ Nj,0 < ∞.
VIII. N UMERICAL R ESULTS
We present the numerical results for models with the
following parameters. The largest channel index is represented
by the variable r. Hence the cardinality of the set of channel
states is r + 1. The total number of power levels is equal
to q + 1, where the variable q represents the largest power
index. We represent the set of channel states and power levels
as {0, r1 , 2r , · · · , 1} and{0, 1, · · · , q} . We consider a Markov
fading model [10] with transition probabilities as λ0,0 = 1/2,
λ0,1 = 1/2; λq,q−1 = 1/2, λq,q = 1/2; λi,i = 1/3,
λi,i+1 = 1/3, λi,i−1 = 1/3 (1≤i≤q−1), where λi,j represents
the probability of going from state index i to state index j. N
is the number of users in the system and 1(k) is a k length row
vector of all ones. Table I shows the various model parameters
and the minimum number of users at which the infinitely
invariant Nash equilibrium becomes the equilibrium policy.
Each element of table II contains values of two stationary
policies. The first policy is the Nash equilibrium policy of the
first user in the models when only two users are in the system
(we neglect the last five users in all models). The second policy
is the infinitely invariant Nash equilibrium policy of the first
user. We observe that the infinitely invariant Nash equilibrium
policy is different from the NE policy in all the models. Indeed,
more number of users are required for the infinitely invariant
Nash equilibrium policies to become a Nash equilibrium.
To further illustrate this point, we plot in figure (1) the
two norm distance between the NE policy and the infinitely
invariant Nash equilibrium policy of the first user against the
total number of users in the system. We can observe from the
figure that for models 1 and 2, it requires at least four users for
the infinitely invariant Nash equilibrium policy to become a
12
TABLE II
S TATIONARY POLICIES ui (pi , hi )
0
1,1
0,0
0,0
0
1,1
.3333,1
0,0
0,0
0
1,1
1,1
0,.15
0,0
0,0
0
1,1
1,1
0,.15
0,0
0,0
0
3
0,0
.2333,0
0,0
3
0,0
0,0
1,.3333
0,0
3
0,0
0,0
0,0
0,1
1,1
3
0,0
0,0
0,0
0,1
1,1
4
0,0
0,0
1,0
4
0,0
0,0
0,.6667
1,1
5
0,0
0,0
0,1
-2
-4
-6
-8
-10
-12
2
2.5
3
3.5
4
4.5
5
5.5
6
6.5
Total No. of users (N)
Figure (2): Plot of the difference of the rewards of the first user, when all
users use their Nash equilibrium policies and when all users use their infinitely
invariant Nash equilibrium policies, against the number of users.
A PPENDIX A
P ROOF OF T HEOREM 3.
Fig. 1. l2 norm difference
1.8
Model1
Model2
Model3
Model4
1.6
A. Proof of Theorem 3, part (1)
Proof: The Markov game Γ is a potential game, then from
definition 2 there exist a function T : Z → R, such that for
each user i, Ti (zi , z−i )−Ti (ẑi , z−i ) = T (zi , z−i )−T (ẑi , z−i ),
for all multi-policy z−i of all users other than user i and for
any pair of policies (zi , ẑi ) of user i. It can now be observed
in the best response algorithm, for each index k,
1.4
1.2
Two Norm Difference
× 10 -3
Model1
Model3
Model4
Model2
Power
−→
2
0,0
.7667,.5667
0,0
2
0,0
0,0
0,0
0,0
2
0,0
0,0
.85,0
1,0
0,0
2
0,0
0,0
.85,0
1,0
0,0
1
0,0
0,.4333
0,0
1
0,0
.6667,0
0,0
0,0
1
0,0
0,0
.15,.85
0,0
0,0
1
0,0
0,0
.15,.85
0,0
0,0
2
Reward Difference
Channel
↓
Model1
0
1
2
Model2
0
1
2
3
Model3
0
1
2
3
4
Model4
0
1
2
3
4
Fig. 2. Reward Difference
1
0.8
T (z(k)) ≤ T (zi (k + 1), ẑ−1 (k)) ≤
(23)
T (z2 (k + 1), ẑ−2 (k)) ≤ · · · ≤ T (z(k + 1)). (24)
0.6
0.4
0.2
0
2
2.5
3
3.5
4
4.5
5
5.5
6
6.5
Total No. of Users(N)
Figure (1) :Plot of two norm distance between the NE policy and the infinitely
invariant Nash equilibrium policy of the first user against the total number of
users in the system.
NE, while for models 3 and 4, we need at least 5 and 6 number
of users respectively. Note that the model 3 and 4 differ only
by user 2 and 3 having lesser power constraint. Due to this,
the infinitely invariant Nash equilibrium policies become NE
policies only once the total number of users crosses 6.
In figure (2) we plot the difference of the rewards of the first
user, when all users use their Nash equilibrium policies and
when all users use their infinitely invariant Nash equilibrium
policies, against the number of users. As mentioned in table
(I) and seen in figure (1), we see that the rewards of the users
have become the same, once the total number of users crosses
the integer N ∗ .
7
We first prove that the algorithm terminates. To do so, assume
the contrary, then for any index k, we have z(k) 6= z(k + 1).
Since by assumption, the best response policy of each user i
computed in the algorithm is a vertex of the polyhedron, from
finiteness of the vertices of each polyhedron Zi , we have that
the sequence {z(k)} can only take values from a finite set.
This implies there exist at least two indices k < j such that,
z(k) = z(j). Without loss of generality (WLOG), assume that
k, j are the smallest indices which satisfy the latter property.
Thus we have from (23),
T (z(k)) = T (z(k + 1)) = · · · = T (z(j − 1)) = T (z(j)).
In particular all inequalities in (23) hold with equality for the
index k. As the best response algorithm ensures for all users i,
zi (k + 1) = zi (k) if T (zi (k + 1), ẑ−1 (k)) = T (zi (k), ẑ−1 (k)),
we have z(k) = z(k + 1). This clearly contradicts our
assumption.
We have shown there exist a index k such that z(k) = z(k+
1). Assume it its the smallest such index. Now we have from
(23), T (z(k)) = T (z(k + 1)). In particular, all inequalities in
(23) hold with equality for index k. Hence, z(k) is a CNE.
7
13
B. Proof of Theorem 3, part (2)
Proof: We first show that the sequence {T (z(k))} converges to a finite number. Calculation (23) shows that the
sequence {T (z(k))} either converges to a finite real number or
infinity. Denote this limit by T ∗ . Let z ∗ be a limit point of the
sequence z(k) generated by the best response algorithm and
let z(km ) be the subsequence that converges to z ∗ . As the
function T is continous, the sequence T (z(km )) converges
to T (z ∗ ). As the set Z is compact, by Weierstrass theorem
sup{T (z) : z ∈ Z} is finite. Since z ∗ ∈ Z as the set Z is
closed, it follows that T (z ∗ ) < ∞. As the sequence {T (z(k))}
converges to T ∗ = T (z ∗ ) which is finite.
We now prove that the algorithm will terminate. suppose
not, then for each index k, we have for some user i,
Ti∗ (k) := Ti (zi , ẑ−i (k)) − Ti (zi (k), ẑ−i (k)) ≥ ǫ.
As
T (z(k + 1)) − T (z(k)) =
X
Ti∗ (k) > ǫ,
i∈N
we have from calculation (23) that the number T ∗ = ∞. This
contradicts our previous proof of the statement that number
T ∗ < ∞. Hence the algorithm is guaranteed to terminate. Let
k be the smallest index such that for all users i, we have
Ti∗ (k)
= Ti (zi , ẑ−i (k)) − Ti (zi (k), ẑ−i (k)) ≤ ǫ.
It can now be verified that z(k) is a ǫ− CNE.
C. Proof of Theorem 3, part (2)
We first show that the sequence {T (z(k))} converges to
a finite number. Calculation (23) shows that the sequence
{T (z(k))} either converges to a finite real number or infinity.
Denote this limit by T ∗ . Let z ∗ be a limit point of the sequence
z(k) generated by the best response algorithm and let z(km )
be the subsequence that converges to z ∗ . As the function T is
continous, the sequence T (z(km )) converges to T (z ∗ ). As the
set Z is compact, by Weierstrass theorem sup{T (z) : z ∈ Z}
is finite. Since z ∗ ∈ Z as the set Z is closed, it follows
that T (z ∗ ) < ∞. As the sequence {T (z(k))} converges to
T ∗ = T (z ∗ ) which is finite.
We now prove that the algorithm will terminate. suppose
not, then for each index k, we have for some user i,
Ti∗ (k) := Ti (zi , ẑ−i (k)) − Ti (zi (k), ẑ−i (k)) ≥ ǫ.
As
T (z(k + 1)) − T (z(k)) =
X
Ti∗ (k) > ǫ,
i∈N
we have from calculation (23) that the number T ∗ = ∞. This
contradicts our previous proof of the statement that number
T ∗ < ∞. Hence the algorithm is guaranteed to terminate. Let
k be the smallest index such that for all users i, we have
A PPENDIX B
P ROOF OF THEOREM 6.
We first prove a simple lemma which states that any optimal
policy to the linear program LP (i, 1) will either ensure that the
user transmits with peak power, or the average power utilized
by this policy will be equal to the average power constraint.
Lemma
1. Let zi denote a 1−sensitive optimal policy. If Pi ≥
Pr
x q
x=1 π(hi )pi , then for all index x,
(
π(hxi ), if a = q
x a
zi (hi , pi ) :=
0,
otherwise.
Pr
x q
However if pi < x=1 π(hi )pi , then the expected average
power constraint pi (zi ) under the policy zi is equal to P i .
Proof: We leave the reader to prove P
the first condition.
We prove the second case. Assume pi < rx=1 π(hxi )pqi , let
policy zi∗ be a 1−sensitive optimal policy. Suppose zi∗ does
not satisfy P i (zi∗ ) = P i . It follows then that for at least one
channel index x∗ we have the occupation measure zi∗ (pqi ) <
∗
∗
π(hxi ). Let the set of power indices a for which zi∗ (pai , hxi ) 6=
0, a 6= q be A. Choose ǫ such that
0<ǫ<
Consider the policy ẑi such that
ǫ
∗ x a
∗
zi (hi , pi ) − |A| , if a ∈ A, a 6= q, x = x ,
ẑi (hxi , pai ) := zi∗ (hxi , pqi ) + ǫ,
if x = x∗ , a = q,
∗ x a
zi (hi , pi ),
otherwise.
The reader can now verify the following,
q
q
X
X
∗
∗
∗ x∗
a
zi (hi , pi ) =
ẑi (hxi , pai ) = π(hxi ),
a=0
a=0
X
pi (ẑi ) = pi (zi∗ ) + ǫ pqi −
pxi /|A| ≤ P i ,
x∈A
and
li1 (ẑi ) − li1 (zi∗ ) =
X
∗
∗
ǫ (pqi · hxi ) −
(pai · hxi )/ |A|) > 0.
a∈A
Hence by contradiction zi∗ cannot be optimal.
We now use the above lemma to complete the proof of
theorem 6. The reader is asked to prove
6 using
Pr theorem
x q
π(h
)p
lemma 1 under the condition pi ≥
i
i on the
x=1
average power constraint of user i. We consider the case where
this condition does not hold true.
Proof: Assume that zi∗ be a 1−sensitive optimal policy.
Then by lemma 1, we know that pi (zi ) = P i > 0. We first
∗
consider the case thatP
for some channel state x∗ , zi∗ (pqi , hxi ) <
∗
π(hxi ), even though rj=x∗ π(hji )pqi ≤ P i . We will show this
is not possible. We first observe that,
∗
q
xX
−1 X
Ti∗ (k) = Ti (zi , ẑ−i (k)) − Ti (zi (k), ẑ−i (k)) ≤ ǫ.
It can now be verified that z(k) is a ǫ− CNE.
P i − pi (zi )
P
.
pqi − a∈A pai / |A|
≥
zi∗ (pai , hji )pai
j=0 a=0
r
X
Pi −
j=x∗
= Pi −
q
r X
X
j=x∗ a=0
π(hji )pqi ≥ 0.
zi∗ (pai , hji )pai
14
We note that if the above set of inequalities hold with equal∗
∗
ities, then zi (pqi , hxi ) = π(hxi ), which contradicts our asPx∗ −1 Pq
j a
a
sumption. So we assume that, j=0
a=0 zi (pi , hi )pi > 0.
∗
This implies that there exist a channel index x̂ < x such that
x̂ a∗
for some index a∗ , such that pa∗
i > 0 and zi (hi , pi ) > 0.
Consider the policy ẑi of user i,
∗
zi∗ (hx̂i , pai ) − ǫi , if , a = a∗ , x = x̂,
∗ x̂
if a = 0, x = x̂
zi (hi , 0) + ǫi ,
∗
q
x a
∗
x
ẑi (hi , pi ) := zi (hi , pi ) + ǫ2 , if , a = q, x = x∗ ,
∗
zi∗ (hxi , 0) − ǫ2 , if , a = 0, x = x∗ ,
z ∗ (hx , pa ),
otherwise,
i
i
i
∗
where the constants ǫi > 0 and ǫ2 > 0 satisfies ǫi ·pai = ǫ2 ·pqi .
The reader can verify that for each channel index x,
q
X
ẑi (hxi , pai )
=
q
X
zi∗ (hxi , pai )
π(hxi ).
=
It also follows that
∗
pi (ẑi ) − pi (zi∗ ) = ǫ2 · pqi − ǫi · pai = 0,
∗
li1 (ẑi ) − li1 (zi∗ ) = hxi · ǫ2 · pqi − hx̂i · ǫi · pai > 0,
Hence we have exhibited a policy ẑi that strictly performs better than the optimal policy zi∗ . Therefore, by contradiction, we
∗ q
x∗
x
have for the optimal policy zi∗ , zP
i (pi , hi ) = π(hi ), whenever
r
j q
∗
the channel index x satisfies, j=x∗ π(hi )pi ≤ P i .
x∗ denote the maximal channel
that
Pr index such
PLet
r
j q
j q
P
.
Hence
we
have
π(h
)p
π(h
)p
>
∗
∗
i
i i ≤
i i
j=x +1
j=x
∗
P i If zi is an optimal policy of the linear program LP (i, 1),
then it should satisfy zi∗ (hji , pqi ) = π(hji ) for all channel
indices j ≥ x∗ +1. We also note that the value of the objective
function li1 (zi ) evaluated at the policy zi∗ is equal to
li1 (zi∗ ) =
q
r X
X
∗
zi∗ (hri , pqi )hji pli
≤
zi (hji , pli )hji pli
+
≤
≤
q
X
π(hji )hji pqi
+
∗
hxi (P i
q
X
−
j=x∗ +1
Pi =
=
j=x∗ +1
+
q
x X
X
l=0
∗
zi∗ (hxi , pli )pli = P i −
(25)
zi∗ (hji , pli )pli .
j=0 l=0
q
X
j=x∗ +1
|Ki |−1
X
log2 1 +
X
j=1
π(hji )pqi ),
One can get equalities in the expressions (25) if the optimal
policy zi∗ is such that
q
X
Let us consider the function
f (x) =
∗
π(hji )pqi
ti (kl , Ij )α(j) = 0.
j=1
|Ki |−1
j=x∗ +1
q
X
X
1
where the last inequality follows from,
pi (zi∗ )
|Ki |−1
π(hji )hji pqi
j=x∗ +1
j=0 l=0
We prove the result by contradiction. Suppose the matrix Mi
is not invertible, then we can find real numbers α(j), 1 ≤ j ≤
|Ki | − 1, not all of them zero, such that for all non zero snr
values kl ∈ Ki /{0},
π(hji )pqi ).
kl
α(j).
1 + Ij
Hence the function f (x) is zero at all points x ∈ Ki . By
the mean value theorem, we have the derivative f 1 of the
function f is zero at |Ki | − 1 distinct real numbers. However
the derivative of f (x) is‘
π(hji )hji pqi
q
X
∗
zi (hji , pli )hix pli +
Mi (l, j) = ti (kl , Ij ),
j=1
j=x∗ +1
j=0 l=0
q
x∗ X
X
q
X
Lemma 2. Let Mi denote a (|Ki | − 1) × (|Ki | − 1) matrix
with the (l, j) entry as
f (x) =
j=0 l=0
q
x X
X
We first prove a result which is used to prove theorem 8.
Proof: Recall that, the one step reward
kl
ti (kj , Il ) = log2 1 +
1 + Ij
and
∗
A PPENDIX C
P ROOF OF THEOREM 8
where kl are the distinct non zero snr terms in the set Ki and
Ij , 1 ≤ j ≤ |Ki | − 1, are distinct interference values. Then
the matrix Mi is invertible.
a=0
a=0
In such a case, one can verify that the policy zi∗ belongs to
the set Qi . This proves that the set of 1−sensitive optimal
solutions Si1 is a subset of the set of structural results Qi .
To prove the converse, observe that for any two distinct
solutions zi1 and zi2 belonging to the structural results Qi ,
we have li1 (zi1 ) = li1 (zi2 ), hence, all policies in the set Qi are
1−sensitive optimal policies.
α(j)
.
(1 + x + Ij )
Clearly the derivative is a rational function and can have at
most |Ki | − 2 real numbers x where f 1 (x) evaluates to zero.
Hence f 1 (x) = 0 for all real numbers x, therefore f (x) is a
constant function. However the function f (x) is zero at x = 0
and nonzero at any x strictly positive if the numbers α(j) are
not all zero. Hence by contradiction, the real numbers α(j)
are all zero for all 1 ≤ j ≤ |Ki | − 1. Hence the matrix Mi is
invertible.
We now prove theorem 8.
Proof: Let zi1 and zi2 be two policies of user i such that
the corresponding snr random variables Zi1 and Zi2 have the
distribution. Then the time average expected reward of user
15
i for any interference I evaluated at the first policy zi1 is
ti (zi1 , I) =
X
X
ti (ki , I)zi1 (ki ) =
ti (ki , I)zi2 (ki ) = ti (zi2 , I).
Lemma 3. Let Pi = {0, p1i , · · · , pqi } denote the set of power
of user i. If x is a constant such that 0 ≤ x ≤ pqi , denote ⌊x⌋
to be the index j which satisfies,
This proves the first part of theorem 8. Now suppose the time
average expected reward of user i evaluated at the two polices
are the same for |Ki |− 1 distinct interference values {Ik }, 1 ≤
k ≤ |Ki |, then the vector {(zi1 (kl ) − z12 (kl )), kl ∈ Ki /{0}}
belongs to the null set of the transpose of the matrix Mi . As
the matrix Mi is invertible by the previous lemma, we have
for all snr terms kl ∈ Ki /{0}, zi1 (kl ) = z12 (kl ). This implies
that the two snr distribution are the same proving the second
part of theorem 8.
Then the unique probability distribution {α(j) : 0 ≤ j ≤ |Pi |}
which minimizes,
X j
(pi )2 α(j),
ki ∈Ki
ki ∈Ki
A PPENDIX D
P ROOF OF THEOREM 9
Proof: Let E(Zi ) = {zi1 , zi2 , · · · , zim } denote the set of
vertices of the polyhedron Zi of feasible occupation measures
of user i. As the set Zi is the convex closure of its vertices, we
have Zi = conv(E(Zi )). We first show that given any distinct
two vertices zij and zik of the previously defined set, there exist
a finite real number I(zij , zik ) such that for all interference
I ≥ I(zij , zik ), either the condition ti (zik , I) ≥ ti (zij , I) or
ti (zik , I) ≤ ti (zij , I). To prove this claim, assume the result is
not true, then we can find at least |Ki | − 1 distinct values of
interference Ik , 1 ≤ k ≤ |Ki | such that ti (zik , Ik ) = ti (zij , Ik )
for all these values of interference.
However by theorem 8 part (2), we have that the two snr
distributions are equal, i.e zik = zij . Again by theorem 8 part
(1), we have for all interference I ≥ 0, ti (zik , Ik ) = ti (zij , Ik )
which contradicts our assumption. Hence, by contradiction,
there exist a finite finite real number I(zij , zik ) such that for
all interference I ≥ I(zij , zik ), either the condition ti (zik , I) ≥
ti (zij , I) or ti (zik , I) ≤ ti (zij , I). We can now see that for all
interference
n
o
I ≥ I0 = max I(zij , zik ) : (zij , zik ) ∈ E(Zi ) ,
there exist a vertex zil such that the condition ti (zil , I) ≥
ti (zik , I) for all vertices zik ∈ E(Zi ) holds true. As the function
ti (zi , I) is a linear function in the variable zi and the set Zi
of feasible occupation measures is the convex closure of its
vertices, we have vertex zil belongs to the set Si (I) of optimal
solutions of the linear program LP (i, I) for all interference
I ≥ I0 . This shows the existence of an infinitely invariant
optimal solution for user i. Suppose there exists two infinitely
invariant optimal solutions, then from theorem 8, part (2) and
the definition of an infinitely invariant optimal solution, we
can show the corresponding snr distribution are the same.
A PPENDIX E
P ROOF OF THEOREM 11.
The following lemma is required for the proof of theorem
11.
⌊x⌋ = min{j ∈ {0, 1, · · · , q} : pji ≥ x}.
j∈Pi
subject to,
X
pji α(j) = α,
(26)
j∈Pi
where 0 ≤ α ≤ pqi , is the probability distribution
⌊α⌋+1
pi
−α
, if j = ⌊α⌋
pi⌊α⌋+1 −p⌊α⌋
i
∗
α (j) = 1 − α∗ (⌊α⌋), if j = ⌊α⌋ + 1
0,
otherwise.
Proof: We first note that if X is a real valued random
variable such that it lies between two real number a and b,
then
V ar(X) ≤
(b − a)2
.
4
(27)
To prove this claim observe that V ar(X) ≤ E(X − c)2 for
2
and using the
all real numbers c, by choosing c = (b+a)
2
fact that the random variable X lies between a and b almost
surely, the result follows. Note that if we denote Xα to be the
random variable that takes values in the set Pi with probability
distribution α(·), then the linear program in lemma 3 can be
restated as, finding a random variable Xα with distribution
α(·) which minimizes,
V ar(Xα ) : Xα ∈ Sα ,
(28)
where the set Sα denotes the set of random variables Xα that
takes values in the set Pi with probability distribution α(·)
and has mean α, i.e. E(Xα ) = α. Let us denote aα and bα
as the numbers belonging to the set P i such that the random
variable Xα with distribution α(·) lies almost surely between
aα and bα . Assume aα < aα , then we have from 27,
(bα − aα )2
: Xα ∈ Sα }.
4
One can easily show that the right hand side in the above
expression is minimized when aα = ⌊α⌋ and bα = ⌊α⌋ + 1.
The distribution α(·) which minimizes the left hand side in the
above expression can now be evaluated to be the distribution
stated in lemma 3.
We now provide the of theorem 11.
min{V ar(Xα ) : Xα ∈ Sα } ≤ min{
A. Proof of theorem 11
Proof: We have already established from theorem 5 and
theorem 10 that any infinitely invariant optimal solution is
1−sensitive optimal. Let zi1 be an infinitely invariant optimal
16
solution and zi2 be any other 1−sensitive optimal policy. Then,
for any interference I greater than interference threshold I0 ,
X
hi p i
1
zi1 (hi , pi )
log2 1 +
ti (zi , I) =
1+I
(hi ,pi )
X
hi p i
2
log2 1 +
≥ ti (zi , I) =
zi2 (hi , pi ).
1+I
(hi ,pi )
At sufficiently large interference I, we have
X
k
∞
(−1)k+1 hi pi
hi p i
=
.
log2 1 +
1+I
k
1+I
k=1
As both policies
zi1
and
zi2
are 1−sensitive optimal, we have
0 ≤ ti (zi1 , I) − ti (zi2 , I) =
k !
∞
X X
(−1)k+1 hi pi
hi pi
k=2
k
1+I
(zi1 (hi , pi ) − zi2 (hi , pi )).
hi pi
Hence the infinitely invariant optimal solution is also
2−sensitive optimal. Define the index k ∗ such that,
k ∗ = max{k :
r
X
pqi π(hji ) > P i }.
j=k
By theorem 6 we know that for any 1−sensitive optimal policy,
we have for channel indices k > k ∗ the policy zi satisfies
zi (hki , pqi ) = π(hki ). Hence any policy zi which is 2−sensitive
optimal minimizes,
∗
q
k X
X
(hli pki )2 zi (hli , pki )
r
X
+
π(hli )pqi ,
l=k∗ +1
l=0 k=0
over all policies zi in the set Si1 of 1−sensitive optimal
policies. Hence any such policy zi minimizes,
q
X
∗
∗
(hki pki )2 zi (hki , pki ),
k=0
subject to
q
X
r
X
π(hji )pqi }.
j=l+1
The uniqueness of this solution follows from lemma 3 and
theorem 6. The reader can verifying the correctness of algorithm 2 by checking that it computes the unique 2−sensitive
optimal policy.
A PPENDIX F
P ROOF OF THEOREM 13
We first characterize the set Bi of all best responses of user
i. The first lemma shows that set is a compact set.
Proof: Recall that that set Bi of all best responses of user
i, was defined as,
[
[
Bi (z−i ),
Bi =
N
N ⊆Z+ z−i ∈Z−i
N
where Z−i
denotes the the set of all feasible multi-policies of
users other than user i, when the set N containing the number
i denotes the set of all users in the game ΓN . Let C(Zi ) denote
the class of all the subsets of the set Zi , which are obtained
as convex closure of finitely many vertices of the polyhedron
Zi . As the vertices of the polyhedron Zi is finite, we have
the class C(Zi ) itself as finite. We note that the set Bi (z−i )
contains all the best response policies of user i, when users
other than user i play multi-policy z−i . As this set contains all
the set of solutions of a linear program, we have that this set
is the convex closure of finitely many points, each point being
a vertex of the set Zi of feasible occupation measures of user
i. This implies that the best response set Bi (z−i ) belongs to
the class C(Zi ) for each multi-policy z−i of users other than
user i. As the class C(Zi ) is finite, we have that the set Bi is
a finite union of compact sets. Hence Bi is compact.
We now define the set M∗i (0) as the set of all policies of
user i, for which the corresponding snr random variable Zi
has a mean strictly greater than zero, i.e,
o
n
X
ki zi (ki ) > 0 .
M∗i (0) = zi ∈ Zi : E(Zi ) =
ki ∈Ki
∗
(pki )2 zi (hki , pki ) = P i −
r
X
π(hli )pqi ,
l=k∗ +1
k=0
and
a = min{l ∈ {0, 1, · · · , q} : pli ≥ P i −
Lemma 4. The set Bi of all best responses of user i is a
compact set.
Multiplying the above expression by I 2 and letting I tend
toward infinity, we have
X
X
−(hi pi )2 zi2 (hi , pi )
−(hi pi )2 zi1 (hi , pi ) ≥
hi pi
where the index a itself satisfies,
q
X
∗
∗
zi (hki , pki ) = π(hki ).
k=0
Using lemma 3, the 2−sensitive optimal policy zi satisfies
∗
zi (hki , pli ) :=
l a+1 Pr
π(hi )pi + j=l+1 π(hji )pqi −P i
, if l = a
pa+1
−pa
i
i
k∗
a
1 − zi (hi , pi ),
if l = a + 1
0,
otherwise,
The next lemma shows that that set Bi of all best responses
is a subset of the previously defined set.
Lemma 5. The set Bi is a subset of the set M∗i (0) of the
set of all policies of user i, for which the corresponding snr
random variable Zi has a mean strictly greater than zero.
Proof: Let zi be a policy belonging to the set Bi of all
best response policies of user i. Then for some finite subset
N ⊆ N and for some multi-policy z−i of users other than user
i, the policy zi belongs
to the set Bi (z−i ) of best responses of
P
user i. Suppose ki ∈Ki ki zi (ki ) = 0, we shall then prove that
this policy cannot belong to the set Bi (z−i ) by exhibiting a
policy zi∗ which provides a strictly better time average reward,
i.e, Ti (zi∗ , z−i ) > Ti (zi , z−i ). Since the policy zi satisfies
17
ki zi (ki ) = 0, then we have that zi (hxi , pai ) > 0 if
and only if x = 0 or x 6= 0, a = 0. This implies that the
average reward Ti (zi , z−i ) at policy zi is equal to 0. Consider
the policy zi∗ of user i which satisfies,
P
i
if x = r, a = q
pqi ,
∗ x a
x
zi (hi , pi ) := π(hi ), if 0 ≤ x ≤ r − 1, a = 0,
0,
otherwise.
P
ki ∈Ki
Note that the above policy allows user i to only transmit whenever her/his channel gain is the best, i.e., x = q. It can be easily
verified that Pi (zi∗ ) = Pi , hence the policy is feasible and that
the average reward obtained at this policy Ti (zi∗ , z−i ) > 0.
Hence we have shown that Ti (zi , z−i ) = 0 < Ti (zi∗ , z−i ).
The next lemma shows that any policy belonging to the set
Bi of all best responses of user i will either ensure that the
user transmits with peak power whenever her/his channel gain
is strictly positive or the expected time average power utilized
by this policy is equal to the average power constraint P i .
Lemma 6. Let N ∈ N denote any subset of users containing
user i. Let z−i denote any multi-policy of all users other
than user i, then any optimal best response zi∗ ∈ Bi (z−i )
has
property. If the power constraint Pi ≥
Pr the following
x q
π(h
)p
,
then
for all index x > 0,
i
i
x=1
(
π(hxi ), if a = q
∗ x a
zi (hi , pi ) :=
0,
otherwise.
Pr
However if Pi < x=1 π(hxi )pqi , then the expected average
power constraint pi (zi∗ ) under the policy zi∗ is equal to P i .
Also, the optimal policy zi∗ satisfies zi∗ (0, 0) = πi (0).
The proof is similar to proof of lemma 1 in Appendix B
and hence omitted. We now give proof of theorem 13
Proof: Consider the function f : Zi :→ R defined as,
X X
f (zi ) =
hi pi zi (hi , pi ).
hi ∈Hi pi ∈Pi
Note that the function f (zi ) is the expected value of the snr
random variable Zi corresponding to the occupation measure
zi of user i. As the set Bi of all best responses of user i is a
compact set (Lemma 4 ) and the function f (zi ) is continuous,
by Weierstrass theorem , there exist a policy zi∗ ∈ Bi such
that,
f (zi∗ ) ≤ f (zi ), ∀ zi ∈ Bi .
As the policy zi∗ ∈ Bi , by lemma 5, we have f (zi∗ ) > 0,
hence we choose the interference multiple γi,0 = f (zi∗ ). Then
we have the set Bi of all best responses is subset of the set
Mi (γi,0 ) of all policies of user i for which the corresponding
snr random variable Zi has mean at least γi,0 . We will now
complete the proof by providing the lower bound on the
interference multiple γi,0 . Consider first the case that the
power constraint P i of user i satisfies,
Pi ≥
r
X
whenever her/his channel gain is strictly positive. In this case
we have for all policies zi ∈ Bi , the function f (zi ) satisfies,
f (zi ) ≥
r
X
π(hxi )hxi pqi > (1 − π(0))h1i p1i .
x=1
On the other hand if the condition 29 does not hold true, then
by lemma 6, we have for every policy zi in the set Bi of
all best responses satisfies zi (0, 0) = π(0). Also by lemma 6
every policy zi in the set satisfies Pi (zi ) = P i . As the set Bi
of all best responses is a subset of the set Zi of all feasible
policies, we have,
(
)
XX
f (zi∗ ) ≥ min
hi pi zi (hi , pi ) : zi ∈ Zi∗ ,
hi
pi
where the set Zi∗ is described as,
n
Zi∗ = zi ∈ R|Hi |×|Pi | , Pi (zi ) = P i
XX
hi
pi
zi (0, pli ) = 0, ∀ 1 ≤ l ≤ q,
o
zi (hi , pi ) = 1, zi (0, p0i ) = π(0) .
It can now be verified that in this case, f (zi∗ ) > h1i P i . Hence
we have that the interference multiple γi,0 for user i satisfies,
γi,0 > min(h1i p1i (1 − π(0), h1i P i ).
A PPENDIX G
P ROOF OF THEOREM 14
In this appendix, for a finite set N of users containing user
i, and other users j ∈ N /{i} using policy zj , we denote
the expected time average reward of user i and immediate
reward of user i at channel state-power action pair (hi , pi )
z
as Ti (zi , z−i , N ) and Ri −i (hi , pi , N ). These changes make
explicit the dependence of these functions on the set N which
will be varying in this appendix. As the proof of the theorem
14 is involved, we first provide a brief outline of the argument.
Let policy zj of user j 6= i belong to his/her set Bj of all best
responses. Let zi and zi∗ denote any policy and the infinitely
sensitive optimal policy of user i. We first show that there exist
a natural number N0,i (zi ) depending upon the policy zi of user
i, such that for all sets N of users satisfying |N | ≥ N0,i (zi ),
the difference between the expected time average rewards
when user i employs his/her infinitely sensitive optimal policy
zi∗ and policy zi is strictly positive, i.e., for all z−i ∈ B−i ,
|N | ≥ N0,i (zi )
Ti (zi∗ , z−i , N ) − Ti (zi , z−i , N ) > 0.
To show this result, we first observe the above expression can
be expressed as,
π(hxi )pqi .
Ti (zi∗ , z−i , N ) − Ti (zi , z−i , N ) =
x=1
Then by lemma 6, all the policies in the set Bi of all best
responses ensure that user i transmits with peak power pqi
(29)
X
hi ,pi
z
Ri −i (hi , pi , N )(zi∗ (hi , pi ) − zi (hi , pi )).
18
We divide the above expression into two terms at sufficiently
large
P number of users depending on the condition that the sum
j6=i Zj of the snr random variables j 6= i corresponding to
the policies zj , exceeds the interference threshold I0,i of user
i, i.e.,
Ti (zi∗ , z−i , N ) − Ti (zi , z−i , N )
= A(zi∗ , zi , N ) + B(zi∗ , zi , N ).
(hi ,pi )∈Ki (ki )
where
A(zi∗ , zi , N ) =
X
hi ,pi
X
ti (h, p)1{Pj6=i Zj ≤I0,i }
h−i ,p−i
Y
zj (hj , pj )
j6=i
∗
(zi (hi , pi ) − zi (hi , pi )) ,
and
B(zi∗ , zi , N ) =
X
hi ,pi
We state some definitions and lemmas required to prove this
theorem. These definitions and theorems convert the tedious
calculations involved in this proof in terms of the snr random
variables which are more easier to handle. Recall that the
snr distribution of the snr random variable Zi of user i
corresponding to the policy zi was given as
X
zi (hi , pi ).
zi (ki ) = Pr(Zi = ki ) =
X
ti (h, p)1{Pj6=i Zj >I0,i }
h−i ,p−i
Y
j6=i
zj (hj , pj )
(zi∗ (hi , pi ) − zi (hi , pi )) .
The idea behind the above breakup comes from two observations. The first observation is that for all interference I
greater than I0,i , the infinitely sensitive optimal policy zi
is the optimal solutions of the linear program LP (i, I) and
hence the term B(zi∗ , zi , N ) will be strictly positive. The
other observation comes from theorem 13, that when all users
other than user i use policies from their set of best response
policies, they will always transmit, causing interference of
at least γj,0 . By ensuring that the interference multiple γj,0
greater than P
some strictly positive constant γ0 , we ensure
that the sum j6=i Zj will eventually exceed the interference
threshold I0,i . By using Hoeffding inequality and this second
observation, we will show that the term A(zi∗ , zi , N ) can be
made arbitrary small. Eventually at large number of users,
the term A(zi∗ , zi , N ) will become much smaller than the
term B(zi∗ , zi , N ), and hence their sum will become positive.
Rigorously, we will show that,
A(zi∗ , zi , N ) = Ω(− exp(−(|N | − 1)),
=Ω
βl+1 (zi , zi∗ )
l+1
(|N | − 1)
−
1
|N | − 1
l+2 !
,
where l ∈ {0, 1} depending upon the policy zi of user i. If
the policy zi is not 1−sensitive optimal, then l = 0 otherwise
l = 1. One can now easily observe the existence of the number
N0,i (zi ). The functions βl+1 (zi , zi∗ ) of the policies zi and zi∗
will be shown to strictly positive for each l ∈ {0, 1}. Then to
complete the proof, we observe that for all sets N of users
satisfying |N | ≥ max{N0,i (zi ) : zi ∈ Ei (Zi )} where Ei (Zi )
is the set of all vertices of the polyhedron Zi , we have for all
policies zj ∈ Bj , j ∈ N /{i}, for all policies zi ∈ Zi ,
Ti (zi∗ , z−i , N ) − Ti (zi , z−i , N ) > 0.
(30)
and the snr random variable Zi takes values in the set Ki of
possible distinct snr values. Given policies zi of each user i
belonging to the finite set N of users, and the corresponding
snr random variables Zi , we define,
1) The immediate reward of user i at snr ki when users
{j ∈ N , j 6= i} has snr random variables Zj as,
!#
"
ki
Z−i
P
,
Ri (ki , N ) = EZ−i log2 1 +
N0 + j6=i Zj
where the expectation is taken with respect to the snr
distributions of the random variables Zj , j 6= i.
2) The expected time average reward of user i, when every
user {j ∈ N } has snr random variables Zj as,
i
h
Z
Ti (Zi , Z−i , N ) = EZi Ri −i (ki , N ) ,
where the expectation is taken with respect to the snr
distribution of the random variables Zi .
The next lemma shows that the above definitions are just
another way of stating the previously defined immediate
reward and The expected time average reward of user i when
all users j ∈ N use policy zj .
Lemma 7. Let each user j ∈ N use policy zj and Zj denote
the snr random variables corresponding to these policy zj .
1) Let the channel state-power action pair (hi , pi ) satisfy
hi pi = ki , then we have
z
Z
Ri −i (ki , N ) = Ri −i (hi , pi , N ).
2) For each user i, we have
Ti (Zi , Z−i , N ) = Ti (zi , z−i , N ).
and
B(zi∗ , zi , N )
The set Ki (ki ) was defined as
o
n
Ki (ki ) = (hi , pi ) ∈ Hi × Pi : hi pi = ki
Proof: Recall that the immediate reward of user i at
channel state hi and power action pi , when other users use
multi-policy z−i is
X
X Y
z
Ri −i (hi , pi , N ) =
[ zl (pl , hl )]ti (h, p). (31)
h−i ∈Hi p−i ∈P−i l6=i
Here the function ti (h, p) is the immediate reward of user i,
when the signals from the other users is treated as noise, i.e.,
!
hi p i
P
.
ti (h, p) = log2 1 +
N0 + j6=i hj pj
z
One can represent the immediate reward Ri −i (hi , pi , N ) as
i
h
hi pi
, (32)
Ez|N | Ez|N |−1 · · · Ez1 log2 1 + N0 +P
hj pj
j6=i
19
where each expectation is successively taken with respect to
the policy {zj , j 6= i}, from j = 1 to |N |. Recall also that the
snr random variable Zj takes value kj = hj pj with probability
X
zj (hj , kj ),
P(Zj = kj ) =
(hj ,pj )∈Kj (kj )
and alternates its sign if it is even. From theorem 11, we note
that policy zi ∈ Sil /Sil+1 for any l ∈ {0, 1} if and only if
βl (zi , zi∗ ) = 0 and βl+1 (zi , zi∗ ) > 0.
We also define the following functions and constants,
where the set Kj (kj ) = {(hj , pj ) : hj pj = kj }. One can now
show that the expression (32) is equal to
i
h
EZ|N | EZ|N |−1 · · · EZ1 log2 1 + N0 +hPi pi Zj ,
X
g(Z−i , N ) = N0 +
Zj .
j∈N /{i}
j6=i
k i = max {ki : ki ∈ Ki } , k = max k i .
where each expectation is successively taken with respect to
the distribution of the snr random variable {Zj , j 6= i}, from
j = 1 to |N |. Hence we have,
Z
Ri −i (ki , N )
=
i
a = max{k, I0,i }, fi (ki ) = zi∗ (ki ) − zi (ki ).
z
Ri −i (hi , pi , N ).
The expected time average reward of user i at policy zi when
other users use multipolicy z−i is,
X z
Ri −i (hi , pi , N )zi (hi , pi )
Ti (zi , z−i , N ) =
c1 (a, N ) = exp 4γ0 a −
hi .pi
c2 = 2
and thus
γ0
k
2
t2
(k)2 (|N | − 1)
.
( 3 )
k
, c5 = max 1,
N0
Ti (zi , z−i , N ) = Ti (Zi , Z−i , N ).
c5 (a) = −c5 (1 + 2 exp(4γ0 a)) .
We now provide the proof of theorem 14. Certain tedious
calculations in this proof are provided later for ease of the
proof.
Proof: We denote the set Si0 as the set of all feasible
policies Zi of user i. Recall that the infinitely sensitive optimal
policy of user i is the 2−sensitive optimal policy and belongs
to the set Si2 . As this set is a subset of the set Si1 of all
1−sensitive optimal policies, we have that
Si2 ⊆ Si1 ⊆ Si0 and Si2 ∪ Si1 /Si2 ∪ Si0 /Si1 = Si0 .
Let Zj be the snr random variable corresponding to the policy
zj of user j 6= i. We note that given the assumptions stated
in theorem 14, the policy zj belongs to the set Bj of all
best responses of user j. Hence by theorem 13 and the stated
conditions in theorem 14, the mean E[Zj ] of the snr random
variable Zj satisfies E[Zj ] ≥ γ0 . We denote Zi∗ and Zi to be
the snr random variable of user i corresponding to the infinitely
optimal sensitive policy zi∗ and any arbitrary policy zi ∈ Si0 .
Note that, by the uniqueness of the 2−sensitive optimal policy,
if the policy zi ∈ Si2 , then the policies zi and zi∗ are the same,
hence we consider only policies zi ∈
/ Si2 . We also define the
following functions of the policies zi and zi∗ for user i. Define
for l ∈ {0, 1, 2},
X
(ki )l (zi∗ (ki ) − zi (ki )),
βl (zi , zi∗ ) = (−1)l+1
ki
where
P(Zi = ki ) = zi (ki ) and P(Zi∗ = ki ) = zi∗ (ki ).
Note that, the function βl (zi , zi∗ ) computes the difference in
the l-th moment between the snr random variables Zi∗ and Zi
l+2
k
N0
l+1
k
N0
−
c4 (a, N ) = c3 (a, N ) −
c3 (a, N ) = −
k
N0
c1 (a, N ).
c1 (a, N ).
Now using lemma 7, the difference between the expected
average rewards when user i uses policy zi∗ and policy zi ,
when other users use policy z−i is,
Ti (zi∗ , z−i , N ) − Ti (zi , z−i , N ) =
Ti (Zi∗ , Z−i , N ) − Ti (Zi , Z−i , N ) =
X
ki
EZ−i log2 1 +
ki
g(Z−i , N )
(zi∗ (ki ) − zi (ki )).
We divide the above summation into two parts,
Ti (Zi∗ , Z−i , N ) − Ti (Zi , Z−i , N )
= A(zi∗ , zi , N ) + B(zi∗ , zi , N ),
where A(zi∗ , zi , N ) =
X
ki
1{Pj6=i Zj ≤a} fi (ki ),
E log2 1 +
g(Z−i , N )
ki
B(zi∗ , zi , N ) =
X
E log2 1 +
ki
ki
g(Z−i , N )
1{Pj6=i Zj >a}
fi (ki ),
20
We provide lower bounds on the two terms when the set N
satisfies
a
+ 1.
|N | ≥
γ0
These bounds are obtained from the lemmas provided after
the end of this proof. We have, from lemma 9
k
c1 (a, N ) exp (−c2 (|N | − 1)) .
N0
We also have from lemma 12, that if the policy zi of user i
belongs to the set Sil /Sil+1 , then term
" l+1
#
X
βr (zi∗ , zi )
∗
B(zi , zi , N ) ≥
r(N0 + (|N | − 1)k)r
r=1
A(zi∗ , zi , N ) ≥ −
−
2k
2N0 + (|N | − 1)γ0
l+2
+ c3 (a, N ) exp(−
c2
(|N | − 1)).
4
the difference
− Ti (zi , z−i ) ≥
" l+1
X
r=1
−
2k
2N0 + (|N | − 1)γ0
l+2
We now have from lemma 13,
" l+1
X
∗
Ti (zi , z−i ) − Ti (zi , z−i ) ≥
r=1
−
2k
2N0 + (|N | − 1)γ0
l+2
βr (zi∗ , zi )
r(N0 + (|N | − 1)k)r
+ c4 (a) exp(−
#
c2
(|N | − 1)).
4
βr (zi∗ , zi )
r(N0 + (|N | − 1)k)r
+ c5 (a) exp(−
#
c2
(|N | − 1)).
4
We let E(Zi ) = {zi1 , zi2 , · · · , zim } denote the vertices of the
polyhedron Zi and define the following numbers,
Ni1 = inf {(N − 1)γ0 > a} .
N ∈N
Ni2 = inf
N ∈N
(
sup
zi ∈E(Zi )
" l(zi )+1
X
r=1
j6=i
where
c1 (t, N ) = exp 4γ0 t −
t2
(k)2 (|N | − 1)
and
c2 = 2(
γ0 2
) .
k
2) And for all finite sets N ,
c
X
E[Zj ]
2
) ≤ 0) ≤ exp − (|N | − 1) .
P( (Zj −
2
4
j6=i
Hence we have for all subsets N satisfying
a
+ 1,
|N | ≥
γ0
Ti (zi∗ , z−i )
Lemma 8. Let N denote any finite set of users containing
user i. Let user j ∈ N use policy zj ∈ Bj . Let Zj be the snr
random variable corresponding to the policy zj .
1) Then for any positive real number t, we have for all sets
N satisfying (|N | − 1)γ0 > t ,
X
P(
Zj ≤ t) ≤ c1 (t, N ) exp (−c2 (|N | − 1)) ,
βr (zi∗ , zi )
−
r(N0 + (|N | − 1)k)r
Proof: Note that the snr random variable Zj takes values
between 0 and k j = max{kj ∈ Kj }. Hence the random variable Zj −E(Zj ) takes values between −E(Zj ) and kj −E(Zj ).
We then have by Hoeffding inequality for all sets N satisfying
(|N | − 1)γ0 > t,
P
P (Z −E(Z ))
P
t− j6=i E(Zj )
j
j
j6=i
P( j6=i Zj ≤ t) = P
≤
|N |−1
|N |−1
P
−2(t− j6=i E(Zj )2
P
.
≤ exp
(k )2
j6=i
j
By theorem 13 and assumptions stated in theorem 14, we have,
X
E[Zj ] ≥ (|N | − 1)γ0 and kj ≤ k.
j6=i
As a consequence we have from the above expression,
2 !
X
−2((|N | − 1)
t
P(
Zj ≤ t) ≤ exp
γ0 −
.
|N | − 1
(k)2
j6=i
It can be observed on expanding the square term, that the
above expression is the desired result. The proof of the second
part of this lemma is similar to the first one and hence it is
omitted.
Lemma 9. We have for all sets N satisfying (|N | − 1)γ0 > a
the term
A(zi∗ , zi , N ) ≥ −
k
c1 (a, N ) exp (−c2 (|N | − 1)) ,
N0
l(zi )+2 #
)
where c2 , c1 (a) constants as defined in lemma 8 and the
c2
≥ −c5 (a) exp(− (|N | − 1)) constant
,
a = max{Ii,0 , k}.
4
Proof: We have A(zi∗ , zi , N ) ≥
#
!
"
where l(zi ) ∈ {0, 1} with l(zi ) = 0 if policy zi is not
X
k
1
2
i
1−sensitive. Define Ni,0 = max{Ni , Ni }. We then have −
P
1{Pj6=i Zj ≤a} zi (ki ).
E log2 1 +
N0 + j6=i Zj
for all sets N ∈ N such that |N | ≥ Ni,0 , Ti (zi∗ , z−i ) −
ki
Ti (zi , z−i ) > 0, for all policies zi ∈ Zi and all policies
Using the facts that the snr random variables Zj are positive
zj ∈ Bj of user j ∈ N /{i}.
and k = maxj {max{kj : kj ∈ Kj }} we have,
The next lemma provides an upper bounds on the proba
bility of sum of independent snr random variables Zj correX
k
A(zi∗ , zi , N ) ≥ − P
Z j ≤ a
sponding to policies zj ∈ Bj of user j.
N0
2k
2N0 + (|N | − 1)γ0
j6=i
21
Using lemma 8, we get the desired result.
The next lemma provides bounds on partial sums of alternating series. ln(1 + x) in the interval x ∈ (−1, 1). The proof
of this lemma is given in theorem 8.16 [29].
Lemma 10. Let {an } be a decreasing sequence of real
numbers
converging to 0. Then the alternating series
P∞
n+1
(−1)
an converges. Let s represent the series and
n=1
sn is its partial sums, i.e„
s=
∞
X
(−1)n+1 an ,
n=1
n
X
sn =
(−1)k+1 ak .
Then,
B(zi , zi∗ , N ) =
Then for each integer n ∈ N,
0 < (−1)n (s − sn ) < an+1
ki
where
fi (ki ) = zi∗ (ki ) − zi (ki ).
We define the term Sl (Z−i , ki ) for each snr ki and each
l ∈ {0, 1} as,
r
l+1
X
(−1)r+1
ki
.
Sl (Z−i , ki ) =
g(Z−i , N )
r
r=1
We use the above lemma to prove a technical result required
to prove theorem 14.
Lemma 11. Let x be a real number in the interval [0, 1), then
we have
2
X
(x)k
k=1
k
(−1)k+1 <
x3
,
3
∞
X
(x)k
k=1
k
where B1 (zi , zi∗ , N ) =
X
E[(ln(Z−i , ki ) − Sl (Z−i , ki )) 1{Pj6=i Zj >a} ]fi (ki ),
ki
and
X
E[Sl (Z−i , ki )1{Pj6=i Zj >a} ]fi (ki ).
ki
x2
.
2
B1 (zi , zi∗ , N )
We have
≥
X
E[(ln(Z−i , ki ) − Sl (Z−i , ki )) 1{Pj6=i Zj >a} ]zi (ki ).
−
Proof: At x ∈ [0, 1), we have
log(1 + x) =
B(zi , zi∗ , N ) = B1 (zi , zi∗ , N ) + B2 (zi , zi∗ , N ),
B2 (zi , zi∗ , N ) =
and
log(1 + x) − x <
E[ln(Z−i , ki )1{Pj6=i Zj >a} ]fi (ki ),
We now write the term B(zi , zi∗ , N ) as sum of two terms and
provide lower bounds on each of them. This breakup depends
upon whether the policy zi is 1− sensitive optimal or not.
If the policy zi is not 1− sensitive optimal, then set l = 0,
otherwise l = 1. Now let,
k=1
0 < log(1 + x) −
X
ki
(−1)k+1 .
The results stated in the lemma follow from lemma 10.
Lemma 12. Suppose the policy zi is not 1−sensitive optimal,
then set l = 0, otherwise set l = 1. We then have for all sets
N satisfying |N | > γa0 + 1, the term
#
" l+1
∗
X
β
(z
,
z
)
r
i
i
B(zi , zi∗ , N ) ≥
r(N
+
(|N
| − 1)k)r
0
r=1
Using lemma 11 and the fact that the indicator function is
bounded above by 1, we have
"
l+2 #
X
ki
∗
E
B1 (zi , zi , N ) ≥ −
zi (ki ),
g(Z−i , N )
ki
=−
X h
E
ki
ki
g(Z−i , N )
2k
2N0 + (|N | − 1)γ0
l+2
P
k
N0
{
P
j6=i (Zj −
E(Zj )
)>0}
2
i
zi (ki ).
{
c2
(|N | − 1)).
4
Using lemma 8 and the properties that ki ≤ k and E[Zj ] ≥ γ0 ,
we have
where
l+2
l+2 l+1
2k
k
k
∗
B
(z
,
z
,
N
)
≥
−
1 i i
c3 (a, N ) = −
−
c1 (a, N ).
2N0 + (|N | − 1)γ0
N0
N0
−
+1
l+2
1
+ c3 (a, N ) exp(−
Proof: We define
ln(Z−i , ki ) = log2 1 +
ki
g(Z−i , N )
where
g(Z−i , N ) = N0 +
X
j6=i
Zj .
.
−
E(Zj )
)≤0}
j6=i (Zj −
2
l+2
exp(−
c2
(|N | − 1)).
4
To get a lower bound on the term B2 (zi , zi∗ , N ), observe that
#
" l+1
X βr (z ∗ , zi )
i
∗
P
1{ j6=i Zj >a} .
B2 (zi , zi , N ) = E
r(g(Z−i , N ))r
r=1
22
Observing that the random variable Zj ≤ k for all users j, we
have from 8, we have
#
" l+1
X
βr (zi∗ , zi )
∗
B2 (zi , zi , N ) ≥
r(N0 + (|N | − 1)k)r
r=1
R EFERENCES
[1] S. Hanly and D. Tse, “Multi access fading channels. II. Delay-limited
capacities,” IEEE Trans. on Info. Theory , vol. 44, no. 7, 1998
[2] J. Edmonds, “Submodular functions, matroids and certain polyhedra,”
in Proc. Calgary Int. Conf. Combinatorial Structures and Applications .
(Calgary, Alta, June 1969), pp. 69–87
[3] Z Han, D Niyato, W Saad, T Basar, A Hjorungnes,“Game Theory in
Wireless and Communication Networks"
[4] L. Lai and H. El Gamal, “The water-filling game in fading multiple·(1 − c1 (a, N ) exp (−c2 (|N | − 1))).
access channels, IEEE Transaction of Information Theory, vol. 54, no.
5, pp. 2110?122, May 2008.
[5] G. Scutari, D P. Palomar, and S Barbarossa, “Optimal Linear PrecodHence we have that the term B(zi , zi∗ , N ) ≥
ing Strategies for Wideband Noncooperative Systems Based on Game
" l+1
#
Theory Part I: Nash Equilibria" IEEE Transaction on Signal Processing,
∗
X
βr (zi , zi )
56, no. 3, pp. 1230-1249, March 2008.
(1 − c1 (a, N ) exp (−c2 (|N | − 1))) [6] vol.
G.
Scutari,
D P. Palomar, and S Barbarossa, “Optimal Linear Precoding
r
r(N0 + (|N | − 1)k)
r=1
Strategies for Wideband Noncooperative Systems Based on Game Thel+2 l+1
ory Part II: Algorithms" IEEE Transaction on Signal Processing, vol.
c2
k
2k
56, no. 3, pp. 1250-1267, March 2008.
−
exp(− (|N | − 1)).
−
[7] G. He, M. Debbah, E. Altman , “A Bayesian Game-Theoretic Approach
2N0 + (|N | − 1)γ0
N0
4
for Distributed Resource Allocation in Fading Multiple Access Channels” EURASIP Journal on Wireless Communications and Networking
On further manipulation, we have
Volume 2010.
" l+1
#
[8] T. Heikkinen, “A potential game approach to distributed power control
X
βr (zi∗ , zi )
and scheduling”, Computer Networks, vol. 50, no. 13, pp. 2295 2?311,
B(zi , zi∗ , N ) ≥
September 2006.
r
r(N
+
(|N
|
−
1)k)
0
r=1
[9] E. Altman, K. Avrachenkov, G. Miller, and B. Prabhu, “Uplink dynamic
discrete power control in cellular networks”,INRIA Technical Report
5818, 2006
l+2
[10] E Altman, K Avrachenkov, N Bonneau, M Debbah, R El-Azouzi,
c2
2k
D Menasch, “Constrained Stochastic Games in Wireless Networks”
+ c3 (a, N ) exp(− (|N | − 1)),
−
Globecomm, 2008
2N0 + (|N | − 1)γ0
4
[11] E. Altman, “Constrained cost coupled stochastic games with independent
state information to be published” Operations Research Letters, vol. 36,
which is the required result.
no. 2, Mar. 2008.
[12]
Lasry, P.L. Lions, “Mean Field Games”
Lemma 13. For all subsets N ⊆ N of users satisfying |N | > [13] J.M.
O. Gu’eant, J.M Lasry, P.L. Lions, “Mean field games and applications
a
”
γ0 + 1, the term
[14] Y. Hayel, H. Tembine, E. Altman, and R. El-Azouzi, “A Markov
Decision Evolutionary Game for Individual Energy Management.”
c4 (a, N ) ≥ c5 (a).
[15] H Tembine, P Vilanova, M Assaad, M Debbah, “Mean field stochastic
games for SINR-based Medium Access Control.”
Proof: Note that we have
[16] F.Mériaux, S.Lasaulce, H.Tembine, “Stochastic Differential Games
and Energy-Efficient Power Control.” Dynamic Games Applications
(2013) 3:3–23
t2
c1 (a, N ) = exp 4γ0 a −
,
[17] H.Tembine, “Energy-constrained mean field games in wireless net2
(k) (|N | − 1)
works.”
[18] H.Tembine, “Risk-Sensitive Mean-Field Games” in IEEE Transactions
hence
on Automatic Control, Vol. 59, Issue 4.
[19] F. Ḿeriaux * , V.Varma, S.Lasaulce, “Mean Field Energy Games in
Wireless Networks”
c1 (a, N ) ≤ exp (4γ0 a) .
[20] E. Altman, K. Avrachenkov, N. Bonneau, M. Debbah, R. El-Azouzi, and
D. Menasch, “Constrained stochastic games: Dynamic control in wireAs,
less networks” Tech Report, 2007, www-net.cs.umass.edu/sadoc/mdp.
[21] D. Monderer and L. S. Shapley, "Potential games", Games and Economic
l+2 l+1
Behavior, vol. 14, no. 1, pp 124-143 May 1996.
k
k
[22] N Prashant and Vinod Sharma, “Stochastic Games on the Multiple
−
c1 (a, N ),
c3 (a, N ) = −
N0
N0
Access Channel” in arXiv:1210.7859, October 2008, “unpublished”.
[23] M. L. Puterman, “Markov Decision Process, Discrete Stochastic Dynamic Programming”
where the index l ∈ {0, 1}, we have
[24] G. B. Dantzig, J. Folkman, N. Shapiro , “On the Continuity of the
Minimum Set ofa Continuous Function”, Journal Of Mathematical
c3 (a, N ) ≥ −c5 (1 + exp (4γ0 a)) .
Analysis and Applications 17, 519-548 (1967)
[25] F. Berggren and R. Jantti, “Asymptotically Fair Transmission Scheduling
over Fading Channels”, IEEE Transactions on Wireless CommunicaHence we have,
tions, Vol. 3, No. 1, January 2004.
[26] A. Fu. E. Modiano and J. Tsitsiklis, “Optimal Transmission Scheduling
k
over a Fading Channel with Energy and Deadline Constraints”, IEEE
c4 (a, N ) = c3 (a, N ) −
c1 (a, N ),
Transactions on Wireless Communications, Vol. 5, No. 3, March 2006.
N0
[27] M. Shaked, J.G. Shanthikumar “Stochastic Orders-Springer(2007)”
≥ c3 (a, N ) − c5 exp (4γ0 a)
[28] D.P.Bertsekas, J.N.Tsitsiklis, “Parallel and Distributed Computation:
Numerical Methods”
≥ −c5 (1 + 2 exp (4γ0 a))
[29] T.M. Apostol “Mathematical Analysis”
= c5 (a).
[30] D. Bertsimas, J. N. Tsitsiklis “Introduction to Linear Optimization”
[31] D. P. Bertsekas “Convex Optimization Theory”
| 3cs.SY
|
Skin Lesion Segmentation: U-Nets versus Clustering
arXiv:1710.01248v1 [cs.CV] 27 Sep 2017
Bill S. Lin1 , Kevin Michael2 , Shivam Kalra3 , H.R. Tizhoosh3
1 Department of Mechanical and Mechatronics, University of Waterloo, Waterloo, Canada
2 Department of Systems Design Engineering, University of Waterloo, Waterloo, Canada
3 KIMIA Lab, University of Waterloo, Waterloo, Canada
Abstract— Many automatic skin lesion diagnosis systems use
segmentation as a preprocessing step to diagnose skin conditions
because skin lesion shape, border irregularity, and size can
influence the likelihood of malignancy. This paper presents,
examines and compares two different approaches to skin lesion
segmentation. The first approach uses U-Nets and introduces
a histogram equalization based preprocessing step. The second
approach is a C-Means clustering based approach that is much
simpler to implement and faster to execute. The Jaccard Index
between the algorithm output and hand segmented images by
dermatologists is used to evaluate the proposed algorithms. While
many recently proposed deep neural networks to segment skin
lesions require a significant amount of computational power
for training (i.e., computer with GPUs), the main objective
of this paper is to present methods that can be used with
only a CPU. This severely limits, for example, the number of
training instances that can be presented to the U-Net. Comparing
the two proposed algorithms, U-Nets achieved a significantly
higher Jaccard Index compared to the clustering approach.
Moreover, using the histogram equalization for preprocessing step
significantly improved the U-Net segmentation results.
Keywords— Skin lesion Segmentation; U-Nets; C-Means Clustering; Melanoma; Histogram Equalization; Color Space
sub-regions are just part of the surrounding skin. Finally, lack
of large labeled datasets (segmented by dermatologists) makes
training sophisticated networks rather difficult.
This paper presents and compares multiple approaches to
automatic segmentation of dermoscopic skin lesion images.
Firstly, we present some different segmentation approaches
using U-nets including one with a novel histogram-based
preprocessing technique. After this, the U-net approach is
compared with a much simpler clustering algorithm. While Unets and other deep networks often require a lot of training and
computational power, the approaches presented in this paper
are designed such that they do not require GPU power. All
training and testing in this paper have been done with a single
CPU.
The paper is organized as follows: Section II briefly reviews the literature of existing techniques for skin lesion
segmentation. Section III describes the proposed dataset used
to benchmark proposed methods. Section IV explains the
proposed methods; results and conclusion are discussed in
Section V and Section VI, respectively.
I. I NTRODUCTION
II. BACKGROUND R EVIEW
With between 2 and 3 million cases occurring globally
each year, skin cancer is the most common cancer worldwide
[1], [2]. Since skin cancer occurs at the surface of the skin,
melanomas, one of the world’s most deadly cancer, can be
diagnosed with visual inspection by a dermatologist [1], [3].
To improve the accuracy, dermatologists use a dermatoscope,
which eliminates some of the surface reflection and enhances
the deeper layers of skin [4]. Dermatologists often look for
five specific signs when classifying melanomas. These include
skin lesion asymmetry, irregular borders, uneven distribution
of color, diameter, and evolution of moles over time [5]. Standard skin cancer computer-aided diagnosis systems include
five important steps: acquisition, preprocessing, segmentation,
feature extraction, and finally, classification [6]. Since the
shape and border of the skin lesion is extremely important
in the diagnosis, automatic diagnosis systems must be able to
accurately identify and segment the skin lesion in the image.
Automatic skin lesion segmentation contains many different
challenges. The lesion color, texture, and size can all vary
considerably for different patients. On top of this, medical
gauzes, hair, veins, and light reflections in the image make it
even more difficult to identify the skin lesion. Finally, in the
areas surrounding a skin lesion, there are often sub-regions that
are darker than others. This makes it difficult to consistently
identify which sub-regions are part of the lesion and which
Machine learning schemes like reinforcement learning and
neural networks have been used for segmentation of medical
images [19], [20], [21], [22], [23]. Since skin lesion segmentation is frequently used as an important first step in many
automated skin lesion recognition systems, there have been
many different approaches. A summary of several skin lesion
segmentation methods by Celebi et al. [7] reviews many recent
techniques proposed in literature. Frequently, proposed lesion
segmentation techniques include a preprocessing step which
removes unwanted artifacts such as hair or light reflections.
Another very common preprocessing step involves performing
color space transformation in order to make segmentation
easier. Common techniques that have recently been proposed
in the literature include, among others, histogram thresholding,
clustering and active contours. Most recently, deep neural
networks have led to many breakthroughs, producing excellent
results in image segmentation [8].
Color space transformations can help facilitate skin lesion
detection. For instance, Garnavi et al. [9], take advantage
of RGB, HSV, HSI, CIE-XYZ, CIE-LAB, and YCbCr color
spaces and use a thresholding approach to separate the lesion
from the background skin. Similarly, Yuan et al. [10] use
RGB, HSV, and CIE-LAB color spaces as input channels to
a fully convolutional-deconvolutional neural network in order
to segment the images.
To appear in proceedings of The IEEE Symposium Series on Computational Intelligence (IEEE SSCI 2017), Honolulu, Hawaii, USA, Nov. 27 – Dec 1, 2017
Hair removal is a very common preprocessing step used
to facilitate skin lesion segmentation. The presence of dark
hairs on the skin often significantly degrades the performance
of most automatic segmentation algorithms. For example, one
very popular software, Dullrazor [11], identifies the dark hair
locations using morphological operations. Afterwards, the hair
pixels are replaced with nearby skin pixels and the resulting
image is smoothed in order to remove the hair artifacts.
Many recent techniques uses similar morphology or diffusion
methods to remove hair [6], [9], [12].
Once the image has been preprocessed, many different
methods have been proposed for segmentation. Since most
skin lesions are darker than the surrounding skin, thresholding
is a very commonly used technique. However, further image
enhancement algorithms need to be performed in order to
make thresholding effective. For example, Fan et al. [6]
propose first enhancing the skin lesion image using color and
brightness saliency maps before applying a modified Otsu
threshold method. Similarly, Garnavi et al. [9] use a thresholding based approach on 25 color channels from a variety of
different color spaces in order to segment the images. Often
with histogram thresholding, a connected components analysis
needs to be performed after thresholding in order to remove
medical gauzes, pen marks, or other artifacts that have been
wrongly classified by thresholding.
Clustering-based methods are commonly used as well. For
instance, Schmid-Saugeon et al. [12] propose using a modified Fuzzy C-Means clustering algorithm on the two largest
principle components of the image in the LUV color space.
The number of clusters for the Fuzzy C-Mean algorithm is
determined by examining the histogram computed using the
two principal components of the image. Melli et al. [13]
propose using mean shift clustering to get a set of redundant
classes which are then merged into skin and lesion regions.
More recently, convolutional neural networks and other deep
artificial neural networks are proposed. Commonly, the image
is first preprocessed to remove artifacts. Afterwards, a neural
network is used to identify the lesion. For instance, in a work
proposed by Jafari et al. [14], the image is first preprocessed
using a guided filter before a CNN uses both local and global
information to output a label for each pixel.
Fig. 1. Sample skin lesion images from public data set used for ISBI 2017
challenge.
and trained a convolutional-deconvolutional neural network
with 500 epochs. Clustering was also used in some submissions, although the performance of clustering algorithms
were generally worse compared with that of the deep artificial
networks. A variety of different evaluation metrics have been
used in the literature. This includes the XOR measure, and
Specificity, Sensitivity, Precision, Recall, and error probability
[7]. The ISBI 2017 challenge used the Jaccard Index in order
to evaluate methods. In order to compare the algorithms and
methods proposed in this paper, the average Jaccard Index is
used to evaluate the proposed method’s performance.
The ISIC 2017 training set was used to both train and validate our approaches. This publicly available dataset contains a
total of 2000 images including 374 melanoma, 254 seborrheic
keratosis and 1372 benign nevi skin lesion images in JPG
format. The images are in many different dimensions and
resolutions. Each skin lesion has also been segmented by a
dermatologist; the 2000 binary masks corresponding to each
skin lesion image is provided in PNG format.
Figure 1 shows a collection of sample images from the
data set. As can be seen from the images, there are many
unwanted artifacts including reflections, hair, black borders,
pen marks, dermatoscope gel bubbles, medical gauzes, and
other instrument markings that make segmentation difficult.
Furthermore, there is a high variability in skin lesion shape,
size, color and location.
III. I MAGE DATA
The 2017 IEEE International Symposium on Biomedical
Imaging (ISBI) organized a skin lesion analysis challenge for
melanoma detection, ISIC 2017 (International Skin Imaging
Collaboration). The challenge released a public set of 2000
skin lesion images that have been segmented and classified
by dermatologists. There is also a validation and a test set
that is not publicly available for the competition. Lesion
segmentation is one of three parts of the competition. For
evaluation, the Jaccard Index is used to rank the algorithms.
This year, the vast majority of the participants, including the
top submissions, used a deep neural networks to segment
the images. For example, Yuan et al. [10] augmented the
2000 images using rotation, flipping, rotating, and scaling
IV. M ETHODOLOGY
The main objective of this paper is to develop different
approaches that can process skin lesion images, and produce
a binary mask with 1 corresponding to pixels that are part of
2
To appear in proceedings of The IEEE Symposium Series on Computational Intelligence (IEEE SSCI 2017), Honolulu, Hawaii, USA, Nov. 27 – Dec 1, 2017
the skin lesion, and 0 corresponding to pixels that are not part
of the skin lesion. The output format should be identical to
the binary masks segmented by the dermatologists in order to
evaluate the performance of the segmentation algorithms.
In order to compare the performance of algorithms with that
of other submissions in the ISBI challenge, the Jaccard Index,
which was the challenge evaluation metric, is used. Since all
of the approaches in this paper are performed without a GPU,
some approaches, in particular, the ones involving a U-Net,
take around several hours (in our case 8 hours) to train for each
validation fold. In order for the validation to complete within
a reasonable amount of time, 5 random cross validation folds
were performed with a 90% training (1800 images) and 10%
validation (200 images) split. In this methodology section, we
first discuss the performance of U-Nets followed by clustering.
The loss function used is the cross entropy between the
prediction and the ground truth. The Adam optimizer with a
constant learning rate of 0.0002 is used to train the network.
A dropout rate of 0.5 is used.
2) Pre- and Post-Processing: Before the data is sent to
the U-Net, a few pre-processing steps are performed. Firstly,
in order to increase the amount of contrast in the image,
the image is converted to HSI color space, and histogram
equalization is applied to the intensity (I) channel. After this,
the image is converted back to RGB format. The reason
for this equalization is that the contrast between the skin
lesions (usually darker) is more noticeable compared to the
surrounding skin (usually brighter). On top of this, as a
4th channel, the original intensity (I) channel with values
normalized to between 0 and 1 is added. Finally, as a 5th
channel, a 2-dimensional Gaussian centered at the middle of
the image is generated. The full-width half maximum of the
2D Guassian was set to be 125 pixels, which was roughly
where most of the skin lesions lie within the center from a
visual inspection of the training data.
To the authors’ best knowledge, this is the first time that
histogram equalization and an extra Gaussian channel is used
as an input to a U-Net. The image needs to be rescaled and
zero-padded so that the U-Net can perform segmentation. Each
image is rescaled such that the largest dimension is 250 pixels.
The smaller dimension is filled with white padding so that
the image becomes 250 × 250. Finally, in order to account
for cropping due to the unpadded convolutions, the image is
white padded with a border of 46 pixels so that the final size
is 342 × 342 pixels.
After training, some post-processing steps are applied to the
testing set in order to improve the result. Firstly, a morphology operation (using the “scipy” binary fill holes function) is
applied to the image in order to fill the holes in the segmented
image. After this, in order to find the best threshold to set for
the U-Net output, a gradient descent is applied to the threshold
using in order to determine the best threshold.
3) U-Net Algorithm: In order to test the effect of the preprocessing and post-processing, two different approaches are
compared. One with the pre-/post-processing and one without pre-/post-processing. The exact same U-Net architecture,
which is described in an earlier section, is used for both cases.
The pseudo-code in Figure 3 shows the algorithm without
pre-/post-processing. The pseudo-code in Figure 4 shows the
algorithm with pre/post-processing. The results obtained from
the two approaches are discussed in a later section.
A. U-Net Approach
The U-Net architecture used in image segmentation uses a
Python library for U-Nets by Akeret et al. [15] with the open
source library TensorFlow. The architecture is based on the
network proposed by Ronneberger et al. [16] for biomedical
image segmentation (Figure 2). As discussed in the paper by
Ronneberger et al., this network architecture can be trained
with very few images and is relatively fast to train and test.
Because the goal of this paper is to develop an algorithm
that works without the need of a GPU, the number of training
iterations that can be done on the U-Net is severely limited.
Unlike the top ISBI submissions that used Neural Networks,
which augment the data to tens of thousands of images and
set the number of epochs to a few hundred [10], the proposed
network was trained with a lot less training iterations.
The 1800 images are first augmented to 7200 images by
a combination of flips and rotations. Afterwards, just over
one epoch (10,000 training iterations) is applied to train
the network. This is roughly 1% of the number of training
iterations used by the top submissions in the ISBI challenge.
1) Architecture and Training: Similar to convolutional neural networks, which repeatedly apply convolutions, activation
functions and downsampling, the U-Net architecture extends
this by including symmetric expansive layers which contain
upsampling operators. The proposed architecture contains 4
contracting steps and 4 expansion steps. Each contraction step
contains two unpadded 3 × 3 convolutions each followed by
a ReLU activation function. Finally, a 2 × 2 max pooling
operation is used for down sampling. In each down sampling
step, the number of feature channels doubles.
In each expansion step, two unpadded 3 × 3 convolutions
each followed by a ReLU activation function is used. However,
instead of the 2 × 2 max pooling, an up-convolution layer
is added which halves the number of feature channels. The
features from the corresponding contraction layer are also
concatenated in the expansion layer.
At the end, a final layer of a 1 × 1 convolution is used to
map the feature vector to a desired binary decision. A pixelwise soft-max is applied in order to determine the degree of
which a certain pixel belongs to the skin lesion.
B. Clustering Algorithm
This approach leverages an unsupervised clustering technique to analyze the image and segment the lesion of interest.
It uses Fuzzy C-means clustering in order to split the image
up into a number of distinct regions of interest. It then
leverages k-means clustering to group the clusters based on
color features. Ultimately it selects the group of clusters with
the darkest color features and identifies it to be the mole.
3
To appear in proceedings of The IEEE Symposium Series on Computational Intelligence (IEEE SSCI 2017), Honolulu, Hawaii, USA, Nov. 27 – Dec 1, 2017
Fig. 2. Schematic illustration of U-Net according to [16].
Fig. 3. U-Net Algorithm without Pre-/Post-Processing (Algorithm 1A)
1) Pre-Processing: When conducting clustering to the images, there are two major sources of disturbance within the
image. Firstly, the presence of hair on-top of the skin causes
discrepancies in the algorithms ability to cluster, therefore
strands of hair were removed in a pre-processing step which
utilizes a morphological top-hat and thresholding technique.
This technique is outlined in pseudo code in Figure 5.
Secondly, a subset of the images were captured using
imaging devices with a circular field of view. Resultant images
captured using this type of device were circular in nature
with black pixels filling the outer boundaries of the circle.
These pixels were removed from the area of analysis and the
algorithm for doing so can be seen in Figure 6.
2) Clustering Algorithm Summary: The algorithm outlined
in the above sections can be found summarized in the pseudocode in Figure 7 and includes the pre-processing steps of
hair removal and outer-circle removal. This algorithm will be
referenced as Algorithm 2 for future sections. Experimentally,
when C = 5, at least one of the clusters is the mole itself. For
this reason, C = 5 was used for the algorithm.
Fig. 4. U-Net Approach with Pre/Post-Processing (Algorithm 1B)
4
To appear in proceedings of The IEEE Symposium Series on Computational Intelligence (IEEE SSCI 2017), Honolulu, Hawaii, USA, Nov. 27 – Dec 1, 2017
Fig. 5. Hair Removal Algorithm
Fig. 7. Clustering Algorithm
2017 ISBI competition, the results were similar to many other
submissions using deep networks. For example, the top ISIC
2017 challenge submission, Yuan et al. [10] who used fully
convolutional-deconvolutional networks obtained an average
Jaccard Index of 0.765. Similarly, Berseth [17] who also used
a U-Net obtained the second place in the 2017 challenge with
a Jaccard Index of 0.762. There were also many challenge
submissions that received a Jaccard Index between 0.444 and
0.665, some of which also used deep neural networks.
The biggest advantage using the proposed algorithm is
the number of training iterations required. For example,
Yuan et al. trained with 500 epochs on the 2,000 images.
Similarly, Berseth trained 200 epochs using 20,000 images
(after data augmentation). This means that the two networks
above required 1,000,000 and 4,000,000 training iterations,
respectively.
In comparison, the networks presented in this paper was
trained with only 10,000 training iterations (just over 5 epochs
of 1800 images). This means that the network only required
1% of the number of training iterations compared with the
other deep network methods. A low number of training iterations is used because the algorithms presented in this paper
is designed to work when only low computational power is
available (computers without GPUs).
Figure 8 shows a comparison of the skin lesion segmentation
algorithms for a few select skin lesion images. The first column
in the figure shows the input RGB image. The second column
contains the Ground Truth (binary mask manually segmented
by a dermatologist). Finally, the third and fourth columns
show the output of the U-Net trained with algorithms 1A
and 1B respectively. As is evident from the figure, while both
Fig. 6. Dark outer Circle Removal Algorithms
V. R ESULTS AND D ISCUSSIONS
As discussed in the Methodology section, the segmented
images are assessed using the Jaccard Index through 5 random
folds validation with 90%-10% train-test split. For the clustering method, since it is completely unsupervised, the average
Jaccard Index for the entire data set is also calculated.
The Jaccard index and dice values for 5 different folds are
presented in Table 1. Figure 8 shows sample segmentations.
Using U-Nets, for each random fold, the network is trained
using the 1800 training images. After training, the average
Jaccard Index score is obtained for the remaining 200 test
images. This is done for both the U-Net algorithm without
pre-/post-processing (Algorithm 1A) and the more complex
algorithm with pre-/post-processing (Algorithm 1B).
Examining the Table 1, it is clear that the preprocessing
steps (performing histogram equalization, etc.) significantly
improved the segmentation results. The average Jaccard Index
for Algorithm 1A was 53% with a standard deviation of 2%
while the average Jaccard Index for Algorithm 1B was 62%
with a standard deviation of 4.7%.
Comparing the U-Net results with state of the art neural
networks while it is lower than the top performers at the
5
To appear in proceedings of The IEEE Symposium Series on Computational Intelligence (IEEE SSCI 2017), Honolulu, Hawaii, USA, Nov. 27 – Dec 1, 2017
Table 1. Jaccard index (and Dice coefficient in parentheses) for 5 different folds (m is the average and σ the standard deviation).
Fold
1
2
3
4
5
m
σ
Algorithm 1A
0.54 (0.70)
0.51 (0.68)
0.55 (0.71)
0.54 (0.70)
0.50 (0.67)
0.53 (0.69)
0.02 (0.02)
Algorithm 1B
0.67 (0.80)
0.61 (0.76)
0.55 (0.71)
0.63 (0.77)
0.65 (0.79)
0.62 (0.77)
0.05 (0.04)
Algorithm 2
0.46 (0.63)
0.41 (0.58)
0.44 (0.61)
0.44 (0.61)
0.43 (0.60)
0.44 (0.61)
0.02 (0.02)
Fig. 8. Sample skin lesion segmentations. Top three rows: Typical examples of most skin lesion images; 3rd and 4rd row: Color variation in skin lesions;
5th row: Medical gauze example.
algorithm 1A and 1B are able to determine the location of the
lesion, algorithm 1B is better at determining the border of the
skin lesion since algorithm 1A produced a smaller image.
Given that the clustering based approach was unsupervised
and did not contain any “training” components, it was not
necessary to conduct cross-validation over multiple folds. The
output was always going to be the same for every image.
However, for the purposes of comparison against the U-nets
based approach, the values were calculated for each of the
folds used in the earlier results. The Jaccard Index calculated
can be found summarized in the Table 1, last column. The
Jaccard Index for the entire data set was calculated to be 0.443.
This value is quite low in comparison to the values obtained
The last three rows show some pathological cases where
one or both of the algorithms are unable to properly segment
the image. This includes the skin lesion being very small, the
presence of another object such as medical gauze or sleeve,
and the skin lesion being extremely faint. Looking at the
pathological cases, it is clear that the algorithm that performs
histogram equalization performs better in most circumstances.
6
To appear in proceedings of The IEEE Symposium Series on Computational Intelligence (IEEE SSCI 2017), Honolulu, Hawaii, USA, Nov. 27 – Dec 1, 2017
using U-nets. However, it is important to remember that this
algorithm is much simpler and quicker to execute and does
not require expensive training.
Figure 8 contains a visualization from the clustering algorithm alongside the output from U-Nets for various image
types. Clearly the clustering results are good for images in
which no external disturbance exists and color variation is
minimized. However, the algorithm performs poorly when
artifacts are present. For example, the algorithm incorrectly
identifies the black border, the medical gauze, and other dark
objects as the skin lesion.
When comparing this with other submissions in the ISIC
challenge, this approach would be on par with the 21st
place submission. Amongst the competition entries, only one
submission followed a clustering based approach submitted
by Alvarez et al. [18]. Their solution produced a significantly
better result with a Jaccard Index of approximately 0.679.
The main reason for this is that their approach uses a much
more complex technique to classify the clusters involving
an ensemble method which leveraged random forests and
support vector machines. When comparing the results from
both approaches it is clear by virtue of the Jaccard Index that
the U-Net based approach (with preprocessing) is superior.
[3] R. Amelard, ”High-Level Intuitive Features (HLIFs) for Melanoma
Detection,” University of Waterloo, Waterloo, 2013.
[4] D. Gutman, C. C. F. Noel, E. Celebi, B. Helba, M. Marchetti, N. Mishra
and A. Halpern, ”Skin Lesion Analysis toward Melanoma Detection:
A Challenge at the International Symposium on Biomedical Imaging
(ISBI),” 2016.
[5] Melanoma Research Foundation, ”The ABCDEs of Melanoma,”
2017. [Online]. Available: https://www.melanoma.org/understandmelanoma/diagnosing-melanoma/detection-screening/abcdes-melanoma.
[Accessed 2017 February 2017].
[6] H. Fan, F. Xie, Y. Li, Z. Jiang and J. Liu, ”Automatic Segmentation of
Dermoscopy Images using Saliency Combined with Otsu Threshold,”
Computers in Biology and Medicine, vol. in press, 2017.
[7] M. E. Celebi, G. Schaefer, H. Iyatomi and W. V. Stoecker, ”Lesion Border Detection in Dermoscopy Images,” Computerized Medical Imaging
and Graphics, vol. 33, no. 2, pp. 148-153, 2009.
[8] L. Yu, H. Chen, J. Qin and P.-A. Heng, ”Automated Melanoma Recognition in Dermoscopy Images via Very Deep Residual Networks,” IEEE
Transactions on Medical Imaging, vol. 36, no. 4, pp. 994-1004, April
2017.
[9] R. Garnavi, M. Aldeen, M. E. Celebi, A. Bhuiyan, C. Dolianitis and
G. Varigos, ”Automatic Segmentation of Dermoscopy Images Using
Histogram Thresholding on Optimal Color Channels,” International
Journal of Medicine and Medical Sciences, vol. 1, no. 2, pp. 126-34,
2010.
[10] Y. Yuan, M. Chao and Y.-C. Lo, ”Automatic Skin Lesion segmentation
with fully convolutional-deconvolutional networks,” New York, 2017.
[11] T. Lee, V. Ng, R. Gallagher, A. Coldman and D. McLean, ”Dullrazor: A
software approach to hair removal from images,” Computers in Biology
and Medicine, vol. 27, no. 6, pp. 533-543, 1997.
[12] P. Schmid-Saugeon, J. Guillod and J.-P. Thiran, ”Towards a computeraided diagnosis system for pigmented skin lesions,” Computerized
Medical Imaging and Graphics, vol. 27, no. 1, pp. 65-78, 2003.
[13] R. Melli, G. Costantino and R. Cucchiara, ”Comparison of color clustering algorithms for segmentation of dermatological images,” Medical
Imaging, pp. 61443S-61443S, 15 march 2006.
[14] M. H. Jafari, N. Karimi, E. N. Esfahani, Samavi, Shadrokh, S. M. R.
Soroushmehr, K. R. Ward and K. Najarian, ”Skin Lesion Segmentaion
in Clinical Images Using Deep Learning,” in IEEE International Conference on Pattern Recognition (ICPR), Cancun, Mexico, 2016.
[15] J. Akeret, C. Chang, A. Lucchi and A. Refregier, ”Radio frequency
interference mitigation using deep convolutional neural networks,” Astronomy and Computing, vol. 18, pp. 35-39, 2017.
[16] O. Ronneberger, P. Fischer and T. Brox, ”U-Net: Convolutional Networks for Biomedical Image Segmentation,” in International Conference
on Medical Image Computing and Computer-Assisted Intervention,
Springer International Publishing, 2015.
[17] M. Berseth, ”ISIC 2017 - Skin Lesion Analysis Towards Melanoma
Detection,” 2017.
[18] D. Alvarez and M. Iglesias, ”k-Means Clustering and Ensemble of
Regressions: An Algorithm for the ISIC 2017 Skin Lesion Segmentation
Challenge,” Oviedo, Spain, 2017.
[19] Sahba, Farhang, Hamid R. Tizhoosh, and Magdy MMA Salama. ”Application of opposition-based reinforcement learning in image segmentation.” Computational Intelligence in Image and Signal Processing, 2007.
CIISP 2007. IEEE Symposium on. IEEE, 2007.
[20] Shokri, Maryam, and Hamid R. Tizhoosh. ”Using reinforcement learning
for image thresholding.” Electrical and Computer Engineering, 2003.
IEEE CCECE 2003. Canadian Conference on. Vol. 2. IEEE, 2003.
[21] Othman, Ahmed, and Hamid Tizhoosh. ”Segmentation of breast ultrasound images using neural networks.” Engineering Applications of
Neural Networks (2011): 260-269.
[22] Cheng, Kuo-Sheng, Jzau-Sheng Lin, and Chi-Wu Mao. ”The application
of competitive Hopfield neural network to medical image segmentation.”
IEEE transactions on medical imaging 15.4 (1996): 560-567.
[23] Ozkan, Mehmet, Benoit M. Dawant, and Robert J. Maciunas. ”Neuralnetwork-based segmentation of multi-modal medical images: a comparative and prospective study.” IEEE transactions on Medical Imaging 12.3
(1993): 534-544.
VI. C ONCLUSION AND F UTURE W ORK
In conclusion, this paper presented an effective approach
to automatically segmenting skin lesions particularly when
computers with GPUs may not be available. We described
two algorithms, one leveraging U-Nets and another using unsupervised clustering, and put forward effective pre-processing
techniques for both approaches to improve the segmentation
performance. It is noted that the solutions outlined in this
paper are on par with the mid-tier submissions in the ISIC
competition while utilizing significantly less training resources
than many submissions.
After comparing the clustering and U-Net algorithms, it
was clear that U-Nets produce a much better segmentation
result, especially using the histogram equalization based preprocessing step. The network only required 1% of the number
of training iterations compared with the other deep network
methods. This could be particularly beneficial when only low
computational power is available (computers without GPUs)
In the future, it would be interesting to find out if the
histogram equalization based pre-processing step also significantly improves segmentation performance when the U-Net is
trained over hundreds of epochs, instead of just one epoch,
with a more powerful GPU similar to the top ISIC 2017
challenge submissions.
R EFERENCES
[1] R. A. Schwartz, Skin Cancer: Recognition and Management, 2nd
Edition, Massachusetts: Blackwell Publishing Inc., 2008.
[2] World Health Organization, ”Skin Cancers,” Ultraviolet radiation and
the INTERSUN Programme, 2017. [Online].
7
| 1cs.CV
|
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE
GROUPS OF LIE TYPE
arXiv:1601.04566v1 [math.GR] 18 Jan 2016
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Abstract. For a finite group G of Lie type and a prime p, we compare the automorphism
groups of the fusion and linking systems of G at p with the automorphism group of G itself.
When p is the defining characteristic of G, they are all isomorphic, with a very short list
of exceptions. When p is different from the defining characteristic, the situation is much
more complex, but can always be reduced to a case where the natural map from Out(G)
to outer automorphisms of the fusion or linking system is split surjective. This work is
motivated in part by questions involving extending the local structure of a group by a group
of automorphisms, and in part by wanting to describe self homotopy equivalences of BG∧
p
in terms of Out(G).
When p is a prime, G is a finite group, and S ∈ Sylp (G), the fusion system of G at S is
the category FS (G) whose objects are the subgroups of S, and whose morphisms are those
homomorphisms between subgroups induced by conjugation in G. In this paper, we are
interested in comparing automorphisms of G, when G is a simple group of Lie type, with
those of the fusion system of G at a Sylow p-subgroup of G (for different primes p).
Rather than work with automorphisms of FS (G) itself, it turns out to be more natural
in many situations to study the group Outtyp (LcS (G)) of outer automorphisms of the centric
linking system of G. We refer to Section 1 for the definition of LcS (G), and to Definition 1.2
for precise definitions of Out(S, FS (G)) and Outtyp (LcS (G)). These are defined in such a way
that there are natural homomorphisms
κ
µ
G
G
−→ Outtyp (LcS (G)) −−−−
−→ Out(S, FS (G))
Out(G) −−−−
and
κG = µG ◦ κG .
For example, if S controls fusion in G (i.e., if S has a normal complement), then Out(S, FS (G)) =
Out(S), and κG is induced by projection to S. The fusion system FS (G) is tamely realized by
G if κG is split surjective, and is tame if it is tamely realized by some finite group G∗ where
S ∈ Sylp (G∗ ) and FS (G) = FS (G∗ ). Tameness plays an important role in Aschbacher’s
program for shortening parts of the proof of the classification of finite simple groups by
classifying simple fusion systems over finite 2-groups. We say more about this later in the
introduction, just before the statement of Theorem C.
By [BLO1, Theorem B], Outtyp (LcS (G)) ∼
= Out(BG∧p ): the group of homotopy classes
of self homotopy equivalences of the p-completed classifying space of G. Thus one of the
2000 Mathematics Subject Classification. Primary 20D06. Secondary 20D20, 20D45, 20E42, 55R35.
Key words and phrases. groups of Lie type, fusion systems, automorphisms, classifying spaces.
C. Broto is partially supported by MICINN grant MTM2010-20692 and MINECO grant MTM2013-42293P.
J. Møller is partially supported by the Danish National Research Foundation through the Centre for Symmetry and Deformation (DNRF92) and by Villum Fonden through the project Experimental Mathematics
in Number Theory, Operator Algebras, and Topology.
B. Oliver is partially supported by UMR 7539 of the CNRS, and by project ANR BLAN08-2 338236,
HGRT.
All three authors wish to thank Københavns Universitet, the Universitat Autònoma de Barcelona, and
especially the Centre for Symmetry and Deformation in Copenhagen, for their hospitality while much of this
work was carried out.
1
2
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
motivations for this paper is to compute Out(BG∧p ) when G is a finite simple group of Lie
type (in characteristic p or in characteristic different from p), and compare it with Out(G).
Following the notation used in [GLS3], for each prime p, we let Lie(p) denote the class of
finite groups of Lie type in characteristic p, and let Lie denote the union of the classes Lie(p)
for all primes p. (See Definition 2.1 for the precise definition.) We say that G ∈ Lie(p) is of
adjoint type if Z(G) = 1, and is of universal type if it has no nontrivial central extensions
which are in Lie(p). For example, for n ≥ 2 and q a power of p, PSLn (q) is of adjoint type
and SLn (q) of universal type.
Our results can be most simply stated in the “equi-characteristic case”: when working
with p-fusion of G ∈ Lie(p).
Theorem A. Let p be a prime. Assume that G ∈ Lie(p) and is of universal or adjoint type,
and also that (G, p) ∼
6= (Sz(2), 2). Fix S ∈ Sylp (G). Then the composite homomorphism
κ
µ
G
G
κG : Out(G) −−−−−
−→ Outtyp (LcS (G)) −−−−−
−→ Out(S, FS (G))
is an isomorphism, and κG and µG are isomorphisms except when G ∼
= PSL3 (2).
Proof. Assume G is of adjoint type. When G ∼
6= GL3 (2), µG is an isomorphism by [O1,
Proposition 4.3]1 or [O2, Theorems C & 6.2]. The injectivity of κG = µG ◦ κG (in all cases)
is shown in Lemma 4.3. The surjectivity of κG is shown in Proposition 4.5 when G has Lie
rank at least three, and in Proposition 4.8 when G has Lie rank 1 and G ∼
6= Sz(2). When G
∼
has Lie rank 2, κG is onto (when G 6= SL3 (2)) by Proposition 4.12, 4.14, 4.15, 4.16, or 4.17.
(See Notation 4.1(H) for the definition of Lie rank used here.)
If G is of universal type, then by Proposition 3.8, G/Z(G) ∈ Lie(p) is of adjoint type
where Z(G) has order prime to p. Also, Out(G) ∼
= Out(G/Z(G)) by [GLS3, Theorem
c
∼
2.5.14(d)]. Hence FS (G) = FS (G/Z(G)) and LS (G) ∼
= LcS (G/Z(G)); and κG and/or κG is
an isomorphism if κG/Z(G) and/or κG/Z(G) , respectively, is an isomorphism.
When G = PSL3 (2) and p = 2, Out(G) ∼
=
= C2 , while Outtyp (LcS (G)) ∼
= Out(S, FS (G)) ∼
C22 . When G = Sz(2) ∼
=
= Aut(C4 ) ∼
= C5 ⋊ C4 and p = 2, Out(G) = 1, while Outtyp (LcS (G)) ∼
C2 . Thus these groups are exceptions to Theorem A.
To simplify the statement of the next theorem, for finite groups G and H, we write G ∼p H
to mean that there are Sylow subgroups S ∈ Sylp (G) and T ∈ Sylp (H), together with an
∼
=
isomorphism ϕ : S −−−→ T which induces an isomorphism of categories FS (G) ∼
= FT (H)
(i.e., ϕ is fusion preserving in the sense of Definition 1.2).
Theorem B. Fix a pair of distinct primes p and q0 , and a group G ∈ Lie(q0 ) of universal
or adjoint type. Assume that the Sylow p-subgroups of G are nonabelian. Then there is a
prime q0∗ 6= p, and a group G∗ ∈ Lie(q0∗ ) of universal or adjoint type, respectively, such that
G∗ ∼p G and κG∗ is split surjective. If, furthermore, p is odd or G∗ has universal type, then
µG∗ is an isomorphism, and hence κG∗ is also split surjective.
Proof. Case 1: Assume p is odd and G is of universal type. Since µG is an isomorphism
by [O1, Theorem C], κG or κG∗ is (split) surjective if and only if κG or κG∗ is.
1Steve
Smith recently pointed out to the third author an error in the proof of this proposition. One can
get around this problem either via a more direct case-by-case argument (see the remark in the middle of
page 345 in [O1]), or by applying [O4, Theorem C]. The proof of the latter result uses the classification of
finite simple groups, but as described by Glauberman and Lynd [GLn, § 3], the proof in [O4] (for odd p)
can be modified to use an earlier result of Glauberman [Gl2, Theorem A1.4], and through that avoiding the
classification.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
3
By Proposition 6.8, we can choose a prime q0∗ and a group G∗ ∈ Lie(q0∗ ) such that either
(1.a) G∗ ∼
= G(q ∗ ) or 2 G(q ∗ ), for some G with Weyl group W and q ∗ a power of q0∗ , and has
a σ-setup which satisfies the conditions in Hypotheses 5.1 and 5.10, and
(1.a.1) −Id ∈
/ W and G∗ is a Chevalley group, or
(1.a.2) −Id ∈ W and q ∗ has even order in F×
p , or
(1.a.3) p ≡ 3 (mod p) and p|(q ∗ − 1); or
(1.b) p = 3, q0∗ = 2, G ∼
= 3D4 (q) or 2F4 (q) for q some power of q0 , and G∗ ∼
= 3D4 (q ∗ ) or
2
F4 (q ∗ ) for q ∗ some power of 2.
Also (by the same proposition), if p = 3 and G∗ = F4 (q ∗ ), then we can assume q0∗ = 2.
In case (1.b), κG∗ is split surjective by Proposition 6.9. In case (1.a), it is surjective by
Proposition 5.14. In case (1.a.1), κG∗ is split by Proposition 5.15(b,c). In case (1.a.3), κG∗
is split by Proposition 5.15(b). In case (1.a.2), if G∗ is a Chevalley group, then κG∗ is split
by Proposition 5.15(c).
This leaves only case (1.a.2) when G∗ is a twisted group. The only irreducible root systems
which have nontrivial graph automorphisms and for which −Id ∈ W are those of type Dn for
∗
even n. Hence G∗ = Spin−
2n (q ) for some even n ≥ 4. By the last statement in Proposition
∗
6.8, G is one of the groups listed in Proposition 1.10, and so q n ≡ −1 (mod p). Hence κG∗
is split surjective by Example 6.6(a), and we are done also in this case.
Case 2: Now assume p = 2 and G is of universal type. By Proposition 6.2, there is an odd
prime q0∗ , a group G∗ ∈ Lie(q0∗ ), and S ∗ ∈ Sylp (G∗ ), such that FS (G) ∼
= FS ∗ (G∗ ) and G∗ has
a σ-setup which satisfies Hypotheses 5.1 and 5.10. By the same proposition, if G∗ ∼
= G2 (q ∗ ),
∗
∗
∗ ∼
then we can arrange that q = 5 or q0 = 3. If G = G2 (5), then by Propositions 6.3 and
A.12, G∗ ∼2 G2 (3), κG2 (3) is split surjective, and µG2 (3) is injective.
In all remaining cases (i.e., G∗ ∼
6= G2 (q ∗ ) or q0∗ = 3), κG∗ is split surjective by Proposition
5.15(a). By Proposition A.3 or A.12, µG∗ is injective, and hence κG∗ is also split surjective.
Case 3: Now assume G is of adjoint type. Then G ∼
= Gu /Z for some Gu ∈ Lie(q0 ) of
universal type and Z ≤ Z(Gu ). By Proposition 3.8, Z = Z(Gu ) and has order prime to q0 .
By Case 1 or 2, there is a prime q0∗ 6= p and a group G∗u ∈ Lie(q0∗ ) of universal type such
that G∗u ∼p Gu and κG∗u is split surjective. Also, G∗u is p-perfect by definition of Lie(q0∗ )
(and since q0∗ 6= p), and H 2 (G∗u ; Z/p) = 0 by Proposition 3.8. Set G∗ = G∗u /Z(G∗u ). By
Proposition 1.7, with G∗u /Op′ (G∗u ) in the role of G, κG∗ is also split surjective.
It remains to check that G ∼p G∗ . Assume first that Gu and G∗u have σ-setups which
satisfy Hypotheses 5.1. Fix S ∈ Sylp (Gu ) and S ∗ ∈ Sylp (G∗u ), and a fusion preserving isomorphism ϕ : S −−−→ S ∗ (Definition 1.2(a)). By Corollary 5.9, Z(FS (Gu )) = Op (Z(Gu ))
and Z(FS ∗ (G∗u )) = Op (Z(G∗u )). Since ϕ is fusion preserving, it sends Z(FS (Gu )) onto
Z(FS ∗ (G∗u )), and thus sends Op (Z(Gu )) onto Op (Z(G∗u )). Hence ϕ induces a fusion preserving isomorphism between Sylow subgroups of G = Gu /Z(Gu ) and G∗ = G∗u /Z(G∗u ).
The only cases we considered where G or G∗ does not satisfy Hypotheses 5.1 were those in
case (1.b) above. In those cases, G ∼
= 2F4 (q) or 3D4 (q) and G∗ ∼
= 2F4 (q ∗ ) or 3D4 (q ∗ ) for some
∗
∗
q and q , hence G and G are also of universal type (d = 1 in the notation of [Ca, Lemma
14.1.2(iii)]), and so there is nothing more to prove.
The last statement in Theorem B is not true in general when G∗ is of adjoint type. For
example, if G∗ ∼
=
= Outtyp (LcS ∗ (G∗ )) ∼
= PSL2 (9), p = 2, and S ∗ ∈ Syl2 (G∗ ), then Out(G∗ ) ∼
4
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
e∗ ∼
C22 , while Out(S ∗ , FS ∗ (G∗ )) ∼
= C2 . By comparison, if G
= SL2 (9) is the universal group,
2
e ∗ )) ∼
C
,
and
κ
and
µ
are
isomorphisms.
then Out(Se∗ , FSe∗ (G
= 2
e∗
e∗
G
G
As noted briefly above, a fusion system FS (G) is called tame if there is a finite group
G∗ such that G∗ ∼p G and κG∗ is split surjective. In this situation, we say that G∗ tamely
realizes the fusion system FS (G). By [AOV, Theorem B], if FS (G) is not tame, then some
extension of it is an “exotic” fusion system; i.e., an abstract fusion system not induced by
any finite group. (See Section 1 for more details.) The original goal of this paper was to
determine whether all fusion systems of simple groups of Lie type (at all primes) are tame,
and this follows as an immediate consequence of Theorems A and B. Hence this approach
cannot be used to construct new, exotic fusion systems.
Determining which simple fusion systems over finite 2-groups are tame, and tamely realizable by finite simple groups, plays an important role in Aschbacher’s program for classifying
simple fusion systems over 2-groups (see [AKO, Part II] or [A2]). Given such a fusion system F over a 2-group S, and an involution x ∈ S, assume that the centralizer fusion system
CF (x) contains a normal quasisimple subsystem E E CF (x). If E is tamely realized by a
finite simple group K, then under certain additional assumptions, one can show that the
entire centralizer CF (x) is the fusion system of some finite extension of K. This is part of
our motivation for looking at this question, and is also part of the reason why we try to give
as much information as possible as to which groups tamely realize which fusion systems.
Theorem C. For any prime p and any G ∈ Lie of universal or adjoint type, the p-fusion
system of G is tame. If the Sylow p-subgroups of G are nonabelian, or if p is the defining
characteristic and G ∼
6= Sz(2), then its fusion system is tamely realized by some other group
in Lie.
Proof. If S ∈ Sylp (G) is abelian, then the p-fusion in G is controlled by NG (S), and FS (G)
is tame by Proposition 1.6. If p = 2 and G ∼
= SL3 (2), then the fusion system of G is tamely
realized by PSL2 (9). In all other cases, the claims follow from Theorems A and B.
We have stated the above three theorems only for groups of Lie type, but in fact, we
proved at the same time the corresponding results for the Tits group:
Theorem D. Set G = 2F4 (2)′ (the Tits group). Then for each prime p, the p-fusion system
of G is tame. If p = 2 or p = 3, then κG is an isomorphism.
Proof. The second statement is shown in Proposition 4.17 when p = 2, and in Proposition
6.9 when p = 3. When p > 3, the Sylow p-subgroups of G are abelian (|G| = 211 · 33 · 52 · 13),
so G is tame by Proposition 1.6(b).
As one example, if p = 2 and G = PSL2 (17), then κG is not surjective, but G∗ = PSL2 (81)
(of adjoint type) has the same 2-fusion system and κG∗ is an isomorphism [BLO1, Proposition
7.9]. Also, κG∗ is non-split surjective with kernel generated by the field automorphism of
e∗ = SL2 (81),
order two by [BLO1, Lemma 7.8]. However, if we consider the universal group G
then κGe∗ and κGe∗ are both isomorphisms by [BL, Proposition 5.5] (note that Out(S, F ) =
Out(S) in this situation).
As another, more complicated example, consider the case where p = 41 and G = Spin−
4k (9).
By [St1, (3.2)–(3.6)], Outdiag(G) ∼
= C2 × C4 is generated by a diagonal
= C2 , and Out(G) ∼
element of order 2 and a field automorphism of order 4 (whose square is a graph automorphism of order 2). Also, µG is an isomorphism by Proposition A.3, so κG is surjective, or
split surjective, if and only if κG is. We refer to the proof of Lemma 6.5, and to Table 6.1 in
that proof, for details of a σ-setup for G in which the normalizer of a maximal torus contains
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
5
a Sylow p-subgroup S. In particular, S is nonabelian if k ≥ 41. By Proposition 5.15(d) and
Example 6.6(a,b), when k ≥ 41, κG is surjective, κG is split (with Ker(κG ) = Outdiag(G))
when k is odd, and κG is not split (Ker(κG ) ∼
= C2 × C2 ) when k is even. By Proposition
∗
∗
1.9(c), when k is even, G ∼41 G for G = Spin4k−1 (9), and κG∗ is split surjective (with
Ker(κG∗ ) = Outdiag(G∗ )) by Proposition 5.15(c). Thus FS (G) is tame in all cases: tamely
realized by G itself when k is odd and by Spin4k−1 (9) when k is even. Note that when k is
odd, since the graph automorphism does not act trivially on any Sylow p-subgroup, the pfusion system of G (equivalently, of SO−
4k (9)) is not isomorphic to that of the full orthogonal
−
group O4k
(9), so by [BMO, Proposition A.3(b)], it is not isomorphic to that of Spin4k+1 (9)
either (nor to that of Spin4k−1 (9) since its Sylow p-subgroups are smaller).
Other examples are given in Examples 5.16 and 6.6. For more details, in the situation of
Theorem B, about for which groups G the homomorphism κG is surjective or split surjective,
see Propositions 5.14 and 5.15.
The following theorem was shown while proving Theorem B, and could be of independent
interest. The case where p is odd was handled by Gorenstein and Lyons [GL, 10-2(1,2)].
Theorem E. Assume G ∈ Lie(q0 ) is of universal type for some odd prime q0 . Fix S ∈
Syl2 (G). Then S contains a unique abelian subgroup of maximal order, except when G ∼
=
Sp2n (q) for some n ≥ 1 and some q ≡ ±3 (mod 8).
Proof. Assume S is nonabelian; otherwise there is nothing to prove. Since q0 is odd, and since
the Sylow 2-subgroups of 2 G2 (32k+1 ) are abelian for all k ≥ 1 [Ree, Theorem 8.5], G must be
a Chevalley or Steinberg group. If G ∼
= 3D4 (q), then (up to isomorphism) S ∈ Syl2 (G2 (q))
by [BMO, Example 4.5]. So we can assume that G ∼
= r G(q) for some odd prime power q,
some G, and r = 1 or 2.
If q ≡ 3 (mod 4), then choose another prime power q ∗ ≡ 1 (mod 4) such that v2 (q ∗ − 1) =
v2 (q + 1) (where v2 (m) = k if 2k |n and 2k+1 ∤ n). Then hq ∗ i = h − qi and h − q ∗ i = hqi as
closed subgroups of (Z2 )× . By [BMO, Theorem A] (see also Theorem 1.8), there is a group
G∗ ∼
= t G(q ∗ ) (where t ≤ 2) whose 2-fusion system is equivalent to that of G. We can thus
assume that q ≡ 1 (mod 4). So by Lemma 6.1, G has a σ-setup which satisfies Hypotheses
5.2. By Proposition 5.12(a), S contains a unique abelian subgroup of maximal order, unless
q ≡ 5 (mod 8) and G ∼
= Sp2n (q) for some n ≥ 1.
In fact, when G ∼
= Sp2n (q) for q ≡ ±3 (mod 8), then S ∈ Syl2 (G) is isomorphic to
(Q8 )n ⋊ P for P ∈ Syl2 (Σn ), S contains 3n abelian subgroups of maximal order 22n , and all
of them are conjugate to each other in NG (S).
The main definitions and results about tame and reduced fusion systems are given in
Section 1. We then set up our general notation for finite groups of Lie type in Sections 2 and
3, deal with the equicharacteristic case in Section 4, and with the cross characteristic case
in Sections 5 and 6. The kernel of µG , and thus the relation between automorphism groups
of the fusion and linking systems, is handled in an appendix.
Notation: In general, when C is a category and x ∈ Ob(C), we let AutC (x) denote the
group of automorphisms of x in C. When F is a fusion system and P ∈ Ob(F ), we set
OutF (P ) = AutF (P )/Inn(P ).
For any group G and g ∈ G, cg ∈ Aut(G) denotes the automorphism cg (h) = ghg −1. Thus
for H ≤ G, gH = cg (H) and H g = c−1
g (H). When G, H, K are all subgroups of a group Γ,
6
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
we define
TG (H, K) = {g ∈ G | gH ≤ K}
HomG (H, K) = {cg ∈ Hom(H, K) | g ∈ TG (H, K)} .
We let AutG (H) be the group AutG (H) = HomG (H, H). When H ≤ G (so AutG (H) ≥
Inn(H)), we also write OutG (H) = AutG (H)/Inn(H).
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
7
1. Tame and reduced fusion systems
Throughout this section, p always denotes a fixed prime. Before defining tameness of
fusion systems more precisely, we first recall the definitions of fusion and linking systems of
finite groups, and of automorphism groups of fusion and linking systems.
Definition 1.1. Fix a finite group G and a Sylow p-subgroup S ≤ G.
(a) The fusion system of G is the category FS (G) whose objects are the subgroups of S, and
where MorFS (G) (P, Q) = HomG (P, Q) for each P, Q ≤ S.
(b) A subgroup P ≤ S is p-centric in G if Z(P ) ∈ Sylp (CG (P )); equivalently, if CG (P ) =
Z(P ) × CG′ (P ) for a (unique) subgroup CG′ (P ) of order prime to p.
(c) The centric linking system of G is the category LcS (G) whose objects are the p-centric
subgroups of G, and where MorLcS (G) (P, Q) = TG (P, Q)/CG′ (P ) for each pair of objects
P, Q. Let π : LcS (G) −−−→ FS (G) denote the natural functor: π is the inclusion on
objects, and sends the class of g ∈ TG (P, Q) to cg ∈ MorFS (G) (P, Q).
(d) For P, Q ≤ S p-centric in G and g ∈ TG (P, Q), we let [[g]]P,Q ∈ MorLcS (G) (P, Q) denote
the class of g, and set [[g]]P = [[g]]P,P if g ∈ NG (P ). For each subgroup H ≤ NG (P ),
[[H]]P denotes the image of H in AutL (P ) = NG (P )/CG′ (P ).
The following definitions of automorphism groups are taken from [AOV, Definition 1.13
& Lemma 1.14], where they are formulated more generally for abstract fusion and linking
systems.
Definition 1.2. Let G be a finite group with S ∈ Sylp (G), and set F = FS (G) and L =
LcS (G).
∼
=
(a) If H is another finite group with T ∈ Sylp (H), then an isomorphism ϕ : S −−−→ T is
called fusion preserving (with respect to G and H) if for each P, Q ≤ S,
HomH (ϕ(P ), ϕ(Q)) = ϕ ◦ HomG (P, Q) ◦ ϕ−1 .
(Composition is from right to left.) Equivalently, ϕ is fusion preserving if it induces an
∼
=
isomorphism of categories FS (G) −−−→ FT (H).
(b) Let Aut(S, F ) ≤ Aut(S) be the group of fusion preserving automorphisms of S. Set
Out(S, F ) = Aut(S, F )/AutF (S).
(c) For each pair of objects P ≤ Q in L, set ιP,Q = [[1]]P,Q ∈ MorL (P, Q), which we call the
inclusion in L of P in Q. For each P , we call [[P ]] = [[P ]]P ≤ AutL (P ) the distinguished
subgroup of AutL (P ).
(d) Let AutItyp (L) be the group of automorphisms α of the category L such that α sends
inclusions to inclusions and distinguished subgroups to distinguished subgroups. For
γ ∈ AutL (S), let cγ ∈ AutItyp (L) be the automorphism which sends an object P to
π(γ)(P ), and sends ψ ∈ MorL (P, Q) to γ ′ ψ(γ ′′ )−1 where γ ′ and γ ′′ are appropriate
restrictions of γ. Set
Outtyp (L) = AutItyp (L) {cγ | γ ∈ AutL (S)} .
(e) Let κG : Out(G) −−−−→ Outtyp (L) be the homomorphism which sends the class [α], for
α ∈ Aut(G) such that α(S) = S, to the class of the induced automorphism of L =
LcS (G).
8
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
(f) Define µG : Outtyp (L) −−−−→ Out(S, F ) by setting µG ([β]) = [βS |S ] for β ∈ AutItyp (LcS (G)),
where βS is the induced automorphism of AutL (S), and βS |S ∈ Aut(S) is its restriction
to S when we identify S with its image in AutL (S) = NG (S)/CG′ (S).
(g) Set κG = µG ◦ κG : Out(G) −−−→ Out(S, F ): the homomorphism which sends the class
of α ∈ NAut(G) (S) to the class of α|S .
By [AOV, Lemma 1.14], the above definition of Outtyp (L) is equivalent to that in [BLO2],
and by [BLO2, Lemma 8.2], both are equivalent to that in [BLO1]. So by [BLO1, Theorem 4.5(a)], Outtyp (LcS (G)) ∼
= Out(BG∧p ): the group of homotopy classes of self homotopy
equivalences of the space BG∧p .
We refer to [AOV, § 2.2] and [AOV, § 1.3] for more details about the definitions of κG and
µG and the proofs that they are well defined. Note that µ is defined there for an arbitrary
linking system, not necessarily one realized by a group.
We are now ready to define tameness. Again, we restrict attention to fusion systems of
finite groups, and refer to [AOV, § 2.2] for the definition in the more abstract setting.
Definition 1.3. For a finite group G and S ∈ Sylp (G), the fusion system FS (G) is tame if
there is a finite group G∗ which satisfies:
∼
=
• there is a fusion preserving isomorphism S −−−→ S ∗ for some S ∗ ∈ Sylp (G∗ ); and
• the homomorphism κG∗ : Out(G∗ ) −→ Outtyp (Lc (G∗ )) ∼
= Out(BG∗∧ ) is split surjective.
S
p
In this situation, we say that G∗ tamely realizes the fusion system FS (G).
The above definition is complicated by the fact that two finite groups can have isomorphic
fusion systems but different outer automorphism groups. For example, set G = P SL2 (9) ∼
=
∼
A6 and H = P SL2 (7) = GL3 (2). The Sylow subgroups of both groups are dihedral of
order 8, and it is not hard to see that any isomorphism between Sylow subgroups is fusion
preserving. But Out(G) ∼
= C2 (see Theorem 3.4 below). Also, κG is
= C22 while Out(H) ∼
an isomorphism, while κH fails to be onto (see [BLO1, Proposition 7.9]). In conclusion, the
2-fusion system of both groups is tame, even though κH is not split surjective.
This definition of tameness was motivated in part in [AOV] by an attempt to construct
new, “exotic” fusion systems (abstract fusion systems not realized by any finite group) as extensions of a known fusion system by an automorphism. Very roughly, if α ∈ AutItyp (LcS (G))
is not in the image of κG , and not in the image of κG∗ for any other finite group G∗ which
has the same fusion and linking systems, then one can construct and extension of FS (G) by
α which is not isomorphic to the fusion system of any finite group. This shows why we are
interested in the surjectivity of κG ; to see the importance of its being split, we refer to the
proof of [AOV, Theorem B].
It is usually simpler to work with automorphisms of a p-group which preserve fusion than
with automorphisms of a linking system. So in most cases, we prove tameness for the fusion
system of a group G by first showing that κG = µG ◦ κG is split surjective, and then showing
that µG is injective. The following elementary lemma will be useful.
Lemma 1.4. Fix a finite group G and S ∈ Sylp (G), and set F = FS (G). Then
(a) κG is surjective if and only if each ϕ ∈ Aut(S, F ) extends to some ϕ ∈ Aut(G), and
(b) Ker(κG ) ∼
= CAut(G) (S)/AutC (S) (G).
G
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
9
Proof. This follows from the following diagram
0
AutNG (S) (G)
/
NAut(G) (S)
/
/
Out(G)
restr
0
/
κG
AutNG (S) (S)
/
0
/
Aut(S, F )
/
Out(S, F )
with exact rows.
/
0
The next lemma can be useful when κG or κG is surjective but not split.
Lemma 1.5. Fix a prime p, a finite group G, and S ∈ Sylp (G).
b ≥ G is such that G E G,
b p ∤ |G/G|,
b
(a) Assume G
and OutGb (G) ≤ Ker(κG ). Then
c
c
∼
b = FS (G) and L (G)
b = L (G).
FS (G)
S
S
b ≥ G as in (a)
(b) If κG is surjective and Ker(κG ) has order prime to p, then there is G
such that κGb is split surjective. In particular, FS (G) is tame, and is tamely realized by
b
G.
b contains an element which
Proof. (a) Since OutGb (G) ≤ Ker(κG ), each coset of G in G
centralizes S. (Recall that κG is induced by the restriction homomorphism from NAut(G) (S)
b = FS (G) and Lc (G)
b = Lc (G).
to Aut(S, F ).) Thus FS (G)
S
S
(b) Since G and G/Op′ (Z(G)) have isomorphic fusion systems at p, we can assume that
Z(G) is a p-group. Set K = Ker(κG ) ≤ Out(G). Since H i (K; Z(G)) = 0 for i = 2, 3, by the
b
obstruction theory for group extensions [McL, Theorems IV.8.7–8], there is an extension G
∼
b G/G
b
of G by K such that G E G,
= K, and OutGb (G) = K. Since K = Ker(κG ) ≤ Ker(κG ),
c b
c
b
FS (G) = FS (G), and LS (G) = LS (G) by (a).
By [OV, Lemma 1.2], and since K E Out(G) and H i (K; Z(G)) = 0 for i = 1, 2, each
b which is unique modulo inner auautomorphism of G extends to an automorphism of G
b
tomorphisms. Thus Out(G) contains a subgroup isomorphic to Out(G)/K, and κGb sends
b
this subgroup isomorphically onto Outtyp (LcS (G)).
So κGb is split surjective, and FS (G) is
tame.
The next proposition is really a result about constrained fusion systems (cf. [AKO, Definition I.4.8]): it says that every constrained fusion system is tame. Since we are dealing
here only with fusion systems of finite groups, we state it instead in terms of p-constrained
groups.
Proposition 1.6. Fix a finite group G and a Sylow subgroup S ∈ Sylp (G).
(a) If CG (Op (G)) ≤ Op (G), then κG and µG are both isomorphisms:
κ
µ
G
G
Out(G) −−−−
−→ Outtyp (LcS (G)) −−−−
−→ Out(S, FS (G)) .
∼
∼
=
=
(b) If S is abelian, or more generally if NG (S) controls p-fusion in G, then FS (G) is tame,
and is tamely realized by NG (S)/Op′ (CG (S)).
Proof. (a) Set Q = Op (G), F = FS (G), and L = LcS (G). Then AutL (Q) = G, so (α 7→ αQ )
defines a homomorphism Φ : AutItyp (L) −−−→ Aut(G, S). For each α ∈ Ker(Φ), αQ = IdG
and hence α = IdL . (Here, it is important that α sends inclusions to inclusions.) Thus Φ is an
isomorphism. Also, α = cγ for some γ ∈ AutL (S) if and only if αQ = cg for some g ∈ NG (S),
10
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
so Φ factors through an isomorphism from Outtyp (L) to Aut(G, S)/AutG (S) ∼
= Out(G), and
this is an inverse to κG . Thus κG is an isomorphism.
In the terminology in [AKO, § I.4], G is a model for F = FS (G). By the uniqueness of
models (cf. [AKO, Theorem III.5.10(c)]), each β ∈ Aut(S, F ) extends to some χ ∈ Aut(G),
and χ is unique modulo AutZ(S) (G). Hence κG is an isomorphism, and so is µG .
(b) If NG (S) controls p-fusion in G, then NG (S) ∼p G. Also, NG (S) ∼p G∗ where G∗ =
NG (S)/Op′ (CG (S)), G∗ satisfies the hypotheses of (a), and hence tamely realizes FS (G). In
particular, this holds whenever S is abelian by Burnside’s theorem.
When working with groups of Lie type when p is not the defining characteristic, it is easier
to work with the universal groups rather than those in adjoint form (µG is better behaved
in such cases). The next proposition is needed to show that tameness for fusion systems of
groups of universal type implies the corresponding result for groups of adjoint type.
Proposition 1.7. Let G be a finite p-perfect group such that Op′ (G) = 1 and H2 (G; Z/p) = 0
(i.e., such that each central extension of G by a finite p-group splits). Choose S ∈ Sylp (G),
and set Z = Z(G) ≤ S. If FS (G) is tamely realized by G, then FS/Z (G/Z) is tamely realized
by G/Z.
Proof. Let H be the set of all P ≤ S such that P ≥ Z and P/Z is p-centric in G/Z, and let
c
H
LH
S (G) ⊆ LS (G) be the full subcategory with object set H. By [AOV, Lemma 2.17], LS (G)
is a linking system associated to FS (G) in the sense of [AOV, Definition 1.9]. Hence the
homomorphism
∼
=
R : Outtyp (LcS (G)) −−−−−−→ Outtyp (LH
S (G))
induced by restriction is an isomorphism by [AOV, Lemma 1.17].
c
Set F = FS (G), L = LH
S (G), G = G/Z, S = S/Z, F = FS (G), and L = LS (G) for short.
Consider the following square:
Out(G)
κG
Outtyp (L) ∼
= Outtyp (LcS (G))
/
O
µ
Out(G)
1−1 ν
κG
/
(1)
Outtyp (L) .
Here, µ sends the class of an automorphism of G to the class of the induced automorphism
of G = G/Z(G).
Assume that ν has been defined so that (1) commutes and ν is injective. If κG is onto, then
ν is onto and hence an isomorphism, so κG is also onto. Similarly, if κG is split surjective,
then κG is also split surjective. Thus F is tamely realized by G if F is tamely realized by
G, which is what we needed to show.
It thus remains to construct the monomorphism ν, by sending the class of α ∈ AutItyp (L)
to the class of a lifting of α to L. So in the rest of the proof, we show the existence and
uniqueness of such a lifting.
Let pr : L −−−→ L denote the projection. Let EndItyp (L) be the monoid of functors from L
to itself which send inclusions to inclusions and distinguished subgroups into distinguished
subgroups. (Thus AutItyp (L) is the group of elements of EndItyp (L) which are invertible.) We
will prove the following two statements:
e ∈ EndItyp (L) such that pr ◦ α
e = α ◦ pr.
(2) For each α ∈ AutItyp (L), there is a functor α
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
11
(3) If β ∈ EndItyp (L) is such that pr ◦ β = pr, then β = IdL .
Assume that (2) and (3) hold; we call α
e a “lifting” of α in the situation of (2). For each
I
α ∈ Auttyp (L), there are liftings α
e of α and α
e∗ of α−1 in EndItyp (L), and these are inverses
to each other by (3). Hence α
e ∈ AutItyp (L), and is the unique such lifting of α by (3) again.
α] when α
e is the unique lifting
Define ν : Outtyp (L) −−−→ Outtyp (L) by setting ν([α]) = [e
I
of α. This is well defined as a homomorphism on Auttyp (L) by the existence and uniqueness
of the lifting; and it factors through Outtyp (L) since conjugation by γ ∈ AutL (S) lifts to
conjugation by γ ∈ AutL (S) for any γ ∈ pr−1
S (γ).
Thus ν is a well defined homomorphism, and is clearly injective. The square (1) commutes
since for each β ∈ Aut(G) such that β(S) = S, κG ([β]) and νκG µ([β]) are the classes of
liftings of the same automorphism of L.
It remains to prove (2) and (3).
Proof of (2): For each α ∈ AutItyp (L), consider the pullback diagram
Le
ρ1
/5 L
❦ ❦
❦
α
e ❦ ❦
pr
ρ2
❦ ❦
❦
❦ pr
❦
α
❦
/L
/ L.
L
(4)
∼
=
Each functor in (4) is bijective on objects, and the diagram restricts to a pullback square of
e
morphism sets for each pair of objects in L (and their inverse images in L and L).
Since the natural projection G −−−→ G is a central extension with kernel Z, the projection
functor pr : L −−−→ L is also a central extension of linking systems in the sense of [5a2,
Definition 6.9] with kernel Z. Since ρ2 is the pullback of a central extension, it is also a
central extension of linking systems by [5a2, Proposition 6.10], applied with ω = pr∗ α∗ (ω0 ) ∈
Z 2 (L; Z), where ω0 is a 2-cocycle on L which determines the extension pr. By [BLO1,
Proposition 1.1], H 2 (|L|; Fp) ∼
= H 2 (G; Fp ), where the last group is zero by assumption. Hence
2
H (|L|; Z) = 0, so ω is a coboundary, and ρ2 is the product extension by [5a2, Theorem
6.13]. In other words, Le ∼
= LcZ (Z)×L, where LcZ (Z) has one object and automorphism group
Z, and there is a subcategory L0 ⊆ Le (with the same objects) which is sent isomorphically
to L by ρ2 . Set α
e = ρ1 ◦ (ρ2 |L0 )−1 .
We first check that α
e sends distinguished subgroups to distinguished subgroups. Let
e(P ). Then
prS : S −−−→ S = S/Z be the projection. Fix an object P in L, and set Q = α
−1
Q/Z = α(P/Z), and αP/Z ([[P/Z]]) = [[Q/Z]], so α
eP ([[P ]]) ≤ prS ([[Q/Z]]) = [[Q]].
For each subgroup P ∈ Ob(L), there is a unique element zP ∈ Z such that α
e(ιP,S ) =
ιαe(P ),S ◦ [[zP ]]αe(P ) . Note that zS = 1. Define a new functor β : L −−→ L by setting β(P ) = α
e(P )
on objects and for each ϕ ∈ MorL (P, Q), β(ϕ) = [[zQ ]]αe(Q) ◦ α
e(ϕ) ◦ [[zP ]]−1
.
Then
β
is
still
a
α
e(P )
lifting of α, and for each P :
−1
β(ιP,S ) = [[zS ]]S ◦ α
e(ιP,S ) ◦ [[zP ]]−1
e(P ),S ◦ [[zP ]]α
e(P ) ◦ [[zP ]]α
e(P ),S .
α
e(P ) = ια
e(P ) = ια
For arbitrary P ≤ Q, since ιαe(P ),eα(Q) is the unique morphism whose composite with ιαe(Q),S
is ιαe(P ),S (see [BLO2, Lemma 1.10(a)]), β sends ιP,Q to ιαe(P ),eα(Q) .
Thus, upon replacing α
e by β, we can assume that α
e sends inclusions to inclusions. This
finishes the proof of (2).
12
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Proof of (3): Assume that β ∈ EndItyp (L) is a lift of the identity on L. Let B(Z) be the
category with one object
* and with morphism group Z. Define a functor χ : L −−−→ B(Z)
by sending all objects in L to , and by sending a morphism [[g]] ∈ MorL (P, Q) to the unique
*
element z ∈ Z such that βP,Q ([[g]]) = [[gz]] = [[zg]]. (Recall that Z ≤ Z(G).)
Now,
H 1 (|L|; Fp ) ∼
= H 1 (G; Fp ) = 0 ,
= H 1 (BG; Fp ) ∼
= H 1 (|LcS (G)|; Fp ) ∼
where the first isomorphism holds by [5a1, Theorem B] and the second by [BLO1, Proposition
1.1]. Hence Hom(π1 (|L|), Fp) ∼
= H 1 (|L|; Fp) = 0, where the second
= Hom(H1 (|L|), Fp ) ∼
isomorphism holds by the universal coefficient theorem (cf. [McL, Theorem III.4.1]), and
so Hom(π1 (|L|), Z) = 0. In particular, the homomorphism χ
b : π1 (|L|) −−−→ π1 (|B(Z)|) ∼
=Z
induced by χ is trivial.
Thus for each ψ ∈ MorL (P, Q), the loop in |L| formed by ψ and the inclusions ιP,S and
ιQ,S is sent to 1 ∈ Z. Since β sends inclusions to inclusions, this proves that χP,Q (ψ) = 1,
and hence that βP,Q (ψ) = ψ. Thus β = IdL .
By Proposition 1.7, when proving tameness for fusion systems of simple groups of Lie type,
it suffices to look at the universal groups (such as SLn (q), SUn (q)) rather than the simple
groups (PSLn (q), PSUn (q)). However, it is important to note that the proposition is false if
we replace automorphisms of the linking systems by those of the fusion system. For example,
set G = SL2 (34 ) and G = P SL2 (34 ). Then S ∼
= Q32 and S ∼
= D16 , Out(S, FS (G)) =
Out(S) ∼
= C4 × C2 (and κG is an isomorphism), while Out(G) ∼
= Out(G) ∼
= C4 × C2 and
Out(S, FS (G)) = Out(S) ∼
= C2 × C2 .
We already gave one example of two groups which have the same fusion system but different
outer automorphism groups. That is a special case of the main theorem in our earlier paper,
where we construct many examples of different groups of Lie type with isomorphic fusion
systems. Since this plays a crucial role in Section 6, where we handle the cross characteristic
case, we restate the theorem here.
As in the introduction, we write G ∼p H to mean that there is a fusion preserving
isomorphism from a Sylow p-subgroup of G to one of H.
Theorem 1.8 ([BMO, Theorem A]). Fix a prime p, a connected reductive group scheme G
over Z, and a pair of prime powers q and q ′ both prime to p. Then the following hold.
(a) G(q) ∼p G(q ′ ) if hqi = hq ′ i as subgroups of Z×
p.
(b) If G is of type An , Dn , or E6 , and τ is a graph automorphism of G, then τG(q) ∼p τG(q ′ )
if hqi = hq ′ i as subgroups of Z×
p.
(c) If the Weyl group of G contains an element which acts on the maximal torus by inverting
all elements, then G(q) ∼p G(q ′ ) (or τG(q) ∼p τG(q ′ ) for τ as in (b)) if h − 1, qi =
h − 1, q ′ i as subgroups of Z×
p.
(d) If G is of type An , Dn for n odd, or E6 , and τ is a graph automorphism of G of order
two, then τG(q) ∼p G(q ′ ) if h − qi = hq ′ i as subgroups of Z×
p.
The next proposition is of similar type, but much more elementary.
Proposition 1.9. Fix an odd prime p, a prime power q prime to p, n ≥ 2, and ε ∈ {±1}.
Then
(a) Sp2n (q) ∼p SL2n (q) if ordp (q) is even;
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
13
(b) Sp2n (q) ∼p Spin2n+1 (q); and
(c) Spinε2n (q) ∼p Spin2n−1 (q) if q is odd and q n 6≡ ε (mod p).
±
Proof. If we replace Spin±
m (q) by SOm (q) in (b) and (c), then these three points are shown
in [BMO, Proposition A.3] as points (d), (a), and (c), respectively. When q is a power of
2, (b) holds because the groups are isomorphic (see [Ta, Theorem 11.9]). So it remains to
show that
Spinεm (q) ∼p Ωεm (q) ∼p SOεm (q)
for all m ≥ 3 (even or odd) and q odd. The first equivalence holds since p is odd and Ωεm (q) ∼
=
Spinεm (q)/K where |K| = 2. The second holds by Lemma 1.5(a), and since OutSOεm (q) (Ωεm (q))
is generated by the class of a diagonal automorphism of order 2 (see, e.g., [GLS3, § 2.7]) and
hence can be chosen to commute with a Sylow p-subgroup. This last statement is shown in
Lemma 5.8 below, and holds since for appropriate choices of algebraic group G containing
the given group G, and of maximal torus T ≤ G, a Sylow p-subgroup of G is contained in
NG (T ) (see [GLS3, Theorem 4.10.2]) and the diagonal automorphisms of G are induced by
conjugation by elements in NT (G) (see Proposition 3.5(c)).
Theorem 1.8 and Proposition 1.9, together with some other, similar relations in [BMO],
lead to the following proposition, which when p is odd provides a relatively short list of
“p-local equivalence class representatives” for groups of Lie type in characteristic different
from p.
Proposition 1.10. Fix an odd prime p, and assume G ∈ Lie(q0 ) is of universal type for
some prime q0 6= p. Assume also that the Sylow p-subgroups of G are nonabelian. Then
there is a group G∗ ∈ Lie(q0′ ) of universal type for some q0′ 6= p, such that G∗ ∼p G and G∗
is one of the groups in the following list:
(a) SLn (q ′ ) for some n ≥ p; or
(b) Spinε2n (q ′ ), where n ≥ p, ε = ±1, (q ′ )n ≡ ε (mod p), and ε = +1 if n is odd; or
(c) 3D4 (q ′ ) or 2F4 (q ′ ), where p = 3 and q ′ is a power of 2; or
(d) G(q ′ ), where G = G2 , F4 , E6 , E7 , or E8 , p |W (G)|, and q ′ ≡ 1 (mod p); or
(e) E8 (q ′ ), where p = 5 and q ′ ≡ ±2 (mod 5).
Furthermore, in all cases except (c), we can take q0′ to be any given prime whose class
generates (Z/p2 )× , and choose G∗ so that q ′ = (q0′ )b where b|(p − 1)pk for some k.
Proof. Let q be such that G ∼
= τG(q) for some τ and some G. Thus q is a power of q0 . Fix a
′
prime q0 as specified above. By Lemma 1.11(a), there are positive integers b, c, and powers
q ′ = (q0′ )b and q ∨ = (q0′ )c such that hqi = hq ′ i, h − qi = hq ∨ i, and b, c|(p − 1)pℓ for some ℓ ≥ 0.
(i)
(ii)
Assume G ∼
= 3D4 (q). Since p 6= q0 , and since S ∈ Sylp (G)
= Sz(q), 2 G2 (q), 2F4 (q), or G ∼
is nonabelian, p divides the order of the Weyl group W of G by [GL, 10-1]. The Weyl
group of B2 is a 2-group, and 2 and 3 are the only primes which divide the orders of
the Weyl groups of G2 , F4 , and D4 . Hence p = 3, G ∼
6= 2 G2 (q) since that is defined
only in characteristic 3, and so G ∼
= 2F4 (q) or 3D4 (q). Set G∗ = 2F4 (q ′ ) or 3D4 (q ′ ),
′
∗
respectively, where q0 = 2. Then G ∼p G, and we are in case (c).
If G = SUn (q) or 2E6 (q), then by Theorem 1.8(d), G ∼p G∗ where G∗ ∼
= SLn (q ∨ ) or
E6 (q ∨ ), respectively. So we can replace G by a Chevalley group in these cases.
14
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
(iii) Assume G = Sp2n (q) for some n and q. If ordp (q) is even, then by Proposition 1.9(a),
G ∼p SL2n (q). If ordp (q) is odd, then ordp (q ∨ ) is even (recall that q ∨ ≡ −q (mod p)),
and G ∼p Sp2n (q ∨ ) by Theorem 1.8(c). So G is always p-locally equivalent to a linear
group in this case.
(iv) Assume G = Spin2n+1 (q) for some n and q. Then G ∼p Sp2n (q) by Proposition 1.9(b).
So G is p-locally equivalent to a linear group by (iii).
(v)
If G = SLn (q), set G∗ = SLn (q ′ ). Then G∗ ∼p G by Theorem 1.8(a), n ≥ p since the
Sylow p-subgroups of G are nonabelian, and we are in the situation of (a).
(vi) Assume G = Spinε2n (q) for some n and q, and ε = ±1. If q is a power of 2, then by
using point (a) or (b) of Theorem 1.8, we can arrange that q be odd. If q n 6≡ ε (mod p),
then G ∼p Spin2n−1 (q) by Proposition 1.9(c), and this is p-equivalent to a linear group
by (iv). So we are left with the case where q n ≡ ε (mod p). If n is odd and ε = −1,
ε
∨
∗
′
set G∗ = Spin+
2n (q ) ∼p G (Theorem 1.8(d)). Otherwise, set G = Spin2n (q ) ∼p G
(Theorem 1.8(a,b)). In either case, we are in the situation of (b).
We are left with the cases where G = G(q) for some exceptional Lie group G. By [GL, 101(2)] and since the Sylow p-subgroups of G are nonabelian, p |W (G)|. If ordp (q) = 1, then
G∗ = G(q ′ ) ∼p G by Theorem 1.8(a). If ordp (q) = 2 and G 6= E6 , then G∗ = G(q ∨ ) ∼p G by
Theorem 1.8(c), where q ∨ ≡ 1 (mod p). In either case, we are in the situation of (d).
If ordp (q) = 2 and G = E6 (q), then hqi = h − q 2 i as closed subgroups of Z×
p (note
2
2 2
that vp (q − 1) = vp ((−q ) − 1)). So by Theorem 1.8(d) and Example 4.4 in [BMO],
G = E6 (q) ∼p 2E6 (q 2 ) ∼p F4 (q 2 ). So we can choose G∗ satisfying (d) as in the last paragraph.
Assume ordp (q) > 2. By [GL, 10-1(3)], for S ∈ Sylp (G) to be nonabelian, there must be
some n ≥ 1 such that p · ordp (q) n, and such that q n − 1 appears as a factor in the formula
for |G(q)| (see, e.g., [GL, Table 4-2] or [Ca, Theorem 9.4.10 & Proposition 10.2.5]). Since
ordp (q)|(p − 1), this shows that the case ordp (q) > 2 appears only for the group E8 (q), and
only when p = 5 and m = 4. In particular, q, q ′ ≡ ±2 (mod 5). Set G∗ = E8 (q ′ ); then
G∗ ∼p G by Theorem 1.8(a), and we are in the situation of (e).
The following lemma was needed in the proof of Proposition 1.10 to reduce still further
the prime powers under consideration.
Lemma 1.11. Fix a prime p, and an integer q > 1 prime to p.
(a) If p is odd, then for any prime r0 whose class generates (Z/p2 )× , there is b ≥ 1 such
that hqi = h(r0 )b i, and b|(p − 1)pℓ for some ℓ.
(b) If p = 2, then either hqi = h3i, or hqi = h5i, or there are ε = ±1 and k ≥ 1 such that
k
ε ≡ q (mod 8) and hqi = hε · 32 i.
Proof. Since q ∈ Z and q > 1, hqi is infinite.
(a) If p is odd, then for each n ≥ 1, (Z/pn )× ∼
= (Z/p)× × (Z/pn−1 ) is cyclic and generated
×
×
ℓ
∼
by the class of r0 . Hence Z×
p = (Z/p) × (Zp , +), and hr0 i = Zp . Also, hqi ≥ 1 + p Zp for
some ℓ ≥ 1, since each infinite, closed subgroup of (Zp , +) contains pk Zp for some k.
ℓ ×
ℓ
ℓ−1
. Then hqi = h(r0 )b i.
Set b = [Z×
p : hqi] = [(Z/p ) : hq+p Zi] (p − 1)p
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
15
∼
(b) If p = 2, then Z×
2 = {±1} × h3i, where h3i = (Z2 , +). Hence the only infinite closed
k
k
subgroups of h3i are those of the form h32 i for some k ≥ 0. So hqi = hε · 32 i for some k ≥ 0
and some ε = ±1, and the result follows since h5i = h − 3i.
We also note, for use in Section 4, the following more technical result.
Lemma 1.12. Let G be a finite group, fix S ∈ Sylp (G), and set F = FS (G). Let P ≤ S be
such that CG (P ) ≤ P and NS (P ) ∈ Sylp (NG (P )). Then for each ϕ ∈ Aut(S, F ) such that
ϕ(P ) = P , ϕ|NS (P ) extends to an automorphism ϕ of NG (P ).
Proof. Since CG (P ) ≤ P and NS (P ) ∈ Sylp (NG (P )), NG (P ) is a model for the fusion system
E = FNS (P ) (NG (P )) in the sense of [AKO, Definition I.4.8]. By the strong uniqueness
property for models [AKO, Theorem I.4.9(b)], and since ϕ|NS (P ) preserves fusion in E, ϕ|NS (P )
extends to an automorphism of the model.
The following elementary lemma will be useful in Sections 5 and 6; for example, when
computing orders of Sylow subgroups of groups of Lie type.
Lemma 1.13. Fix a prime p. Assume q ≡ 1 (mod p), and q ≡ 1 (mod 4) if p = 2. Then
for each n ≥ 1, vp (q n − 1) = vp (q − 1) + vp (n).
Proof. Set r = vp (q − 1), and let k be such that q = 1 + pr k. Then q n = 1 + npr k + ξ, where
vp (npr k) = vp (n) + r, and where each term in ξ has strictly larger valuation.
16
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
2. Background on finite groups of Lie type
In this section and the next, we fix the notation to be used for finite groups of Lie type,
and list some of the (mostly standard) results which will be needed later. We begin by
recalling the following concepts used in [GLS3]. We do not repeat the definitions of maximal
tori and Borel subgroups in algebraic groups, but refer instead to [GLS3, §§ 1.4–1.6].
Definition 2.1 ([GLS3, Definitions 1.7.1, 1.15.1, 2.2.1]). Fix a prime q0 .
(a) A connected algebraic group G over Fq0 is simple if [G, G] 6= 1, and all proper closed
normal subgroups of G are finite and central. If G is simple, then it is of universal type
if it is simply connected, and of adjoint type if Z(G) = 1.
(b) A Steinberg endomorphism of a connected simple algebraic group G is a surjective algebraic endomorphism σ ∈ End(G) whose fixed subgroup is finite.
(c) A σ-setup for a finite group G is a pair (G, σ), where G is a simple algebraic group over
′
Fq0 , and where σ is a Steinberg endomorphism of G such that G = O q0 (CG (σ)).
(d) Let Lie(q0 ) denote the class of finite groups with σ-setup (G, σ) where G is simple and is
defined in characteristic q0 , and let Lie be the union of the classes Lie(q0 ) for all primes
q0 . We say that G is of universal (adjoint) type if G is of universal (adjoint) type.
If G is universal, then CG (σ) is generated by elements of q0 -power order (see [St3, Theorem
12.4]), and hence G = CG (σ) in (c) above. In general, CG (σ) = G · CT (σ) (cf. [GLS3,
Theorem 2.2.6]).
A root group in a connected algebraic group G over Fq0 with a given maximal torus T
is a one-parameter closed subgroup (thus isomorphic to Fq0 ) which is normalized by T .
The roots of G are the characters for the T -actions on the root groups, and lie in the Zlattice X(T ) = Hom(T , F×
q0 ) of characters of T . (Note that this is the group of algebraic
× ∼
homomorphisms, and that Hom(F×
q0 , Fq0 ) = Z.) The roots are regarded as lying in the Rvector space V = R⊗Z T ∗ . We refer to [GLS3, § 1.9] for details about roots and root subgroups
of algebraic groups, and to [Brb, Chapitre VI] for a detailed survey of root systems.
The following notation and hypotheses will be used throughout this paper, when working
with a finite group of Lie type defined via a σ-setup.
Notation 2.2. Let (G, σ) be a σ-setup for the finite group G, where G is a connected, simple
algebraic group over Fq0 for a prime q0 . When convenient, we also write G = G(Fq0 ), where
G is a group scheme over Z.
(A) The maximal torus and Weyl group of G. Fix a maximal torus T in G such
that σ(T ) = T . Let W = NG (T )/T be the Weyl group of G (and of G).
(B) The root system of G. Let Σ be the set of all roots of G with respect to T , and let
X α < G denote the root group for the root α ∈ Σ. Thus X α = {xα (u) | u ∈ Fq0 } with
respect to some fixed Chevalley parametrization of G. Set V = R ⊗Z T ∗ : a real vector
space with inner product (−, −) upon which the Weyl group W acts orthogonally. Let
Π ⊆ Σ be a fundamental system of roots, and let Σ+ ⊆ Σ be the set of positive roots
with respect to Π. For each α ∈ Σ+ , let ht(α) denote the height of α: the number of
summands in the decomposition of α as a sum of fundamental roots.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
17
For each α ∈ Σ, let wα ∈ W be the reflection in the hyperplane α⊥ ⊆ V .
For α ∈ Σ and λ ∈ F×
q0 , let nα (λ) ∈ hX α , X −α i and hα (λ) ∈ T ∩ hX α , X −α i
be as defined in [Ca, § 6.4] or [GLS3, Theorem 1.12.1]: the images of −λ0−1 λ0 and
λ 0
1 u to
under
the
homomorphism
SL
(F
)
−
−
−
→
G
which
sends
2
q
0
0
1
0 λ−1 , respectively,
xα (u) and v1 10 to x−α (v). Equivalently, nα (λ) = xα (λ)x−α (−λ−1 )xα (λ) and hα (λ) =
nα (λ)nα (1)−1 .
(C) The maximal torus, root system and Weyl group of G. Set T = T ∩ G.
Let τ ∈ Aut(V ) and ρ ∈ Aut(Σ) be the orthogonal automorphism and permutation,
respectively, such that for each α ∈ Σ, σ(X α ) = X ρ(α) and ρ(α) is a positive multiple
of τ (α). Set W0 = CW (τ ).
If ρ(Π) = Π, then set V0 = CV (τ ), and let pr⊥
V0 be the orthogonal projection of V onto
b be the set of equivalence classes in Σ determined by τ , where α, β ∈ Σ are
V0 . Let Σ
⊥
equivalent if pr⊥
V0 (α) is a positive scalar multiple of prV0 (β) (see [GLS3, Definition 2.3.1]
b ⊆Σ
b + denote the images in Σ
b of Π ⊆ Σ+ .
or [Ca, § 13.2]). Let Π
b set X αb = hX α | α ∈ α
For each α
b ∈ Σ,
bi and Xαb = C (σ). When α ∈ Σ is of minimal
Xα
b
b and q ′ = |X ab |, then for u ∈ Fq′ , let x
bα (u) ∈ Xαb be an element
height in its class α
b ∈ Σ,
α
b
whose image under projection to Xα is xα (u) (uniquely determined modulo [Xαb , Xαb ]).
hα (λ) ∈ T be an element in G ∩ hhβ (F× ) | β ∈ α
bi whose
For α ∈ Π and λ ∈ F× , let b
q0
component in
hα (F×
q0 )
q0
is hα (λ) (if there is such an element).
To see that τ and ρ exist as defined in point (C), recall that the root groups X α for α ∈ Σ
are the unique closed subgroups of G which are isomorphic to (Fq0 , +) and normalized by T
(see, e.g., [GLS3, Theorem 1.9.5(a,b)]). Since σ is algebraic (hence continuous) and bijective,
σ −1 sends root subgroups to root subgroups, and σ permutes the root subgroups (hence the
roots) since there are only finitely many of them. Using Chevalley’s commutator formula,
one sees that this permutation ρ of Σ preserves angles between roots, and hence (up to
positive scalar multiple) extends to an orthogonal automorphism of V .
These definitions of x
bα (u) ∈ Xαb and b
hα (λ) ∈ T are slightly different from the definitions
in [GLS3, § 2.4] of elements xαb (u) and hαb (λ). We choose this notation to emphasize that
b This will be
these elements depend on the choice of α ∈ Σ, not only on its class α
b ∈ Σ.
important in some of the relations we need to use in Section 4.
Lemma 2.3. Under the assumptions of Notation 2.2, the action of W on T restricts to an
action of W0 on T , and the natural isomorphism NG (T )/T ∼
= W restricts to an isomorphism
NG (T ) ∩ NG (T ) T ∼
= CW (τ ) = W0 .
Proof. For each α ∈ Σ, nα (1) = xα (1)x−α (−1)xα (1) represents the reflection wα ∈ W , and
hence σ(nα ) ∈ hXρ(α) , X−ρ(α) i ∩ NG (T ) represents the reflection wρ(α) = τ(wα ). Since W is
generated by the wα for α ∈ Σ, we conclude that σ and τ have the same action on W .
Thus the identification NG (T )/T ∼
= W restricts to the following inclusions:
NG (T ) ∩ NG (T ) T ≤ CN (T ) (σ)/CT (σ) ≤ CN (T )/T (σ) ∼
= CW (τ ) = W0 .
G
G
If w ∈ W0 represents the coset xT ⊆ NG (T ), then x−1 σ(x) ∈ T . By the Lang-Steinberg
theorem, each element of T has the form t−1 σ(t) for some t ∈ T , and hence we can choose
18
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
′
x such that σ(x) = x. Then x ∈ CG (σ), and hence x normalizes G = O q0 (CG (σ)) and
T = G ∩ T . Since CG (σ) = GCT (σ) (see [GLS3, Theorem 2.2.6(g)] or [St3, Corollary
12.3(a)]), some element of xT lies in NG (T ). So the above inclusions are equalities.
The roots in G are defined formally as characters of its maximal torus T . But it will be
useful to distinguish the (abstract) root α ∈ Σ from the character θα ∈ Hom(T , F×
q0 ) ⊆ V .
For each root α ∈ Σ ⊆ V , let α∨ ∈ V ∗ be the corresponding co-root (dual root): the
unique element such that (α∨ , α) = 2 and wα is reflection in the hyperplane Ker(α∨). Since
we identify V = V ∗ via a W -invariant inner product, α∨ = 2α/(α, α). Point (c) of the next
∗
lemma says that α∨ = hα , when we regard hα ∈ Hom(F×
q0 , T ) as an element in V .
Lemma 2.4. Assume we are in the situation of (A) and (B) in Notation 2.2.
(a) We have CG (T ) = T . In particular, Z(G) ≤ T , and is finite of order prime to the
defining characteristic q0 .
(b) The maximal torus T in G is generated by the elements hα (λ) for α ∈ Π and λ ∈ F×
q0 .
Q
If G is universal, and λα ∈ Fq0 are such that α∈Π hα (λα ) = 1, then λα = 1 for each
α ∈ Π. Thus
Y
T =
hα (F×
q0 ),
α∈Π
and hα is injective for each α.
(c) For each β ∈ Σ, let θβ ∈ X(T ) = Hom(T , F×
q0 ) be the character such that
t
xβ (u) = xβ (θβ (t)·u)
for t ∈ T and u ∈ Fq0 . Then
∨
θβ (hα (λ)) = λ(α ,β)
for β, α ∈ Σ, λ ∈ F×
q0 .
Q
Q
The product homomorphism θΠ = θβ : T −−−−→ β∈Π F×
q0 is surjective, and Ker(θΠ ) =
Z(G).
(d) If α, β1 , . . . , βk ∈ Σ and n1 , . . . , nk ∈ Z are such that α∨ = n1 β1∨ + . . . + nk βk∨, then for
n1
nk
each λ ∈ F×
q0 , hα (λ) = hβ1 (λ ) · · · hβk (λ ).
(e) For each w ∈ W , α ∈ Σ, and λ ∈ F×
q0 , and each n ∈ NG (T ) such that nT = w ∈
NG (T )/T ∼
= W , n(X α ) = X w(α) and n(hα (λ)) = hw(α) (λ). For each α, β ∈ Σ and each
λ ∈ F×
q0 ,
∨
wα (hβ (λ)) = hwα (β) (λ) = hβ (λ)hα (λ−(β ,α) ) .
Hence wα (t) = t · hα (θα (t))−1 for each t ∈ T .
Proof. (a) By [Hu, Proposition 24.1.A], the maximal torus T is regular (i.e., contained in only
finitely many Borel subgroups). So CG (T ) = T by [Hu, Corollary 26.2.A]. Hence Z(G) ≤ T ,
it is finite since G is assumed simple, and so it has order prime to the defining characteristic
q0 .
We claim that it suffices to prove the relations in (c)–(e) in the adjoint group G/Z(G),
and hence that we can use the results in [Ca, §§ 7.1–2]. For relations in T , this holds since
T is infinitely divisible and Z(G) is finite (thus each homomorphism to T /Z(G) has at most
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
19
one lifting to T ). For relations in a root group X α , this holds since each element of X α Z(G)
of order q0 lies in X α , since |Z(G)| is prime to q0 by (a).
(b) This is stated without proof in [GLS3, Theorem 1.12.5(b)], and with a brief sketch of
a proof in [St4, p. 122]. We show here how it follows from the classification of reductive
algebraic groups in terms of root data (see, e.g., [Sp, § 10]).
Consider the homomorphism
def
hΠ : Te =
Q
Y
α∈Π
F×
−−−−−−→ T
q0 −
which sends (λα )α∈Π to α hα (λα ). Then hΠ is surjective with finite kernel (see [Ca, § 7.1]).
It remains to show that it is an isomorphism when G is of universal type.
We recall some of the notation used in [Sp, § 7]. To G is associated the root datum
X(T ), Σ, X ∨ (T ), Σ∨ , where
X(T ) = Hom(T , F×
q0 ),
X ∨ (T ) = Hom(F×
q0 , T ),
Σ∨ = {α∨ = hα | α ∈ Σ} ⊆ X ∨ (T ) .
As noted before, X(T ) and X ∨ (T ) are groups of algebraic homomorphisms, and are free
abelian groups of finite rank dual to each other. Recall that Σ ⊆ X(T ), since we identify a
root α with the character θα .
Set Y ∨ = ZΣ∨ ⊆ X ∨ (T ), and let Y ⊇ X(T ) be its dual. Then (Y, Σ, Y ∨ , Σ∨ ) is still
a root datum as defined in [Sp, § 7.4]. By [Sp, Proposition 10.1.3] and its proof, it is
e with maximal torus Te, which lies in a central
realized by a connected algebraic group G
e −−−→ G which extends hΠ . Since G is of universal type, f and hence hΠ are
extension f : G
isomorphisms.
(c) Let ZΣ ≤ V be the additive subgroup generated by Σ. In the notation of [Ca, pp.
97–98], for each α ∈ Σ and λ ∈ F×
q0 , hα (λ) = h(χα,λ ) where
χα,λ ∈ Hom(ZΣ, F×
q0 )
is defined by
∨ ,v)
χα,λ (v) = λ2(α,v)/(α,α) = λ(α
.
h(χ)
xβ (u) =
Also, by [Ca, p. 100], for each χ ∈ Hom(ZΣ, F×
q0 ), β ∈ Σ, and u ∈ Fq0 ,
×
xβ (χ(β)·u). Thus there are homomorphisms θβ ∈ Hom(T , Fq0 ), for each β ∈ Σ, such that
t
xβ (u) = xβ (θβ (t)·u), and θβ (h(χ)) = χ(β) for each χ. For each α ∈ Σ and λ ∈ F×
q0 ,
∨ ,β)
θβ (hα (λ)) = θβ (h(χα,λ )) = χα,λ (β) = λ(α
.
(1)
Assume t ∈ Ker(θΠ ). Thus t ∈ Ker(θα ) for all α ∈ Π, and hence for all α ∈ Σ ⊆ ZΠ. So
[t, Xα ] = 1 for all α ∈ Σ, these root subgroups generate G (see [Sp, Corollary 8.2.10]), and
this proves that t ∈ Z(G). The converse is clear: t ∈ Z(G) implies t ∈ T by (a), and hence
θβ (t) = 1 for all β ∈ Π by definition of θβ .
Q
It remains to show that θΠ sends T onto β∈Π Fq0 . Consider the homomorphisms
Y
Y
hΠ
θΠ
def
−−−
−−→ T −−−−−
F×
−−→
F×
(2)
Te =
q0 −−
q0 ,
α∈Π
β∈Π
where hΠ was defined in the proof of (b). We just saw that θΠ ◦ hΠ has matrix (α∨ , β)
∨
∗
F×
q0
α,β∈Π
,
which has nonzero determinant since Π ⊆ V and Π ⊆ V are bases. Since
is divisible
and its finite subgroups are cyclic, this implies that θΠ ◦ hΠ is onto, and hence θΠ is onto.
20
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
(d) This follows immediately from (c), where we showed, for α ∈ Σ, that α∨ can be identified
∗
with hα in Hom(F×
q0 , T ) ⊆ V .
(e) The first statement (n(X α ) = X w(α) and n(hα (λ)) = hw(α) (λ)) is shown in [Ca, Lemma
7.2.1(ii) & Theorem 7.2.2]. By the usual formula for an orthogonal reflection, wα (β) =
β − 2(α,β)
α = β − (α∨ , β)α. Here, we regard wα as an automorphism of V (not of T ). Since
(α,α)
wα (β) and β have the same norm,
wα (β)∨ =
2β
2(α, β)
2α
2wα (β)
=
−
·
= β ∨ − (β ∨ , α) · α∨ ,
(β, β)
(β, β)
(β, β) (α, α)
and by (d),
wα (hβ (λ)) = hwα (β) (λ) = hβ (λ)hα (λ−(β
∨ ,α)
) = hβ (λ)hα (θα (hβ (λ))−1 )
where the last equality follows from (c). Since T is generated by the hβ (λ) by (b), this
implies that wα (t) = t · hα (θα (t))−1 for all t ∈ T .
For any algebraic group H, H 0 denotes its identity connected component. The following
proposition holds for any connected, reductive group, but we state it only in the context of
Notation 2.2. Recall the homomorphisms θβ ∈ Hom(T , F×
q0 ), defined for β ∈ Σ in Lemma
2.4(c).
Proposition 2.5. Assume Notation 2.2. For any subgroup H ≤ T , CG (H) is an algebraic
group, CG (H)0 is reductive, and
CG (H)0 = hT , Xα | α ∈ Σ, H ≤ Ker(θα )i
CG (H) = CG (H)0 · {g ∈ NG (T ) | [g, H] = 1} .
(3)
If, furthermore, G is of universal type, then Z(G) = CT (W ).
Proof. The description of CG (H)0 is shown in [Ca2, Theorem 3.5.3] when H is finite and
cyclic, and the proof given there also applies in the more general case. For each g ∈ CG (H),
cg (T ) is another maximal torus in CG (H)0 , so gh ∈ CN (T ) (H) for some h ∈ CG (T )0 , and
thus CG (H) = CG (H)0 · CN
G
(T )
G
(H).
Assume G is of universal type. Since Z(G) ≤ T by Lemma 2.4(a), we have Z(G) ≤ CT (W ).
Conversely, by Lemma 2.4(b), for each t ∈ T and each α ∈ Σ, t(xα (u)) = xα (θα (t)u), and
θ−α (t) = θα (t)−1 . Hence also t(nα (1)) = nα (θα (t)) (see the formula for nα (λ) in Notation
2.2(B)). If t ∈ CT (W ), then [t, nα (1)] = 1 for each α, and since G is of universal type,
hX α , X −α i ∼
= SL2 (Fq0 ). Thus θα (t) = 1 for all α ∈ Σ, t acts trivially on all root subgroups,
and so t ∈ Z(G).
We now look more closely at the lattice ZΣ∨ generated by the dual roots.
Lemma 2.6. Assume Notation 2.2(A,B), and also that G (and hence G) is of universal
type.
(a) There is an isomorphism
−−−−−→ T
Φ : ZΣ∨ ⊗Z F×
q0 −
with the property that Φ(α∨ ⊗ λ) = hα (λ) for each α ∈ Σ and each λ ∈ F×
q0 .
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
21
∨
Fix some λ ∈ F×
−−→ T .
q0 , and set m = |λ|. Set Φλ = Φ(−, λ) : ZΣ −
(b) The map Φλ is Z[W ]-linear, Ker(Φλ ) = mZΣ∨ , and Im(Φλ ) = {t ∈ T | tm = 1}.
(c) Fix t ∈ T and x ∈ ZΣ∨ such that Φλ (x) = t, and also such that
kxk < 12 m · min kα∨ k α ∈ Π .
Then CW (t) = CW (x).
(d) If m = |λ| ≥ 4, then for each α ∈ Σ, CW (hα (λ)) = CW (α).
Proof. (a,b) Identify ZΣ∨ as a subgroup of Hom(F×
q0 , T ), and let
Φ : ZΣ∨ × F×
−−−−→ T
q0 −−
be the evaluation pairing. This is bilinear, hence induces a homomorphism on the tensor
product, and Φ(α∨ , λ) = hα (λ) by Lemma 2.4(c). Since {α∨ | α ∈ Π} is a Z-basis for ZΣ∨
(since Σ∨ is a root system by [Brb, § VI.1, Proposition 2]), and since G is of universal type,
Φ is an isomorphism by Lemma 2.4(b).
In particular, for fixed λ ∈ F×
q0 of order m, Φ(−, λ) induces an isomorphism from the
∨
∨
quotient group ZΣ /mZΣ onto the m-torsion subgroup of T .
(c) Clearly, CW (x) ≤ CW (t); it remains to prove the opposite inclusion. Fix w ∈ CW (t).
By (a), w(x) ≡ x (mod mZΣ∨ ).
√
Set r = min kα∨ k α ∈ Π . For each α ∈ Σ, kα∨k= k · r for some k = 1, 2, 3, and
hence (α∨ , α∨) ∈ r 2 Z. For each α, β ∈ Σ, 2(α∨, β ∨ ) (α∨ , α∨ ) ∈ Z (cf. [Ca, Definition
1 2
∨
∨
2
∨
2.1.1]),
and hence (α ∨, β ) ∈ 2 r Z. Thus (x, x) ∈ r Z for each x ∈ ZΣ , and in particular,
min kxk 0 6= x ∈ ZΣ = r.
By assumption, kw(x)k = kxk < mr/2, so kw(x) − xk < mr. Since each nonzero element
in mZΣ∨ has norm at least mr, this proves that w(x) − x = 0, and hence that w ∈ CW (x).
(d) This is the special case of (b), where x = α∨ and t = hα (λ).
Lemma 2.7. Assume Notation 2.2, and assume also that G is of universal type. Let Γ <
Aut(V ) be any finite group of isometries of (V, Σ). Then there is an action of Γ on T , where
g(hα (u)) = hg(α) (u) for each g ∈ Γ, α ∈ Σ, and u ∈ F×
q0 . Fix m ≥ 3 such that q0 ∤ m, and
m
set Tm = {t ∈ T | t = 1}. Then Γ acts faithfully on Tm . If 1 6= g ∈ Γ and ℓ ∈ Z are such
that g(t) = tℓ for each t ∈ Tm , then ℓ ≡ −1 (mod m).
Proof. The action of Γ on T is well defined by the relations in Lemma 2.4(d,b).
Now fix m ≥ 3 prime to q0 , and let Tm < T be the m-torsion subgroup. It suffices to
prove the rest of the lemma when m = p is an odd prime, or when m = 4 and p = 2. Fix
∨
λ ∈ F×
−−−→ T be the homomorphism of Lemma 2.6(a). By
q0 of order m, and let Φλ : ZΣ −
definition of Φλ , it commutes with the actions of Γ on ZΣ∨ < V and on Tm .
Assume 1 6= g ∈ Γ and ℓ ∈ Z are such that g(t) = tℓ for each t ∈ Tm . Set r = dim(V ), and
let B ∈ GLr (Z) be the matrix for the action of g on ZΣ∨ , with respect to some Z-basis of
ZΣ∨ . Then |g| = |B|, and B ≡ ℓI (mod mMr (Z)). If p = 2 (m = 4), let µ ∈ {±1} be such
that ℓ ≡ µ (mod 4). If p is odd (so m = p), then let µ ∈ (Zp )× be such that µ ≡ ℓ (mod p)
and µp−1 = 1. Set B ′ = µ−1 B ∈ GLr (Zp ). Thus B ′ also has finite order, and B ′ ≡ I (mod
mMr (Zp )).
22
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
The logarithm and exponential maps define inverse bijections
ln
−
−
−
−
−
→
I + mMr (Zp ) −
←
−
−
−
−
−
− mMr (Zp ) .
exp
They are not homomorphisms, but they do have the property that ln(M k ) = k ln(M) for
each M ∈ I + mMr (Zp ) and each k ≥ 1. In particular, the only element of finite order in
I + mMr (Zp ) is the identity. Thus B ′ = I, so B = µI. Since µ ∈ Z and B 6= I, we have
µ = −1 and B = −I.
The following lemma about the lattice ZΣ∨ will also be useful when working with the
Weyl group action on certain subgroups of T .
Lemma 2.8. Assume Notation 2.2(A,B). Set Λ = ZΣ∨ : the lattice in V generated by the
dual roots. Assume that there are b ∈ W of order 2, and a splitting Λ = Λ+ × Λ− , such that
Λ+ , Λ− 6= 0 and b acts on Λ± via ±Id. Then G ∼
= Cn (= Sp2n ) for some n ≥ 2.
Proof. Fix b ∈ w and a splitting Λ = Λ+ × Λ− as above. When considering individual cases,
we use the notation of Bourbaki [Brb, Planches I–IX] to describe the (dual) roots, lattice,
and Weyl group.
• If G = An (n ≥ 2), then Λ = (a0 , . . . , an ) ∈ Zn+1 a0 + . . . + an = 0 , and b exchanges
certain coordinates pairwise. Choose v ∈ Λ with coordinates 1, −1, and otherwise 0;
where the two nonzero entries are in separate orbits of b of which at least one is nonfixed.
Then v ∈
/ Λ+ × Λ− , a contradiction.
• If G = G2 , then as described in [Brb, Planche IX], Λ is generated by the dual fundamental
roots (1, −1, 0) and ( 23 , − 13 , − 31 ), and does not have an orthogonal basis.
• If G = Bn (n ≥ 3), Dn (n ≥ 4), or F4 , then Λ < Zn is the sublattice of n-tuples the sum
of whose coordinates is even. Also, b acts by permuting the coordinates and changing
sign (or we can assume it acts this way in the F4 case). Choose v with two 1’s and the
rest 0, where the 1’s are in separate b-orbits, of which either at least one is nonfixed, or
both are fixed and exactly one is negated. Then v ∈
/ Λ+ × Λ− , a contradiction.
• If G = E8 , then Λ = Λ(E8 ) < R8 is generated by 12 (1, 1, . . . , 1) and the n-tuples of
integers whose sum is even. We can assume (up to conjugation) that b acts as a signed
permutation. Choose v as in the last case.
• If G = E7 , then Λ < R8 is the lattice of all x = (x1 , . . . , x8 ) ∈ Λ(E8 ) such that x7 = −x8 .
Up to conjugation, b can be again be assumed to act on A via a signed permutation
(permuting only the first six coordinates), and v can be chosen as in the last case.
• If G = E6 , then Λ < R8 is the lattice of all x = (x1 , . . . , x8 ) ∈ Λ(E8 ) such that x6 = x7 =
−x8 . Also, W contains a subgroup isomorphic to 24 : S5 with odd index which acts on
the remaining five coordinates via signed permutations. So b and v can be taken as in the
last three cases.
We finish the section with a very elementary lemma.
It will be useful to know, in certain situations, that each coset of T in NG (T ) contains
elements of G.
Lemma 2.9. Assume that we are in the situation of Notation 2.2(A,B). Assume also that
σ acts on T via (t 7→ tm ) for some 1 6= m ∈ Z. Then for each g ∈ NG (T ), gT ∩ CG (σ) 6= ∅.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
23
Proof. Since σ|T ∈ Z(Aut(T )), we have g −1 σ(g) ∈ CG (T ) = T , the last equality by Lemma
r
2.4(a). So for each t ∈ T , σ(gt) = gt if and only if g −1 σ(g) = t1−m . Since T ∼
= (F×
q0 ) for
some r, and Fq0 is algebraically complete (and 1 − m 6= 0), this always has solutions.
24
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
3. Automorphisms of groups of Lie type
Since automorphisms of G play a central role in this paper, we need to fix our notation
(mostly taken from [GLS3]) for certain subgroups and elements of Aut(G). We begin with
automorphisms of the algebraic group G.
Definition 3.1. Let G and its root system Σ be as in Notation 2.2(A,B).
(a) When q is any power of q0 (the defining characteristic of G), let ψq ∈ End(G) be the
field endomorphism defined by ψq (xα (u)) = xα (uq ) for each α ∈ Σ and each u ∈ Fq0 .
Set ΦG = {ψq0b | b ≥ 1}: the monoid of all field endomorphisms of G.
(b) Let ΓG be the group or set of graph automorphisms of G as defined in [GLS3, Definition
1.15.5(e)]. Thus when (G, q0 ) 6= (B2 , 2), (G2 , 3), nor (F4 , 2), ΓG is the group of all
γ ∈ Aut(G) of the form γ(xα (u)) = xρ(α) (u) (all α ∈ ±Π and u ∈ Fq0 ) for some
isometry ρ of Σ such that ρ(Π) = Π. If (G, q0 ) = (B2 , 2), (G2 , 3), or (F4 , 2), then
ΓG = {1, ψ}, where for the angle-preserving permutation ρ of Σ which exchanges long
and short roots and sends Π to itself, ψ(xα (u)) = xρ(α) (u) when α is a long root and
ψ(xα (u)) = xρ(α) (uq0 ) when α is short.
(c) A Steinberg endomorphism σ of G is “standard” if σ = ψq ◦ γ = γ ◦ ψq , where q is a
power of q0 and γ ∈ ΓG . A σ-setup (G, σ) for a finite subgroup G < G is standard if σ
is standard.
By [GLS3, Theorem 2.2.3], for any G with σ-setup (G, σ) as in Notation 2.2, G is Gconjugate to a subgroup G∗ which has a standard σ-setup. This will be made more precise
in Proposition 3.6(a).
Most of the time in this paper, we will be working with standard σ-setups. But there are
a few cases where we will need to work with setups which are not standard, which is why
this condition is not included in Notation 2.2.
Following the usual terminology, we call G a “Chevalley group” if it has a standard σsetup where γ = Id in the notation of Definition 3.1; i.e., if G ∼
= G(q) where q is some power
of q0 . In this case, the root groups Xαb are all abelian and isomorphic to Fq . When G has a
standard σ-setup with γ 6= Id, we refer to G as a “twisted group”, and the different possible
structures of its root groups are described in [GLS3, Table 2.4]. We also refer to G as a
“Steinberg group” if γ 6= Id and is an algebraic automorphism of G; i.e., if G is a twisted
group and not a Suzuki or Ree group.
The following lemma will be useful in Sections 5 and 6.
Lemma 3.2. Assume G is as in Notation 2.2(A,B). Then for each algebraic automorphism
γ of G which normalizes T , there is an orthogonal automorphism τ of V such that τ (Σ) = Σ,
and
and
γ(hα (λ)) = hτ (α) (λ)
γ(X α ) = X τ (α)
for each α ∈ Σ and each λ ∈ F×
q0 . In particular, γ|T = |τ | < ∞. If, in addition, γ
normalizes each of the root groups X α (i.e., τ = Id), then γ ∈ AutT (G).
Proof. By [GLS3, Theorem 1.15.2(b)], and since γ is an algebraic automorphism of G, γ =
cg ◦ γ0 for some g ∈ G and some γ0 ∈ ΓG . Furthermore, γ0 has the form: γ0 (xα (u)) = xχ(α) (u)
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
25
for all α ∈ Σ and u ∈ Fq0 , and some isometry χ ∈ Aut(V ) such that χ(Π) = Π. Since γ and
γ0 both normalize T , we have g ∈ NG (T ).
Thus by Lemma 2.4(e), there is τ ∈ Aut(V ) such that τ (Σ) = Σ, and γ(X α ) = X τ (α) and
γ(hα (λ)) = hτ (α) (λ) for each α ∈ Σ and λ ∈ F×
q0 . In particular, γ|T = |τ |.
If τ = Id, then γ0 = Id and g ∈ T . Thus γ ∈ AutT (G).
We next fix notation for automorphisms of G.
Definition 3.3. Let G and G be as in Notation 2.2(A,B,C), where in addition, we assume
the σ-setup is standard.
(a) Set
Inndiag(G) = AutT (G)Inn(G)
and
Outdiag(G) = Inndiag(G)/Inn(G) .
(b) Set ΦG = ψq |G q = q0b , b ≥ 1 , the group of field automorphisms of G.
(c) If G is a Chevalley group, set ΓG = γ|G γ ∈ ΓG , the group of graph automorphisms
of G. Set ΓG = 1 if G is a twisted group (a Steinberg, Suzuki, or Ree group).
Note that in [GLS3, Definition 2.5.13], when G has a standard σ-setup (G, σ), Inndiag(G)
is defined to be the group of automorphisms induced by conjugation by elements of CG/Z(G) (σ)
(lifted to G). By [GLS3, Lemma 2.5.8], this is equal to Inndiag(G) as defined above when
G is of adjoint form, and hence also in the general case (since Z(G) ≤ T ).
Steinberg’s theorem on automorphisms of groups of Lie type can now be stated.
Theorem 3.4 ([St1, § 3]). Let G be a finite group of Lie type. Assume that (G, σ) is
a standard σ-setup for G, where G is in adjoint or universal form. Then Aut(G) =
Inndiag(G)ΦG ΓG , where Inndiag(G) E Aut(G) and Inndiag(G) ∩ (ΦG ΓG ) = 1.
Proof. See, e.g., [GLS3, Theorem 2.5.12(a)] (together with [GLS3, Theorem 2.5.14(d)]).
Most of this follows from the main result in [St1], and from [St2, Theorems 30 & 36].
We also need the following characterizations of Inndiag(G) which are independent of the
choice of σ-setup.
Proposition 3.5. Assume the hypotheses and notation in 2.2. Then
(a) CG (G) = Z(G);
(b) NG (G) = GNT (G);
(c) Inndiag(G) = AutT (G)Inn(G) = AutG (G), and hence Outdiag(G) = OutT (G).
In fact, (b) and (c) hold if we replace T by any σ-invariant maximal torus in G.
Proof. (a) Since the statement is independent of the choice of σ-setup, we can assume that
Q
Q
σ is standard. Set U = α∈Σ+ X α and U ∗ = α∈Σ+ X −α .
Fix g ∈ CG (G). Since G has a BN-pair (see [Ca, Proposition 8.2.1]), it has a Bruhat
decomposition G = BNB = U NU [Ca, Proposition 8.2.2(i)], where B = T U and N =
NG (T ). Write g = unv, where u, v ∈ U and n ∈ N . For each x ∈ U ∩ G, gx = u(nvx) ∈ U
implies that nvx = n(vx) ∈ U .
26
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Since n ∈ NG (T ), conjugation by n permutes the root groups of G, in a way determined by
the class w = nT ∈ W = NG (T )/T . Thus w sends each (positive) root in the decomposition
of vx to a positive root. For each α ∈ Σ+ , x
bα (1) ∈ G, v(b
xα (1)) has α in its decomposition,
and hence w(α) ∈ Σ+ .
Thus w sends all positive roots to positive roots, so w(Π) = Π, and w = 1 by [Ca, Corollary
2.2.3]. So n ∈ T , and g = unv ∈ T U .
By the same argument applied to the negative root groups, g ∈ T U ∗ . Hence g ∈ T .
For each α ∈ Σ, g ∈ T commutes with x
bα (1) ∈ G, and hence g centralizes X β for each
β∈α
b (Lemma 2.4(c)). Thus g centralizes all root groups in G, so g ∈ Z(G).
(b) Let T ∗ be any σ-invariant maximal torus in G. Fix g ∈ NG (G). Then g −1 · σ(g) ∈
CG (G) = Z(G) ≤ T ∗ by (a). By Lang’s theorem [GLS3, Theorem 2.1.1], there is t ∈ T ∗ such
that g −1 · σ(g) = t−1 · σ(t). Hence gt−1 ∈ CG (σ) = G · CT ∗ (σ), where the last equality holds
by [GLS3, Theorem 2.2.6(g)]. So g ∈ GT ∗ , and g ∈ GNT ∗ (G) since g normalizes G.
(c) By (b), AutG (G) = AutT ∗ (G)Inn(G) for each σ-invariant maximal torus T ∗ . By definition, Inndiag(G) = AutT ∗ (G)Inn(G) when T ∗ is the maximal torus in a standard σ-setup
for G. Hence Inndiag(G) = AutG (G) = AutT ∗ (G)Inn(G) for all such T ∗ .
We refer to [GLS3, Definitions 1.15.5(a,e) & 2.5.10] for more details about the definitions
of ΦG and ΓG . The next proposition describes how to identify these subgroups when working
in a nonstandard setup.
Proposition 3.6. Assume G, T , and the root system of G, are as in Notation 2.2(A,B).
′
Let σ be any Steinberg endomorphism of G, and set G = O q0 (CG (σ)).
′
(a) There is a standard Steinberg endomorphism σ ∗ of G such that if we set G∗ = O q0 (CG (σ ∗ )),
then there is x ∈ G such that G = x(G∗ ).
Fix G∗ , σ ∗ , and x as in (a). Let Inndiag(G∗ ), ΦG∗ , and ΓG∗ be as in Definition 3.3 (with
−1
respect to the σ-setup (G, σ ∗ )). Set Inndiag(G) = cx Inndiag(G∗ )c−1
x , ΦG = cx ΦG∗ cx , and
−1
ΓG = cx ΓG∗ cx , all as subgroups of Aut(G). Then the following hold.
(b) Inndiag(G) = AutG (G).
(c) For each α ∈ ΦG ΓG such that α|G∗ ∈ ΦG∗ ΓG∗ , and each β ∈ α · Inn(G) such that
β(G) = G, β|G ≡ cx (α)c−1
x (mod Inndiag(G)).
(d) If ψq0 normalizes G, then Inndiag(G)ΦG = Inndiag(G)hψq0 |G i.
Thus the subgroups ΦG and ΓG are well defined modulo Inndiag(G), independently of the
choice of standard σ-setup for G.
Proof. (a) See, e.g., [GLS3, Theorem 2.2.3]: for any given choice of maximal torus, positive roots, and parametrizations of the root groups, each Steinberg automorphism of G is
conjugate, by an element of Inn(G), to a Steinberg automorphism of standard type.
(b) This follows immediately from Proposition 3.5(c).
−1
(c) By assumption, β ≡ α ≡ cx αc−1
x (mod Inn(G)). Since β and cx αcx both normalize G,
β|G ≡ cx α∗ c−1
x modulo AutG (G) = Inndiag(G).
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
27
(d) If ψq0 normalizes G, then (c), applied with α = β = ψq0 , implies that as elements of
Aut(G)/Inndiag(G), [ψq0 |G ] = [cx (ψq0 |G∗ )c−1
x ] generates the image of ΦG .
′
Lemma 3.7. Assume G, T , σ, G = O q0 (CG (σ)), and the root system of G, are as in
Notation 2.2(A,B). Assume that ϕ ∈ Aut(T ) is the restriction of an algebraic automorphism
of G such that [ϕ, σ|T ] = 1. Then there is an algebraic automorphism ϕ ∈ Aut(G) such that
ϕ|T = ϕ, [ϕ, σ] = 1, and ϕ(G) = G.
Proof. By assumption, there is ϕ ∈ Aut(G) such that ϕ|T = ϕ. Also, [ϕ, σ] is an algebraic
automorphism of G by [GLS3, Theorem 1.15.7(a)], it is the identity on T , and hence [ϕ, σ] =
ct for some t ∈ T by Lemma 3.2. Using the Lang-Steinberg theorem, upon replacing ϕ by
cu ϕ for appropriate u ∈ T , we can arrange that [ϕ, σ] = 1. In particular, ϕ(G) = G.
The following proposition is well known, but it seems to be difficult to find references
where it is proven.
Proposition 3.8. Fix a prime q0 , and a group G ∈ Lie(q0 ) of universal type. Then Z(G)
has order prime to q0 , G/Z(G) ∈ Lie(q0 ) and is of adjoint type, and Z(G/Z(G)) = 1. If
G/Z(G) is simple, then each central extension of G by a group of order prime to q0 splits
(equivalently, H 2 (G; Z/p) = 0 for all primes p 6= q0 ).
Proof. Let (G, σ) be a σ-setup for G, and choose a maximal torus and positive roots in G.
We can thus assume Notation 2.2. By Lemma 2.4(a), Z(G) is finite of order prime to q0 .
Since Z(G) ≤ CG (G) = Z(G) by Proposition 3.5(a), Z(G) also has order prime to q0 .
Set Ga = G/Z(G) and let Ga < Ga be the image of G under projection. Thus Ga is an
′
algebraic group of adjoint type, and Ga = O q0 (CGa (σa )) ∈ Lie(q0 ) where σa ∈ End(Ga ) is
induced by σ. Also, Z(Ga ) ≤ Z(Ga ) = 1 by Proposition 3.5(a) again.
It remains to prove the statement about central extensions. When G is a Chevalley group,
this was shown in [St4, Théorème 4.5]. It was shown in [St6, Corollary 6.2] when G ∼
= 2An (q)
for n even, and in [AG] when G ∼
= 2 G2 (q) or Sz(q). The remaining cases follow by similar
arguments (see [St5, 9.4 & 12.4]). (See also [Cu, § 1].)
The next proposition shows that in most cases, CG (T ) = T . In the next section, we will
see some conditions which imply that CG (Op (T )) = T when p is a prime different from the
defining characteristic.
Proposition 3.9. Let (G, σ) be a σ-setup for G, where G and G are of universal type.
Assume Notation 2.2, and in particular, that we have fixed a maximal torus T and a root
system Σ in G.
(a) Assume that CG (T )0 T , where (−)0 denotes the connected component of the identity.
Then there is α ∈ Σ+ such that θα (T ) = 1. Also, there is β ∈ Hom(T , F×
q0 ) such that
−1 ∗
−1
θα = β σ (β); i.e., θα (t) = β(t σ(t)) for each t ∈ T .
(b) If the σ-setup is standard, then CG (T )0 = T except possibly when G ∼
= r G(2) for some
2
G and some r ≤ 3, or when G ∼
= A1 (3), Cn (3) for n ≥ 2, or G2 (3).
(c) If CG (T )0 = T , then NG (T )/T ∼
= W0 .
28
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Proof. (a) Since G is of universal type, G = CG (σ) and T = CT (σ). Hence there is a short
exact sequence
t7→t−1 σ(t)
1 −−−→ T −−−−−→ T −−−−−−−−−→ T ,
where the last map is onto by the Lang-Steinberg theorem. Upon dualizing, and regarding
Hom(T , F×
q0 ) additively, we get an exact sequence
σ∗ −Id
restr
0 −−−→ Hom(T , F×
−−−−−→ Hom(T , F×
−−−−→ Hom(T, F×
q0 ) −
q0 ) −
q0 )
(see also [Ca2, Proposition 3.2.3]), where Hom(T , F×
q0 ) is the group of algebraic homomorphisms. Since θα is in the kernel of the restriction map, by assumption, it has the form
β −1 σ ∗ (β) for some β ∈ Hom(T , F×
q0 ).
(b) Let P (Σ) and Q(Σ) be as in [Brb, § VI.1.9] (but with Σ in place of R to denote the
root system). Thus Q(Σ) = ZΣ, the integral lattice generated by Σ, and
P (Σ) = {v ∈ V | (v, α∨) ∈ Z for all α ∈ Σ} ≥ Q(Σ) .
∨
(v,α )
for
For each v ∈ P (Σ), define θv ∈ X(T ) = Hom(T , F×
q0 ) by setting θv (hα (λ)) = λ
α ∈ Π and λ ∈ Fq0 . Since G is of universal type, this is a well defined homomorphism by
Lemma 2.4(b), and the same formula holds for all α ∈ Σ by Lemma 2.4(d). By Lemma
2.4(c), this extends our definition of θβ for β ∈ Σ ⊆ P (Σ).
Recall that Hom(F× , F× ) ∼
= Z. For each θ ∈ X(T ) and each α ∈ Σ, let nθ,α ∈ Z be such
q0
q0
∨
that θ(hα (λ)) = λ
for all λ ∈ F×
q0 . For given θ, there is v ∈ P (Σ) such that (v, α ) = nθ,α
for all α ∈ Π, and hence (by Lemma 2.4(d)) for all α ∈ Σ. Then θ = θv as defined above. In
this way, we identify P (Σ) with the lattice X(T ) of characters for T , while identifying Q(Σ)
with ZΣ.
nθ,α
From the appendix to Chapter VI in [Brb] (Planches I–IX), we obtain the following table:
root system Σ
min{kvk | v ∈ P (Σ)}
max{kαk | α ∈ Σ}
p
An
n/(n+1)
√
2
Cn
1
2
Bn , Dn
G2 F4
E6
E7 E8
p
p
√
√ √
min{ n/4, 1}
2 1
4/3
2
2
√
√ √
√
√ √
2
6
2
2
2
2
Here, the norms are given with respect to the descriptions of these lattices in [Brb] as
subgroups of Euclidean spaces.
T . By (a), there are α ∈ Σ+ and β ∈ Hom(T , F×
Assume CG (T )0
q0 ) such that α =
β σ (β). If we regard α and β as elements in the normed vector space V , then kαk =
kσ ∗ (β)−βk ≥ kσ ∗ (β)k−kβk. If G = rG(q) (and σ is a standard setup), then kσ ∗ (β)k = qkβk,
√
except when G is a Suzuki or Ree group in which case kσ ∗ (β)k = qkβk. Thus
(
kαk
q
if G is a Chevalley or Steinberg group
+1≥ √
kβk
q if G is a Suzuki or Ree group.
−1 ∗
By the above table, this is possible only if q = 2, or if G is isomorphic to one of the groups
A1 (3), B2 (3), Cn (3) (n ≥ 3), 2 G2 (3), or 2B2 (8).
Assume G ∼
= Sz(8). It is most convenient to use the root system for C2 con= 2B2 (8) ∼
structed in [Brb]: P (Σ) = Z2 , and Σ = {(±2, 0), (0, ±2), (±1, ±1)}.
Then α and β satisfy
√
the above inequality only if kαk = 2, kβk = 1, and kα + βk = 8. So (α, β) = 32 , which is
impossible for α, β ∈ Z2 . Hence CG (T )0 = T in this case.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
(c) If CG (T )0 = T , then NG (T ) ≤ NG (T ), and so NG (T )/T ∼
= W0 by Lemma 2.3.
29
The following, more technical lemma will be needed in Section 6.
Lemma 3.10. Assume the hypotheses and notation in 2.2, and also that the σ-setup (G, σ)
b each orbit contains elements of Π.
b
is standard. Then under the action of W0 on Σ,
Proof. When ρ = Id, this is [Ca, Proposition 2.1.8]. When ρ 6= Id, it follows from the
b in [Ca, §§ 13.2–13.3].
descriptions of W0 and Σ
30
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
4. The equicharacteristic case
The following notation will be used in this section.
Notation 4.1. Assume the notation in 2.2, and also that ρ(Π) = Π, q0 = p, and Z(G) = 1.
Thus G = G(Fp ) is a connected, simple group over Fp in adjoint form, σ is a Steinberg
′
endomorphism of G of standard form, and G = O p (CG (σ)).
def
(D) Set U = X α α ∈ Σ+ and B = NG (U ) = U T (the Borel subgroup of G). Set
b +i ,
U = CU (σ) = hXαb | α
b∈Σ
B = NG (U) ,
and
T = T ∩ G.
Q
Thus U = αb∈Σb + Xαb ∈ Sylp (G), and B = UT . (See, e.g., [GLS3, Theorems 2.3.4(d)
& 2.3.7], or [Ca, Theorems 5.3.3(ii) & 9.4.10] in the case of Chevalley groups.) When
b is the image in Σ
b + of a τ -invariant subset J $ Π, let U b ≤ U be the subgroup
Jb $ Π
J
generated by root groups for positive roots in Σ+ rhJi (the unipotent radical subgroup
b and set P b = NG (U b) = B X−bα α ∈ hJi (the parabolic subgroup
associated to J),
J
J
b Thus U = U∅ and B = P∅ . We also write Uαb = U{bα} and Pαb = P{bα}
associated to J).
b
for each α
b ∈ Π.
P
(E) The height of a positive root α =
γ∈Π nγ γ ∈ Σ+ (nγ ≥ 0) is defined by ht(α) =
P
b + is the minimum of the heights of
α) of a class of roots α
b∈Σ
γ∈Π nγ . The height ht(b
roots in the class α
b.
(F) Set F = FU (G) and L = LcU (G).
b +, α
(G) Set U0 = hXαb | α
b∈Σ
b ∩ Π = ∅i = hXαb | ht(b
α) ≥ 2i.
b equivalently, to the number of maximal parabolic
(H) The Lie rank of G is equal to |Π|;
subgroups containing B.
For example, assume σ = ψq ◦ γ, where γ ∈ Aut(G) is a graph automorphism which
b
induces ρ ∈ Aut(Σ+ ), and ψq is the field automorphism induced by t 7→ tq . Then for α
b ∈ Σ,
i
Xαb ∼
b = {ρ (α)} is the ρ-orbit of α
b = {α} contains only one root, Xαb ∼
= Fq when α
= Fqa if α
with length a, and Xαb is nonabelian if α
b contains a root α and sums of roots in its ρ-orbit.
We need the following, stronger version of Theorem 3.4.
Theorem 4.2 ([St1, § 3]). Assume G is as in Notation 2.2 and 4.1. If α ∈ Aut(G) is such
that α(U) = U, then α = cu df g for unique automorphisms cu ∈ AutU (G), d ∈ Inndiag(G) =
AutT (G), f ∈ ΦG , and g ∈ ΓG .
Proof. Let NAut(G) (U) ≤ Aut(G) and NInndiag(G) (U) ≤ Inndiag(G) be the subgroups of those
automorphisms which send U to itself. Since ΦG ΓG ≤ NAut(G) (U) by definition, Theorem 3.4
implies that NAut(G) (U) = NInndiag(G) (U) · (ΦG ΓG ), a semidirect product. Since ΦG ∩ ΓG = 1,
it remains to show that NInndiag(G) (U) = AutU (G)AutT (G) and AutU (G) ∩ AutT (G) = 1.
The first is immediate: since AutT (G) ≤ NAut(G) (U) and NG (U) = T U,
NInndiag(G) (U) = Inn(G)AutT (G) ∩ NAut(G) (U)
= AutNG (U ) (G)AutT (G) = AutU (G)AutT (G) .
Finally, if cu = ct ∈ Aut(G) where u ∈ U and t ∈ T , then cu = IdG , since u has p-power
order and t has order prime to p.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
31
Lemma 4.3. Assume G ∈ Lie(p). Then for U ∈ Sylp (G), κG sends Out(G) injectively into
Out(U, F ).
Proof. Assume that κG/Z(G) is injective. Since Z(G) is a p′ -group (since Z(G) ≤ T ), and
′
since O p (G) = G by definition of Lie(p), Aut(G) injects into Aut(G/Z(G)), and hence κG
is injective. It thus suffices to prove the lemma when G is in adjoint form.
We can thus assume Notation 4.1. By Lemma 1.4, it will suffice to prove that CAut(G) (U) ≤
Inn(G). Fix β ∈ Aut(G) such that β|U = IdU . By Theorem 4.2, there are unique automorphisms cu ∈ AutU (G), d ∈ AutT (G), f ∈ ΦG , and g ∈ ΓG such that β = cu df g.
If g 6= Id, then it permutes the fundamental root groups nontrivially, while cu df |U sends
each such group to itself modulo higher root groups and commutators. Hence g = Id.
Similarly, f = Id, since otherwise β would act on the fundamental root groups (modulo
higher root groups) via some automorphism other than a translation.
Thus β = cu d, where d = ct for some t ∈ NT (G). Then u has p-power order while t has
order prime to p, so d|U = ct |U = Id. By Lemma 2.4(c), ct sends each root group in U to
itself via xα (u) 7→ xα (θα (t)·u) for some character θα ∈ Hom(T , F×
p ) which is linear in α. For
b
each α
b ∈ Σ+ , ct |Xαb = Id implies that θα (t) = 1 for all α ∈ α
b. Thus θα (t) = 1 for all α ∈ Σ+ ,
so ct = IdG , and β = cu ∈ Inn(G).
It now remains, when proving Theorem A, to show the surjectivity of κG . This will be
done case-by-case. We first handle groups of Lie rank at least three, then those of rank one,
and finally those of rank two.
For simplicity, we state the next two propositions only for groups of adjoint type, but they
also hold without this restriction. The first implies that each element of Aut(U, F ) permutes
the subgroups UJb (as defined in Notation 4.1), and that each element of AutItyp (LcS (G))
b
induces an automorphism of the amalgam of parabolics P b for Jb $ Π.
J
Proposition 4.4. Assume Notation 4.1. For 1 6= P ≤ U, the following are equivalent:
(i)
(ii)
b
P = UJb for some Jb $ Π;
P E B, CU (P ) ≤ P , and Op (OutF (P )) = 1; and
(iii) P E B, CG (P ) ≤ P , and Op (NG (P )) = P .
Hence for each ϕ ∈ Aut(U, F ), ϕ permutes the subgroups UJb, and in particular permutes the
b
subgroups Uαb for α
b ∈ Π.
b CG (U b) = Z(U b) by [GLS3, Theorem 2.6.5(e)] (recall
Proof. (i) =⇒ (iii): For each Jb $ Π,
J
J
that G is of adjoint type). Also, Op (NG (UJb)) = Op (PJb) = UJb, and UJb is normal in B since
NG (UJb) = PJb ≥ B.
(iii) =⇒ (ii): This holds since OutF (P ) ∼
= NG (P )/P CG (P ).
(ii) =⇒ (i):
In this case, P E B, so NG (P ) ≥ B, and NG (P ) = PJb for some
b (cf. [Ca, Theorem 8.3.2]). Then P ≤ Op (P b) = U b. Also, U bCG (P ) P CG (P ) ≤
Jb $ Π
J
J
J
Op (NG (P )/P CG(P )) = 1, so UJb ≤ P CG (P ). Since UJb ≤ U, this implies that UJb ≤
P CU (P ) = P ; i.e., that P = UJb. So (i) holds.
The last statement follows from the equivalence of (i) and (ii).
When G has large Lie rank, Theorem A now follows from properties of Tits buildings.
32
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Proposition 4.5. Assume G ∈ Lie(p) is of adjoint type and has Lie rank at least 3. Fix
U ∈ Sylp (G). Then κG is split surjective.
Proof. Set L = LcU (G). By Proposition 4.4, for each α ∈ AutItyp (L), α permutes the subb For each such J,
b CG (U b) = Z(U b), so AutL (U b) = NG (U b) = P b.
groups UJb for Jb $ Π.
J
J
J
J
J
Thus α induces an automorphism of the amalgam of parabolic subgroups PJb. Since G is
the amalgamated sum of these subgroups by a theorem of Tits (see [Ti, Theorem 13.5] or
[Se, p. 95, Corollary 3]), α extends to a unique automorphism α of G.
Thus α 7→ α defines a homomorphism sb: AutItyp (L) −−−→ Aut(G). If α = cγ for γ ∈
s factors
AutL (U) = NG (U), then α is conjugation by γ ∈ G and hence lies in Inn(G). Hence b
through s : Outtyp (L) −−−→ Out(G), κG ◦ s = IdOuttyp (L) , and thus κG is split surjective.
Before we can handle the rank 1 case, two elementary lemmas are needed.
Lemma 4.6. Let G be a finite group with normal Sylow p-subgroup S E G. Fix subgroups
1 = S0 < S1 < · · · < Sk = S normal in G such that the following hold:
(i)
(ii)
Sk−1 ≤ Fr(S);
CG (S) ≤ S; and
(iii) for each 1 ≤ i ≤ k − 1, Si is characteristic in G, [S, Si ] ≤ Si−1 , Si /Si−1 has exponent
p, and HomFp [G/S] (S/Fr(S), Si /Si−1 ) = 0 (i.e., no irreducible Fp [G/S]-submodule of
Si /Si−1 appears as a submodule of S/Fr(S)).
Let α ∈ Aut(G) be such that [α, S] ≤ Sk−1 . Then α ∈ AutS (G).
Proof. For 1 6= g ∈ G of order prime to p, the conjugation action of g on S is nontrivial since
CG (S) ≤ S, and hence the conjugation action on S/Fr(S) is also nontrivial (see [G, Theorem
5.3.5]). Thus G/S acts faithfully on S/Fr(S). Since α induces the identity on S/Fr(S), α
also induces the identity on G/S.
Assume first that α|S = Id. Since S is a p-group and G/S has order prime to p,
H 1 (G/S; Z(S)) = 0. So by [OV, Lemma 1.2], α ∈ Inn(G). If g ∈ G is such that α = cg ,
then [g, S] = 1 since α|S = Id, and g ∈ S since G/S acts faithfully on S/Fr(S). Thus
α ∈ AutS (G) in this case.
In particular, this proves the lemma when k = 1. So assume k ≥ 2. We can assume
inductively that the lemma holds for G/S1 , and hence can arrange (after composing by an
appropriate element of AutS (G)) that α induces the identity on G/S1 .
Let ϕ ∈ Hom(S, S1 ) be such that α(x) = xϕ(x) for each x ∈ S (a homomorphism since
S1 ≤ Z(S)). Then ϕ factors through ϕ ∈ Hom(S/Fr(S), S1) since S1 is elementary abelian,
and ϕ is a homomorphism of Fp [G/S]-modules since α(g) ≡ g (mod S1 ) for each g ∈ G
(and S1 ≤ Z(S)). Thus ϕ = 1 since HomG/S (Sk /Sk−1 , S1 ) = 0 by (iii), so α|S = Id, and we
already showed that this implies α ∈ AutS (G).
The next lemma will be useful when checking the hypotheses of Lemma 4.6.
Lemma 4.7. Fix a prime p and e ≥ 1, and set q = pe and Γ = F×
q . For each a ∈ Z, set
a
Va = Fq , regarded as an Fp Γ-module with action λ(x) = λ x for λ ∈ Γ and x ∈ Fq .
(a) For each a, Va is Fp Γ-irreducible if and only if a/ gcd(a, q − 1) does not divide pt − 1 for
any t|e, t < e.
(b) For each a, b ∈ Z, Va ∼
= Vb as Fp Γ-modules if and only if a ≡ bpi (mod q − 1) for some
i ∈ Z.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
33
Proof. (a) Set d = gcd(a, q − 1), and let t be the order of p in (Z/ q−1
)× . Thus t|e since
d
q−1
(pe − 1). If t < e, then λa ∈ Fpt for each λ ∈ Fq , so 0 6= Fpt $ Va is a proper
d
Fp Γ-submodule, and Va is reducible.
Conversely, if Va is reducible, then it contains a proper submodule 0 6= W $ Va of
, so
dimension i, some 0 < i < e. All Γ-orbits in Va r0, hence in W r0, have length q−1
d
q−1
i
(p − 1), and t ≤ i < e.
d
(b) For each a ∈ Z, let V a ∼
= Fq be the Fq Γ-module where Γ acts via λ(x) = λa x. Then
Fq ⊗Fp Va ∼
= V a if and only if b ≡ a
= V a ⊕ V ap ⊕ · · · ⊕ V ape−1 as Fq Γ-modules. Since V b ∼
i
∼
(mod q − 1), Vb = Va if and only if b ≡ ap (mod q − 1) for some i.
In principle, we don’t need to look at the fusion systems of the simple groups of Lie rank 1
if we only want to prove tameness. Their fusion is controlled by the Borel subgroup, so their
fusion systems are tame by Proposition 1.6. But the following proposition is needed when
proving Theorem A in its stronger form, and will also be used when working with groups of
larger Lie rank.
Proposition 4.8. Fix a prime p, and a group G ∈ Lie(p) of Lie rank 1. Assume (G, p) ∼
6=
(Sz(2), 2). Then each ϕ ∈ Aut(U, F ) extends to an automorphism of G. Also, if [ϕ, U] ≤
[U, U], then ϕ ∈ Inn(U).
Proof. If G is of universal form, then Z(G) is cyclic of order prime to p by Proposition 3.8. For each Z ≤ Z(G), Out(G/Z) ∼
= Out(G) by [GLS3, Theorem 2.5.14(d)], and
∼
Out(U, FU (G/Z)) = Out(U, FU (G)) since G and G/Z have the same p-fusion systems. It
thus suffices to prove the proposition when G has adjoint form.
Assume first G = PSL2 (q). Thus U ∼
= Fq (as an additive group), T ∼
= C(q−1)/ε where
def
ε = gcd(q − 1, 2), and Γ = AutT (U) is the subgroup of index ε in F×
q . If ϕ ∈ Aut(U) is
fusion preserving, then under these identifications, there is α ∈ Aut(Γ) such that α(u)ϕ(v) =
ϕ(uv) for each u ∈ Γ ≤ F×
q and v ∈ Fq . After composing with an appropriate diagonal
automorphism (conjugation by a diagonal element of P GL2 (q)), we can assume that ϕ(1) =
1. Hence the above formula (with v = 1) implies that α = ϕ|Γ , and thus that ϕ(uv) =
ϕ(u)ϕ(v) for each u, v ∈ Fq with u ∈ Γ. If ε = 1, then ϕ acts as a field automorphism on
U, hence is the restriction of a field automorphism of G, and we are done. Otherwise, there
is u ∈ Γ such that Fq = Fp (u), u and ϕ(u) have the same minimal polynomial over Fp , and
there is ψ ∈ Aut(Fq ) (a field automorphism) such that ψ(u) = ϕ(u). Thus ψ(ui ) = ϕ(ui )
for each i, so ψ = ϕ since both are additive homomorphisms, and hence ϕ extends to a field
automorphism of G. (Note that this argument also holds when q = 3 and Γ = 1.)
Next assume G = PSU3 (q). Following the conventions in [H, Satz II.10.12(b)], we identify
1 a b
where
[[a, b]] = 0 1 −aq ;
U = [[a, b]] a, b ∈ Fq2 , b + bq = −aq+1
0 0 1
×
where
d(λ) = diag(λ−q , λq−1 , λ).
T = d(λ) λ ∈ Fq2
Here, whenever we write a matrix, we mean its class in PSU3 (q). Then B = UT = NG (U) ≤
G (see [H, Satz II.10.12(b)]), and
[[a, b]] · [[c, d]] = [[a + c, b + d − acq ]]
and
d(λ)
[[a, b]] = [[λ1−2q a, λ−1−q b]] .
Set ε = gcd(2q − 1, q 2 − 1) = gcd(2q − 1, q 2 − 2q) = gcd(q + 1, 3). Then d(λ) = 1 exactly
when λε = 1, CT (U) = 1, and hence |T | = |AutB (U/Z(U))| = (q 2 − 1)/ε. If q > 2, then
|T | does not divide pi − 1 for any power 1 < pi < q 2 , and by Lemma 4.7(a), U/Z(U) and
34
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Z(U) are both irreducible as Fp [T ]-modules. (Note, in particular, the cases q = 5 and q = 8,
where (U/Z(U), T ) is isomorphic to (F25 , C8 ) and (F64 , C21 ), respectively.)
Fix ϕ ∈ Aut(U, F ), and extend it to α ∈ Aut(B) (Lemma 1.12). Via the same argument
as that used when G = PSL2 (q), we can arrange (without changing the class of ϕ modulo
Im(κG )) that ϕ ≡ Id (mod [U, U]). If q > 2, then the hypotheses of Lemma 4.6 hold (with
[U, U] < U < B in the role of S1 < S2 = S < G), so α ∈ AutU (B) and ϕ ∈ Inn(U).
If G ∼
= Q8 and T = 1, so Out(U, F ) =
= C32 ⋊Q8 (cf. [Ta, p. 123–124]), then U ∼
= PSU3 (2) ∼
∼
Out(U) = Σ3 . By Theorem 3.4 (or by direct computation), Out(G) = Outdiag(G)ΦG has
order six, since |Outdiag(G)| = gcd(3, q + 1) = 3 and |ΦG | = 2. Thus κG is an isomorphism,
since it is injective by Lemma 4.3.
√
The proof when G = Sz(q) is similar. Set θ = 2q. We follow the notation in [HB,
§ XI.3], and identify U as the group of all S(a, b) for a, b ∈ Fq and T < B = NG (U) as the
group of all d(λ) for λ ∈ F×
q , with relations
S(a, b) · S(c, d) = S(a + c, b + d + aθ c) and
d(λ)
S(a, b) = S(λa, λ1+θ b) .
As in the last case, we can arrange that ϕ ∈ Aut(U, F ) is the identity modulo [U, U]. Since
q ≥ 8 (q 6= 2 by hypothesis), Z(U) and U/Z(U) are nonisomorphic, irreducible F2 T -modules
by Lemma 4.7(a,b) (and since Z(U) ∼
= V1 in the notation of that
= V1+θ and U/Z(U) ∼
lemma). We can thus apply Lemma 4.6 to show that ϕ ∈ Inn(U).
It remains
to handle the Ree groups 2 G2 (q), where q = 3m for some odd m ≥ 1. Set
√
θ = 3q. We use the notation in [HB, Theorem XI.13.2], and identify U = (Fq )3 with
multiplication given by
(x1 , y1 , z1 )·(x2 , y2, z2 ) = (x1 + x2 , y1 + y2 + x1 ·xθ2 , z1 + z2 − x1 ·y2 + y1 ·x2 − x1 ·xθ1 ·x2 ) .
2
Note that xθ = x3 . Let T ≤ B = NG (U) be the set of all d(λ) for λ ∈ F×
q , acting on U via
d(λ)
(x, y, z) = (λx, λθ+1 y, λθ+2z).
Again, we first reduce to the case where ϕ ∈ Aut(U, F ) is such that [ϕ, U] ≤ [U, U], and
extend ϕ to α ∈ Aut(B). If q > 3, then U/[U, U] ∼
= V1 , [U, U]/Z(U) ∼
= Vθ+1 , and Z(U) ∼
= Vθ+2
are irreducible and pairwise nonisomorphic as F3 T -modules by Lemma 4.7 (for Va as defined
in that lemma), since neither θ + 1 nor θ + 2 is a power of 3. So ϕ ∈ Inn(U) by Lemma 4.6.
If q = 3, then U = ha, bi, where |a| = 9, |b| = 3, and [a, b] = a3 . Set Qi = habi i ∼
= C9
0
(i = 0, 1, 2): the three subgroups of U isomorphic to C9 . Let Aut (U) ≤ Aut(U) be the
group of those α ∈ Aut(U) which send each Qi to itself. For each such α, the induced
action on U/Z(U) sends each subgroup of order three to itself, hence is the identity or
(g 7→ g −1 ), and the latter is seen to be impossible using the relation [a, b] = a3 . Thus each
α ∈ Aut0 (U) induces the identity on U/Z(U) and on Z(U), and has the form α(g) = gϕ(g)
for some ϕ ∈ Hom(U/Z(U), Z(U)). So Aut0 (U) = Inn(U) since they both have order 9
(and clearly Inn(U) ≤ Aut0 (U)). The action of Aut(U) on {Q0 , Q1 , Q2 } thus defines an
embedding of Out(U) into Σ3 , and the automorphisms (a, b) 7→ (ab, b) and (a, b) 7→ (a−1 , b)
show that Out(U) ∼
= Σ3 . Since |OutF (U)| = 2 and AutF (U) E Aut(U, F ), it follows that
Out(U, F ) = 1 = Out(G). (See also [BC, Theorem 2] for more discussion about Aut(U).)
It remains to show that κG (at the prime p) is surjective when G ∈ Lie(p) has Lie rank
2, with the one exception when G ∼
= SL3 (2). Our proof is based on ideas taken from the
article of Delgado and Stellmacher [DS], even though in the end, we do not actually need to
refer to any of their results in our argument. The third author would like to thank Richard
Weiss for explaining many of the details of how to apply the results in [DS], and also to
Andy Chermak and Sergey Shpectorov for first pointing out the connection.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
35
Fix a prime p, and a finite group G ∈ Lie(p) of Lie rank two. We assume Notation 2.2
and 4.1. In particular, (G, σ) is a σ-setup for G, T ≤ G is a maximal torus, U ∈ Sylp (G)
is generated by the positive root subgroups, and B = NG (U) is a Borel subgroup. Set
b = {b
Π
α1 , α
b2 }, and set P1 = Pαb1 = hB, X−bα1 i and P2 = Pαb2 = hB, X−bα2 i: the two maximal
parabolic subgroups of G containing B. Our proofs are based on the following observation:
Lemma 4.9. Assume, for G ∈ Lie(p) of rank 2 and its amalgam of parabolics as above, that
each automorphism of the amalgam (P1 > B < P2 ) extends to an automorphism
of G.
(∗)
Then κG is surjective.
Here, by an automorphism of the amalgam, we mean a pair (χ1 , χ2 ), where either χi ∈
Aut(Pi ) for i = 1, 2 or χi ∈ Iso(Pi , P3−i ) for i = 1, 2, and also χ1 |B = χ2 |B .
Proof. Set L = LcU (G) and Ui = Op (Pi ). By Proposition 4.4, each χ ∈ AutItyp (L) either
sends U1 and U2 to themselves or exchanges them. For each i = 1, 2, CG (Ui ) ≤ Ui , so
AutL (Ui ) = NG (Ui ) = Pi . Thus χ induces an automorphism of the amalgam (P1 > B <
P2 ). By assumption, this extends to an automorphism χ of G, and κG (χ) = ξ.
Set G = P1 B P2 : the amalgamated free product over B. Let ρ : G −−−→ G be the
*
natural surjective homomorphism. Since each automorphism of the amalgam induces an
automorphism of G, (∗) holds if for each automorphism of (P1 > B < P2 ), the induced
automorphism of G sends Ker(ρ) to itself.
Let Γ be the tree corresponding to the amalgam (P1 > B < P2 ). Thus Γ has a vertex
[gPi ] for each coset gPi (for all g ∈ G and i = 1, 2), and an edge g(eB ) connecting [gP1 ] to
[gP2 ] for each coset gB in G. Also, G acts on Γ via its canonical action on the cosets, and
in particular, it acts on g(eB ) with stabilizer subgroup gB.
Similarly, let ΓG be the graph of G with respect to the same amalgam: the graph with
vertex set (G/P1 ) ∪ (G/P2 ) and edge set G/B. Equivalently, since P1 , P2 , and B are
self-normalizing, ΓG is the graph whose vertices are the maximal parabolics in G and whose
edges are the Borel subgroups. Let ρb: Γ −−−→ ΓG be the canonical map which sends a vertex
[gPi ] in Γ to the vertex in ΓG corresponding to the image of gPi in G.
Fix a subgroup N ≤ G such that (B, N) is a BN-pair for G, and such that B ∩N = T and
N/T ∼
= W0 (where T and W0 are as defined in Notation 2.2). We refer to [Ca, §§ 8.2, 13.5]
for the definition of BN-pairs, and the proof that G has a BN-pair (B, N) which satisfies
these conditions. In order to stay close to the notation in [DS], we also set T : their notation
for the Cartan subgroup. For i = 1, 2, choose ti ∈ (N ∩ Pi )rB = (N ∩ Pi )rT . Since
(N ∩ Pi )/T ∼
= C2 and N = hN ∩ P1 , N ∩ P2 i, we have N = T ht1 , t2 i, consistent with the
notation in [DS]. Note that T can be the trivial subgroup. We also regard the ti ∈ Pi as
elements of G, and T ≤ B as a subgroup of G, when appropriate.
Let T be the union of the edges in the T ht1 , t2 i-orbit of eB . Thus T is a path of infinite
length in Γ of the following form:
···
t1 t2 t1 (eB )
•
t1 t2 (eB )
[t1 t2 P1 ]
•
[t1 P2 ]
t1 (eB )
•
[P1 ]
eB
•
[P2 ]
t2 (eB )
•
[t2 P1 ]
t2 t1 (eB )
•
t2 t1 t2 (eB )
[t2 t1 P2 ]
···
Thus ρb(T ) is an apartment in the building ΓG under Tits’s definition and construction of
these structures in [Ti, 3.2.6].
A path in Γ is always understood not to double back on itself.
36
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Lemma 4.10. Let G and Γ be as above. Let n ∈ {3, 4, 6, 8} be such that W0 ∼
= D2n . Then
each path in Γ of length at most n + 1 is contained in g(T ) for some g ∈ G.
Proof. A path of length 1 is an edge, and is in the G-orbit of eB which has stabilizer group
B. If eB is extended to a path of length 2 with the edge ti (eB ) (i = 1 or 2), then this path
has stabilizer group
Y
B ∩ tiB =
Xαb · T .
b + r{b
α
b∈Σ
αi }
(Recall that tiXαbi = X−bαi , and X−bαi ∩ B = 1 by [Ca, Lemma 7.1.2].) Thus the stabilizer
subgroup has index pj in B, where pj = |Xαbi |. Furthermore, |Pi /B| = 1 + pj , since by [Ca,
Proposition 8.2.2(ii)],
Pi = B ∪ (Bti B) where |Bti B| = |B| · |B/(B ∩ tiB)| = |B| · pj .
Hence there are exactly pj extensions of eB to a path of length 2 containing the vertex [Pi ]
in the interior, and these are permuted transitively by B.
Upon continuing this argument, we see inductively that for all 2 ≤ k ≤ n + 1, the paths of
length k starting at eB with endpoint [P3−i ] are permuted transitively by B, and of them,
the one contained in T has stabilizer subgroup the product of T with (n + 1 − k) root
subgroups in U. (Recall that B = T U, and U is the product of n root subgroups.) Since G
acts transitively on the set of edges in Γ, each path of length k is in the G-orbit of one which
begins with eB (and with endpoint [P1 ] or [P2 ]), and hence in the G-orbit of a subpath of
T.
Proposition 4.11. Let G, G, and (T, t1 , t2 ) be as above, and let n be such that W0 ∼
= D2n .
Assume that
for each (χ1 , χ2 ) ∈ Aut P1 > B < P2 , where χi ∈ Aut(Pi ) or χi ∈
(†)
Iso(Pi , P3−i ) for i = 1, 2, we have (χ1 (t1 )χ2 (t2 ))n ∈ χ1 (T ).
Then (∗) holds (each automorphism of (P1 > B < P2 ) extends to an automorphism of G),
and hence κG is onto.
Proof. Let ≈ be the equivalence relation on the set of vertices in Γ generated by setting x ≈ y
if x and y are of distance 2n apart in some path in the G-orbit of T . Since T ht1 , t2 i/T ∼
= D2n
as a subgroup of NG (T )/T , the natural map ρb: Γ −−−→ ΓG sends T to a loop of length 2n,
and hence sends all apartments in the G-orbit of T to loops of length 2n. Hence Γ −−−→ ΓG
factors through Γ/≈.
We claim that
ΓG contains no loops of length strictly less than 2n; and
(1)
each pair of points in Γ/≈ is connected by a path of length at most n.
(2)
Assume (1) does not hold: let L be a loop of minimal length 2k (k < n), and fix edges
σi = [xi , yi ] in L (i = 1, 2) such that the path from xi to y3−i in L has length k − 1. Since
ΓG is a building whose apartments are loops of length 2n [Ti, 3.2.6], there is an apartment
Σ which contains σ1 and σ2 . By [Ti, Theorem 3.3] or [Br, p. 86], there is a retraction of ΓG
onto Σ. Hence the path from xi to y3−i in Σ (for i = 1, 2) has length at most k − 1, these
two paths must be equal to the minimal paths in L since there are no loops of length less
than 2k, and this contradicts the assumption that L and Σ are loops of different lengths.
(See also [Br, § IV.3, Exercise 1]. Point (1) also follows since ΓG is a generalized n-gon in
the sense of Tits [Br, p. 117], and hence any two vertices are joined by at most one path of
length less than n.)
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
37
Now assume (2) does not hold: let x, y be vertices in Γ such that the shortest path between
their classes in Γ/≈ has length k ≥ n + 1. Upon replacing x and y by other vertices in their
equivalence classes, if needed, we can assume that the path [x, y] in Γ has length k. Let x∗
be the vertex in the path [x, y] of distance n + 1 from x. By Lemma 4.10, [x, z] is contained
in g(T ) for some g ∈ G; let x′ be the vertex in g(T ) of distance 2n from x and distance
n − 1 from z. Then x′ ≈ x, and [x′ , y] has length at most (n − 1) + (k − n − 1) = k − 2, a
contradiction. This proves (2).
Assume the map (Γ/≈) −−−→ ΓG induced by ρb is not an isomorphism of graphs, and let
x and y be distinct vertices in Γ/≈ whose images are equal in ΓG . By (2), there is a path
from x to y of length at most n, and of even length since the graph is bipartite. This path
cannot have length 2 since ρb: Γ −−−→ ΓG preserves valence, so its image in ΓG is a loop of
length at most n, and this contradicts (1). We conclude that ΓG ∼
= Γ/≈.
Now let (χ1 , χ2 ) be an automorphism of the amalgam (P1 > B < P2 ). Let χ ∈ Aut(G)
be the induced automorphism of the amalgamated free product, and let χ
b ∈ Aut(Γ) be
the automorphism which sends a vertex [gPi ] to [χ(gPi )]. Since (χ1 (t1 )χ2 (t2 ))n = 1 in
G by assumption, ρb(b
χ(T )) is a loop of length 2n in ΓG . Hence ρb ◦ ψb factors through
(Γ/≈) ∼
= Γ/≈, the induced
= ΓG , and by an argument similar to that used to show that ΓG ∼
ρ
map ΓG −−−→ ΓG is an automorphism of ΓG . So χ sends Ker[G −−−→ G] to itself, and thus
induces an automorphism of G. The last statement (κG is onto) now follows from Lemma
4.9.
It remains to find conditions under which (†) holds. The following proposition handles all
but a small number of cases.
Proposition 4.12. Assume N = NG (T ) (and hence NG (T )/T is dihedral of order 2n).
Then (†) holds, and hence each automorphism of the amalgam (P1 > B < P2 ) extends to
an automorphism of G. In particular, (†) and (∗) hold, and hence κG is onto, whenever
G = rXn (q) ∈ Lie(p) has Lie rank 2 for q > 2 and G ∼
6= Sp4 (3).
Proof. Assume that NG (T ) = N = T ht1 , t2 i. Then the choices of the ti are unique modulo
T . Also, any two choices of T are B-conjugate, so each automorphism of the amalgam is Bconjugate to one which sends T to itself. Thus (†) holds, and so (∗) follows from Proposition
4.11.
The last statement now follows from Proposition 3.9. Note that if (†) holds for G of
universal type, then it also holds for G/Z(G) of adjoint type.
What can go wrong, and what does go wrong when G = SL3 (2), is that an automorphism
of the amalgam can send t1 , t2 to another pair of elements whose product (modulo T ) has
order strictly greater than 2n. This happens when T is sent to another path not in the
G-orbit of T : one whose image in ΓG is a loop of a different length.
Example 4.13. Assume G = SL3 (2). In particular, T = 1. Let B be the group of upper
triangular matrices, let t1 and t2 be the permutation matrices for (1 2) and (2 3), respectively,
and set Pi = hB, ti i.
Consider the automorphism α of the amalgam which is the identity on P1 (hence on B),
and which is conjugation by e13 (the involution in Z(B)) on P2 . Set t′i = α(ti ). Thus
1 1 1
0 1 0
and
t′2 = 0 0 1 .
t′1 = 1 0 0
0 0 1
0 1 0
38
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
One checks that t′1 t′2 has order 4, so that ht′1 , t′2 i ∼
= D8 while ht1 , t2 i ∼
= D6 . In other words, α
sends the lifting (from ΓG to Γ) of a loop of length 6 to the lifting of a loop of length 8, hence
is not compatible with the relation ≈, hence does not extend to an automorphism of G.
We are left with seven cases: four cases with n = 4, two with n = 6, and one with n = 8.
Those with n = 4 are relatively easy to handle.
Proposition 4.14. Assume G is one of the groups Sp4 (2), PSp4 (3), PSU4 (2), or PSU5 (2).
Then (†) holds, and hence (∗) also holds and κG is onto.
Proof. In all cases, we work in the universal groups Sp4 (q) and SUn (2), but the arguments
are unchanged if we replace the subgroups described below by their images in the adjoint
group. Recall that p is always the defining characteristic, so the second and third cases are
distinct, even though PSp4 (3) ∼
= SU4 (2) (see [Wi, § 3.12.4] or [Ta, Corollary 10.19]).
Let (χ1 , χ2 ) be an automorphism of (P1 > B < P2 ). Since all subgroups of B isomorphic
to T are conjugate to T by the Schur-Zassenhaus theorem, we can also assume that χi (T ) =
T . Set χ0 = χ1 |B = χ2 |B and t∗i = χi (ti ) for short; we must show that |t∗1 t∗2 | = n = 4. Note
that t∗1 t∗2 has order at least 4, since otherwise ΓG would contain a loop of length strictly less
than 8 = 2n, which is impossible by point (1) in the proof of Proposition 4.11.
G = Sp4 (2) ∼
= Σ6 : Set G′ = [G, G]: the subgroup of index 2. The elements xγ (1) for
γ ∈ Σ are all Aut(G)-conjugate: the long roots and the short roots are all W -conjugate and
a graph automorphism exchanges them. Since these elements generate G, none of them are
in G′ . Hence for i = 1, 2, all involutions in
xα (1), x−α (1) ∼
= Σ3
= GL2 (2) ∼
i
′
i
′
lie in GrG , and in particular, ti ∈ GrG .
Each automorphism of the amalgam sends the focal subgroup to itself (as a subgroup of
B), and hence also sends the intersections Pi ∩ G′ to themselves. So t∗1 , t∗2 ∈ GrG′ , and
t∗1 t∗2 ∈ G′ ∼
= A6 . It follows that |t∗1 t∗2 | ≤ 5, and |t∗1 t∗2 | = 4 since every dihedral subgroup of
order 10 in Σ6 is contained in A6 .
G = Sp4 (3) : In this case, T ∼
= A4 ≀ C2
= SL2 (3) ≀ C2 . Hence NG (T )/T ∼
= C22 , and NG (T ) ∼
contains elements of order 2, 3, 4, and 6, but no dihedral subgroups of order 12. Since t∗1 t∗2
has order at least 4, |t∗1 t∗2 | = 4, and condition (†) holds.
G = SUn (2) for n = 4 or 5 : We regard these as matrix groups via
SUn (2) = M ∈ SLn (4) M t = M −1
aij t = an+1−j,n+1−i ,
where
and where x = x2 for x ∈ F4 . We can then take B to be the group of upper triangular
matrices in SUn (2), U the group of strict upper triangular matrices, and T the group of
diagonal matrices. We thus have
T = diag(x, x−1 , x−1 , x) x ∈ F4 ∼
if n = 4
= C3
2
T = diag(x, y, xy, y, x) x, y ∈ F4 ∼
if n = 5.
= C3
Since NG (T ) must permute the eigenspaces of the action of T on Fn4 , we have NGUn (2) (T ) ∼
=
GU2 (2) ≀ C2 (if n = 4) or (GU2 (2) ≀ C2 ) × F×
(if
n
=
5).
So
in
both
cases,
4
∼
NG (T )/T = PGU2 (2) ≀ C2 ∼
= C 2 ⋊ D8 .
= Σ3 ≀ C2 ∼
3
Set Q = NG (T )/O3 (NG (T )) ∼
= D8 , and let ψ : NG (T ) −−−→ Q be the natural projection.
Set Q0 = ψ(CG (T )). Since CG (T )/T ∼
= Σ3 × Σ3 (the subgroup of elements which send each
2
∼
eigenspace to itself), Q0 = C2 and CG (T ) = ψ −1 (Q0 ).
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
39
Choose the indexing of the parabolics such that P1 is the subgroup of elements which fix
an isotropic point and P2 of those which fix an isotropic line. Thus
n
o
A
X
A
∈
SL
(4)
if n = 4
t
−1
2
0 (A )
o
n u v x
0 A w
A ∈ GUn−2 (2)
and P2 = A v X
P1 =
0 0 u
0
u
w
A ∈ GL2 (4)
if n = 5
t −1
0 0 (A )
Then ψ(NP1 (T )) ≤ Q0 : no matrix in P1 can normalize T and exchange its eigenspaces. Also,
NB (T ) contains CU (T ) = he1,n (1), e2,n−1 (1)i, where ei,j (u) denotes the elementary matrix
with unique off-diagonal entry u in position (i, j). Thus Q0 ≥ ψ(NP1 (T )) ≥ ψ(NB (T )) ∼
= C22 ,
so these inclusions are all equalities. Also, P2 contains the permutation matrix for the
permutation (1 2)(n−1 n), this element exchanges the eigenspaces of rank 2 for T , and so
ψ(NP2 (T )) = Q.
Since T ht1 , t2 i/T ∼
= D8 , hψ(t1 ), ψ(t2 )i = Q, and so ψ(t1 ) ∈ Q0 rZ(Q) and ψ(t2 ) ∈ QrQ0 .
Since (χ1 , χ2 ) induces an automorphism of the amalgam (Q > Q0 = Q0 ), this implies that
ψ(t∗1 ) ∈ Q0 rZ(Q) and ψ(t∗2 ) ∈ QrQ0 . But then hψ(t∗1 ), ψ(t∗2 )i = Q since these elements
generate modulo Z(Q), so |t∗1 t∗2 | ∈ 4Z, and |t∗1 t∗2 | = 4 since NG (T )/T ∼
= Σ3 ≀ C2 contains no
elements of order 12.
It remains to handle the groups G2 (2), 3D4 (2), and 2F4 (2). In the first two cases, if t∗i is
an arbitrary involution in NPi (T )rNB (T ) for i = 1, 2, then t∗1 t∗2 can have order 6 or 8 (or
order 7 or 12 when G = G2 (2)), and there does not seem to be any way to prove condition
(†) short of analyzing automorphisms of the amalgam sufficiently to prove (∗) directly.
Let {α, β} be a fundamental system in the root system of G2 where α is the long root.
Let α, α′ , α′′ be the three long positive roots, and β, β ′ , β ′′ the three short positive roots, as
described in (3) below.
Let γ0 , γ1 , γ2 , γ3 denote the four fundamental roots in the D4 root system, where γ0 is in
the center of the Dynkin diagram, and the other three are permuted cyclically by the triality
automorphism. Set γij = γi + γj (when it is a root), etc. We identify the six classes of
positive roots in 3D4 with the roots in G2 by identifying the following two diagrams:
γ00123
αO′′
α
′
′′
γ0
′
β
α
f▼▼▼ β X✶✶
q8
✌F
q
q
▼▼▼
✶✶
✌
▼▼▼ ✶✶ ✌✌ qqqqq
▼▼▼✶ ✌✌qqq
▼✶q✌✌q
/ β
o
−β
G2
γ ,
01
γ02 ,
γ03
O
γ ,
012
γ023 ,
γ013
γ
0123
X✶
f▼▼▼
q8
✌F
q
q
▼▼▼ ✶✶✶
✌
▼▼▼ ✶✶ ✌✌ qqqqq
▼▼▼✶ ✌✌qqq
▼✶q✌✌q
/ {γ ,γ ,γ }
o
−b
γ
1
1
2
3
(3)
3
&
D4
−α
&
−γ0
The following list gives all nontrivial commutator relations among root subgroups of G2 (q)
or 3D4 (q) (see [GLS3, Theorems 1.12.1(b) & 2.4.5(b)]):
[xα (u), xβ (v)] ≡ xβ ′ (±uv)xβ ′′ (±uv 1+q )
q
q
[xβ ′ (u), xβ (v)] ≡ xβ ′′ (±(uv + u v))
[xα (u), xα′ (v)] = xα′′ (±uv)
q
(mod Xα′ Xα′′ )
(4)
(mod Xα′ Xα′′ )
(5)
(6)
[xβ ′ (u), xβ ′′ (v)] = xα′′ (±Tr(uv ))
(7)
[xβ ′′ (u), xβ (v)] = xα′ (±Tr(uq v)) .
(8)
40
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Again, Tr : Fq3 −−−→ Fq denotes the trace. Note that when G = G2 (q), then u, v ∈ Fq in all
2
2
cases, and hence uq = uq = u, uq+q = u2 , and Tr(u) = 3u. When G = 3D4 (q), the notation
xβ (−), xβ ′ (−), and xβ ′′ (−) is somewhat ambiguous (and formula (5) depends on making the
right choice), but this doesn’t affect the arguments given below.
Proposition 4.15. Assume p = 2 and G = G2 (2). Then (∗) holds: each automorphism
of the amalgam (Pα > B < Pβ ) extends to an automorphism of G. (In fact, each automorphism of the amalgam is conjugation by some element of B.) In particular, κG is
onto.
Proof. In this case, T = 1, and
Pα ∼
= (C4 × C4 ) ⋊ D12
and
Pβ ∼
= (Q8 ×C2 Q8 ) ⋊ Σ3 .
Also, B = U has presentation U = A ⋊ hr, ti, where
A = ha, bi ∼
= C22 , ra = a−1 , rb = b−1 , ta = b, tb = a .
= C4 × C4 , hr, ti ∼
In terms of the generators xγ = xγ (1) for γ ∈ Σ+ , we have A = xβ ′ xβ , xβ ′′ xβ and Ω1 (A) =
xα′ , xα′′ , and we can take r = xβ ′′ , t = xα , and a = xβ xβ ′′ (and then b = ta). Note that (5)
takes the more precise form [xβ ′ , xβ ] = xα′ xα′′ in this case. Also,
Uα = Ahri ∼
= (C4 × C4 ) ⋊ C2
Uβ = hab−1 , a2 ti ×ha2 b2 i hab, a2 rti ∼
= Q8 ×C2 Q8
′
U ∩ G = Ahti ∼
= C4 ≀ C2 .
′
The last formula holds since G = [G, G] ∼
= SU3 (3) has index two in G (see [Wi, § 4.4.4] or
[Di, pp. 146–150]), since xα , xα′ , xα′′ ∈ G′ (note that xα = [x−β , xβ ′ ]), and since xβ , xβ ′ , and
xβ ′′ are all G-conjugate and hence none of them lies in G′ .
Fix an automorphism (χα , χβ ) of the amalgam (Pα > B < Pβ ), and set χ0 = χα |B = χβ |B .
Then χ0 ∈ Aut(U) sends each of the subgroups Uα , Uβ , and U ∩ G′ to itself. Hence it sends
each quaternion factor in Uβ to itself, and sends Uα ∩G′ = ha, bi to itself. After composing by
an appropriate element of AutU (Pβ ), we can arrange that χ0 (ab) = ab and χ0 (ab−1 ) = ab−1 .
In particular, χ0 induces the identity on Ω1 (A) and hence also on A/Ω1 (A).
Let g ∈ Pα be an element of order 3, chosen so that g(a2 ) = b2 and g(b2 ) = a2 b2 . The
image of hgi in Pα /A ∼
= D12 is normal, so χα (g) ∈ Ag. Let x ∈ Ω1 (A) be such that
χα (b) = χ0 (b) = ax. Then gb ∈ hab, b2 i ≤ CA (χ0 ), so gb = χα (gb) = g(bx) implies that gx = 1
and hence x = 1. Thus χ0 |A = Id. Also, χα (hgi) ∈ Syl3 (Pα ) is conjugate to hgi by an
element of A, so we can arrange that χα (hgi) = hgi and hence that χα |Ahgi = Id. But then
χα is the identity modulo CPα (Ahgi) = Z(Ahgi) = 1, so χα = IdPα .
Since χβ |Uβ = Id, χβ induces the identity modulo CPβ (Uβ ) = Z(Uβ ) ∼
= C2 . It thus has the
form χβ (x) = xψ(x) for some ψ ∈ Hom(Pβ , Z(Uβ )). Hence χβ = Id, since it is the identity
on U ∈ Syl2 (Pβ ).
Proposition 4.16. Assume p = 2 and G = 3D4 (2). Then (∗) holds, and κG is onto.
∼
∼
∼
Proof. In this case, T ∼
= F×
8 = C7 , Pα /Uα = C7 × Σ3 , and Pβ /Uβ = SL2 (8). Also, by (6)
and (7), Uβ is extraspecial with center Xα′′ . Fix an automorphism (χα , χβ ) of the amalgam
(Pα > B < Pβ ), and set χ0 = χα |B = χβ |B . We must show that χα and χβ are the
restrictions of some automorphism of G.
By Theorem 3.4, and since Outdiag(SL2 (8)) = 1 = ΓSL2 (8) , Out(Pβ /Uβ ) ∼
= Out(SL2 (8))
is generated by field automorphisms, and hence automorphisms which are restrictions of
field automorphisms of G. So we can compose χβ and χα by restrictions of elements of
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
41
AutB (G)ΦG = NAutPβ (G) (U)ΦG , to arrange that χβ induces the identity on Pβ /Uβ . Then,
upon composing them by some element of AutU (G), we can also arrange that χ0 (T ) = T .
Since Xβ ′ and Xβ ′′ are dual to each other by (7) and hence nonisomorphic as F2 [T ]-modules,
χ0 sends each of them to itself.
Since χ0 (T ) = T , χ0 sends CU (T ) = Xα Xα′ Xα′′ ∼
= D8 to itself. It cannot exchange the
two subgroups Xα Xα′′ and Xα′ Xα′′ (the first is not contained in Uα and the second is),
so χ0 |CU (T ) ∈ Inn(CU (T )). Hence after composing by an element of AutCU (T ) (G), we can
arrange that χ0 is the identity on this subgroup. Also, by applying (4) with u = 1, and since
χ0 |Xβ ≡ Id (mod Uβ ) and [Xα , Uβ ] ≤ Xα′′ , we see that χ0 is the identity on Xβ ′ Xβ ′′ . We
conclude that χ0 is the identity on Uβ .
Since χβ induces the identity on Uβ and on Pβ /Uβ , it has the form χβ (x) = xψ(x) (all
x ∈ Pβ ) for some
ψ ∈ Hom(Pβ /Uβ ; Z(Uβ )) ∼
= Hom(SL2 (8), C2 ) = 1 .
So χβ = IdPβ .
Now, CPα (T ) ∼
= Σ4 × C7 , and Out(Σ4 ) = 1. Hence χα |CPα (T ) must be conjugation by some
element z ∈ Z(CU (T )) = Xα′′ = Z(Pβ ). After composing χα and χβ by restrictions of cz ,
we can thus assume that χα is the identity on CPα (T ) (and still χβ = IdPβ ). Since χα |U = Id
and Pα = hU, CPα (T )i, we have χα = IdPα .
It remains only to handle 2F4 (2) and the Tits group.
Proposition 4.17. Assume G = 2F4 (2)′ or 2F4 (2). Then κG is an isomorphism.
Proof. By the pullback square in [AOV, Lemma 2.15] (and since Outtyp (L) is independent of
the choice of objects in L by [AOV, Lemma 1.17]), κG is an isomorphism when G = 2F4 (2)
if it is an isomorphism when G is the Tits group. So from now on, we assume G = 2F4 (2)′ .
We adopt the notation for subgroups of G used by Parrott [Pa]. Fix T ∈ Syl2 (G), and
set Z = Z(T ) ∼
= C2 , H = CG (Z), and J = O2 (H). Let z ∈ Z be a generator. Then H
is the parabolic subgroup of order 211 · 5, |J| = 29 , and H/J ∼
= C5 ⋊ C4 . Set E = [J, J].
By [Pa, Lemma 1], E = Z2 (J) = Fr(J) ∼
= C25 , and by the proof of that lemma, the Sylow
5-subgroups of H act irreducibly on J/E ∼
= C24 . Since each element of
= C24 and on E/Z ∼
AutH/J (J/E) sends CJ/E (T /J) ∼
= C2 to itself,
AutH/J (J/E) = {IdJ/E } and |HomH/J (J/E, E/Z)| ≤ |HomH/J (J/E, J/E)| = 2 .
(9)
Let N > T be the other parabolic, and set K = O2 (N). Thus N/K ∼
= Σ3 , and [T : K] = 2.
∼
Fix P ∈ Syl5 (H) ⊆ Syl5 (G) (so P = C5 ). By [Pa, p. 674], H/E = (J/E) · (NG (P )/Z),
where NG (P )/Z ∼
= C5 ⋊ C4 . For each β ∈ Aut(H) such that β(T ) = T , there is
= H/J ∼
β1 ≡ β (mod AutJ (H)) such that β1 (P ) = P . Since each automorphism of H/J which sends
T /J ∼
= C4 to itself is conjugation by an element of T /J, there is β2 ≡ β1 (mod AutNT (P ) (H))
such that β2 induces the identity on H/J. By (9), β2 also induces the identity on J/E, and
hence on H/E = (J/E) · (NG (P )/Z). Thus
NAut(H) (T ) = AutT (H) · {β ∈ Aut(H) | β(P ) = P, [β, H] ≤ E} .
(10)
Now set L = LcT (G) for short, and identify N = AutL (K) and H = AutL (J). For each
α ∈ AutItyp (L), let αH ∈ Aut(H) and αN ∈ Aut(N) be the induced automorphisms, and set
αT = αH |T = αN |T . Set
A0 = α ∈ AutItyp (L) [αH , H] ≤ E and αH |P = IdP .
By (10), each class in Outtyp (L) contains at least one automorphism in A0 .
42
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Fix α ∈ A0 . Since [αH , H] must be normal in H, we have [αH , H] ∈ {E, Z, 1}. If
[αH , H] = Z, then αH |JP = Id, so [αH , K] = [αN , K] = Z, which is impossible since Z is not
normal in N by [Pa, Lemma 6] (or since z ∈
/ Z(G) and G = hH, Ni). Thus either αH = Id,
or [αH , H] = E.
If αH = IdH , then αN |T = Id. In this case, αN determines an element of H 1 (N/K; Z(K))
whose restriction to H 1 (T /K; Z(K)) is trivial, and since this restriction map for H 1 (−; Z(K))
is injective (since T /K ∈ Syl2 (N/K)), αN ∈ Inn(N) (see, e.g., [OV, Lemma 1.2]). Hence
αN ∈ AutZ (N) since αN |T = Id (and Z = Z(T )). So α ∈ AutZ (L) in this case, and
[α] = 1 ∈ Outtyp (L).
Set H = H/Z, and similarly for subgroups of H. Let αH ∈ Aut(H) and αT ∈ Aut(T )
be the automorphisms induced by αH and αT , and set β = αT |J . Then E = Z(J) since
E = Z2 (J), so β(g) = gϕ(b
g ) for some ϕ ∈ HomH/J (J/E, E). If ϕ = 1, so that [α, J] ≤ Z,
then since α|P = Id, we have [αH , H] < E and so αH = Id.
We have now constructed a homomorphism from A0 to HomH/J (J/E, E) with kernel
AutZ (L). Thus
|Outtyp (L)| ≤ |A0 /AutZ (L)| ≤ |HomH/J (J/E, E)| ≤ 2 .
where the last inequality holds by (9). Since |Out(G)| = 2 by [GrL, Theorem 2], and since
κG is injective by Lemma 4.3, this proves that κG is an isomorphism.
Alternatively, this can be shown using results in [Fn]. Since T /[T, T ] ∼
= C2 × C4 by the
above description of T /E (where E ≤ [T, T ]), Aut(T ) and hence Outtyp (L) are 2-groups. So
each automorphism of the amalgam H > T < N determines a larger amalgam. Since the
only extension of this amalgam is to that of 2F4 (2) by [Fn, Theorem 1], |Outtyp (L)| = 2.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
43
5. The cross characteristic case: I
Throughout this section, we will work with groups G = CG (σ) which satisfy the conditions
in Hypotheses 5.1 below. In particular, 5.1(I) implies that G is not a Suzuki or Ree group.
We will see in Section 6 (Proposition 6.8) that while these hypotheses are far from including
all finite Chevalley and Steinberg groups, their fusion systems at the prime p do include
almost all of those we need to consider.
For any finite abelian group B, we denote its “scalar automorphisms” by
ψkB ∈ Aut(B),
ψkB (g) = g k
for all k such that (k, |B|) = 1
and define the group of its scalar automorphisms
Autsc (B) = ψkB (k, |B|) = 1 ≤ Z(Aut(B)) .
Hypotheses 5.1. Assume we are in the situation of Notation 2.2(A,B,C).
(I)
Let p be a prime distinct from q0 such that p |W0 |. Assume also that σ = ψq ◦ γ =
γ ◦ ψq ∈ End(G), where
• q is a power of the prime q0 ;
• ψq ∈ ΦG is the field automorphism (see Definition 3.1(a)); and
• γ ∈ Aut(G) is an algebraic automorphism of finite order which sends T to itself and
commutes with ψq0 (so that ψq0 (G) = G).
Also, there is a free hτ i-orbit of the form {α1 , α2 , . . . , αs } or {±α1 , ±α2 , . . . , ±αs } in
Σ such that the set {α1 , α2 , . . . , αs } is linearly independent in V .
(II) The algebraic group G is of universal type, and NG (T ) contains a Sylow p-subgroup of
G.
(III) Set A = Op (T ). Assume one of the following holds: either
(III.1) q ≡ 1 (mod p), q ≡ 1 (mod 4) if p = 2, |γ| ≤ 2, and γ ∈ ΓG (thus ρ(Π) = Π); or
(III.2) p is odd, q ≡ −1 (mod p), G is a Chevalley group (i.e., γ ∈ Inn(G)), and
γ(t) = t−1 for each t ∈ T ; or
(III.3) p is odd, |τ | = ordp (q) ≥ 2, CA (Op′ (W0 )) = 1, CS (Ω1 (A)) = A, AutG (A) =
AutW0 (A),
NAut(A) (AutW0 (A)) ≤ Autsc (A)AutAut(G) (A)
where AutAut(G) (A) = δ|A δ ∈ Aut(G), δ(A) = A , and
(
/W
γ|A
if 2 ordp (q) or −Id ∈
AutW0 (A) ∩ Autsc (A) ≤
A
γ|A , ψ−1
otherwise,
Since W0 acts on T by Lemma 2.3, it also acts on A = Op (T ).
We will see in Lemma 5.3 that the conditions CS (Ω1 (A)) = A (or CS (A) = A when p = 2)
and AutG (A) = AutW0 (A), both assumed here in (III.3), also hold in cases (III.1) and (III.2).
Recall, in the situation of (III.3), that |τ | = |γ|T | by Lemma 3.2.
Note that the above hypotheses eliminate the possibility that G be a Suzuki or Ree group.
Since we always assume the Sylow p-subgroups are nonabelian, the only such case which
44
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
needs to be considered here (when q0 6= p) is that of 2F4 (q) when p = 3, and this will be
handled separately.
By Lemma 3.2, whenever σ = ψq ◦ γ, and γ is an algebraic automorphism of G which
normalizes T , there is τ ∈ Aut(V ) such that τ (Σ) = Σ and σ(X α ) = X τ (α) for each α ∈ Σ. So
under Hypotheses 5.1, the condition at the beginning of Notation 2.2(C) holds automatically,
and with ρ = τ |Σ . To simplify the notation, throughout this section and the next, we write
τ = ρ to denote this induced permutation of Σ.
The following notation will be used throughout this section, in addition to that in Notation
b and Σ
b are defined in Notation 2.2(C) only when ρ(Π) = Π, and hence only
2.2. Note that Π
in case (III.1) of Hypotheses 5.1. It will be convenient, in some of the proofs in this section,
to extend this definition to case (III.2).
Recall (Notation 2.2) that for α ∈ Σ, wα ∈ W denotes the reflection in the hyperplane
α⊥ ⊆ V .
Notation 5.2. Assume we are in the situation of Notation 2.2 and Hypotheses 5.1.
b = Σ, Π
b = Π, and V0 = V . Note that W0 = W in this case.
(D) If (III.2) holds, then set Σ
b let wαb ∈ W0 be the element in hwα | α ∈ α
(E) If (III.1) holds, then for each α
b ∈ Σ,
bi which
⊥
acts on V0 as the reflection across the hyperplane hb
αi , and which exchanges the positive
and negative roots in the set hb
αi ∩ Σ. (Such an element exists and lies in W0 by [Ca,
Proposition 13.1.2].)
b set
(F) If (III.1) or (III.2) holds, then for each α ∈ Σ and each α
b ∈ Σ,
K α = hX α , X −α i
T α = hα (F×
q0 )
K αb = hK α | α ∈ α
bi
T αb = hT α | α ∈ α
bi .
(G) Set N = NG (T )/Op′ (T ), and identify A = Op (T ) with T /Op′ (T ) E N. If (III.1) or
b set Aαb = A ∩ T αb .
(III.2) holds, then for α
b ∈ Σ,
(H) Fix S ∈ Sylp (G) such that A ≤ S ≤ NG (T ), and set F = FS (G). (Recall that NG (T )
contains a Sylow p-subgroup of G by Hypotheses 5.1(II).) Set
Aut(A, F ) = β ∈ Aut(A) β = β|A , some β ∈ Aut(S, F ) .
Set Autdiag (S, F ) = CAut(S,F ) (A) = β ∈ Aut(S, F ) β|A = Id , and let Outdiag (S, F )
be the image of Autdiag (S, F ) in Out(S, F ).
Note that when (G, σ) is a standard setup (i.e., in case (III.1)), W0 acts faithfully on V0
(see [Ca, Lemma 13.1.1]).
Recall that N = NG (T )/Op′ (T ). We identify A = Op (T ) with T /Op′ (T ) E N.
Lemma 5.3. Assume Hypotheses 5.1 and Notation 5.2.
(a) If condition (III.1) or (III.2) holds, then CW (A) = 1, CG (A) = CG (T ) = T , CG (A) = T ,
and CS (A) = A. If p is odd, then CW (Ω1 (A)) = 1 and CS (Ω1 (A)) = A.
(b) If CG (A)0 = T (in particular, if (III.1) or (III.2) holds), then NG (A) = NG (T ) ≤
NG (T ), and the inclusion of NG (T ) in NG (T ) induces isomorphisms W0 ∼
=
= NG (T )/T ∼
N/A. Thus AutG (A) = AutW0 (A).
Proof. (a) Assume condition (III.1) or (III.2) holds. We first prove that CW (A) = 1, and
also that CW (Ω1 (A)) = 1 when p is odd.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
45
If p is odd, set A0 = Ω1 (A) and pb = p. If p = 2, set A0 = Ω2 (A) and pb = 4. Thus in all
cases, A0 is the pb-torsion subgroup of A. Set ε = 1 if we are in case (III.1), or ε = −1 in
×
case (III.2). By assumption, pb|(q − ε). Choose λ ∈ F×
b. Set
q (or λ ∈ Fq 2 if ε = −1) of order p
Π = {α1 , . . . , αr }. Fix w ∈ CW (A0 ).
Assume first G = G(q), a Chevalley group. Then T = t ∈ T tq−ε = 1 , and A0 contains
all elements of order pb in T . So w = 1 by Lemma 2.7.
Now assume that Id 6= γ ∈ ΓG ; i.e., G is one of the Steinberg groups 2An (q), 2Dn (q), or
2
E6 (q). Then CG (γ) is a simple algebraic group of type Bm , Cm , or F4 (cf. [Ca, § 13.1–3])
b ⊆ V0 = CV (τ ), and A0 contains all pb-torsion in C (γ). By Lemma 2.7
with root system Σ
T
again, w|V0 = Id. Since w and τ are both orthogonal, w also sends the (−1)-eigenspace for
the action of τ to itself, and thus w ∈ CW (τ ) = W0 . But W0 acts faithfully on V0 (see, e.g.,
[Ca, 13.1.1]), so w = 1.
Thus CW (A0 ) = 1. Hence CG (A0 ) = T by Proposition 2.5, and the other statements
follow immediately.
(b) If CG (A)0 = T , then NG (T ) ≤ NG (A) ≤ NG (T ) (recall that A is the p-power torsion in
T ). If g ∈ NG (T ) and σ(g) = g, then g also normalizes T = CT (σ). Thus NG (T ) = NG (A) ≤
NG (T ), and hence NG (T )/T ∼
= W0 by Lemma 2.3. The identification N/A ∼
= NG (T )/T is
immediate from the definition of N.
Recall (Notation 5.2(F)) that when case (III.1) of Hypotheses 5.1 holds (in particular,
b where K α = hX α , X −α i. The conditions
bi for α
b ∈ Σ,
when p = 2), we set K αb = hK α | α ∈ α
b is of the form {α}, {α, τ (α)}, or {α, τ (α), α + τ (α)} for
in (III.1) imply that each class in Σ
some α. This last case occurs only when G ∼
= SUn (q) for some odd n ≥ 3 and some q ≡ 1
(mod p or mod 4).
Lemma 5.4. Assume Hypotheses 5.1, case (III.1), and Notation 5.2. For each α ∈ Σ,
b K αb ∼
Kα ∼
b ∈ Σ,
= SL2 (Fq0 ). For each α
= SL2 (Fq0 ), SL2 (Fq0 ) × SL2 (Fq0 ), or SL3 (Fq0 ) whenever
the class α
b has order 1, 2, or 3, respectively. Also, G ∩ K αb is isomorphic to SL2 (q), SL2 (q 2 ),
or SU3 (q), respectively, in these three cases.
b is in the W0 -orbit of a class in Π.
b So it suffices to
Proof. By Lemma 3.10, each class in Σ
b is its equivalence
prove the statements about K α and K αb when α ∈ Π, and when α
b∈Π
class.
By Lemma 2.4(b) (and since G is universal), K α ∼
= SL2 (Fq0 ) for each α ∈ Π. So when
∼
α = τ (α) (when |b
α| = 1), K αb = K α = SL2 (Fq ).
0
When α 6= τ (α) and they are not orthogonal, then G ∼
= SL2n+1 (Fq0 ) for some n, and the
inclusion of SL3 (Fq0 ) is clear. When α ⊥ τ (α), then [K α , K τ (α) ] = 1, and K α ∩ K τ (α) = 1
by Lemma 2.4(b) and since G is universal, and since the intersection is contained in the
centers of the two factors and hence in the maximal tori. Hence K̄αb = hX̄±α , X̄±τ (α) i ∼
=
K α × K τ (α) ∼
= SL2 (Fq0 ) × SL2 (Fq0 ).
In all cases, since G is universal, G ∩ K αb = CG (σ) ∩ K αb = CK αb (σ). If α = τ (α), then γ
acts trivially on K αb , and CK αb (σ) ∼
= SL2 (q). If α ⊥ τ (α) then γ exchanges the two factors
2
∼
and CK αb (σ) = SL2 (q ). Finally, if α 6= τ (α) and they are not orthogonal, then γ is the
graph automorphism of SL3 (Fq ), so C (σ) ∼
= SU3 (q).
0
Kα
b
46
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
We also recall here the definition of the focal subgroup of a saturated fusion system F
over a finite p-group S:
foc(F ) = xy −1 x, y ∈ S, x is F -conjugate to y .
By the focal subgroup theorem for groups (cf. [G, Theorem 7.3.4]), if F = FS (G) for some
finite group G with S ∈ Sylp (G), then foc(F ) = S ∩ [G, G].
Lemma 5.5. Assume Hypotheses 5.1, case (III.1) or (III.2), and Notation 5.2. Assume
b ≥ 2. Then the following hold.
also that |Π|
b If p = 2, then for each α
b
(a) If p is odd, then [wαb , A] = Aαb for each α
b ∈ Σ.
b ∈ Σ,
[wαb , A] ≤ Aαb with index at most 2, and [wαb , A] = Aαb with the following exceptions:
• τ = Id, G = Cn (or B2 ), and α
b = {α} where α is a long root; or
• |τ | = 2, G = Dn (or A3 ), and α
b = {α, τ (α)} where α ⊥ τ (α).
• |τ | = 2, G = A2n , and |b
α| = 3.
b if and only if [w, A] is cyclic.
(b) For each w ∈ W0 of order 2, w = wαb for some α
b∈Σ
b
(c) If p = 2, then for each α
b ∈ Σ,
(
CG (CA (wαb )) =
if |b
α| ≤ 2
if α
b = {α, τ (α), α + τ (α)}.
T K αb
T K α+τ (α)
If in addition, |b
α| ≤ 2, then
Aαb = A ∩ CG (CA (wαb )), CG (CA (wαb )) = A ∩ foc(CF (CA (wαb ))).
Proof. (a) If we are in case (III.1) of Hypotheses 5.1, then by Lemma 3.10, each orbit of
b contains an element of Π.
b If we are in case (III.2), then W0 = W
W0 under its action on Σ
b = Σ, so the statement holds by the same lemma. So in either case, it suffices to prove
and Σ
b
this when α
b ∈ Π.
b be its class. Since wαb ∈ hwα , wτ (α) i, [wαb , A] ≤ A ∩ T αb = Aαb
Fix α ∈ Π, and let α
b∈Π
in all cases by Lemma 2.4(e). By the same lemma, wαb (b
hα (λ)) = b
hα (λ−1 ) for all λ ∈ F×
q0 if
×
−q
|b
α| ≤ 2; and wαb (b
hα (λ)) = b
hα (λ ) for λ ∈ F 2 if |b
α| = 3. So [wαb , A] = Aαb if p is odd, and
q
[wαb , A] has index at most 2 in Aαb if p = 2.
Assume now that p = 2, and hence that q ≡ 1 (mod 4). If τ = Id (and hence α
b = {α}),
then for each β ∈ Π and each λ ∈ F×
,
Lemma
2.4(e)
implies
that
q
if β ⊥ α
hβ (λ)
wα (hβ (λ)) = hβ (λ)hα (λ) if β 6⊥ α, kβk ≥ kαk
√
h (λ)h (λk ) if β ⊥
6 α, kαk = k · kβk, k = 1, 2, 3.
β
α
(Note that wα (β ∨ ) = β ∨, β ∨ + α∨ , or β ∨ + kα∨ , respectively, in these three cases.) Since T
is generated by the hβ (λ) for β ∈ Π and λ ∈ F×
that [wα , A] has index 2 in Aα
q ,, it follows
√
exactly when there are roots with two lengths and ratio 2, α is a long root, and orthogonal
to all other long roots in Π. This happens only when G ∼
= Cn or B2 .
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
47
Now assume |τ | = 2. In particular, all roots in Σ have the same length. By Lemma 2.4(e)
b we have
again, for each β ∈ Πrb
α such that β 6⊥ α and with class βb ∈ Π,
b
b
b = 1 and λ ∈ F×
if |β|
q
hβ (λ)hα (λ)
b ≥ 2, |b
wαb (b
hβ (λ)) = b
hβ (λ)b
hα (λ)
if |β|
α| = 2, and λ ∈ F×
q2
b
q+1
b
b
hβ (λ)hα (λ ) if |β| ≥ 2, |b
α| = 1 or 3, and λ ∈ F×2
q
Thus [wαb , A] = Aαb exactly when |b
α| = 1, or |b
α| = 2 and there is some β ∈ Π such that
β 6⊥ α and β 6= τ (β). The only cases where this does not happen are when G = Dn or A3
and |b
α| = 2, and when G = A2n and |b
α| ≥ 3.
b [wαb , A] ≤ Aαb by (a), and hence is cyclic. It remains to prove the
(b) For each α
b ∈ Σ,
converse.
When we are in case (III.2) (and hence the setup is not standard), it will be convenient to
define V0 = V . (Recall that V0 is defined in Notation 2.2(C) only when (G, σ) is a standard
setup.) Note that by assumption, G is always a Chevalley group in this case.
Let w ∈ W0 be an element of order 2 which is not equal to wαb for any α
b. If G is a Chevalley
group (if W0 = W and V0 = V ), then CV (w) contains no points in the interior of any Weyl
chamber, since W permutes freely the Weyl chambers (see [Brb, § V.3.2, Théorème 1(iii)]).
Since w is not the reflection in a root hyperplane, it follows that dim(V /CV (w)) ≥ 2. If G
is a Steinberg group (thus in case (III.1) with a standard setup), then W0 acts on V0 as the
Weyl group of a certain root system on V0 (see [Ca, § 13.3]), so dim(V0 /CV0 (w)) ≥ 2 by a
similar argument.
Set ε = +1 if we are in case (III.1), or ε = −1 if we are in case (III.2). Set m = vp (q − ε),
and choose λ ∈ (Fq2 )× of order pm . Set Λ = ZΣ∨ , regarded as the lattice in V with Z-basis
Π∨ = {α∨ | α ∈ Π}. Let
Φλ : Λ/pm Λ −−−−−−→ T
be the Z[W ]-linear monomorphism of Lemma 2.6(a) with image the pm -torsion in T . Thus
Φλ (α∨ ) = hα (λ) for each α ∈ Σ. Also, σ(hα (λ)) = hτ (α) (λ) for each α ∈ Σ (λq = λ by
assumption), and thus Φλ commutes with the actions of τ on Λ < V and of σ on T .
Set Λ0 = CΛ (τ ) in case (III.1), or Λ0 = Λ in case (III.2). Then CΛ/pm Λ (τ ) = Λ0 /pm Λ0 in
case (III.1), since τ permutes the basis Π∨ of Λ. We claim that Φλ restricts to a Z[W0 ]-linear
isomorphism
∼
=
Φ0 : Λ0 /pm Λ0 −−−−−−→ Ωm (A) ,
where Ωm (A) is the pm -torsion subgroup of A and hence of T = CT (σ). If G is a Chevalley
group (in either case (III.1) or (III.2)), then Λ0 = Λ, so Im(Φ0 ) is the pm -torsion subgroup
of T and equal to Ωm (A). If G is a Steinberg group, then ε = +1, each element of order
dividing pm in T is fixed by ψ q , and hence lies in Ωm (A) if and only if it is fixed by γ (thus
in Φλ (CΛ/pm Λ (τ ))).
Thus [w, A] ≥ [w, Ωm (A)] ∼
= [w, Λ0 /pm Λ0 ]. Set B = Λ0 /pm Λ0 for short; we will show that
[w, B] is noncyclic. Set
r = rk(Λ0 ) = dim(V0 )
and
s = rk(CΛ0 (w)) = dimR (CV0 (w)) ≤ r − 2 .
For each b ∈ CB (w), and each v ∈ Λ0 such that b = v + pm Λ0 , v + w(v) ∈ CΛ0 (w) maps to
2b ∈ CB (w). Thus B ∼
= (Z/pm )r , while {2b | b ∈ CB (w)} is contained in CΛ0 (w)/pm CΛ0 (w) ∼
=
m s
m
(Z/p ) . Since p > 2 by assumption (and r − s ≥ 2), it follows that B/CB (w) ∼
= [w, B] is
not cyclic.
b We set up our notation as follows.
(c) Fix α
b ∈ Σ.
48
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Case (1): |b
α| = 1 or 3. Set α∗ = α if α
b = {α} (where τ (α) = α), or α∗ = α + τ (α) if
α
b = {α, τ (α), α + τ (α)}. Set wαb = wα∗ , Wαb = hwαb i, and ∆ = {±α∗ } ⊆ Σ.
Case (2): α
b = {α, τ (α)} where α ⊥ τ (α). Set wαb = wα wτ (α) , Wαb = hwα , wτ (α) i, and
∆ = {±α, ±τ (α)} ⊆ Σ.
In case (1), by Lemma 2.4(c,e),
CT (wαb ) = CT (wα∗ ) = Ker(θα∗ ) = CT (X α∗ ) = CT (X −α∗ ) .
Hence CG (CA (wαb )) ≥ CG (CT (wαb )) ≥ T hX α∗ , X −α∗ i = T K α∗ . In case (2), by the same
lemma,
CT (wαb ) = CT (hwα , wτ (α) i) = CT (hX α , X −α , X τ (α) , X −τ (α) i) = CT (K α K τ (α) )
so that CG (CA (wαb )) ≥ T K αb . This proves one of the inclusions in the first statement in (c).
By Proposition 2.5, the opposite inclusion will follow once we show that
Fix w ∈ CW (CA (wαb )).
CW (CA (wαb )) ≤ Wαb .
(1)
• Let β ∈ Σ ∩ ∆⊥ be such that β = τ (β). Then hβ (λ) ∈ CA (wαb ) for λ ∈ F×
q0 of order 4, so
w(hβ (λ)) = hβ (λ), and β ∈ CV (w) by Lemma 2.6(c).
• Let β ∈ Σ ∩ ∆⊥ be such that β 6= τ (β), and set β ′ = τ (β) for short. Let r ≥ 2 be such
r+1
that q ≡ 1 + 2r (mod 2r+1 ), and choose λ ∈ F×
. Set a = 1 − 2r , so λa = λq .
q0 of order 2
Then
hβ (λ)hβ ′ (λa ), hβ (λa )hβ ′ (λ) ∈ CA (wαb ) ≤ CT (w) .
Also, kβ + aβ ′ k = kaβ + β ′ k < (1 − a)kβk = 21 |λ|kβk since a < 0 and β ′ 6= −β (since
τ (Σ+ ) = Σ+ ). Thus β + aβ ′ , aβ + β ′ ∈ CV (w) by Lemma 2.6(b), so β, β ′ ∈ CV (w).
• Let β ∈ Σ be such that β = τ (β) and β ∈
/ ∆⊥ , and set η = β +wαb (β). Since wαb τ = τ wαb in
Aut(V ), τ (η) = η. Since β ∈
/ ∆⊥ = CV (wαb ), we have wαb (β) 6= β, and hence kηk < 2kβk.
For λ ∈ F×
(λ) ∈ CA (wαb ), so w(t) = t, and η = β + wαb (β) ∈
q0 of order 4, t = hβ (λ)hwα
b (β)
CV (w) by Lemma 2.6(b).
Consider the set
Σ∗ = Σ ∩ ∆⊥ ∪ β + wαb (β) β ∈ Σ, τ (β) = β, β 6⊥ ∆ ⊆ V .
We have just shown that w(η) = η for each η ∈ Σ∗ , and hence that w|hΣ∗ i = Id. From the
description of the root systems in [Brb, Planches I–IX], we see that Σ ∩ (Σ∗ )⊥ = ∆, except
when G ∼
= A2 and τ 6= Id (i.e., when G ∼
= SU3 (q)).
∼
Thus when G 6= SU3 (q), the only reflection hyperplanes which contain hΣ∗ i are those in
the set {β ⊥ | β ∈ ∆}. Fix a “generic” element v ∈ hΣ∗ i; i.e., one which is not contained in any
of these hyperplanes. In case (1), v is contained in only the one reflection hyperplane α∗⊥ ,
and hence is in the closure of exactly two Weyl chambers for (Σ, W ): chambers which are
exchanged by wαb . In case (2), v is contained in the two reflection hyperplanes α⊥ and τ (α)⊥ ,
and hence in the closure of four Weyl chambers which are permuted freely and transitively
by Wαb = hwα , wτ (α) i. Since W permutes the Weyl chambers freely and transitively (see [Brb,
§ V.3.2, Théorème 1(iii)]), and since hw, Wαb i permutes the chambers whose closures contain
v, we have w ∈ Wαb .
This proves (1) when G 6∼
= SU3 (q). If G ∼
= SU3 (q), then hα∗ (−1) ∈ CA (wαb ). But no
element of order 2 in T < SL3 (Fq0 ) centralizes the full Weyl group W ∼
= Σ3 , so (1) also holds
in this case.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
49
If |b
α| ≤ 2, then
CG (CA (wαb )) = G ∩ CG (CA (wαb )) = T (G ∩ K αb )
where by Lemma 5.4, G ∩K αb ∼
= SL2 (q) or SL2 (q 2 ). Hence CG (CA (wαb )) has commutator sub-
group G∩K αb , and focal subgroup Aαb . Since CF (CA (wαb )) is the fusion system of CG (CA (wαb ))
(cf. [AKO, Proposition I.5.4]), this proves the last statement.
Lemma 5.6. Assume Hypotheses 5.1, case (III.1), and Notation 5.2.
b have order 1 or 2. (Equivalently, τ (α) = α or τ (α) ⊥ α
(a) Assume that all classes in Σ
for each α ∈ Σ.) Then CT (W0 ) = CT (W ) = Z(G), and Z(G) = CT (W0 ).
b contains classes of order 3. Then G ∼
(b) Assume that Σ
= SL2n−1 (Fq0 ) and G ∼
= SU2n−1 (q)
×
−q
∼
for some n ≥ 2. Also, CT (W0 ) = Fq0 , and σ(t) = t for all t ∈ CT (W0 ).
Proof. (a) Assume that τ (α) = α or τ (α) ⊥ α for each α ∈ Σ. We first show, for each
b that C (wαb ) = C (wα , wτ (α) ). This is clear if α = τ (α). If α ⊥ τ (α),
α
b = {α, τ (α)} ∈ Π,
T
T
then wαb = wα wτ (α) , so if t ∈ CT (wαb ), then wα (t) = wτ (α) (t) and t−1 wα (t) = t−1 wτ (α) (t).
Also, t−1 wα (t) ∈ T α and t−1 wτ (α) (t) ∈ T τ (α) by Lemma 2.4(e). Since T α ∩ T τ (α) = 1 by
Lemma 2.4(b), t−1 wα (t) = 1, and hence t ∈ CT (wα , wτ (α) ).
Since W = hwα | α ∈ Πi, this proves that CT (W0 ) = CT (W ). Since G is universal,
CT (W ) = Z(G) by Proposition 2.5. In particular, CT (W0 ) ≤ G ∩ Z(G) ≤ Z(G); while
Z(G) ≤ CT (W0 ) since CG (T ) = T by Lemma 5.3(a).
b contains a class of order 3. Then by [GLS3, (2.3.2)], γ =
(b) Assume Σ
6 Id, G = SL2n−1 ,
∼
and G = SU2n−1 (q) (some n ≥ 2). Also, if we identify
T = diag(λ1 , . . . , λ2n−1 ) λi ∈ F×
q0 , λ1 λ2 · · · λ2n−1 = 1 ,
and identify W = Σ2n−1 with its action on T permuting the coordinates, then
−1
γ diag(λ1 , . . . , λ2n−1 ) = diag(λ−1
2n−1 , . . . , λ1 ),
and W0 ∼
= C2 ≀ Σn−1 is generated by the permutations (i 2n−i) and (i j)(2n−i 2n−j)
i, j < n. So CT (W0 ) is the group of all matrices diag(λ1 , . . . , λ2n−1 ) such that λi = λ1
−(2n−2)
−q
all i 6= n and λn = λ1
, and CT (W0 ) ∼
= F×
q0 . Also, γ inverts CT (W0 ), so σ(t) = t
t ∈ CT (W0 ).
for
for
for
Recall (Notation 5.2(H)) that Aut(A, F ) is the group of automorphisms of A which extend
to elements of Aut(S, F ). The next result describes the structure of Aut(A, F ) for a group G
in the situation of case (III.1) or (III.2) of Hypotheses 5.1. Recall that W0 acts faithfully on
A by Lemma 5.3(a), and hence that W0 ∼
= AutN (A) = AutNG (T ) (A) by Lemma 5.3(b). It will
be convenient to identify W0 with this subgroup of Aut(A). Since each element of Aut(A, F )
is fusion preserving, this group normalizes and hence acts on W0 , and W0 Aut(A, F ) is a
subgroup of Aut(A).
For convenience, we set AutAut(G) (A) = δ|A δ ∈ Aut(G), δ(A) = A .
Lemma 5.7. Assume that G and (G, σ) satisfy Hypotheses 5.1, case (III.1) or (III.2).
Assume also Notation 5.2.
(a) CW0 Aut(A,F ) (W0 ) ≤ W0 Autsc (A).
(b) Aut(A, F ) ≤ Autsc (A)AutAut(G) (A), with the exceptions
50
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
• (G, p) ∼
= (2E6 (q), 3), or
• (G, p) ∼
= (G2 (q), 2) and q0 6= 3, or
• (G, p) ∼
= (F4 (q), 3) and q0 6= 2.
(c) In all cases, Aut(A, F ) ∩ Autsc (A)AutAut(G) (A) has index at most 2 in Aut(A, F ).
b and Π
b are defined when ρ(Π) = Π, and hence
Proof. Recall that in Notation 2.2(C), V0 , Σ,
b = Σ, and Π
b = Π in
in case (III.1) of Hypotheses 5.1. In case (III.2), we defined V0 = V , Σ
Notation 5.2(D). So under the hypotheses of the lemma (and since G is always a Chevalley
b = Π if and only if G is a Chevalley group.
group in case (III.2)), we have V0 = V and Π
If α ∈ Π and α+τ (α) ∈ Σ, then T α+τ (α) ≤ T α T τ (α) by Lemma 2.4(d). Hence T αb = T α T τ (α)
(the maximal torus in K αb ) for each α ∈ Π. So by Lemma 2.4(b), in all cases,
Y
Y
CT (ψq γ)
and hence
A=
Aαb .
(2)
T = CT (ψq γ) =
α
b
b
α
b∈Π
Set
(
+1 if q ≡ 1 (mod p) (case (III.1))
ε=
−1 if q ≡ −1 (mod p) and p is odd (case (III.2))
b
α
b∈Π
and
m = vp (q − ε) .
By assumption, ε = 1 if G is a Steinberg group or if p = 2, and m > 0 in all cases.
If G is a Chevalley group, then τ = ε · IdV , so W0 = W . Also, σ = γψq acts on T via
m
σ(t) = tεq , so T = {t ∈ T | tq−ε = 1}, and A = {t ∈ T | tp = 1}. Thus for each α ∈ Σ,
Tα ∼
= Cp k .
= Cq−ε and Aα ∼
b either
Now assume G is a Steinberg group (τ 6= Id). For each α
b ∈ Π,
• α
b ∩ Π = {α, τ (α)} for some α ∈ Π such that α 6= τ (α), in which case σ = ψq γ acts on
T αb = T α × T τ (α) by sending (a, b) to (bq , aq ), and so CT αb (ψq γ) ∼
= Cq2 −1 ; or
• α
b = {α} for some α ∈ Π such that α = τ (α), in which case ψq γ acts on T αb = T α via
(a 7→ aq ), and CT αb (ψq γ) ∼
= Cq−1 .
Since vp (q 2 − 1) = m (p odd) or m + 1 (p = 2), we have now shown that in all cases,
(
if p = 2 and |b
α| = 1
C2m
Aαb ∼
Aαb ∼
= Cpm if p is odd;
=
α| ≥ 2.
C2m+1 if p = 2 and |b
(3)
Step 1: We first prove that
ϕ ∈ CW0 Aut(A,F ) (W0 )
=⇒
b
ϕ(Aαb ) = Aαb for all α
b ∈ Σ.
If p is odd, then Aαb = [wαb , A] by Lemma 5.5(a), so (4) is immediate.
(4)
Next assume that p = 2, and also that |b
α| ≤ 2. Write ϕ = w ◦ ϕ0 , where w ∈ W0
−1
and ϕ0 ∈ Aut(A, F ). Then ϕ0 (CA (wαb )) = w (CA (wαb )) = CA (wβb), where βb = w −1 (b
α). By
definition of Aut(A, F ) (Notation 5.2), ϕ0 = ϕ0 |A for some ϕ0 ∈ Aut(S, F ). Since ϕ0 is fusion
preserving, it sends foc(CF (CA (wαb ))) onto foc(CF (CA (wβb))). Since these focal subgroups are
Aαb and Aβb, respectively, by Lemma 5.5(c), ϕ(Aαb ) = w(Aβb) = Aw(β)
b = Aα
b also in this case
(the second equality by Lemma 2.4(e)).
It remains to consider the case where p = 2 and |b
α| = 3, and thus where G ∼
= SU2n+1 (q)
for some n ≥ 1. There is a subgroup (H1 × · · · × Hn ) ⋊ Σn < G of odd index, where
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
51
Hi ∼
= GU2 (q). Fix Si ∈ Syl2 (Hi ); then Si ∼
= SD2k where k = v2 (q 2 − 1) + 1 ≥ 4. Let
Ai , Qi < Si denote the cyclic and quaternion subgroups of index 2 in Si . Then we can take
A = A1 × · · · × An ∼
= (C2k−1 )n , N = (S1 × · · · × Sn ) ⋊ Σn , and S ∈ Syl2 (N).
b + of order 3, which we label so that [wαb , A] ≤ Ai
There are exactly n classes α
b1 , . . . , α
bn ∈ Σ
i
([wαbi , A] = A ∩ Qi ). Equivalently, these are chosen so that wαbi acts on A via conjugation by
an element of Si rAi . Let αi∗ ∈ Σ+ be the root in the class α
bi which is the sum of the other
two.
Write ϕ = w ◦ ϕ0 , where w ∈ W0 and ϕ0 ∈ Aut(A, F ), and let ϕ0 ∈ Aut(S, F ) be
such that ϕ0 = ϕ0 |A . For each 1 ≤ i ≤ n, ϕ0 (CA (wαbi )) = w −1 (CA (wαbi )) = CA (wαbf (i) ), where
f ∈ Σn is such that α
bf (i) = w −1 (b
αi ). Since ϕ0 is fusion preserving, it sends foc(CF (CA (wαbi )))
onto foc(CF (CA (wαbf (i) ))). By Lemma 5.5(c), CG (CA (wαbi )) = G ∩ (T K α∗i ), its commutator
subgroup is G ∩ K α∗ ∼
= SL2 (q), and hence foc(CF (CA (wαb ))) = Qi . Thus ϕ0 (Qi ) = Qf (i) .
i
i
∼
For each i, set
= SL3 (q)
(Lemma 5.4) with
Sf (i) , and
hence ϕ0 (Ai ) = Af (i) . So ϕ(Ai ) = w(Af (i) ) = Ai for each i where Ai = Aαbi , and this finishes
the proof of (4).
Q∗i
= hQj | j 6= ii. Then CG (Q∗i ) is the product of G ∩ K αbi
Z(Q∗i ). Thus ϕ0 sends foc(CF (Q∗i )) = Si to foc(CF (Q∗f (i) )) =
Step 2:
We next prove point (a): that CW0 Aut(A,F ) (W0 ) ≤ W0 Autsc (A). Let ϕ ∈
W0 Aut(A, F ) be an element which centralizes AutN (A) ∼
= N/A ∼
= W0 . By (4), ϕ(Aαb ) = Aαb
b
b
for each α
b ∈ Σ. Since Aαb is cyclic for each α
b ∈ Σ+ by (3), ϕ|Aαb is multiplication by some
×
unique uαb ∈ (Z/qαb ) , where qαb = |Aαb |. We must show that uαb is independent of α
b.
m
Assume first that τ = Id. By (3), |Aα | = p for each α ∈ Π. Fix α1 , α2 ∈ Π and β ∈ Σ+
such that k1 β = k1 α1 + α2 , where either
• k = 1 and all three roots have the same length; or
√
• k ∈ {2, 3} and kβk = kα1 k = k · kα2 k.
The relation between the three roots is chosen so that hβ (λ) = hα1 (λ)hα2 (λ) for all λ ∈ F×
q0
by Lemma 2.4(d). Hence uα1 ≡ uβ ≡ uα2 (mod pm ) by (2). By the connectivity of the
Dynkin diagram, the uα for α ∈ Π are all equal, and ϕ ∈ Autsc (A).
Now assume |τ | = 2; the argument is similar but slightly more complicated. By assumption, G is of type An , Dn , or En ; i.e., all roots have the same length. Set m′ = vp (q 2 − 1);
then m′ = m if p is odd, and m′ = m + 1 if p = 2. Fix α1 , α2 ∈ Π such that α1 6= τ (α2 ) and
def
m′
β = α1 + α2 ∈ Σ+ . Choose λ ∈ F×
q0 of order p .
′
If α1 6= τ (α1 ) and α2 6= τ (α2 ), then |Aαb1 | = |Aαb2 | = pm by (3), and
b
hα1 (λ)b
hα2 (λ) = hα1 (λ)hτ (α1 ) (λq )hα2 (λ)hτ (α2 ) (λq ) = hβ (λ)hτ (β) (λq ) = b
hβ (λ) ∈ Aβb .
Hence
b
hα1 (λ)b
hα2 (λ)
uβb
=ϕ b
hα1 (λ)b
hα2 (λ) = b
hα1 (λ)uαb1 · b
hα2 (λ)uαb2 ,
′
and together with (2), this proves that uαb1 ≡ uβb ≡ uαb2 (mod pm ).
If τ (αi ) = αi for i = 1, 2, then a similar argument shows that uαb1 ≡ uβb ≡ uαb2 (mod pm ).
′
It remains to handle the case where α1 6= τ (α1 ) and α2 = τ (α2 ). In this case, |Aαb1 | = pm
and |Aαb2 | = pm by (3), and these groups are generated by b
hα1 (λ) = hα1 (λ)hτ (α1 ) (λq ) and
hα2 (λq+1 ), respectively. Then
b
hα1 (λ)b
hα2 (λq+1 ) = hα1 (λ)hτ (α1 ) (λq )hα2 (λq+1 ) = hβ (λ)hτ (β) (λq ) = b
hβ (λ) ∈ Aβb ,
52
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
so
b
hα1 (λ)b
hα2 (λq+1 )
uβb
=ϕ b
hα1 (λ)b
hα2 (λq+1 ) = b
hα1 (λ)uαb1 · hα2 (λq+1 )uαb2 ,
and uαb1 ≡ uβb ≡ uαb2 (mod pm ) by (2) again.
Since the Dynkin diagram is connected, and since the subdiagram of nodes in free orbits
b
in the quotient diagram is also connected, this shows that the uαb are all congruent for α
b∈Π
m
m′
(modulo p or p , depending on where they are defined), and hence that ϕ ∈ Autsc (A).
b
Step 3:
Consider the subset WΠb = {wαb | α
b ∈ Π}.
We need to study the subgroup
NW0 Aut(A,F ) (WΠb ): the group of elements of W0 Aut(A, F ) which permute the set WΠb . Note
b = Π in
that W0 = hWΠb i (see, e.g., [Ca, Proposition 13.1.2], and recall that W0 = W and Π
case (III.2)). We first show that
Aut(A, F ) ≤ W0 NW0 Aut(A,F ) (WΠb ) .
(5)
b = {b
Write Π
α1 , . . . , α
bk }, ordered so that for each 2 ≤ i ≤ k, α
bi is orthogonal to all but one
of the α
bj for j < i. Here, α
bi ⊥ α
bj means orthogonal as vectors in V0 . Thus wαbi commutes
with all but one of the wαbj for j < i. By inspection of the Dynkin diagram of G (or the
quotient of that diagram by τ ), this is always possible.
Fix ϕ ∈ Aut(A, F ). In particular, ϕ normalizes W0 (recall that we identify W0 =
AutW0 (A)) since ϕ is fusion preserving. (Recall that AutG (A) = AutW0 (A) by Lemma
5.3(b).) We must show that some element of ϕW0 normalizes the set WΠb .
By definition of Aut(A, F ) (Notation 5.2), ϕ = ϕ|A for some ϕ ∈ Aut(S, F ). Since ϕ is
fusion preserving, ϕ normalizes AutF (A) = AutG (A), where AutG (A) ∼
= W0 since
= N/A ∼
CN (A) = A by Lemma 5.3(a). Thus there is a unique automorphism ϕ
b ∈ Aut(W0 ) such that
ϕ(w)
b
= ϕ ◦ w ◦ ϕ−1 for each w ∈ W0 .
For each i, since |ϕ(w
b αbi )| = 2 and [ϕ(w
b αbi ), A] ∼
b αbi ) = wαb′i for some
= [wαbi , A] is cyclic, ϕ(w
b by Lemma 5.5(b), where α
α
bi′ ∈ Σ
bi′ is uniquely determined only up to sign. For i 6= j,
α
bi ⊥ α
bj ⇐⇒ [wαbi , wαbj ] = 1 ⇐⇒ [ϕ(w
b αbi ), ϕ(w
b αbj )] = 1 ⇐⇒ α
bi′ ⊥ α
bj′ .
So using the assumption about orthogonality, we can choose successively α
b1′ , α
b2′ , . . . , α
bk′ so
′
′
bj i ≤ 0 for i 6= j.
αi , α
that ϕ(w
b αbi ) = wαb′i for each i, and hb
bi and α
bj is equal
For each i 6= j, since |wαbi wαbj | = |wαb′i wαb′j |, the angle (in V0 ) between α
′
′
to that between α
bi and α
bj (by assumption, both angles are between π/2 and π). The roots
′
b as a root system on V0 with Weyl group W0 , and hence
α
bi for 1 ≤ i ≤ k thus generate Σ
b (Recall that Σ
b = Σ, V0 = V ,
are the fundamental roots for another Weyl chamber for Σ.
and W0 = W in case (III.2).) Since W0 permutes the Weyl chambers transitively [Brb,
§ VI.1.5, Theorem 2(i)], there is w ∈ W0 which sends the set {wαbi } onto {ϕ(w
b αbi )}. Thus
c−1
◦
ϕ
∈
N
(W
),
so
ϕ
∈
W
N
(W
),
and
this
proves
(5).
b
b
0 W0 Aut(A,F )
W0 Aut(A,F )
w
Π
Π
Step 4: Set AutW0 Aut(A,F ) (WΠb ) = NW0 Aut(A,F ) (WΠb ) CW0 Aut(A,F ) (WΠb ): the group of permutations of the set WΠb which are induced by elements of W0 Aut(A, F ). By (a) (Step 2)
and (5), and since W0 = hWΠb i, there is a surjection
onto
AutW0 Aut(A,F ) (WΠb ) −−−−−։
W0 NW0 Aut(A,F ) (WΠb )
W0 Aut(A, F )
=
.
W0 CW0 Aut(A,F ) (WΠb )
W0 Autsc (A)
(6)
To finish the proof of the lemma, we must show that each element of AutW0 Aut(A,F ) (WΠb ) is
represented by an element of AutAut(G) (A) (i.e., the restriction of an automorphism of G),
with the exceptions listed in point (b).
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
53
In the proof of Step 3, we saw that each element of AutW0 Aut(A,F ) (WΠb ) preserves angles
b and hence induces an automorphism of the Coxeter
between the corresponding elements of Π,
b (i.e., the Dynkin diagram without orientation on the edges).
diagram for (V0 , Σ)
Case 1: Assume G = G(q) is a Chevalley group. The automorphisms of the Coxeter
diagram of G are well known, and we have
6 if G = D4
AutW0 Aut(A,F ) (WΠb ) ≤ 2 if G = An (n ≥ 2), Dn (n ≥ 5), E6 , B2 , G2 , or F4
(7)
1 otherwise.
In case (III.1) (i.e., when the setup is standard), all of these automorphisms are realized
by restrictions of graph automorphisms in ΓG (see [Ca, §§ 12.2–4]), except possibly when
G ∼
= B2 (q), G2 (q), or F4 (q). In case (III.2), with the same three exceptions, each such
automorphism is realized by some graph automorphism ϕ ∈ ΓG , and ϕ|T commutes with
σ|T ∈ Z(Aut(T )). Hence by Lemma 3.7, ϕ|T extends to an automorphism of G whose
restriction to A induces the given symmetry of the Coxeter diagram. Together with (6), this
proves the lemma for Chevalley groups, with the above exceptions.
If G ∼
= B2 (q) or F4 (q) and p 6= 2, then AutW0 Aut(A,F ) (WΠb ) = 2, and the nontrivial element
is represented by an element of AutΓG (A) exactly when q0 = 2. This proves the lemma in
these cases, and a similar argument holds when G ∼
= G2 (q) and p 6= 3.
It remains to check the cases where (G, p) ∼
= (B2 (q), 2), (G2 (q), 3), or (F4 (q), 2). We claim
that AutW0 Aut(A,F ) (WΠb ) = 1 in these three cases; then the three groups in (6) are trivial, and
so Aut(A, F ) ≤ W0 Autsc (A). If (G, p) = (B2 , 2) or (G2 , 3), then with the help of Lemma
2.4(d,b), one shows that the subgroups Ω1 (Aα ) are all equal for α a short root, and are all
distinct for the distinct (positive) long roots. More precisely, of the p + 1 subgroups of order
p in Ω1 (A) ∼
= Cp2 , one is equal to Aα when α is any of the short roots in Σ+ , while each of the
other p is equal to Aα for one distinct long root α. Since Ω1 (Aα ) = Ω1 ([wα , A]) for each α, no
element of NW0 Aut(A,F ) (WΠb ) can exchange the long and short roots, so AutW0 Aut(A,F ) (WΠb ) =
1.
Now assume (G, p) = (F4 , 2). Let α, β ∈ Π be such that α is long, β is short, and α 6⊥ β.
Then α and β generate a root system of type B2 , and by the argument in the last paragraph,
no element of NW0 Aut(A,F ) (WΠb ) can exchange them. Thus no element in NW0 Aut(A,F ) (WΠb )
can exchange the long and short roots in G, so again AutW0 Aut(A,F ) (WΠb ) = 1.
Case 2: Assume G is a Steinberg group. In particular, we are in case (III.1). The Coxeter
b has type Bn , Cn , or F4 (recall that we excluded the
diagram for the root system (V0 , Σ)
3
triality groups D4 (q) in Hypotheses 5.1), and hence has a nontrivial automorphism only
when it has type B2 or F4 . It thus suffices to consider the groups G = 2A3 (q), 2A4 (q), and
2
E6 (q).
For these groups, the elements b
hα (λ) for λ ∈ F×
q , and hence the (q − 1)-torsion in the
b + , have relations similar to those among the corresponding subgroups
subgroups Tαb for α
b∈Σ
of T when G = B2 (q) or F4 (q). This follows from Lemma 2.6(a): if λ ∈ F×
q is a generator,
then Φλ restricts to an isomorphism from CZΣ∨ (τ )/(q − 1) to the (q − 1)-torsion in T , and
b can be identified in a natural way with a basis for CZΣ∨ (τ ). Hence when
the elements in Π
b + , proving that no element in
p = 2, certain subgroups Ω1 (Aαb ) are equal for distinct α
b∈Σ
NW0 Aut(A,F ) (WΠb ) can exchange the two classes of roots. Thus the same argument as that
used in Case 1 when (G, p) = (B2 (q), 2) or (F4 (q), 2) applies to prove that NW0 Aut(A,F ) (WΠb ) =
Autsc (A) in these cases.
54
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Since p |W0 | by Hypotheses 5.1(I), we are left only with the case where p = 3 and G =
b is the root system of F4 , so Aut(A, F ) ∩
E6 (q) for some q ≡ 1 (mod 3). Then (V0 , Σ)
W0 Autsc (A) has index at most 2 in Aut(A, F ) by (6) and (7). Thus (c) holds in this case.
(In fact, the fusion system of G is isomorphic to that of F4 (q) by [BMO, Example 4.4], and
does have an “exotic” graph automorphism.)
2
We now look at groups which satisfy any of the cases (III.1), (III.2), or (III.3) in Hypotheses 5.1. Recall that κG = µG ◦ κG : Out(G) −−−→ Out(S, F ).
Lemma 5.8. Assume Hypotheses 5.1 and Notation 5.2. Then each ϕ ∈ Autdiag (S, F ) is the
restriction of a diagonal automorphism of G. More precisely, κG restricts to an epimorphism
from Outdiag(G) onto Outdiag (S, F ) whose kernel is the p′ -torsion subgroup. Also, CA (W0 ) =
Op (Z(G)).
Proof. In general, whenever H is a group and B E H is a normal abelian subgroup, we let
Autdiag (H, B) be the group of all ϕ ∈ Aut(H) such that ϕ|B = IdB and [ϕ, H] ≤ B, and
let Outdiag (H, B) be the image of Autdiag (H, B) in Out(H). There is a natural isomorphism
ηH,B
Autdiag (H, B)/AutB (H) −−∼
−→ H 1 (H/B; B) (cf. [Sz1, 2.8.7]), and hence H 1 (H/B; B) sur=
jects onto Outdiag (H, B). If B is centric in H (if CH (B) = B), then Outdiag (H, B) ∼
=
1
H (H/B; B) since AutB (H) = Inn(H) ∩ Autdiag (H, B).
In particular, Outdiag (S, A) is a p-group since H 1 (S/A; A) is a p-group. Also, CS (A) = A
by Lemma 5.3(a) (or by assumption in case (III.3)), and hence we have Outdiag (S, A) ∼
=
Autdiag (S, A)/AutA (S). So Autdiag (S, A) is a p-group, and its subgroup Autdiag (S, F ) is a
p-group. It follows that
Autdiag (S, F ) ∩ AutG (S) = Autdiag (S, F ) ∩ Inn(S) = AutA (S) ,
and thus Outdiag (S, F ) ∼
= Autdiag (S, F )/AutA (S).
Since Outdiag(G) = OutT (G) by Proposition 3.5(c), κG (Outdiag(G)) ≤ Outdiag (S, F ),
and in particular, κG sends all torsion prime to p in Outdiag(G) to the identity. It remains
to show that it sends Op (Outdiag(G)) isomorphically to Outdiag (S, F ).
Consider the following commutative diagram of automorphism groups and cohomology
groups:
χ
/ H 1 (AutG (A); A)
Outdiag (S, F ) ∼
= Autdiag (S, F )/AutA (S)
ρ2
incl
Outdiag (S, A) ∼
= Autdiag (S, A)/AutA (S)
∼
=
(8)
ηS,A
/
H 1 (AutS (A); A) .
Here, ρ2 is induced by restriction, and is injective by [CE, Theorem XII.10.1] and since
AutS (A) ∈ Sylp (AutG (A)) (since A E S ∈ Sylp (G)). For each ω ∈ Autdiag (S, F ), since
ω is fusion preserving, ηS,A ([ω]) ∈ H 1 (AutS (A); A) is stable with respect to AutG (A)fusion, and hence by [CE, Theorem XII.10.1] is the restriction of a unique element χ([ω]) ∈
H 1 (AutG (A); A).
The rest of the proof splits into two parts, depending on which of cases (III.1), (III.2), or
(III.3) in Hypotheses 5.1 holds. Recall that AutF (A) = AutG (A) = AutW0 (A): the second
equality by Lemma 5.3(b) in cases (III.1) or (III.2), or by assumption in case (III.3).
Cases (III.2) and (III.3): We show that in these cases, Outdiag(G), Outdiag (S, F ), Z(G),
and CA (W0 ) all have order prime to p. Recall that p is odd in both cases. By hypothesis
in case (III.3), and since γ|T ∈ Op′ (W0 ) inverts T in case (III.2), CA (Op′ (W0 )) = 1. In
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
55
particular, CA (W0 ) = 1. Since Z(G) ≤ Z(G) by Proposition 3.5(a), and Z(G) ≤ T by
Lemma 2.4(a), Z(G) ≤ G ∩ CT (W ) ≤ CT (W0 ), so Op (Z(G)) ≤ CA (W0 ) = 1. This proves
the last statement.
Now, Op (Outdiag(G)) = 1 since Outdiag(G) ∼
= Z(G) (see [GLS3, Theorem 2.5.12(c)]) and
Op (Z(G)) = 1. Also,
H 1 (AutG (A); A) = H 1(AutW0 (A); A) ∼
= H 1 (AutW0 (A)/AutOp′ (W0 ) (A); CA (Op′ (W0 ))) = 0
since A is a p-group and CA (Op′ (W0 )) = 1. Hence Outdiag (S, F ) = 1 by diagram (8).
Case (III.1): Since CW (A) = 1 by Lemma 5.3(a) (and since AutG (A) = AutW0 (A)), we
can identify H 1 (AutG (A); A) = H 1 (W0 ; A). Consider the following commutative diagram of
automorphism groups and cohomology groups
Op (Outdiag(G))
R
/
Op (Outdiag (NG (T ), T ))
ηN(T ),T
∼
=
H 1 (W0 ; T )(p)
/
∼
= σ2
∼
= σ1
χ
κG
Outdiag (N, A)
✐✐4
χ0 ✐✐✐✐✐✐
✐
ρ1
✐
✐✐✐✐
✐✐✐✐
incl
/ Outdiag (S, A)
Outdiag (S, F )
ηN,A
+
∼
=
/
H 1 (W0 ; A)
(9)
ρ2
ηS,A
∼
=
/
H 1 (S/A; A)
where R is induced by restriction to NG (T ). By Lemma 5.3(a), T is centric in NG (T ) and
A is centric in N, so the three η’s are well defined and isomorphisms (i.e., Outdiag (N, A) =
Autdiag (N, A)/AutA (N), etc.). The maps σi are induced by dividing out by Op′ (T ), and are
isomorphisms since A = Op (T ). The maps ρi are induced by restriction, and are injective
since S/A ∈ Sylp (W0 ) (see [CE, Theorem XII.10.1]).
Consider the short exact sequence
Ψ
1 −−−→ T −−−−−→ T −−−−−→ T −−−→ 1,
where Ψ(t) = t−1 · γψq (t) = t−1 γ(tq ) for t ∈ T . Let
Ψ
δ
θ
∗
→ CT (W0 ) −−−−→ H 1 (W0 ; T ) −−−−→ H 1 (W0 ; T ) (10)
1 −−−→ CT (W0 ) −−−−→ CT (W0 ) −−−−
be the induced cohomology exact sequence for the W0 -action, and recall that H 1 (W0 ; A) ∼
=
H 1 (W0 ; T )(p) by (9). We claim that
(11) |Op (Outdiag(G))| = |Im(δ)(p) | = |Op (Z(G))| = |CA (W0 )|;
(12) R is injective; and
(13) χ(Outdiag (S, F )) ≤ Ker(θ).
These three points will be shown below. It then follows from the commutativity of diagram (9) (and since Im(δ) = Ker(θ)) that κG sends Op (Outdiag(G)) isomorphically onto
Outdiag (S, F ).
Proof of (11) and (12): Assume first that γ 6= Id and G = SL2n−1 (some n ≥ 1). Thus
G∼
= SU2n−1 (q). By [St1, 3.4], Outdiag(G) and Z(G) are cyclic of order (q + 1, 2n − 1), and
−q
hence have no p-torsion (recall p|(q − 1)). By Lemma 5.6(b), CT (W0 ) ∼
= F×
q0 , and σ(u) = u
for u ∈ CT (W0 ). Thus Ψ∗ (u) = u−1 σ(u) = u−1−q for u ∈ CT (W0 ), so Ψ∗ is onto, and
Im(δ) = 1 ∼
= Op (Outdiag(G)) in this case. Also, CT (W0 ) = Ker(Ψ∗ ) has order q + 1, so
CA (W0 ) = Op (CT (W0 )) = 1.
56
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Now assume γ = Id or G 6= SL2n−1 . By Lemma 5.6, in all such cases,
CT (W0 ) = CT (W ) = Z(G)
and
CT (W0 ) = Z(G) .
(14)
In particular, these groups are all finite, and hence |Im(δ)| = |Z(G)| by the exactness of (10).
By [GLS3, Theorem 2.5.12(c)], Outdiag(G) ∼
= Z(G) in all cases, and hence |Outdiag(G)| =
|Im(δ)|.
If [ϕ] ∈ Ker(R), then we can assume that it is the class of ϕ ∈ AutT (G). Thus ϕ = cx
for some x ∈ NT (G), and ϕ|NG (T ) = cy for some y ∈ NG (T ) which centralizes A. Then
−1
y ∈ CG (A) = T by Lemma 5.3(a), and upon replacing ϕ by c−1
y ◦ ϕ and x by y x (without
changing the class [ϕ]), we can arrange that ϕ|NG (T ) = Id. Then x ∈ CT (W0 ) since it
centralizes NG (T ) (and since NG (T )/T ∼
= W0 by Lemma 5.3(b)), so x ∈ Z(G) by (14), and
hence ϕ = IdG . Thus R is injective.
Proof of (13): Fix ϕ ∈ Autdiag (S, F ). Choose ϕ ∈ Autdiag (N, A) such that ϕ|S = ϕ (i.e.,
such that [ϕ] = χ0 ([ϕ]) in diagram (9)). Recall that W0 ∼
= N/A by Lemma 5.3(b). Let
∼
c : W0 = N/A −−−→ A be such that ϕ(g) = c(gA)·g for each g ∈ N; thus ηN,A ([ϕ]) = [c]. We
must show that θ([c]) = 1: that this is a consequence of ϕ being fusion preserving.
b set uαb = c(wαb ). Thus for g ∈ N, ϕ(g) = uαb g if g ∈ wαb (as a coset of
For each α
b ∈ Π,
A in N). Since wα2b = 1, g 2 = ϕ(g 2 ) = (uαb g)2, and hence wαb (uαb ) = u−1
α
b . We claim that
b
b ∈ Π.
uαb ∈ Aαb = A ∩ K αb for each α
• If p is odd, then uαb ∈ Aαb , since Aαb = {a ∈ A | wαb (a) = a−1 } by Lemma 2.4(e).
• If p = 2, wαb ∈ S/A, and
α| ≤ 2, choose gαb ∈ S ∩ K αb such that wαb = gαb A. (For
Q |b
example, if we set g = α∈bα nα (1) (see Notation 2.2(B)), then g ∈ NG (T ) represents
the class wαb ∈ W0 , and is T -conjugate to an element of S ∩ K αb .) By Lemma 5.5(c),
CG (CA (wαb )) = G ∩ T K αb , where G ∩ K αb ∼
= SL2 (q) or SL2 (q 2 ) by Lemma 5.4. Hence
foc(CF (CA (wαb ))) = foc(CG (CA (wαb ))) = S ∩ [G ∩ T K αb , G ∩ T K αb ] = S ∩ K αb
(see the remarks before Lemma 5.5), and gαb lies in this subgroup. Since ϕ is fusion
preserving, ϕ(gαb ) ∈ foc(CF (CA (wαb ))). By Lemma 5.5(c) again,
uαb = ϕ(gαb ) · gα−1
b ))) = Aα
b.
b ∈ A ∩ foc(CF (CA (wα
• If p = 2, wαb ∈ S/A, and α
b = {α, τ (α), α∗} where α∗ = α + τ (α), then wαb = wα∗ . Choose
gαb ∈ S ∩ K α∗ such that gαb A = wαb ∈ N/A. (For example, there is such a gαb which is T conjugate to nα∗ (1).) By Lemma 5.5(c), CG (CA (wαb )) = G∩T K α∗ , G∩K α∗ ∼
= SL2 (q), and
hence gαb ∈ foc(CF (CA (wαb ))). So ϕ(gαb ) ∈ foc(CF (CA (wαb ))) since ϕ|S is fusion preserving.
By Lemma 5.5(c),
uαb = ϕ(gαb ) · gα−1
b ))) = A ∩ K α∗ ≤ Aα
b.
b ∈ A ∩ foc(CF (CA (wα
• If p = 2 and wαb ∈
/ S/A ∈ Syl2 (W0 ), then it is W0 -conjugate to some other reflection
b
b + ), c(w b) ∈ A b by the above argument, and hence uαb = c(wαb ) ∈ Aαb .
wβb ∈ S/A (for β ∈ Σ
β
β
Consider the homomorphism
Y
Φ = (Φα )α∈Π : T −−−−−−→
Tα
α∈Π
where
Φα (t) = t−1 wα (t) ∀ t ∈ T , α ∈ Π.
Since W = hwα | α ∈ Πi, we have Ker(Φ) = CT (W ) = Z(G) is finite (Proposition 2.5). Thus
r
Φ is (isomorphic to) a homomorphism from (F×
q0 ) to itself with finite kernel (where r = |Π|),
and any such homomorphism is surjective since F×
q0 has no subgroups of finite index.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
57
Choose elements vα ∈ T α for α ∈ Π as follows.
• If α
b = {α} where τ (α) = α, we set vα = uαb .
• If α
b = {α, τ (α)}, where α ⊥ τ (α), then T αb = T α × T τ (α) , and we let vα , vτ (α) be such that
vα vτ (α) = uαb .
• If α
b = {α, τ (α), α∗} where α∗ = α + τ (α), then uαb = hα (λ)hτ (α) (λ′ ) for some λ, λ′ ∈ F×
q0 ,
wαb (hα (λ)hτ (α) (λ′ )) = hα (λ′−1 )hτ (α) (λ−1 )
by Lemma 2.4(e), and λ = λ′ since wαb (uαb ) = u−1
α
b . Set vα = hα (λ) and vτ (α) = 1. (This
depends on the choice of α ∈ α
b ∩ Π.)
b This
b ∈ Π.
Let t ∈ T be such that Φ(t) = (vα )α∈Π . We claim that t−1 wαb (t) = uαb for each α
∗
is clear when |b
α| ≤ 2. If α
b = {α, τ (α), α } and λ are as above, then
wαb (t) = wα∗ (t) = wτ (α) wα wτ (α) (t) = wτ (α) (wα (t)) = wτ (α) t · hα (λ)
= t · wτ (α) (hα (λ)) = t · hα∗ (λ) = t · uαb .
b Since W0 = hwαb | α
b (and since c and dt are
b ∈ Π.
b ∈ Πi
Thus c(wαb ) = dt(wαb ) for each α
both cocycles), this implies that c = dt, and hence that [c] = 0 in H 1 (W0 ; T ).
As one consequence of Lemma 5.8, the Z ∗ -theorem holds for these groups. This is known
to hold for all finite groups (see [GLS3, § 7.8]), but its proof for odd p depends on the
classification of finite simple groups, which we prefer not to assume here.
Corollary 5.9. Assume that G ∈ Lie(q0 ), p 6= q0 , and S ∈ Sylp (G) satisfy Hypotheses 5.1.
Then Z(FS (G)) = Op (Z(G)).
Proof. By Lemma 5.8, Op (Z(G)) = CA (W0 ), where CS (A) = A and AutG (A) = AutW0 (A)
by Lemma 5.3(a,b) or by hypothesis (in Case 5.1(III.3)). Hence Z(FS (G)) ≤ Op (Z(G)),
while the other inclusion is clear.
We now need the following additional hypotheses, in order to be able to compare Autsc (A)
with the group of field automorphisms of G.
Hypotheses 5.10. Fix a prime p and a prime power q. Assume that q = q0b where q0 is
prime, b ≥ 1, q0 6= p, and
(i)
(ii)
q0 ≡ ±3 (mod 8) if p = 2;
the class of q0 generates (Z/p2 )× if p is odd; and
(iii) b|(p − 1)pℓ for some ℓ ≥ 0.
We will also say that “G satisfies Hypotheses 5.10” (for a given prime p) if G ∼
= t G(q) for
some t and G, and some q which satisfies the above conditions.
By Hypothesis 5.1(I), ψq0 (G) = G, and thus all field endomorphisms of G normalize G.
When G has a standard σ-setup, ΦG was defined to be the group of restrictions of such
endomorphisms ψq0a ∈ ΦG for a ≥ 0. Under our Hypotheses 5.1, this applies only when we
are in case (III.1) (although Proposition 3.6 describes the relation between ΦG and ψq0 in
the other cases). In what follows, it will be useful to set
b G = hψq0 |G i ≤ Aut(G).
Φ
58
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
b G = Inndiag(G)ΦG , although Φ
b G can be strictly larger
By Proposition 3.6(d), Inndiag(G)Φ
b G ∩ Inndiag(G) need not be trivial). Note that since each element of this group
than ΦG (Φ
b G normalizes T and each of its subgroups.
acts on T via (t 7→ tr ) for some r, Φ
Recall that τ ∈ Aut(V ) is the automorphism induced by σ, and also denotes the induced
permutation of Σ.
Lemma 5.11. Assume Hypotheses 5.1 and 5.10 and Notation 5.2. Let
b G −−−−−−−→ Aut(A, F )
χ0 : Φ
be the homomorphism induced by restriction from G to A. Set m = |τ | = γ|T . Then the
following hold.
(a) Either T has exponent q m − 1; or p is odd, m = ordp (q), m is even, and (q m/2 +
1) expt(T ) (q m − 1).
b G ) = Autsc (A). If p = 2, then χ0 (Φ
b G ) has index 2 in Autsc (A),
(b) If p is odd, then χ0 (Φ
A
and Autsc (A) = Im(χ0 )hψ−1 i.
(c) If p = 2, then χ0 is injective. If p is odd, then
(
hψq |G i = hγ|G i
in case (III.1)
Ker(χ0 ) =
m
m
b
h(ψq |G ) i = hγ |G i = ΦG ∩ AutT (G) in cases (III.2) and (III.3).
Proof. We first recall some of the assumptions in cases (III.1–3) of Hypotheses 5.1:
case (III.1)
ordp (q) = 1, m = |γ|, and m ≤ 2
case (III.2)
ordp (q) = m = 2
p is odd
case (III.3)
ordp (q) = m
p is odd
(15)
(Recall that γ is a graph automorphsm in case (III.1), so |γ| = |τ | = m.) In all of these
cases, p|(q m − 1) since ordp (q)|m.
m
(a) For each t ∈ T = CT (ψq ◦ γ), tq = ψq (t) = γ −1 (t). Hence t = γ −m (t) = (ψq )m (t) = tq ,
m
and tq −1 = 1. Thus expt(T ) (q m − 1).
By Hypotheses 5.1(I), there is a linearly independent subset Ω = {α1 , . . . , αs } ⊆ Σ such
that either Ω or ±Ω = {±α1 , . . . , ±αs } is a free hτ i-orbit in Σ. Assume Ω is a free orbit
(this always happens in case (III.1)). In particular, m = |τ | = s. For each 1 6= λ ∈ F×
q0 such
q s −1
that λ
= 1, the element
m−1
Y
i
t(λ) =
hτ i (α1 ) (λq )
i=0
is fixed by σ = ψq ◦ γ (recall σ(hβ (λ)) = hτ (β) (λq ) for each β ∈ Σ by Lemma 3.2). Hence
t(λ) ∈ T , and t(λ) 6= 1 when λ 6= 1 by Lemma 2.4(d,b). Thus T contains the subgroup
m
{t(λ) | λq −1 = 1} of order q m − 1, this subgroup is cyclic (isomorphic to a subgroup of F×
q0 ),
m
and hence expt(T ) = q − 1.
Assume now that ±Ω is a free hτ i-orbit (thus m = |τ | = 2s). In particular, we are not
in case (III.1), so p is odd and m = ordp (q). Then τ i (α1 ) = −α1 for some 0 < i < 2s, and
q s +1
= 1,
i = s since τ 2i (α1 ) = α1 . For each 1 6= λ ∈ F×
q0 such that λ
t(λ) =
s−1
Y
i=0
i
hτ i (α1 ) (λq )
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
59
s
is fixed by σ = ψq ◦ γ by Lemma 3.2 and since hτ s (α1 ) (λq ) = h−α1 (λ−1 ) = hα1 (λ). Hence
e
t(λ) ∈ T , and t(λ) 6= 1 when λ 6= 1 by Lemma 2.4 again. Thus {t(λ) | λq +1 = 1} ≤ T is
cyclic of order q s + 1, and so (q s + 1) expt(T ).
b G ) is generated by χ0 (ψq0 ) = ψq0 |A , which acts on A via
(b) By definition, Im(χ0 ) = χ0 (Φ
(a 7→ aq0 ). If p is odd, then by Hypotheses 5.10(ii), the class of q0 generates (Z/p2 )× , and
hence generates (Z/pk )× for each k > 0. So Im(χ0 ) = Autsc (A) in this case.
If p = 2, then q0 ≡ ±3 (mod 8) by Hypotheses 5.10(i). So for each k ≥ 2, hq0 i has index
A
2 in (Z/2k )× = hq0 , −1i. Hence Im(χ0 ) = hψq0 |A i has index 2 in Autsc (A) = hψq0 |A , ψ−1
i.
b G . Then (φ0 )b = ψq |G = (γ|G )−1 since G = C (ψq ◦ γ),
(c) Set φ0 = ψq0 |G , a generator of Φ
G
and so φ0 |T divides b|γ|T | = bm. Also, (φ0 )bm = (γ|G )−m ∈ AutT (G) by Lemma 3.2.
By (a), either expt(T ) = q m − 1; or m is even, p is odd, ordp (q) = m, and (q m/2 +
1) expt(T ) (q m − 1). In the latter case, vp (q m/2 + 1) = vp (q m − 1) > 0 since p ∤ (q m/2 − 1).
Thus
expt(A) = pe
where
e = vp (q m − 1) = vp (q0 bm − 1) > 0 .
(16)
If p = 2, then we are in case (III.1). In particular, q = q0b ≡ 1 (mod 4), and m ≤ 2. Also, b
(and hence bm) is a power of 2 by Hypotheses 5.10(iii). If bm = 1, then q = q0 ≡ 5 (mod 8),
so e = v2 (q−1) = 2. If bm is even, then e = v2 (q0bm −1) = v2 (q02 −1)+v2(bm/2) = 3+v2 (bm/2)
by Lemma 1.13. Thus in all cases, e = 2 + v2 (bm). So Im(χ0 ) ≤ Autsc (A) ∼
= (Z/2e )× has
order 2e−2 = bm. Since (ψq0 |G )bm = (ψq |G )m = (γ −1 |G )m = IdG (recall m = |γ| in case
(III.1)), χ0 is injective.
Now assume p is odd, and set m0 = ordp (q). Then b|(p−1)pℓ for some ℓ ≥ 0 by Hypotheses
5.10(iii), and q = q0b where the class of q0 generates (Z/pk )× for each k ≥ 1. For r ∈ Z,
q r = q0br ≡ 1 (mod p) if and only if (p − 1)|br. Hence bm0 = b · ordp (q) = (p − 1)pℓ for some
ℓ ≥ 0. Since vp (q0p−1 − 1) = 1, and since m = m0 or 2m0 , Lemma 1.13 implies that
e = vp (q m − 1) = vp (q0bm − 1) = vp (q0bm0 − 1) = 1 + vp (pℓ ) = 1 + ℓ .
Thus ℓ = e − 1, where pe = expt(A) by (16), so |Autsc (A)| = (p − 1)pe−1 = bm0 .
b G to the generator χ0 (φ0 ) of Autsc (A), this proves
Since χ0 sends the generator φ0 of Φ
m0
m0
that Ker(χ0 ) = hψq |G i = hγ |G i. The descriptions in the different cases now follow
−m
immediately. Note that in cases (III.2) and (III.3) (where m = m0 ), φbm
|G ∈ AutT (G)
0 = γ
b
by Lemma 3.2. The converse is immediate: ΦG ∩ AutT (G) ≤ Ker(χ0 ).
Before applying these results to describe Out(S, F ) and the homomorphism κG , we need
to know in which cases the subgroup A is characteristic in S.
Proposition 5.12. Assume Hypotheses 5.1 and Notation 5.2.
(a) If p = 2, then A is characteristic in S, and is the unique abelian subgroup of S of order
|A|, except when q ≡ 5 (mod 8) and G ∼
= Sp2n (q) for some n ≥ 1.
(b) If p is odd, then A is characteristic in S, and Ω1 (A) is the unique elementary abelian
subgroup of S of maximal rank, except when p = 3, q ≡ 1 (mod 3), v3 (q − 1) = 1, and
G∼
= SU3 (q) or G2 (q).
In all cases, each normal subgroup of S isomorphic to A is NG (S)-conjugate to A.
Proof. If p is odd, then by [GL, 10-2(1,2)], there is a unique elementary p-subgroup E ≤ S
of rank equal to that of A (denoted rm0 in [GL]), except when p = 3 and G is isomorphic
to one of the groups SL3 (q) (q ≡ 1 (mod 3)), SU3 (q) (q ≡ −1 (mod 3)), or G2 (q), 3D4 (q),
60
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
or 2F4 (q) (q ≡ ±1 (mod 3)). When there is a unique such subgroup E, then A = CS (E) by
Lemma 5.3(a) (or by assumption in case (III.3)), and hence A is characteristic in S.
Among the exceptions, SL3 (q) and G2 (q) are the only ones which satisfy Hypotheses 5.1.
In both cases, S is an extension of A ∼
= (C3ℓ )2 by C3 , where ℓ = v3 (q − 1), and where
Z(S) = CA (S) has order 3. If ℓ > 1, then A is the unique abelian subgroup of index p in
S. If ℓ = 1, then S is extraspecial of order 33 and exponent 3. By Theorem 1.8(a), we can
assume q = 4 without changing the isomorphism type of the fusion system, so G contains
SU3 (2). This is a semidirect product S ⋊ Q8 (cf. [Ta, p. 123–124]), and hence the four
subgroups of S of order 9 are NG (S)-conjugate.
It remains to prove the proposition when p = 2. We use [O4, § 2] as a reference for
information about best offenders, since this contains what we need in a brief presentation.
Assume A is not the unique abelian subgroup of S of order |A|. Then there is an abelian
subgroup 1 6= B ≤ W0 such that |B| · |CA (B)| ≥ |A|. In other words, the action of the Weyl
group W0 on A has a nontrivial best offender [O4, Definition 2.1(b)]. Hence by Timmesfeld’s
replacement theorem [O4, Theorem 2.5], there is a quadratic best offender 1 6= B ≤ W0 : an
offender such that [B, [B, A]] = 1.
We consider three different cases.
Case 1: G ∼
= G(q) is a Chevalley group, where either q ≡ 1 (mod 8), or G 6∼
=
Sp2n (q) for any n ≥ 1.
Set n = rk(A) = rk(T ): the Lie rank of G (or of G). Set
ℓ = v2 (q − 1) ≥ 2. Then A ∼
= (C2ℓ )n is the group of all 2ℓ -torsion elements in T (or in T ).
Since the result is clear when n = 1 (G ∼
= Q2ℓ+1 ), we
= C2ℓ , and S ∼
= Sp2 (q), A ∼
= SL2 (q) ∼
assume n ≥ 2.
Let Λ = ZΣ∨ be the lattice in V generated by the dual roots. By Lemma 2.6(a), there
are Z[W ]-linear isomorphisms A ∼
= Λ/2Λ.
= Λ/2ℓ Λ and Ω1 (A) ∼
Assume first that B acts faithfully on Ω1 (A). Since B has quadratic action, it is elementary
abelian [O4, Lemma 2.4]. Set k = rk(B); thus B ∼
= C2k and |A/CA (B)| ≤ 2k .
Since the B-action on V is faithful, the characters χ ∈ Hom(B, {±1}) which have nontrivial eigenspace on V generate the dual group B ∗ . So we can choose a basis χ1 , . . . , χk for
B ∗ such that each χi has nontrivial eigenspace. Let b ∈ B be the unique element such that
χi (b) = −1 for each i = 1, . . . , k. Let V+ , V− be the ±1-eigenspaces for the b-action on V ,
and set Λ± = Λ ∩ V± . By construction, dim(V− ) ≥ k.
Let v ∈ Λ be an element whose class modulo 2ℓ Λ is fixed by b, and write v = v+ + v−
where v± ∈ V± . Then 2v− = v − b(v) ∈ 2ℓ Λ ∩ V− = 2ℓ Λ− , so v− ∈ 2ℓ−1 Λ− and v+ = v − v− ∈
Λ ∩ V+ = Λ+ . Thus CΛ/2ℓ Λ (b) = (Λ+ × 2ℓ−1 Λ− )/2ℓ Λ. Set r = rk(Λ− ) = dim(V− ) ≥ k; then
2k ≥ |A/CA (B)| ≥ |A/CA (b)| = |Λ/(Λ+ × 2ℓ−1 Λ− )| = 2r(ℓ−1) · |Λ/(Λ+ × Λ− )|
≥ 2k(ℓ−1) · |Λ/(Λ+ × Λ− )|.
In particular, Λ = Λ+ × Λ− . But then b acts trivially on Λ/2Λ, hence on Ω1 (A), which
contradicts our assumption.
Thus B does not act faithfully on Ω1 (A). Set B0 = CB (Ω1 (A)) ∼
= CB (Λ/2Λ) 6= 1. If
−IdV ∈ B0 , then it inverts A, [B, Ω1 (A)] ≤ [B, [B0 , A]] = 1 since B acts quadratically, so
B = B0 , and |B0 | ≥ |A/CA (B)| ≥ |A/Ω1 (A)| = 2(ℓ−1)n . If b ∈ B0 is such that b2 = −IdV ,
then b defines a C-vector space structure on V , and hence does not induce the identity on
Λ/2Λ, a contradiction.
Thus there is b ∈ B0 which does not act on V via ±Id. Let V± 6= 0 be the ±1-eigenspaces
for the b-action on V , and set Λ± = Λ∩V± . For each v ∈ Λ, v−b(v) ∈ 2Λ since b acts trivially
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
61
on Ω1 (A) ∼
= Λ/2Λ. Set v = v+ + v− , where v± ∈ V± . Then 2v− = v − b(v) ∈ 2Λ ∩ V− = 2Λ−
implies that v− ∈ Λ− , and hence v+ ∈ Λ+ . Thus V ∈ Λ+ × Λ− , so by Lemma 2.8, G = Cn .
By assumption, q ≡ 1 (mod 8), so ℓ ≥ 3, and [b, [b, Λ/2ℓ Λ]] ≥ 4Λ− /2ℓ Λ− 6= 1, contradicting
the assumption that B acts quadratically on A.
Case 2: G ∼
= Sp2n (q) for some n ≥ 1 and some q ≡ 5 (mod 8). Fix subgroups
Hi ≤ G (1 ≤ i ≤ n) and K < G such that Hi ∼
= Σn is the group of
= Sp2 (q) for each i, K ∼
permutation matrices (in 2 × 2 blocks), and K normalizes H = H1 × · · · × Hn and permutes
∼
=
the factors in the obvious way. We can also fix isomorphisms χi : Hi −−−→ Sp2 (q) such that
the action of K on the Hi commutes with the χi .
b ∼
b < Q
b < Sp2 (q), where Q
b ∼
Fix subgroups A
= C4
= Q8 (a Sylow 2-subgroup), and A
−1 b
−1 b
is contained in the maximal torus. Set Qi = χi (Q) and Ai = χi (A), and set Q =
Q1 Q2 · · · Qn and A = A1 A2 · · · An . Thus A = O2 (T ) is as in Hypotheses 5.1(III): the 2power torsion in the maximal torus of G. By [CF, § I], S = QR for some R ∈ Syl2 (K). Also,
W ∼
= C2 ≀ Σn acts on A via signed permutations of the coordinates.
= QK/A ∼
Let B be any nontrivial best offender in W on A. Consider the action of B on the set
{1, 2, . . . , n}, let X1 , . . . , Xk be the set of orbits, and set di = |Xi |. For 1 ≤ i ≤ k, let Ai ≤ A
be the subgroup of elements whose coordinates vanish except for those in
Qkpositions in Xi ;
di
∼
thus Ai = (C4 ) and A = A1 × · · · × Ak . Set Bi = B/CB (Ai ); then |B| ≤ i=1 |Bi |. Since B
is abelian, either |Bi | = di and Bi permutes the coordinates freely, or |Bi | = 2di and there is
a unique involution in Bi which inverts all coordinates in Ai . In the first case, |CAi (Bi )| = 4,
and so |Bi | · |CAi (Bi )| = di · 4 ≤ 4di = |Ai | with equality only if di = 1. In the second case,
|CAi (Bi )| = 2, and again |Bi | · |CAi (Bi )| = 2di · 2 ≤ 4di = |Ai | with equality only if di = 1.
Since
k
Y
i=1
|Ai | = |A| ≤ |B| · |CA (B)| = |B| ·
k
Y
i=1
|CAi (Bi )| ≤
k
Y
i=1
|Bi | · |CAi (Bi )| ,
we conclude that di = 1 for all i, and hence that B acts only by changing signs in certain
coordinates.
For each 1 ≤ i ≤ n, let pri : Q −−−→ Qi be the projection onto the i-th factor. If A∗ ≤ S
is abelian of order 4n , then A∗ A/A is a best offender in W on A, and hence A∗ ≤ Q by the
last paragraph. Also, pri (A∗ ) is cyclic
at most 4 for each i, and since |A∗ | = 4n ,
Qn of order
∗ ∼
∗
∗
pri (A ) = C4 for each i and A = i=1 pri (A ). Thus there are exactly 3n such subgroups.
Now assume A∗ E S, and set A∗i = pri (A∗ ) ≤ Qi for short. Since A∗ is normal, the
b < Sp2 (q) are equal for all i lying in any R-orbit of the set {1, 2, . . . , n}.
subgroups χi (A∗i ) ≤ Q
Hence we can choose elements x1 , x2 , . . . , xn , where xi ∈ NHi (Qi ) ∼
= SL2 (3) and xi(Ai ) = A∗i
for each i, and such that χi (xi ) ∈ Sp2 (q) is constant on each R-orbit. Set x = x1 x2 · · · xn ;
then xA = A∗ , and x ∈ NG (S).
Case 3: G is a Steinberg group. Assume γ ∈ ΓG is a graph automorphism of order 2,
and that G = CG (σ) where σ = γψq . Set G0 = CG (γ, ψq ); thus G0 ≤ G. Set ℓ = v2 (q−1) ≥ 2.
We must again show that the action of W0 on A has no nontrivial best offenders.
∼
If G ∼
= 2E6 (q) or Spin−
2n (q) (n ≥ 4), then G0 = F4 (q) or Spin2n−1 (q), respectively, and W0
is the Weyl group of G0 . If 1 6= B ≤ W0 is a best offender in W0 on A, then it is also a best
offender on Ωℓ (A) ≤ G0 , which is impossible by Case 1.
If G ∼
= (SD2ℓ+2 )n ⋊ R for some R ∈ Syl2 (Σn ) [CF, pp. 143–
= 2A2n (q), then S ∼
= SU2n+1 (q) ∼
n
144]. Thus A ∼
= (C2ℓ+1 ) , W0 ∼
= C2 ≀ Σn , Σn < W0 acts on A by permuting the coordinates,
n
∼
and the subgroup W1 = (C2 ) in W0 has a basis each element of which acts on one coordinate
62
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
ℓ
by (a 7→ a2 −1 ). If B ≤ W0 is a nontrivial quadratic best offender on A, then it is also a best
offender on Ωℓ (A) [O4, Lemma 2.2(a)], hence is contained in W1 by the argument in Case 2,
which is impossible since no nontrivial element in this subgroup acts quadratically. Thus A
is characteristic in this case.
It remains to consider the case where G ∼
=
= 2A2n−1 (q). Since the case SU2 (q) ∼
= SU2n (q) ∼
b = GU2n (q) > G, set
Sp2 (q) has already been handled, we can assume n ≥ 2. Set G
b and set G1 = N b (G0 ) ∼
G0 = GU2 (q) × · · · × GU2 (q) ≤ G,
= GU2 (q) ≀ Σn . Then G1 has
G
b [CF, pp. 143–144], so we can assume S ≤ G1 ∩ G. Fix H0 ∈ Syl2 (G0 );
odd index in G
∼
thus H0 = (SD2ℓ+2 )n . Since v2 (q + 1) = 1, and since the Sylow 2-subgroups of SU2 (q) are
quaternion,
χn
sum
G ∩ H0 = Ker H0 ∼
= (SD2ℓ+2 )n −−−−→ C2n −−−−→ C2 ,
where χ : SD2ℓ+2 −−−→ C2 is the surjection with quaternion kernel. As in the last case,
W0 ∼
= C2 ≀ Σn with normal subgroup W1 ∼
= C2n . If B ≤ W0 is a nontrivial quadratic best
offender on A, then it is also a best offender on Ωℓ (A) [O4, Lemma 2.2(a)], so B ≤ W1 by
the argument used in Case 2. Since no nontrivial element in W1 acts quadratically on A, we
conclude that A is characteristic in this case.
The next lemma is needed to deal with the fact that not all fusion preserving automorphisms of A lie in Aut(A, F ) (since they need not extend to automorphisms of S).
Lemma 5.13. Let G be any finite group, fix S ∈ Sylp (G), and let S0 E S be a normal
subgroup. Let ϕ ∈ Aut(G) be such that ϕ(S0 ) = S0 and ϕ|S0 ∈ NAut(S0 ) (AutS (S0 )). Then
there is ϕ′ ∈ Aut(G) such that ϕ′ |S0 = ϕ|S0 , ϕ′ (S) = S, and ϕ′ ≡ ϕ (mod Inn(G)).
Proof. Since ϕ|S0 normalizes AutS (S0 ), and cϕ(g) = ϕcg ϕ−1 for each g ∈ G, we have
Autϕ(S) (S0 ) = ϕAutS (S0 ) = AutS (S0 ). Hence ϕ(S) ≤ CG (S0 )S. Since S normalizes CG (S0 )
and S ∈ Sylp (CG (S0 )S), we have ϕ(S) = xS for some x ∈ CG (S0 ). Set ϕ′ = c−1
x ◦ ϕ ∈ Aut(G);
′
′
then ϕ (S) = S and ϕ |S0 = ϕ|S0 .
In the next two propositions, we will be referring to the short exact sequence
R
1 −−−→ Autdiag (S, F ) −−−−−−→ NAut(S,F ) (A) −−−−−−→ Aut(A, F ) −−−→ 1 .
(17)
Here, R is induced by restriction, and Aut(A, F ) = Im(R) and Autdiag (S, F ) = Ker(R) by
definition of these two groups (Notation 5.2(H)). By Proposition 5.12, in all cases, each class
in Out(S, F ) is representated by elements of NAut(S,F ) (A).
Proposition 5.14. Assume Hypotheses 5.1 and 5.10 and Notation 5.2. Then κG is surjective, except in the following cases:
• (G, p) ∼
= (2E6 (q), 3), or
• (G, p) ∼
= (G2 (q), 2) and q0 6= 3, or
• (G, p) ∼
= (F4 (q), 3) and q0 6= 2.
In the exceptional cases, |Coker(κG )| ≤ 2.
Proof. We first claim that for ϕ ∈ Aut(S, F ),
ϕ(A) = A and ϕ|A ∈ Autsc (A)AutAut(G) (A)
=⇒
[ϕ] ∈ Im(κG ) .
(18)
A
To see this, fix such a ϕ. By Lemma 5.11(b), each element of Autsc (A), or of Autsc (A)/hψ−1
i
b G . If p = 2, then we are in case (III.1), the
if p = 2, is the restriction of an element of Φ
A
σ-setup is standard, and hence the inversion automorphism ψ−1
is the restriction of an inner
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
63
automorphism of G (if −IdV ∈ W ) or an element of Inn(G)ΓG . Thus ϕ|A extends to an
automorphism of G.
Now, ϕ|A normalizes AutS (A) since ϕ(S) = S. So by Lemma 5.13, ϕ|A is the restriction
of an automorphism of G which normalizes S, and hence is the restriction of an element
ψ ∈ Aut(S, F ) such that [ψ] ∈ Im(κG ). Then ϕψ −1 ∈ Ker(R) = Autdiag (S, F ) by the
exactness of (17), and [ϕψ −1 ] ∈ Im(κG ) by Lemma 5.8. So [ϕ] ∈ Im(κG ), which proves (18).
By Proposition 5.12, each class in Out(S, F ) is represented by an element of NAut(S,F ) (A).
Hence by (18), |Coker(κG )| is at most the index of Aut(A, F ) ∩ Autsc (A)AutAut(G) (A) in
Aut(A, F ). So by Lemma 5.7, |Coker(κG )| ≤ 2, and κG is surjective with the exceptions
listed above.
We now want to refine Proposition 5.14, and finish the proof of Theorem B, by determining
Ker(κG ) in each case where 5.1 and 5.10 hold and checking whether it is split. In particular,
we still want to show that each of these fusion systems is tamely realized by some finite
group of Lie type (and not just an extension of such a group by outer automorphisms).
Since Op′ (Outdiag(G)) ≤ Ker(κG ) in all cases by Lemma 5.8, κG induces a quotient
homomorphism
◦
κG : Out(G)/Op′ (Outdiag(G)) −−−−−−→ Out(S, F ) ,
◦
and it is simpler to describe Ker(κG ) than Ker(κG ). The projection of Out(G) onto the
quotient Out(G)/Op′ (Outdiag(G)) is split: by Steinberg’s theorem (Theorem 3.4), it splits
◦
back to Op (Outdiag(G))ΦG ΓG . Hence κG is split surjective if and only if κG is split surjective.
Proposition 5.15. Assume Hypotheses 5.1 and 5.10 and Notation 5.2. Assume also that
none of the following hold: neither
• (G, p) ∼
= (2E6 (q), 3), nor
• (G, p) ∼
= (G2 (q), 2) and q0 6= 3, nor
• (G, p) ∼
= (F4 (q), 3) and q0 6= 2.
◦
(a) If p = 2, then κG is an isomorphism, and κG is split surjective.
(b) Assume that p is odd, and that we are
√
Then q ∈ N, and
∼
[ψ√q ] = C2
◦
Ker(κG ) =
[γ0 ψ√q ] ∼
= C2
[ψ√ ] ∼ C
= 4
q
in the situation of case (III.1) of Hypotheses 5.1.
if γ = Id and −Id ∈ W
if γ = Id and −Id ∈
/W
if γ 6= Id (G is a Steinberg group)
where in the second case, γ0 ∈ ΓG is a graph automorphism of order 2. Hence κG and
◦
κG are split surjective if and only if either γ = Id and −Id ∈
/ W , or p ≡ 3 (mod 4).
(c) Assume that p is odd, and that we are in the situation of case (III.2) or (III.3) of
Hypotheses 5.1. Assume also that G is a Chevalley group (γ ∈ Inn(G)), and that
ordp (q) is even or −Id ∈
/ W0 . Let ΦG , ΓG ≤ Aut(G) be as in Proposition 3.6. Then
◦
◦
◦
ΦG ∩ Ker(κG ) = 1, so |Ker(κG )| ≤ |ΓG |, and κG and κG are split surjective.
(d) Assume that p is odd, and that we are in the situation of case (III.3) of Hypotheses 5.1.
Assume also that G is a Steinberg group (γ ∈
/ Inn(G)), and that ordp (q) is even. Then
(
[γ|G ] ∼
◦
= C2 if γ|A ∈ AutW0 (A)
Ker(κG ) =
1
otherwise.
64
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
◦
Hence κG and κG are split surjective if and only if q is an odd power of q0 or Ker(κG ) =
Op′ (Outdiag(G)). If κG is not split surjective, then its kernel contains a graph automorphism of order 2 in Out(G)/Outdiag(G).
Proof. In all cases, κG is surjective by Proposition 5.14 (with the three exceptions listed
above).
By definition and Proposition 5.12,
Out(S, F ) = Aut(S, F )/AutF (S) ∼
= NAut(S,F ) (A) NAutF (S) (A) .
Also, Outdiag (S, F ) is the image in Out(S, F ) of Autdiag (S, F ). Since NAutF (S) (A) is the
group of automorphisms of S induced by conjugation by elements in NG (S) ∩ NG (A), the
short exact sequence (17) induces a quotient exact sequence
R
1 −−−→ Outdiag (S, F ) −−−−−−→ Out(S, F ) −−−−−−→ Aut(A, F ) AutNG (S) (A) −−−→ 1 . (19)
We claim that
AutNG (S) (A) = Aut(A, F ) ∩ AutG (A) .
(20)
That AutNG (S) (A) is contained in the two other groups is clear. Conversely, assume α ∈
Aut(A, F ) ∩ AutG (A). Then α = cg |A for some g ∈ NG (A), and α ∈ NAut(A) (AutS (A))
since it is the restriction of an element of Aut(S, F ). Hence g normalizes SCG (A), and since
S ∈ Sylp (SCG (A)), there is h ∈ CG (A) such that hg ∈ NG (S). Thus α = cg |A = chg |A ∈
AutNG (S) (A), and this finishes the proof of (20).
By Lemma 5.8, κG sends Outdiag(G) onto Outdiag (S, F ) with kernel Op′ (Outdiag(G)).
Hence by the exactness of (19), restriction to A induces an isomorphism
◦
R0
Ker(κG ) −−−
−
→
Ker
Out(G)/Outdiag(G)
−
−
−
−
→
Aut(A,
F
)/Aut
(A)
N
(S)
G
∼
=
= Ker Out(G)/Outdiag(G) −−−−→ NAut(A) (AutG (A))/AutG (A) , (21)
where the equality holds by (20).
Recall that for each ℓ prime to p, ψℓA ∈ Autsc (A) denotes the automorphism (a 7→ aℓ ).
(a,b) Under either assumption (a) or (b), we are in case (III.1) of Hypotheses 5.1. In
particular, (G, σ) is a standard σ-setup for G. Set k = vp (q − 1); then k ≥ 1, and k ≥ 2 if
p = 2.
If p is odd, then by Hypotheses 5.10(b), the class of q0 generates (Z/p)× . Since q = q0b ≡ 1
√
b/2
(mod p), this implies that (p − 1)|b. In particular, b is even, and q = q0 ∈ N.
Since Out(G)/Outdiag(G) ∼
= ΦG ΓG by Theorem 3.4, where ΦG ΓG normalizes T and hence
A, (21) takes the form
◦
(22)
Ker(κG ) ∼
= ϕ ∈ ΦG ΓG ϕ|A ∈ AutW0 (A) .
◦
In fact, when Ker(κG ) has order prime to p (which is the case for all examples considered
here), the isomorphism in (22) is an equality since Outdiag(G)/Op′ (Outdiag(G)) is a p-group.
Assume first that G = G(q) is a Chevalley group. Thus σ = ψq where q ≡ 1 (mod
k
p), and A = t ∈ T tp = 1 . By Lemma 2.7 (applied with m = pk ≥ 3), the group
A
AutW (T )AutΓG (T ) acts faithfully on A, and its action intersects Autsc (A) only in hψ−1
i. By
b G isomorphically onto Autsc (A) if p is odd, and
Lemma 5.11(b,c), restriction to A sends Φ
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
65
with index 2 if p = 2. So when G is not one of the groups B2 , F4 , or G2 , then ΦG ΓG acts
faithfully on A, and
ϕ ∈ ΦG ΓG
if p = 2
1
√
ϕ|A ∈ AutW0 (A) = hψ q i
if p is odd and −Id ∈ W
hγ ψ√ i if p is odd and −Id ∈
/W
q
0
where in the last case, γ0 ∈ ΓG is a graph automorphism such that the coset γ0 W contains
−Id. (Note that b = (p − 1)pℓ for some ℓ ≥ 0 by Hypotheses 5.10(b,c) and since p|(q − 1).
√
A
Hence q ≡ −1 modulo pk = expt(A), and ψ√q |A = ψ−1
.)
◦
◦
Thus by (22), κG is injective if p = 2, and |Ker(κG )| = 2 if p is odd. When p is odd, since
◦
Ker(κG ) is normal of order prime to p in Out(G) (hence of order prime to |Op (Outdiag(G))|),
◦
Ker(κG ) is generated by [ψ√q ] if −Id ∈ W (i.e., if there is an inner automorphism which
◦
inverts T and hence A), or by [γ0 ψ√q ] otherwise for γ0 as above. In the latter case, κG is
split since it sends Op (Outdiag(G))ΦG O3 (ΓG ) isomorphically onto Out(S, F ) (recall ΓG ∼
= C2
◦
√
or Σ3 ). When Ker(κG ) = [ψ q ] , the map is split if and only if 4 ∤ |ΦG | = b, and since
b = (p − 1)pm for some m, this holds exactly when p ≡ 3 (mod 4).
If (G, p) ∼
= (B2 (q), 2), (F4 (q), 2), or (G2 (q), 3), then since q0 6= p, ΓG = 1. So similar
◦
◦
arguments show that Ker(κG ) = 1, 1, or [ψ√q ] ∼
= C2 , respectively, and that κG is split in
all cases.
Next assume G = G2 (q), where p = 2, q = 3b , and b is a power of 2. Then b ≥ 2 since
q ≡ 1 (mod 4). The above argument shows that ΦG injects into Out(S, F ). Since Out(G)
is cyclic of order 2b, generated by a graph automorphism whose square generates ΦG (and
since 2|b), Out(G) injects into Out(S, F ).
If G = F4 (q), where p = 3, q = 2b , and b = 2 · 3ℓ for some ℓ ≥ 0, then the same argument
shows that ΦG injects into Out(S, F ). Since Out(G) is cyclic of order 2b = 4 · 3ℓ , generated
by a graph automorphism whose square generates ΦG , Out(G) injects into Out(S, F ).
It remains to handle the Steinberg groups. Let H be such that CG (γ) = H(Fq0 ): a simple
algebraic group by [GLS3, Theorem 1.15.2(d)]. In particular, G ≥ H = H(q). Also, W0
is the Weyl group of H by [GLS3, Theorem 1.15.2(d)] (or by the proof of [St3, Theorem
8.2]). By Lemma 2.7 applied to H(Fq0 ), W0 acts faithfully on A ∩ H = Ωk (A), and intersects
A
Autsc (A) at most in hψ−1
i.
A
If p = 2, then by Lemma 5.11(b), ψ−1
is not the restriction of an element in ΦG . Also,
◦
ΦG ∼
= C2b is sent injectively into Autsc (A) by Lemma 5.11(c), so κG is injective by (22).
If p is odd, then ψq0 |A has order b in Autsc (A) by Lemma 5.11(c). Since (ψq0 )b/2 = ψ√q
√
where q ≡ −1 (mod p) (recall b|(p − 1)pℓ for some ℓ by Hypotheses 5.10(iii)), ψq0 |A has
order b/2 modulo AutW0 (A). So by (22) and the remark afterwards, and since ΦG is cyclic of
◦
◦
order 2b, Ker(κG ) = [ψ√q ] ∼
= C4 . In particular, κG is split only if b/2 is odd; equivalently,
p ≡ 3 (mod 4).
(c,d) In both of these cases, p is odd, ordp (q) is even or −Id ∈
/ W , and we are in the situation
of case (III.2) or (III.3) in Hypothesis 5.1. Then γ|G = (ψq |G )−1 since G ≤ CG (γψq ). Also,
ψq0 (G) = G by 5.1(I), and hence γ(G) = G. Since ψq0 and γ both normalize T by assumption
or by construction, they also normalize T = G ∩ T and A = Op (T ). By Proposition 3.6(d),
[ψq0 ] generates the image of ΦG in Out(G)/Outdiag(G).
66
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
We claim that in all cases,
AutG (A) = AutW0 (A)
and
AutG (A) ∩ Autsc (A) ≤ hγ|A i .
(23)
This holds by assumption in case (III.3), and since ordp (q) is even or −Id ∈
/ W0 . In case
(III.2), the first statement holds by Lemma 5.3(b), and the second by Lemma 2.7 (and since
W0 = W and A contains all pk -torsion in T ).
(c) Assume in addition that G is a Chevalley group. Thus γ ∈ Inn(G), so γ|G ∈
Inndiag(G) = Inn(G)AutT (G) by Proposition 3.6(b), and hence γ|A ∈ AutG (A). Also,
γ|A = (ψq |A )−1 = (ψq0 |A )−b since σ = γψq centralizes G ≥ A. Since ψq0 |A has order
b · ordp (q) in Autsc (A) by Lemma 5.11(c), its class in NAut(A) (AutG (A))/AutG (A) has order
b by (23).
◦
Thus by (21), κG sends Op (Outdiag(G))ΦG injectively into Out(S, F ). Since ΓG is iso◦
◦
morphic to 1, C2 , or Σ3 (and since κG is onto by Proposition 5.14), κG and κG are split.
(d) Assume G is a Steinberg group and ordp (q) is even. In this case, γ ∈
/ Inn(G), and
∼
Out(G)/Outdiag(G) = ΦG is cyclic of order 2b, generated by the class of ψq0 |G . Hence by
◦
(21), Ker(κG ) is isomorphic to the subgroup of those ψ ∈ ΦG such that ψ|A ∈ AutG (A). By
◦
(23) and since ψq |A = γ −1 |A , AutG (A) ∩ Autsc (A) ≤ hψqA i. Thus |Ker(κG )| ≤ 2, and
◦
◦
|Ker(κG )| = 2
◦
⇐⇒
γ|A ∈ AutG (A) = AutW0 (A) .
When Ker(κG ) 6= 1, κG is split if and only if 4 ∤ |ΦG | = 2b; i.e., when b is odd.
◦
In the situation of Proposition 5.15(c), if −Id ∈
/ W , then Ker(κG ) = [γ0 ψ√q ] where γ0
◦
is a nontrivial graph automorphism. If −Id ∈ W (hence ordp (q) is even), then κG is always
injective: either because ΓG = 1, or by the explicit descriptions in the next section of the
setups when ordp (q) = 2 (Lemma 6.4), or when ordp (q) > 2 and G = D2n (Lemma 6.5).
The following examples help to illustrate some of the complications in the statement of
Proposition 5.15.
4
4
4
Example 5.16. Set p = 5. If G = Spin−
4k (3 ), Sp2k (3 ), or SUk (3 ) (k ≥ 5), then by Proposition 5.15(b), κG is surjective but not split. (These groups satisfy case (III.1) of Hypotheses
5.1 by Lemma 6.1.) The fusion systems of the last two are tamely realized by Sp2ℓ (32 )
and SLn (32 ), respectively (these groups satisfy case (III.2) by Lemma 6.4, hence Proposition
−
4
2
5.15(c) applies). The fusion system of Spin−
4k (3 ) is also realized by Spin4k (3 ), but not tamely
(Example 6.6(b)). It is tamely realized by Spin4k−1 (32 ) (see Propositions 1.9(c) and 5.15(c)).
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
67
6. The cross characteristic case: II
In Section 5, we established certain conditions on a finite group G of Lie type in characteristic q0 , on a σ-setup for G, and on a prime p 6= q0 , and then proved that the p-fusion
system of G is tame whenever those conditions hold. It remains to prove that for each G
of Lie type and each p different from the characteristic, there is another group G∗ whose
p-fusion system is tame by the results of Section 5, and is isomorphic to that of G.
We first list the groups which satisfy case (III.1) of Hypotheses 5.1.
Lemma 6.1. Fix a prime p and a prime power q ≡ 1 (mod p), where q ≡ 1 (mod 4) if p = 2.
Assume G ∼
= 2 G(q)
= G(q) for some simple group scheme G over Z of universal type, or G ∼
for G ∼
= An , Dn , or E6 of universal type. Then G has a σ-setup (G, σ) such that Hypotheses
5.1, case (III.1) holds.
Proof. Set G = G(Fq ), and let ψq ∈ ΦG be the field automorphism. Set σ = γψq ∈ End(G),
where γ = Id if G ∼
= 2 G(q).
= G(q), and γ ∈ ΓG has order 2 if G ∼
NG (T ) contains a Sylow p-subgroup of
Q G. If γ = Id, then by [Ca, Theorem 9.4.10]
(and since G is in universal form), |G| = q N ri=1 (q di −1) for some integers N, d1 , . . . , dr (r =
rk(G)), where d1 d2 · · · dr = |W | by [Ca, Theorem 9.3.4]. Also, |T | = (q −1)r , NG (T )/T ∼
= W,
and so
r
r
X
X
di
vp (|G|) =
vp (q − 1) =
vp (q − 1) + vp (di ) = vp (|T |) + vp (|W |) = vp (NG (T )) ,
i=1
i=1
where the second equality holds by Lemma 1.13.
If |γ| = 2, then by [Ca, §§
for N and di Q
as above, there are εi , ηi ∈ {±1} for
Q14.2–3],
r
r
di
N
1 ≤ i ≤ r such that |G| = q
i=1 (q−ηi ). (More precisely, the ηi are
i=1 (q −εi ) and |T | =
the eigenvalues of the γ-action on V , and polynomial generators I1 , . . . , Ir ∈ R[x1 , . . . , xr ]W
can be chosen such that deg(Ii ) = di and τ (Ii ) = εi Ii .) By [Ca, Proposition 14.2.1],
|W0 | = lim
t→1
r
Y
1 − εi tdi
i=1
1 − ηi t
=⇒
{1 ≤ i ≤ r | εi = 1} = {1 ≤ i ≤ r | ηi = 1}
Q
and |W0 | = {di | εi = +1}.
Also, vp (q d + 1) = vp (q + 1) for all d ≥ 1: they are both 0 if p is odd, and both 1 if p = 2.
Hence
r
r
q di − 1
X
X
=
vp (di ) = vp (|W0 |) = vp (|NG (T )|) − vp (|T |)
vp (|G|) − vp (|T |) =
vp
q
−
1
i=1
i=1
εi =+1
εi =+1
by Lemma 1.13 again, and so NG (T ) contains a Sylow p-subgroup of G.
The free hγi-orbit {α} (if γ = Id) or {α, τ (α)} (if |γ| = 2 and α 6= τ (α)), for any α ∈ Σ,
satisfies the hypotheses of this condition.
[γ, ψq0 ] = Id since γ ∈ ΓG .
We are now ready to describe the reduction, when p = 2, to groups with σ-setups satisfying
Hypotheses 5.1.
Proposition 6.2. Assume G ∈ Lie(q0 ) is of universal type for some odd prime q0 . Fix
S ∈ Syl2 (G), and assume S is nonabelian. Then there is an odd prime q0∗ , a group G∗ ∈
Lie(q0∗ ) of universal type, and S ∗ ∈ Syl2 (G∗ ), such that FS (G) ∼
= FS ∗ (G∗ ), and G∗ has a
68
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
σ-setup which satisfies case (III.1) of Hypotheses 5.1 and also Hypotheses 5.10. Moreover,
if G∗ ∼
= G2 (q ∗ ) where q ∗ is a power of q0∗ , then we can arrange that either q ∗ = 5 or q0∗ = 3.
Proof. Since q0 is odd, and since the Sylow 2-subgroups of 2 G2 (32k+1) are abelian for all
k ≥ 1 [Ree, Theorem 8.5], G must be a Chevalley or Steinberg group. If G ∼
= 3D4 (q), then F
is also the fusion system of G2 (q) by [BMO, Example 4.5]. So we can assume that G ∼
= τ G(q)
for some odd prime power q, some G, and some graph automorphism τ of order 1 or 2.
Let ε ∈ {±1} be such that q ≡ ε (mod 4). By Lemma 1.11, there is a prime q0∗ and k ≥ 0
k
such that hqi = hε · (q0∗ )2 i, where either q0∗ = 5 and k = 0, or q0∗ = 3 and k ≥ 1.
k
If ε = 1, then set G∗ = τ G((q0∗ )2 ), and fix S ∗ ∈ Syl2 (G∗ ). Then FS ∗ (G∗ ) ∼
= FS (G)
by Theorem 1.8(a), G∗ satisfies case (III.1) of Hypotheses 5.1 by Lemma 6.1 (and since
k
(q0∗ )2 ≡ 1 (mod 4)), and G∗ also satisfies Hypotheses 5.10.
k
Now assume ε = −1. If −Id is in the Weyl group of G, then set G∗ = τ G((q0∗ )2 ). If −Id
k
is not in the Weyl group, then G = An , Dn for n odd, or E6 , and we set G∗ = G((q0∗ )2 ) if
k
τ 6= Id, and G∗ = 2 G((q0∗ )2 ) if G = G(q). In all cases, for S ∗ ∈ Sylp (G∗ ), FS ∗ (G∗ ) ∼
= FS (G)
by Theorem 1.8(c,d), G∗ satisfies case (III.1) of Hypotheses 5.1 by Lemma 6.1 again, and
also satisfies Hypotheses 5.10.
k
By construction, if G = G2 , then either q0∗ = 3 or (q0∗ )2 = 5.
When G ∼
= G2 (5) and p = 2, G satisfies Hypotheses 5.1 and 5.10, but κG is not shown to
be surjective in Proposition 5.14 (and in fact, it is not surjective). Hence this case must be
handled separately.
Proposition 6.3. Set G = G2 (5) and G∗ = G2 (3), and fix S ∈ Syl2 (G) and S ∗ ∈ Syl2 (G∗ ).
Then FS ∗ (G∗ ) ∼
= FS (G) as fusion systems, and κG∗ = µG∗ ◦ κG∗ is an isomorphism from
∗ ∼
Out(G ) = C2 onto Out(S ∗ , FS ∗ (G∗ )).
Proof. The first statement follows from Theorem 1.8(c). Also, |Out(G)| = 2 and |Out(G∗ )| =
1 by Theorem 3.4, and since G and G∗ have no field automorphisms and all diagonal automorphisms are inner (cf. [St1, 3.4]), and G = G2 (3) has a nontrivial graph automorphism while
G∗ = G2 (5) does not [St1, 3.6]. Since G satisfies Hypotheses 5.1 and 5.10, |Coker(κG )| ≤ 2
by Proposition 5.14, so |Out(S, FS (G))| ≤ 2.
By [O6, Proposition 4.2], S ∗ contains a unique subgroup Q ∼
= Q8 ×C Q8 of index 2.
2
Let x ∈ Z(Q) = Z(S ∗ ) be the central involution. Set G = G2 (F3 ) > G∗ . Then CG (x) is
connected since G is of universal type [St3, Theorem 8.1], so CG (x) ∼
= SL2 (F3 ) ×C2 SL2 (F3 )
by Proposition 2.5. Furthermore, any outer (graph) automorphism which centralizes x exchanges the two central factors SL2 (F3 ). Hence for each α ∈ Aut(G∗ )rInn(G∗ ) which normalizes S ∗ , α exchanges the two factors Q8 , and in particular, does not centralize S ∗ . Thus
κG∗ is injective, and hence an isomorphism since |Out(G∗ )| = 2 and |Out(S ∗ , FS ∗ (G∗ ))| =
|Out(S, FS (G))| ≤ 2.
We now turn to case (III.2) of Hypotheses 5.1.
Lemma 6.4. Fix an odd prime p, and an odd prime power q prime to p such that q ≡ −1
(mod p). Let G be one of the groups Sp2n (q), Spin2n+1 (q), Spin+
4n (q) (n ≥ 2), G2 (q), F4 (q),
E7 (q), or E8 (q) (i.e., G = G(q) for some G whose Weyl group contains −Id), and assume
that the Sylow p-subgroups of G are nonabelian. Then G has a σ-setup (G, σ) such that
Hypotheses 5.1, case (III.2), hold.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
69
Proof. Assume q = q0b where q0 is prime and b ≥ 1. Set G = G(Fq0 ), and let T < G be a
maximal torus. Set r = rk(T ) and k = vp (q + 1).
In all of these cases, −Id ∈ W , so there is a coset w0 ∈ NG (T )/T which inverts T . Fix
g0 ∈ NG (T ) such that g0 T = w0 and ψq0 (g0 ) = g0 (Lemma 2.9). Set γ = cg0 and σ = γ ◦ ψq .
′
We identify G = O q0 (CG (σ)), T = G ∩ T , and A = Op (T ). Since σ(t) = t−q for each t ∈ T ,
T ∼
= (Cq+1 )r is the (q + 1)-torsion subgroup of T , and A ∼
= (Cpk )r .
NG (T ) contains a Sylow p-subgroup ofQG.
In all cases, by [Ca, Theorem 9.4.10]
r
N
di
(and since G is in universal form), |G| = q
i=1 (q − 1), where d1 d2 · · · dr = |W | by [Ca,
Theorem 9.3.4]. Also, the di are all even in the cases considered here (see [St2, Theorem 25]
or [Ca, Corollary 10.2.4 & Proposition 10.2.5]). Hence by Lemma 1.13 and since p is odd,
vp (|G|) =
r
X
i=1
di
vp (q − 1) =
= r · vp (q + 1) +
r
X
2 di /2
vp (q )
i=1
r
X
i=1
−1 =
r
X
i=1
vp (q 2 − 1) + vp (di /2)
vp (di ) = vp (|T |) + vp (|W |) = vp (|NG (T )|) .
[γ, ψq0 ] = Id since γ = cg0 and ψq0 (g0 ) = g0 .
A free hγi-orbit in Σ. For each α ∈ Σ, {±α} is a free hγi-orbit.
We now consider case (III.3) of Hypotheses 5.1. By [GL, 10-1,2], when p is odd, each
finite group of Lie type has a σ-setup for which NG (T ) contains a Sylow p-subgroup of G.
Here, we need to construct such setups explicitly enough to be able to check that the other
conditions in Hypotheses 5.1 hold.
When p is a prime, A is a finite abelian p-group, and Id 6= ξ ∈ Aut(A) has order prime
to p, we say that ξ is a reflection in A if [A, ξ] is cyclic. In this case, there is a direct
product decomposition A = [A, ξ] × CA (ξ), and we call [A, ξ] the reflection subgroup of ξ.
This terminology will be used in the proofs of the next two lemmas.
Lemma 6.5. Fix an odd prime p, and an odd prime power q prime to p such that q 6≡ 1
(mod p). Let G be one of the classical groups SLn (q), Sp2n (q), Spin2n+1 (q), or Spin±
2n (q),
and assume that the Sylow p-subgroups of G are nonabelian. Then G has a σ-setup (G, σ)
such that case (III.3) of Hypotheses 5.1 holds.
Proof. Set m = ordp (q); m > 1 by assumption. We follow Notation 2.2, except that we have
yet to fix the σ-setup for G. Thus, for example, q0 is the prime of which q is a power.
When defining and working with the σ-setups for the spinor groups, it is sometimes easier
to work with orthogonal groups than with their 2-fold covers. For this reason, throughout
the proof, we set Gc = SOℓ when G = Spinℓ , set Gc = SOℓ (Fq0 ) when G = Spinℓ (Fq0 ), and
let χ : G −−−→ Gc be the natural surjection. We then set Gc = CGc (σ) ∼
= SO±
ℓ (q), once σ
has been chosen so that G = CG (σ) ∼
= Spin±
ℓ (q), and set T c = χ(T ) and Tc = CT c (σ). Also,
in order to prove the lemma without constantly considering these groups as a separate case,
we set Gc = G, Gc = G, χ = Id, etc. when G is linear or symplectic. Thus Gc and Gc are
classical groups in all cases.
Regard Gc as a subgroup of Aut(V , b), where V is an Fq0 -vector space of dimension n, 2n,
0 1 ⊕n
or 2n + 1, and b is a bilinear form. Explicitly, b = 0 if G = SLn , and b has matrix −1
0
70
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
⊕n
⊕n
if G = Sp2n , 01 10
if G = Spin2n , or 01 10
⊕ (1) if G = Spin2n+1 . Let T c be the group
of diagonal matrices in Gc , and set
if G = SLn
diag(λ1 , . . . , λn )
−1
[λ1 , . . . , λn ] = diag(λ1 , λ1 , . . . , λn , λ−1
if G = Sp2n or Spin2n
n )
diag(λ , λ−1 , . . . , λ , λ−1 , 1) if G = Spin
1
n
2n+1 .
1
n
n
In this way, we identify the maximal torus T c < Gc with (F×
q0 ) in the symplectic and
n
orthogonal cases, and with a subgroup of (F×
q0 ) in the linear case.
Set W ∗ = W (the Weyl group of G and of Gc ), except when G = Spin2n , in which
case we let W ∗ < Aut(T c ) be the group of all automorphisms which permute and invert
the coordinates. Thus in this last case, W ∗ ∼
= {±1} ≀ Σn , while W is the group of signed
permutations which invert an even number of coordinates (so [W ∗ : W ] = 2). Since W ∗
induces a group of isometries of the root system for Spin2n and contains W with index 2, it
is generated by W and the restriction to T c of a graph automorphism of order 2 (see, e.g.,
[Brb, § VI.1.5, Proposition 16]).
We next introduce some notation in order to identify certain elements in W ∗ . For each r, s
such that rs ≤ n, let τ sr ∈ Aut(T c ) be the Weyl group element induced by the permutation
(1 · · · r)(r+1 · · · 2r) · · · ((s−1)r + 1 · · · sr); i.e.,
τ sr ([λ1 , . . . , λn ]) = [λr , λ1 , . . . , λr−1 , λ2r , λr+1, . . . , λsr , λ(s−1)r+1 , . . . , λsr−1, λsr+1, . . .].
For 1 ≤ i ≤ n, let ξ i ∈ Aut(T ) be the automorphism which inverts the i-th coordinate. Set
τ sr,+1 = τ sr and τ sr,−1 = τ sr ξr ξ 2r · · · ξsr . Thus for θ = ±1,
τ sr,θ ([λ1 , . . . , λn ]) = [λθr , λ1 , . . . , λr−1 , λθ2r , λr+1 , . . . λθsr , λ(s−1)r+1 , . . . , λsr−1, λsr+1, . . .].
Recall that m = ordp (q). Define parameters µ, θ, k, and κ as follows:
if m is odd :
if m is even :
µ=m
θ=1
µ = m/2
θ = −1
k = [n/m]
κ = [n/µ] = [n/m]
κ = [n/µ] = [2n/m] .
We can now define our σ-setups for G and Gc . Recall that we assume m > 1. In Table
6.1, we define an element w0 ∈ W ∗ , and then describe Tc = CT c (w0 ◦ ψq ) and W0∗ = CW ∗ (w0 )
(where W0 = CW (w0 ) has index at most 2 in W0∗ ). In all cases, we choose γ ∈ Aut(Gc ) as
follows. Write w0 = w0′ ◦ γ0 |T c for some w0′ ∈ W and γ0 ∈ ΓGc (possibly γ0 = Id). Choose
g0 ∈ NGc (T c ) such that g0 T c = w0′ and ψq0 (g0 ) = g0 (Lemma 2.9), and set γ = cg0 ◦ γ0 . Then
[γ, ψq0 ] = IdGc , since cg0 and γ0 both commute with ψq0 , and we set σ = γ ◦ ψq = ψq ◦ γ.
When G = Spin2n or Spin2n+1 , since G is a perfect group and Ker(χ) ≤ Z(G), γ and σ lift to
unique endomorphisms of G which we also denote γ and σ (and still [γ, ψq0 ] = 1 in Aut(G)).
Thus G ∼
= CGc (σ) in all cases, and we identify these groups. Set
= CG (σ) and Gc ∼
∗
T = CT (σ), Tc = CT c (σ), W0 = CW ∗ (γ), and W0 = CW (γ). If G = Spin2n+1 or Spin2n , then
χ(T ) is the kernel of the homomorphism Tc → Ker(χ) which sends χ(t) to t−1 σ(t), and thus
has index at most 2 in Tc . Since p is odd, this proves the statement in the last line of Table
6.1.
In the description of W0∗ in Table 6.1, H always denotes a direct factor of order prime to
p. The first factor in the description of W0∗ acts on the first factor in that of T , and H acts
on the other factors.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
Gc
conditions
SLn (q)
w0 = γ|T c
Tc
W0∗
τ km
n−mk−1
(Cqm −1 )k × Cq−1
(Cm ≀ Σk ) × H
τ κµ,θ
n−κµ
(Cqµ −θ )κ × Cq−1
(C2µ ≀ Σκ ) × H
τ κµ,θ ξ n
n−κµ−1
(Cqµ −θ )κ × Cq−1
× Cq+1
(C2µ ≀ Σκ ) × H
τ κ−1
µ,θ
µ
(Cqµ −θ )κ−1 × Cq−1
71
Sp2n (q)
SO2n+1 (q)
ε = θκ
SOε2n (q)
ε 6= θκ , µ∤n
κ
ε 6= θ , µ|n
θ = −1
ε 6= θκ , µ|n
θ = +1
τ κ−1
µ,θ
ξn
(Cqµ −θ )
κ−1
×
µ−1
Cq−1
× Cq+1
(C2µ ≀ Σκ−1 ) × H
χ
In all cases, T −
→ Tc has kernel and cokernel of order ≤ 2, and so A = Op (T ) ∼
= Op (Tc ).
Table 6.1.
−1
When Gc = SLn (q) and m|n, the second factor Cq−1
in the description of T doesn’t make
sense. It should be interpreted to mean that T is “a subgroup of index q − 1 in the first
factor (Cqm −1 )k ”.
µ
Recall that Tc = CT c (γ ◦ ψq ). When U = (F×
q0 ) , then
µ−1
2
µ−1
CU (τ 1µ,θ ◦ ψq ) = (λ, λq , λq , . . . , λq ) (λq )qθ = λ
µ
2
µ−1
= (λ, λq , λq , . . . , λq ) λq −θ = 1 ∼
= Cqµ −θ .
This explains the description of Tc in the symplectic and orthogonal cases: it is always the
direct product of (Cqµ −θ )κ or (Cqµ −θ )κ−1 with a group of order prime to p. (Note that p|(q+1)
only when m = 2; i.e., when θ = −1 and 1 = µ|n.)
Since the cyclic permutation (1 2 · · · µ) generates its own centralizer in Σµ , the centralizer
µ
1
T
1
µ
of τ 1µ,θ in {±1} ≀ Σµ < Aut((F×
q0 ) ) is generated by τ µ,θ and ψ−1 . If θ = −1, then (τ µ,θ ) =
T
, while if θ = 1, then τ 1µ,θ has order µ. Since m = µ is odd in the latter case, the
ψ−1
centralizer is cyclic of order 2µ in both cases. This is why, in the symplectic and orthogonal
cases, the first factor in W0∗ is always a wreath product of C2µ with a symmetric group.
We are now ready to check the conditions in case (III.3) of Hypotheses 5.1.
NG (T ) contains a Sylow p-subgroup of G. Set
e = vp (q m − 1) = vp (q µ − θ) .
The second equality holds since if 2|m, then p ∤ (q µ − 1) and hence e = vp (q µ + 1). Recall
also that m|(p − 1), so vp (m) = 0. Consider the information listed in Table 6.2, where the
formulas for vp (|T |) = vp (|Tc |) and vp (|W0 |) follow from Table 6.1, and those for |G| are
shown in [St2, Theorems 25 & 35] and also in [Ca, Corollary 10.2.4, Proposition 10.2.5 &
Theorem 14.3.2].
For all i > 0, we have
(
e + vp (i/m) if m|i
vp (q i − 1) =
0
if m∤i.
72
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
G
cond.
SLn (q)
vp (q i − 1)
i=2
Sp2n (q)
Pn
i=1
Spin2n+1 (q)
ε = θκ
Spinε2n (q)
vp (|G|)
Pn
vp (|T |)
vp (|W0 |)
ke
vp (k!)
κe
vp (κ!)
vp (q 2i − 1)
vp (q n − ε)
ε 6= θ , µ∤n
Pn−1
+ i=1
vp (q 2i − 1)
κ
ε 6= θ , µ|n
(κ−1)e vp ((κ−1)!)
κ
Table 6.2.
The first case follows from Lemma 1.13, and the second case since m = ordp (q). Using this,
we check that vp (q 2i − 1) = vp (q i − 1) for all i whenever m is odd, and that
(
e + vp (2n/m) if m|2n and ε = (−1)2n/m
vp (q n − ε) =
0
otherwise.
So in all cases, vp (|G|) = vp (|T |) + vp (|W0 |) by the above relations and the formulas in Table
6.2. Since NG (T )/T ∼
= W0 by Lemma 5.3(b), this proves that vp (|G|) = vp (|NG (T )|), and
hence that NG (T ) contains a Sylow p-subgroup of G.
γ|T = |τ | = ordp (q) ≥ 2 and [γ, ψq0 ] = Id by construction. Note, when G is a spinor
group, that these relations hold in G if and only if they hold in Gc .
CS (Ω1 (A)) = A by Table 6.1 and since p ∤ |H|.
CA (Op′ (W0 )) = 1. By Table 6.1, in all cases, there are r, t ≥ 1 and 1 6= s|(p − 1) such
that A ∼
= Cs ≀ Σr acts on A by acting on and permuting the cyclic
= (Cpt )r , and AutW0∗ (A) ∼
factors. In particular, AutOp′ (W0 ) (A) contains a subgroup of index at most 2 in (Cs )r , this
subgroup acts nontrivially on each of the cyclic factors in A, and hence CA (Op′ (W0 )) = 1.
A free hγi-orbit in Σ. This can be defined as described in Table 6.3. In each case, we
use the notation of Bourbaki [Brb, pp. 250–258] for the roots of G. Thus, for example, the
roots of SLn are the ±(εi − εj ) for 1 ≤ i < j ≤ n, and the roots of SO2n the ±εi ± εj . Note
that since S is assumed nonabelian, p |W0 |, and hence n ≥ pm in the linear case, and n ≥ pµ
in the other cases.
G
SLn (q)
θ=1
θ = −1
{εi − εm+i | 1 ≤ i ≤ m}
Sp2n (q)
{2εi | 1 ≤ i ≤ µ}
{±2εi | 1 ≤ i ≤ µ}
Spin2n+1 (q)
{εi | 1 ≤ i ≤ µ}
{±εi | 1 ≤ i ≤ µ}
Spinε2n (q)
{εi − εµ+i | 1 ≤ i ≤ µ} {±(εi − εµ+i ) | 1 ≤ i ≤ µ}
Table 6.3.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
73
(
hγ|A i
if ordp (q) even or −Id ∈
/ W0
AutW0 (A) ∩ Autsc (A) ≤
A
hγ|A , ψ−1 i otherwise.
∗
Set K = AutW0∗ (A) ∩ Autsc (A) and K = AutW0 (A) ∩ Autsc (A) for short. By Table 6.1,
|K ∗ | = m if G ∼
= SLn (q), and |K ∗ | = 2µ otherwise. Also, hγ|A i = hψq−1 |A i has order ordp (q).
∗
Thus K ≤ K = hγ|A i except when G is symplectic or orthogonal and m = ordp (q) is odd.
In this last case, K = K ∗ (so |K| = 2µ = 2m) if W0 contains an element which inverts A
(hence which inverts T and T ); and |K ∗ /K| = 2 (|K| = m) otherwise.
AutG (A) = AutW0 (A). Since A = Op (T ) ∼
= Op (Tc ) by Table 6.1, it suffices to prove
g
this for Gc . Fix g ∈ NGc (A). Since T c is a maximal torus in the algebraic group CGc (A)
(Proposition 2.5), there is b ∈ CGc (A) such that bT c = gT c . Set a = b−1 g ∈ NGc (T c ); thus
ca = cg ∈ Aut(A). Set w = aT c ∈ W = NGc (T c )/T c ; thus w ∈ NW (A), and w|A = cg |A .
By the descriptions in Table 6.1, we can factor T c = T 1 × T 2 , where γ and each element of
NW (A) send each factor to itself, γ|T 2 = Id, A ≤ T 1 , and [CW (A), T 1 ] = 1. Since σ(g) = g,
σ(a) ≡ a (mod CGc (A)), and so τ (w) ≡ w (mod CW (A)). Thus τ (w)|T 1 = w|T 1 since CW (A)
acts trivially on this factor, τ (w)|T 2 = w|T 2 since γ|T 2 = Id, and so w ∈ W0 = CW (τ ).
By Table 6.1, for some r, t ≥ 1,
NAut(A) (AutW0 (A)) ≤ Autsc (A)AutAut(G) (A).
A = A1 ×· · ·×Ar , where Ai ∼
= Cs ≀Σr
= Cpt for each i. Also, for some 1 6= s|(p−1), AutW0∗ (A) ∼
acts on A via faithful actions of Cs on each Ai and permutations of the Ai .
Let Aut0W0∗ (A) E AutW0∗ (A) and Aut0W0 (A) E AutW0 (A) be the subgroups of elements
which normalize each cyclic subgroup Ai . Thus Aut0W0∗ (A) ∼
= (Cs )r , and contains Aut0W0 (A)
with index at most 2.
Case 1:
Assume first that Aut0W0 (A) is characteristic in AutW0 (A). Fix some α ∈
NAut(A) (AutW0 (A)). We first show that α ∈ AutW0∗ (A)Autsc (A).
Since α normalizes AutW0 (A), it also normalizes Aut0W0 (A). For each β ∈ Aut0W0 (A),
[β, A] is a product of Ai ’s. Hence the factors Ai are characterized as the minimal nontrivial
intersections of such [β, A], and are permuted by α. So after composing with an appropriate
element of AutW0∗ (A), we can assume that α(Ai ) = Ai for each i.
After composing α by an element of Autsc (A), we can assume that α|A1 = Id. Fix i 6= 1
(2 ≤ i ≤ r), let u ∈ Z be such that α|Ai = ψuAi = (a 7→ au ), and choose w ∈ AutW0 (A)
−1
−1
such that w(A
1 ) = AAi . Then w s αwα ∈ Autt W0 (A) since α normalizes AutW0 (A), and
−1
−1
w αwα
= ψu 1 . Hence u ≡ 1 (mod p = |A1 |), and since this holds for each i,
A1
α ∈ AutW0∗ (A).
Thus NAut(A) (AutW0 (A)) ≤ AutW0∗ (A)Autsc (A). By Table 6.1, each element of AutW0∗ (A)
extends to some ϕ ∈ AutW ∗ (T ) which commutes with σ|T . So AutW0∗ (A) ≤ AutAut(G) (A) by
Lemma 3.7, and this finishes the proof of the claim.
Case 2: Now assume that Aut0W0 (A) is not characteristic in AutW0 (A). Then r ≤ 4, and
since p ≤ r, we have p = 3 and r = 3, 4. Also, s = 2 since s|(p − 1) and s 6= 1. Thus r = 4,
since Aut0W0 (A) = O2 (AutW0 (A)) if r = 3. Thus AutW0 (A) ∼
= C23 ⋊ Σ4 : the Weyl group of
D4 . Also, m = 2 since p = 3, so (in the notation used in the tables) µ = 1, θ = −1, and
κ = n. By Table 6.1, G ∼
= SO8 (q) for some q ≡ 2 (mod 3) (and W0 = W ).
∼
Now, O2 (W ) = Q8 ×C2 Q8 , and so Out(O2 (W )) ∼
=
= Σ3 ≀C2. Under the action of W/O2(W ) ∼
Σ3 , the elements of order 3 act on both central factors and those of order 2 exchange
the factors. (This is seen by computing their centralizers in O2 (W ).) It follows that
74
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
NOut(O2 (W )) (OutW (O2 (W )))/OutW (O2 (W )) ∼
= Σ3 ∼
= ΓG , and all classes in this quotient
∼
extend to graph automorphisms of G = Spin8 (q). So for each α ∈ NAut(A) (AutW (A)), after
composing with a graph automorphism of G we can arrange that α commutes with O2 (W ),
and in particular, normalizes Aut0W (A). Hence by the same argument as used in Case 1,
α ∈ Autsc (A)AutAut(G) (A).
This finishes the proof that this σ-setup for G satisfies case (III.3) of Hypotheses 5.1.
Example 6.6. Fix distinct odd primes p and q0 , and a prime power q = q0b where b is even
and ordp q is even. Set G = Spin−
4k (q) for some k ≥ 2. Let (G, σ) be the setup for G of
Lemma 6.5, where σ = ψq γ for γ ∈ Aut(G). In the notation of Table 6.1, m = ordp (q),
µ = m/2, θ = −1 = ε, n = 2k, and κ = [2k/µ] = [4k/m]. There are three cases to consider:
(a) If q 2k ≡ −1 (mod p); equivalently, if m|4k and κ = 4k/m is odd, then ε = θκ , w0 =
γ|T c = τ κµ,θ , rk(A) = κ, and W0∗ ∼
= Cm ≀ Σκ . Then W0∗ acts faithfully on A while
w0 ∈ W0∗ rW0 , and so γ|A ∈
/ AutW0 (A). Hence by Proposition 5.15(d), κG is split.
(b) If q 2k ≡ 1 (mod p); equivalently, if m|4k and κ = 4k/m is even, then ε 6= θκ , γ|T c =
∗ ∼
∼
τ κ−1
µ,θ , rk(A) = κ − 1, and W0 = (Cm ≀ Σκ−1 ) × H where H = (C2 ≀ Σµ ). Then H acts
trivially on A and contains elements in W0∗ rW0 , so γ|A ∈ AutW0 (A). Hence κG is not
split.
(c) If q 2k 6≡ ±1 (mod p); equivalently, if m ∤ 4k, then in either case (κ even or odd), the
factor H in the last column of Table 6.1 is nontrivial, acts trivially on A, and contains
elements in W0∗ rW0 . Hence γ|A ∈ AutW0 (A) in this case, and κG is not split.
We also need the following lemma, which handles the only case of a Chevalley group of
exceptional type which we must show satisfies case (III.3) of Hypotheses 5.1.
Lemma 6.7. Set p = 5, let q be an odd prime power such that q ≡ ±2 (mod 5), and set
G = E8 (q). Then G has a σ-setup which satisfies Hypotheses 5.1 (case (III.3)).
Proof. We use the notation in 2.2, where q is a power of the odd prime q0 , and G = E8 (Fq0 ).
By [Brb, Planche VII], the of roots of E8 can be given the following form, where {ε1 , . . . , ε8 }
denotes the standard orthonormal basis of R8 :
8
8
n
o n1 X
o
X
mi
Σ = ±εi ± εj 1 ≤ i < j ≤ 8 ∪
(−1) εi
mi even ⊆ R8 .
2 i=1
i=1
By the same reference, the Weyl group W is the group of all automorphisms of R8 which
permute Σ (A(R) = W (R) in the notation of [Brb]). Give R8 a complex structure by setting
iε2k−1 = ε2k and iε2k = −ε2k−1 , and set ε∗k = ε2k−1 for 1 ≤ k ≤ 4. Multiplication by i
permutes Σ, and hence is the action of an element w0 ∈ W . Upon writing the elements of Σ
with complex coordinates, we get the following equivalent subset Σ∗ ⊆ C4 :
n
o n
o
∗
∗
m ∗
n ∗
Σ = (±1 ± i)εk 1 ≤ k ≤ 4 ∪ i εk + i εℓ 1 ≤ k < ℓ ≤ 4, m, n ∈ Z
∪
4
n1 + i X
2
k=1
imk ε∗k
X
mk even
o
.
Let ZΣ ⊆ R8 be the lattice generated by Σ. By Lemma 2.4(d) (and since (α, α) = 2 for
×
all α ∈ Σ), we can identify T ∼
= ZΣ ⊗Z F×
q0 by sending hα (λ) to α ⊗ λ for α ∈ Σ and λ ∈ Fq0 .
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
75
Set Λ0 = ZΣ ∩ Z8 , a lattice in R8 of index 2 in ZΣ and in Z8 . The inclusions of lattices
induce homomorphisms
χ1
χ2
∼ × 8
−−−−− Λ0 ⊗Z F×
−−−−−→ Z8 ⊗Z F×
T ∼
= ZΣ ⊗Z F×
q0 ←−
q0 −
q0 = (Fq0 )
∼
each of which is surjective with kernel of order 2 (since Tor1Z (Z/2, F×
q0 ) = Z/2). We can thus
8
identify T = (F×
q0 ) , modulo 2-power torsion, in a way so that
α=
8
X
i=1
ki εi ∈ Σ, λ ∈ F×
q0
=⇒
hα (λ) = (λk1 , . . . , λk8 ) .
Under this identification, by the formula in Lemma 2.4(c),
β=
8
X
i=1
ℓi ε i ∈ Σ
=⇒
θβ (λ1 , . . . , λ8 ) = λℓ11 · · · λℓ88
(1)
for λ1 , . . . , λ8 ∈ F×
q0 . Also,
−1
−1
w0 (λ1 , . . . , λ8 ) = (λ−1
2 , λ1 , λ4 , λ3 , . . . , λ8 , λ7 )
for each (λ1 , . . . , λ8 ).
Choose g0 ∈ NG (T ) such that g0 T = w0 and ψq0 (g0 ) = g0 (Lemma 2.9), and set γ = cg0 ∈
Inn(G). Thus σ = ψq ◦ γ = γ ◦ ψq , G = CG (σ), and T = CT (σ). By the Lang-Steinberg
theorem [St3, Theorem 10.1], there is h ∈ G such that g = hψq (h−1 ); then σ = ch ψq c−1
h and
∼
G = CG (ψq ) = E8 (q). It remains to check that the setup (G, σ) satisfies the list of conditions
in Hypotheses 5.1.
We identify W0 = CW (w0 ) with the group of C-linear automorphisms of C4 which permute
∗
Σ . The order of W0 is computed in [Ca3, Table 11] (the entry Γ = D4 (a1 )2 ), but since we
need to know more about its structure, we describe it more precisely here. Let W2 ≤ GL4 (C)
be the group of monomial matrices with nonzero entries ±1 or ±i, and with determinant ±1.
Then W2 ≤ W0 , |W2 | = 21 · 44 · 4! = 210 · 3, and W2 acts on Σ∗ with three orbits corresponding
to the three subsets in the above description of Σ∗ . The (complex) reflection of order 2 in
(ε∗1 + ε∗2 + ε∗3 + ε∗4 ) sends (1 + i)ε∗1 to 1+i
(ε∗1 − ε∗2 − ε∗3 − ε∗4 ),
the hyperplane orthogonal to 1+i
2
2
and it sends (ε∗1 + iε∗2 ) to 1+i
(i3 ε∗1 + iε∗2 − ε∗3 − ε∗4 ). Thus W0 acts transitively on Σ∗ .
2
Let Σ ⊆ P (C4 ) be the set of projective points representing elements of Σ∗ , and let [α] ∈ Σ
denote the class of α ∈ Σ∗ . To simplify notation, we also write [x] = [α] for x ∈ C4
representing the same point, also when x ∈
/ Σ∗ . Let ∼ denote the relation on Σ: [α] ∼ [β] if
α = β, or if α ⊥ β and the projective line [α], [β] ⊆ P (C4 ) contains four other points in Σ.
By inspection, [ε∗j ] ∼ [ε∗k ] for all j, k ∈ {1, 2, 3, 4}, and these are the only elements [α] such
that [α] ∼ [ε∗j ] for some j. Since this relation is preserved by W0 , and W0 acts transitively
on Σ, we see that ∼ is an equivalence relation on Σ with 15 classes of four elements each.
Set ∆ = Σ/∼, and let [α]∆ denote the class of [α] in ∆. Thus |Σ| = 41 |Σ| = 60 and |∆| = 15.
Since W2 is the stabilizer subgroup of [ε∗1 ]∆ under the transitive W0 -action on ∆, we have
|W0 | = |W2 | · 15 = 210 · 32 · 5.
Let W1 E W0 be the subgroup of elements which act trivially on ∆. By inspection,
W1 ≤ W2 , |W1 | = 26 , and W1 is generated by w0 = diag(i, i, i, i), diag(1, 1, −1, −1),
diag(1, −1, 1, −1), and the permutation matrices for the permutations (1 2)(3 4) and (1 3)(2 4).
Thus W1 ∼
= C4 ×C2 D8 ×C2 D8 .
76
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
By the above computations, |W0 /W1 | = 24 · 32 · 5 = |Sp4 (2)|. There is a bijection from
∆ to the set of maximal isotropic subspaces in W1 /Z(W1 ) which sends a class [α]∆ to the
subgroup of those elements in W1 which send each of the four projective points in [α]∆ to
itself. Hence for each w ∈ CW0 (W1 ), w acts via the identity on ∆, and so w ∈ W1 by
definition. Thus W0 /W1 injects into Out(W1 ) ∼
= Σ6 × C2 , and injects into the first factor
∼
since Z(W1 ) = Z(W0 ) (= C4 ). So by counting, W0 /W1 ∼
= Σ6 . Also, W1 = O2 (W0 ).
a
Set a = v5 (q 4 − 1) = v5 (q 2 + 1), and fix u ∈ F×
q0 of order 5 . Let A be as in Notation
5.2(G): the subgroup of elements in T of 5-power order. Thus
(2)
A = (u1 , uq1 , u2, uq2 , u3 , uq3, u4 , uq4) u1 , u2, u3 , u4 ∈ hui ∼
= (C5a )4 .
By (2) and (1), there is no β ∈ Σ such that A ≤ Ker(θβ ). Hence CG (A)0 = T by Proposition
2.5. So by Lemma 5.3(b),
NG (A) = NG (T ) and NG (T )/T = W0 .
(3)
We are now ready to check the conditions in Case (III.3) of Hypotheses 5.1.
NG (T ) contains a Sylow p-subgroup of G. Let S be a Sylow p-subgroup of NG (T )
which contains A. Since NG (T )/T = W0 by (3), A ∼
= Σ6 ,
= (C5a )4 , and W0 /O2 (W0 ) ∼
|S| = 54a+1 . By [St2, Theorem 25] or [Ca, Corollary 10.2.4 & Proposition 10.2.5], and since
v5 (q k − 1) = 0 when 4 ∤ k and v5 (q 4ℓ − 1) = a + v5 (ℓ) (Lemma 1.13),
v5 (|G|) = v5 (q 24 − 1)(q 20 − 1)(q 12 − 1)(q 8 − 1) = 4a + 1 .
Thus S ∈ Sylp (G).
γ|T = ordp (q) ≥ 2 and [γ, ψq0 ] = Id. The first is clear, and the second holds since
γ = cg0 where ψq0 (g0 ) = g0 .
CS (Ω1 (A)) = A by the above description of the action of W0 on A.
CA (Op′ (W0 )) = 1 since w0 ∈ O5′ (W0 ) and CA (w0 ) = 1.
A free hγi-orbit in Σ. The subset {±(ε1 + ε3 ), ±(ε2 + ε4 )} ⊆ Σ is a free hγi-orbit.
AutW0 (A) ∩ Autsc (A) ≤ hγ|A i. Recall that γ|T = 4 and |Autsc (A)| = 4 · 5k for some
k, and W0 acts faithfully on A. So if this is not true, then there is an element of order 5 in
Z(W0 ), which is impossible by the above description of W0 .
AutG (A) = AutW0 (A) by (3).
NAut(A) (AutW0 (A)) ≤ Autsc (A)AutW0 (A). For j = 1, 2, 3, 4, let Aj < A be the cyclic
subgroup of all elements as in (2) where uk = 1 for k 6= j. The group W0 contains as
subgroup C2 ≀ Σ4 : the group which permutes pairs of coordinates up to sign. So each of the
four subgroups Aj is the reflection subgroup of some reflection in W0 .
For each ϕ ∈ CAut(A) (AutW0 (A)), ϕ(Aj ) = Aj for each j, and ϕ(a) = anj for some
nj ∈ (Z/5a )× . Also, n1 = n2 = n3 = n4 since the Aj are permuted transitively by elements
of W0 , and hence ϕ ∈ Autsc (A).
Now assume ϕ ∈ NAut(A) (AutW0 (A)). Since ϕ centralizes Z(W1 ) = hw0 i = hdiag(i, i, i, i)i
(since diag(i, i, i, i) ∈ Z(Aut(A))), cϕ |W1 ∈ Inn(W1 ), and we can assume (after composing
by an appropriate element of W1 ) that [ϕ, W1 ] = 1. So cϕ ∈ Aut(W0 ) has the form cϕ (g) =
gχ(g), where g ∈ W0 /W1 ∼
=
= Σ6 is the class of g ∈ W0 , and where χ ∈ Hom(W0 /W1 , Z(W1 )) ∼
∼
Hom(Σ6 , C4 ) = C2 is some homomorphism. Since (w0 )2 inverts the torus T , composition
with (w0 )2 does not send reflections (in A) to reflections, and so we must have cϕ = IdW0 .
Thus ϕ ∈ CAut(A) (AutW0 (A)) = Autsc (A) (modulo AutW0 (A)).
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
77
The following lemma now reduces the proof of Theorem B to the cases considered in
Section 5, together with certain small cases handled at the end of this section. As before,
when p is a prime and p ∤ n, ordp (n) denotes the multiplicative order of n in F×
p.
Proposition 6.8. Fix an odd prime p, and assume G ∈ Lie(q0 ) is of universal type for some
prime q0 6= p. Fix S ∈ Sylp (G), and assume S is nonabelian. Then there is a prime q0∗ 6= p,
a group G∗ ∈ Lie(q0∗ ) of universal type, and S ∗ ∈ Sylp (G∗ ), such that FS (G) ∼
= FS ∗ (G∗ ), and
one of the following holds: either
(a) G∗ has a σ-setup which satisfies Hypotheses 5.1 and 5.10, G∗ ∼
= G(q ∗ ) or 2 G(q ∗ ) where
q ∗ is a power of q0∗ , and
(a.1) −Id ∈
/ W and G∗ is a Chevalley group, or
(a.2) −Id ∈ W and ordp (q ∗ ) is even, or
(a.3) p ≡ 3 (mod p) and ordp (q ∗ ) = 1
where W is the Weyl group of G; or
(b) p = 3, q0∗ = 2, G ∼
= 3D4 (q) or 2F4 (q) for q some power of q0 , and G∗ ∼
= 3D4 (q ∗ ) or 2F4 (q ∗ )
for q ∗ some power of 2.
Moreover, if p = 3 and G∗ = F4 (q ∗ ) where q ∗ is a power of q0∗ , then we can assume q0∗ = 2.
In all cases, we can choose G∗ to be either one of the groups listed in Proposition 1.10(a–e),
or one of E7 (q ∗ ) or E8 (q ∗ ) for some q ∗ ≡ −1 (mod p).
Proof. We can assume that G = G(q) is one of the groups listed in one of the five cases
(a)–(e) of Proposition 1.10. In all cases except 1.10(c), we can also assume that G satisfies
Hypotheses 5.10, with q0 = 2 if p = 3 and G = F4 , and with q0 odd in cases (a) and (b)
of 1.10. If G = SLn (q) or Spin±
2n (q) where p|(q − 1), or G is in case (d), then G satisfies
Hypotheses 5.1 by Lemma 6.1. If G ∼
= SLn (q) or Spin±
2n (q) where p ∤ (q − 1), then G
satisfies Hypotheses 5.1 by Lemma 6.5. This leaves only case (c) in Proposition 1.10, which
corresponds to case (b) here, and case (e) (p = 5, G = E8 (q), q ≡ ±2 (mod 5)) where G∗
satisfies Hypotheses 5.1 by Lemma 6.7.
We next show, in cases (a,b,d,e) of Proposition 1.10, that we can arrange for one of the
conditions (a.1), (a.2), or (a.3) to hold. If −Id ∈
/ W , then G = An , Dn for n odd, or E6 ,
and G is a Chevalley group by the assumptions in cases (a,b,d) of Proposition 1.10. So
(a.1) holds. If −Id ∈ W and ordp (q) is even, then (a.2) holds, while if p ≡ 3 (mod 4) and
ordp (q) = 1, then (a.3) holds.
By inspection, we are left with the following two cases:
n
(b′ ) G = G(q), where G = Spin+
2n , n ≥ p is even, ordp (q) is odd, and q ≡ 1 (mod p); or
(d′ ) G = G(q) where G = G2 , F4 , E6 , E7 , or E8 , p ≡ 1 (mod 4), p|(q − 1), and p |W (G)|.
In case (d′ ), the above conditions leave only the possibility p = 5 and G = E7 or E8 (see
the computations of |W | in, e.g., [Brb, § VI.4]). In either case, by Lemma 1.11(a), we can
choose a prime power q ∨ which satisfies Hypotheses 5.10 and such that hq ∨ i = h − qi in Z×
p,
∨
∨
∨
∨
and set G = G(q ). Then G ∼p G by Theorem 1.8(c), ordp (q ) is even, so (a.2) holds,
and G∨ satisfies Hypotheses 5.1 by Lemma 6.4 or 6.5.
We now consider the two families of groups which appear in Proposition 6.8(b): those not
covered by Hypotheses 5.1.
Proposition 6.9. Let G be one of the groups 3D4 (q) where q is a prime power prime to 3,
2
F4 (22m+1 ) for m ≥ 0, or 2F4 (2)′ . Then the 3-fusion system of G is tame. If G ∼
= 3D4 (2n )
78
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
(n ≥ 1), 2F4 (22m+1 ) (m ≥ 0), or 2F4 (2)′ , then κG is split surjective, and Ker(κG ) is the
subgroup of field automorphisms of order prime to 3.
Proof. Fix S ∈ Syl3 (G), and set F = FS (G).
If G is the Tits group 2F4 (2)′ , then S is extraspecial of order 33 and exponent 3, so
Out(S) ∼
= SD16 , since the normalizer in
= D8 and OutAut(G) (S) ∼
= GL2 (3). Also, OutG (S) ∼
2
F4 (2) of an element of order 3 (the element t4 in [Sh]) has the form SU3 (2) : 2 ∼
: SD16
= 31+2
+
by [Sh, Table IV] or [Ma, Proposition 1.2]. Hence Out(S, F ) ≤ NOut(S) (OutG (S))/OutG (S)
has order at most 2, and κG sends Out(G) ∼
= C2 ([GrL, Theorem 2]) isomorphically to
SD
Out(S, F ). If G = 2F4 (2), then OutG (S) ∼
=
16 , so Out(S, F ) = 1 by a similar argument,
and κG is an isomorphism between trivial groups.
Assume now that G ∼
= 2F4 (2n ) for odd n ≥ 3 or G ∼
= 3D4 (q) where 3 ∤ q. In order to
describe the Sylow 3-subgroups of these groups, set ζ = e2πi/3 , R = Z[ζ], and p = (1 − ζ)R.
Let Sk be the semidirect product R/pk ⋊ C3 , where the quotient acts via multiplication by
ζ. Explicitly, set
Sk = {(x, i) | x ∈ R/pk , i ∈ Z/3}
and
Ak = R/pk × {0},
where (x, i)(y, j) = (x + ζ i y, i + j). Thus |Sk | = 3k+1. Set s = (0, 1), so that s(x, 0)s−1 =
(ζx, 0) for each x ∈ R/pk .
Assume k ≥ 3, so that Ak is the unique abelian subgroup of index three in Sk . Set S = Sk
and A = Ak for short. We want to describe Out(S). Define automorphisms ξa (a ∈ (R/pk )× ),
ω, η, and ρ by setting
ξa (x, i) = (xa, i),
η = ξ−1 ,
ω(x, i) = (−x, −i),
ρ(x, i) = (x + λ(i), i).
(4)
Here, x 7→ x means complex conjugation, and λ(i) = 1 + ζ + . . . + ζ i−1. Note, when checking
that ρ is an automorphism, that λ(i) + ζ i λ(j) = λ(i + j). Note that ρ3 ∈ Inn(S): it is (left)
conjugation by (1 − ζ 2, 0).
Let Aut0 (S) E Aut(S) be the subgroup of automorphisms which induce the identity on
S/[S, S] = S/[s, A], and set Out0 (S) = Aut0 (S)/Inn(S). Each element in s·[s, A] is conjugate
to s, and thus each class in Out0 (S) is represented by an automorphism which sends s to
itself, which is unique modulo hcs i. If ϕ ∈ Aut(S) and ϕ(s) = s, then ϕ|A commutes with cs ,
thus is R-linear under the identification A ∼
= R/pk , and hence ϕ = ξa for some a ∈ 1 + p/pk .
Moreover, since
(1 + p/pk )× = (1 + p2 /pk )× × hζi = (1 + 3R/pk )× × hζi
as multiplicative groups (just compare orders, noting that the groups on the right have trivial
intersection), each class in Out0 (S) is represented by ξa for some unique a ∈ 1 + 3R/pk .
Since the images of η, ω, and ρ generate Aut(S)/Aut0 (S) (the group of automorphisms of
S/[s, A] ∼
= C3 ), this shows that Out(S) is generated by the
= C32 which normalize A/[s, A] ∼
classes of the automorphisms in (4). In fact, a straightforward check of the relations among
them shows that
0
∼
where
Out0 (S) = [ξa ] a ∈ (1 + 3R/pk )× .
Out(S) = Out (S) ⋊ C2 × Σ3
[ω]
[ρ],[η]
Also, ωξa ω −1 = ξā for a ∈ (1 + 3R/pk )× .
For each x ∈ 1 + 3R such that x ≡ x (mod pk ), we can write x = r + sζ with r, s ∈ Z, and
then s(ζ − ζ) ∈ pk , so s ∈ pk−1, and x ∈ r + s + pk ⊆ 1 + 3Z + pk . This proves that
COut(S) (ω) = [ξa ] a ∈ Z × [ω] × [ρ], [η] .
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
79
For any group G with S ∈ Syl3 (G) and S ∼
= Sk , OutG (S) has order prime to 3, and hence
is a 2-group and conjugate to a subgroup of hω, ηi ∈ Syl2 (Out(S)). If |OutG (S)| = 4, then
we can identify S with Sk in a way so that OutG (S) = [ω], [η] . Then
Out(S, F ) ≤ NOut(S) [ω], [η] / [ω], [η]
= COut(S) [ω], [η] / [ω], [η] = [ξa ] a ∈ Z = [ξ2 ] ,
where the first equality holds since O3 (Out(S)) has index four in Out(S).
We are now ready to look at the individual groups. Assume G = 2F4 (q), where q = 2n
and n ≥ 3 is odd. By [St1, 3.2–3.6], Out(G) is cyclic of order n, generated by the field
automorphism ψ2 . By the main theorem in [Ma], there is a subgroup NG (T8 ) ∼
= (Cq+1 )2 ⋊
GL2 (3), the normalizer of a maximal torus, which contains a Sylow 3-subgroup. Hence if we
set k = v3 (q + 1) = v3 (4n − 1) = 1 + v3 (n) (Lemma 1.13), we have S ∼
= (C3k )2 ⋊ C3 ,
= S2k ∼
and OutG (S) = hω, ηi up to conjugacy. So Out(S, F ) is cyclic, generated by ξ2 = κG (ψ2 ).
Since A ∼
= (C3k )2 , and since ξ−1 ∈ OutG (S), |Out(S, F )| = |[ξ2 ]| = 3k−1 where k − 1 = v3 (n).
Thus κG is surjective, and is split since the Sylow 3-subgroup of Out(G) ∼
= Cn is sent
isomorphically to Out(S, F ).
Next assume G = 3D4 (q), where q = 2n for n ≥ 1. By [St1, 3.2–3.6], Out(G) is cyclic of
order 3n, generated by the field automorphism ψ2 (and where the field automorphism ψ2n
of order three is also a graph automorphism). Set k = v3 (q 2 − 1) = v3 (22n − 1) = 1 + v3 (n)
(Lemma 1.13). Then S ∼
= S2k+1 : this follows from the description of the Sylow structure
in G in [GL, 10-1(4)], and also from the description (based on [Kl]) of its fusion system in
[O5, Theorem 2.8] (case (a.ii) of the theorem). Also, OutG (S) = hω, ηi up to conjugacy.
So Out(S, F ) is cyclic, generated by ξ2 = κG (ψ2 ). Since A ∼
= C3k × C3k+1 , and since ξ−1 ∈
OutG (S), |Out(S, F )| = |[ξ2 ]| = 3k . Thus κG is surjective, and is split since the Sylow
3-subgroup of Out(G) ∼
= C3n is sent isomorphically to Out(S, F ).
By Theorem 1.8(b) and Lemma 1.11(a), for each prime power q with 3 ∤ q, the 3-fusion
system of 3D4 (q) is isomorphic to that of 3D4 (2n ) for some n. By [O1, Theorem C], µG is
injective in all cases. Thus the 3-fusion systems of all of these groups are tame.
80
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Appendix A. Injectivity of µG
Bob Oliver
Recall that for any finite group G and any S ∈ Sylp (G),
µG : Outtyp (LcS (G)) −−−−−−−→ Out(S, FS (G))
is the homomorphism which sends the class of β ∈ AutItyp (LcS (G)) to the class of βS |S , where
βS is the induced automorphism of AutLcS (G) (S) = NG (S)/Op′ (CG (S)). We need to develop
tools for computing Ker(µG ), taking as starting point [AOV, Proposition 4.2].
As usual, for a finite group G and a prime p, a proper subgroup H < G is strongly pembedded in G if p |H|, and p ∤ |H ∩ gH| for g ∈ GrH. The following properties of groups
with strongly embedded subgroups will be needed.
Lemma A.1. Fix a prime p and a finite group G.
(a) If G contains a strongly p-embedded subgroup, then Op (G) = 1.
(b) If H < G is strongly p-embedded, and K E G is a normal subgroup of order prime to p
such that KH < G, then HK/K is strongly p-embedded in G/K.
Proof. (a) See, e.g., [AKO, Proposition A.7(c)].
(b) Assume otherwise. Thus there is g ∈ GrHK such that p |(gHK/K) ∩ (HK/K)|, and
hence x ∈ gHK ∩ HK of order p. Then H ∩ Khxi and gH ∩ Khxi have order a multiple of
p, so there are elements y ∈ H and z ∈ gH of order p such that y ≡ x ≡ z (mod K).
Since hyi, hzi ∈ Sylp (Khxi), there is k ∈ K such that hyi = khzi. Then y ∈ H ∩ kgH,
and kg ∈
/ H since k ∈ K and g ∈
/ HK. But this is impossible, since H is strongly pembedded.
For the sake of possible future applications, we state the next proposition in terms of
abstract fusion and linking systems. We refer to [AOV], and also to Chapters I.2 and III.4
in [AKO], for the basic definitions. Recall that if F is a fusion system over a finite p-group
S, and P ≤ S, then
• P is F -centric if CS (Q) ≤ Q for each Q which is F -conjugate to P ;
• P is fully normalized in F if |NS (P )| ≥ |NS (Q)| whenever Q is F -conjugate to P ; and
• P is F -essential if P < S, P is F -centric and fully normalized in F , and if OutF (P )
contains a strongly p-embedded subgroup.
For any saturated fusion system F over a finite p-group S, set
b ) = E ≤ S E elementary abelian, fully normalized in F ,
Z(F
E = Ω1 (Z(CS (E))), AutF (E) has a strongly p-embedded subgroup .
The following proposition is our main tool for proving that µL is injective in certain cases.
Point (a) will be used to handle the groups Spin±
n (q), point (c) the linear and symplectic
groups, and point (b) the exceptional Chevalley groups.
Proposition A.2. Fix a saturated fusion system F over a p-group S and an associated
b ) be such that each E ∈ Z(F
b ) is F -conjugate
centric linking system L. Let E1 , . . . , Ek ∈ Z(F
to Ei for some unique i. For each i, set Pi = CS (Ei ) and Zi = Z(Pi ). Then the following
hold.
b ) = ∅), then Ker(µL ) = 1.
(a) If k = 0 (Z(F
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
81
(b) If k = 1, E1 E S, and AutF (Ω1 (Z(S))) = 1, then Ker(µL ) = 1.
Q
(c) Assume, for each (gi )ki=1 ∈ ki=1 CZi (AutS (Pi )), that there is g ∈ CZ(S) (AutF (S)) such
that gi ∈ g · CZi (AutF (Pi )) for each i. Then Ker(µL ) = 1.
Proof. We first show that (a) and (b) are special cases of (c), and then prove (c). That (a)
follows from (c) is immediate.
(b) If k = 1, E1 E S, and AutF (Ω1 (Z(S))) = 1, then the group OutF (S) of order prime
to p acts trivially on Ω1 (Z(S)), and hence acts trivially on Z(S) (cf. [G, Theorem 5.2.4]).
Also, P1 = CS (E1 ) E S, so CZ1 (AutS (P1 )) = Z(S) = CZ(S) (AutF (S)), and Ker(µL ) = 1 by
(c).
(c) Fix a class [α] ∈ Ker(µL ). By [AOV, Proposition 4.2], there is an automorphism
α ∈ AutItyp (L) in the class [α] such that αS = IdAutL (S) . By the same proposition, there
are elements gP ∈ CZ(P ) (AutS (P )), defined for each P ≤ S which is fully normalized and
F -centric, such that
(i)
(ii)
αP ∈ Aut(AutL (P )) is conjugation by δP (gP );
αP = Id if and only if gP ∈ CZ(P ) (AutF (P )); and
(iii) if Q < P are both fully normalized and F -centric, then
gP ≡ gQ
(mod CZ(Q)(NAutF (P ) (Q))).
Furthermore,
(iv) [α] = 1 ∈ Outtyp (L) if and only if there is g ∈ CZ(S) (AutF (S)) such that gP ∈
g · CZ(P ) (AutF (P )) for each P < S such that P is F -essential and P = CS (Ω1 (Z(P ))).
Set gi = gPi (1 ≤ i ≤ k) for short.
By hypothesis, we can assume there is an element g ∈ CZ(S) (AutF (S)) such that gi ∈
g·CZ(Pi ) (AutF (Pi )) for each i. Upon replacing α by its composite with c−1
δS (g) , we can assume
gi ∈ CZ(Pi ) (AutF (Pi )), and hence αPi = IdAutL (Pi ) for each i.
We claim that αP = Id for all P ≤ S, and hence that [α] = 1 by (iv) and (ii). Assume
otherwise, and choose Q < S which is fully normalized and of maximal order among all
subgroups such that αQ 6= Id. Thus αR = Id for all R ≤ S with |R| > |Q|. By Alperin’s
fusion theorem (cf. [AKO, Theorem I.3.6]), Q is F -essential, and α is the identity on
MorL (P, P ∗) for all P, P ∗ ∈ Ob(L) such that |P |, |P ∗| > |Q|. Also, for each Q∗ ∈ QF , there
is (by Alperin’s fusion theorem again) an isomorphism χ ∈ IsoL (Q, Q∗ ) which is a composite
of isomorphisms each of which extends to an isomorphism between strictly larger subgroups,
and hence αQ,Q∗ (χ) = χ. Thus
Q∗ ∈ QF , Q∗ fully normalized
=⇒
αQ∗ 6= Id .
(1)
Set E = Ω1 (Z(Q)). Let ϕ ∈ HomF (NS (E), S) be such that ϕ(E) is fully normalized (cf.
[AKO, Lemma I.2.6(c)]). Then NS (Q) ≤ NS (E), so |NS (ϕ(Q))| ≥ |NS (Q)|, and ϕ(Q) is
fully normalized since Q is. Since αQ∗ 6= Id by (1), we can replace Q by Q∗ and E by E ∗ ,
and arrange that Q and E are both fully normalized in F (and Q is still F -essential).
Set Γ = AutF (Q), and set
Γ0 = CΓ (E) = ϕ ∈ AutF (Q) ϕ|E = IdE
Γ1 = ϕ ∈ AutF (Q) ϕ = ϕ|Q for some ϕ ∈ HomF (R, S), R > Q .
82
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Let πQ : AutL (Q) −−−→ AutF (Q) be the homomorphism induced by the functor π. For each
−1
ϕ ∈ Γ = AutF (Q), and each ψ ∈ πQ
(ϕ), we have
ϕ(gQ ) = gQ
⇐⇒
[ψ, δQ (gQ )] = Id
⇐⇒
αQ (ψ) = ψ :
(2)
the first by axiom (C) in the definition of a linking system (see, e.g., [AKO, Definition
III.4.1]) and since δQ is injective, and the second by point (i) above.
Now, AutS (Q) ≤ Γ1 , since each element of AutS (Q) extends to NS (Q) and NS (Q) > Q
(see [Sz1, Theorem 2.1.6]). Hence
Γ0 Γ1 ≤ O p (Γ0 ) · AutS (Q) · Γ1 = O p (Γ0 )Γ1 .
For each ϕ ∈ Γ0 of order prime to p, ϕ|Z(Q) = IdZ(Q) since ϕ is the identity on E =
Ω1 (Z(Q)) (cf. [G, Theorem 5.2.4]). Thus gQ ∈ CZ(Q) (O p (Γ0 )). If ϕ ∈ AutF (Q) extends
to ϕ ∈ HomF (R, S) for some R > Q, then by the maximality of Q, α(ψ) = ψ for each
ψ ∈ MorL (R, S) such that π(ψ) = ϕ, and since α commutes with restriction (it sends
−1
inclusions to themselves), αQ is the identity on ψ|Q,Q ∈ πQ
(ϕ). So by (2), ϕ(gQ ) = gQ .
Thus ϕ(gQ ) = gQ for all ϕ ∈ Γ1 . Since αQ 6= Id by assumption, there is some ϕ ∈ AutF (Q)
such that ϕ(gQ ) 6= gQ (by (2) again), and we conclude that
gQ ∈ CZ(Q) (Γ0 Γ1 )
and
Γ0 Γ1 < Γ = AutF (Q) .
(3)
Set Q∗ = NCS (E) (Q) ≥ Q. Then AutQ∗ (Q) = Γ0 ∩ AutS (Q) ∈ Sylp (Γ0 ) since AutS (Q) ∈
Sylp (Γ), and by the Frattini argument, Γ = NΓ (AutQ∗ (Q))Γ0 . If Q∗ > Q, then for each ϕ ∈
NΓ (AutQ∗ (Q)), ϕ extends to ϕ ∈ AutF (Q∗ ) by the extension axiom. Thus NΓ (AutQ∗ (Q)) ≤
Γ1 in this case, so Γ = Γ1 Γ0 , contradicting (3). We conclude that Q∗ = Q.
The homomorphism Γ = AutF (Q) −−−→ AutF (E) induced by restriction is surjective by
the extension axiom, so AutF (E) ∼
= Γ/Γ0 . By [AKO, Proposition I.3.3(b)], Γ1 /Inn(Q) is
strongly p-embedded in Γ/Inn(Q) = OutF (Q); and Γ0 Γ1 < Γ by (3). Also, p ∤ |Γ0 /Inn(Q)|,
since otherwise we would have Γ1 ≥ NΓ (T ) for some T ∈ Sylp (Γ0 ), in which case Γ1 Γ0 ≥
NΓ (T )Γ0 = Γ by the Frattini argument. Thus Γ1 Γ0 /Γ0 is strongly p-embedded in Γ/Γ0 ∼
=
AutF (E) by Lemma A.1(b).
Now, CS (E) = Q since NCS (E) (Q) = Q (cf. [Sz1, Theorem 2.1.6]). Thus Ω1 (Z(CS (E))) =
b ). Then E ∈ (Ei )F for some unique 1 ≤ i ≤ k,
Ω1 (Z(Q)) = E, and we conclude that E ∈ E(F
F
and Q ∈ (Pi ) by the extension axiom (and since E and Ei are both fully centralized). But
then αPi 6= Id by (1), contradicting the original assumption about αPi . We conclude that
α = Id.
A.1. Classical groups of Lie type in odd characteristic.
Throughout this subsection, we fix an odd prime power q and an integer n ≥ 1. We want
to show Ker(µG ) = 1 when G is one of the quasisimple classical groups of universal type
over Fq . By Theorem 1.8(d), we need not consider the unitary groups.
Proposition A.3. Fix an odd prime power q. Let G be isomorphic to one of the quasisimple
groups SLn (q), Spn (q) (n = 2m), or Spin±
n (q) (n ≥ 3). Then Ker(µG ) = 1.
b = Aut(V, b) be such that G = [G,
b G]
b if G ∼
Proof. Let V , b, and G
= SLn (q),
= Spn (q) or G ∼
±
b G]
b for some z ∈ Z(G)
b if G ∼
Spin
(q)
(where
z
∈
Z(G)).
Thus
V is a
and G/hzi = [G,
=
n
vector space of dimension n over the field K = Fq , b is a trivial symplectic, or quadratic
b is one of the groups GLn (q), Sp2n (q), or O ± (q).
form, and G
n
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
83
b ) for short.
Fix S ∈ Syl2 (G), and set F = FS (G). Set Zb = Z(F
Case 1: Assume G = Spin(V, b), where b is nondegenerate and symmetric. Set Z = Z(G),
and let z ∈ Z be such that G/hzi = Ω(V, b). We claim that Zb = ∅ in this case, and hence
that Ker(µG ) = 1 by Proposition A.2(a).
Lm
Fix an elementary abelian 2-subgroup E ≤ G where E ≥ Z. Let V =
i=1 Vi be the
decomposition as a sum of eigenspaces for the action of E on V . Fix indices j, k ∈ {1, . . . , m}
such that either dim(Vj ) ≥ 2, or the subspaces have the same discriminant (modulo squares).
(Since dim(V ) ≥ 3, this can always be done.) Then there is an involution γ ∈ SO(V, b) such
that γ(Vi ) = Vi for all i, γ|Vi = Id for i 6= j, k, det(γ|Vj ) = det(γ|Vk ) = −1, and such that the
(−1)-eigenspace of γ has discriminant a square. This last condition ensures that γ ∈ Ω(V, b)
(cf. [LO, Lemma A.4(a)]), so we can lift it to g ∈ G. Then for each x ∈ E, cg (x) = x if
x has the same eigenvalues on Vj and Vk , and cg (x) = zx otherwise (see, e.g., [LO, Lemma
A.4(c)]). Since z is fixed by all elements of AutF (E), cg ∈ O2 (AutF (E)), and hence AutF (E)
b
has no strongly 2-embedded subgroups by Lemma A.1(a). Thus E ∈
/ Z.
Case 2: Now assume G is linear or symplectic, and fix S ∈ Syl2 (G). For each V =
L
{V1 , . . . , Vk } such that V = ki=1 Vi , and such that Vi ⊥ Vj for i 6= j if G is symplectic, set
E(V) = ϕ ∈ G ϕ|Vi = ±Id for each i .
b and let V =
We claim that each subgroup in Zb has this form. To see this, fix E ∈ Z,
{V1 , . . . , Vk } be the eigenspaces for the nonzero characters of E. Then E ≤ E(V), V =
Lk
b (E) is the
i=1 Vi , and this is an orthogonal decomposition if G is symplectic. Also, CG
product of the groups Aut(Vi , b|Vi ). Since E = Ω1 (Z(P )) where P = CS (E), E contains the
2-torsion in the center of CG (E), and thus E = E(V). Furthermore, the action of S on each
Vi must be irreducible (otherwise Ω1 (Z(CS (E))) > E), so dim(Vi ) is a power of 2 for each i.
Again assume E = E(V) ∈ Zb for some V. Then AutGb (E) is a product of symmetric
Qgroups:
if V contains ni subspaces of dimension i for each i ≥ 1, then AutGb (E(V)) ∼
= i≥1 Σni .
Each such permutation can be realized by a self map of determinant one (if G is linear), so
AutG (E) = AutGb (E). Since AutG (E) contains a strongly 2-embedded subgroup by definition
of Zb (and since a direct product of groups of even order contains no strongly 2-embedded
subgroup), AutG (E) = AutGb (E) ∼
= Σ3 .
Write n = dim(V ) = 2k0 + 2k1 L
+ . . . + 2km , where 0 ≤ k0 < k1 < · · · < km . There is an
(orthogonal) decomposition V = m
i=0 Vm , where S acts irreducibly on each Vi , and where
dim(Vi ) = 2ki (see [CF, Theorem 1]). For each 1 ≤ i ≤ m, fix an (orthogonal) decomposition
Wi of Vi whose components have dimensions 2ki−1 , 2ki−1 , 2ki−1 +1 , . . . , 2ki −1 , and set
Vi = {Vj | j 6= i} ∪ Wi
and Ei = E(Vi ). Thus Vi contains exactly three subspaces of dimension 2ki−1 , and the dimenb Conversely, by
sions of the other subspaces are distinct. Hence AutG (Ei ) ∼
= Σ3 , and Ei ∈ Z.
the above analysis (and since the conjugacy class of E ∈ Zb is determined by the dimensions
of its eigenspaces), each subgroup in Zb is G-conjugate to one of the Ei .
For each 1 ≤ i ≤ m, set Pi = CS (Ei ) and Zi = Z(Pi ) (so Ei = Ω1 (Zi )). Since each
element of NG (Pi ) = NG (Ei ) permutes members of Vi of equal dimension, and the elements
84
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
of NS (Pi ) do so only within each
Zi = z ∈ G
CZi (AutS (Pi )) = z ∈ Zi
CZi (AutG (Pi )) = z ∈ Zi
of the Vj , we have
(z)
(z)
z|X = λX IdX for all X ∈ Vi , some λX ∈ O2 (F×
q )
(z)
(z)
λ Xi = λ X ′
(4)
i
(z)
(z)
(z)
λXi = λX ′ = λVi−1 ,
i
where Xi , Xi′ , and Vi−1 are the three members of the decomposition Vi of dimension 2ki−1
(and Xi , Xi′ ∈ Wi ).
Qm
(gi )
(gi )
Fix (gi )m
i=1 ∈
i=1 CZi (AutS (Pi )). Then gi ∈ CZi (AutG (Pi )) if and only if λVi−1 = λXi .
b such that g|V = ηi · Id for each i, where the ηi ∈ O2 (F× ) are chosen so
Choose g ∈ G
q
i
(g )
(g )
k
i
that ηi /ηi−1 = λXii /λVi−1
for each 1 ≤ i ≤ m. If G is linear, then det(g) = θ2 0 for some
−2k0 /n
θ ∈ O2 (F×
IdV (recall k0 = v2 (n)) we can assume g ∈ G.
q ), and upon replacing g by g ◦ θ
Then g ∈ CZ(S) (AutG (S)) since it is a multiple of the identity on each Vi and has 2-power
order. By construction and (4), g ≡ gi (mod CZi (AutG (Pi ))) for each i; so Ker(µG ) = 1 by
Proposition A.2(c).
A.2. Exceptional groups of Lie type in odd characteristic.
Throughout this subsection, q0 is an odd prime, and q is a power of q0 . We show that
Ker(µG ) = 1 when G is one of the groups G2 (q), F4 (q), E6 (q), E7 (q), or E8 (q) and is of
universal type.
The following proposition is a special case of [GLS3, Theorem 2.1.5], and is stated and
proven explicitly in [O2, Proposition 8.5]. It describes, in many cases, the relationship
between conjugacy classes and normalizers in a connected algebraic group and those in the
subgroup fixed by a Steinberg endomorphism.
Proposition A.4. Let G be a connected algebraic group over Fq0 , let σ be a Steinberg
endomorphism of G, and set G = CG (σ). Let H ≤ G be any subgroup, and let H be the
set of G-conjugacy classes of subgroups G-conjugate to H. Let NG (H) act on π0 (CG (H)) by
sending g to xgσ(x)−1 (for x ∈ NG (H)). Then there is a bijection
∼
=
ω : H −−−−−−−→ π0 (CG (H))/NG (H),
defined by setting ω([xH]) = [x−1 σ(x)] whenever xH ≤ CG (σ). Also, for each x ∈ G such that
H ≤ G, AutG (xH) is isomorphic to the stabilizer of [x−1 σ(x)] ∈ π0 (CG (H))/CG (H) under
the action of AutG (H) on this set.
x
Since we always assume G is of universal type in this section, the group G = CG (σ) of
′
Proposition A.4 is equal to the group G = O q0 (CG (σ)) of Definition 2.1 and Notation 2.2.
The following definitions will be useful when applying Proposition A.4. For any finite
group G, set
SE(G) = H ≤ G H has a strongly 2-embedded subgroup
(
min [G : H] H ∈ SE(G)
if SE(G) 6= ∅
δ(G) =
∞
if SE(G) = ∅.
Thus by Proposition A.4, if H < G is such that |π0 (CG (H))| > δ(OutG (H)), then no
subgroup H ∗ ≤ CG (σ) which is G-conjugate to H has the property that AutCG (σ) (H ∗ ) has
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
85
a strongly 2-embedded subgroup. The next lemma provides some tools for finding lower
bounds for δ(G).
Lemma A.5. (a) For any finite group G, δ(G) ≥ |O2(G)| · δ(G/O2(G)).
(b) If G = G1 × G2 is finite, and δ(Gi ) < ∞ for i = 1, 2, then
δ(G) = min δ(G1 ) · η(G2 ) , δ(G2 ) · η(G1 ) ,
where η(Gi ) is the smallest index of any odd order subgroup of Gi .
(c) If δ(G) < ∞, and there is a faithful F2 [G]-module V of rank n, then 2v2 (|G|)−[n/2] δ(G).
(d) More concretely, δ(GL3 (2)) = 28, δ(GL4 (2)) = 112, δ(GL5 (2)) = 28 · 7 · 31, and
−
+
4
6
δ(SO+
4 (2)) = 2 = δ(SO4 (2)). Also, 2 ≤ δ(SO6 (2)) < ∞ and 2 ≤ δ(SO7 (2)) < ∞.
Proof. (a) If H ∈ SE(G), then H ∩O2 (G) = 1 by Lemma A.1(a). Hence there is a subgroup
H ∗ ≤ G/O2 (G) isomorphic to H, and
[G : H] = |O2 (G)| · [G/O2 (G) : H ∗ ] ≥ |O2 (G)| · δ(G/O2 (G)) .
(b) If a finite group H has a strongly 2-embedded subgroup, then so does its direct product
with any odd order group. Hence δ(G) ≤ δ(Gi )η(G3−i ) for i = 1, 2.
Assume H ≤ G has a strongly 2-embedded subgroup K < H. Set Hi = H ∩ Gi for
i = 1, 2. Since all involutions in H are H-conjugate (see [Sz2, 6.4.4]), H1 and H2 cannot
both have even order. Assume |H2 | is odd. Let pr1 be projection onto the first factor.
If pr1 (K) = pr1 (H), then there is x ∈ (HrK) ∩ H2 , and this commutes with all Sylow
2-subgroups of H since they lie in G1 , contradicting the assumption that K is strongly 2embedded in H. Thus pr1 (K) < pr1 (H). Then pr1 (H) has a strongly 2-embedded subgroup
by Lemma A.1(b), and hence
[G : H] = [G1 : pr1 (H)] · [G2 : H2 ] ≥ δ(G1 ) · η(G2 ) .
So δ(G) ≥ δ(Gi )η(G3−i ) for i = 1 or 2.
(c) This follows from [OV, Lemma 1.7(a)]: if H < G has a strongly 2-embedded subgroup,
T ∈ Syl2 (H), and |T | = 2k , then dim(V ) ≥ 2k.
+
∼
(d) The formulas for δ(SO±
4 (2)) hold since SO4 (2) = Σ3 ≀C2 contains a subgroup isomorphic
−
2
to C3 ⋊ C4 and SO4 (2) ∼
= Σ5 a subgroup isomorphic to A5 . Since 4|δ(GL3 (2)) by (c), and
since 7|δ(GL3 (2)) (there are no subgroups of order 14 or 42), we have 28|δ(GL3 (2)), with
equality since Σ3 has index 28. The last two (very coarse) estimates follow from (c), and the
6- and 7-dimensional representations of these groups.
Fix n = 4, 5, and set Gn = GLn (2). Assume H ≤ Gn has a strongly embedded subgroup,
where 7 |H| or 31 |H|. By (c), 24 |δ(G4 ) and 28 |δ(G5 ), and thus 8 ∤ |H|. If H is almost simple,
then H ∼
= A5 by Bender’s theorem (see [Sz2, Theorem 6.4.2]), contradicting the assumption
about |H|. So by the main theorem in [A1], H must be contained in a member of one of the
classes Ci (1 ≤ i ≤ 8) defined in that paper. One quickly checks that since (7 · 31, |H|) 6= 1,
H is contained in a member of C1 . Thus H is reducible, and since O2 (H) = 1, either H is
isomorphic to a subgroup of GL3 (2) × GLn−3 (2), or n = 5 and H < GL4 (2). By (b) and
since 7 |δ(GL3 (2)), we must have H ∼
= Σ3 × (C7 ⋊ C3 ), in which case |H| < 180 = |GL2 (4)|.
Thus 7|δ(Gn ) for n = 4, 5, and 31|δ(G5 ). Since GL4 (2) contains a subgroup isomorphic to
GL2 (4) ∼
= C3 × A5 , we get δ(G4 ) = 24 · 7 and δ(G5 ) = 28 · 7 · 31.
We illustrate the use of the above proposition and lemma by proving the injectivity of µG
when G = G2 (q).
86
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Proposition A.6. If G = G2 (q) for some odd prime power q, then Ker(µG ) = 1.
Proof. Assume q is a power of the prime q0 , set G = G2 (Fq0 ), and fix a maximal torus T .
We identify G = CG (ψq ), where ψq is the field automorphism, and acts via (t 7→ tq ) on T .
b S (G)).
Fix S ∈ Syl2 (G), and set Zb = Z(F
Let E ∼
= C 2 be the 2-torsion subgroup of T . By Proposition 2.5, C (V ) = T hθi where
2
G
θ ∈ NG (T ) inverts the torus. Thus by Proposition A.4, there are two G-conjugacy classes of
subgroups G-conjugate to E, represented by E ± (E + = E), where AutG (E ± ) = Aut(E ± ) ∼
=
Σ3 and CG (E ± ) = (Cq∓1 )2 ⋊ C2 . The subgroups in one of these classes have centralizer in S
b while those in the other class do lie in Z.
b The latter
isomorphic to C23 , hence are not in Z,
2
also have normalizer of order 12(q ± 1) and hence of odd index in G, and thus are normal
in some choice of Sylow 2-subgroup.
By [Gr, Table I], for each nontoral elementary abelian 2-subgroup E ≤ G, rk(E) = 3,
CG (E) = E, and AutG (E) ∼
= GL3 (2). By Proposition A.4, and since δ(AutG (E)) = 28 >
|CG (E)| by Lemma A.5, AutG (E) contains no strongly 2-embedded subgroup, and thus
b
E∈
/ Z.
Thus Zb is contained in a unique G-conjugacy class of subgroups of rank 2, and Ker(µG ) = 1
by Proposition A.2(b).
Throughout the rest of this section, fix an odd prime power q, and let G be one of the
groups F4 , E6 , E7 , or E8 .
Hypotheses A.7. Assume G = G(Fq0 ) and G ∼
= G(q), where q is a power of the odd prime
q0 , and where G = F4 , E6 , E7 , or E8 and is of universal type. Fix a maximal torus T < G.
(I) Set T(2) = {t ∈ T | t2 = 1}. Let 2A and 2B denote the two G-conjugacy classes of
noncentral involutions in G, as defined in [Gr, Table VI], except that when G = E7 , they
denote the classes labelled 2B and 2C, respectively, in that table. For each elementary
abelian 2-subgroup E < G, define
qE : E −−−−−−→ F2
by setting q(x) = 0 if x ∈ 2B ∪ {1}, and q(x) = 1 if x ∈ 2A ∪ (Z(G)r1).
(II) Assume G = CG (ψq ), where ψq is the field endomorphism with respect to some root
structure with maximal torus T . Thus ψq (t) = tq for all t ∈ T . Fix S ∈ Syl2 (G), and
b S (G)).
set Zb = Z(F
By [Gr, Lemma 2.16], qT(2) is a quadratic form on T(2) in all cases, and hence qE is
quadratic for each E ≤ T(2) . In general, qE need not be quadratic when E is not contained
in a maximal torus. In fact, Griess showed in [Gr, Theorems 7.3, 8.2, & 9.2] that in many
(but not all) cases, E is contained in a torus if and only if qE is quadratic (cx(E) ≤ 2 in his
terminology).
With the above choices of notation for noncentral involutions, all of the inclusions F4 ≤
E6 ≤ E7 ≤ E8 restrict to inclusions of the classes 2A and of the classes 2B. This follows
since the forms are quadratic, and also (for E7 < E8 ) from [Gr, Lemma 2.16(iv)].
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
87
Lemma A.8. Assume Hypotheses A.7, and let b be the bilinear form associated to q. Define
V0 = v ∈ T(2) b(v, T(2) ) = 0, q(v) = 0
γx = v 7→ v + b(v, x)x ∈ Aut(T(2) , q)
for x ∈ T(2) with q(x) = 1, q 6⊥ T(2)
Then the following hold.
(a) AutG (T(2) ) = Aut(T(2) , q).
⊥
(b) For each nonisotropic x ∈ T(2) rT(2)
, γx is the restriction to T(2) of a Weyl reflection on
T . If α ∈ Σ is such that γx = wα |T(2) , then θα (v) = (−1)b(x,v) for each v ∈ T(2) .
(c) If G = Er (r = 6, 7, 8), then q is nondegenerate (V0 = 0), and the restriction to T(2) of
⊥
each Weyl reflection is equal to γx for some nonisotropic x ∈ T(2) rT(2)
.
(d) If G = F4 , then dim(V0 ) = 2, and q(v) = 1 for all v ∈ T(2) rV0 .
Proof. (a) Since AutG (T(2) ) has to preserve G-conjugacy classes, it is contained in Aut(T(2) , q).
Equality will be shown while proving (c) and (d).
(c) If G = Er for r = 6, 7, 8, then q is nondegenerate by [Gr, Lemma 2.16]. Hence the
only orthogonal transvections are of the form γx for nonisotropic x, and each Weyl reflection
restricts to one of them. By a direct count (using the tables in [Brb]), the number of
pairs {±α} of roots in G (hence the number of Weyl reflections) is equal to 36, 63, or
⊥
120, respectively. This is equal to the number of nonisotropic elements in T(2) r T(2)
=
T(2) r Z(G) (see the formula in [Ta, Theorem 11.5] for the number of isotropic elements).
So all transvections are restrictions of Weyl reflections, and AutG (T(2) ) = Aut(T(2) , q).
(d) Assume G = F4 . Then dim(V0 ) = 2 and q−1 (1) = T(2) rV0 by [Gr, Lemma 2.16]. Thus
|Aut(T(2) , q)| = 42 · |GL2 (2)|2 = 26 · 32 = 12 |W | (see [Brb, Planche VIII]), so AutW (T(2) ) =
Aut(T(2) , q) since W also contains −Id.
There are three conjugacy classes of transvections γ ∈ Aut(T(2) , q): one of order 36 containing those where γ|V0 6= Id (and hence [γ, T(2) ] ≤ V0 ), and two of order 12 containing
those where γ|V0 = Id (one where [γ, T(2) ] ≤ V0 and one where [γ, T(2) ] V0 ). Since there
are two W -orbits of roots (long and short), each containing 12 pairs ±α, the corresponding
Weyl reflections must restrict to the last two classes of transvections, of which one is the set
of all γx for x ∈ T(2) r V0 .
(b) We showed in the proofs of (c) and (d) that each orthogonal transvection γx is the
restriction of a Weyl reflection. If γx = wα |T(2) for some root α ∈ Σ, then θα ∈ Hom(T , F×
q0 )
(Lemma 2.4(c)), so [T(2) : Ker(θα |T(2) )] ≤ 2. Also, Ker(θα ) ≤ CT (wα ) by Lemma 2.4(e),
so Ker(θα |T(2) ) ≤ CT(2) (wα ) = CT(2) (γx ) = x⊥ , with equality since [T(2) : x⊥ ] = 2. Since
θα (T(2) ) ≤ {±1}, it follows that θα (v) = (−1)b(x,v) for each v ∈ T(2) .
b
We are now ready to list the subgroups in Z(G(q))
in all cases. The proof of the following
lemma will be given at the end of the section.
b
Lemma A.9. Let G = G(Fq0 ) and G = G(q) be as in Hypotheses A.7. Assume E ∈ Z(G).
Then either G 6= E7 , rk(E) = 2, and qE = 0; or G = E7 , Z = Z(G) ∼
= C2 , and E = Z × E0
where rk(E0 ) = 2 and qE0 = 0. In all cases, AutG (E) ∼
= Σ3 .
Proof. This will be shown in Lemmas A.14 and A.15.
88
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
The next two lemmas will be needed to apply Proposition A.2(b) to these groups. The
first is very elementary.
Lemma A.10. Let V be an F2 -vector space of dimension k, and let q : V −−−→ F2 be a
quadratic form on V . For m ≥ 1 such that k > 2m, the number of totally isotropic subspaces
of dimension m in V is odd.
Proof. This will be shown by induction on m, starting with the case m = 1. Since k ≥ 3,
there is an orthogonal splitting V = V1 ⊥ V2 where V1 , V2 6= 0. Let ki be the number of
isotropic elements in Vi (including 0), and set ni = |Vi |. The number of isotropic elements
in V is then k1 k2 + (n1 − k1 )(n2 − k2 ), and is even since the ni are even. The number of
1-dimensional isotropic subspaces is thus odd.
Now fix m > 1 (such that k > 2m), and assume the lemma holds for subspaces of
dimension m − 1. For each isotropic element x ∈ V , a subspace E ≤ V of dimension m
containing x is totally isotropic if and only if E ≤ x⊥ and E/hxi is isotropic in x⊥ /hxi with
the induced quadratic form. By the induction hypothesis, and since
2 · dim(E/hxi) = 2(m − 1) < k − 2 ≤ dim(x⊥ /hxi),
the number of isotropic subspaces of dimension m which contain x is odd. Upon taking the
sum over all x, and noting that each subspace has been counted 2m − 1 times, we see that
the number of isotropic subspaces of dimension m is odd.
Lemma A.11. Assume Hypotheses A.7(I). Let σ be a Steinberg endomorphism of G such
that for some ε = ±1, σ(t) = tεq for each t ∈ T . Set G = CG (σ). Fix E ≤ T(2) of rank
2 such that q(E) = 0. Then the set of subgroups of G which are G-conjugate to E, and
the set of subgroups which are G-conjugate to E, both have odd order and contain all totally
isotropic subgroups of rank 2 in T(2) .
Proof. Let X ⊇ X be the sets of subgroups of G which are G-conjugate to E or G-conjugate
to E, respectively. Let X0 be the subset of all totally isotropic subgroups of T(2) of rank 2. If q
is nondegenerate, then by Witt’s theorem (see [Ta, Theorem 7.4]), AutW (T(2) ) = Aut(T(2) , q)
permutes X0 transitively, and hence all elements in X0 are G-conjugate to E by Lemma 2.9.
If in addition, dim(T(2) ) ≥ 5, then |X0 | is odd by Lemma A.10. Otherwise, by Lemma
A.8(c,d), G = F4 and X0 = {E}. Thus in all cases, X0 ⊆ X and |X0 | is odd.
Assume G = E6 . Then CG (T(2) ) = T by Proposition 2.5. Consider the conjugation action
of T(2) on X, and let X1 be its fixed point set. Since T(2) ≤ G by the assumptions on σ, this
action also normalizes X. For F ∈ X1 , either the action of T(2) fixes F pointwise, in which
case F ∈ X0 , or there are x, y ∈ F such that [x, T(2) ] = 1 and [y, T(2) ] = hxi. In particular,
cy ∈ AutG (T(2) ) = SO(T(2) , q). For each v ∈ T(2) such that [y, v] = x, q(v) = q(vx) and
q(x) = 0 imply x ⊥ v, so x ⊥ T(2) since T(2) is generated by those elements. This is impossible
since q is nondegenerate by Lemma A.8(c), and thus X1 = X0 .
Now assume G = F4 , E7 , or E8 . Then −Id ∈ W , so there is θ ∈ NG (T ) which inverts
T . Then CG (T(2) ) = T hθi. By the Lang-Steinberg theorem, there is g ∈ G such that
g −1 σ(g) ∈ θT ; then σ(gtg −1) = gt∓q g −1 for t ∈ T , and thus σ acts on gT g −1 via t 7→ t∓q .
We can thus assume T was chosen so that G ∩ T = CT (σ) contains the 4-torsion subgroup
T (4) ≤ T . Let X1 ⊆ X be the fixed point set of the conjugation action of T (4) on X.
For F ∈ X1 , either the action of T (4) fixes F pointwise, in which case F ∈ X0 , or there
are x, y ∈ F such that [x, T (4) ] = 1 and [y, T (4) ] = hxi. But then [F, T ∗(4) ] = 1 for some
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
89
T ∗(4) < T (4) of index two, [F, T(2) ] = 1 implies F ≤ T(2) hθi; and F ≤ T(2) since no element in
T (4) rT(2) commutes with any element of T(2) θ. So X1 = X0 in this case.
Thus in both cases, X0 is the fixed point set of an action of a 2-group on X which normalizes
X. Since |X0 | is odd, so are |X| and |X|.
We are now ready to prove:
Proposition A.12. Fix an odd prime power q. Assume G is a quasisimple group of universal
type isomorphic to G2 (q), F4 (q), E6 (q), E7 (q), or E8 (q). Then Ker(µG ) = 1.
Proof. This holds when G ∼
= G2 (q) by Proposition A.6, so we can assume Hypotheses A.7.
Let X be the set of all elementary abelian 2-subgroups E ≤ G such that either G 6= E7 ,
rk(E) = 2, and qE = 0; or G = E7 , rk(E) = 3, and E = Z(G) × E0 where qE0 = 0. By
b
Lemma A.11, |X| is odd. In all cases, by Lemma A.9, Z(G)
⊆ X. By Proposition A.2(a,b),
b
b
to prove µG is injective, it remains to show that if Z(G) 6= ∅, then Z(G)
has odd order and
is contained in a single G-conjugacy class, and AutG (Z(S)) = 1.
Fix E ∈ X such that E ≤ T(2) . We first claim that if G = F4 , E6 , or E7 , then CG (E)
is connected, and hence all elements in X are G-conjugate to E by Proposition A.4. If
G = E7 , then CG (E) is connected by [Gr, Proposition 9.5(iii)(a)]. If G = F4 or E6 , then for
x ∈ E, CG (x) ∼
= Spin9 (Fq0 ) or Fq0 ×C4 Spin10 (Fq0 ), respectively (see [Gr, Table VI]). Since
the centralizer of each element in the simply connected groups Spin9 (Fq0 ) and Spin10 (Fq0 ) is
connected [St3, Theorem 8.1], CG (E) is connected in these cases.
Now assume G = E8 . We can assume G = CG (ψq ), where ψq is the field automorphism;
in particular, ψq (t) = tq for t ∈ T . Fix x, y ∈ E such that E = hx, yi. By [Gr, Lemma
2.16(ii)], (T(2) , q) is of positive type (has a 4-dimensional totally isotropic subspace). Hence
E ⊥ = E × V1 S
× V2 , where dim(V
i ) = 2 and q(Vi r1) = 1 for i = 1, 2, and V1 ⊥ V2 . Thus
2
−1
(qE ⊥ ) (1) = i=1 (Vi r1) × E , and by Lemma A.8(b,c), these are the restrictions to T(2)
of Weyl reflections wα for α ∈ Σ such that E ≤ Ker(θα ). Also, CW (E) ∼
= W (D4 ) ≀ C2 . By
0
Proposition 2.5, CG (E) has type D4 × D4 and |π0 (CG (E))| = 2. More precisely, CG (E) =
(H 1 ×E H 2 )hδi, where H i ∼
= Spin8 (Fq0 ) and Z(H i ) = E for i = 1, 2, and conjugation by
δ ∈ NG (T ) exchanges V1 and V2 and hence exchanges H 1 and H 2 .
By Proposition A.4, the two connected components in the centralizer give rise to two
G-conjugacy classes of subgroups which are G-conjugate to E, represented by E and gEg −1
where g −1σ(g) lies in the nonidentity component of CG (E). Then CG (E) contains a subgroup
+
Spin+
8 (q)×C22 Spin8 (q) with index 8 (the extension by certain pairs of diagonal automorphisms
of the Spin+
8 (q)-factors, as well as an automorphism which switches the factors). So E =
b
Z(T ) for T ∈ Syl2 (CG (E)), and E ∈ Z(G).
Also, gyg −1 ∈ CG (gEg −1) if and only if
y ∈ CG (E) and τe(y) = y where τe = cg−1 σ(g) ◦ σ. Then τe switches the central factors in CG (E),
τ ) splits as a product of E times the group of elements which are
and the group CCG (E) (e
invariant after lifting τe to the 4-fold cover Spin8 (Fq0 )≀C2. Since gEg −1 intersects trivially with
the commutator subgroup of CG (gEg −1), Ω1 (Z(T )) > gEg −1 for any T ∈ Syl2 (CG (gEg −1))
b
b
(since Z(T ) ∩ [T, T ] 6= 1); and thus gEg −1 ∈
/ Z(G).
Thus Z(G)
is the G-conjugacy class of
E, and has odd order by Lemma A.11.
b
Thus, in all cases, if Z(G)
is nonempty, it has odd order and is contained in one Gb
conjugacy class. Also, Z(S) ≤ CE (AutS (E)) < E for E ∈ Z(G),
so either |Z(S)| = 2, or
90
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
G = E7 , Z(S) ∼
= C22 , and the three involutions in Z(S) belong to three different G-conjugacy
classes. Hence AutG (Z(S)) = 1.
It remains to prove Lemma A.9, which is split into the two Lemmas A.14 and A.15. The
next proposition will be used to show that certain elementary abelian subgroups are not in
b
Z.
Proposition A.13. Assume Hypotheses A.7. Let E ≤ T(2) and x ∈ T(2) rE be such that
the orbit of x under the CW (E)-action on T(2) has odd order. Then no subgroup of S which
b More generally, if E ≥ E is also elementary abelian, and is
is G-conjugate to E is in Z.
such that x is not CG (E)-conjugate to any element of E, then for any L E G which contains
b
{gxg −1 | g ∈ G} ∩ G, no subgroup of S which is G-conjugate to E is in Z.
Proof. In [O2], an elementary abelian p-subgroup E < G is called pivotal if Op (AutG (E)) =
1, and E = Ω1 (Z(P )) for some P ∈ Sylp (CG (E)). In particular, by Lemma A.1(a), the
subgroups in Zb are all pivotal. Note that T(2) ≤ G by Hypotheses A.7. By [O2, Proposition
8.9], no subgroup satisfying the above conditions can be pivotal, and hence they cannot be
b
in Z.
In the next two lemmas, we show that in all cases, E ∈ Zb implies rk(E) = 2 and qE = 0 if
G 6= E7 , with a similar result when G = E7 . We first handle those subgroups which are toral
(contained in a maximal torus in G), and then those which are not toral. By a 2Ak -subgroup
or subgroup of type 2Ak (2Bk -subgroup or subgroup of type 2Bk ) is meant an elementary
abelian 2-subgroup of rank k all of whose nonidentity elements are in class 2A (class 2B).
Lemma A.14. Assume Hypotheses A.7. Fix some E ∈ Zb which is contained in a maximal
torus of G. Then either G 6= E7 , rk(E) = 2, and qE = 0; or G = E7 , Z = Z(G) ∼
= C2 , and
E = Z × E0 where rk(E0 ) = 2 and qE0 = 0. In all cases, AutG (E) ∼
Σ
.
= 3
Proof. Set Z = O2 (Z(G)) ≤ T(2) . Thus |Z| = 2 if G = E7 , and |Z| = 1 otherwise. Recall
that AutG (T(2) ) = AutG (T(2) ) = Aut(T(2) , q) by Lemmas 2.9 and A.8(a).
The following notation will be used to denote isomorphism types of quadratic forms over
F2 . Let [n]± denote the isomorphism class of a nondegenerate form of rank n. When n
is even, [n]+ denotes the hyperbolic form (with maximal Witt index), and [n]− the form
with nonmaximal Witt index. Finally, a subscript “(k)” denotes sum with a k-dimensional
−
+
trivial form. By [Gr, Lemma 2.16], qT(2) has type [2]−
(2) , [6] , [7], or [8] when G = F4 , E6 ,
E7 , or E8 , respectively.
b Set
Fix E ≤ T(2) ; we want to determine whether E can be G-conjugate to an element of Z.
⊥
E1 = E ∩ E (the orthogonal complement taken with respect to q), and set E0 = Ker(qE1 ).
Note that E1 > E0 if G = E7 (E ≥ Z).
Assume first that E0 = 1. If G = F4 , then T(2) ∩ 2B is a CW (E)-orbit of odd order. If
G = Er and E1 = 1, then E × E ⊥ , E ⊥ is CW (E)-invariant, and hence there is 1 6= x ∈ E ⊥
whose CW (E)-orbit has odd order. If G = Er and rk(E1 ) = 1, then E ∩ E ⊥ = E1 , there
is an odd number of involutions in E ⊥ rE1 of each type (isotropic or not), and again there
is 1 6= x ∈ E ⊥ whose CW (E)-orbit has odd order. In all cases, x has the property that
CW (hE, xi) has odd index in CW (E). So by Proposition A.13, no subgroup of G which is
b
G-conjugate to E can be in Z.
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
91
Thus E0 6= 1. Set k = rk(E0 ). Then
π0 (CG (E)) = CW (E) wα α ∈ Σ, E ≤ Ker(θ(α))
(Proposition 2.5)
⊥
≤ CW (T(2) ) · CSO(T(2) ,q) (E) γv v ∈ 2A ∩ E
(Lemma A.8(a,b))
≤ CW (T(2) ) · CSO(T(2) ,q) (E0⊥ ) · CSO(E0⊥ ,q) (E) γv v ∈ 2A ∩ E ⊥ .
(5)
The first factor is easily described:
CW (T(2) ) = 2ε
(
1 if −Id ∈ W (if G = F4 , E7 , E8 )
where ε =
0 if −Id ∈
/ W (if G = E6 ).
(6)
We next claim that
CSO(T(2) ,q) (E0⊥ ) ≤ 2(2) ,
k
(7)
with equality except possibly when G = F4 . To see this, let F1 < T(2) be a subspace
complementary to E0⊥ . Each α ∈ CAut(T(2) ) (E0⊥ ) has the form α(x) = xψ(x) for some
ψ ∈ Hom(F1 , E0 ), and α is orthogonal if and
only if x ⊥ ψ(x) for each x. The space of such
k
homomorphisms has dimension at most 2 (corresponding to symmetric k ×k matrices with
zeros on the diagonal); with dimension equal to k2 if dim(F1 ) = dim(E0 ) (which occurs if q
is nondegenerate).
Write (E0 )⊥ = E × F2 , where E ⊥ = E0 × F2 and the form qF2 is nondegenerate. By
[Ta, Theorem 11.41], SO(F2 , qF2 ) is generated by transvections unless qF2 is of type [4]+ ,
in which case the reflections generate a subgroup of SO(F2 , qF2 ) ∼
= Σ3 ≀ C2 isomorphic to
Σ3 × Σ3 . Also, F2 is generated by nonisotropic elements except when qF2 is of type [2]+ ,
and when this is the case, all automorphisms of (E0 )⊥ which induce the identity on E and
on (E0 )⊥ /E0 are composites of transvections. (Look at the composites γvx ◦ γv for v ∈ F2
and x ∈ E0 .) Hence
CSO(E0⊥ ,q) (E) γv v ∈ 2A ∩ E ⊥ ≤ 2η
+
where η = 1 if qE ⊥ has type [4]+
(k) , η = k if qE ⊥ has type [2](k) , and η = 0 otherwise.
Together with (5), (6), and (7), this proves that
|π0 (CG (E))| ≤ 2(2)+ε+η
k
where ε ≤ 1.
(8)
Now, NG (E) ≤ CG (E)0 NG (T ) by the Frattini argument: each maximal torus which contains E lies in CG (E)0 and hence is CG (E)0 -conjugate to T . So each element of AutG (E) is
represented by a coset of T in NG (T ), and can be chosen to lie in G by Lemma 2.9. Thus
the action described in Proposition A.4 which determines the automizers AutG (E ∗ ) for E ∗
G-conjugate to E is the conjugation action of AutG (E) on the set of conjugacy classes in
π0 (CG (E)). In particular, this action is not transitive, since the identity is fixed.
Set ℓ = rk(E/E0 ) − 1 if G = E7 and ℓ = rk(E/E0 ) otherwise. Every automorphism of E
which induces the identity on E0 Z and on E/E0 is orthogonal, and hence the restriction of
an element of O2 (CW (E)). Thus |O2 (OutG (E))| ≥ 2kℓ . If E ∗ ∈ Z is G-conjugate to E, then
since AutG (E ∗ ) has a strongly 2-embedded subgroup, 2kℓ ≤ δ(AutG (E)) < π0 (CG (E))
by Proposition A.4 and Lemma A.5(a), with strict inequality since the action of NG (E)
on π0 (CG (E)) is not transitive. Together with (8), and since ε ≤ 1, this implies that
, and ℓ ≤ k−1
if η = 0. By definition, η = 0 whenever
kℓ ≤ k2 + η ≤ k2 + k. Thus ℓ ≤ k+1
2
2
rk(E1 /E0 ) = 1, which is the case if G = E7 or ℓ is odd. Since 2k + ℓ ≤ 8, we are thus left
with the following possibilities.
92
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
⊥
• If (k, ℓ) = (3, 2), then G = E8 , E has form of type [2]+
= E0 has trivial form,
(3) , so E
k
and η = 0. Thus kℓ 2 + η, so this cannot occur.
• If (k, ℓ) = (3, 1), then G = E8 , E has form of type ∗ + 3, and rk(E) = rk(E1 ) = 4.
Then AutG (E) ∼
= C23 ⋊ GL3 (2), so δ(AutG (E)) ≥ 23 · 28 by Lemma A.5(a,d). Since
|π0 (CG (E))| ≤ 16, this case is also impossible.
• If (k, ℓ) = (4, 0), then G = E8 and E = E0 is isotropic of rank 4. By Proposition 2.5
and Lemma A.8(c), CG (E)0 = T . By [CG, Proposition 3.8(ii)], π0 (CG (E)) is extraspecial
of order 27 and AutG (E) ∼
= GL4 (2). (This is stated for subgroups of E8 (C), but the
same argument applies in our situation.) In particular, π0 (CG (E)) has just 65 conjugacy
classes. Since δ(GL4 (2)) = 112 by Lemma A.5(d), Proposition A.4 implies that AutG (E ∗ )
cannot have a strongly 2-embedded subgroup.
• If (k, ℓ) = (3, 0), then E = Z ×E0 where dim(E0 ) = 3, and Aut (E) ∼
= GL3 (2). If G = E6
G
or E7 , then E ⊥ = E, and |π0 (CG (E))| ≤ 16 by (8).
If G = E8 , then (E ⊥ , qE ⊥ ) has type [2]+
(3) . By the arguments used to prove (8),
|CW (E)| = |CW (T(2) )| · |CSO(T(2) ,q) (E0⊥ )| · |CSO(E0⊥ ,q) (E)| = 2 · 23 · 27 = 211 .
Also, E ⊥ contains exactly 8 nonisotropic elements, they are pairwise orthogonal, and
hence determine 8 pairwise commuting transvections on T(2) . These extend to 8 Weyl
reflections which are pairwise commuting since no two can generate a dihedral subgroup
of order 8 (this would imply two roots of different lengths). Hence by Proposition 2.5,
CG (E)0 has type (A1 )8 and |π0 (CG (E))| = 211 /28 = 23 . Since δ(GL3 (2)) = 28 by Lemma
A.5(c), this case cannot occur.
• If (k, ℓ) = (2, 0), then E = Z × E0 where dim(E0 ) = 2. Then E is as described in the
statement of the lemma.
It remains to handle the nontoral elementary abelian subgroups.
Lemma A.15. Assume Hypotheses A.7. Let E ≤ G be an elementary abelian 2-group which
b
is not contained in a maximal torus of G. Then E ∈
/ Z.
Proof. To simplify notation, we write K = Fq0 . Set Z = O2 (Z(G)) ≤ T(2) . Thus |Z| = 2 if
G = E7 , and |Z| = 1 otherwise. The maximal nontoral subgroups of G are described in all
cases by Griess [Gr].
(A) If G = F4 or E6 , then by [Gr, Theorems 7.3 & 8.2], G contains a unique conjugacy
class of maximal nontoral elementary abelian 2-subgroups, represented by W5 of rank
five. There is a subgroup W2 ≤ W5 of rank two such that W5 ∩ 2A = W5 rW2 . Also,
AutG (E5 ) = Aut(E5 , qE5 ): the group of all automorphisms of W5 which normalize W2 .
A subgroup E ≤ W5 is nontoral if and only if it contains a 2A3 -subgroup.
When G = F4 , we can assume W5 = T(2) hθi, where θ ∈ NG (T ) inverts the torus.
(B) If G = E7 , then by [Gr, Theorem 9.8(i)], G contains a unique maximal nontoral elementary abelian 2-subgroup W6 , of rank six. For any choice of E6 (K) < G, W5 < E6 (K)
(as just described) has rank 5, is nontoral since it contains a 2A3 -subgroup, and so we
can take W6 = Z × W5 .
Each coset of Z of involutions in GrZ contains one element of each class 2A and
2B. Together with the above description of E5 , this shows that all 2A2 -subgroups of
W6 are contained in W5 . Hence for each nontoral subgroup E ≤ W6 which contains
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
93
Z, E ∩ W5 is the subgroup generated by 2A2 -subgroups of E, thus is normalized by
AutG (E), and so
AutG (E) ∼
= Aut(E, qE )
= AutG (E ∩ W5 ) = Aut(E ∩ W5 , qE∩W5 ) ∼
6
Aut (W6 ) ∼
= C ⋊ (Σ3 × GL3 (2))
= Aut(W6 , qW ) ∼
G
6
2
For Z ≤ E ≤ W6 , the subgroup E is nontoral exactly when it contains a 2A3 subgroup. This is immediate from the analogous statement in (A) for E6 (K).
(C) If G = E8 , then by [Gr, Theorem 2.17], G contains two maximal elementary abelian
subgroups W8 and W9 , neither of which is toral [Gr, Theorem 9.2]. An elementary
abelian 2-subgroup E ≤ G is nontoral if and only if qE is not quadratic or E has type
2B5 [Gr, Theorem 9.2].
We refer to [Gr, Theorem 2.17] for descriptions of W8 and W9 . There are subgroups
F0 ≤ F1 , F2 ≤ W8 such that rk(F0 ) = 2, rk(F1 ) = rk(F2 ) = 5, F1 ∩ F2 = F0 , and
W8 ∩ 2A = (F1 rF0 ) ∪ (F2 rF0 ). Also, AutG (W8 ) is the group of those automorphisms
of W8 which leave F0 invariant, and either leave F1 and F2 invariant or exchange them.
We can assume that W9 = T(2) hθi, where θ ∈ NG (T ) inverts T . Also, W9 rT(2) ⊆ 2B.
Hence T(2) = hW9 ∩ 2Ai is AutG (W9 )-invariant. Each automorphism of W9 which is the
identity on T(2) is induced by conjugation by some element of order 4 in T , and thus
AutG (W9 ) is the group of all automorphisms whose restriction to T(2) lies in AutG (T(2) ).
We next list other properties of elementary abelian subgroups of G, and of their centralizers
and normalizers, which will be needed in the proof.
(D) If G = E8 , E ≤ G, E ∼
= C2r , and |E ∩ 2A| = m, then dim(CG (E)) = 28−r + 25−r m − 8.
This follows from character computations:Pif g denotes the Lie algebra of G = E8 (K),
then dim(CG (E)) = dim(Cg (E)) = |E|−1 x∈E χg (x). By [Gr, Table VI], χg (1) =
dim(G) = 248, and χg (x) = 24 or −8 when x ∈ 2A or 2B, respectively.
(E) If G = E8 , E ≤ G is an elementary abelian 2-group, and Et < E has index 2 and is
such that ErEt ⊆ 2B, then there is g ∈ G such that gE ≤ W9 = T(2) hθi and gEt ≤ T(2) .
It suffices to prove this when E is maximal among such such pairs Et < E. We can
assume that E is contained in W8 or W9 .
If E ≤ W8 , then in the notation of (C), F0 ≤ E (since E is maximal), and either
rk(E ∩ Fi ) = 3 for i = 1, 2 and rk(E) = 6, or rk(E ∩ Fi ) = 4 for i = 1, 2 and
rk(E) = 7. These imply that |E ∩ 2A| = 8 or 24, respectively, and hence by (D)
that dim(CG (Et )) = 8 (CG (Et )0 = T ) and dim(CG (E)) = 0. Hence in either case,
if g ∈ G is such that gEt ≤ T(2) , then gErgEt ⊆ θT , and there is t ∈ T such that
tg
E ≤ T(2) hθi = W9 .
If E ≤ W9 , set E2 = hE ∩ 2Ai. Then E2 ≤ E ∩ T(2) and E2 ≤ Et , so there is nothing
to prove unless rk(E/E2 ) ≥ 2. In this case, from the maximality of E, we see that
Et = Ea × Eb , where Ea ∼
= C22 has type 2ABB, Eb is a 2B3 -group, and Ea ⊥ Eb with
respect to the form q. Thus rk(E) = 6, |E ∩ 2A| = 8, and the result follows by the
same argument as in the last paragraph.
(F) If G = E8 , and E ≤ G is a nontoral elementary abelian 2-group, then either E contains
a 2A3 -subgroup, or E is G-conjugate to a subgroup of W9 .
94
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
Assume E ≤ W8 is nontoral and contains no 2A3 -subgroup. We use the notation
F0 < F1 , F2 < W8 of (C). Set Ei = E ∩ Fi for i = 0, 1, 2. Then qE1 E2 is quadratic: it
is the orthogonal direct sum of qE0 , qE1 /E0 , and qE2 /E0 , each of which is quadratic since
rk(Ei /E0 ) ≤ 2 for i = 1, 2 (E has no 2A3 -subgroup). Hence E > E1 E2 ≥ hE ∩ 2Ai
since E is nontoral, so E is conjugate to a subgroup of W9 by (E).
(G) Let E ≤ G be an elementary abelian 2-subgroup, and let Et ≤ E be maximal among
toral subgroups of E. Assume that Et ∩Et⊥ ∩2B = ∅, and that either rk(T )−rk(Et ) ≥ 2
b
or Et ∩ Et⊥ = 1. Then E ∈
/ Z.
To see this, choose F ≥ Ft which is G-conjugate to E ≥ Et and such that Ft =
F ∩ T(2) . By maximality, no element of F rFt is CG (Ft )-conjugate to an element of
T . If Ft ∩ Ft⊥ = 1, then some CW (Ft )-orbit in Ft⊥ r1 has odd order. Otherwise,
since q is linear on Ft ∩ Ft⊥ , we have Ft ∩ Ft⊥ = hyi for some y ∈ 2A, in which case
|q−1
(0)| = |Ft⊥ |/2 is even since rk(Ft⊥ ) ≥ rk(T ) − rk(Ft ) ≥ 2. So again, some CW (Ft )Ft⊥
orbit in Ft⊥ r1 has odd order in this case. Point (G) now follows from Proposition
A.13.
(H) Assume G = E8 . Let 1 6= E0 ≤ E ≤ G be
rk(E) = 3, and E ∩ 2A = E0 r1. Then
E × F4 (K)
∼
CG (E) = E × PSp8 (K)
E × PSO (K)
8
elementary abelian 2-subgroups, where
if rk(E0 ) = 3
if rk(E0 ) = 2
if rk(E0 ) = 1.
To see this, fix 1 6= y ∈ E0 , and identify CG (y) ∼
= SL2 (K) ×C2 E7 (K). For each
x ∈ Erhyi, since x and xy are G-conjugate, x 6= (1, b) for b ∈ E7 (K). Thus x = (a, b)
for some a ∈ SL2 (K) and b ∈ E7 (K) both of order 4, and (in the notation of [Gr, Table
VI]) b is in class 4A or 4H since b2 ∈ Z(E7 (K)). By (D) and [Gr, Table VI],
(
80 = dim(CE7 (K) (4H)) + 1 if E has type 2AAA
dim(CG (E)) =
64 = dim(CE7 (K) (4A)) + 1 if E has type 2ABB,
and thus x ∈ 2A if b ∈ 4H and x ∈ 2B if b ∈ 4A. Thus if E = hy, x1 , x2 i, and
xi = (ai , bi ), then ha1 , a2 i ≤ SL2 (K) and hb1 , b2 i ≤ E7 (K) are both quaternion of order
8. Point (H) now follows using the description in [Gr, Proposition 9.5(i)] of centralizers
of certain quaternion subgroups of E7 (K). When combined with the description in [Gr,
Table VI] of CE7 (K) (4A), this also shows that
F ∼
= C22 of type 2ABB =⇒ CG (F )0 is of type A7 T 1
(9)
(i.e., CG (F )0 ∼
= (SL8 (K) × K× )/Z, for some finite subgroup Z ≤ Z(SL8 (K)) × K× ).
(I) If U < G is a 2A3 -subgroup, then CG (U) = U × H, where H is as follows:
G
H
F4
E6
E7
E8
SO3 (K) SL3 (K) Sp6 (K) F4 (K)
When G = E8 , this is a special case of (H). For x ∈ 2A ∩ F4 (K), CE8 (K) (x) ∼
=
SL2 (K) ×C2 E7 (K) by [Gr, 2.14]. Since CF4 (K) (x) ∼
SL
(K)
×
Sp
(K),
this
shows
that
= 2
C2
6
CE7 (K) (U) ∼
= U × Sp6 (K).
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
95
Similarly, CE8 (K) (y) ∼
= SL3 (K) ×C3 E6 (K) by [Gr, 2.14] again (where y is in class
3B in his notation). There is only one class of element of order three in F4 (K) whose
centralizer contains a central factor SL3 (K) — CF4 (K) (y) ∼
= SL3 (K) ×C3 SL3 (K) for y of
∼
type 3C in F4 (K) — and thus CE6 (K) (U) = U × SL3 (K).
If G = F4 , then by [Gr, 2.14], for y ∈ 3C, CG (y) ∼
= SL3 (K) ×C3 SL3 (K). Also,
the involutions in one factor must all lie in the class 2A and those in the other in
2B. This, together with Proposition 2.5, shows that for U2 < U of rank 2, CG (U2 ) ∼
=
(T 2 ×C3 SL3 (K))hθi, where θ inverts a maximal torus. Thus CG (U) = U × CSL3 (K) (θ),
where by [Gr, Proposition 2.18], CSL3 (K) (θ) ∼
= SO3 (K). This finishes the proof of (I).
For the rest of the proof, we fix a nontoral elementary abelian 2-subgroup E < G. We
b In almost all cases, we do this either by showing that the hypotheses
must show that E ∈
/ Z.
of (G) hold, or by showing that δ(AutG (E)) > |π0 (CG (E))| (where δ(−) is as in Lemma A.5),
in which case AutG (E) has no strongly 2-embedded subgroup by Proposition A.4, and hence
b
E∈
/ Z.
By (A), (B), and (F), either E contains a 2A3 -subgroup of rank three, or G = E8 and E
is G-conjugate to a subgroup of W9 . These two cases will be handled separately.
Case 1: Assume first that E contains a 2A3 -subgroup U ≤ E. From the lists in (A,B,C)
of maximal nontoral subgroups, there are the following possibilities.
G = F4 , E6 , or E7 : By (A,B), we can write E = U × E0 × Z, where E0 is a 2Bk subgroup
(some k ≤ 2) and UE0 rE0 ⊆ 2A (and where Z = 1 unless G = E7 ). If k = 0, then
E ∈
/ Zb by (G), so assume k ≥ 1. By (I), and since each elementary abelian 2-subgroup
of SL3 (K) and of Sp6 (K) has connected centralizer, π0 (CG (E)) ∼
= U if G = E6 or E7 . If
∼
G = F4 , then by (I) again, and since the centralizer in SO3 (K) = PSL2 (K) of any C2k has
2k components, |π0 (CG (E))| = 23+k .
By (A,B) again, AutG (E) is the group of all automorphisms which normalize E0 and
UE0 and fix Z. Hence
|O2(Aut (E))| = 23k
and
Aut (E)/O2(Aut (E)) ∼
= GL3 (2) × GLk (2) .
G
So δ(AutG (E)) ≥ 2
G
3k+3
G
b
> |π0 (CG (E))| by Lemma A.5, and E ∈
/ Z.
G = E8 : By (I), CG (U) = U ×H where H ∼
= F4 (K). Set E2 = E ∩H, and let E0 = hE2 ∩2Bi.
Set k = rk(E0 ) and ℓ = rk(E2 /E0 ).
If k = 0, then E2 has type 2Aℓ , and Er(U ∪ E2 ) ⊆ 2B. So each maximal toral
subgroup Et < E has the form Et = U1 × U2 , where rk(U1 ) = 2, rk(U2 ) ≤ 2, and
b
Et ∩ 2A = (U1 ∪ U2 )r1. The hypotheses of (G) thus hold, and so E ∗ ∈
/ Z.
Thus k = 1, 2. If ℓ ≤ 2, then E2 is toral, and |π0 (CG (E))| = 8 · |π0 (CH (E2 ))| ≤ 23+k
by formula (8) in the proof of Lemma A.14. (Note that ε = 1 and η = 0 in the notation
of that formula.) If ℓ = 3, then |π0 (CG (E))| = 26+k by the argument just given for
F4 (K). Also, AutG (E) contains all automorphisms of E which normalize E0 , and either
normalize UE0 and E2 or (if ℓ = 3) exchange them: since in the notation of (C), each
such automorphism extends to an automorphism of W8 which normalizes F1 and F2 . So
|O2 (AutG (E))| ≥ 2k(3+ℓ) , and AutG (E)/O2 (AutG (E)) ∼
= GL3 (2) × GLk (2) × GLℓ (2) or (if
ℓ = 3) (GL3 (2) ≀ C2 ) × GLk (2). In all cases, δ(AutG (E)) ≥ 23k+ℓk+3 > |π0 (CG (E))|, so
E∈
/ Z.
Case 2: Now assume that G = E8 , and that E is G-conjugate to a subgroup of W9 . To
simplify the argument, we assume that E ≤ W9 , and then prove that no subgroup E ∗ ∈ Zb
96
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
can be G-conjugate to E. Recall that W9 = T(2) hθi, where θ ∈ NG (T ) inverts the torus and
θT(2) ⊆ 2B.
If E ∩ 2A = ∅, then rk(E) = 5. In this case, AutG (E) ∼
= GL5 (2) and |CG (E)| = 215 [CG,
Proposition 3.8]. (Cohen and Griess work in E8 (C), but their argument also holds in our
situation.) Since δ(GL5 (2)) > 215 by Lemma A.5(d), no E ∗ ∈ Zb can be G-conjugate to E.
Now assume E has 2A-elements, and set E2 = hE ∩ 2Ai. Then E2 ≤ T(2) (hence qE2 is
quadratic) by the above remarks. Set E1 = E2⊥ ∩ E2 and E0 = Ker(qE1 ). If E0 = 1 and
b
rk(E2 ) 6= 7, then by (G), no subgroup of S which is G-conjugate to E lies in Z.
It remains to consider the subgroups E for which E0 6= 1 or rk(E2 ) = 7. Information
about |O2(AutG (E))| and |π0 (CG (E))| for such E is summarized in Table A.1. By the “type
of qE ” is meant the type of quadratic form, in the notation used in the proof of Lemma A.14.
Case nr.
1
2
3
4
5
6
7
8
9
10
11
12
13
rk(E/E2 )
1
1
1
1
1
1
1
1
1
1
2
2
2
rk(E2 /E0 )
7
6+
5
4+
4+
4−
3
3
2−
2−
1
1
1
rk(E0 )
0
1
1
2
1
1
2
1
2
1
3
2
1
type of qE2
[7] [6](1) [5](1) [4](2) [4](1) [4](1) [3](2) [3](1) [2](2) [2](1) [1](3) [1](2) [1](1)
|π0 (CG (E ∗ ))| ≤
29
29
|O2 (AutG (E ∗ ))| 27
213
+
+
+
−
210
210
28
27
26
211
214
29
29
δ(AutG (E ∗ ))≥ 213 217 3·213 216
210
210
−
−
25
25
24
212
28
25
211
27
28
25
211
28
25
212
27
29
25
214
29
25
Table A.1.
We first check that the table includes all cases. If rk(E/E2 ) = 1, then E2 = E ∩ T(2) , and
the table lists all types which the form qE2 can have. Note that since E2 is generated by
nonisotropic vectors, qE2 cannot have type [2]+
(k) . If rk(E/E2 ) = 2, then qE2 is linear, and
must be one of the three types listed. Since qE∩T(2) is quadratic and qE is not, E2 has index
at most 2 in E ∩ T(2) .
We claim that
E, F < W9 , α ∈ Iso(E, F ) such that α(E ∩ T(2) ) = F ∩ T(2) and α(E ∩ 2A) =
F ∩ 2A =⇒ α = ctg for some t ∈ T and some g ∈ NG (T ) = G ∩ NG (T ).
(10)
By (C) and Witt’s theorem (see [Ta, Theorem 7.4]), there is g ∈ NG (T ) such that α|E∩T(2) =
cg , and we can assume g ∈ G by Lemma 2.9. Then gErg(E ∩ T(2) ) ≤ θT since θT ∈
Z(NG (T ))/T , so α = ctg for some t ∈ T . This proves (10). In particular, any two subgroups
of W9 which have the same data as listed in the first three rows of Table A.1 are G-conjugate.
By (10), together with (E) when rk(E/E2 ) = 2, we have AutG (E) = Aut(E, qE ) in all
cases. Thus AutG (E) is the group of all automorphisms of E which normalize E0 and E2
and preserve the induced quadratic form on E2 /E0 . This gives the value for |O2 (AutG (E))|
in the table, and the lower bounds for δ(AutG (E)) then follow from Lemma A.5.
In cases 1–6, the upper bounds for |π0 (CG (E))| given in the table are proven in [O2, p.
78–79]. In all cases, |π0 (CG (E2 ))| is first computed, using Proposition 2.5 or the upper bound
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
97
given in formula (8) in the proof of the last lemma, and then [O2, Proposition 8.8] is used
to compute an upper bound for |π0 (CG (E))| |π0 (CG (E2 ))|. There is in fact an error in the
table on [O2, p. 79] (the group CG (E0 )0s in the third-to-last column should be SL2 × SL2 up
to finite cover), but correcting this gives in fact a better estimate |π0 (CG (E))| ≤ 29 .
Case nr. 11 can be handled in a similar way. Set Et = E ∩ T(2) < E, so that |E/Et | = 2 =
3
⊥
4
|Et /E2 |. The form qEt has type [2]+
(3) , while Et has type 2B . Hence |π0 (CG (Et ))| ≤ 2 by
(8). By [O2, Proposition 8.8], |π0 (CG (E))| ≤ 24+r , where r = dim(T ) = 8.
To handle the remaining cases, fix rank 2 subgroups F1 , F2 ≤ T(2) < G with involutions
of type AAA and ABB, respectively, and consider the information in Table A.2. The
dim(CG (Fi )hθ, gi) for g as follows:
i
CG (Fi hθi)
1
F1 hθi × PSp8 (K)
20
24
16
16
20
2 F2 hθi × PSO8 (K)
12
16
16
16
12
−I4 ⊕ I4 −I2 ⊕ I6 order 4 2A 2B
Table A.2.
description of CG (Fi hθi) follows from (H). The third through fifth columns give dimensions
of centralizers of Fi hθihgi, for g as described after lifting to Sp8 (K) or SO8 (K). (Here,
Im denotes the m × m identity matrix.) The last two columns do this for g ∈ 2A or 2B,
respectively, when g ∈ T(2) is orthogonal to Fi with respect to the form q, and the dimensions
follow from (D). Thus elements of class 2B lift to involutions in Sp8 (K) or SO8 (K) with
4-dimensional (−1)-eigenspace, while for i = 1 at least, elements of class 2A lift to elements
of order 4 in Sp8 (K).
Thus in all of the cases nr. 7–13 in Table A.1, we can identify E = Fi hθi × F ∗ , where
i = 1 in nr. 7–10 or i = 2 in nr. 11–13, and where F ∗ lifts to an abelian subgroup of Sp8 (K)
or SO8 (K) (elementary abelian except for nr. 7–8). This information, together with the
following:
H a group, Z ≤ Z(H), |Z| = p, Z ≤ P ≤ H a p-subgroup
=⇒ CH/Z (P ) CH (P )/Z ≤ |P/Fr(P )|
(applied with H = Sp8 (K) or SO8 (K)), imply the remaining bounds in the last line of Table
A.1.
In all but the last case in Table A.1, δ(AutG (E)) > |π0 (CG (E))|, so no E ∗ ∈ Zb is Gconjugate to E by Proposition A.4. In the last case, by the same proposition, E can be
G-conjugate to some E ∗ ∈ Zb only if AutG (E) acts transitively on π0 (CG (E)) ∼
= C25 with
point stabilizers isomorphic to Σ3 . By (10), each class in O2 (AutG (E)) is represented by some
element tg ∈ NG (E), where g ∈ NG (T ) and t ∈ T . In particular, (tg)σ(tg)−1 = tσ(t)−1 ∈ T .
So each class in the O2 (AutG (E))-orbit of 1 ∈ π0 (CG (E)) has nonempty intersection with
T . But by (9), CG (F2 )0 ∩ θT = ∅, so θCG (E)0 ∩ T = ∅. Thus the action is not transitive
b
/ Z.
on π0 (CG (E)), and hence E ∗ ∈
98
CARLES BROTO, JESPER M. MØLLER, AND BOB OLIVER
References
[AG]
J. Alperin & D. Gorenstein, The multiplicators of certain simple groups, Proc. Amer. Math. Soc. 17
(1966), 515–519
[AOV] K. Andersen, B. Oliver, & J. Ventura, Reduced, tame, and exotic fusion systems, Proc. London
Math. Soc. 105 (2012), 87–152
[A1] M. Aschbacher, On the maximal subgroups of the finite classicial groups, Invent. math. 76 (1984),
469–514
[A2] M. Aschbacher, The generalized Fitting subsystem of a fusion system, Memoirs Amer. Math. Soc.
209 (2011), nr. 986
[AKO] M. Aschbacher, R. Kessar, & B. Oliver, Fusion systems in algebra and topology, Cambridge Univ.
Press (2011)
[BC] J. Bidwell & M. Curran, Corrigendum to “The automorphism group of a split metacyclic p-group”,
Arch. Math. 92 (2009), 14–18
[Brb] N. Bourbaki, Groupes et algèbres de Lie, Chapitres 4–6, Hermann (1968)
[5a1] C. Broto, N. Castellana, J. Grodal, R. Levi, & B. Oliver, Subgroup families controlling p-local finite
groups, Proc. London Math. Soc. 91 (2005), 325–354
[5a2] C. Broto, N. Castellana, J. Grodal, R. Levi, & B. Oliver, Extensions of p-local finite groups, Trans.
Amer. Math. Soc. 359 (2007), 3791–3858
[BL] C. Broto & R. Levi, On spaces of self-homotopy equivalences of p-completed classifying spaces of
finite groups and homotopy group extensions, Topology 41 (2002), 229–255
[BLO1] C. Broto, R. Levi, & B. Oliver, Homotopy equivalences of p-completed classifying spaces of finite
groups, Invent. math. 151 (2003), 611–664
[BLO2] C. Broto, R. Levi, & B. Oliver, The homotopy theory of fusion systems, J. Amer. Math. Soc. 16
(2003), 779–856
[BMO] C. Broto, J. Møller, & B. Oliver, Equivalences between fusion systems of finite groups of Lie type,
Jour. Amer. Math. Soc. 25 (2012), 1–20
[Br]
K. Brown, Buildings, Springer-Verlag (1989)
[BW] N. Burgoyne & C. Williamson, On a theorem of Borel and Tits, Arch. math. 27 (1976), 489–491
[Bu] W. Burnside, The Theory of Groups of Finite Order, Cambridge Univ. Press (1897).
[CE] H. Cartan & S. Eilenberg, Homological algebra, Princeton Univ. Press (1956)
[Ca] R. Carter, Simple groups of Lie type, Wiley (1972)
[Ca2] R. Carter, Finite groups of Lie type: conjugacy classes and complex characters, Wiley (1985)
[Ca3] R. Carter, Conjugacy classes in the Weyl group, Compositio math. 25 (1972), 1–59
[CF] R. Carter & P. Fong, The Sylow 2-subgroups of the finite classical groups, J. Algebra 1 (1964),
139–151
[CG] A. Cohen & R. Griess, On finite simple subgroups of the complex Lie group of type E8 , Proc. Symp.
Pure Math. 47 (1987), 367–405
[Cu] C. Curtis, Central extensions of groups of Lie type, J. Reine Angew. Math. 220 (1965), 174–185
[DS] A. Delgado & B. Stellmacher, Weak (B, N )-pairs of rank 2, Groups and graphs: new results and
methods, Birkhäuser (1985), 58–244
[Di]
L. Dickson, A new system of simple groups, Math. Ann. 60 (1905), 137–150
[Fn]
P. Fan, Amalgams of prime index, J. Algebra 98 (1986), 375–421
[Gl]
G. Glauberman, Central elements in core-free groups, J. Algebra 4 (1966), 403–420
[Gl2] G. Glauberman, Global and local properties of finite groups, Finite simple groups, Academic Press
(1971), 1–64
[GLn] G. Glauberman & J. Lynd, Control of fixed points and existence and uniqueness of centric linking
systems (preprint)
[Gd] D. Goldschmidt, Automorphisms of trivalent graphs, Annals of Math. 111 (1980), 377–406
[G]
D. Gorenstein, Finite groups, Harper & Row (1968)
[GL] D. Gorenstein & R. Lyons, The local structure of finite groups of charateristic 2 type, Memoirs Amer.
Math. Soc. 276 (1983)
[GLS3] D. Gorenstein, R. Lyons, & R. Solomon, The classification of the finite simple groups, nr. 3, Amer.
Math. Soc. surveys and monogr. 40 #3 (1997)
[Gr]
R. Griess, Elementary abelian p-subgroups of algebraic groups, Geometriae Dedicata 39 (1991), 253–
305
[GrL] R. Griess & R. Lyons, The automorphism group of the Tits simple group, Proc. Amer. Math. Soc.
52 (1975), 75–78
AUTOMORPHISMS OF FUSION SYSTEMS OF FINITE SIMPLE GROUPS OF LIE TYPE
99
[Hu]
[H]
[HB]
[Kl]
J. Humphreys, Linear algebraic groups, Springer-Verlag (1975)
B. Huppert, Endliche Gruppen I, Springer-Verlag (1967)
B. Huppert & N. Blackburn, Finite groups III, Springer-Verlag (1982)
P. Kleidman, The maximal subgroups of the Steinberg triality groups 3D4 (q) and of their automorphism groups, J. Algebra 115 (1988), 182–199
[LO] R. Levi & B. Oliver, Construction of 2-local finite groups of a type studied by Solomon and Benson,
Geom. Topol. 6 (2002), 917–990.
[McL] S. Mac Lane, Homology, Springer-Verlag (1975)
[Ma] G. Malle, The maximal subgroups of 2F4 (q 2 ), J. Algebra 139 (1991), 52–69
[O1] B. Oliver, Equivalences of classifying spaces completed at odd primes, Math. Proc. Camb. Phil. Soc.
137 (2004), 321–347
[O2] B. Oliver, Equivalences of classifying spaces completed at the prime two, Amer. Math. Soc. Memoirs
848 (2006)
[O4] B. Oliver, Existence and uniqueness of linking systems: Chermak’s proof via obstruction theory, Acta
Math. 211 (2013), 141–175
[O5] B. Oliver, Simple fusion systems over p-groups with abelian subgroup of index p: I, J. Algebra 398
(2014), 527–541
[O6] B. Oliver, Reduced fusion systems over 2-groups of sectional rank at most 4, Memoirs Amer. Math.
Soc. 239 (2015), nr. 1131
[OV] B. Oliver & J. Ventura, Saturated fusion systems over 2-groups, Trans. Amer. Math. Soc. 361 (2009),
6661–6728
[Pa]
D. Parrott, A characterization of the Tits’ simple group, Can. J. Math. 24 (1972), 672–685
[Ree] R. Ree, A family of simple groups associated with the simple Lie algebra of type (G2 ), Amer. J.
Math. 83 (1961), 432–462
[Se]
J.-P. Serre, Trees, Springer-Verlag (1980)
[Sh]
K. Shinoda, The conjugacy classes of the finite Ree groups of type (F4 ), J. Fac. Sci. Univ. Tokyo 22
(1975), 1–15
[Sp]
T. A. Springer, Linear algebraic groups, Birkhäuser (1981, 1998)
[St1] R. Steinberg, Automorphisms of finite linear groups, Canad. J. Math. 12 (1960), 606–615
[St4] R. Steinberg, Gnrateurs, relations et revtements de groupes algbriques, Coll. on Algebraic groups,
Brussels (1962), 113–127
[St5] R. Steinberg, Representations of algebraic groups, Nagoya Math. J. 22 (1963), 33–56
[St2] R. Steinberg, Lectures on Chevalley groups, Yale Univ. (1967)
[St3] R. Steinberg, Endomorphisms of linear algebraic groups, Memoirs Amer. Math. Soc. 80 (1968)
[St6] R. Steinberg, Generators, relations, and coverings of algebraic groups, II, J. Algebra 71 (1981),
527–543
[Sz1] M. Suzuki, Group theory I, Springer-Verlag (1982)
[Sz2] M. Suzuki, Group theory II, Springer-Verlag (1986)
[Ta]
D. Taylor, The geometry of the classical groups, Heldermann Verlag (1992)
[Ti]
J. Tits, Buildings of spherical type and finite BN -pairs, Lecture Notes in Math. 386, Springer-Verlag
(1974)
[Wg] A. Wagner, The subgroups of PSL(5, 2a ), Resultate Math. 1 (1978), 207–226
[Wi] R. Wilson, The finite simple groups, Springer-Verlag (2009)
Departament de Matemàtiques, Universitat Autònoma de Barcelona, E–08193 Bellaterra,
Spain
E-mail address: [email protected]
Matematisk Institut, Universitetsparken 5, DK–2100 København, Denmark
E-mail address: [email protected]
Université Paris 13, Sorbonne Paris Cité, LAGA, UMR 7539 du CNRS, 99, Av.
Clément, 93430 Villetaneuse, France.
E-mail address: [email protected]
J.-B.
| 4math.GR
|
arXiv:1801.09887v1 [math.PR] 30 Jan 2018
Contribution of the Extreme Term in the Sum of
Samples with Regularly Varying Tail
Van Minh Nguyen
∗
January 31, 2018
Abstract
For a sequence of random variables (X1 , X2 , . . . , Xn ), n ≥ 1, that are
independent and identically distributed with a regularly varying tail with
index −α, α ≥ 0, we show that the contribution of the maximum term
Mn , max(X1 , . . . , Xn ) in the sum Sn , X1 + · · · + Xn , as n → ∞,
decreases monotonically with α in stochastic ordering sense.
Keywords Extreme · Sum · Regular variation · Stochastic ordering ·
Wireless network modeling
Mathematics Subject Classification (2000) 60G70 · 60G50 · 60E15
· 90B15
1
Introduction
Let (X1 , X2 , . . .) be a sequence of random variables that are independent and
identically distributed (i.i.d.) with distribution F . For n ≥ 1, the extreme term
and the sum are defined as follows:
n
Mn , max Xi ,
i=1
Sn ,
n
X
Xi .
(1)
n=1
The influence of the extreme term in the sum has various implications in both
theory and applications. In particular, it has been used to characterize the
nature of possible convergence of the sums of i.i.d. random variables (Darling
1952). On the other hand, besides well-known applications to risk management,
insurance and finance (Embrechts et al. 1997), it has been recently applied to
wireless communications for characterizing a fundamental parameter which is
the signal-to-interference ratio (Nguyen et al. 2010; Nguyen & Kountouris 2017).
Denote by F̄ (x) , 1 − F (x) the tail distribution of F , a primary result of
this question is due to (Darling 1952) in which it is shown that
∗ Huawei
Technologies, France. Email: [email protected]
1
Property 1 ((Darling 1952)). Suppose that Xi ≥ 0. Then E(Sn /Mn ) → 1 as
n → ∞ if for every t > 0 we have
lim
x→∞
F̄ (tx)
= 1.
F̄ (x)
(2)
In Darling’s result, condition (2) is of particular interest as it is a specific
case of a more general class called regularly varying tails which is defined in the
following (Bingham et al. 1989):
Definition 1. A positive, Lebesgue measurable function h on (0, ∞) is called
h(tx)
= tα for every constant
regularly varying with index α ∈ R at ∞ if lim
x→∞ h(x)
0 < t < ∞.
In the sequel, Rα denotes the class of regularly varying functions with index
α, and in particular R0 is referred to as the class of slowly varying functions. In
p
d
a.s.
addition, →, →, and → stands for the convergence in distribution, convergence
in probability, and almost sure convergence, respectively.
Since the work of (Darling 1952), there has been subsequent extensions which
in particular investigated other cases of regularly varying tails. Among those,
(Arov & Bobrov 1960) derived the characteristic function and limits of the
jointed sum and extreme term for a regularly varying tail. (Teugels 1981) derived
the limiting characteristic function of the ratio of the sum to order statistics,
and moreover investigated norming sequences for its convergence to a constant
or a normal law. (Chow & Teugels 1978; Anderson & Turkman 1991, 1995)
investigated the asymptotic independence of normed extreme and normed sum.
a.s.
Unlike the slowly varying case, (O’Brien 1980) showed that Mn /Sn → 0 ⇔
EX1 < ∞. For F̄ regularly varying with index −α, α > 0, (Bingham & Teugels
1981) showed that the extreme term only contributes a proportion to the sum:
Property 2 ((Bingham & Teugels 1981)). The following are equivalent:
1. F̄ ∈ R−α for 0 < α < 1;
d
2. Mn /Sn → R where R has a non-degenerate distribution;
3. E(Sn /Mn ) → (1 − α)−1 .
Property 3 ((Bingham & Teugels 1981)). Let µ = EX1 . The following are
equivalent:
1. F̄ ∈ R−α for 1 < α < 2;
d
2. (Sn − (n − 1)µ)/Mn → D where D has a non-degenerate distribution;
3. E((Sn − (n − 1)µ)/Mn ) → c where c is a constant.
(Maller & Resnick 1984) extended Darling’s convergence in mean to convergence in probability of Sn /Mn :
2
Property 4 ((Maller & Resnick 1984)).
p
Sn /Mn → 1 ⇔ F̄ ∈ R0 .
The ratio of the extreme to the sum has been further studied with the
following result:
Property 5 ((Downey & Wright 2007)). If either one of the following conditions:
1. F̄ ∈ R−α for α > 1,
2. F has finite second moment,
holds, then
E(Mn /Sn ) =
EMn
(1 + o(1)),
ESn
as n → ∞.
To this end, the contribution of the extreme term in the sum has been
investigated and classified in the following cases:
• F̄ ∈ R0 ;
• F̄ ∈ R−α , 0 < α < 1;
• F̄ ∈ R−α , 1 < α < 2;
• F̄ ∈ R−α , α > 2.
Nevertheless, how the influence of the extreme term in the sum gradually varies
with the regular variation index has not been quantified and remains an open
question. Precisely, consider two cases with F̄ ∈ R−α1 and F̄ ∈ R−α2 in which
0 ≤ α1 < α2 , which case results in larger Mn /Sn ? This question is particularly important for analysis and design of wireless communication networks.
In this context, random variables (X1 , X2 , . . .) are used to model the signal
that a user receives from base stations. It has been proven that the tail distribution of Xi is regularly varying (Nguyen & Kountouris 2017) either due
to the effect of distance-dependent propagation loss or due to fading (Tse &
Viswanath 2005) that is regular varying (Rajan et al. 2017) as a consequence
of advances in communication and signal processing techniques such as massive multiple-input-multiple-output transmission, coordinated multipoint and
millimeter wave systems. Meanwhile, Mn expresses the useful signal power and
(Sn − Mn ) is the total interference due to the other transmitters (Nguyen 2011).
Mn /(Sn − Mn ) is hence the signal-to-interference ratio (SIR), and its limit as
n → ∞ happens for a dense or ultra-dense network (Nguyen & Kountouris 2016;
Nguyen 2017). Capacity of a communication channel is expressed in term of the
well-known Shannon’s capacity limit of log(1 + SIR) (Shannon 1948; Cover &
Thomas 2006) considering that thermal noise is negligible in comparison to the
interference. Therefore, Mn /Sn (or Sn /Mn ) is a fundamental parameter of wireless network engineering. From the perspective of capacity, a primary purpose
3
is to design the network such that Xi possesses properties that make Mn /Sn as
large as possible. In particular, how Mn /Sn varies according to the tail of Xi
turns out to be a critical question.
In this paper, we establish a stochastic ordering for Sn /Mn and show that
between two cases with F̄ ∈ R−α1 and F̄ ∈ R−α2 in which 0 ≤ α1 < α2 , the
contribution of Mn in Sn as n → ∞ is larger in the former than in the latter
case in stochastic ordering sense.
2
Main Result
In the following, for n ≥ 1 we define
Rn , Sn /Mn .
(3)
We also restrict our consideration to non-negative random variables, i.e., Xi ≥ 0,
and consider F̄ ∈ R−α with α ≥ 0. In the context where α is analyzed, a
variable v is written as vα , e.g., write Sα,n , Mα,n , and Rα,n for Sn , Mn , and
Rn , respectively.
Lemma 1. For s ∈ C with ℜ(s) ≥ 0, define LRn (s) , E e−sRn . If F̄ ∈ R−α
with α ≥ 0 then:
e−s
LRn (s) =
, n → ∞,
(4)
1 + φα (s)
where
Z 1
dt
(5)
φα (s) , α
(1 − e−st ) 1+α .
t
0
Proof. Let G(x1 , · · · , xn ) be the joint distribution of (X1 , · · · , Xn ) given that
Mn = X1 , it is given as follows:
(
F (dx1 ) · · · F (dxn ) if x1 = maxni=1 xi
G(dx1 , · · · , dxn ) =
.
(6)
0
otherwise
Since (X1 , · · · , Xn ) are i.i.d., Mn = X1 with probability 1/n. Thus, the (unconditional) joint distribution of (X1 , · · · , Xn ) is nG(x1 , · · · , xn ). Hence,
Z
Z
−sRn
) = n · · · e−s(x1 +x2 +···+xn )/x1 G(dx1 , · · · , dxn ),
LRn (s) = E(e
and using G from (6), we obtain
Z ∞ Z x
Z
LRn (s) = n
···
0
= ne
−s
= ne−s
0
Z
∞
Z0 ∞
x
e
−s
0
Z
−st
0
(ϕ(x))
e
−sxi /x
n−1
F (xdt)
F (dx)
n−1
0
4
!
F (dxi ) F (dx)
i=2
1
e
n
Y
F (dx),
(7)
where
ϕ(x) ,
1
Z
e−st F (xdt).
(8)
0
Given that ℜ(s) ≥ 0, we can see that
Z
|ϕ(x)| ≤
1
0
Hence,
F (xdt) = F (x) < 1,
for x < ∞.
0
T
Z
1
Z
|e−st F (xdt)| ≤
(ϕ(x))
n−1
F (dx) → 0,
as n → ∞ for T < ∞.
0
As a result, we only need to consider the contribution of large x in (7) for
LRn (s). An integration by parts with e−st and F (xdt) yields:
ϕ(x) = 1 − e−s F̄ (x) −
Z
1
se−st F̄ (xt)dt
0
Z
= 1 − F̄ (x) +
1
0
se−st F̄ (x) − F̄ (tx) dt.
(9)
For F̄ ∈ R−α with α ≥ 0, we can write
F̄ (tx) ∼ t−α F̄ (x),
Thus
Z
0
as x → ∞,
1
se
−st
F̄ (x) − F̄ (tx) dt ∼ F̄ (x)
Z
0 < t < ∞.
1
se−st (1 − t−α )dt,
as x → ∞.
0
Using an integration by parts with (1 − t−α ) and d(1 − e−st ) we obtain
F̄ (x)
Z
1
se−st (1 − t−α )dt = −F̄ (x)
0
Put
φα (s) ,
Z
Z
1
α(1 − e−st )
0
1
α(1 − e−st )
0
dt
t1+α
dt
t1+α
.
(10)
,
and substitute it back in (10) and (9), we obtain
ϕ(x) = 1 − (1 + φα (s))F̄ (x),
for x → ∞.
Substitute ϕ(x) back in the expression of LRn (s) in (7), we obtain
Z ∞
n−1
−s
1 − (1 + φα (s))F̄ (x)
F (dx),
as n → ∞.
LRn (s) ∼ ne
0
5
(11)
Here, we resort to a change of variable with v = nF̄ (x) and obtain:
Z n
n−1
v
LRn (s) ∼ e−s
1 − (1 + φα (s))
dv
n
Z0 ∞
(a) −s
→e
e−v(1+φα (s)) dv
0
e−s
=
,
1 + φα (s)
as n → ∞
where (a) is due to the formula (1 + nx )n → ex as n → ∞.
To present the main result, we use the following notation. For two random
variables U and V , U is said to be smaller than V in Laplace transform order,
denoted by U Lt V , if and only if LU (s) = E(e−sU ) ≥ E(e−sV ) = LV (s) for
all positive real number s.
Theorem 1. Let Rα1 ,n and Rα2 ,n be as defined in (3) for F̄ ∈ R−α1 with
α1 ≥ 0 and for F̄ ∈ R−α2 with α2 ≥ 0, respectively. Then
α1 ≤ α2 ⇒ Rα1 ,n Lt Rα2 ,n ,
n → ∞.
(12)
Proof. The proof is direct from Lemma 1. By noting that α/t1+α is increasing
with respect to (w.r.t.) α ≥ 0 for t ∈ [0, 1], we have φα (s) in (5) is increasing
w.r.t. α ≥ 0. Thus, LRα,n (s) as n → ∞ and ℜ(s) > 0 is decreasing w.r.t. α ≥ 0.
It follows that, for 0 ≤ α1 ≤ α2 , we have LRα1 ,n (s) ≥ LRα2 ,n (s) as n → ∞
for all s with ℜ(s) > 0, thus Rα1 ,n Lt Rα2 ,n as n → ∞.
Theorem 1 dictates that the more slowly F̄ decays at ∞, i.e., smaller α, the
smaller is the ratio of the sum to the extreme, thus the bigger is the contribution
of the extreme term in the sum. This contribution of the extreme term to the
sum increases to the ceiling limit 1 when α gets close to 0 as we have known
from (Darling 1952; Maller & Resnick 1984).
Since we have established the Laplace transform ordering for Rn , an immediate application is related to completely monotonic and Bernstein functions.
Let us recall:
• Completely monotonic functions: A function g : (0, ∞) → R+ is said to be
completely monotonic if it possesses derivatives of all orders k ∈ N ∪ {0}
which satisfy (−1)k g (k) (x) ≥ 0, ∀x ≥ 0, where the derivative of order
k = 0 is defined as g(x) itself. We denote by CM the class of completely
monotonic functions.
• Bernstein functions: A function h : (0, ∞) → R+ with dh(x)/dx being
completely monotonic is called a Bernstein function. We denote by B the
class of Bernstein functions.
Note that a completely monotonic function is positive, decreasing and convex,
whereas a Bernstein function is positive, increasing and concave.
6
It is well known that for all completely monotonic functions g, we have
that U Lt V ⇔ E(g(U )) ≥ E(g(V )), whereas for all Bernstein functions h,
U Lt V ⇔ E(h(U )) ≤ E(h(V )). Hence, we can have a direct corollary of
Theorem 1 as follows.
Corollary 1. With the same notation and assumption of Theorem 1, if 0 ≤
α1 ≤ α2 , then:
∀g ∈ CM :
∀h ∈ B :
E(g(Rα1 ,n )) ≥ E(g(Rα2 ,n )),
E(h(Rα1 ,n )) ≤ E(h(Rα2 ,n )),
as n → ∞,
as n → ∞.
Note that h(x) = 1, ∀x > 0, is a Bernstein function, whereas g(x) = 1/x,
∀x > 0, is a completely monotonic function. For two cases with F̄ ∈ R−α1 and
F̄ ∈ R−α2 with 0 ≤ α1 ≤ α2 , Corollary 1 gives:
Sα1 ,n
Sα2 ,n
E
≤E
, n → ∞,
Mα1 ,n
Mα2 ,n
Mα2 ,n
Mα1 ,n
≥E
, n → ∞.
E
Sα1 ,n
Sα2 ,n
Application Example We now can show an application of the results developed above to the context of wireless communication networks. The signalto-interference ratio SIR as described in the Introduction can be expressed in
terms of Rn as follows:
Sn − M n
1
= Rn − 1 := Zn .
=
SIR
Mn
(13)
Assume that F̄ ∈ R−α , α ≥ 0, the Laplace transform of Zn can be directly
obtained from that of Rn as given by Lemma 1 as follows:
LZα,n (s) = es LRα,n (s) = (1 + φα (s))−1 ,
n → ∞,
(14)
for all s ∈ C, ℜs > 0. It is easy to see that LZα,n (s) is also decreasing with
respect to α ≥ 0 (see the proof of Theorem 1).
Now, consider two cases for the distribution of the signal Xi which are F̄ ∈
R−α1 and F̄ ∈ R−α2 with 0 ≤ α1 ≤ α2 , we firstly have:
Zα1 ,n Lt Zα2 ,n ,
n → ∞.
(15)
Then, by noting that functions 1/x and log(1 + 1/x) with x > 0 both are
completely monotonic, we immediately have for n → ∞:
E(SIRα1 ) ≥ E(SIRα2 ),
(16)
E(log(1 + SIRα1 )) ≥ E(log(1 + SIRα2 )).
(17)
This says that it is beneficial to the communication quality and capacity to
design the network such that the signal received from a transmitting base station
Xi admits a regularly varying tail with as small variation index as possible (i.e.,
as close to a slowly varying tail as possible).
7
References
Anderson, C. W., & Turkman, K. F. 1991. The joint limiting distribution of
sums and maxima of stationary sequences. J. Appl. Probab., 28(1), 33–44.
Anderson, C. W., & Turkman, K. F. 1995. Sums and maxima of stationary
sequences with heavy tailed distributions. Sankhyā: The Indian Journal of
Statistics, Series A, 57(1), 1–10.
Arov, D. Z., & Bobrov, A. A. 1960. The extreme terms of a sample and their role
in the sum of independent variables. Theory Probab. Appl., 5(4), 377–396.
Bingham, N. H., & Teugels, J. L. 1981. Conditions implying domains of attraction. Pages 23–34 of: Proc. 6th Conf. Prob. Theory.
Bingham, N. H., Goldie, C. M., & Teugels, J. L. 1989. Regular Variation.
Cambridge, U.K.: Cambridge University Press.
Chow, T., & Teugels, J. L. 1978. The sum and the maximum of i.i.d. random
variables. Pages 81–92 of: Proc. Second Prague Symp. Asympt. Statist.
Cover, T. M., & Thomas, J. A. 2006. Elements of Information Theory. 2nd edn.
John Wiley & Sons.
Darling, D. A. 1952. The influence of the maximum term in the addition of
independent random variables. Trans. Amer. Math. Soc., 73(1), 95–107.
Downey, P. J., & Wright, P. E. 2007. The ratio of the extreme to the sum in a
random sequence. Extremes, 10(4), 249–266.
Embrechts, P., Klüppelberg, C., & Mikosch, T. 1997.
Events. Berlin, Germany: Springer-Verlag.
Modelling Extremal
Maller, R. A., & Resnick, S. I. 1984. Limiting behaviour of sums and the term
of maximum modulus. Proc. London Math. Soc., 3(3), 385–422.
Nguyen, V. M. 2011 (Jun.). Wireless Link Modelling and Mobility Management
for Cellular Networks. Ph.D. thesis, Telecom ParisTech, France.
Nguyen, V. M. 2017. Network Densification. IET. Chap. 1, pages 3–24.
Nguyen, V. M., & Kountouris, M. 2016 (May). Coverage and capacity scaling
laws in downlink ultra-dense cellular networks. Pages 1–7 of: IEEE Int. Conf.
Commun. (ICC).
Nguyen, V. M., & Kountouris, M. 2017. Performance limits of network densification. IEEE J. Sel. Areas Commun., 35(6), 1294–1308.
Nguyen, V. M., Baccelli, F., Thomas, L., & Chen, C. S. 2010. Best signal quality in cellular networks: asymptotic properties and applications to mobility
management in small cell networks. EURASIP J. Wirel. Commun. Netw.,
2010(May), 690161.
8
O’Brien, G. L. 1980. A limit theorem for sample maxima and heavy branches
in Galton-Watson trees. J. Appl. Probab., 17(2)(Jun.), 539–545.
Rajan, A., Tepedelenlioğlu, C., & Zeng, R. 2017. A unified approach for modeling fading channels using infinitely divisible distributions. IEEE Trans. Veh.
Technol., 66(11), 10194–10206.
Shannon, C. E. 1948. A mathematical theory of communication. Bell Syst.
Tech. J., 27(3), 379–423.
Teugels, J. L. 1981. Limit theorems on order statistics. Ann. Probab., 9(5),
868–880.
Tse, David, & Viswanath, Pramod. 2005. Fundamentals of Wireless Communication. 1st edn. New York, U.S.A.: Cambridge University Press.
9
| 7cs.IT
|
arXiv:1611.04479v1 [math.AC] 14 Nov 2016
Giesbrecht’s algorithm, the HFE cryptosystem
and Ore’s ps-polynomials
Robert S. Coulter∗
George Havas
Marie Henderson
Abstract
We report on a recent implementation of Giesbrecht’s algorithm for
factoring polynomials in a skew-polynomial ring. We also discuss the
equivalence between factoring polynomials in a skew-polynomial ring and
decomposing ps -polynomials over a finite field, and how Giesbrecht’s algorithm is outlined in some detail by Ore in the 1930’s. We end with
some observations on the security of the Hidden Field Equation (HFE)
cryptosystem, where p-polynomials play a central role.
1
Introduction and Background
Let Fq denote the finite field with q = pe elements, p a prime. We use F∗q to
denote the non-zero elements of Fq . The polynomial ring in an indeterminate
X over any field K will be denoted by K[X] and for f, g ∈ K[X], f ◦ g = f (g)
represents the composition of f with g. We recall that a permutation polynomial
is a polynomial which permutes the elements of the finite field under evaluation.
A p-polynomial (sometimes called an additive or linearised polynomial) is a
polynomial L ∈ Fq [X] of the shape
X
i
L(X) =
ai X p
i
with ai ∈ Fq . More specifically, for any integer s, a ps -polynomial is a ppolynomial where ai = 0 whenever i is not a multiple of s. We note that
ps -polynomials are closed under composition (this is simply established).
The problem of completely decomposing a polynomial f ∈ K[X] into indecomposable factors, where K is a field, has a long and rich history. When K
is the complex plane, Ritt [23] showed that there exists an essentially unique
decomposition for any chosen polynomial. It is unique in the sense that for any
f ∈ K[X] in a complete decomposition of f : the number of factors is invariant;
∗ author now at Department of Mathematical Sciences, University of Delaware, Newark,
DE, 19716, U.S.A. A version of this paper appeared in the Proceedings of the 5th ASCM,
2001, pages 36–45, but is not widely available.
1
and the degrees of the factors are unique up to permutation. So, if we have two
complete decompositions
f = f1 ◦ · · · ◦ fm
= g1 ◦ · · · ◦ gn ,
then m = n and deg(fi ) = deg(gπ(i) ) for some permutation π of {1, . . . , m}.
Any class of polynomials defined over a field for which this property holds is
commonly said to satisfy Ritt’s theorem. The generalisation of Ritt’s theorem
to all fields of characteristic zero was carried out by Engstrom[10], and Levi[18].
However, for fields of non-zero characteristic, the situation is not so clearcut.
A polynomial is called wild if its degree is divisible by the characteristic p, and
tame otherwise. Any non-linear ps -polynomial is therefore a wild polynomial. A
distinction between the behaviour of wild and tame polynomials arises when one
considers Ritt’s theorem in the context of a finite field. Fried and MacRae[11]
showed that any tame polynomial satisfies Ritt’s theorem. However, Dorey and
Whaples[9] gave an example which showed that not all wild polynomials satisfied
Ritt’s Theorem. Other properties (not discussed in this article) of tame and wild
polynomials are also distinct. However, not all wild polynomials deviate from
tame polynomial behaviour. Specific to this question, Ore[19] showed in the
1930’s that p-polynomials satisfy Ritt’s theorem.
It is interesting to note that p-polynomials over a finite field appear to be
the second class of polynomials shown to satisfy Ritt’s theorem, after Ritt had
established the complex field case. This was not noted by Ore but is evident from
his work: see Ore[19] (Chapter 2, Theorem 4) which gives a statement equivalent
to Ritt’s theorem for p-polynomials. A further class of wild polynomials, known
as (ps , d)-polynomials (or, sub-linearised polynomials) can be shown to satisfy
Ritt’s theorem by using results of Henderson and Matthews[15].
Exponential-time algorithms for determining the complete decomposition of
polynomials were first given by Alagar and Thanh[1], and Barton and Zippel[2].
The first polynomial-time algorithm was published by Kozen and Landau[17],
and separately by Gutierrez, Recio and Ruiz de Velasco[14]. These results were
improved for the tame case over a finite field by von zur Gathen[12]. A general
purpose polynomial-time algorithm for finding a complete decomposition of a
rational function over an arbitrary field was given by Zippel[26]. This last
algorithm provides a method for decomposing any polynomial, wild or tame,
over a finite field. However, one should note that in the wild case, the algorithm
simply finds any complete decomposition, as there does not necessarily exist an
essentially unique decomposition.
2
Although p-polynomials were the first polynomials over a finite field shown
to satisfy Ritt’s theorem, they are the latest class of polynomials for which a
polynomial-time decomposition algorithm has been given. The algorithm we
refer to was described and analysed by Giesbrecht[13]. Giesbrecht presents his
algorithm in terms of factoring in skew-polynomial rings but it is well known
(and we later show) that the problem he considers is equivalent to decomposing
p-polynomials over a finite field. We note that any decomposition algorithm
for ps -polynomials can be adapted, at no computational cost, to decomposing
(ps , d)-polynomials. For (p, d)-polynomials this was shown by the authors[5],
following earlier work of Henderson and Matthews[15]. This can be extended
to all (ps , d)-polynomials using the work of Ore[19]. This subject is covered in
another paper under preparation by the authors.
In this article, we report on a successful implementation of Giesbrecht’s
algorithm, making some specific comments concerning the probabilistic part of
the algorithm. We also recall the work of Oystein Ore, showing how Giesbrecht’s
algorithm is equivalent to an algorithm described by Ore sixty years earlier. We
also consider implications of Ore’s work to the security of the Hidden Field
Equations (HFE) cryptosystem.
2
Giesbrecht’s algorithm and the work of Ore
Giesbrecht[13] introduces a probabilistic polynomial-time algorithm for obtaining a complete (essentially unique) factorisation of a polynomial in some classes
of skew-polynomial ring defined over a finite field. This problem is intimately
connected to the problem of determining an essentially unique complete decomposition of p-polynomials, a class of wild polynomials. In fact, there is a one-one
correspondence between factoring in a particular skew-polynomial ring over a
finite field and decomposing ps -polynomials over a finite field.
The skew-polynomial ring Fq [Y ; σ], where Y is an indeterminate and σ is
an automorphism of Fq , is a ring of polynomials with the usual componentwise addition, and with multiplication defined by Y a = σ(a)Y for any a ∈ Fq
(we simply use juxtaposition to represent multiplication in Fq [X] and Fq [Y ; σ]).
s
Since σ is an automorphism of Fq , we must have σ(a) = ap for some integer
s. Given the definition of multiplication above, it is easily seen that the skewpolynomial ring Fq [Y ; σ] is isomorphic to the ring of ps -polynomials over Fq with
the operations of polynomial addition and composition. Explicitly, the required
isomorphism Φ satisfies Φ(X p ) ◦ Φ(aX) = Y a = ap Y = Φ(ap X p ). From this it
follows that composition of ps -polynomials acts in exactly the same manner as
multiplication in the skew-polynomial ring Fq [Y, σ].
The theory introduced by Giesbrecht[13] is developed in its entirety in the
works of Ore[19, 20, 21]. It may be more efficient to implement Giesbrecht’s
algorithm using the ps -polynomial representation of the ring rather than the
skew-polynomial ring representation as set out in Giesbrecht’s article but this
is yet to be tested. While Giesbrecht refers to Ore[20], it is in Ore’s other
two papers that he develops the algorithm which Giesbrecht has rediscovered.
3
Giesbrecht’s key contribution is to find a way of computing the crucial step,
which is to find non-zero zero divisors in a small algebra. He does this by using
what he refers to as Eigen rings. Ore[19] discusses the same method in Chapter
2, Section 6 where he uses invariant rings. In particular, Ore’s Theorem 12 of
that section is the key idea in Giesbrecht’s algorithm. Of course, Ore develops
his theory in terms of ps -polynomials rather than skew-polynomial rings. Ore
obtains these results using an earlier paper, Ore[20], where he developed theory
on factoring and primality of polynomials in more general skew-polynomial rings
than discussed here. The problem of developing an algorithm for factoring
polynomials over any skew-polynomial ring remains open.
Recently, a successful implementation of Giesbrecht’s algorithm was produced by Larissa Meinecke at the University of Queensland using the Magma[4]
algebra package. There is one step in Giesbrecht’s algorithm which is probabilistic in nature, the rest of the algorithm is strictly deterministic. Giesbrecht
gives a lower bound for the probability of this step being successful as 1/9.
We have carried out some testing regarding this step which suggests this lower
bound is very conservative. While we have been unable to determine a worstcase scenario, in almost all cases tested, the step has been successful on the first
attempt.
3
HFE and p-polynomials
The Hidden Field Equation (HFE) cryptosystem was introduced by Patarin[22].
HFE is a public key cryptosystem and can be described as follows:
1. Choose a finite field Fq , q = pe , and a basis (β1 , . . . , βe ) for Fq over Fp .
2. Select a polynomial D of “relatively small degree” with the shape
X
i
j
aij X p +p
D(X) =
i,j
where aij ∈ Fq for all i, j.
3. Choose two p-polynomials, S and T , that permute Fq .
4. Calculate E(X) = S ◦ D ◦ T (X) mod (X q − X).
5. Calculate n1 , . . . , ne ∈ Fp [X1 , . . . , Xe ] satisfying
E(X) =
e
X
βi ni (X1 , . . . , Xe )
i=1
and publish Fq and the ni , 1 ≤ i ≤ e. The polynomials S, T and D are
the secret keys.
If someone wishes to send a message m to the owner of E(X), then they simply calculate E(m) = y and send y. Decryption is carried out by performing the
4
following steps. As S and T are permutation polynomials, they have functional
and compositional (modulo X q − X) inverses. As S and T are known to the
owner, they can determine the inverse polynomials modulo X q − X (note that
these inverses are also p-polynomials). Thus the recipient of the message y knows
S, D, T , S −1 and T −1 . They determine z satisfying S −1 (y) = z = D(T (m)).
Next they determine any m1 ∈ Fq so that D(m1 ) = z. Once m1 is chosen they
determine m = T −1 (m1 ). The middle step is only computationally feasible
because the degree of D is chosen to be “small”.
The security of the system relies on the assumption that if deg(E) is large,
then solving for m in E(m) = y is computationally infeasible. Note that several
m1 ∈ Fq may need to be tried to find a “sensible” message m. This is because
D is not necessarily chosen to be a permutation of Fq as the authors of HFE
assumed that this may be too difficult. However, Blokhuis et al.[3] have since
given examples of permutation polynomials from this class.
Note that it makes no difference whether the polynomial E or the set of e
polynomials ni is published if the basis used is known. In fact, an attacker need
not know the basis chosen as they may choose any basis to reconstruct a different, but effectively equivalent encryption function (see the discussion below). If
E is constructed from the e polynomials ni using a different basis, alternative
secret keys S, T and D may be obtained and used to decipher messages.
The HFE system is one of a family of cryptosystems which use functional
composition. Recently, some general attacks for these systems were developed
by Ye, Dai and Lam[25]. An attack which targets HFE specifically has been
published by Kipnis and Shamir[16]. This is general in nature and is quite
successful, but does not break HFE in all cases. This attack has since been
improved by Courtois[7].
Polynomials with the shape D are known as Dembowski-Ostrom (DO) polynomials, see Dembowski[8], Coulter and Matthews[6] and Blokhuis et al.[3]. For
any p-polynomial L ∈ Fq [X] and any DO polynomial D ∈ Fq [X], L◦D and D◦L
are both DO polynomials. In other words, DO polynomials are closed under
composition with p-polynomials. Also, it can be established that the reduction
of a DO polynomial modulo X q − X is again a DO polynomial. The HFE description given above works in exactly the same way as that given by Patarin[22]
precisely because of the above comments, coupled with the well known fact that
any function over Fq can be represented by a polynomial in Fq [X] of degree less
than q and a well known result concerning linear operators (discussed below).
Kipnis and Shamir[16] note several problems an attacker faces when they
consider this scheme. We address some of their concerns here. In the original
description of HFE, two linear transformations (or linear operators) over the
vector space Fep are chosen, rather than two linearised polynomials as described
above. Kipnis and Shamir comment that “these mixing operations have natural
interpretation over Fp but not over Fpe , and it is not clear apriori that the e
published polynomials over Fp can be described by a single univariate polynomial G over Fpe ”. In fact, there is a natural interpretation. Roman[24] (pages
184-5) shows that every linear operator on Fep can be represented by a linearised
polynomial over Fpe . So the description of HFE as given above is equivalent.
5
As DO polynomials are closed under composition with linearised polynomials
and their reduction modulo X q − X still results in a DO, we are guaranteed
that the published polynomials can be described by a single univariate polynomial: it must be a DO. Kipnis and Shamir continue “Even if it exists (a single
univariate polynomial), it may have an exponential number of coefficients, and
even if it is sparse, it may have an exponentially large degree which makes it
practically unsolvable”. As the resulting polynomial is a DO polynomial, it has
O(e2 ) terms (compare to a random polynomial which has O(pe ) terms), which
is not exponential. Certainly, the degree may be large. It remains our objective,
then, of finding a method of reducing the size of the degree.
We can make more comments concerning the univariate description of HFE
given above. Let E(X) be the public key, which is a DO polynomial. Suppose
we can determine p-polynomials L1 and L2 which are permutation polynomials
and satisfy L1 ◦ f ◦ L2 = E. Clearly, f must also be a DO polynomial. Then
we can decrypt any message sent to the owner of E using exactly the same
method used to decrypt in the standard way, but using the polynomials L1 , L2
and f , providing the degree of f is sufficiently small. Of course, it may not
be possible to determine p-polynomials that permute Fq which are left or right
decompositional factors of E(X). However, when considering this problem, the
following result by Coulter and Matthews[6], immediately draws our attention.
For any a ∈ Fq and any polynomial t ∈ Fq [X], define the difference polynomial
of t with respect to a by ∆t,a (X) = t(X + a) − t(X) − t(a).
Theorem 1 Let f ∈ Fq [X] with deg(f ) < q. The following conditions are
equivalent.
(i) f = D + L, where D is a Dembowski-Ostrom polynomial and L is a ppolynomial.
(ii) For each a ∈ F∗q , ∆f,a = La where La is a p-polynomial depending on a.
This result provides an alternative definition of DO polynomials and establishes
an important connection between DO polynomials and p-polynomials.
Let E be the published DO polynomial used in the HFE cryptosystem. We
wish to find L1 , L2 and D satisfying E = L1 ◦ D ◦ L2 . For the remainder, we set
f = D ◦ L2 so that E = L1 ◦ f and underline that f is also a DO polynomial.
Our objective is to determine some information regarding L1 . By Theorem 1,
∆E,a is a p-polynomial for any choice of a. Moreover, we have
∆E,a (X) = E(X + a) − E(X) − E(a)
= L1 (f (X + a)) − L1 (f (X)) − L1 (f (a))
= L1 (f (X + a) − f (X) − f (a))
= L1 ◦ ∆f,a .
Thus for any non-zero choice of a, the polynomial L1 is a left decompositional
factor of ∆E,a . Ore[20] shows that there exists a left and right decomposition algorithm similar to the well known greatest common divisor algorithm for a large
class of non-commutative polynomial rings (note that, in general, commutativity for composition does not hold). He uses these results in Ore[19] to establish
6
and describe such algorithms for p-polynomials specifically. In particular, using
a variant of the Euclidean algorithm, we can determine the Greatest Common
Left-Decompositional Factor (GCLDF) of two p-polynomials. This suggests the
following method of attack to determine the polynomial L1 .
1. Choose distinct elements a1 , a2 ∈ F∗q .
2. Calculate L(X) = GCLDF(∆E,a1 (X), ∆E,a2 (X)).
3. Test to see if L is a left decompositional factor of E. If it is, then L1 = L
and we are done.
4. If L is not a left decompositional factor of E, then choose a new a ∈ F∗q , distinct from previous choices, and calculate L(X) = GCLDF(L(X), ∆E,a (X)).
Return to Step 3.
We make the following observations. Step 3 can be carried out in time O(logp (deg(E)))
so has complexity much less than the Euclidean algorithm calculation required
in step 2 or 4. Note also that as Ore’s work does not extend to DO polynomials,
one cannot simply calculate GCLDF(L(X), E(X)) to obtain L1 .
As mentioned, Giesbrecht’s algorithm determines a complete decomposition
of a p-polynomial in probabilistic polynomial-time. However, this does not
mean we can determine L1 methodically by completely decomposing L after
step 2. Due to the nature of Ritt’s theorem, we are not guaranteed that in a
full decomposition the proper factors of L1 would be determined strictly on the
left. Further, the number of possible full decompositions is exponential in the
number of indecomposable factors. We make no claims at this point concerning
the number of GCLDF calculations required in step 4 to determine L1 . It may
require O(q) such calculations, making the algorithm no better than exhaustive
search. Finally, we note that this attack does not necessarily break HFE as
the DO polynomial may not have a non-trivial GCLDF and even if it did then
the resulting DO polynomial may not be of “sufficiently small” degree. We
are undertaking further research to analyse this attack and to determine other
methods of attacking HFE using the connections between the DO polynomial
and ps -polynomial classes.
Acknowledgments
This work was supported by the Australian Research Council.
References
[1] V.S. Alagar and M. Thanh, Fast polynomial decomposition algorithms, Eurocal ’85 (B. Buchberger and B.F. Caviness, eds.), Lecture Notes in Computer Science, vol. 204, 1985, pp. 150–153.
7
[2] D.R. Barton and R. Zippel, Polynomial decomposition algorithms, J. Symbolic Comput. 1 (1985), 159–168.
[3] A. Blokhuis, R.S. Coulter, M. Henderson, and C.M. O’Keefe, Permutations
amongst the Dembowski-Ostrom polynomials, Finite Fields and Applications: proceedings of the Fifth International Conference on Finite Fields
and Applications (D. Jungnickel and H. Niederreiter, eds.), 2001, pp. 37–42.
[4] W. Bosma, J. Cannon, and C. Playoust, The Magma algebra system I: The
user language, J. Symbolic Comput. 24 (1997), 235–265.
[5] R.S. Coulter, G. Havas, and M. Henderson, Functional decomposition of a
class of wild polynomials, J. Combin. Math. Combin. Comput. 28 (1998),
87–94.
[6] R.S. Coulter and R.W. Matthews, Planar functions and planes of LenzBarlotti class II, Des. Codes Cryptogr. 10 (1997), 167–184.
[7] N.T. Courtois, The security of Hidden Field Equations (HFE), Topics in
Cryptology - CT - RSA (D.N. Naccache, ed.), Lecture Notes in Computer
Science, vol. 2020, 2001, pp. 266–281.
[8] P. Dembowski and T.G. Ostrom, Planes of order n with collineation groups
of order n2 , Math. Z. 103 (1968), 239–258.
[9] F. Dorey and G. Whaples, Prime and composite polynomials, J. Algebra
28 (1974), 88–101.
[10] H.T. Engstrom, Polynomial substitutions, Amer. J. Math. 63 (1941), 249–
255.
[11] M.D. Fried and R.E. MacRae, On the invariance of chains of fields, Illinois
J. Math. 13 (1969), 165–171.
[12] J. von zur Gathen, Functional decomposition of polynomials: the tame case,
J. Symbolic Computation 9 (1990), 281–299.
[13] M. Giesbrecht, Factoring in skew-polynomial rings over finite fields, J. Symbolic Comput. 26 (1998), 463–486.
[14] J. Gutierrez, T. Recio, and C. Ruiz de Velasco, A polynomial decomposition algorithm of almost quadratic complexity, AAECC 1988: proceedings
(T. Mora, ed.), Lecture Notes in Computer Science, vol. 357, 1989, pp. 471–
476.
[15] M. Henderson and R. Matthews, Composition behaviour of sub-linearised
polynomials over a finite field, Finite Fields: Theory, Applications and Algorithms (R.C. Mullin and G.L. Mullen, eds.), Contemporary Mathematics,
vol. 225, American Mathematical Society, 1999, pp. 67–75.
8
[16] A. Kipnis and A. Shamir, Cryptanalysis of the HFE public key cryptosytem
by relinearization, Advances in Cryptology – Crypto ’99 (M. Wiener, ed.),
Lecture Notes in Computer Science, vol. 1666, 1999, pp. 19–30.
[17] D. Kozen and S. Landau, Polynomial decomposition algorithms, J. Symbolic
Computation 7 (1989), 445–456.
[18] H. Levi, Composite polynomials with coefficients in an arbitrary field of
characteristic zero, Amer. J. Math. 64 (1942), 389–400.
[19] O. Ore, On a special class of polynomials, Trans. Amer. Math. Soc. 35
(1933), 559–584, Errata, ibid. 36, 275 (1934).
[20] O. Ore, Theory of non-commutative polynomials, Annals of Math. 34
(1933), 480–508.
[21] O. Ore, Contributions to the theory of finite fields, Trans. Amer. Math. Soc.
36 (1934), 243–274.
[22] J. Patarin, Hidden fields equations (HFE) and isomorphisms of polynomials
(IP): two new families of asymmetric algorithms, Advances in Cryptology
– Eurocrypt ’96 (U. Maurer, ed.), Lecture Notes in Computer Science, vol.
1070, 1996, pp. 33–48.
[23] J.F. Ritt, Prime and composite polynomials, Trans. Amer. Math. Soc. 23
(1922), 51–66.
[24] S. Roman, Field Theory, Graduate Texts in Mathematics, vol. 158,
Springer-Verlag, 1995.
[25] D.F. Ye, Z.D. Dai, and K.Y. Lam, Decomposing attacks on asymmetric
cryptography based on mapping compositions, J. Cryptology 14 (2001), 137–
150.
[26] R. Zippel, Rational function decomposition, Proc. ACM Symp. Symbolic
and Algebraic Computation, 1991, pp. 1–6.
9
| 0math.AC
|
arXiv:1801.07197v1 [math.GR] 22 Jan 2018
PROBABILISTICALLY NILPOTENT GROUPS
ANER SHALEV
Abstract. We show that, for a finitely generated residually finite
group Γ, the word [x1 , . . . , xk ] is a probabilistic identity of Γ if and
only if Γ is virtually nilpotent of class less than k.
Related results, generalizations and problems are also discussed.
1. Introduction
A well known result of Peter Neumann [N] shows that a finite group
G in which the probability that two random elements commute is at
least ǫ > 0 is bounded-by-abelian-by-bounded; this means that there
are normal subgroups N, K of G such that K ≤ N, N/K is abelian,
and both |G/N| and |K| are bounded above by some function of ǫ.
The probability that two elements commute received considerable
attention over the years, see for instance [G], [J], [LP], [GR], [GS], [H],
[NY], [E]. However, the natural extension to longer commutators and
the probability of them being 1 remained unexplored. Here we shed
some light on this problem, providing some results, characterizations
and directions for further investigations.
Neumann’s result, as well as a similar result of Mann on groups with
many involutions [M1], can be viewed in the wider context of the theory
of word maps (see for instance the survey paper [S] and the references
therein) and the notion of probabilistic identities.
A word w = w(x1 , . . . , xk ) is an element of the free group Fk on
x1 , . . . , xk . Given a group G, the word w induces a word map wG :
Gk → G induced by substitution. We denote the image of this word
map by w(G).
If G is finite then w induces a probability distribution PG,w on G,
given by
−1
PG,w (g) = |wG
(g)|/|G|k .
A similar distribution is defined on profinite groups G, using their
normalized Haar measure.
2010 Mathematics Subject Classification: Primary 20E26; Secondary 20P05.
The author was partially supported by ERC advanced grant 247034, BSF grant
2008194, ISF grant 1117/13 and the Vinik Chair of mathematics which he holds.
1
2
ANER SHALEV
This distribution has been studied extensively in recent years, with
particular emphasis on the case where G is a finite simple group – see
[DPSS], [GS], [LS1], [LS2], [LS3], [LS4], [B]. Here we focus on general
finite groups and residually finite groups, and the proofs of our results
do not use the Classification of finite simple group.
Let Γ be a residually finite group. Recall that a word w 6= 1 is
said to be a probabilistic identity of Γ if there exists ǫ > 0 such that
for each finite quotient H = Γ/∆ of Γ we have PH,w (1) ≥ ǫ. This
b of Γ, we
amounts to saying that, in the profinite completion G = Γ
have PG,w (1) > 0. We shall also be interested in finite and profinite
groups G in which PG,w (g) ≥ ǫ > 0 for some element g ∈ G.
While Neumann’s result deals with the commutator word [x1 , x2 ] of
length two, here we consider commutator words of arbitrary length.
Define inductively w1 = x1 and wk+1 = [wk , xk+1 ]. Thus wk is the left
normed commutator [x1 , . . . , xk ].
Our main result characterizes finitely generated residually finite groups
in which such a word is a probabilistic identity. Clearly, if Γ has a nilpob then
tent normal subgroup of finite index m and class < k, and G = Γ,
−k
PG,wk (1) ≥ m > 0, so wk is a probabilistic identity of Γ. It turns out
that the converse is also true.
Theorem 1.1. Let Γ be a finitely generated residually finite group, and
let k be a positive integer. Then the word [x1 , . . . , xk ] is a probabilistic
identity of Γ if and only if Γ has a finite index normal subgroup ∆
which is nilpotent of class less than k.
In fact our proof shows more, namely: if Γ (or its profinite completion) is generated by d elements, and for some fixed ǫ > 0 and every
finite quotient H of Γ there exists h ∈ H such that PH,wk (h) ≥ ǫ, then
d
the index |Γ : ∆| of the nilpotent subgroup ∆ above divides n!n! , where
n = ⌊k/ǫ⌋.
Since a coset identity is a probabilistic identity we immediately obtain the following.
Corollary 1.2. Let Γ be a finitely generated residually finite group, and
let k be a positive integer. Suppose there exist a finite index subgroup ∆
of Γ, and elements g1 , . . . , gk ∈ Γ satisfying [g1 ∆, . . . , gk ∆] = 1. Then
there exists a finite index subgroup ∆0 of Γ such that γk (∆0 ) = 1.
Here and throughout this paper γk (G) denotes the kth term of the
lower central series of a group G. Theorem 1.1 follows from an effective
result on finite groups, which is of independent interest. To state it we
need some notation.
PROBABILISTICALLY NILPOTENT GROUPS
3
For groups G and H we define
G(H) = ∩φ:G→H ker φ,
namely, the intersection of all kernels of homomorphisms from G to
H. Then G(H) ⊳ G, G/G(H) ≤ H l for some l (which is finite if G is
finite). Moreover, G/G(H) satisfies all the identities of H, namely it
lies in the variety generated by H.
The case H = Sn , the symmetric group of degree n, will play a role
below. For a positive integer n define
L(n) = lcm(1, 2, . . . , n),
where lcm stands for the least common multiple. Clearly, L(n) is
the exponent of Sn . It is well known (and easy to verify using the
Prime Number Theorem) that L(n) = e(1+o(1))n . There are considerably shorter identities for Sn , but the length of its shortest identity is
still unknown.
Theorem 1.3. Fix a positive integer k and a real number ǫ > 0.
Let wk = [x1 , . . . , xk ]. Let G be a finite group and suppose that for
some g ∈ G we have PG,wk (g) ≥ ǫ. Set n = ⌊k/ǫ⌋ and let N = G(Sn ).
Then N is nilpotent of class less than k.
Furthermore, G/N ≤ Snl for some l, and it satisfies all the identities
of Sn . In particular, G/N has exponent dividing L(⌊k/ǫ⌋).
The following is an immediate consequence for residually finite groups
which are not necessarily finitely generated.
Corollary 1.4. If wk is a probabilistic identity of a residually finite
group Γ, then Γ is an extension of a nilpotent group of class less than
k by a group of finite exponent.
In [LS3] the following problem is posed.
Problem 1.5. Do all finitely generated residually finite groups Γ which
satisfy a probabilistic identity w satisfy an identity?
This seems to be a rather challenging problem. Till recently the only
non-trivial cases where a positive answer was known were w = [x1 , x2 ]
and w = x21 .
In [LS3] an affirmative answer to Problem 1.5 is given for all words
w, provided the group Γ is linear.
In [LS4] it is shown that, if a residually finite group Γ (not necessarily finitely generated) satisfies a probabilistic identity, then the
non-abelian upper composition factors of Γ have bounded size. This
leads to solutions of problems from [DPSS] and [B].
4
ANER SHALEV
The next result provides a positive answer to Problem 1.5 for additional words w; it also deals with groups which are not finitely generated.
Corollary 1.6. Every residually finite group which satisfies the probabilistic identity wk satisfies an identity.
Indeed, Corollary 1.4 shows that our group satisfies the identity
[xc1 , . . . , xck ] for some positive integer c.
Definition 1.7. A word 1 6= w ∈ Fk is said to be good if for any real
number ǫ > 0 there exists a word 1 6= v ∈ Fm (for some m) depending
only on w and ǫ such that, if G is a finite group satisfying PG,w (g) ≥ ǫ
for some g ∈ G, then v is an identity of G.
For example, results from [N], [M1] and [M2] imply that the words
[x1 , x2 ] and x21 are good. It is easy to see that, if w is a good word,
and v is any word disjoint from w (namely, their sets of variables are
disjoint), then the words wv and vw are also good.
Theorem 1.3 above shows that the words [x1 , . . . , xk ] are good for all
k. In fact, we can generalize the latter result as follows.
Proposition 1.8. Let w(x1 , . . . , xk ) be a word, and let w ′ (x1 , . . . , xk+1 ) =
[w(x1 , . . . , xk ), xk+1]. Suppose w is good. Then so is w ′ .
Induction on k immediately yields the following.
Corollary 1.9. The words [x21 , x2 , . . . , xk ] (k ≥ 2) are good.
In fact we also obtain a related structure theorem as follows.
Theorem 1.10. (i) Let G be a finite group, let w = [x21 , x2 , . . . , xk ],
and suppose PG,w (g) ≥ ǫ > 0 for some g ∈ G. Then there exists
n = n(k, ǫ) depending only on k and ǫ, such that G(Sn ) is nilpotent of
class at most k.
(ii) If w above is a probabilistic identity of a finitely generated residually finite group Γ, then Γ has a finite index subgroup which is nilpotent
of class at most k.
2. Proofs
In this section we prove Proposition 1.8, Theorem 1.3, Theorem 1.1
and Theorem 1.10, which in turn imply the other results stated in the
Introduction.
PROBABILISTICALLY NILPOTENT GROUPS
5
Lemma 2.1. Let w(x1 , . . . , xk ) be a word, and let w ′(x1 , . . . , xk+1 ) =
[w(x1 , . . . , xk ), xk+1]. Let G be a finite group, and suppose PG,w′ (g) ≥
ǫ > 0 for some g ∈ G. Choose g1 , . . . , gk ∈ G uniformly and independently. Then, for every 0 < δ < ǫ,
P rob(|G : CG (w(g1 , . . . , gk ))| < 1/δ) > ǫ − δ.
Proof. Choose gk+1 ∈ G also uniformly and independently. Then
P rob([w(g1, . . . , gk ), gk+1] = g) = PG,w′ (g) ≥ ǫ.
Given g1 , . . . , gk , g ∈ G, the number of elements gk+1 ∈ G satisfying
[w(g1, . . . , gk ), gk+1] = g is at most |CG (w(g1 , . . . , gk ))|. This yields
X
|CG (w(g1, . . . , gk ))|.
ǫ|G|k+1 ≤
g1 ,...,gk ∈G
Let
p = P rob(|CG (w(g1, . . . , gk ))| > δ|G|) = P rob(|G : CG (w(g1, . . . , gk ))| < 1/δ).
Then we obtain
ǫ ≤ p + (1 − p)δ,
so p ≥ (ǫ − δ)/(1 − δ) > ǫ − δ, as required.
Proposition 2.2. Let G, k, w, w ′, ǫ be as in Lemma 2.1. Suppose
PG,w′ (g) ≥ ǫ for some g ∈ G and let 0 < δ < ǫ. Let N = G(Sn ),
where n = ⌊1/δ⌋, and set M = CG (N). Then PG/M,w (1) > ǫ − δ.
Proof. Using Lemma 2.1 we obtain
P rob(|G : CG (w(g1 , . . . , gk ))| < 1/δ) > ǫ − δ.
Clearly, if |G : CG (w(g1, . . . , gk ))| < 1/δ then the permutation representation of G on the cosets of CG (w(g1, . . . , gk )) gives rise to a homomorphism φ : G → Sn satisfying G(Sn ) ≤ ker(φ) ≤ CG (w(g1 , . . . , gk )).
This implies that
P rob(N ≤ CG (w(g1 , . . . , gk )) > ǫ − δ.
Since M = CG (N) we obtain
P rob(w(g1, . . . , gk ) ∈ M) > ǫ − δ,
so PG/M,w (1) > ǫ − δ, as required.
We now prove Proposition 1.8.
6
ANER SHALEV
Proof. Recall that w ∈ Fk is a good word and w ′ = [w, xk+1 ].
To show that w ′ is good, suppose PG,w′ (g) ≥ ǫ > 0 for some g ∈ G.
Set δ = ǫ/2, n = ⌊1/δ⌋ = ⌊2/ǫ⌋ and apply Proposition 2.2. We obtain
PG/M,w (1) > ǫ/2, where M = CG (G(Sn )). Since w is good there is a
word 1 6= v ∈ Fm depending on w and ǫ such that v(G) ≤ M.
L(n)
Since g L(n) ∈ G(Sn ) for all g ∈ G, it follows that v ′ = [v, xm+1 ] is an
identity of G, which depends on w ′ and ǫ. Therefore w ′ is good.
Since x1 is a good word, it now follows from Proposition 1.8 by
induction on k that all commutator words [x1 , . . . , xk ] are good. We
can now also prove the more refined Theorem 1.3.
Proof. We prove, by induction on k ≥ 1, that, under the assumptions
of the theorem, for n = ⌊k/ǫ⌋ and N = G(Sn ), we have γk (N) = 1.
The other statements of the theorem follow immediately.
If k = 1 then |G|−1 = PG,x1 (g) ≥ ǫ for some g ∈ G. This yields
|G| ≤ 1/ǫ. Let n = ⌊1/ǫ⌋. Then N = G(Sn ) = 1, which yields the
induction base.
Now, suppose the theorem holds for k and we prove it for k + 1.
We assume PG,wk+1 (g) ≥ ǫ and let n = ⌊(k + 1)/ǫ⌋, N = G(Sn ) and
M = CG (N).
Using Proposition 2.2 with w = wk , w ′ = wk+1 and δ = ǫ/(k + 1) we
obtain
PG/M,wk (1) > kǫ/(k + 1).
By induction hypothesis this implies that (G/M)(S⌊k/(kǫ/(k+1))⌋ ) is nilpotent of class less than k. Since k/((kǫ/(k + 1)) = (k + 1)/ǫ we see that
(G/M)(Sn ) is nilpotent of class less than k. Clearly (G/M)(Sn ) ≥
G(Sn )M/M, and this yields
γk (G(Sn )) ≤ M.
Therefore
γk+1(N) = [γk (N), N] ≤ [M, N] = 1.
This completes the proof.
We can now prove Theorem 1.1. This result follows easily from Corollary 1.4 using Zelmanov’s solution to the Restricted Burnside Problem,
which, for general exponents, also relies on the Classification of finite
simple groups. However, we are able to provide an elementary proof of
Theorem 1.1 which avoids these very deep results.
PROBABILISTICALLY NILPOTENT GROUPS
7
Proof. It suffices to show that, if G is a d-generated finite group satisfying PG,wk (1) ≥ ǫ, then G has a normal subgroup N which is nilpotent
of class less than k, such that |G/N| is bounded above in terms of d, k, ǫ
only.
Using Theorem 1.3 and its notation, N = G(Sn ) is nilpotent of class
less than k. We also have G/N ≤ Snl for some l. Thus G/N is a
d-generated group lying in the variety generated by Sn .
A classical result of B.H. Neumann [Ne, 14.3] states that, for every
finite group H and a positive integer d, the free d-generated group in
d
the variety generated by H is finite of order dividing |H||H| . This
implies that
d
|G/N| ≤ n!n! .
Since n = ⌊k/ǫ⌋, |G/N| is bounded above in terms of d, k and ǫ.
This completes the proof of Theorem 1.1.
Note that the conclusion of Theorem 1.1 holds under the weaker
assumption that, for every finite quotient H of Γ there exists an element
h ∈ H such that PH,wk (h) ≥ ǫ > 0. Indeed this follows from Theorem
1.3 as above, replacing PG,wk (1) by PG,wk (g).
Finally, we prove Theorem 1.10.
Proof. We prove part (i) of the theorem by induction on k, starting
with k = 1 and w = x21 .
By Proposition 5 of [M2], if PG,x21 (g) ≥ ǫ > 0 then PG,[x1,x2 ] (1) ≥ ǫ2 .
By Theorem 1.3 this implies that G(Sn ) is abelian, where n = ⌊2/ǫ2 ⌋.
Now suppose the result holds for k ≥ 1 and let us prove it for k + 1.
Set w = [x21 , x2 , . . . , xk ] and w ′ = [w, xk+1].
We assume PG,w′ (g) ≥ ǫ > 0. Choose 0 < δ < ǫ (depending on
ǫ), and set m = ⌊1/δ⌋ and M = CG (G(Sm )). Then, by Proposition
2.2 we have PG/M,w (1) > ǫ − δ. By induction hypothesis there exists
n = n(k, ǫ − δ) such that (G/M)(Sn ) is nilpotent of class at most k.
This yields
γk+1(G(Sn )) ≤ M.
Therefore
[γk+1(G(Sn )), G(Sm )] ≤ [M, N] = 1.
Define n(k + 1, ǫ) = max(n(k, ǫ − δ), ⌊1/δ⌋). Then it follows that, for
n′ = n(k + 1, ǫ) = max(n, m) we have
γk+2(G(Sn′ )) = 1,
proving part (i).
Part (ii) follows from part (i) as in the proof of Theorem 1.1.
8
ANER SHALEV
By choosing δ to be a suitable function of ǫ (so that m = n at each
inductive step) one may obtain explicit good bounds on n(k, ǫ). We
leave this for the interested reader.
3. Related problems
We conclude with some natural questions and directions for further
research.
Problem 3.1. Characterize residually finite groups in which [x1 , . . . , xk ]
is a probabilistic identity.
In particular, are these groups virtually nilpotent of class less than
k?
The answer is positive for k = 2. Indeed, a result of Lévai and
Pyber [LP, 1.1(iii)] shows that a profinite group G in which [x1 , x2 ] is a
probabilistic identity has an open abelian subgroup, whose finite index
need not be bounded in terms of PG,[x1 ,x2 ] (1). This implies a similar
result for residually finite groups.
By Theorem 1.1, the answer to the question above is positive for
all k if the ambient group Γ (or its profinite completion) is finitely
generated. In the general case it follows that, for some n, Γ(Sn ) (which
may have infinite index in Γ) is nilpotent of class less than k. But the
reverse implication does not hold, as the product of infinitely many
copies of Sn demonstrates.
Problem 3.2. Are all non-identity words good?
This does not seem likely (or provable), so it would be nice to find
an example of a non-identity word which is not good.
Problem 3.3. Characterize the good words, or at least find more examples of them.
This is particularly interesting for some specific words.
Problem 3.4. Are power words xk1 good?
Let us say that general commutator words are words constructed
from the variables xk (k ≥ 1) in finitely many steps in which we pass
from previously constructed words w1 , w2 in disjoint sets of variables
to the word [w1 , w2 ].
For examples, let δ1 (x1 ) = x1 and
δk+1 (x1 , . . . , x2k ) = [δk (x1 , . . . , x2k−1 ), δk (x2k−1 +1 , . . . , x2k )].
PROBABILISTICALLY NILPOTENT GROUPS
9
Thus G satisfies the identity δk+1 if and only if it is solvable of derived
length at most k.
Problem 3.5. Are general commutator words good? Are the words δk
good?
A positive answer would of course follow from a positive answer to
the following.
Problem 3.6. Suppose w1 , w2 are good words in disjoint sets of variables. Does it follow that the word [w1 , w2 ] is good?
It would be nice to find analogues of Theorem 1.1 where we replace
nilpotency by solvability.
Problem 3.7. Let Γ be a finitely generated residually finite group and
suppose the word δk is a probabilistic identity of Γ. Does it follow that
Γ has a solvable subgroup ∆ of finite index? Can we further require
that the derived length of ∆ is less than k?
Finally, for a finite group G, set
P rk (G) = PG,wk (1),
the probability that [g1 , . . . , gk ] = 1 in G. Note that P r2 (G), denoted
in the literature by P r(G) and cp(G), was widely studied by various
authors. It was shown in [G] that the maximal value of P r(G) for G
non-abelian is 5/8.
Problem 3.8. Given k ≥ 3, find the maximal value of P rk (G) for
finite groups G satisfying γk (G) 6= 1.
The set {P r(G) : G a finite group} also received considerable attention; see [J], [H] and [E]. In the latter paper, Eberhard shows that this
set is well ordered by > and that its limit points are all rational.
Problem 3.9. For k ≥ 3, study the set {P rk (G) : G a finite group}.
Are its limit points all rational? Is it well ordered by >?
References
[B]
A. Bors, Fibers of automorphic word maps and an application to composition factors, J. Group Theory 20 (2017), 1103–1134.
[DPSS] J.D. Dixon, L. Pyber, Á. Seress, A. Shalev, Residual properties of free
groups and probabilistic methods, J. reine angew. Math. (Crelle’s) 556
(2003), 159–172.
[E]
S. Eberhard, Commuting probabilities of finite groups, Bull. London Math.
Soc. 47 (2015), 796–808.
10
ANER SHALEV
[GS]
S. Garion and A. Shalev, Commutator maps, measure preservation, and
T -systems, Trans. Amer. Math. Soc. 361 (2009), 4631–4651.
[GR]
R.M. Guralnick and G.R. Robinson, On the commutator probability in
finite groups, J. Algebra 300 (2006), 509–528; Addendum, J. Algebra 319
(2008), 18–22.
[G]
W.H. Gustafson, What is the probability that two group elements commute? Amer. Math. Monthly 80 (1973), 1031–1034.
[H]
P. Hegarty, Limit points in the range of the commuting probability function
on finite groups, J. Group Theory 16 (2013), no. 2, 235–247.
[J]
K. Joseph, Several conjectures on commutativity in algebraic structures,
Amer. Math. Monthly 84 (1977), 550–551.
[LS1] M. Larsen and A. Shalev, Fibers of word maps and some applications, J.
Algebra 354 (2012), 36–48.
[LS2] M. Larsen and A. Shalev, On the distribution of values of certain word
maps, Trans. Amer. Math. Soc. 368 (2016), 1647–1661.
[LS3] M. Larsen and A. Shalev, A probabilistic Tits alternative and probabilistic
identities, Algebra and Number Theory 10 (2016), 1359–1371.
[LS4] M. Larsen and A. Shalev, Words, Hausdorff dimension and randomly free
groups, Math. Ann., to appear.
[LP]
L. Lévai and L. Pyber, Profinite groups with many commuting pairs or
involutions, Arch. Math. (Basel) 75 (2000), no. 1, 1–7.
[M1]
A. Mann, Finite groups containing many involutions, Proc. Amer. Math.
Soc. 122 (1994), 383–385.
[M2]
A. Mann, Groups satisfying identities with high probability, Internat. J.
Algebra Comput., to appear.
[NY]
R.K. Nath and M.K. Yadav, On the probability distribution associated to
commutator word map in finite groups, Internat. J. Algebra Comput. 25
(2015), no. 7, 1107–1124.
[Ne]
B.H. Neumann, Identical relations in groups. I, Math. Ann. 114 (1937),
no. 1, 506–525.
[N]
P.M. Neumann, Two combinatorial problems in group theory, Bull. London
Math. Soc. 21 (1989), 456–458.
[S]
A. Shalev, Some problems and results in the theory of word maps, Erdős
Centennial, eds: Lovász et al., Bolyai Soc. Math. Studies 25 (2013), pp.
611–649.
E-mail address: [email protected]
Einstein Institute of Mathematics, Hebrew University, Givat Ram,
Jerusalem 91904, Israel
| 4math.GR
|
arXiv:1709.09503v1 [cs.CV] 27 Sep 2017
Modeling the Resource Requirements of Convolutional Neural
Networks on Mobile Devices
Zongqing Lu
Swati Rallapalli
Peking University
[email protected]
IBM Research
[email protected]
Kevin Chan
Thomas La Porta
Army Research Laboratory
[email protected]
Pennsylvania State University
[email protected]
ABSTRACT
Convolutional Neural Networks (CNNs) have revolutionized the
research in computer vision, due to their ability to capture complex patterns, resulting in high inference accuracies. However, the
increasingly complex nature of these neural networks means that
they are particularly suited for server computers with powerful
GPUs. We envision that deep learning applications will be eventually and widely deployed on mobile devices, e.g., smartphones,
self-driving cars, and drones. Therefore, in this paper, we aim to
understand the resource requirements (time, memory) of CNNs on
mobile devices. First, by deploying several popular CNNs on mobile CPUs and GPUs, we measure and analyze the performance and
resource usage for every layer of the CNNs. Our findings point
out the potential ways of optimizing the performance on mobile
devices. Second, we model the resource requirements of the different CNN computations. Finally, based on the measurement, profiling, and modeling, we build and evaluate our modeling tool, Augur,
which takes a CNN configuration (descriptor) as the input and estimates the compute time and resource usage of the CNN, to give
insights about whether and how efficiently a CNN can be run on
a given mobile platform. In doing so Augur tackles several challenges: (i) how to overcome profiling and measurement overhead;
(ii) how to capture the variance in different mobile platforms with
different processors, memory, and cache sizes; and (iii) how to account for the variance in the number, type and size of layers of the
different CNN configurations.
KEYWORDS
Convolutional neural networks; modeling; mobile devices
1
INTRODUCTION
Deep learning has become the norm of state-of-the-art learning
systems, especially in computer version. Convolutional Neural
Networks (CNNs) have demonstrated impressive performance on
various computer vision tasks from classification and detection to
segmentation and captioning. A CNN consists of different types
of layers (e.g., convolutional, pooling, fully connected), where each
layer performs certain transform on the input data and outputs the
data to the next layer. Different CNNs for computer vision tasks
have been designed, from a few layers to a thousand layers. But,
the core of these networks naturally are the convolutional layers,
which consist of a set of learnable kernels that are convolved across
the length and width of the input image to produce output features.
There are several frameworks that support the training (forward
and backward pass) and inference (only forward pass) phases of
CNNs, including Caffe [1], TensorFlow [6], Torch [8], Theano [7],
etc. All of these frameworks are designed and optimized for both
training and inference on computers with powerful GPUs.
However, we envision that deep learning applications will be
eventually and widely deployed on mobile devices. It is also expected that for computer vision tasks mobile devices will only perform inference (forward pass), since training can be carried out
offline by computers with powerful GPUs. In the rest of this paper,
the terms “inference”, “test” or “forward pass”, mean the same.
Since both the frameworks, as well as the CNN models are designed for computers with powerful GPUs, they may not effectively and efficiently work on mobile devices due to several factors,
e.g., constrained memory and limited computing capability. CNNs
for vision tasks are very complex – for example, VGGNet [21] has
528M parameters and requires over 15G FLOPs (FLoating-point
OPerations) to classify a single image. Due to the large amount
of parameters and FLOPs, and the need to enable running these
CNNs on resource-constrained mobile devices, several works focus on accelerating the computing of CNNs on mobile devices by
compressing parameters [16, 23], by cloud offload [11], and by distributing computation to heterogeneous processors on-board [18].
However, complementary to these techniques, our goal is to model
the resource requirements of CNNs accurately. Motivation for this
is that our system can serve guidelines to decide when performance
optimizations, offloading, etc. are required to successfully run analytics tasks on mobile devices. For instance, using the output of our
models, one could decide to run all the convolutional layers on the
mobile device while offloading the fully connected layers to the cloud
so as to cut down on the memory requirement on the mobile device.
Although accurately modeling the resource requirements of CNNs
is very hard, we make progress towards achieving it.
This paper overviews the workflow of CNNs, shares the experiences of deploying CNNs on mobile devices, gives the performance
measurements and analysis, and models the resource requirements
of the inference phase of the CNNs on mobile devices. In doing so
we face significant challenges. (i) Profiling overhead: to measure
timing of GPU computations, we need to add a synchronization
call that waits for all the results to come back before recording
the time. As pointed out by [3], this causes an overhead, as some
cores may be idling while waiting for the rest of the cores to complete the computation. We address this challenge by amortizing
this measurement cost by executing the computing task a large
number of times and averaging the running time. This ensures
that the overhead per iteration is negligible. (ii) Different types of
layers: CNNs are composed of various types of layers, so to model
the resource requirements of all the different types is challenging.
On the other hand, since main computation of all these layers boils
down to matrix multiplication, we are able to model the different
layers by abstracting out the details and focusing on the core of the
computation. (iii) How matrix multiplication scales: as the core of
the computation of CNNs, it is important to understand how the
computation scales with the sizes of matrices in terms of the resource requirements. Due to the large number of combinations of
matrix sizes, this can be very challenging. However, by extracting
the matrix multiplication sizes of popular CNNs, we observe that
all of them result into a small set of matrix sizes and thus we are
able to accurately model them for different mobile platforms.
Contributions: (i) We deploy the popular CNN models including AlexNet [17], VGGNet [21], GoogleNet [22], and ResNet [12]
using the Caffe framework [15] on mobile platforms (i.e., NVIDIA
TK1 and TX1), where the inference phase is run on both CPUs
and GPUs (§3). (ii) We measure and analyze the performance and
resource usage of the inference phase of these CNN models on a
layerwise granularity. Our findings point out the potential ways
of optimizing the computing of CNNs on mobile devices (§4). (iii)
We profile and model the resource requirements of CNNs. We also
build a modeling tool, Augur, which takes a CNN model descriptor
as the input and estimates the resource requirements of the CNN
so as to give insights on how well the CNN can be run on a mobile
platform without having to implement and deploy it (§5).
2 BACKGROUND
2.1 Overview of CNNs
Our goal is to model the resource requirements of the forward pass
of a CNN. The CNN architecture is typically composed of convolutional, normalization, and subsampling layers optionally followed
by fully connected layers. We overview these layers below, as it
lays the foundations for modeling the resource requirements.
Convolutional Layer: The convolutional (CONV) layers form the
core of CNNs. The parameters of this layer are a set of kernels
(weights) and biases learned during the training phase. During the
forward pass, kernels are convolved across the width, height, and
depth of the input, computing the dot product between the kernel
and the input and producing the output volume. Since the main
operation is dot product between the kernels and local regions of
the input, the forward pass of a CONV layer can be formulated as
a matrix multiplication. For the input volume, each local region
(a block of pixels) is stretched into a column of a matrix, and the
number of columns is the total number of local regions. The kernel
is stretched into a column of another matrix, and the number of
columns is the number of kernels. Finally, the product of the matrix
multiplication is reshaped to the output volume with a depth equal
to the number of kernels. For example, the input of AlexNet [227 ×
227 × 3] (width × height × depth) is convolved with 96 kernels
at size [11 × 11 × 3] and with a stride 4, and hence there are 55
locations along both width and height. So, the matrix for the input
is [3025 × 363], the matrix of the kernels is [363 × 96], and the
produced matrix is [3025× 96] and finally reshaped to [55× 55× 96].
The CONV layer is commonly implemented using the matrix
multiplication function of Basic Linear Algebra Subprograms (BLAS)
on CPUs and cuBLAS [2] on CUDA GPUs for acceleration. However, as many values in the input volume are replicated multiple
times in the matrix stretched from the input volume, it uses more
memory than the input volume itself.
Pooling Layer: The pooling (POOL) layer commonly sits between
CONV layers and performs downsampling to reduce the spatial
size (width and height). The pooling is performed on local regions
with the kernel size defined by a CNN model. The most common
pooling operation in the state-of-the-art CNN models is max pooling. The pooling layer independently operates on the input volume
without parameters, and hence its implementation is simple.
Normalization Layer: Two types of normalization layers are commonly used in CNNs: local response normalization (LRN) and batch
normalization (BatchNorm). However, LRN’s role has been outperformed by other techniques, such as BatchNorm, and thus here we
only detail BatchNorm.
BatchNorm is introduced to reduce the internal covariant shift
during training [14]. During test phase, BatchNorm normalizes the
input volume on each dimension (weight × height), e.g., for the i-th
dimension, as follows,
x (i ) − E[x (i ) ]
xb(i ) = p
,
Var[x (i ) ]
where E[x (i ) ] and Var[x (i ) ] are learned during the training phase
for dimension i.
Fully Connected Layer: Each neuron in a fully connected (FC)
layer is connected to all activations in the previous layer. Due
to the full connectivity, there are a huge number of parameters,
which places heavy burden on memory usage and computation.
Recently, FC layers have fallen out of favor, e.g., the latest CNNs,
i.e., GoogleNet and ResNet, only have one fully connected layer
as the classifier. This dramatically reduces the number of parameters, e.g., 26MB parameters in GoogleNet while 233MB in AlexNet.
Moreover, it was found that FC layers of VGGNet can be removed
with no performance reduction. Therefore, it is anticipated that
CNNs will eliminate the use of FC layers. The forward pass of FC
layers is also implemented as a matrix multiplication.
Besides these four layers, rectified linear unit (ReLU) layer that
applies an elementwise function, e.g., max(0, x), on the input volume, is also commonly used in CNNs. However, ReLU is simple,
has no parameters, and does not change the size of input volume.
Thus we skip the detail of ReLU layer.
2.2 Related Work
Although CNNs have been applied to various computer vision applications on different computing platforms, only a few works consider running CNNs on mobile devices, which we envision to be a
significant future area for the deployment of deep learning applications.
Among these works, many focus on accelerating the computing
of CNNs, e.g., by compressing parameters [16, 23], by cloud offload
[11], and by distributing computation to heterogeneous processors
on-board [18]. Some consider reducing the memory usage to better fit mobile devices while maintaining high inference accuracy,
Table 2: Timing benchmarks on AlexNet
Table 1: CNN models
Layer
CONV
POOL
NORM
ReLU
FC
Concat
Scale
Eltwise
Total
AlexNet
5
3
2
7
3
20
VGGNet
13
5
15
3
36
GoogLeNet
57
14
2
57
1
9
ResNet
53
2
53
49
1
140
53
16
227
Platform
CONV
318.7±0.2
51.42%
24.6±3.5
33.53%
66.9±5.3
7.48%
24.2±8.3
45.79%
666M
91.36%
CPU
TK1
GPU
CPU
TX1
GPU
FLOPs
Layerwise Pass (ms)
POOL
LRN
ReLU
6.1±0.1
103.8±0.0
4.6±0.0
0.99%
16.74%
0.75%
2.3±0.6
2.4±0.5
5.2±1.2
3.15%
3.22%
7.11%
7.6±0.0
172.4±0.3
2.4±0.0
0.85%
19.28%
0.27%
1.3±2.6
2.7±3.0
5.9±5.9
2.51%
5.12%
11.23%
1M
2M
0.7M
0.14%
0.27%
0.10%
FC
186.3±0.1
30.05%
35.1±5.9
47.95%
644.7±5.3
72.09%
15.2±4.7
28.76%
59M
8.09%
Total (ms)
Forward Pass (ms)
619.8±0.2
619.5±0.2
73.3±10.7
54.7±2.4
894.3±4.8
892.7±2.3
52.8±15.7
29.3±6.5
729M
Table 3: Timing benchmarks on VGGNet
Platform
CPU
TK1
GPU
CPU
TX1
GPU
FLOPs
CONV
7160.5±0.7
93.02%
263.1±19.3
75.68%
1952.9±12.2
69.14%
136.3±5.4
73.98%
15360M
99.08%
Layerwise Pass (ms)
POOL
ReLU
60.1±0.1
95.6±0.1
0.78%
1.24%
7.2±0.5
17.5±1.2
2.06%
5.03%
71.3±1.5
52.5±1.9
2.52%
1.86%
3.4±1.6
9.9±4.9
1.84%
5.35%
6M
14M
0.04%
0.09%
FC
381.6±0.2
4.96%
57.6±0.5
16.58%
747.7±24.9
26.47%
32.8±1.3
17.82%
124M
0.79%
Total (ms)
Forward Pass (ms)
7697.9±0.6
7697.8±0.5
347.6±20.1
326.7±2.1
2824.6±23.2
2809.1±10.6
184.2±7.4
175.3±2.0
Type/Platform
Weights & Biases
Data
Workspace
CPU
TK1
GPU
CPU
TX1
GPU
AlexNet
233
8
11
324
560
362
589
VGGNet
528
110
168
972
1508
1013
1537
GoogleNet
26
53
46
161
196
200
226
ResNet
97
221
79
409
533
453
562
15503M
e.g., [10, 13]. The resource bottlenecks of running CNNs on mobile devices are preliminarily investigated in [19]. Different CNNs
are benchmarked in [9], but it does not consider how to model the
resource requirements of CNNs.
While CNNs grow from a few layers to a thousand layers, the
computational capability of mobile devices continues to improve.
As a result, different mobile devices perform differently on different CNNs, and hence custom optimization and offloading may or
may not be needed. It depends on whether and how efficiently a
CNN can be run on a given mobile platform. This question motivates our work.
3
Table 4: Memory of CNN models on platforms (MB)
MEASUREMENT SET-UP
To understand the resource requirement of the forward pass of
CNNs, we deployed several CNN models on two mobile platforms
using the popular deep learning framework – Caffe.
Platforms: Although some frameworks (e.g., Caffe, Torch) can
run on Android and iOS, they do not support GPU acceleration
on off-the-shelf mobile devices, such as smartphones or tablets. To
understand the performance of CNNs on both mobile CPUs and
GPUs, in this paper, we focus on two developer kits for low power
edge devices – NVIDIA TK1 and TX1.
TK1 is equipped with a 2.3GHz quad-core ARM Cortex-15A 32bit
CPU, 192 CUDA cores Kepler GPU, and 2GB DDR3L RAM. TX1 is
more powerful and has a 1.9GHz quad-core ARM Cortex-A57 64bit
CPU, 256 CUDA cores Maxwell GPU, and 4GB LPDDR4 RAM. The
system-on-chip (including CPU and GPU) of TK1 and TX1 also appears in many off-the-shelf mobile devices, such as Google Nexus
9 and Pixel C. However, none of these devices are enabled to support CUDA, on which deep learning frameworks are built for GPU
acceleration. Thus, for ease of experimentation we choose NVIDIA
TK1 and TX1, the results of which should indicate the performance
of CNNs on mobile devices.
Framework: There are several frameworks for deep neural networks. As mentioned before, most of the frameworks use BLAS
on CPU and cuBLAS on GPUs for the CNN computations and thus
show similar performance. In this paper, we use the popular Caffe
framework, where the choice of BLAS is OpenBLAS [5].
CNN Models: For the measurement, we consider the most popular CNN models including AlexNet, VGGNet (VGG-16), GoogleNet,
and ResNet (ResNet-50). Although the architectures of these models are quite different, from several layers to more than one hundred layers and from regular stacked layers to branched and stacked
layers, they are mainly built on the basic layers of CNNs. Table 1
shows how many these layers each model contains.
4 INITIAL MEASUREMENT STUDY
In this section, we investigate the resource requirements and bottlenecks of running several well known CNN models on mobile
platforms.
4.1 Timing
First, we measure the timing of each model on different platforms
using CPU and GPU in terms of (i) complete forward pass: i.e., timing is measured for the entire forward pass and (ii) as summation
of individual layer times. We also calculate the number of FLOPs
for each model and each type of layer.
AlexNet has the least number of layers among these models and
indeed requires the least amount of computation in terms of FLOPs,
i.e., 729M. As shown in Table 2, on the CPU of both TK1 and TX1,
the summation of layerwise timing perfectly matches with that of
a full forward pass, which are about 600ms (on TK1) and 900ms
(on TX1). Surprisingly, although TX1 has a more powerful CPU, the
forward pass on TX1 is slower than TK1. The CONV layers on TX1
run much faster than on TK1 (more than 4x), but the FC layers are
much slower (more than 3x). Since the basic computation of both
Table 5: Timing benchmarks on GoogleNet
Platform
CPU
TK1
GPU
CPU
TX1
GPU
FLOPs
CONV
755.3±0.2
70.84%
186.9±45.0
69.40%
174.4±3.6
27.64%
165.9±48.8
64.28%
1585M
98.80%
POOL
68.8±0.1
6.45%
20.6±4.9
7.65%
89.9±0.2
14.24%
18.5±11.2
7.16%
13M
0.80%
Layerwise Pass (ms)
LRN
ReLU
214.3±0.2
22.8±0.0
20.10%
2.14%
6.5±1.5
35.3±9.9
2.40%
13.10%
349.4±0.6
9.5±0.1
55.38%
1.50%
3.3±2.3
49.5±31.2
1.28%
19.16%
3M
3M
0.20%
0.20%
Concat
2.0±0.0
0.19%
13.0±4.4
4.81%
1.7±0.1
0.36%
15.4±9.8
5.96%
FC
2.7±0.0
0.26%
2.4±0.8
0.90%
5.7±0.0
0.90%
1.2±1.1
0.46%
1M
0.06%
Total (ms)
Forward Pass (ms)
1066.2±0.3
1065.6±0.2
269.3 ±65.6
167.0±44.3
630.9±3.5
637.9±14.7
258.1±89.8
143.9±59.2
1606M
Table 6: Timing benchmarks on ResNet
Platform
CPU
TK1
GPU
CPU
TX1
GPU
FLOPs
CONV
1830.4±0.4
88.31%
245.8±16.3
36.53%
362.3±5.4
63.83%
279.4±42.6
42.03%
3866M
98.59%
POOL
8.8±0.0
0.42%
5.5±0.6
0.81%
13.7±0.2
2.41%
3.0±2.7
0.45%
2M
0.05%
Layerwise Pass (ms)
BatchNorm
ReLU
Scale
97.1±0.1
64.0±0.1
42.0±0.1
4.68%
3.09%
2.03%
249.5±11.6
38.7±2.0
76.0±3.3
37.08%
5.75%
11.29%
83.5±0.3
33.2±0.1
31.9±3.6
14.7%
5.86%
5.62%
198.1±36.8
63.6±31.3
79.8±24.2
29.80%
9.57%
12.01%
32M
9M
11M
0.81%
0.23%
0.27%
CONV and FC is matrix multiplication, the results seem contradictory at first. However, we investigate and explain the reasons for
the behavior below.
First, even though the clock is slower on TX1 compared to TK1,
i.e., 1.9 GHz vs. 2.3 GHz, TX1 runs more instructions per clock
cycle compared to TK1 (3 vs. 2) and hence the performance of
TX1 CPU is expected to be better than TK1 CPU as we see for
the CONV layers. Second, FC layers have many more parameters
than the CONV layers. Therefore, FC layers are bottlenecked by
the memory whereas CONV layers are compute bound. Third, the
L1 data cache size is 32 KB on both and L2 cache is larger on TK1
compared to TX1. Even if cache size is same on both – because the
address is longer on TX1 (64 bit vs. 32 bit), more memory is used
up for the addressing and we have lesser memory available to save
the data itself on the cache. This means that we need to fetch data
from RAM to the cache more often while executing the FC layers
on TX1 due to the large number of parameters which causes the
slow down.
GPUs can significantly accelerate the computation of a CNN
and thus improve the performance over CPUs. More advanced
TX1 GPU outperforms TK1 GPU as expected. However, we face
one challenge: the summation of layerwise timing does not match
the timing of the full forward pass on GPUs. The reason for the
mismatch is that CUDA supports asynchronous programming. Before time measurement, an API (cudaDeviceSynchronize) has to
be called to make sure that all cores have finished their tasks. This
explicit synchronization is the overhead of measuring time on the
GPUs. Therefore, the sum of layerwise timing on GPUs is longer
than a full forward pass.
VGGNet has 2x CONV layers compared to AlexNet (Table 1). However, the number of operations is 20x that of AlexNet because VGGNet uses much larger feature maps. While other results follow
similar pattern as AlexNet, the throughput of both CPU and GPU
Eltwise
24.8±0.1
1.20%
47.0±2.7
6.98%
20.4±4.2
3.59%
34.8±12.9
5.24%
6M
0.14%
FC
5.4±0.0
0.26%
3.9±0.1
0.58%
22.2±0.1
3.92%
1.8±2.4
0.27%
2M
0.05%
Total (ms)
Forward Pass (ms)
2072.7±0.4
2072.2±0.3
673.0±33.4
149.4±4.9
567.6±7.6
566.8±9.7
664.7±116.5
104.4±14.0
3922M
on VGGNet is higher than on AlexNet. For example, the throughput of TK1 CPU on AlexNet is 1 GFLOPS (GFLOPs per Second)
and of VGGNet is 2 GFLOPS. This is mainly because both CPU and
GPU have better throughput on matrix multiplication with larger
size.
GoogleNet has more than 50 CONV layers, many more than AlexNet.
However, the CONV layers have only two times more FLOPs than
that of AlexNet. The main reason is that the size of the kernels
and feature maps is small, which dramatically reduces the number
of operations. Similar to AlexNet, GoogleNet also employs LRN
that significantly affects the performance on CPU for both TK1
and TX1. For example, it takes more than 55% of total time on TX1
CPU. GoogleNet has a layer, named Concat, that does not involve
any computation, but concatenates the outputs from previous layers, thus involving memory operations only.
The difference between layerwise timing and full forward pass
on GoogleNet is much larger than AlexNet and VGGNet as shown
in Table 5. GoogleNet has many more layers than AlexNet and
VGGNet and thus much more measuring overhead on GPUs. The
measuring overhead may be larger than the compute time when
the computation of a layer does not cost much time, e.g., ReLU
layers. Due to this measurement artifact, in Table 5, ReLU layers
cost more time on GPUs than CPUs. This is a motivation for us to
devise measurement techniques that can overcome these measurement overheads as we see later in §5.
ResNet has more than two hundred layers. ResNet includes BatchNorm, Scale, and Eltwise that are not commonly used by other
models. These layers are not expensive in terms of FLOPs as shown
in Table 6. We observe that the computation of Scale and Eltwise
costs more on GPUs than CPUs, which is again due to the measurement overhead on GPUs as discussed above. Interestingly, although ResNet has more FLOPs (2x) than GoogleNet, a full forward pass is faster than GoogleNet on TX1. This is because LRN
of GoogleNet is very expensive: (55% of total time) on TX1 CPU.
Moreover, GoogleNet has more CONV layers and the underlying
matrix multiplication is smaller than that of ResNet. As GPU throughput is higher on matrix multiplication with larger size, ResNet is
faster than GoogleNet on TX1 GPU.
4.2
Memory
The memory requirement to run a CNN comes from three major
sources: (i) the memory that holds the parameters of the CNN; (ii)
the memory that stores intermediate data of the CNN; and (iii) the
workspace for computation. A majority of the CNN parameters
come from CONV and FC layers (i.e., weights and biases). Intermediate data is the output of each layer (i.e., the input of next layer),
e.g., feature maps. Some types of layers require additional space to
perform computation, e.g., on CONV layers, the memory is needed
to hold the matrix stretched from the input data for matrix multiplication. The workspace memory is mostly consumed by the matrix
multiplication of CONV layers. The NVIDIA CUDA Deep Neural
Network library (cuDNN) [4] can reduce the workspace by sacrificing the speed of computing on GPUs. However, as the workspace
is not the most significant part, cuDNN cannot reduce the memory
usage of CNNs significantly.
Table 4 shows the memory requirement of weights and biases of
CONV and FC layers, intermediate data, and workspace of CONV
layers for each CNN – by parsing the model descriptor (e.g., a prototxt file in Caffe). Table 4 also gives the measured memory usage of Caffe, running each CNN on these platforms. One can see
that deeper CNNs (from AlexNet to ResNet) may not require more
memory, especially for GoogleNet, which requires the least memory among them. Memory usage on TX1 is more than TK1, because
TK1 is running a 32-bit OS while TX1 is running a 64-bit OS, which
incurs more memory usage for the framework itself.
To speed up the computation of CNNs, all memory should be
allocated beforehand and not released during the computation. Although existing frameworks (e.g., Caffe1 ) follow this rule, they are
designed for training and testing (scoring) on workstations with
powerful GPUs, and thus not quite suitable for mobile devices in
terms of memory management.
Unified Memory Architecture: Unlike workstations2 where GPUs
have their dedicated memory, mobile platforms usually have a unified memory architecture, where GPU shares system memory with
CPU. On workstations, in the current implementation of Caffe,
data is transferred to and from the memory of GPU for access,
which is efficient on workstations. However, on united memory architecture, e.g., TK1 and TX1, memory transfer from CPU to GPU
simply generates a redundant data copy on system memory. As
shown in Table 4, on both TK1 and TX1, the memory usage on
GPU is always more than CPU, and the additional memory is actually used to hold a redundant copy of the parameters of each CNN
(mostly weights and biases). For example, running AlexNet on TK1
GPU takes 560MB memory, which is 236MB more than TK1 CPU,
1 Caffe allocates
the memory for intermediate data on demand (lazily) during the first
run, and thus it takes longer time than later runs.
2 Although GPUs on workstations can also directly access host memory over PCIe, e.g.,
CUDA kernels, reading data over PCIe is limited by PCIe bandwidth (up to 32GB/s)
which is much slower than reading data from GPU memory (limit 200GB/s).
while weights and biases of AlexNet are 233MB in total. This also
stands for other CNNs.
Mobile GPUs can directly access data by mapping host memory
without degrading performance and incurring memory transfer
overhead (i.e., zero-copy memory). Existing frameworks, including Caffe, Torch, and Theano, do not take into consideration the
unified memory architecture for mobile platforms. On the contrary, the unified memory architecture can be exploited to design
a tailored computing framework for mobile devices. (i) We can
eliminate memory transfers between CPU and GPU. (ii) We can
compute a CNN in the most efficient way; i.e., each layer can be
executed on the most efficient unit, switching back and forth between GPU and CPU, without incurring additional memory transfer overhead.
4.3 Analysis
FLOPs. As the throughput of both CPU and GPU is higher on the
CNN with more FLOPs and a significant amount of memory operations are involved in a CNN computation, FLOPs cannot accurately
reflect the compute time of a CNN. For example, ResNet is faster
than GoogleNet on GPUs, though it involves more FLOPs. Therefore, estimating the compute time of CNNs directly from their FLOPs
is not feasible.
CONV and FC Layer. The computation of CONV and FC layers
in most models accounts for a majority of FLOPs. Therefore, can
one measure these layers instead of the entire network? However,
this approach encounters other difficulties, i.e., layerwise measuring overhead on GPUs, and we have no way to know the exact
overhead for each layer, which is hidden by GPUs.
Matrix Multiplication. The core of CONV and FC layers are matrix multiplications. Therefore, rather than going into the details
of each of the individual layers, if we are able to extract the matrix multiplication part of the layer, we will be able to accurately
capture the resource requirements of these layers.
5 AUGUR
We aim to build a modeling tool that can estimate the resource
requirements of any given CNN descriptor on specific mobile platforms without implementation and deployment. This way, we can
take the costs into consideration during the design of a CNN. This
is critical when designing CNNs for resource-constrained mobile
devices.
5.1 Profiling
The basic idea is simple. We first find the matrix multiplications
that form the core of the CNN computation. Then we measure
their performance based on the BLAS and cuBLAS libraries, which
are commonly used for matrix multiplications on CPUs and GPUs
respectively.
Extract matrix sizes: To find all matrix multiplications and their
sizes, we need to parse the descriptor of a CNN. The dimension of
input (e.g., images and feature maps) and network parameters (e.g.,
convolution kernels) determines two matrix sizes (that are to be
multiplied) at a CONV or FC layer. As the dimension of feature
maps can be changed by some other layers, e.g., POOL layers, we
Forward Pass
2
ResNet
compute time (ms)
(a) TK1 CPU
·103
88.85%
3
2495.9
2
78.01%
1
696.4 892.7
compute time (ms)
Matrix Mul.
24.93%
637.9
159
Forward Pass
71.45%
405
1
(64, 562 )
2
(256,
(384,28
142))
0 72 )
(96,
6
VGGNet
0
2000
n
ResNet
(b) TX1 CPU
·102
80.99%
4
264.6
2
GoogleNet
326.7
Matrix Mul.
19.40%
167
70.75%
38.7
54.7
4
GoogleNet
78.44%
2
137.5
143.9
175.3
29.3
28.4
VGGNet
GoogleNet
(96, 64 × 12 )
3 × 103
2
5 × 103
0
2000
m
× 103
(64 × 52 , 282 )
0
(64 × 32 , 72 )
3 × 103
(256 × 32 , 142 )
2 × 103
0
1 × 103
1 × 103
k
n
·103
ResNet
Forward Pass
60.35%
63
2
2 × 103
3 × 103
1 × 103
m
k
282
(c) effect of m and k , where n = 256
·102
(64, 562 )
(256,
(384,28
1422))
0 72 )
(96,
10 × 103
0
2000
n
ResNet
(d) TX1 GPU
·103
6
(512, 1122 )
(128, 1122 )
(32, 1122 )
1
0
AlexNet
0
1
149.4
104.4
73.72%
21.6
Matrix Mul.
19.74%
2
57.03%
(c) TK1 GPU
·102
3
1
VGGNet
(64, 256 × 32 )
(384, 128 × 32 )
(32, 64 × 52 )
(256, 64 × 32 )
4
(64 × 72 , 562 )
(128 × 32 , 562 )
(64 × 12 , 562 )
Figure 2: Matrix multiplication on TK1 CPU with varying n, m, and k.
0
AlexNet
(128, 64 × 72 )
(a) effect of n and m , where k = 576 (b) effect of n and k , where m =
Forward Pass
85.2
32.4
10 × 103
566.8
·103
6
(128, 1122 )
(32, 1122 )
0
AlexNet
compute time (ms)
2809.1
·102
compute time (ms)
GoogleNet
·103
compute time (ms)
VGGNet
91.16%
1889 2072.2
compute time (ms)
747.5 1065.6
493.2 619.5
AlexNet
4
(512, 64 × 72 )
(512, 1122 )
70.15%
79.61%
compute time (ms)
Matrix Mul.
4
(512, 64 × 72 )
(128, 64 × 72 )
(64, 256 × 32 )
(32, 64 × 52 )
2
0
(384, 128 × 32 )
(256, 64 × 32 )
(96, 64 × 12 )
5 × 103
0
3 × 103
2 × 103
2000
m
n
1
compute time (ms)
96.01%
7390.6 7697.8
compute time (ms)
compute time (ms)
·103
10
8
6
4
2
0
1
0
(64 × 72 , 562 )
(64 × 32 , 72 )
k
(64 × 72 , 562 )
(64 × 52 , 282 )
(256 × 32 , 142 )
0
1 × 103
× 103
(128 × 32 , 562 )
2 × 103
3 × 103
3 × 103
2 × 103
1 × 103
m
k
(a) effect of n and m , where k = 576 (b) effect of n and k , where m = 282 (c) effect of m and k , where n = 256
Figure 1: Matrix multiplication and forward pass of AlexNet, VGGNet, GoogleNet,
and ResNet on mobile platforms.
Figure 3: Matrix multiplication on TX1 CPU with varying n, m, and k.
need to trace the dimension of feature maps layer by layer. However, this can be easily done by parsing the parameter settings at
each layer, such as zero-padding (P), stride (S), the number of output feature maps (N ). For instance, in case of a CONV layer, let I
denote the spatial dimension of the input feature map, O denote
the spatial dimension of the output feature map, K denote the 3D
volume of the convolution kernels. Then, we have:
Ow = ⌊(Iw − K w + 2P)/S⌋ + 1
Oh = ⌊(Ih − K h + 2P)/S⌋ + 1.
Then, the matrix multiplication at the CONV layer is [(Ow · Oh ) ×
(K w · K h · Kd )][(K w · K h · Kd ) × N ].
Mitigate measurement overhead: Layerwise timing measurement incurs heavy overhead on GPUs and causes a large deviation
from a full forward pass. Moreover, the overhead is not fixed and
varies over each measurement. As illustrated in Table 5 and 6, the
measurement overhead (the difference between the sum of layerwise measurements and full forward pass) of GoogleNet (131 measurements) on TX1 GPU is 128 ms, while the overhead of ResNet
(227 measurements) is 595 ms. Therefore, we need a way to mitigate the overhead for accurate timing of matrix multiplications.
Timing measurements on GPUs can only been recorded after
all cores finish their tasks. In a full forward pass, timing is only
recorded at the last layer. Therefore, a core may be assigned with
the computation of following layers and thus it can continuously
perform the computation without synchronization. For example,
after finishing the multiply-add operations for the matrix multiplication at a CONV layer, a core can continue to calculate the max
function of next ReLU layer on the output of multiply-add operations. If layerwise timing is recorded, all cores have to wait until all
multiply-add operations of the CONV layer have been completed.
The idea of mitigating the measurement overhead is simple. To
benchmark a matrix multiplication, we keep GPUs iteratively running the matrix multiplication in a way that GPU cores can continuously perform multiply-add operations without synchronization,
before recording the end time. Then, the measurement overhead is
amortized over all the iterations, giving accurate timing estimates.
When the number of iterations is large enough, the overhead is
negligible. In our experiments we measure the timing of a large
number of computing iterations on a matrix multiplication and
use the averaged value of each iteration as the compute time of
the matrix multiplication.
Fraction of forward pass spent by matrix multiplication: In
Figure 1, we study the fraction of forward pass time spent by matrix multiplication (matmul) operations. We do so, by extracting
the matmul operations, measuring them, and then comparing with
the full forward pass measurement. Note that due to the above explained averaging methodology, measurement overhead for matmul operations in this section is negligible.
First, as seen in Figure 1a, matmul operations on TK1 CPU take
a large portion of forward pass time – 79.61%, 96.01%, 70.15%, and
91.16% for AlexNet, VGGNet, GoogleNet, and ResNet, respectively.
Note that this also approximates the time taken by CONV and FC
layers from Table 2, 3, 5, and 6 (81.47%, 97.98%, 71.1%, and 88.57%).
Second, the trend is similar on TX1 CPU, as depicted in Figure 1b,
except GoogleNet (only about 25% time spent on matmul operations), which is caused by the particular combination of the architecture of TX1 CPU and GoogleNet as discussed in §4.1. Third, the
trend on TK1 and TX1 GPUs is similar to the trend on TX1 CPU,
as seen in Figure 1c and 1d. One thing to note is that while matmul
operations of GoogleNet only take about 20% of the total time of
forward pass, our previous measurement in Table 5, showed that
CONV and FC layers take about 60% of the total forward pass time.
2242
n = 96
100
40
104
64
128
384
20
102
k = 64 × 12
64 × 32
128 × 32
64 × 52
256 × 32
64 × 72
memory usage (MB)
1122
n = 32
96
256
512
32 64
128
256
384
64
512
576
n
1152 1600
2304
10−2
3136
142
282
k
562
1122
Figure 4: Matrix multiplication on TK1 GPU with varying n, m, and k.
282
562
1122
2242
n = 96
100
40
104
64
128
384
compute time (ms)
200
n = 32
96
256
512
142
compute time (ms)
compute time (ms)
m = 72
20
102
k = 64 × 12
64 × 32
128 × 32
64 × 52
256 × 32
64 × 72
32 64
128
256
n
384
512
64
576
1152 1600
2304
3136
400
60
100
10−2
k
72
142
282
562
1122
2242
m
Figure 5: Matrix multiplication on TX1 GPU with varying n, m, and k.
We believe this is because the matmul operations are run without
taking into account dependencies, whereas, GoogleNet consists of
inception components, each of which has four branches of CONV
layers in parallel. Before proceeding to next inception component,
all four branches of CONV layers have to be competed. How to
handle such dependencies is part of our future work.
In summary, for most cases, matmul operations take a large proportion (more than 60%) of the compute time of a CNN on mobile
platforms. Thus, we can predict matmul time, to be able to approximately estimate the compute time of a CNN.
Modeling
So far, we have exactly measured matmul time. In this section, we
aim to model this time, to be able to predict the compute time, just
from the matrix sizes. To do so, we benchmark several matrix sizes,
as explained below to understand the relationship between the size
of the matrices and the compute time.
Given the matmul of [n ×k] and [k ×m] (the number of FLOPs is
n ×m ×k) performed by a CONV layer, n is the number of kernels, k
is the size of a kernel in 3D (width × height × depth, where depth is
the number of input feature maps), and m is the spatial size (width
× height) of output feature maps.
CNNs follow special rules on these parameters of CONV layers. The number of kernels n is usually a multiple of 16, commonly
from 32 to 512. The spatial size of a kernel is commonly 12 , 32 , 52 ,
72 , or 112 . The depth of a kernel is usually the number of kernels in
the previous CONV layer and hence also a multiple of 16; except
the first CONV layer, where the depth is the number of channels of
the input image, typically equal to three. The spatial size of output
feature maps of a CNN m gradually reduces; it is common to have
2242 , 1122 , 562 , 282 , 142 , or 72 , though AlexNet has slightly different ones, i.e., 552 , 272 and 132 . Based on these typical parameter
settings, we carried out experiments on matmul with varying n, m,
and k. The FC layer is currently used in CNNs only as a classifier
156
116
89
150
675
625
705
486
181
0
NIN on GPU
VGG19M on CPU VGG19M on GPU
Memory estimate of NIN and
800
94.64%
Estimate
513.6 542.7
400
87.67%
106.7 121.7
80.44%
TX1 CPU
Forward Pass
78.45%
48.1 59.8
18.2 23.2
TK1 GPU
TX1 GPU
0
TK1 CPU
(a) effect of n and m , where k = 576 (b) effect of k and n , where m = 562 (c) effect of m and k , where n = 256
5.2
581
Figure 7: Timing estimate of NIN.
0
0
TX1
2242
m
60
TK1
398
Figure 6:
VGG19M.
72
(a) effect of n and m , where k = 576 (b) effect of k and n , where m = 562 (c) effect of m and k , where n = 256
300
Estimate
NIN on CPU
0
0
800
100
compute time (ms)
562
compute time (ms)
142
282
compute time (ms)
compute time (ms)
200
m = 72
compute time (ms)
60
300
105
104
94.18%
8818 9363
Estimate
Forward Pass
85.60%
2193 2562
84.27%
103
276.5 328.1
102
TK1 CPU
TX1 CPU
TK1 GPU
79.47%
141.7 178.3
TX1 GPU
Figure 8: Timing profiling of VGG19M.
(e.g., in GoogleNet and ResNet) and thus its compute time is negligible compared to the forward pass. Therefore, we do not consider
the size of matrices for FC layers in the modeling.
Simple linearity on CPU: Figure 2 and 3 illustrate the performance of matmul on TK1 CPU and TX1 CPU, respectively. The
settings of n, m, and k are: n = [32, 64, 96, 128, 256, 512], m =
[72 , 142 , 282 , 562 , 1122 ], and k = [64 × 12 , 64 × 32 , 128 × 32 , 64 ×
52 , 256 × 32 , 64 × 72 ]. In each figure, we fix one of three parameters and vary other two; data points are shown as small circles;
black circles are labeled with coordinates to highlight the setting
of varying parameters.
From Figure 2a, 2b, and 2c, it is observed that the compute time
of matmul on TK1 CPU scales linearly with n, m, and k. The linearity can also be observed on TX1 CPU as depicted in Figure 3a,
3b, and 3c. Thus, we have a linear model per CPU device, which
predicts the matmul time, given the matrix sizes.
Complex linearity on GPU: Figure 4 and 5 illustrate the performance of matmul with varying settings of n, m, and k on TK1
GPU and TX1 GPU, respectively. The compute time of matmul on
GPUs exhibits more complex relationship with n, m, and k.
Figure 4a mainly depicts the effect of n, which is bipartite. For
all the settings of m, the compute time has a monotonic relationship with n from n = 32 to 128, except n = 96 which incurs even
longer compute time than n = 128, while, from n = 128 to 512, the
compute time exhibits a perfect linear relationship with n. Similar
result is also found on TX1 GPU as shown in Figure 5a. Although
TX1 GPU has more CUDA cores (256 compare to 192 cores in TK1
GPU) and generally computes matmuls faster than TK1 GPU, it
also exhibits this pattern at n = 96. This artifact is related to the
algorithm that determines how the CUDA cores compute matmul
in parallel. Since cuBLAS is not an open-source library, it is hard
to trace the exact reason. However, it is indicated [2] that matmul
works best if n and m are multiples of 128 on Maxwell architecture
(TX1 GPU) and if n is multiple of 256 and m multiple of 192 on
Kepler architecture (TK1 GPU). This may explain why it behaves
differently when n is small.
For given values of n and m, the compute time linearly increases
with k on TK1 GPU and TX1 GPU as depicted in Figure 4b and 5b,
respectively. While the compute time increases with m on both
TK1 GPU and TX1 GPU as depicted in Figure 4c and 5c, the effect
of m is tripartite. The compute time has three separate linear relationships with k (different coefficients), e.g., from 72 to 282 , from
282 to 562 , and from 562 to 2242 on TK1 GPU, as highlighted by different regions in Figure 4c. In each such region, the compute time
on different values of k linearly scales with m at mostly the same
coefficient. Moreover, in the middle region (i.e., between 282 and
562 in Figure 4c and between 142 and 282 in Figure 5c, the compute
time increases with m slower than other two regions. This is especially true on TX1 GPU, where the region is much more flat and
tends to plateau. This region should be the transition area, where
cuBLAS adopts different schemes based on m and the number of
CUDA cores to assign the workload of matmul to CUDA cores. The
transition area is different on TK1 GPU and TX1 GPU, mainly because they have different number of CUDA Cores.
Based on the characteristics discussed above, we are able to
model the compute time of matmul on a specific GPU, though we
need more data points than that on a CPU.
5.3
Accuracy
Based on the measurement, profiling, and modeling of CNNs on
mobile devices, we built the modeling tool, Augur, which estimates
the compute time and memory usage for any given CNN. Augur first
parses the descriptor of a CNN. Based the type and setting of each
layer, it calculates the minimal memory needed to run the CNN.
The memory includes data, parameters, and workspace. Then, Augur extracts matmuls from the computation of the CNN. Based on
the models of TK1 and TX1 on matmul, i.e., the linear fits obtained
from Figure 2 and 4 for TK1, and Figure 3 and 5 for TX1, Augur calculates the compute time of individual matmuls and then uses their
summation as the estimate of the compute time of the CNN.
To verify the accuracy of Augur, we model two CNNs (i.e., NIN
[20] and VGG19M3 ) and compare the estimates to the measured
memory usage and compute time using Caffe. Figure 6 depicts
the memory usage of NIN and VGG19M on different processing
units. The estimate of memory usage is always less than the actual
usage, because the estimate does not take into account the memory
usage of Caffe itself, which is framework-dependent. However, it
is easy to incorporate that if a specific framework is targeted to
perform the CNN computation. Note that the estimate of Augur is
accurate on the memory usage of data, parameters, and workspace
as discussed in §4.2.
Figure 7 and 8 evaluate the accuracy of Augur’s compute time
estimation of NIN and VGG19M, respectively. From Figure 7 and
8, we observe that the estimate based on only matmul can approximate the compute time of NIN and VGG19M on both CPUs and
GPUs, with more than 78% accuracy for all the cases. Since matmul generally takes a larger proportion of the compute time on
CPUs than on GPUs as discussed in §4.1, the estimate on CPUs (up
3 VGG19M is a modified version of VGGNet with more CONV layers. The FC layers
in the original VGGNet are replaced by a CONV layer and a POOL layer to reduce
memory usage.
to 94%) is closer to the actual compute time than on GPUs (up to
84%). Moreover, more powerful processing unit can perform matmul faster, but the speed up is not the same across all operations.
Therefore, the matmul of a CNN takes a smaller proportion of the
compute time on a more powerful processing unit. This explains
why the estimate on TK1 CPU (or TK1 GPU) is more accurate than
TX1 CPU (TK1 GPU) for the same CNN.
In summary, Augur can estimate whether and how efficiently a
CNN can be run on mobile devices before any deployment. It can
also help the design of CNNs for resource-constrained mobile devices. When designing a CNN model using Augur, designers can
estimate the resource usage and compute time without implementation and deployment and tune the model to satisfy their specific
needs.
6 DISCUSSION
Augur can be extended to support additional mobile platforms by
simply profiling matrix multiplication operations on them. Matrix
multiplications of a CNN take most computation (more than 90%
of FLOPs from Table 2, 3, 5, and 6), which commonly takes a dominant proportion of the compute time. Thus, matrix multiplication
is currently exploited by Augur to estimate the compute time of
a CNN. To obtain a more precise estimate, additional factors need
to be taken into consideration, e.g., memory operations and CNN
architectures (stacked or branched). Augur will be enhanced with
these features and this will be our future work.
Moreover, we observe that a framework customized for running CNNs on mobile platforms is highly desired. The framework
should be optimized for performing the test phase of CNNs and
tailored for the characteristics of mobile platforms, e.g., the unified
memory architecture.
7 CONCLUSION
In this paper, we aim to model the resource requirements of CNNs
on mobile devices. By deploying several popular CNNs on mobile
CPUs and GPUs, we measured and analyzed the performance and
resource usage at a layerwise granularity. Our findings pointed
out the potential ways of optimizing the performance of CNNs on
mobile devices. As matrix multiplications form the core computations of a CNN, we profiled and modeled matrix multiplications
on mobile platforms. Based on the measurement, profiling, and
modeling, we built Augur that can estimate the compute time and
memory usage of the CNN so as to give insights on whether and
how efficiently the CNN can be run on a mobile platform without
implementation and deployment. Therefore, it is a power tool that
helps the design of CNNs for resource-constrained mobile devices.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
Caffe. http://caffe.berkeleyvision.org/.
cuBLAS. https://developer.nvidia.com/cublas.
CUDA C Programming Guide. https://docs.nvidia.com/cuda/.
cuDNN. https://developer.nvidia.com/cudnn/.
OpenBLAS. http://www.openblas.net/.
TensorFlow. http://www.tensorflow.org/.
Theano. http://deeplearning.net/software/theano/.
Torch. http://torch.ch/.
Alfredo Canziani, Adam Paszke, and Eugenio Culurciello. 2016. An Analysis of Deep Neural Network Models for Practical Applications. arXiv preprint
arXiv:1605.07678 (2016).
[10] Yunchao Gong, Liu Liu, Ming Yang, and Lubomir Bourdev. 2014. Compressing deep convolutional networks using vector quantization. arXiv preprint
arXiv:1412.6115 (2014).
[11] Seungyeop Han, Haichen Shen, Matthai Philipose, Sharad Agarwal, Alec Wolman, and Arvind Krishnamurthy. 2016. MCDNN: An Approximation-Based Execution Framework for Deep Stream Processing Under Resource Constraints.
In International Conference on Mobile Systems, Applications, and Services (MobiSys’16).
[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep Residual
Learning for Image Recognition. In IEEE Conference on Computer Vision and
Pattern Recognition (CVPR’16).
[13] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J
Dally, and Kurt Keutzer. 2016. SqueezeNet: AlexNet-level accuracy with 50x
fewer parameters and < 0.5MB model size. arXiv preprint arXiv:1602.07360
(2016).
[14] Sergey Ioffe and Christian Szegedy. 2015. Batch Normalization: Accelerating
Deep Network Training by Reducing Internal Covariate Shift. In International
Conference on Machine Learning (ICML’15).
[15] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long,
Ross Girshick, Sergio Guadarrama, and Trevor Darrell. 2014. Caffe: Convolutional Architecture for Fast Feature Embedding. In ACM International Conference on Multimedia (MM’14).
[16] Yong-Deok Kim, Eunhyeok Park, Sungjoo Yoo, Taelim Choi, Lu Yang, and
Dongjun Shin. 2016. Compression of Deep Convolutional Neural Networks for
Fast and Low Power Mobile Applications. In International Conference on Learning Representations (ICLR’16).
[17] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. Imagenet Classification with Deep Convolutional Neural Networks. In Neural Information Processing Systems Conference (NIPS’12).
[18] Nicholas D Lane, Sourav Bhattacharya, Petko Georgiev, Claudio Forlivesi, Lei
Jiao, Lorena Qendro, and Fahim Kawsar. 2016. Deepx: A Software Accelerator
for Low-Power Deep Learning Inference on Mobile Devices. In International
Conference on Information Processing in Sensor Networks (IPSN’16).
[19] Nicholas D Lane, Sourav Bhattacharya, Petko Georgiev, Claudio Forlivesi, and
Fahim Kawsar. 2015. An Early Resource Characterization of Deep Learning on
Wearables, Smartphones and Internet-of-Things Devices. In International Workshop on Internet of Things towards Applications (IoT-App’15).
[20] Min Lin, Qiang Chen, and Shuicheng Yan. 2014. Network in Network. In International Conference on Learning Representations (ICLR’14).
[21] Karen Simonyan and Andrew Zisserman. 2015. Very Deep Convolutional Networks for Large-Scale Image recognition. In International Conference on Learning Representations (ICLR’15).
[22] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir
Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. 2015.
Going Deeper with Convolutions. In IEEE Conference on Computer Vision and
Pattern Recognition (CVPR’15).
[23] Jiaxiang Wu, Cong Leng, Yuhang Wang, Qinghao Hu, and Jian Cheng. 2016.
Quantized Convolutional Neural Networks for Mobile Devices. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR’16).
| 1cs.CV
|
ON THE RECOGNITION PROBLEM FOR VIRTUALLY
SPECIAL CUBE COMPLEXES
arXiv:1408.2325v2 [math.GR] 19 Jun 2017
MARTIN R. BRIDSON AND HENRY WILTON
Abstract. We address the question of whether the property of
being virtually special (in the sense of Haglund and Wise) is algorithmically decidable for finite, non-positively curved cube complexes. Our main theorem shows that it cannot be decided by examining one hyperplane at a time. Specifically, we prove that there
does not exist an algorithm that, given a compact non-positively
curved squared 2-complex X and a hyperplane H in X, will decide
whether or not there is a finite-sheeted cover of X in which no lift
of H self-osculates.
1. Introduction
Haglund and Wise’s concept of a special cube complex [6] has had
profound ramifications in geometric group theory and low-dimensional
topology, most notably in Agol’s solution to the Virtually Haken Conjecture [2]. A compact, non-positively curved cube complex is special if
it admits a local isometry to the canonical cube complex associated to
a right-angled Artin group. When phrased thus, this is not a condition
that one expects to be able to check in a practical manner, given a compact cube complex X. But Haglund and Wise proved that being special
is equivalent to a purely combinatorial condition which requires that
each of the finitely many hyperplanes in X should not behave in one
of a small number of forbidden ways: a hyperplane should not cross
itself, it should be 2-sided, it should not self-osculate, and it should
not inter-osculate with another hyperplane (see [6] for details). We say
that a hyperplane is clean if it does not exhibit any of these forbidden
behaviours.
Given the combinatorial nature of these conditions, it is easy to check
if a given hyperplane is clean and hence whether a given compact cube
complex is special or not. However, in applications, it is not specialness
that matters so much as virtual specialness, i.e. whether the complex
has a finite-sheeted covering space that is special. Agol [1, p. 33] has
promoted the study of the following question.
2010 Mathematics Subject Classification. 20F10, 20F67, 57M07.
Both authors are supported by the EPSRC. Bridson is also supported by a
Wolfson Research Merit Award from the Royal Society.
1
2
MARTIN R. BRIDSON AND HENRY WILTON
Question 1. Is there an algorithm that, given a finite, non-positively
curved cube complex, will determine whether or not that complex is
virtually special?
This is equivalent to the question:
Question 11 . Is there an algorithm that, given a finite, non-positively
curved cube complex X (whose hyperplanes are H1 , . . . , Hn , say), will
p Ñ X in which
determine the existence of a finite-sheeted covering X
each lift of each Hi is clean.
One has to think clearly for a moment to see that this question is
not equivalent to:
Question 2. Is there an algorithm that, given a finite, non-positively
curved cube complex X and a collection of hyperplanes H1 , . . . , Hn , will
p Ñ X in which
determine the existence of a finite-sheeted covering X
each lift of each Hi is clean.
One’s first thought is that a positive solution to Question 11 should
proceed via a positive solution to Question 2. But on reflection one
realises that the latter would be significantly more powerful because
certain of the complexes X for which one did establish the existence
p would not be virtually special. Focusing on this difference, we
of X
shall answer Question 2 in the negative. The following theorem is an
immediate consequence of our main technical result, Theorem 2.4.
Theorem 1.1. There does not exist an algorithm that, given a finite,
non-positively curved cube complex X and a hyperplane H in X, can
p ÑX
determine whether or not there exists a finite-sheeted covering X
in which each lift of H is clean.
Note that this theorem is equivalent to the corresponding statement
p Ñ X is normal, and
with ‘each’ replaced by ‘some’: this is clear if X
since the lift of a clean hyperplane is clean, we are free to pass to a
finite-sheeted normal covering.
The main ingredient in our proof of Theorem 2.4 (and hence Theorem 1.1) is our work in [3], in which we showed that there is no algorithm that determines whether or not a finite complex has a proper
finite-sheeted cover (see Theorem 3.2). In [3] we also proved that this
undecidability persists in the setting of finite, non-positively curved
cube complexes (see Theorem 3.1). The point of this paper is to note
that, with care, the same procedure can be made to translate the nonexistence of a proper finite-sheeted cover into the failure of a certain
hyperplane to be virtually clean.
The obstruction to cleanness that we focus on to prove Theorem 2.4
is self-osculation. Our construction shows that the question of whether
one can remove self-osculations in finite-sheeted covers is already undecidable in dimension 2, and this is the context in which we shall work
ON RECOGNIZING VIRTUALLY SPECIAL CUBE COMPLEXES
3
for the remainder of this paper. Although this theorem does not settle
Question 1, it strongly constrains the nature of any possible positive
solution: a putative recognition algorithm would have to take account
of all the hyperplanes of X simultaneously. We therefore believe that
Question 1 has a negative solution:
Conjecture 1.2. Virtual specialness is an undecidable property of nonpositively curved cube complexes.
2. VH-complexes
Rather than working with non-positively curved cube complexes in
full generality, we restrict attention to the class of VH-complexes.
Definition 2.1. A square complex is VH if there is a partition of the
edges of the 1-skeleton into vertical and horizontal subgraphs, such
that the boundary cycle of each 2-cell crosses vertical and horizontal
edges alternately.
We refer the reader to [4, 7] for the basic theory of VH-complexes, but
we recall some salient facts here. In common with all cube complexes,
much of the geometry of a VH-complex X is encoded in its set of
hyperplanes Y Ď X. (A hyperplane is a maximal connected union of
midcubes.) We are concerned with certain regularity hypotheses on
the hyperplanes.
Definition 2.2. Let Y be a hyperplane embedded in a VH-complex
X. The union of all the open cells intersecting Y , together with the
closest-point projection map, defines a bundle N̊ Ñ Y with fibre p0, 1q,
which has a natural completion to an I-bundle denoted by N Ñ Y
equipped with a map ι : N Ñ X extending the inclusion of N̊ into X.
If N is a trivial bundle then Y is called non-singular (or 2-sided ). In
this case, N – Y ˆ r0, 1s, and the two maps
Bt : y ÞÑ ιpy, tq
for t “ 0, 1, are called the pushing maps associated to Y .
The hyperplane Y is called clean if the pushing maps B0 , B1 are both
embeddings.
If every hyperplane is non-singular then the complex X is nonsingular, and if every hyperplane is clean then the whole complex X
is called clean. The following theorem of Haglund and Wise reconciles
this terminology with that of virtually special complexes.
Theorem 2.3. [6, Theorem 5.7] A VH-complex is virtually special if
and only if it is virtually clean.
Thus, we may specialize Question 1 to the case of VH-complexes as
follows.
4
MARTIN R. BRIDSON AND HENRY WILTON
Question 3. Is there an algorithm that decides whether or not a given
finite VH-complex is virtually clean?
As in the introduction, our focus here is a local version of this question. We call a hyperplane Y in a VH-complex X virtually clean if
p Ñ X and a connected
there exists a finite-sheeted covering map p : X
p is clean. Our
component Yp Ď p´1 Y such that the hyperplane Yp of X
main theorem can now be stated as follows.
Theorem 2.4. There is a recursive sequence of pairs of compact VHcomplexes Xn and hyperplanes Yn Ď Xn such that the set of natural
numbers n for which Yn is virtually clean is recursively enumerable but
not recursive.
The complexes Xn that we shall construct to prove Theorem 2.4 are
never (globally) virtually clean—indeed, π1 Xn is not residually finite—
so they are not themselves candidates to answer Question 3.
3. Undecidability and finite covers
The proof of Theorem 2.4 relies on our recent work on the triviality
problem for profinite completions, in which we proved the following
theorem [3, Theorem B].
Theorem 3.1. There is a recursive sequence of compact, non-positively
curved square complexes Xn such that the set
tn P N | πz
1 Xn fl 1u
is recursively enumerable but not recursive.
p denotes the profinite completion of a group Γ. By definition,
Here, Γ
p is trivial if and only if every finite quotient of Γ is trivial.
Γ
The proof of this theorem begins with a non-positively curved square
complex J that has no proper finite-sheeted covers. With a careful
choice of J, we can ensure that, in the output sequence pXn q, each Xn is
a VH-complex. The main result of this section is Theorem 3.7, a slight
refinement of Theorem 3.1. We will start by explaining the construction
that proves Theorem 3.1, and then observe some additional properties
of the examples that we use to deduce Theorem 3.7.
The first ingredient in the proof of Theorem 3.1 is the corresponding
result for finitely presented groups (without any non-positive curvature
condition).
Theorem 3.2. There is a recursive sequence of finite presentations Pn
of groups Γn such that the set
pn fl 1u
tn P N | Γ
is recursively enumerable but not recursive.
For J we will use a complex that was constructed by Wise.
ON RECOGNIZING VIRTUALLY SPECIAL CUBE COMPLEXES
5
Theorem 3.3 (Wise [7]). There exists a compact VH-complex J such
that π1 J is infinite but has no proper finite quotients.
Burger and Mozes exhibited compact VH-complexes with the even
more striking property that their fundamental groups are infinite and
simple [5].
We next perform a construction reminiscent of the well known Kan–
Thurston construction.
Definition 3.4. Let J be as above and fix a simple closed loop c in a
vertical component of the 1-skeleton of J. Let K be the presentation 2complex associated to a finite presentation P ” xa1 , . . . , an | r1 , . . . , rm y
of a group Γ, and let V :“ K p1q . We define JP to be the space obtained
by attaching m copies of J to V , with the j-th copy attached by a
cylinder joining c to the edge-loop in V labelled rj . More formally,
writing ρj : S1 Ñ V for this last loop, we define „ to be the equivalence
relation on
ž
ž`
˘
S1 ˆ r0, 1s ˆ t1, . . . , mu
J ˆ t1, . . . , mu
V
defined by
@t P S1 @j P t1, . . . , mu : ρj ptq „ pt, 0, jq and pt, 1, jq „ pcptq, jq,
and define JP to be the quotient space.
J
c
φ
J
c
V
K
Figure 1. The complex JP and the map φ.
The construction of JP is illustrated in Figure 1. The next proposition collects together some easy properties of this construction.
Proposition 3.5. Let JP be as above. Then we have the following
properties:
(1) JP has the structure of a VH-complex;
(2) there is a natural continuous map φ : JP Ñ K, with the property
that every homomorphism from π1 JP to a finite group factors
through φ˚ ;
(3) the composition
φ˚
π1 V Ñ π1 JP Ñ Γ
6
MARTIN R. BRIDSON AND HENRY WILTON
is surjective.
Proof. To prove (1), we subdivide the copies of J and V so that the
attaching maps of the cylinders have the same length at either end.
The gluing cylinders can then be subdivided into squares, giving the
whole space the structure of a square complex, which is obviously VH.
The map φ is defined to crush each copy of J to a point; since
a cylinder with one end identified with a point is homeomorphic to
a disc, the resulting space is homeomorphic to Y . The kernel of φ˚
is normally generated by the copies of π1 J. Since π1 J has no finite
quotients, if f is a homomorphism from π1 JP to a finite group, then f
kills the conjugacy classes defined by the loops in each copy of J, and
therefore factors through φ˚ . This proves (2).
Since V is the entire 1-skeleton of K, item (3) is immediate.
Definition 3.6. A pointed VH-pair is a triple pL, V, ˚q where L is a
compact VH-complex, V is a vertical component of the 1-skeleton and
˚ is a vertex of V .
Theorem 3.7. There is a recursive sequence of pointed VH-pairs pLn , Vn , ˚n q
such that the following set is recursively enumerable but not recursive:
tpn, γq | n P N, γ̂ ‰ 1 in πy
1 Ln u
where γ runs through simple loops in Vn based at ˚n and γ̂ denotes the
element of πy
1 Ln defined by γ.
Proof. Let Pn and Γn be as in Theorem 3.2 and let Kn be the presentation complex for Pn . Set Ln “ JPn , let Vn be the vertical subgraph
of JPn from Proposition 3.5 and let ˚n be the unique vertex of Vn .
Note that the construction of pLn , Vn , ˚n q from Pn is algorithmic, and
therefore the sequence is recursive.
Consider now a finite set of simple based loops γ1 , . . . , γk that generate π1 pVn , ˚n q. If there were an algorithm to determine whether or
not such a loop γ survives in some quotient of π1 pLn , ˚n q then, applying the algorithm to each γi in turn, we would obtain an algorithm
to determine whether or not π1 pVn , ˚n q has non-trivial image in some
finite quotient of π1 pLn , ˚n q. But the latter occurs if and only if Γn has
a non-trivial finite quotient (by items (2) and (3) of Proposition 3.5),
which is undecidable.
This proves that the set of pairs pn, γq for which γ survives in some
finite quotient of π1 pLn , ˚n q is not recursive. However, if γ does survive
in some finite quotient then a systematic search will eventually discover
this. Therefore, the set of such pn, γq is recursively enumerable.
It will be convenient in what follows to have a fixed enumeration of
the pairs pn, γq with γ a simple loop in Vn based at ˚n : denote this
by m ÞÑ pnm , γm q. Correspondingly, we write pLm , Vm , ˚m q in place of
pLnm , Vnm , ˚nm q, so γm is a simple loop in Vm based at ˚m .
ON RECOGNIZING VIRTUALLY SPECIAL CUBE COMPLEXES
7
4. Proof of Theorem 2.4
In this final section, we modify the output of Theorem 3.7 in order
to deduce Theorem 2.4. First we construct L1n from Ln by attaching
a vertical loop α (consisting of a single edge) at ˚n . We also let Vn1 “
Vn Y α. For each simple based loop γ in Vn we define γ 1 to be the
concatenation γ 1 ¨ α in L1n .
Each loop γn1 is a local isometry
γn1 : S1 Ñ L1n
which induces a cellular structure on S1 , sending vertices to vertices and
edges to edges. To construct the complex Xn , take two copies of L1n ,
and attach an annulus An – S1 ˆ I, where the two ends of the cylinder
An are attached by the two copies of the map γn1 . The construction is
illustrated in Figure 2.
Yn
γn
γn1
Ln
˚n
L1n
An
α
L1n
Figure 2. The construction of the complex Xn .
Since the loops γ 1 are vertical and locally geodesic, we can extend the
VH-structures of the two copies of L1n to a non-singular VH-structure
on Xn by dividing the annulus An into squares.
The hyperplane Yn is defined to be the unique vertical hyperplane
in the annulus An . It is homeomorphic to a circle, and two-sided (i.e.
non-singular) by definition.
Our final lemma relates the geometry of Yn in finite-sheeted covers of
Xn to the group-theoretical properties of φ˚ pγn q P Γn (in the notation
of Proposition 3.5 with Ln “ JPn ).
Lemma 4.1. The following statements are equivalent.
(1) The hyperplane Yn Ď Xn is virtually clean.
(2) There is a based, finite-sheeted covering space pR, :q Ñ pLn , ˚n q
so that the lift at : of γn is not a loop.
(3) There is a subgroup N of finite index in π1 Ln such that γn R N .
(4) The group element rγn s survives in some finite quotient of π1 Ln .
(5) The element φ˚ prγn sq survives in some finite quotient of Γn .
Proof. We first show that (1) implies (2). Suppose that p : Z Ñ Xn is
a finite-sheeted covering map and that some component W of p´1 Yn is
8
MARTIN R. BRIDSON AND HENRY WILTON
clean. Consider the pushing map B0 : W Ñ Z. Choose some : P p´1 p˚n q
that lies in B0 pW q. Let R be the component of p´1 Ln that contains :.
Because B0 is an embedding, it follows that the lift of γn at : is not a
loop.
We next show that (2) implies (1); to this end, we define a certain
retraction ρ : Xn Ñ Ln . Consider the natural retraction Xn Ñ L1n
which identifies the two copies of L1n and maps the gluing annulus via
γn1 . Next, consider the retraction L1n Ñ Ln which maps αptq to γ n ptq,
where γ n is the path γn in the opposite direction. The concatenation
Xn Ñ L1n Ñ Ln
defines the retraction ρ. To show that (2) implies (1), we pull the
covering map R Ñ Ln back along ρ to obtain a covering space Z Ñ Xn .
Let W be the unique component of the preimage of Yn such that : is
contained in Bt pW q for some t “ 0 or 1. By construction, W is a circle
and Bt is given by the lift of γn1 “ γn ¨ α. Since the lift of γn is an
embedded arc with distinct endpoints, the lift of γn1 is an embedded
loop, as required.
Items (2) and (3) are equivalent by standard covering-space theory:
in one direction, take N “ π1 pR, :q, and in the other take pR, :q to be
the based covering space corresponding to the subgroup N .
The equivalence of items (3) and (4) is also standard. To show that
(4) implies (3), just take N to be the kernel of the finite quotient map.
For the converse, the action of π1 Ln by left translation on πLn {N
defines a homomorphism π1 Ln Ñ SympπLn {N q in which γn survives.
Finally, that (4) and (5) are equivalent follows from item (2) of
Proposition 3.5.
Lemma 4.1 tells us that deciding if the hyperplane Yn Ď Xn is virtually clean is equivalent to deciding if pn, γn q lies in the set considered in
Theorem 3.7; and this set is recursively enumerable but not recursive.
This completes the proof of Theorem 2.4.
References
[1] Ian Agol. Virtual properties of 3-manifolds. https://dl.dropboxusercontent.
com/u/8592391/virtualspecialICM.pdf.
[2] Ian Agol. The virtual Haken conjecture. Doc. Math., 18:1045–1087, 2013. With
an appendix by Agol, Daniel Groves, and Jason Manning.
[3] Martin R. Bridson and Henry Wilton. The triviality problem for profinite completions. Invent. Math., 202(2):839–874, 2015.
[4] Martin R. Bridson and Daniel T. Wise. V H complexes, towers and subgroups
of F ˆ F . Mathematical Proceedings of the Cambridge Philosophical Society,
126(3):481–497, 1999.
[5] Marc Burger and Shahar Mozes. Finitely presented simple groups and products
of trees. C. R. Acad. Sci. Paris Sér. I Math., 324(7):747–752, 1997.
[6] Frédéric Haglund and Daniel T. Wise. Special cube complexes. Geometric and
Functional Analysis, 17(5):1551–1620, 2008.
ON RECOGNIZING VIRTUALLY SPECIAL CUBE COMPLEXES
9
[7] Daniel T. Wise. Complete square complexes. Comment. Math. Helv., 82(4):683–
724, 2007.
Martin R. Bridson, Mathematical Institute, University of Oxford,
Andrew Wiles Building, Oxford, OX2 6GG, UK
E-mail address: [email protected]
Henry Wilton, DPMMS, Centre for Mathematical Sciences, Wilberforce Road, Cambridge, CB3 0WB, UK
E-mail address: [email protected]
| 4math.GR
|
GROUPS QUASI-ISOMETRIC TO RAAG’S
arXiv:1601.00946v2 [math.GR] 16 Mar 2016
JINGYIN HUANG AND BRUCE KLEINER
Abstract. We characterize groups quasi-isometric to a right-angled
Artin group G with finite outer automorphism group. In particular all such groups admit a geometric action on a CAT(0) cube
complex that has an equivariant “fibering” over the Davis building
of G.
1. Introduction
Overview. In this paper we will study right angled Artin groups
(RAAG’s). Like other authors, our motivation for considering these
groups stems from the fact that they are an easily defined yet remarkably rich class of objects, exhibiting interesting features from many
different vantage points: algebraic structure (subgroup structure, automorphism groups) [Dro87, Ser89, Lau95, CCV07], finiteness properties [BB97, BM01], representation varieties [KM98], CAT (0) geometry
[CK00], cube complex geometry [Wis11, HW08], and coarse geometry [Wis96, BM00, BKS08a, BN08, BJN10, Hua14b, Hua14a]. Further
impetus for studying RAAG’s comes from their role in the theory of
special cube complexes, which was a key ingredient in Agol’s spectacular solution of Thurston’s virtual Haken and virtual fibered conjectures
[AGM13, Wis11, HW08, Sag95, KM12].
Our focus here is on quasi-isometric rigidity, which is part of Gromov’s program for quasi-isometric classification of groups and metric
spaces. In this paper we build on [BKS08b, BKS08a, Hua14b, Hua14a],
which analyzed the structure of individual quasi-isometries G → G,
where G is a RAAG with finite outer automorphism group. Our main
results are a structure theorem for groups of quasi-isometries (more
precisely quasi-actions), and a characterization of finitely generated
groups quasi-isometric to such RAAG’s. Both are formulated using a
Date: March 14, 2018.
The second author was supported by a NSF grant DMS-1405899 and a Simons
Fellowship.
1
2
JINGYIN HUANG AND BRUCE KLEINER
geometric description in terms of Caprace-Sageev restriction quotients
[CS11] and the Davis building [Dav98].
Background. Prior results on quasi-isometric classification of RAAG’s
may be loosely divided into two types: internal quasi-isometry classification among (families of) RAAG’s, and quasi-isometry rigidity results characterizing arbitrary finitely generated groups quasi-isometric
to a given RAAG. In the former category, it is known that to classify
RAAG’s up to quasi-isometry, it suffices to consider the case when the
groups are 1-ended and do not admit any nontrivial direct product
decomposition, or equivalently, when their defining graphs are connected, contain more than one vertex, and do not admit a nontrivial
join decomposition ([Hua14a, Theorem 2.9], [PW02, KKL98]). This
covers, for instance, the classification up to quasi-isometry of RAAG’s
that may be formed inductively by taking products or free products,
starting from copies of Z. Beyond this, internal classification is known
for RAAG’s whose defining graph is a tree [BN08] or a higher dimension analog [BJN10], or when the outer automorphism group is finite
[Hua14a, BKS08a]. General quasi-isometric classification results in the
literature are much more limited; if H is a finitely generated group
quasi-isometric to a RAAG G then:
(i) If G is free or free abelian, H is virtually free or free abelian,
respectively [Sta68, Dun85, Bas72, Gro81a].
(ii) If G = Fk × Z` , then H is virtually Fk × Z` [Why10].
(iii) If the defining graph of G is a tree, then H is virtually the
fundamental group of a non-geometric graph manifold that has
nonempty boundary in every Seifert fiber space component, and
moreover H is virtually cocompactly cubulated [BN08, KL97a,
HP15].
(iv) If G is a product of free groups, then H acts geometrically on
a product of trees [Ahl02, KKL98, MSW03].
Unlike (i)-(iii), which give characterizations up to commensurability,
the characterization in (iv) only asserts the existence of an action on
a good geometric model; the stronger commensurability assertion is
false, in view of examples of Wise and Burger-Mozes [Wis96, BM00].
The setup. We now recall some terminology and notation; see Section
3 for more detail.
GROUPS QUASI-ISOMETRIC TO RAAG’S
3
If Γ is a finite simplicial graph with vertex set V (Γ), we denote the
associated right-angled Artin group by G(Γ). This is the fundamental group of the Salvetti complex S(Γ), a nonpositively curved cube
complex that may be constructed by choosing a pointed unit length
circle Q
(Sv1 , ?v ) for every vertex v ∈ V (Γ), forming the pointed product
torus v (Sv1 , ?v ), and passing to the union of the product subtori corresponding to the cliques (complete subgraphs) in Γ. The clique subtori
are the standard tori in S(Γ).
We denote the universal covering by X(Γ) → S(Γ); here X(Γ) is a
CAT(0) cube complex on which G(Γ) acts geometrically by deck transformations. The inverse image of a standard torus in S(Γ) under the
universal covering X(Γ) → S(Γ) breaks up into connected components;
these are the standard flats in X(Γ) which we partial order by inclusion.
Note that we include standard tori and standard flats of dimension 0.
The poset of standard flats in X(Γ) turns out to be crucial to our
story. Using it one may define a locally infinite CAT(0) cube complex
|B|(Γ) whose cubes of dimension k ≥ 0 are indexed by inclusions F1 ⊂
F2 , and F1 , F2 are standard flats where dim F2 = dim F1 + k. Elements
of the 0-skeleton |B|(0) (Γ) correspond to the trivial inclusions F ⊂
F where F is a standard flat, so we will identify |B|(0) (Γ) with the
collection of standard flats, and define the rank of a vertex of |B|(Γ) to
be the dimension of the corresponding standard flat; in particular we
may identify the 0-skeleton X (0) (Γ) with the set of rank 0 vertices of
|B|(0) . Since G(Γ) y X(Γ) preserves the collection of standard flats,
there is an induced action G(Γ) y |B|(Γ) by cubical isomorphisms.
The above description is a slight variation on the original construction
of the same object given by Davis, in which one views |B|(Γ) as the
Davis realization of a certain right-angled building B(Γ) associated with
G(Γ), where the apartments of B(Γ) are modelled on the right-angled
Coxeter group W (Γ) with defining graph Γ; see [Dav98] and Section 3.
By abuse of terminology we will refer to this cube complex as the Davis
building associated with G(Γ); it has been called the modified Deligne
complex in [CD95b] and flat space in [BKS08b].
The following lemma is not difficult to prove.
Lemma 1.1.
• Every isomorphism |B|(0) (Γ) → |B|(0) (Γ) of the poset of standard flats extends to a unique cubical isomorphism |B|(Γ) →
|B|(Γ) (Section 3.4).
4
JINGYIN HUANG AND BRUCE KLEINER
• Every cubical isomorphism of |B| → |B| induces a poset isomorphism |B|(0) → |B|(0) (Lemma 3.15).
• A bijection φ(0) : |B|(0) (Γ) ⊃ X (0) (Γ) → X (0) (Γ) ⊂ |B|(0) (Γ)
induces/extends to a poset isomorphism |B|(0) (Γ) → |B|(0) (Γ)
iff it is flat-preserving in the sense that for every standard flat
(0)
F1 ⊂ X(Γ), the 0-skeleton F1 is mapped bijectively by φ(0)
onto the 0-skeleton of some standard flat F2 (Section 5.1).
Remark 1.2. We caution the reader that a cubical isomorphism |B|(Γ) →
|B|(Γ) need not arise from an isomorphism B(Γ) → B(Γ) of the rightangled building.
Rigidity and flexibility. We now fix a finite graph Γ such that the
outer automorphism group Out(G(Γ)) is finite; by work of [CF12,
Day12], one may view this as the generic case. The reader may find it
helpful to keep in mind the case when Γ is a pentagon.
Since there is no ambiguity in Γ we will often suppress it in the
notation below.
It is known that in this case X = X(Γ) is not quasi-isometrically
rigid: there are quasi-isometries that are not at finite sup distance from
isometries, and there are finitely generated groups H that are quasiisometric to X, but do not admit geometric actions on X (Corollary
6.12). On the other hand, quasi-isometries exhibit a form of partial
rigidity that is captured by the building |B|:
Theorem 1.3 ([Hua14a, BKS08a]). Suppose Out(G(Γ)) is finite and
G(Γ) 6' Z. If φ : X (0) → X (0) is an (L, A)-quasi-isometry, then there
is a unique cubical isomorphism |B| → |B| such that associated flatpreserving bijection φ̄ : X (0) → X (0) is at finite sup distance from φ,
and moreover
d(φ̄, φ) = sup{v ∈ X (0) | d(φ̄(v), φ(v))} < D = D(L, A) .
By the uniqueness assertion, we obtain a cubical action QI(X) y |B|
of the quasi-isometry group of X on |B|.
We point out that the partial rigidity statement of the theorem does
not hold for general RAAG’s: it only holds for the RAAG’s covered by
the theorem in [Hua14a].
GROUPS QUASI-ISOMETRIC TO RAAG’S
5
The main results. We will produce good geometric models quasiisometric to X(Γ) that are simultaneously compatible with group actions, the underlying building |B|, and cubical structure. The key idea
for expressing this is:
Definition 1.4. A cubical map q : Y → Z between CAT(0) cube
complexes (see Definition 3.4) is a restriction quotient if it is surjective,
and the point inverse q −1 (z) is a convex subset of Y for every z ∈ Z.
It turns out that restriction quotients as defined above are essentially equivalent to the class of mappings introduced by Caprace-Sageev
[CS11] with a different definition (see Section 4 for the proof that the
definitions are equivalent). Restriction quotients Y → |B| provide a
means to “resolve” or “blow-up” the locally infinite building |B| to a
locally finite CAT(0) cube complex.
Theorem 1.5. (See Section 3 for definitions.) Let H y X be a quasiaction of an arbitrary group on X = X(Γ), where Out(G(Γ)) is finite
and G(Γ) 6' Z. Then there is an H-equivariant restriction quotient
q
H y Y −→ H y |B| where:
(a) H y |B| is the cubical action arising from the quasi-action
H y X using Theorem 1.5, and H y Y is a cubical action.
(b) The point inverse q −1 (v) of every rank k vertex v ∈ |B|(0) is a
copy of Rk with the usual cubulation.
(c) H y X is quasiconjugate to the cubical action H y Y .
Theorem 1.6. If | Out(G(Γ))| < ∞ and G(Γ) 6' Z, then a finitely generated group H is quasi-isometric to G(Γ) iff there is an H-equivariant
q
restriction quotient H y Y −→ H y |B| where
(a) H y Y is a geometric cubical action.
(b) H y |B| is cubical.
(c) The point inverse q −1 (v) of every rank k vertex v ∈ |B|(0) is a
copy of Rk with the usual cubulation.
Remark 1.7. In fact the restriction quotient Y → |B| in Theorems 1.5
and 1.6 has slightly more structure, see Theorem 6.4.
In particular, we have:
Corollary 1.8. Any group quasi-isometric to G is cocompactly cubulated, i.e. it has a geometric cubical action on a CAT(0) cube complex.
One may compare Theorem 1.6 with rigidity theorems for symmetric
spaces or products of trees, which characterize a quasi-isometry class
6
JINGYIN HUANG AND BRUCE KLEINER
of groups by the existence of a geometric action on a model space of
a specified type [Sul81, Gro81b, Tuk86, Pan89, KL97b, Sta68, Dun85,
KKL98, MSW03, Ahl02]. As in the case of products of trees — and
unlike the case of symmetric spaces — there are finitely generated
groups H as in Theorem 1.6 which do not admit a geometric action
on the original model space X, so one is forced to pass to a different
space Y [BKS08a, Hua14a]. Also, Theorems 1.5 and 1.6 fail for general
RAAG’s, for instance for free abelian
Q groups of rank ≥ 2, and for
products of nonabelian free groups 1≤j≤k Gj , for k ≥ 1.
The quasi-isometry invariance of the existence of a cocompact cubulation as asserted in Corollary 1.8 is false in general. Some groups
quasi-isometric to H2 × R admit a cocompact cubulation, while others
are not virtually CAT (0) [BH99]. Combining [Lee95], [BN08] and [HP],
it follows that there is a pair of quasi-isometric CAT (0) graph manifold groups, one of which is the fundamental group of a compact special
cube complex, while the other is not virtually cocompactly cubulated.
The quasi-isometry invariance of cocompact cubulations fails to hold
even among RAAG’s: for n > 1 there are groups quasi-isometric Rn
that are not cocompactly cubulated [Hag14].
Earlier cocompact cubulation theorems in the spirit of Corollary
1.8 include the cases of groups quasi-isometric to trees, products of
trees, and hyperbolic k-space Hk for k ∈ {2, 3} [Sta68, Dun85, KKL98,
MSW03, Ahl02, GMRS98, KM12, BW12]. It is worth noting that each
case requires different ingredients that are specific to the spaces in
question.
Further results. We briefly discuss some further results here, referring the reader to the body of the paper for details.
One portion of the proof of Theorem 1.5 has to do with the geometry of restriction quotients, and more specifically, restriction quotients
with a right-angled building as target. We view this as a contribution
to cube complex geometry, and to the geometric theory of graph products; beyond the references mentioned already, our treatment has been
influenced by the papers of Januszkiewicz-Swiatkowski and Haglund
[JŚ01, Hag08]. The main results on this are:
(a) We show in Section 4 that restriction quotients may be characterized in several different ways.
(b) We show that having a restriction quotient q : Y → Z is equivalent to knowing certain “fiber data” living on the target complex
Z.
GROUPS QUASI-ISOMETRIC TO RAAG’S
7
(c) When |B| is the Davis realization of a right-angled building B
and Y → |B| is a restriction quotient whose fibers are copies
of Rk with dimension specified as in Theorems 1.5 and 1.6, the
fiber data in (b) may be distilled even more, leading to what
we call “blow-up data”.
As by-products of (a)-(c), we obtain:
• A characterization of the quasi-actions H y X(Γ) that are
quasiconjugate to isometric actions H y X(Γ) (Section 6.2).
• A characterization of the restriction quotients Y → |B| satisfying (b) of Theorem 1.5 for which Y is quasi-isometric to X
(Corollary 6.5 and Theorem 6.6).
• A proof of uniqueness of the right-angled building modelled on
the right-angled Coxeter group W (Γ) with defining graph Γ,
with countably infinite rank 1 residues (Corollary 5.23).
• Applications to more general graph products (Theorem 8.8).
It follows from [KL01] that a finitely generated group H quasiisometric to a symmetric space of noncompact type X admits an epimorphism H → Λ with finite kernel, where Λ is a cocompact lattice in
the isometry group Isom(X). In contrast to this, we have the following
result, which is inspired by [MSW03, Theorem 9, Corollary 10]:
Theorem 1.9. (See Section 6.2) Suppose G is a RAAG with | Out(G)| <
∞. Then there are finitely generated groups H and H 0 quasi-isometric
to G that do not admit discrete, virtually faithful cocompact representations into the same locally compact topological group.
Open questions. As mentioned above, Corollary 1.8 may be considered part of the quasi-isometry classification program for finitely
generated groups. The leads to:
Question 1.10. If Out(G(Γ)) is finite, what is the commensurability
classification of groups quasi-isometric to G(Γ)? Are they all commensurable to G(Γ)? What about cocompact lattices in the automorphism
group of X(Γ)?
For comparison, we recall that any group quasi-isometric to a tree is
commensurable to a free group, but there are groups quasi-isometric to
a product of trees that contain no nontrivial finite index subgroups, and
are therefore not commensurable to a product of free groups [Wis96,
BM00].
We mention that Theorem 1.6 will be used in forthcoming work to
answer Question 1.10 in certain cases.
8
JINGYIN HUANG AND BRUCE KLEINER
Discussion of the proofs. Before sketching the arguments for Theorems 1.5 and 1.6, we first illustrate them in the tautological case when
H = G and the quasi-action is the deck group action G y X. In this
case we cannot take Y = X, as there is no H-equivariant restriction
quotient H y X → H y |B| satisfying (c) of Theorem 1.6. Instead,
we use a different geometric model.
Definition 1.11 (Graph products of spaces [Hag08]). For every vertex
v ∈ V (Γ), choose a pointed geodesic metric space (Zv , ?v ). The Γgraph
product of {(Zv , ?v )}v∈V (Γ) is obtained by forming the product
Q
the subproducts corresponding
v (Zv , ?v ), and passing to the union of Q
to the cliques in Γ. We denote this by Γ (Zv , ?v ). When the Zv ’s are
nonpositively curved, then so is the graph product [Hag08, Corollary
4.6].
There are three graph products that are useful here:
Q
(1) The Salvetti complex S(Γ) is the graph product Γ (Sv1 , ?v ),
where (Sv1 , ?v ) is a pointed unit circle.
(2) For every v ∈ V (Γ), let (Lv , ?v ) be a pointed lollipop, i.e. Lv is
the wedge of the unit circle Sv1 and a unit interval Iv , and the
basepointQ?v ∈ Lv is the vertex of valence 1. Then the graph
product Γ (Lv , ?v ) is the exploded Salvetti complex Se = Se (Γ).
We denote its universal covering by Xe → Se .
(3) If (Zv , ?v ) is a unit interval and ?v ∈ Q
Zv is an endpoint for
every v ∈ V (Γ), then the graph product Γ (Zv , ?v ) is the Davis
chamber, i.e. it is a copy of the Davis realization |c| of a chamber
c in |B|(Γ); for this reason we will denote it by |c|Γ .
By collapsing the interval Iv in each lollipop Lv to a point, we obtain
a cubical map Se → S; this has contractible point inverses, and is
therefore a homotopy equivalence. If we collapse the circles Sv1 ⊂ Lv
to points instead, we get a map Se → |c|Γ to the Davis chamber whose
point inverses are closed, locally convex tori. The point inverses of
the composition Xe → Se → |c|Γ cover the torus point inverses of
Se → |c|Γ , and their connected components form a “foliation” of Xe
by flat convex subspaces. It turns out that by collapsing Xe along
these flat subspaces, we obtain a copy of |B|, and the quotient map
Xe → |B| is a restriction quotient Xe → |B|. Alternately, one may
take the collection K of hyperplanes of Xe dual to edges σ ⊂ Xe whose
projection under Xe → |c|Γ is an edge, and form the restriction quotient
using the Caprace-Sageev construction.
GROUPS QUASI-ISOMETRIC TO RAAG’S
9
Remark 1.12. The exploded Salvetti complex and the restriction quotient Xe → |B| were discussed in [BKS08a] in the 2-dimensional case,
using an ad hoc construction that was initially invented for “ease of
visualization”. However, the authors were unaware of the general description above, and the notion of restriction quotient had not yet appeared.
We now discuss the proofs of Theorem 1.5 and the forward direction
of 1.6.
The forward direction of Theorem 1.6 reduces to Theorem 1.5, by
qi
the standard observation that a quasi-isometry H → G ' X allows us
to quasiconjugate the left translation action H y H to a quasi-action
H y X. Therefore we focus on Theorem 1.5.
Let H y X be as in Theorem 1.5. By a bounded perturbation, we
may assume that this quasi-action preserves the 0-skeleton X (0) ⊂ X.
Applying Theorem 1.3, we may further assume that we have an action
H y X (0) by flat-preserving quasi-isometries. The fact the we have an
action, rather than just a quasi-action, comes from the uniqueness in
Theorem 1.5; this turns out to be a crucial point in the sequel.
Before proceeding further, we remark that if one is only interested in
Corollary 1.8 as opposed to the more refined statement in Theorem 1.6,
then there is an alternate approach using wallspaces. This is treated
in Sections 9 and 10.
Given a standard geodesic ` ⊂ X, the parallel set P` ⊂ X decomposes as a product R` × Q` , where R` is a copy of R; likewise there
(0)
(0)
is a product decomposition of 0-skeleta P` ' Z` × Q` . One argues
(0)
that the action H y X (0) permutes the collection of 0-skeleta {P` }` ,
(0)
(0)
and that for any `, the stabilizer Stab(P` , H) of P` in H acts on
(0)
(0)
P` ' Z` × Q` preserving the product structure. We call the action
(0)
Stab(P` , H) y Z` a factor action. The factor actions are by bijections with quasi-isometry constants bounded uniformly independent of
`.
It turns out that factor actions play a central role in the story. For
instance, when the action H y X (0) is the restriction of an action
H y X by cubical isometries, then the factor actions H[`] y Z[`]
are also actions by isometries. In general the factor actions can be
arbitrary: up to isometric conjugacy, any action A y Z by quasiisometries with uniform constants can arise as a factor action for some
10
JINGYIN HUANG AND BRUCE KLEINER
action as in Theorem 1.5. A key step in the proof is to show that such
actions have a relatively simple structure:
ρ0
Proposition 1.13 (Semiconjugacy). Let U y Z be an action of an
arbitrary group by (L, A)-quasi-isometries. Then there is an isometric
ρ1
action U y Z and surjective equivariant (L0 , A0 )-quasi-isometry
ρ0
ρ1
U y Z −→ U y Z ,
where L0 and A0 depend only on L and A.
The assumption that ρ0 is an action, as opposed to a quasi-action, is
crucial: if a group U has a nontrivial quasihomomorphism α : U → R,
α̂
then the translation quasi-action U y R defined by α̂(u)(x) = x+α(u)
is quasiconjugate to a quasi-action on Z, but not to an isometric action
on Z.
ρ0
It follows immediately from the Proposition 1.13 that U y Z is
quasiconjugate to an isometric action on the tree R. In that respect
Proposition 1.13 is similar to the theorem of Mosher-Sageev-Whyte
about promoting quasi-actions on bushy trees to isometric actions on
trees [MSW03, Theorem 1]. Since R is not bushy [MSW03, Theorem
1] does not apply, and indeed the example above shows that the assumption of bushiness is essential in that theorem.
Continuing with the proof of Theorem 1.5, Proposition 1.13 gives a
(0)
good geometric model for the factor action Stab(P` , H) y Z` : we
simply extend each isometry Z` → Z` to an isometry R` → R` , thereby
(0)
obtaining a cubical action Stab(P` , H) y R` . In vague terms, the
remainder of the proof is concerned with combining these cubical models into models for the fibers of a restriction quotient Z → |B|, in an
H-equivariant way. This portion of the proof is covered by more general results about restriction quotients, see (b)-(c) in the subsection on
further results above.
Organization of the paper. The paper is divided into three relatively independent parts. We first outline the content of each part,
then give suggestions for readers who want to focus on one particular
part.
A summary of notation can found in Section 2. Section 3 contains
some background material on quasi-actions, CAT (0) cube complexes,
RAAG’s and buildings. One can proceed directly to later sections with
Section 2 and Section 3 as references.
GROUPS QUASI-ISOMETRIC TO RAAG’S
11
The main part of the paper is Section 4 to Section 7, where we prove
Theorem 1.6. In Section 4 we discuss restriction quotients, showing
how to construct a restriction quotient Y → Z starting from the target
Z and an admissible assignment of fibres to the cubes of Z. Then we
discuss equivariance properties and the coarse geometry of restriction
quotients.
In Section 5, we introduce blow-ups of buildings based on Section
4. These are restriction quotients Y → |B| where the target is a
right-angled building and the fibres are Euclidean spaces of varying
dimension. We motivate our construction in Section 5.1 and Section
5.2. Blow-ups of buildings are constructed in Section 5.3. Several
properties of them are discussed in Section 5.4 and Section 5.5. We
incorporate a group action into our construction in Section 5.6.
In Section 6.1, we apply the construction in Section 5.6 to RAAG’s
and prove Theorem 1.6 modulo Theorem 1.13, which is postponed until
Section 7. In Section 6.2 we answer several natural questions motivated
by Theorem 1.6, and prove Theorem 1.9.
The second part of the paper is Section 8. We discuss an alternative
construction of blow-ups of buildings. In certain cases, this is more
general than the construction in Section 5. We also discuss several
applications of this construction to graph products.
The third part of the paper is Section 9 and Section 10. Using
wallspaces we give an alternative way to cubulate groups quasi-isometric
to RAAG’s, and prove a weaker version of Theorem 1.6.
The reader can proceed directly to Section 8 with reference to Section
3.4 and Definition 5.35. The reader can also start with Section 9 with
reference to Section 3.3, and come back to Section 7 when needed.
Acknowledgements. The second author would like to thank Mladen
Bestvian and Michah Sageev for many discussions related to atomic
RAAG’s, which planted the seeds for the cubulation result in this paper. In particular, Proposition 6.2 originated in these discussions.
Contents
1. Introduction
1
2. Index of notation
12
3. Preliminaries
13
12
JINGYIN HUANG AND BRUCE KLEINER
Part I: A fibration approach to cubulating RAAG’s
21
4. Restriction quotients
21
5. The Z-blow-up of right-angled Building
31
6. Quasi-actions on RAAG’s
49
7. Actions by quasi-isometries on Z
58
Part II: An alternate blow-up construction for buildings
62
8. Blowing-up buildings by metric spaces
62
Part III: A wallspace approach to cubulating RAAG’s
71
9. Construction of the wallspace
71
10. Properties of the cubulation
76
References
81
2. Index of notation
•
•
•
•
•
•
•
•
•
•
•
•
•
B: A combinatorial building (Section 3.4).
|B|: The Davis realization of a building (Section 3.4).
Chambers in the combinatorial building B are c, c0 , d.
|c|Γ : the Davis chamber (the discussion after Definition 1.11,
Section 3.4).
Sr : the collection of all spherical residues in the building B.
projR : B → R: the nearest point projection from B to a residue
R (Section 3.4).
ΛB : the collection of parallel classes of rank 1 residues in the
combinatorial building B. We also write Λ when the building
B is clear (Section 5.3).
T : a type map which assigns each residue of B a subset of ΛB
(Section 5.3).
β: a branched line (Section 9).
CCC: the category of nonempty CAT(0) cube complexes with
morphisms given by convex cubical embeddings.
PC : the parallel set of a closed convex subset of a CAT(0) space
(Section 3.2).
W (Γ): The right-angled Coxeter group with defining graph Γ
(Section 3.4).
G(Γ) the right-angled Artin group with defining graph Γ.
GROUPS QUASI-ISOMETRIC TO RAAG’S
13
• X(Γ) → S(Γ) the universal covering of the Salvetti complex
(Section 3.3)
• Xe (Γ) → Se (Γ) the universal covering of the exploded Salvetti
complex (after Definition 1.11 and Section 5.1). We also write
Xe → Se when the graph Γ is clear.
• P(Γ): the extension complex (Definition 3.5).
• X → X(K): the restriction quotient arising from a set K of
hyperplanes in a CAT(0) cube complex (Definition 4.1).
• Lk(x, X) or Lk(c, X): the link of a vertex x or a cell c in a
polyhedral complex X.
• Γ1 ◦ Γ2 : the join of two graphs.
• K1 ∗ K2 : the join of two simplicial complexes.
3. Preliminaries
3.1. Quasi-actions. We recall several definitions from coarse geometry.
Definition 3.1. An (L, A)-quasi-action of a group G on a metric space
Z is a map ρ : G × Z → Z so that ρ(γ, ·) : Z → Z is an (L, A) quasiisometry for every γ ∈ G, d(ρ(γ1 , ρ(γ2 , z)), ρ(γ1 γ2 , z)) < A for every
γ1 , γ2 ∈ G, z ∈ Z, and d(ρ(e, z), z) < A for every z ∈ Z.
The action ρ is discrete if for any point z ∈ Z and any R > 0, the set
of all γ ∈ G such that ρ(γ, z) is contained in the ball BR (z) is finite; ρ
is cobounded if Z coincides with a finite tubular neighbourhood of the
“orbit” ρ(G, z). If ρ is a discrete and cobounded quasi-action of G on Z,
then the orbit map γ ∈ G → ρ(γ, z) is a quasi-isometry. Conversely,
given a quasi-isometry between G and Z, it induces a discrete and
cobounded action of G on Z.
Two quasi-actions ρ and ρ0 are equivalent if there exists a constant
D so that d(ρ(γ), ρ0 (γ)) < D for all γ ∈ G.
Definition 3.2. Let ρ and ρ0 be quasi-actions of G on Z and Z 0 respectively, and let φ : Z → Z 0 be a quasi-isometry. Then ρ is quasiconjugate
to ρ0 via φ if there is a D so that d(φ ◦ ρ(γ), ρ0 (r) ◦ φ) < D for all γ ∈ G.
3.2. CAT(0) cube complexes. We refer to [BH99] for background
about CAT(0) spaces (Chapter II.1) and cube complexes (Chapter
II.5), and [Sag95, Sag12] for CAT(0) cube complexes and hyperplanes.
A unit Euclidean n-cube is [0, 1]n with the standard metric. A midcube is the set of fixed points of a reflection with respect to some [0, 1]
14
JINGYIN HUANG AND BRUCE KLEINER
factor of [0, 1]n . A cube complex Y is obtained by taking a collection
of unit Euclidean cubes and gluing them along isometric faces. The
gluing metric on Y is CAT(0) if and only if Y is simply connected and
the link of each vertex in Y is a flag simplicial complex ([Gro87]), in
this case, Y is called a CAT(0) cube complex.
Let X be a CAT(0) space and let C ⊂ X be a closed convex subset.
Then there is a well-defined nearest point projection from X to C,
which we denote by πC : X → C. Two convex subsets C1 and C2
are parallel if d(·, C2 )|C1 and d(·, C1 )|C2 are constant functions. In this
case, the convex hull of C1 and C2 is isometric to C1 × [0, d(C1 , C2 )].
For closed convex subset C ⊂ X, we define PC , the parallel set of
C, to be the union of all convex subsets of X which are parallel to C.
If C has geodesic extension property, then PC is also a closed convex
subset and admits a canonical splitting PC ∼
= C × C ⊥ ([BH99, Chapter
II.2.12]).
Suppose Y is a CAT(0) cube complex. Then two edges e and e0 are
parallel if and only if there exists sequences of edges {ei }ni=1 such that
e1 = e, en = e0 , and ei , ei+1 are the opposite sides of a 2-cube in Y . For
each edge e ⊂ Y . Let Ne be the union of cubes in Y which contain an
edge parallel to e. Then Ne is a convex subcomplex of Y , moreover,
Ne has a natural splitting Ne ∼
= he × [0, 1], where [0, 1] corresponds to
the e direction. The subset he × {1/2} is called the hyperplane dual
to e, and Ne is called the carrier of this hyperplane. Each hyperplane
is a union of mid-cubes, hence has a natural cube complex structure,
which makes it a CAT(0) cube complex. The following are true for
hyperplanes:
(1) Each hyperplane h is a convex subset of Y . Moreover, Y \ h
has exactly two connected components. The closure of each
connected component is called a halfspace. Each halfspace is
also a convex subset.
(2) Pick an edge e ⊂ Y . We identify e with [0, 1] and consider the
CAT(0) projection πe : Y → e ∼
= [0, 1]. Then h = πe−1 (1/2)
is the hyperplane dual to e, and πe−1 ([0, 1/2]), πe−1 ([1/2, 1]) are
two halfspaces associated with h. The closure of πe−1 ((0, 1)) is
the carrier of h.
Let Y be a CAT(0) cube complex and let l ∈ Y be a geodesic line
(with respect to the CAT(0) metric) in the 1-skeleton of Y . Let e ⊂ l
be an edge and pick x ∈ e. We claim that if x is in the interior of
e, then πl−1 (x) = πe−1 (x). It is clear that πl−1 (x) ⊂ πe−1 (x). Suppose
GROUPS QUASI-ISOMETRIC TO RAAG’S
15
y ∈ πe−1 (x). It follows from the splitting Ne ∼
= he × [0, 1] as above that
the geodesic segment xy is orthogonal to l, i.e. ∠x (y, y 0 ) = π/2 for any
y 0 ∈ l \ {x}, thus y ∈ πl−1 (x).
The above claim implies πl−1 (x) is a convex subset for any x ∈ l.
Moreover, the following lemma is true.
Lemma 3.3. Let Y and l be as before. Pick an edge e ⊂ Y . If e is
parallel to some edge e0 ⊂ l, then πl (e) = e0 , otherwise πl (e) is a vertex
of l.
Now we define an alternative metric on the CAT(0) cube complex
Y , which is called the l1 -metric. One can view the 1-skeleton of Y as a
metric graph with edge length = 1, and this metric extends naturally
to a metric on Y . The distant between two vertices in Y with respect
to this metric is equal to the number of hyperplanes separating these
two vertices.
A combinatorial geodesic in Y is an edge path in Y (1) which is a
geodesic with respect to the l1 metric. However, we always refer to the
CAT(0) metric when we talk about a geodesic.
If Y is finite dimensional, the l1 metric and the CAT(0) metric on
Y are quasi-isometric ([CS11, Lemma 2.2]). In this paper, we will use
the CAT(0) metric unless otherwise specified.
Definition 3.4. ([CS11, Section 2.1]) A cellular map between cube
complexes is cubical if its restriction σ → τ between cubes factors as
σ → η → τ , where the first map σ → η is a natural projection onto a
face of σ and the second map η → τ is an isometry.
3.3. Right-angled Artin groups. Pick a finite simplicial graph Γ,
recall that G(Γ) is the right-angled Artin group with defining graph Γ.
Let S be a standard generating set for G(Γ) and we label the vertices
of Γ by elements in S. G(Γ) has a nice Eilenberg-MacLane space S(Γ),
called the Salvetti complex
(see [CD95a, Cha07]). Recall that S(Γ) is
Q
the graph product Γ (Sv1 , ?v ), where (Sv1 , ?v ) is a pointed unit circle
(see Definition 1.11).
The 2-skeleton of S(Γ) is the usual presentation complex of G(Γ), so
π1 (S(Γ)) ∼
= G(Γ). The 0-skeleton of S(Γ) consists of one point whose
link is a flag complex, so S(Γ) is non-positively curved and S(Γ) is an
Eilenberg-MacLane space for G(Γ) by the Cartan-Hadamard theorem
([BH99, Theorem II.4.1]).
The closure of each k-cell in S(Γ) is a k-torus. Tori of this kind are
called standard tori. There is a 1-1 correspondence between the k-cells
16
JINGYIN HUANG AND BRUCE KLEINER
(or standard torus of dimension k) in S(Γ) and k-cliques in Γ. We
define the dimension of G(Γ) to be the dimension of S(Γ).
Denote the universal cover of S(Γ) by X(Γ), which is a CAT(0) cube
complex. Our previous labelling of vertices of Γ induces a labelling of
the standard circles of S(Γ), which lifts to a labelling of edges of X(Γ).
A standard k-flat in X(Γ) is a connected component of the inverse
image of a standard k-torus under the covering map X(Γ) → S(Γ).
When k = 1, we also call it a standard geodesic.
For each simplicial graph Γ, there is a simplicial complex P(Γ) called
the extension complex, which captures the combinatorial pattern of
how standard flats intersect each other in X(Γ). This object was first
introduced in [KK13]. We will define it in a slightly different way (see
[Hua14a, Section 4.1] for more discussion).
Definition 3.5 (Extension complex). The vertices of P(Γ) are in 1-1
correspondence with the parallel classes of standard geodesics in X(Γ).
Two distinct vertices v1 , v2 ∈ P(Γ) are connected by an edge if and
only if there is standard geodesic li in the parallel class associated with
vi (i = 1, 2) such that l1 and l2 span a standard 2-flat. Then P(Γ) is
defined to be the flag complex of its 1-skeleton, namely we build P(Γ)
inductively from its 1-skeleton by filling a k-simplex whenever we see
the (k − 1)-skeleton of a k-simplex.
Since each complete subgraph in the 1-skeleton of P(Γ) gives rise to
a collection of mutually orthogonal standard geodesics lines, there is
a 1-1 correspondence between k-simplexes in P(Γ) and parallel classes
of standard (k + 1)-flats in X(Γ). In particular, there is a 1-1 correspondence between maximal simplexes in P(Γ) and maximal standard
flats in X(Γ). Given standard flat F ⊂ X(Γ), we denote the simplex
in P(Γ) associated with the parallel class containing F by ∆(F ).
Lemma 3.6. Pick non-adjacent vertices v, u ∈ P(Γ) and let lu be a
standard geodesic such that ∆(lu ) = u. Then for any standard geodesic
line l with ∆(l) = v, πlu (l) is a vertex of lu which does not depend on
the choice of l in the parallel class. Thus we define this vertex to be the
projection of v on lu and denote it by πlu (v).
Proof. It follows from Lemma 3.3 that if two standard geodesic lines
l1 and l2 are not parallel, then πl1 (l2 ) is a vertex of l2 . Thus πlu (l) is
a vertex of lu . Suppose l0 is parallel to l, then there exists a chain of
parallel standard geodesic lines {li }ni=1 such that l1 = l, ln = l0 and
li , li+1 are in the same standard 2-flat Fi . Let l⊥ ⊂ Fi be a standard
GROUPS QUASI-ISOMETRIC TO RAAG’S
17
geodesic orthogonal to li and li+1 . Since v and u are not adjacent in
P(Γ), l⊥ and lu are not parallel, thus πlu (l⊥ ) is a point. It follows that
πlu (li ) = πlu (li+1 ), hence πlu (l) = πlu (l0 ).
3.4. Right-angled buildings. We will follow the treatment in [Dav98,
AB08, Ron09]. In particular, we refer to Section 1.1 to Section 1.3 of
[Dav98] for the definitions of chamber systems, galleries, residues, Coxeter groups and buildings. We will focus on right-angled buildings, i.e.
the associated Coxeter group is right-angled, though most of the discussion below is valid for general buildings.
Let W = W (Γ) be a right-angled Coxeter group with (finite) defining
graph Γ. Let B = B(Γ) be a right-angled building with the associated
W -distance function denoted by δ : B × B → W . We will also call B(Γ)
a right-angled Γ-building for simplicity.
Let I be the vertex set of Γ. Recall that a subset J ⊂ I is spherical
if the subgroup of W generated by J is finite. Let S be the poset of
spherical subsets of I (including the empty set) and let |S|∆ be the
geometric realization of S, i.e. |S|∆ is a simplicial complex such that its
vertices are in 1-1 correspondence to elements in S and its n-simplices
are in 1-1 correspondence to (n + 1)-chains in S. Note that |S|∆ is
isomorphic the simplicial cone over the barycentric subdivision of the
flag complex of Γ.
Recall that for elements x ≤ y in S, the interval Ixy between x and
y is a poset consist of elements z ∈ S such that x ≤ z ≤ y with
the induced order from S. There is a natural simplicial embedding
|Ixy |∆ ,→ |S|∆ . Each |Ixy |∆ is a simplicial cone over the barycentric
subdivision of a simplex, thus can be viewed a subdivision of a cube
into simplices. It is not hard to check the collection of all intervals
in S gives rise to a structure of cube complex on |S|∆ . Let |S| be the
resulting cube complex, then |S| is CAT (0).
A residue is spherical if it is a J-residue with J ∈ S. Let Sr be
the poset of all spherical residues in B. For x ∈ Sr which comes from
a J-residue, we define the rank of x to be the cardinality of J, and
define a type map t : Sr → S which maps x to J ∈ S. Let |Sr |∆ be
the geometric realization of Sr , then the type map induces a simplicial
map t : |Sr |∆ → |S|∆ . For x ∈ Sr , let Srx be the sub-poset made of
elements in Sr which is ≥ x. If x is of rank 0, then Srx is isomorphic
to S, moreover, there is a natural simplicial embedding |Srx |∆ → |Sr |∆
and t maps the image of |Srx |∆ isomorphically onto |S|∆ .
18
JINGYIN HUANG AND BRUCE KLEINER
As before, the geometric realization of each interval in Sr is a subdivision of a cube into simplices. Moreover, the intersection of two
intervals in Sr is also an interval. Thus one gets a cube complex |B|
whose cubes are in 1-1 correspondence with intervals in Sr . |B| is called
the Davis realization of the building B and |B| is a CAT (0) cube complex by [Dav98]. Moreover, the above type map induces a cubical map
t : |B| → |S|. Let R ⊂ B be a residue. Since R also has the structure
of a building, there is an isometric embedding |R| → |B| between their
Davis realizations. |R| is called a residue in |B|.
In the special case when B is equal to the associated Coxeter group
W , there is a natural embedding from the Cayley graph of W to |B|
such that vertices of Cayley graph are mapped to vertices of rank 0
in |B|. And |B| can be viewed as the first cubical subdivision of the
cubical completion of the Cayley graph of W (the cubical completion
means we attach an n-cube to the graph whenever there is a copy of
the 1-skeleton of an n-cube inside the graph).
Each vertex of |B| corresponds to a J-residue in B, thus has a welldefined rank. For a vertex x of rank 0, the space |Srx |∆ discussed in the
previous paragraph induces a subcomplex |Bx | ⊂ |B|. Note that t maps
|Bx | isomorphically onto |S|. |Bx | is called a chamber in |B|, and there is
a 1-1 correspondence between chambers in |B| and chambers in B. Let
|Bx | and |By | be two chambers in |B|. Since there is an apartment A ⊂
B which contains both x and y, this induces an isometric embedding
|A| → |B| whose image contains |Bx | and |By |, here |A| is isomorphic
to the Davis realization of the Coxeter group W . |A| is called an
apartment in |B|.
Definition 3.7. For c1 , c2 ∈ B, define d(c1 , c2 ) to be the minimal length
of word in W (with respect to the generating set I) that represents
δ(c1 , c2 ). For any two residues R1 , R2 ∈ B, we define d(R1 , R2 ) =
min{d(c, d) | c ∈ R1 , d ∈ R2 }. It turns out that for any c ∈ R1 and
d ∈ R2 with d(c, d) = d(R1 , R2 ), δ(c, d) gives rise to the same element
in W ([AB08, Chapter 5.3.2]), this element is defined to be δ(R1 , R2 ).
Lemma 3.8. d(c1 , c2 ) = 2dl1 (c1 , c2 ), here dl1 means the l1 -distance in
|B|. Since c1 and c2 can be also viewed as vertex of rank 0 in |B|,
dl1 (c1 , c2 ) makes sense.
Proof. If B = W , then this lemma follows from the above description
of the Davis realization of a Coxeter group. The general case can be
reduced to this case by considering an apartment |A| ⊂ |B| which
contains c1 and c2 . Note that |A| is convex in |B|.
GROUPS QUASI-ISOMETRIC TO RAAG’S
19
Given a residue R ⊂ B, there is a well-defined nearest point projection map as follows.
Theorem 3.9 (Proposition 5.34, [AB08]). Let R be a residue and c a
chamber. Then there exists a unique c0 ∈ R such that d(c, c0 ) = d(R, c).
This projection is compatible with several other projections in the
following sense. Let |R| ⊂ |B| be the convex subcomplex corresponding
to R. Let c and c0 be as above. We also view them as vertex of rank 0
in |B|. Let c1 be the combinatorial projection of c onto |R| (see [HW08,
Lemma 13.8]) and let c2 be the CAT (0) projection of c onto |R|.
Lemma 3.10. c0 = c1 = c2 .
Proof. c1 = c2 is actually true for any CAT (0) cube complexes. By
[Hua14a, Lemma 2.3], c2 is a vertex. If c2 6= c1 , by [HW08, Lemma
13.8], the concatenation of the combinatorial geodesic ω1 which connects c2 and c1 and the combinatorial geodesic ω2 which connects c1 and
c is a combinatorial geodesic connecting c and c2 . Note that ω1 ⊂ |R|.
Let e ⊂ ω1 be the edge that contains c2 and let x be the other endpoint
of e. Then x and c are in the same side of the hyperplane dual to e. It
is easy to see d(x, c) < d(c2 , c) (here d denotes the CAT (0) distance),
which yields a contradiction.
To see c0 = c1 , by Lemma 3.8, it suffices to prove c1 is of rank 0.
When B = W , this follows from c1 = c2 , since we can work with the
cubical completion of the Cayley graph of W instead of |W | (the latter
is the cubical subdivision of the former) and apply [Hua14a, Lemma
2.3]. The general case follows by considering an apartment |A| ⊂ |B|
which contains c1 and c, note that in this case |A| ∩ |R| can be viewed
as a residue in |A|.
Definition 3.11. Let projR be the map defined in Theorem 3.9. Two
residues R1 and R2 are parallel if projR1 (R2 ) = R1 and projR2 (R1 ) =
R2 . In this case projR1 and projR2 induce mutually inverse bijections
between R1 and R2 . These bijections are called parallelism maps between R1 and R2 . They are also isomorphisms of chamber system i.e.
they map residues to residues ([AB08, Proposition 5.37]).
It follows from the uniqueness of the projection map that if f : R →
R is the parallelism map between two parallel residues and R1 ⊂ R
is a residue, then R1 and f (R1 ) are parallel, and the parallelism map
between R1 and f (R1 ) is induced by f .
0
Lemma 3.12. If R1 and R2 are parallel, then |R1 | and |R2 | are parallel with respect to the CAT (0) metric on |B|. Moreover, the parallelism
20
JINGYIN HUANG AND BRUCE KLEINER
maps between R1 and R2 induces by projR1 and projR2 is compatible
with the CAT (0) parallelism between |R1 | and |R2 | induced by CAT (0)
projections.
Proof. By Lemma 3.10, it suffices to show for any residue R ∈ B, |R| is
the convex hull of the vertices of rank 0 inside |R|. This is clear when
B = W if one consider the cubical completion of the Cayley graph of
W . The general case also follows since |R| is a union of apartments in
|R|, and |R| is convex in |B|.
It follows that if R1 and R2 are parallel residues, and R2 and R3 are
parallel residues, then R1 is parallel to R3 . Moreover, let fij be the
parallelism map from Ri to Rj induced by the projection map, then
f13 = f23 ◦ f12 .
Given chamber systems C1 , · · · , Ck over I1 , · · · , Ik , their direct product C1 × · · · × Ck is a chamber system over the disjoint union I1 t · · · t
Ik . Its chambers are k-tuples (c1 , · · · , ck ) with ct ∈ Ct . For i ∈ It ,
(c1 , · · · , ck ) is i-adjacent to (d1 , · · · , dk ) if cj = dj for j 6= t and ct and
dt are i-adjacent.
Suppose the defining graph Γ of the right-angled Coxeter group W
admits a join decomposition Γ = Γ1 ◦Γ2 ◦· · ·◦Γk . Let I = ∪ki=1 Ii be the
Q
corresponding decomposition of the vertex set of Γ and W = ki=1 Wi
be the induced product decomposition of W . Pick chamber c ∈ B,
and let Bi be the Ii -residue that contains c. Define a map φ : B →
B1 × B2 × · · · × Bk by φ(d) = (projB1 (d), projB2 (d), · · · , projBk (d)) for
any chamber d ∈ B.
Theorem 3.13 (Theorem 3.10, [Ron09]). The definition of φ does not
depend on the choice of c, and φ is an isomorphism of buildings.
It follows from the definition of the Davis realization that there is a
natural isomorphism |B1 × B2 × · · · × Bk | ∼
= |B1 | × |B2 | × · · · × |Bk |, thus
we have a product decomposition |B| ∼
|B
= 1 | × |B2 | × · · · × |Bk |, where
the isomorphism is induced by CAT (0) projections from |B| to |Bi |’s
(this is a consequence of Lemma 3.12).
We define the parallel set of a residue R ⊂ B to be the union of all
residues in B that are parallel to R.
Lemma 3.14. Suppose R is a J-residue. Let J ⊥ ⊂ I be the collection
of vertices in Γ which are adjacent to every vertex in J. Then:
(1) If R0 is parallel to R, then R0 is a J-residue.
(2) The parallel set of R is the J ∪ J ⊥ -residue that contains R.
GROUPS QUASI-ISOMETRIC TO RAAG’S
21
Note that this lemma is not true if the building under consideration
is not right-angled.
Proof. Suppose R0 is a J1 -residue. Let w = δ(R, R0 ) (see Definition
3.7). It follows from (2) of [AB08, Lemma 5.36] that R0 is a (J ∩
wJ1 w−1 )-residue. Since R and R0 are parallel, they have the same
rank, thus J = wJ1 w−1 . By considering the abelianization of the
right-angled Coxeter group W , we deduce that J = J1 (this proves
the first assertion of the lemma) and w commutes with each element
in J. Thus w belongs to the subgroup generated by J ⊥ and R0 is in
the J ∪ J ⊥ -residue S that contains R. Then the parallel set of R is
contained in S. It remains to prove every J-residue in S is parallel to
R, but this follows from Theorem 3.13.
Q
Pick a vertex v ∈ |B| of rank k and let R = ki=1 Ri be the associated residue with its product decomposition. Let {vλ }λ∈Λ be the
collection of vertices that are adjacent to v. Then there is a decomposition {vλ }λ∈Λ = {vλ ≤ v} t {vλ > v}, where {vλ > v} denotes the
collection of vertices whose associated residues contain R. This induces
a decomposition Lk(v, |B|) = K1 ∗ K2 of the link of v in |B| ([BH99,
Definition I.7.15]) into a spherical join of two CAT (1) all-right spherical complexes. Note that K2 is finite, since {vλ>v } is finite. Moreover,
Qk
K1 ∼
= Lk(v, |R|). However, |R| ∼
= i=1 |Ri |, thus K1 is the spherical
join of k discrete sets such that elements in each of these discrete sets
are in 1-1 correspondence to elements in some Ri . Now we can deduce
from this the following result.
Lemma 3.15. Suppose B is a right-angled building such that each of its
residues of rank 1 contains infinitely many elements. If α : |B| → |B|
is a cubical isomorphism, then α preserves the rank of vertices in |B|.
Part I: A fibration approach to cubulating RAAG’s
4. Restriction quotients
In this section we study restriction quotients, a certain type of mapping between CAT(0) cube complexes introduced by Caprace and Sageev
in [CS11]. These play a central role in our story.
We first show in Subsection 4.1 that restriction quotients can be
characterized in several different ways, see Theorem 4.4. We then show
in Subsection 4.2 that a restriction quotient f : Y → Z determines
fiber data that satisfies certain conditions; conversely, given such fiber
22
JINGYIN HUANG AND BRUCE KLEINER
data, one may construct a restriction quotient inducing the given data,
which is unique up to equivalence. This correspondence will later be
applied to construct restriction quotients over right-angled buildings.
Subsections 4.3 and 4.4 deals with the behavior of restriction quotients
under group actions and quasi-isometries.
4.1. Quotient maps between CAT(0) cube complexes. We recall
the notion of restriction quotient from [CS11, Section 2.3]; see Section
9.1 for the background on wallspaces.
Definition 4.1. Let Y be a CAT(0) cube complex and let H be the
collection of walls in the 0-skeleton Y (0) corresponding to the hyperplanes in Y . Pick a subset K ⊂ H and let Y (K) be the CAT(0) cube
complex associated with the wallspace (Y (0) , K). Then every 0-cube of
the wallspace (Y (0) , H) gives rise to a 0-cube of (Y (0) , K) by restriction.
This can be extended to a surjective cubical map q : Y → Y (K), which
is called the restriction quotient arising from the subset K ⊂ H.
The following example motivates many of the constructions in this
paper:
Example 4.2 (The canonical restriction quotient of a RAAG). For a
fixed graph Γ, let Se → |c|Γ and Xe → Se be the mappings associated
with the exploded Salvetti complex, as defined in the introduction after
Definition 1.11. Let K be the collection of hyperplanes in Xe (Γ) dual
to edges e ⊂ Xe that project to edges under the composition Xe →
Se → |c|Γ . Then the canonical restriction quotient of G = G(Γ) is the
restriction quotient arising from K.
Let q : Y → Y (K) be a restriction quotient. Pick an edge e ⊂ Y . If
e is dual to some element in K, then q(e) is an edge, otherwise q(e) is
a point. The edge e is called horizontal in the former case, and vertical
in the latter case. We record the following simple observation.
Lemma 4.3. Let α : Y → Y be a cubical CAT(0) automorphism
of Y that maps vertical edges to vertical edges and horizontal edges to
horizontal edges. Then α descends to an automorphism Y (K) → Y (K).
The following result shows that restriction quotients may be characterized in several different ways.
Theorem 4.4. If f : Y → Z is a surjective cubical map between two
CAT(0) cube complexes, then the following conditions are equivalent:
(1) The inverse image of each vertex of Z is convex.
GROUPS QUASI-ISOMETRIC TO RAAG’S
(2)
(3)
(4)
(5)
23
The inverse image of every point in Z is convex.
The inverse image of every convex subcomplex of Z is convex.
The inverse image of every hyperplane in Z is a hyperplane.
f is equivalent to a restriction quotient, i.e. for some set of
walls K in Y , there is a cubical isomorphism φ : Z → Y (K)
making the following diagram commute:
Y
q
f
>
∨
Z
> Y (K)
φ
The proof of Theorem 4.4 will take several lemmas. For the remainder of this subsection we fix CAT(0) cube complexes Y and Z and a
(not necessarily surjective) cubical map f : Y → Z.
Lemma 4.5. Let σ ⊂ Z be a cube and let Yσ be the be the union of
cubes in Y whose image under f is exactly σ. Then:
(1) If y ∈ σ is an interior point, then f −1 (y) ⊂ Yσ .
(2) f −1 (y) has a natural induced structure as a cube complex; moreover, there is a natural isomorphism of cube complexes Yσ ∼
=
−1
f (y) × σ.
(3) If σ1 ⊂ σ2 are cubes of Z and yi ∈ σi are interior points, then
there is a canonical embedding f −1 (y2 ) ,→ f −1 (y1 ). Moreover,
these embeddings are compatible with composition of inclusions.
Lemma 4.6.
(1) For every y ∈ Z, every connected component of f −1 (y) is a
convex subset of Y .
(2) For every convex subcomplex A ⊂ Z, every connected component of f −1 (A) is a convex subcomplex of Y .
Proof. First we prove (1). Let σ be the support of y and let Yσ ∼
=
−1
f (y) × σ be the subcomplex defined as above. It suffices to show Yσ
is locally convex. Pick vertex x ∈ Yσ , and let {ei }ni=1 be a collection
of edges in Yσ that contains x. It suffices to show if these edges span
an n-cube η ⊂ Y , then η ⊂ Yσ . It suffices to consider the case when
all ei ’s are orthogonal to σ, in which case it follows from Definition 3.4
that η × σ ⊂ Yσ .
To see (2), pick an n-cube η ⊂ Y and let {ei }ni=1 be the edges of η at
one corner c ⊂ η. It suffices to show if f (ei ) ⊂ A, then f (η) ⊂ A. Note
that f (η) is a cube, and every edge of this cube which emanates from
24
JINGYIN HUANG AND BRUCE KLEINER
the corner f (c) is contained in A. Thus f (η) ⊂ A by the convexity of
A.
Lemma 4.7. Let f : Y → Z be a cubical map as above. Then:
(1) The inverse image of each hyperplane of Z is a disjoint union
of hyperplanes in Y .
(2) If the inverse image of each hyperplane of Z is a single hyperplane, then for each point y ∈ Z, the point inverse f −1 (y) is
connected, and hence convex.
Proof. It follows from Definition 3.4 that the inverse image of each
hyperplane of Z is an union of hyperplanes. If two of them were to
intersect, then there would be a 2-cube in Y with two consecutive
edges mapped to the same edge in Z, which is impossible.
Now we prove (2). It suffices to consider the case that y is the
center of some cube in Z. In this case, y is a vertex in the first cubical
subdivision of Z, and f can viewed as a cubical map from the first
cubical subdivision of Y to the first cubical subdivision of Z such that
the inverse image of each hyperplane is a single hyperplane, thus it
suffices to consider the case that y is a vertex of Z.
Suppose f −1 (y) contains two connected components A and B. Pick
a combinatorial geodesic ω of shortest distant that connects vertices
in A and vertices in B. Note that f (ω) is a non-trivial edge-loop in
Z, otherwise we will have ω ⊂ f −1 (y). It follows that there exists two
different edges e1 and e2 of ω mapping to parallel edges in Y . The
hyperplanes dual to e1 and e2 are different, yet they are mapped to the
same hyperplane in Y , which is a contradiction.
Lemma 4.8. If f is surjective, and for any vertex v ∈ Z, f −1 (v) is
connected, then the inverse image of each hyperplane of Z is a single
hyperplane.
Proof. Let h ⊂ Z be a hyperplane, by Lemma 4.7, f −1 (h) = tλ∈Λ hλ
where each hλ is a hyperplane in Y . Since f is surjective, {f (hλ )}λ∈Λ
is a collection of subcomplexes of h that cover h. Thus there exists
h1 , h2 ∈ {hλ }λ∈Λ and vertex u ∈ h such that u ⊂ f (h1 ) ∪ f (h2 ). Let
e ⊂ Z be the edge such that u = e∩h, then there exist edges e1 , e2 ⊂ Y
such that ei ∩ hi 6= ∅ and f (ei ) = e for i = 1, 2. Since h1 ∩ h2 = ∅, a
case study implies there exist x1 and x2 which are endpoints of e1 and
e2 respectively such that
(1) these two points are separated by at least one of h1 and h2 ;
GROUPS QUASI-ISOMETRIC TO RAAG’S
25
(2) they are mapped to the same end point y ∈ e.
It follows that f −1 (y) is disconnected, which is a contradiction.
Remark 4.9. If f is not surjective, then the above conclusion is not
necessarily true. Consider the map from A = [0, 3] × [0, 1] to the
unit square which collapses the [0, 1] factor in A and maps [0, 3] to 3
consecutive edges in the boundary of the unit square.
Lemma 4.10. If q : Y → Y (K) is the restriction quotient as Definition 4.1, then the inverse image of each hyperplane in Y (K) is a single
hyperplane in Y . Conversely, suppose f : Y → Z is a surjective cubical map between CAT(0) cube complexes such that the inverse image of
each hyperplane is a hyperplane. Let K be the collection of walls arising from inverse images of hyperplanes in Z. Then there is a natural
isomorphism i : Z ∼
= Y (K) which fits into the following commutative
diagram:
f
Y
>Z
q > i∨
Y (K)
Proof. Define two vertices of Y to be K-equivalent if and only if they
are not separated by any wall in K. This defines an equivalence relation
on vertices of Y , and the corresponding equivalent classes are called Kclasses. For each K-class C and every wall in K, we may choose the
halfspace that contains C; it follows that the points in C are exactly
the set of vertices contained in the intersection of such halfspaces, and
thus C is the vertex set of a convex subcomplex of Y . Note that each
K-class determines a 0-cube of (Y 0 , K), hence is mapped to this 0-cube
under q. It follows that the inverse image of every vertex in Y (K) is
convex, thus by Lemma 4.8, the inverse image of a hyperplane is a
hyperplane.
It remains to prove the converse. Note that the inverse image of each
halfspace in Z under f is a halfspace of Y . Moreover, the surjectivity
of f implies that f maps hyperplane to hyperplane and halfspace to
halfspace. Pick vertex y ∈ Z, let {Hλ }λ∈λ be the collection of hyperplanes in Z that contains y. Then f −1 (y) ⊂ ∩λ∈Λ f −1 (Hλ ), and every
vertex of ∩λ∈Λ f −1 (Hλ ) is mapped to y by f , and thus the vertex set of
f −1 (y) is a K-class. This induces a bijective map from Z (0) to the vertex set of Y (K), which extends to an isomorphism. The above diagram
commutes since it commutes when restricted to the 0-skeleton.
26
JINGYIN HUANG AND BRUCE KLEINER
Proof of Theorem 4.4. The equivalence of (4) and (5) follows from Lemma
4.10. (1) ⇒ (4) follows from Lemma 4.8, (4) ⇒ (2) follows from Lemma
4.7, (3) ⇒ (1) is obvious. It suffices to show (2) ⇒ (3). Pick a convex
subcomplex K ⊂ Z and let {Rλ }λ∈Λ be the collection of cubes in K.
For each Rλ , let YRλ be the subcomplex defined after Definition 3.4.
YRλ 6= ∅ since f is surjective and YRλ is connected by (2). If Rλ ⊂ Rλ0 ,
then YRλ ∩ YRλ0 6= ∅. Thus f −1 (K) = ∪λ∈Λ YRλ is connected, hence
convex.
4.2. Restriction maps versus fiber functors. If q : Y → Z is a
restriction quotient between CAT(0) cube complexes, then we may express the fiber structure in categorical language as follows. Let Face(Z)
denote the face poset of Z, viewed as a category, and let CCC denote
the category whose objects are nonempty CAT(0) cube complexes and
whose morphisms are convex cubical embeddings. By Lemma 4.5, we
obtain a contravariant functor Ψq : Face(Z) → CCC.
Definition 4.11. The contravariant functor Ψq is the fiber functor of
the restriction quotient q : Y → Z.
For notational brevity, for any inclusion i : σ1 → σ2 , we will often
denote the map Ψ(i) : Ψ(σ2 ) → Ψ(σ1 ) simply by Ψ(σ2 ) → Ψ(σ1 ),
suppressing the name of the map.
Note that if σ1 ⊂ σ2 ⊂ σ3 , then the functor property implies that
the image of Ψ(σ3 ) → Ψ(σ1 ) is a convex subcomplex of the image of
Ψ(σ2 ) → Ψ(σ1 ). In particular, if v is a vertex of a cube σ, then the
image of Φ(σ) → Ψ(v) is a convex subcomplex of the intersection
\
Im(Ψ(e) → Ψ(v))
v(e⊂σ (1)
Definition 4.12. Let Z be a cube complex. A contravariant functor
Ψ : Face(Z) → CCC is 1-determined if for every cube σ ∈ Face(Z),
and every vertex v ∈ σ (0) ,
\
(4.13)
Im(Ψ(σ) −→ Ψ(v)) =
Im(Ψ(e) → Ψ(v)) .
v(e⊂σ (1)
Lemma 4.14. If q : Y → Z is a restriction quotient, then the fiber
functor Ψ : Face(Z) → CCC is 1-determined.
Proof. Pick σ ∈ Face(Z), v ∈ σ (0) . We know that Im(Ψ(σ) → Ψ(v))
is a nonempty convex subcomplex of ∩v(e⊂σ(1) Im(Ψ(e) → Ψ(v)), so to
establish (4.13) we need only show that the two convex subcomplexes
have the same 0-skeleton.
GROUPS QUASI-ISOMETRIC TO RAAG’S
27
Pick a vertex w ∈ Im(Ψ(σ) → Ψ(v)), and let w0 ∈ ∩v(e⊂σ(1) Im(Ψ(e) →
Ψ(v)) be a vertex adjacent to w. We let τ ∈ Face(Y ) denote the edge
spanned by w, w0 . For every edge e of Z with v ( e ⊂ σ (1) , let ê ⊂ Y (1)
denote the edge with q(ê) = e that contains w. By assumption, the
collection of edges {τ } ∪ {ê}v(e⊂σ(1) determines a complete graph in the
link of w, and therefore is contained in a cube σ̂ of dimension 1+dim σ.
Then q(σ̂) = σ and τ ⊂ σ̂; this implies that τ ⊂ Im(Ψ(σ) → Ψ(v)).
Since the 1-skeleton of ∩v(e⊂σ(1) Im(Ψ(e) → Ψ(v)) is connected, we
conclude that it coincides with the 1-skeleton of Im(Ψ(σ) → Ψ(v)). By
convexity, we get (4.13).
Theorem 4.15. Let Z be a CAT(0) cube complex, and Ψ : Face(Z) →
CCC be a 1-determined contravariant functor. Then there is a restriction quotient q : Y → Z such that the associated fiber functor
Ψq : Face(Z) → CCC is equivalent by a natural transformation to Ψ.
Proof. We first construct the cube complex Y , and then verify that it
has the desired properties.
F
We begin with the disjoint union σ∈Face(Z) (σ × Ψ(σ)), and for every
inclusion σ ⊂ τ , we glue the subset σ × Ψ(τ ) ⊂ τ × Ψ(τ ) to σ × Ψ(σ)
by using the map
idσ ×Ψ(σ⊂τ )
σ × Ψ(τ ) −−−−−−−→ σ × Ψ(σ) .
F
One checks that the cubical structure on σ∈Face(Z) (σ ×Ψ(σ)) descends
to the quotient Y , the projection maps σ × Ψ(σ) → σ descend to a
cubical map q : Y → Z, and for every σ ∈ Face(Z), the union of the
cubes σ̂ ⊂ Y such that f (σ̂) = σ is a copy of σ × Ψ(σ).
We now verify that links in Y are flag complexes.
Let v be a 0-cube in Y , and suppose σ1 , . . . , σk are 1-cubes containing
v, such that for all 1 ≤ i 6= j ≤ k the 1-cubes σi , σj span a 2-cube σij
in the link of v. We may assume after reindexing that for some h ≥ 0
the image q(σi ) is a 1-cube in Z if i ≤ h and a 0-cube if i > h.
Since Ψ(v) is a CAT(0) cube complex, the edges {σi }i>h span a cube
σvert ⊂ q −1 (v).
For 1 ≤ i 6= j ≤ h, the 2-cube σij projects to a 2-cube q(σij ) spanned
by the two edges q(σi ), q(σj ). Since Z is a CAT(0) cube complex,
the edges {q(σi )}i≤h span an h-cube σ̄hor ⊂ Z. By the 1-determined
property, we get that Im(Ψ(σ̄hor ) → Ψ(v)) contains v, and so there is
an h-cube σhor ⊂ Y containing v such that q(σhor ) = σ̄hor .
28
JINGYIN HUANG AND BRUCE KLEINER
Fix 1 ≤ i ≤ h. Then for j > h, the 2-cube σij projects to q(σi ), and
hence σj belongs to Im(Ψ(σi ) → Ψ(v)). If j, k > h, then σj , σk both
belong to Im(Ψ(σi ) → Ψ(v)), and by the convexity of Im(Ψ(σi ) →
Ψ(v)) in Ψ(v), we get that σjk also belongs to Im(Ψ(σi ) → Ψ(v)).
Applying convexity again, we get that σvert ⊂ Im(Ψ(σi ) → Ψ(v)). By
the 1-determined property, it follows that σvert ⊂ Im(Ψ(σ̄hor ) → Ψ(v)).
This yields a k-cube σ ⊂ Y containing σhor ∪ σvert , which is spanned
by σ1 , . . . , σk .
Thus we have shown that links in Y are flag complexes. The fact that
the fibers of f : Y → Z are contractible implies that Y is contractible
(in particular simply connected), so Y is CAT(0).
We now observe that the construction of restriction quotients is compatible with product structure:
Lemma 4.16 (Behavior under products). For i ∈ {1, 2} let qi : Yi →
Zi be a restriction quotient with fiber functor Ψi : Face(Zi ) → CCC.
Then the product q1 × q2 : Y1 × Y2 → Z1 × Z2 is a restriction quotient
with fiber functor given by the product:
Ψ ×Ψ
×
2
Face(Z1 × Z2 ) ' Face(Z1 ) × Face(Z2 ) −−1−−→
CCC × CCC −
→ CCC .
In particular, if one starts with CAT(0) cube complexes Zi and fiber
functors Ψi : Zi → CCC for i ∈ {1, 2}, then the product fiber functor
defined as above is the fiber functor of the product of the restriction
quotients associated to the Ψi ’s.
4.3. Equivariance properties. We now discuss isomorphisms between restriction quotients, and the naturality properties of the restriction quotient associated with a fiber functor.
Suppose we have a commutative diagram
α̂
Y1
> Y2
q1
q2
∨ α ∨
Z1
> Z2
where the qi ’s are restriction quotients and α, α̂ are cubical isomorphisms. Let Ψi : Face(Zi ) → CCC be the fiber functor associated with
qi . Notice that the pair α, α̂ allows us to compare the two fiber functors,
since for every σ ∈ Face(Z1 ), the map α̂ induces a cubical isomorphism
between Ψ1 (σ) and Ψ2 (α(σ)), and this is compatible with maps induced
GROUPS QUASI-ISOMETRIC TO RAAG’S
29
with inclusions of faces. This may be stated more compactly by saying
that α̂ induces a natural isomorphism between the fiber functors Ψ1
and Ψ2 ◦ Face(α), where Face(α) : Face(Z1 ) → Face(Z2 ) is the poset
isomorphism induced by α. Here the term natural isomorphism is being used in the sense of category theory, i.e. a natural transformation
that has an inverse that is also a natural transformation.
Now suppose that for i ∈ {1, 2} we have a CAT(0) cube complex Zi
and a 1-determined fiber functor Ψi : Face(Zi ) → CCC. Let fi : Yi →
Zi be the associated restriction quotients. If we have a pair α, β, where
α : Z1 → Z2 is a cubical isomorphism, and β is a natural isomorphism
between the fiber functors Ψ1 and Ψ2 ◦ Face(α), then we get an induced
map α̂ : Y1 → Y2 , which may be defined by using the description of Yi
as the quotient of the disjoint collection {σ × Ψi (σ)}σ∈Face(Zi ) .
As a consequence of the above, having an action of a group G on
a restriction quotient f : Y → Z is equivalent to having an action
G y Z together with a compatible “action” on the fiber functor Ψf , i.e.
a family {(α(g), β(g))}g∈G as above that also satisfies an appropriate
composition rule.
4.4. Quasi-isometric properties. We now consider the coarse geometry of restriction quotients; this amounts to a “coarsification” of the
discussion in the preceding subsection.
The relevant definition is a coarsification of the natural isomorphisms
between fiber functors.
Definition 4.17. Let Z be a CAT(0) cube complex and Ψi : Face(Z) →
CCC be fiber functors for i ∈ {1, 2}. An (L, A)-quasi-natural isomorphism from Ψ1 to Ψ2 is a collection {φ(σ) : Ψ1 (σ) → Ψ2 (σ)}σ∈Face(Z)
such that φ(σ) is an (L, A)-quasi-isometry for all σ ∈ Face(Z), and for
every inclusion σ ⊂ τ , the diagram
φ(τ )
Ψ1 (τ )
> Ψ2 (τ )
∨ φ(σ) ∨
Ψ1 (σ)
> Ψ2 (σ)
commutes up to error L.
Now for i ∈ {1, 2} let fi : Yi → Z be a finite dimensional restriction
quotient, with respective fiber functor Ψi : Face(Z) → CCC. For
any σ ∈ Face(Z), we identify Ψi (σ) with fi (bσ ), where bσ ∈ σ is the
barycenter.
30
JINGYIN HUANG AND BRUCE KLEINER
Lemma 4.18. Suppose we have a commutative diagram
φ
Y1 <
> Y2
<
φ
>
f1
0
f2
Z
where φ, φ0 are (L, A)-quasi-isometries that are A-quasi-inverses, i.e.
the compositions φ ◦ φ0 , φ0 ◦ φ are at distance < A from the identity
maps. Then the collection
φ
{Ψ1 (σ) =
f1−1 (bσ )
|
f1−1
(bσ )
−→
f2−1 (bσ ) = Ψ2 (σ)}σ∈Face(Z)
is an (L0 , A0 )-quasi-natural isomorphism where L0 = L0 (L, A, dim Yi ),
A0 = A0 (L, A, dim Yi ).
Proof. By Theorem 4.4, the fiber fi−1 (bσ ) is a convex subset of Yi , and
hence is isometrically embedded. Therefore φ and φ0 induce (L, A)quasi-isometric embeddings f1−1 (bσ ) → f2−1 (bσ ), f2−1 (σb ) → f1−1 (bσ ). If
σ ⊂ τ , then any point x ∈ fi−1 (bτ ) lies at distance < C = C(dim Yi )
from a point in fi−1 (bσ ), and this implies that the collection of maps
{Ψ1 (σ) → Ψ2 (σ)}σ∈Face(Z) is an (L0 , A0 )-quasi-natural isomorphism as
claimed.
Lemma 4.19. If {φ(σ) : Ψ1 (σ) → Ψ2 (σ)}σ∈Face(Z) is an (L, A)-quasinatural isomorphism from Ψ1 to Ψ2 , then it arises from a commutative diagram as in the previous lemma, where φ, φ0 are (L0 , A0 )-quasiisometries that are A0 -quasi-inverses, and L0 , A0 depend only on L, A,
and dim Yi .
Proof. For every σ ∈ Face(Z), we may choose a quasi-inverse φ0 (σ) :
Ψ2 (σ) → Ψ1 (σ) with uniform constants; this is also a quasi-natural
isomorphism. Identifying fi−1 (Int(σ)) with the product Int(σ) × Ψi (σ),
we define φ|f −1 (Int(σ)) by
1
f1−1 (Int(σ)) = Int(σ) × Ψ1 (σ)
idInt(σ) ×φ(σ)
−→
Int(σ) × Ψ2 (σ) = f2−1 (σ) ,
and φ0 similarly using {φ0 (σ)}σ∈Face(Z) . One readily checks that φ, φ0
are quasi-isometric embeddings that are also quasi-inverses, where the
constants depend on L, A, and dim Yi .
GROUPS QUASI-ISOMETRIC TO RAAG’S
31
5. The Z-blow-up of right-angled Building
In this section Γ will be an arbitrary finite simplicial graph, and all
buildings will be right-angled buildings modelled on the right-angled
Coxeter group W (Γ) with defining graph Γ. The reader may wish to
review Section 3.4 for terminology and notation regarding buildings,
before proceeding.
The goal of this section is examine restriction quotients q : Y → |B|,
where the fibers are Euclidean spaces satisfying a dimension condition
as in Theorem 1.5 or 1.6. For such restriction quotients, the fiber
functor may be distilled down to 1-data, see Definition 5.3; this is
discussed in Subsection 5.2. Conversely, given a building B and certain
blow-up data (Definition 5.8), one can construct a corresponding 1determined fiber functor as in Section 4.2; see Subsection 5.3.
5.1. The canonical restriction quotient for a RAAG. Let G(Γ)
be the RAAG with defining graph Γ and let B(Γ) be the building
associated with G(Γ) (see [Dav98, Section 5]). Then G(Γ) can identified
with the set of chambers of B(Γ). Under this identification, the Jresidues of B, for J a collection of vertices in Γ, are the left cosets
of the standard subgroups of G(Γ) generated by J. Thus the poset of
spherical residues is exactly the poset of left cosets of standard Abelian
subgroups of G(Γ), which is also isomorphic to the poset of standard
flats in X(Γ).
We now revisit the discussion after Definition 1.11 and Example 4.2
in more detail, and relate them to buildings. To simplify notation, we
will write G = G(Γ), B = B(Γ) and X = X(Γ).
Let |B| be the Davis realization of the building B. Then we have
an induced isometric action G y |B|, which is cocompact, but not
proper. It turns out there is natural way to blow-up |B| to obtain a
space Xe = Xe (Γ) such that there is a geometric action G y Xe and a
G-equivariant restriction quotient map Xe → |B|.
Xe can be constructed as follows. First we constructed the exploded
Salvetti complex Se = Se (Γ), which was introduced in [BKS08a], see
also the discussion after Definition 1.11. Suppose L is the “lollipop”,
which is the union of a unit circle S and a unit interval I along one
point. For each vertex v in the vertex set V (Γ) of Γ, we associate
a copy
Q of Lv = Sv ∪ Iv , and let ?v ∈ Lv be the free end of Iv . Let
T = v∈V (Γ) Lv . Each clique ∆ ⊂ Γ gives rise to a subcomplex T∆ =
Q
Q
v∈∆ Lv ×
v ∈∆
/ {?v }. Then Se is the subcomplex of T which is the
32
JINGYIN HUANG AND BRUCE KLEINER
union of all such T∆ ’s, here ∆ is allowed to be empty. It is easy to
check Se is a non-positively curved
A standard torus
Q cube complex.
Q
in Se is a subcomplex of form v∈∆ Sv × v∈∆
/ {?v }, where ∆ ⊂ Γ
is a clique. Note that there is a unique standard torus of dimension
0, which corresponds to the empty clique. There is a natural map
Se = Se (Γ) → S(Γ) by collapsing the Iv -edge in each Lv -factor. This
maps induces a 1-1 correspondence between standard tori in Se and
standard tori in S(Γ). Notice that there is also a 1-1 correspondence
between vertices in Se and standard tori in Se .
Let Xe be the universal cover of Se . Then Xe is a CAT (0) cube
complex and the action G y Xe is geometric. The inverse images
of standard tori in Se are called standard flats. Note that each vertex in Xe is contained in a unique standard flat. We define a map
(0)
between the 0-skeletons p : Xe (Γ) → |B|(0) as follow. Pick a Gequivariant identification between 0-dimensional standard flats in X
and elements in G, and pick a G-equivariant map φ : Xe → X induced by Se = Se (Γ) → S(Γ) described as above. Note that c induces
a 1-1 correspondence between standard flats in Xe and standard flats
in X. This gives rise to a 1-1 correspondence between standard flats
in Xe and left cosets of standard Abelian subgroups of G. For each
(0)
x ∈ Xe (Γ), we define p(x) to be the vertex in |B|(0) that represents
the left coset of the standard Abelian subgroup of G which corresponds
to the unique standard flat that contains x.
A vertical edge of Xe is an edge which covers some Sv -circle in Se .
A horizontal edge of Xe is an edge which covers some Iv -interval in
Se . Two endpoints of every vertical edge are in the same standard
flat, thus they are mapped by p to the same point in |B|(0) . More
generally, for any given vertical cube, i.e. every edge in this cube is a
vertical edge, its vertex set is mapped by p to one point in |B|(0) . Pick
a horizontal edge and let F1 , F2 ⊂ Xe be standard flats which contain
the two endpoints of this edge respectively. Then φ(F1 ) and φ(F2 ) are
two standard flats in X such that one is contained as a codimension 1
flat inside another. More generally, if σ is a horizontal cube, i.e. each
edge of σ is a horizontal edge, then by looking the image of σ under
the covering map Xe → Se , we know the vertex set of σ corresponds
to an interval in the poset of standard flats of X. Every cube in Xe
splits as a product of a vertical cube and a horizontal cube (again this
is clear by looking at cells in Se ). Thus we can extend p to a cubical
map p : Xe → |B|.
GROUPS QUASI-ISOMETRIC TO RAAG’S
33
By construction, for a vertex v ∈ |B| of rank k, p−1 (v) is isometric
to En . It follows from Theorem 4.4 that p arises from a restriction
quotient, and this is called the canonical restriction quotient for the
RAAG G. This restriction quotient is exactly the one described in
Example 4.2, since hyperplanes in K of Example 4.2 are those which are
dual to horizontal edges. We record following immediate consequence
of the this construction.
Lemma 5.1. Let σ ⊂ |B| be a cube and let v ∈ σ be the vertex of
minimal rank in σ. Then for any interior point x ∈ σ, p−1 (x) is
isometric to Erank(σ) .
Remark 5.2. In the literature, there is a related cubical map X →
|B| defined as follows. First we recall an alternative description of
X. Actually similar spaces can be defined for all Artin groups (not
necessarily right-angled) and was introduced by Salvetti. We will follow
the description in [Cha]. Let G → W (Γ) be the natural projection map.
This map has a set theoretic section defined by representing an element
w ∈ W by a minimal length positive word with respect to the standard
generating set and setting σ(w) to be the image of this word G. It
follows from fundamental facts about Coxeter groups that σ is welldefined. Let I be the vertex set of Γ, and for any J ⊂ I, let W (J)
be the subgroup of W (Γ) generated by J. Let K be the geometric
realization of the following poset:
{gσ(W (J)) | g ∈ G, J ⊂ I, W (J) is finite}.
It turns out that K is isomorphic to the first barycentric subdivision
of X. Let G(J) ≤ G be the subgroup generated by J. We associate
each gσ(W (J)) with the left coset gG(J), and this induces a cubical
map from the first cubical subdivision of X to |B|. However, this map
is not a restriction quotient, since it has a lot of foldings (think of the
special case when G ∼
= Z).
5.2. Restriction quotients with Euclidean fibers. We reminder
the reader that in this section, W = W (Γ) will be the right-angled
Coxeter group with defining graph Γ and standard generating set I.
Let B be an arbitrary right-angled building modelled on W . Let S be
the poset of spherical subsets of I and let |B| be the Davis realization
of B.
Our next goal is to generalize the canonical restriction quotient mentioned in the previous subsection. However, to motivate our construction, we will first consider a restriction quotient q : Y 0 → |B| which
34
JINGYIN HUANG AND BRUCE KLEINER
satisfies the conclusion of Lemma 5.1, and identify several key features
of q.
Let Φ be the fiber functor associated with q (see Section 4.2). For
any vertices v, w ∈ |B|, we will write v ≤ w if and only if the residue
associated with v is contained in the residue associated with w.
Let Sr be the poset of spherical residues in B. Then Φ induces a
functor Φ0 from Sr to CCC (Section 4.2) as follows. Each element in
Sr is associated with the fiber of the corresponding vertex in |B|. If
s, t ∈ Sr are two elements such that rank(t) = rank(s) + 1 and s < t,
then the associated vertices in vs , vt ∈ |B| are joined by an edge est . In
this case Φ(est ) → Φ(vs ) is an isomorphism, so we define the morphism
Φ0 (s) → Φ0 (t) to be the map induced by Φ(est ) → Φ(vt ). If s, t ∈ Sr are
arbitrary two elements with s ≤ t, then we find an ascending chain from
s to t such that the difference between the ranks of adjacent elements
in the chain is 1, and define Φ0 (s) → Φ0 (t) be the composition of those
maps induced by the chain. It follows from the functor property of Φ
that Φ0 (s) → Φ0 (t) does not depend on the choice of the chain, and Φ0 is
a functor. Recall that there is a 1-1 correspondence between elements
in Sr and vertices of |B|, so we will also view Φ0 as a functor from the
vertex set of |B| to CCC. Let σ1 ⊂ σ2 be faces in |B| and let vi be the
vertex of minimal rank in σi for i = 1, 2. Then by our construction,
then morphism Φ(σ2 ) → Φ(σ1 ) is the same as Φ0 (v2 ) → Φ0 (v1 ).
Definition 5.3 (1-data). Pick a vertex v ∈ |B| of rank 1, and let Rv be
the associated residue. Let {vλ }λ∈Λ be the collection of vertices in |B|
which is < v and let eλ be the edge joining v and vλ . Then there is a 1-1
correspondence between elements in Rv and vλ ’s. Each vλ determines
a point in Φ(v) by consider the image of Φ(eλ ) → Φ(v). This induced a
map fRv : Rv → Φ(v). The collection of all such fRv ’s with v ranging
over all rank 1 vertices of |B| is called the 1-data associated with the
restriction quotient q : Y 0 → |B|.
Lemma 5.4. Pick two vertices v, u ∈ |B| of rank 1, and let Rv , Ru
be the corresponding residues. Suppose these two residues are parallel
with the parallelism map given by p : Rv → Ru . Then:
(1) Φ(v) and Φ(u), considered as convex subcomplexes of Y 0 , are
parallel.
(2) If p0 : Φ(v) → Φ(u) is the parallelism map, then the following
diagram commutes:
GROUPS QUASI-ISOMETRIC TO RAAG’S
35
p
Rv −−−→ Ru
fR v y
fRu y
p0
Φ(v) −−−→ Φ(u)
Proof. It follows from Lemma 3.14 that there is a finite chain of residues,
starting at Rv and ending at Ru , such that adjacent elements in the
chain are parallel residues in a spherical residue of rank 2. Thus we
can assume without loss of generality that Rv and Ru are contained in
the a spherical residue S of type J = {j, j 0 }, and we assume both Rv
and Ru are j-residues.
Pick x ∈ Rv . By Theorem 3.13, there is a j 0 -residue W which contains both x and p(x). Let s, w ∈ |B| be the vertex corresponding
to S and W. Note that there is a 2-cube in |B| such that v, w, s
are its vertices. Since Φ is 1-determined, Im(Φ0 (v) → Φ0 (s)) and
Im(Φ0 (w) → Φ0 (s)) are orthogonal lines in the 2-flat Φ0 (s). Moreover, the intersection these two lines is the image of fRv (x) under the
morphism Φ0 (v) → Φ0 (s). Similarly, the images of Φ0 (u) → Φ0 (s)
and Φ0 (w) → Φ0 (s) are orthogonal lines Φ0 (s), and their intersection is the image of fRu (p(x)) under Φ0 (v) → Φ0 (s). It follows that
Im(Φ0 (v) → Φ0 (s)) and Im(Φ0 (u) → Φ0 (s)) are parallel, hence Φ(v) and
Φ(u), considered as convex subcomplexes of Y 0 , are parallel. Moreover,
since image of fRv (x) under Φ0 (v) → Φ0 (s) and the image of fRu (p(x))
under Φ0 (v) → Φ0 (s) are in the line Im(Φ0 (w) → Φ0 (s)), the diagram in
(2) commutes.
Pick a vertex u ∈ |B| of rank = k and let Ru be the corresponding
J-residue with J = ∪ki=1 ji . Then there is a map fRu : Ru → Φ0 (Ru ) =
Φ(u) defined by considering Φ0 (x) → Φ0 (Ru ) for each element x ∈ Ru .
This map coincides with the fRu defined before when u is rank 1. For
1 ≤ i ≤ k, let Ri be a ji -residue in Ru . Since Φ is 1-determined,
{Im(Φ0 (Ri ) → Φ0 (Ru ))}ki=1 are mutually orthogonal lines in Φ0 (Ru ).
This induces an isomorphism
(5.5)
i:
k
Y
Φ0 (Ri ) → Φ0 (Ru ).
i=1
The following is a consequence of (2) of Lemma 5.4.
Q
Corollary 5.6. The map fRu satisfies fRu = i ◦ ( ki=1 fRi ) ◦ g, where
Q
g : Ru → ki=1 Ri is the map in Theorem 3.13. In this case, we will
Q
write fRu = ki=1 fRi for simplicity.
36
JINGYIN HUANG AND BRUCE KLEINER
Pick Q
J 0 ⊂ J andQ
let R0u be a J 0 -residue in Ru . By Theorem 3.13,
= i∈J 0 Ri × i∈J
/ 0 {ai } for ai ∈ Ri . Then the following is a
consequence of Corollary 5.6 and the functorality of Φ0 .
R0u
CorollaryQ5.7. Let h be the morphism between Φ0 (Ru0 ) and Φ0 (Ru ) =
0
0
0
Φ
/ 0 Φ (Ri ). Then for x ∈ Φ (Ru0 ), we have h(x) = {x} ×
Q(Ru0 ) × i∈J
i∈J
/ 0 {fRi (ai )}.
5.3. Construction of the Z-blow-up. In the previous section, we
started from a restriction quotient q : Y 0 → |B|, and produced associated 1-data (Definition 5.3), which is compatible with parallelism in
the sense of Lemma 5.4. In this section, we will consider the inverse,
namely we want construct a restriction quotient from this data.
Let ΛB be the collection of parallel sets of i-residues in B (i could be
any element in I). There is another type map T which maps a spherical
J-residue R to {λ
Q ∈ ΛB | λ contains a representative in R}. In other
words, let R ∼
= i∈I Ri be the product decomposition as in Theorem
3.13, where each Ri is an i-residue in R (i ∈ J). Then T (R) is the
collection of parallel sets represented by those Ri ’s. Let ZT (R) be the
collection of maps from T (R) to Z, and let Z∅ be a single point.
Our goal in this section is to construct a restriction quotient from
the following data.
Definition 5.8 (Blow-up data). For each i-residue R ⊂ B, we associate a map hR : R → ZT (R) such that if two i-residues R1 and
R2 are parallel, let h12 : R1 → R2 be the parallelism map, then
hR1 = hR2 ◦ h12 .
∼
Q If R is a spherical residue with product decomposition given by TR(R)=
.
i∈I Ri , then the maps hRi : Ri → Z induces a map hR : R → Z
It follows from the definition of hR , and the discussion after Definition
3.11 that if R, R0 ∈ C are parallel and let h : R → R0 be the parallelism
map, then hR = hR0 ◦ h.
The following result is a consequence of Theorem 3.13:
Qn
Lemma 5.9. Let T ∈ C be an H-residue. Let g : T ∼
= i=1 Ti be
the product decomposition
induced by H = tni=1 Hi (see Theorem 3.13).
Qn
Then hT = ( i=1 hTi ) ◦ g.
Q
simplify notation, we will write hT = ni=1 hTi instead of hT =
QTo
( ni=1 hTi ) ◦ g.
Q
Let J and R =Q i∈J Ri be Q
as before. A J 0 -residue R0 ⊂ R can
be expressed as ( i∈J 0 Ri ) × ( i∈J\J 0 {ci }), here ci is a chamber in
GROUPS QUASI-ISOMETRIC TO RAAG’S
37
0
T (R )
T (R)
Ri . We
by hR0 R (a) =
Q define an inclusion hR0 R : Z
Q→ Z
0
{a} × i∈J\J 0 {hRi (ci )}. Since hR = hR × ( i∈J\J 0 hRi ), hR0 R fits into
the following commutative diagram:
R0
hR0 y
−−−→
0
h
R
hR y
0
R
ZT (R ) −−R−→
ZT (R)
00
SupposeQR00 is a J 00 -residue such
R0 ⊂ R. Since hR =
Q that R ⊂ Q
hR0 × ( i∈J\J 0 hRi ) = hR00 × ( i∈J 0 \J 00 hRi ) × ( i∈J\J 0 hRi ), we have
(5.10)
hR00 R = hR0 R ◦ hR00 R0 .
Now we define a contravariant functor Ψ : Face(|B|) → CCC as
follows. Let f be a face of |B| and let vf ∈ f be unique vertex which
has minimal rank among the vertices of f . Let Rf ⊂ B be the residue
associated with vf . We define Ψ(f ) = RT (Rf ) (R∅ is a single point),
here RT (Rf ) is endowed with the standard cubical structure and we
identify ZT (Rf ) with the 0-skeleton of RT (Rf ) .
An inclusion of faces f → f 0 induces an inclusion Rf 0 → Rf . We
define the morphism Ψ(f 0 ) → Ψ(f ) to be the embedding induce by
hRf 0 Rf : ZT (Rf 0 ) → ZT (Rf ) .
Lemma 5.11. Ψ is contravariant functor.
Proof. It is easy to check that passing from an inclusion of faces f → f 0
to Rf 0 → Rf is a functor. And it follows from (5.10) that passing from
Rf 0 → Rf to hRf 0 Rf : ZT (Rf 0 ) → ZT (Rf ) is a functor.
Lemma 5.12. Ψ is 1-determined.
Proof. Let σ ⊂ |B| be a face and pick a vertex v ∈ σ. Let {vi }ki=1 be
the vertices in σ that are adjacent to v along an edge ei . Let σ≤v be
the sub-cube of σ that is spanned by ei ’s such that vi ≥ v. We define
σ>v similarly (σ>v could be empty). Then σ = σ≤v × σ>v . Moreover,
v is the maximal vertex in σ≤v and the minimum vertex in σ>v . Note
that Ψ(ei ) → Ψ(v) is an isometry if vi > v. Thus it suffices to consider
the case where v is the maximal vertex of σ.
Let vm be the minimal vertex of σ. Note that Im(Ψ(σ) → Ψ(v)) ⊂
Im(Ψ(e) → Ψ(v)) is a cubical convex embedding of Euclidean subspaces, it suffices to show they have the same dimension. Let R(v) ⊂ C
be the residue corresponding to the vertex v. Note that T (R(vm )) =
∩ki=1 T (R(vi )) (T is the type map defined on the beginning of Section
∩ki=1
38
JINGYIN HUANG AND BRUCE KLEINER
5.3). Thus the dimension of ∩ki=1 Im(Ψ(e) → Ψ(v)) equals to the cardinality of T (R(vm )), which is the dimension of Im(Ψ(σ) → Ψ(v)).
Ψ is called the fiber functor associated with the blow-up data {hR },
and the restriction quotient q : Y → |B| which arises from the fiber
functor Ψ (see Theorem 4.15) is called the restriction quotient associated with the blow-up data {hR }. It is clear from the construction that
the 1-data of q (Definition 5.3) is the blow-up data {hR } (we naturally
identify ZT (R) ’s in the blow-up data with the 0-skeleton of the q-fibers
of rank 1 vertices in |B|). We summarize the above discussion in the
following theorem.
Theorem 5.13. Given the blow-up data {hR } as in Definition 5.8,
there exists a restriction quotient q : Y → |B| whose 1-data is the
blow-up data we start with.
Remark 5.14. Here we blow up the building B with respect to a collection of Z’s since we want to apply the construction for RAAG’s.
However, in other cases, it may be natural to blow up with respect to
other objects. Here is a variation. For each parallel class of rank 1
residues λ ∈ ΛB , we associate a CAT (0) cube complex Zλ . For each
rank 1 residue R in the class λ, we define a map hR which assigns each
element of R a convex subcomplex of Zλ . We require these {hR } to
be compatible with parallelism between rank 1 residues. Given this set
of blow-up data, we can repeat the previous construction to obtain a
restriction quotient over |B|. It is also possible to blow-up buildings
with respect to spaces more general than CAT (0) cube complexes. We
give a slightly different approach in Section 8.
Now we show that the construction in this section is indeed a converse
to Section 5.2 in the following sense. Let q : Y 0 → |B| be a restriction
quotient as in Section 5.2 and let Φ and Φ0 be the functors introduced
there. For each vertex v ∈ |B| of rank 1 and its associated residue
Rv , we pick an isometric embedding ηv : ZT (Rv ) → Φ(v) such that
its image is vertex set of Φ(v). We also require these ηv ’s respect
parallelism. More precisely, let u ∈ |B| be a vertex of rank 1 such that
Φ(v) and Φ(u) (understood as subcomplexes of Y 0 ) are parallel with
the parallelism map given by p : Φ(v) → Φ(u). Then p ◦ ηv = ηu (note
that T (Rv ) = T (Ru ) by Lemma 5.4).
Let Ψ be the functor constructed in this section from the blow-up
data {hRv = ηv−1 ◦ fRv : Rv → ZT (Rv ) }v∈|B| , here v ranges over all
vertices of rank 1 in |B|, Rv is the residue associated with v and fRv
is the map in Definition 5.3. Pick a face σ ∈ |B| and let u ∈ σ be
GROUPS QUASI-ISOMETRIC TO RAAG’S
39
the vertex of minimal rank. Let Ru be the Q
associated J-residue with
its product decomposition given by Ru = j∈J Rvj (vj ’s are rank 1
vertices ≤ u). Let ξσ : Ψ(σ) → Φ(σ) be the isometry induced by
Q
Q
T (Ru )
→ j∈J Φ(vj )
j∈J ηvj : Z
Q
and the product decomposition j∈J Φ(vj ) ∼
= Φ(u) ∼
= Φ(σ) which
comes from (5.5). The following is a consequence of Corollary 5.6,
Corollary 5.7 and the discussion in this section.
Corollary 5.15. The maps {ξσ }σ∈Face(|B|) induce a natural isomorphism between Φ and Ψ. Thus for any restriction quotient q : Y 0 → |B|
which satisfies the conclusion of Lemma 5.1, if q 0 is the restriction quotient whose blow-up data is the 1-data of q, then q 0 is equivalent to q
up to a natural isomorphism between their fiber functors.
Corollary 5.16. Let q : Y → |B| be a restriction quotient which satisfies the conclusion of Lemma 5.1. Let B ∼
= B1 × B2 be a product decomposition of the building B induced by the join decomposition Γ = Γ1 ◦ Γ2
of the defining graph of the associated right-angled Coxeter group. Then
there are two restriction quotients q1 : Y1 → |B1 | and q2 : Y2 → |B2 |
such that Y = Y1 × Y2 and q = q1 × q2 . Moreover, q1 and q2 also satisfy
the conclusion of Lemma 5.1.
Proof. By Corollary 5.15, we can assume q is the restriction quotient
associated with a set of blow-up data {hR }. For every B1 -slice in B, we
can restrict {hR } to B1 to obtain a blow-up data for B1 . This does not
depend on our choice of the B1 -slice, since the blow-up data respects
parallelism. We obtain a blow-up data for B2 in a similar way. It follows
from the above construction that the fiber functor associated with {hR }
is the product of the fiber functors associated the blow-up data on B1
and B2 . Thus this corollary is a consequence of Lemma 4.16.
5.4. More properties of the blow-up buildings. In this section,
we look at the restriction quotient q : Y → |B| associated with the
blow-up data {hR } as in Definition 5.8 (or equivalently, a restriction
quotient q : Y → |B| which satisfies the conclusion of Lemma 5.1) in
more detail, and record several basic properties of Y . A hurried reader
can go through Definition 5.17, then proceed directly to Section 5.5
and come back to this part later.
Definition 5.17. A vertex y ∈ Y is of rank k if p(y) is a vertex of rank
k. Thus q induces a bijection between rank 0 vertices in Y and rank
0 vertices in |B|. Since rank 0 vertices in |B| can be identified with
40
JINGYIN HUANG AND BRUCE KLEINER
chambers in B, q −1 induces a well-defined map q −1 : B → Y from the
set of chambers of B (or rank 0 vertices of |B|) to rank 0 vertices in Y .
Lemma 5.18. For any residue R ⊂ B, we view R as a building
and restrict the blow-up data over B to a blow-up data over R. Let
qR : YR → |R| be the associated restriction quotient. Then there exists an isometric embedding i : YR → Y which fits into the following
commutative diagram:
i
YR −−−→
qR y
Y
qy
i0
|R| −−−→ |B|
Moreover, i(YR ) = q −1 (i0 (|R|)).
The lemma is a direct consequence of the construction in Section 5.3.
Pick a vertex v ∈ |B|. The downward complex of v is the smallest
convex subcomplex of |B| which contains all vertices which are ≤ v.
If Rv is the residue associated with v, then the downward complex is
the image of the embedding |Rv | ,→ |B|. The next result follows from
Lemma 5.18 and Corollary 5.16.
Lemma 5.19.
Q Let Dv be the downward complex of a vertex v ∈ B and
let Rv = ki=1 Ri be the product decomposition of residue associated
with v. Then q −1 (Dv ) is isomorphic to the product of the mapping
hR
i
ZT (Ri ) → RT (Ri ) (1 ≤ i ≤ k).
cylinders of Ri −−→
Lemma 5.20.
T (R)
(1) If h−1
, then
R (x) is finite for any rank 1 residue R and x ∈ Z
Y is locally finite. If there is a uniform upper bound for the
cardinality of h−1
R (x), then Y is uniformly locally finite.
(2) If there exists D > 0 such that the image of each hR is Ddense in ZT (R) , then there exists D0 which depends on D and
the dimension of |B| such that the collection of inverse images
of rank 0 vertices in |B| is D0 -dense in Y .
Proof. We prove (1) first. Pick a vertex y ∈ Y . Let v = q(y). It suffices
to show the set of edges in |B| which contain v, and can be lifted to an
edge in Y that contains y, is finite. Since there are only finitely many
vertices in |B| which are ≥ v, it suffices to consider the edges of the
form vλ v with vλ < v. It follows from our assumption and Lemma 5.19
GROUPS QUASI-ISOMETRIC TO RAAG’S
41
that there are only finitely many such edges which have the required
lift. The proof of uniform local finiteness is similar.
To see (2), notice that ∪v∈|B| Ψ(v) is 1-dense in Y , here v ranges over
all vertices of |B|. It follows from Lemma 5.19 that every point in Ψ(v)
be can approximated by the inverse image of some rank 0 vertex up to
distance D0 .
Next we discuss the relation between Y and the exploded Salvetti
complex Se = Se (Γ) introduced in Section 5.1. Let Ψ be the fiber
functor associated with q : Y → |B|.
First we label each vertex v ∈ Y by a clique in Γ as follows. Recall
that q(v) is associated with a J-residue R ⊂ B, where J is the vertex
set of a clique in Γ. Thus we label v by this clique. We also label each
vertex of Se by a clique. Any vertex v ∈ Se is contained in a unique
standard torus. Recall that a standard torus arises from a clique in
Γ, thus we label v by this clique. Note some vertices of Y and Se
are labelled by the empty set. There is a unique label-preserving map
(0)
p : Y (0) → Se (Γ).
An edge in Y or Se is horizontal if the labels on its two endpoints are
different, otherwise, this edge is vertical. When Y = Xe , this definition
coincides with the one in Section 5.1. Moreover, horizontal (or vertical)
edges in Xe are lifts of horizontal (or vertical) edges in Se .
Horizontal edges in Y are exactly those ones whose dual hyperplanes
are mapped by q to hyperplanes in |B|, and the q-image of any vertical
edge is a point. Now we label each edge vertical edge of Y by vertices
in Γ as follows. Pick vertical edge e ⊂ Y and let v = q(e). Let R =
Qk
with
i=1 Ri be the product decomposition of the residue associated
Qk
v. There is a corresponding product decomposition Ψ(v) = i=1 `i ,
where `i is a line which is parallel to Ψ(vi ), here vi ∈ |B| is the vertex
associated with Ri , and we view Ψ(vi ) and Ψ(v) as subcomplexes of Y .
If e is in the `i -direction, then we label e by the type of Ri , which is a
vertex in Γ. A case study implies if two vertical edges are the opposite
sides of a 2-cube, then they have the same label. Hence all parallel
vertical edges have the same label. Now we label vertical edges in Se .
Recall that the map Se → S(Γ) induces a 1-1 correspondence between
vertical edges in Se and edges in S(Γ), and edges in S(Γ) are labelled
by vertices of Γ. This induces a labelling of vertical edges in Se .
We pick an orientation for each vertical edge in Se , and orient every
vertical edge in Y in the following way. A vertical line is a geodesic line
made of vertical edges. It is easy to see every vertical edge is contained
42
JINGYIN HUANG AND BRUCE KLEINER
in a vertical line. For two vertical `1 and `2 , if there exist edges ei ∈ `i
for i = 1, 2 such that they are parallel, then `1 and `2 are parallel.
To see this, it suffices to consider the case where e1 and e2 are the
opposite sides of a 2-cube, and this follows from a similar case study
as before. Now we pick an orientation for each parallel class of vertical
lines, and this induces well-define orientation on each vertical edge of
Y , moreover, this orientation respects parallelism of edges.
(0)
There is a unique way to extend p : Y (0) → Se (Γ) to p : Y (1) →
(1)
Se (Γ) such that p preserves the orientation and labelling of vertical
edges. One can further extend p to higher-dimensional cells as follows.
A cube σ ⊂ Y is of type (m, n) if σ is the product of m vertical edges
and n horizontal edges. We extend p according to the type:
(1) If σ is of type (m, 0), then we can define p on σ since the orientation of vertical edges in Y respects parallelism, and p preserves
labelling and orientation of vertical edges. In this case, p(σ) is
an m-dimensional standard torus.
(2) If σ is of type (0, n), then we can define p on σ since p preserves
labelling of vertices. In this case, p(σ) ∼
= [0, 1]n .
(3) If σ is of type (m, n), then we can define p on σ for similar
reasons as before. In this case, p(σ) ∼
= Tm × [0, 1]n .
Pick vertex y ∈ Y , then p induces a simplicial map between the
vertex links py : Lk(y, Y ) ∼
= Lk(p(y), Se ). The above case study implies
py is a combinatorial map, i.e. py maps each simplex isomorphically
onto its image.
Theorem 5.21. If each map hR in the blow-up data is a bijection,
then Y is isomorphic to Xe = Xe (Γ), which is the universal cover of
the exploded Salvetti complex Se = Se (Γ).
Proof. We prove the theorem by showing p : Y → Se is a covering
map. It suffices to show for each vertex y ∈ Y , the above map py is
an isomorphism. Suppose y is labelled by a clique ∆ ⊂ Γ. We look at
edges which contain y, which fall into three classes:
(1) vertical edges;
(2) horizontal edges whose other endpoints are labelled by cliques
in ∆;
(3) horizontal edges whose other endpoints are labelled by cliques
that contain ∆.
Note that there is a 1-1 correspondence between edges in (3) and cliques
which contain ∆ and have exactly one vertex not in ∆. For any clique
GROUPS QUASI-ISOMETRIC TO RAAG’S
43
∆0 ⊂ ∆ which contains all but one vertex of ∆, there exists a unique
edge in (2) such that its other endpoint is labelled by ∆0 , since if such
edge does not exist, then some hR will not be surjective; if there exists
more than one such edges, then some hR will not be injective. Thus
there is a 1-1 correspondence between horizontal edges which contains
y and horizontal edges which contains p(y). Hence py induces bijection
between the 0-skeletons. Moreover, edges in (3) are orthogonal to edges
in (1) and (2), so a case study implies if two edges at p(y) form the
corner of a 2-cube, then their lifts at y (if any exist) also form the
corner of a 2-cube. It follows that py induces isomorphism between the
1-skeletons. Since both Lk(y, Y ) and Lk(p(y), Se ) are flag complexes,
py is an isomorphism.
Remark 5.22. If each map hR is injective (or surjective), then p is
locally injective (or locally surjective).
Corollary 5.23. Let B1 = B1 (Γ) and B2 = B2 (Γ) be two right-angled
Γ-buildings with countably infinite rank 1 residues. Then they are isomorphic as buildings.
Proof. We pick a blow-up for B1 such that each map in the blow-up
data is a bijection. Let Y → |B1 | be the associated restriction quotient
and let p : Y → Se be the covering map as in Theorem 5.21. Note that p
sends vertical edges to vertical edges and horizontal edges to horizontal
edges, and p preserves the labelling of vertices and edges. So does the
lift p̃ : Y → Xe of p. Lemma 4.3 implies p̃ descends to a cubical
isomorphism |B1 | → |B|, where |B| is the building associated with
G(Γ). Since p̃ is label-preserving, this cubical isomorphism induces a
building isomorphism B1 → B. Similarly, we can obtain a building
isomorphism B2 → B. Hence the corollary follows.
Theorem 5.24. Suppose Γ does not admit a join decomposition Γ =
Γ1 ◦ Γ2 where that Γ1 is a discrete graph with more than one vertex.
If B is a Γ-building and q : Y → |B| is a restriction quotient with
blow-up data {hR }, then any automorphism α : Y → Y descends to an
automorphism α0 : |B| → |B|.
Proof. By Lemma 4.3, it suffices to show α preserves the rank (Definition 5.17) of vertices of Y . Let F (Γ) be the flag complex of Γ. Here
we change the label of each vertex in Y from some clique in Γ to the
associated simplex in F (Γ). Suppose y ∈ Y is vertex of rand k labelled by ∆. Then Lemma 5.19 and the proof of Theorem 5.21 imply
Lk(y, Y ) ∼
= K1 ∗ K2 ∗ · · · ∗ Kk ∗ Lk(∆, F (Γ)), where each Ki is discrete
with cardinality ≥ 2, and Lk(∆, F (Γ)) is understood to be F (Γ) when
44
JINGYIN HUANG AND BRUCE KLEINER
∆ = ∅. Note that {Ki }ki=1 comes from vertices adjacent to y of rank
≤ k, and Lk(∆, F (Γ)) comes from vertices adjacent to y of rank > k.
Thus α preserves the collection of rank 0 vertices.
Now we assume α preserves the collection of rank i vertices for i ≤
k − 1. A rank k vertex in Y is of type I if it is adjacent to a vertex of
rank k − 1, otherwise it is a vertex of type II. It is clear that α preserves
the collection of rank k vertices of type I. Before we deal with type II
vertices, we need the following claim. Suppose w ∈ Y is a vertex of
rank k such that α(w) is also of rank k. If there exist k vertices {zi }ki=1
adjacent to w such that
(1) rank(zi ) ≤ k and rank(α(zi )) ≤ k;
(2) the edges {zi w}ki=1 are mutually orthogonal,
then rank(α(z)) ≤ k for any z adjacent to w with rank(z) ≤ k.
Let w0 = α(w). Suppose w and w0 are labelled by ∆ and ∆0 . Then
α induces an isomorphism between the links of w and w0 in Y :
α∗ : K1 ∗ · · · ∗ Kk ∗ Lk(∆, F (Γ)) → K10 ∗ · · · ∗ Kk0 ∗ Lk(∆0 , F (Γ)).
Each edge zi w gives rise to a vertex in Ki , and each edge α(zi )w0 gives
rise to a vertex in Ki0 . Thus α∗ (K1 ∗ · · · ∗ Kk ) = K10 ∗ · · · ∗ Kk0 . Since the
edge zw gives rise to a vertex in K1 ∗ · · · ∗ Kk , the edge α(z)w0 gives
rise to a vertex in K10 ∗ · · · ∗ Kk0 . Then α(z) is of rank ≤ k.
Let y ∈ Y be a rank k vertex of type II. Then there exists an edge
path ω from y to a type I vertex y1 such that every vertex in ω is of
rank k. Let {yi }m
i=1 be consecutive vertices in ω such that ym = y. Note
that there are k vertices of rank k − 1 adjacent to y1 . By the induction
assumption, they are send to vertices of rank k − 1 by α. Moreover,
rank(α(y1 )) = k since y1 is of type I. Thus the assumption of the claim
is satisfied for y1 . Then rank(α(y2 )) ≤ k, hence rank(α(y2 )) = k by
the induction assumption. Next we show y2 satisfies the assumption
of the claim. Let {zi }ki=1 be vertices of rank k such that they are
adjacent to y1 and {zi y1 }ki=1 are mutually orthogonal. We also assume
y2 = z1 . Then rank(α(zi )) = k for all i. Hence all α(zi y1 )’s are
vertical edges. For i ≥ 2, let zi0 be the vertex adjacent to y2 such that
zi0 y2 and zi y1 are parallel. Then α(zi0 y2 ) is a vertical edge for i ≥ 2.
Thus rank(α(zi0 )) = k and the assumption of the claim is satisfies for
y2 . We can repeat this argument finite many times to deduce that
rank(α(y)) = k.
GROUPS QUASI-ISOMETRIC TO RAAG’S
45
Remark 5.25. If the assumption on Γ in Theorem 5.24 is not satisfied,
then there exists a blow-up Y → |B| and an automorphism of Y such
that it does not descend to an automorphism of |B|. By Corollary 5.16,
it suffices to construct an example in the case when Γ be a discrete
graph with n vertices with n ≥ 2. If n ≥ 3, then we define each hR to
be a surjective map such that the inverse image of each point has n − 2
points. Then Y is a tree with valence = n. If n = 2, then we define hR
to be an injective map whose image is the set of even integers. Then
Y is isomorphic to the first subdivision of a tree of valence 3. In both
case, it is not hard to find an automorphism of Y which maps some
vertex of rank 0 to a vertex of rank 1.
5.5. Morphisms between blow-up data. Let B and B 0 be two buildings modelled on the same right-angled Coxeter group W (Γ). An isomorphism η : |B| → |B 0 | is rank-preserving if for each vertex v ∈ |B|,
v and η(v) have the same rank. Note that such η induces a bijection
η 0 : B → B 0 which preserves the spherical residues. Conversely, every bijection B → B 0 which preserves the spherical residues induces a
rank-preserving isomorphism |B| → |B 0 |. Note that η 0 maps parallel
residues of rank 1 to parallel residues of rank 1, thus η 0 induces a bijection η̄ : ΛB → ΛB0 , where ΛB and ΛB0 denote the collection of parallel
classes of residues of rank 1 in B and B 0 respectively (see Section 5.3).
Definition 5.26 (η-isomorphism). Suppose the blow-up data (Definition 5.8) of |B| and |B 0 | are given by {hR } and {h0R } respectively. An
η-isomorphism between the blow-up data is defined to be a collection
of isometries {fλ : Zλ → Zη̄(λ) }λ∈ΛB such that the following diagram
commutes for every rank 1 residue R ⊂ B:
R
η0 y
h
R
−−−
→
ZT (R)
fT (R) y
h0η0 (R)
η 0 (R) −−−→ Zη̄(T (R))
Here T is the type map defined in the beginning of Section 5.3. The
map hR is nondegenerate if its image contains more than one point. In
this case, if fT (R) exists, then it is unique. If hR is degenerate, then we
have two choices for fT (R) .
Let η1 : |B1 | → |B2 |, η2 : |B2 | → |B3 | and η : |B1 | → |B3 | be rankpreserving isomorphisms such that η = η2 ◦ η1 . We fix a blow-up data
for each Bi . Let {fλ : Zλ → Zη̄1 (λ) }λ∈ΛB1 and {gλ : Zλ → Zη̄2 (λ) }λ∈ΛB2
be the η1 -isomorphism and η2 -isomorphism between the corresponding
46
JINGYIN HUANG AND BRUCE KLEINER
blow-up data. We define the composition of them to be {gη̄1 (λ) ◦fλ }λ∈Λ ,
which turns out to be an η-isomorphism.
Let Ψ and Ψ0 be the fiber functor associated with the blow-up data
{hR } and {h0R } respectively, and let Y → |B| and Y 0 → |B 0 | be the
associated restriction quotient.
Lemma 5.27. Every η-isomorphism induces a natural isomorphism
from Ψ to Ψ0 , hence by Section 4.3, it induces an isomorphism Y →
Y 0 which is a lift of η : |B| → |B 0 |. Moreover, composition of ηisomorphisms gives rise to composition of natural transformations of
the associated fiber functors.
Proof. For every spherical residue R ⊂ B, η 0 respects the product decomposition of R. Thus the following diagram commutes:
R
η0 y
0
h
R
−−−
→
ZT (R)
Q
f
λ
λ∈T (R)
y
h0η0 (R)
η (R) −−−→ Zη̄(T (R))
Q
Here λ∈T (R) fλ induces an isometry RT (R) → Rη̄(T (R)) . This gives rise
to a collection of isometries between objects of Ψ and Ψ0 . It follows from
the construction in Section 5.3 that these isometries give the required
natural isomorphism between Ψ and Ψ0 . The second assertion in the
lemma is straightforward.
Remark 5.28. If we weaken the assumption of Definition 5.26 by assuming each fλ is a bijection, then we can obtain a bijection between the
vertex sets of Y and Y 0 . This bijection preserves the fibers, however,
we may not be able to extend it to a cubical map.
Theorem 5.29. If each map hR in the blow-up data is a bijection,
then Y is isomorphic to Xe = Xe (Γ), which is the universal cover of
the exploded Salvetti complex.
Definition 5.30 (η-quasi-morphism). We follow the notation in Definition 5.26. An (η, L, A)-quasi-morphism between the blow-up data
{hR } and {h0R } is a collection of (L, A)-quasi-isometries {fλ : Zλ →
Zη̄(λ) }λ∈ΛB such that the diagram in Definition 5.26 commutes up to
error A.
Lemma 5.31. Each (η, L, A)-quasi-morphism between {hR } and {h0R }
induces an (L0 , A0 )-quasi-isometry Y → Y 0 with L0 , A0 depending on
L, A, and the dimension of |B|.
GROUPS QUASI-ISOMETRIC TO RAAG’S
47
Proof. By Lemma 4.19, it suffices to produce an (L0 , A0 )-quasi-natural
isomorphism
from Ψ to Ψ0 . This can be done by considering maps of
Q
form λ∈T (R) fλ as in Lemma 5.27.
Remark 5.32 (A nice representative). Let Y0 be the collection of rank
0 vertices in Y (Definition 5.17). We define Y00 similarly. If the assumption in Lemma 5.20 (2) is satisfied, then Y0 and Y00 are D-dense
in Y and Y 0 respectively. In this case, the quasi-isometry Y → Y 0 in
Lemma 5.31 can be represented by φ : Y0 → Y00 , where φ is the bijection induced by η : |B| → |B 0 | (recall that we can identify Y0 and Y00
with rank 0 vertices in |B| and |B 0 | respectively, see Definition 5.17).
The fact that φ is a quasi-isometry follows from the construction in the
proof Lemma 4.19.
Corollary 5.33. If there exists constant D > 0 such that each map
hR in the blow-up data satisfies:
(1) For any x ∈ ZT (R) , |h−1
R (x)| ≤ D.
(2) The image of hR is D-dense in ZT (R) .
Then Y is quasi-isometric to G(Γ).
Proof. By the assumptions, there exists another set of blow-up data
{h0R } such that each h0R is a bijection, and an (η, L, A)-quasi-isomorphism
{fλ }λ∈ΛB from {h0R } to {hR } where η is the identity map. It follows
from Lemma 5.31 and Theorem 5.21 that Y is quasi-isometric to Xe ,
the universal cover of the exploded Salvetti complex; hence Y is quasiisometric to G(Γ).
In the rest of this section, we look at the special case when B = B(Γ)
is the Davis building of G(Γ) (see the beginning of Section 5.1), and
record an observation for later use. In this case, we identify points in
G(Γ) with chambers in B.
We denote the word metric on G(Γ) by dw . If we identify G(Γ) with
chambers of the building B = B(Γ), then there is another metric on
G(Γ) defined in Definition 3.7. We caution the reader that these two
metrics are not the same. We pick a set of blow-up data {hR } on B
and let q : Y → |B| be the associated restriction quotient. Recall that
vertices of rank 0 on |B| can be identified with chambers in B, hence can
be identified with G(Γ). Thus the map q −1 : G(Γ) → Y is well-defined.
Lemma 5.34. If there exist L, A > 0 such that all {hR : R → ZT (R) }
are (L, A)-quasi-isometries (here we identify chambers in R with a subset of G(Γ), hence R is endowed with an induced metric from dw ), then
48
JINGYIN HUANG AND BRUCE KLEINER
q −1 : (G(Γ), dw ) → Y is an (L0 , A0 )-quasi-isometry with its constants
depending on L, A and Γ.
Proof. Let q 0 : Xe → |B| be the G(Γ)-equivariant canonical restriction
quotient constructed in Section 5.1. In this case, (q 0 )−1 : G(Γ) → Xe is
a quasi-isometry whose constants depend on Γ. Let h0R be the blow-up
data which arises from the 1-data (Definition 5.3) of q 0 . Then each
h0R is an isometry. It follows from the assumption that there exists an
(η, L, A)-quasi-isomorphism from the blow-up data {h0R } to {hR } with
η being the identity map. Thus there exists a quasi-isometry Xe → Y
which can be represented by a map φ of the form in Remark 5.32. Since
q −1 = φ ◦ (q 0 )−1 , the lemma follows.
5.6. An equivariant construction. Let B = B(Γ) be a right-angled
building. Let K be group which acts on |B| by automorphisms which
preserve the rank of its vertices and let K y B and K y ΛB be the
induced actions (ΛB is defined in the beginning of Section 5.3).
Definition 5.35 (Factor actions). Pick λ ∈ Λ and let Rλ ⊂ B be a
residue of rank 1 such that T (Rλ ) = λ (T is the type map defined in
Section 5.3). Let Kλ be the stabilizer of λ with respect to the action
K y ΛB and let P (Rλ ) = Rλ × R⊥
λ be the parallel set of Rλ with
its product decomposition (see Lemma 3.14 and Theorem 3.13). Then
P (Rλ ) is Kλ -invariant, and Kλ respects the product decomposition of
P (Rλ ). Let ρλ : Kλ y Rλ be the action of Kλ on the Rλ -factor. This
action ρλ is called a factor action.
We construct equivariant blow-up data as follows. Pick one representative from each K-orbit of K y ΛB and form the set {λu }u∈U . Let
Ku be the stabilizer of λu . Pick residue Ru ⊂ B of rank 1 such that
T (Ru ) = λu and let ρu : Ku y Ru be the factor action defined as
above.
To obtain a K-equivariant blow-up data, we pick an isometric action
Ku y Zλu and a Ku -equivariant map hRu : Ru → Zλu . If R is parallel
to Ru with the parallelism map given by p : R → Ru , we define
hR = hRu ◦ p. By the previous discussion, there is a factor action
Ku y R, and hR is Ku -equivariant. We run this process for each
element in {λu }u∈U . If λ ∈
/ {λu }u∈U , then we fix an element gλ ∈ K
such that gλ (λ) ∈ {λu }u∈U . For rank 1 element R with T (R) = λ,
we define hR = Id ◦ hgλ (R) ◦ gλ , here Id : Zgλ (λ) → Zλ is the identity
map. Let Kλ = gλ−1 Kgλ (λ) gλ be the stabilizer of λ. We define the action
GROUPS QUASI-ISOMETRIC TO RAAG’S
49
Kλ y Zλ by letting gλ−1 ggλ acts on Zλ by Id ◦ g ◦ Id−1 (g ∈ Kgλ (λ) ).
Then hR becomes Kλ -equivariant.
It follows from the above construction that we can produce an isometry fg,R : ZT (R) → ZT (g(R)) for each g ∈ K and rank 1 residue R ∈ B
such that the following diagram commutes
R
gy
h
R
−−−
→ ZT (R)
fg,R y
hg(R)
g(R) −−−→ ZT (g(R))
and fg1 g2 ,R = fg1 ,g2 (R) ◦ fg2 ,R for any g1 , g2 ∈ K. Let Ψ be the fiber
functor associated with the above blow-up data and let q : Y → |B| be
the corresponding restriction quotient. Lemma 5.27 implies K acts on
Ψ by natural transformations, hence there is an induced action K y Y
and q is K-equivariant.
Remark 5.36. The previous construction depends on several choices:
(1) The choice of the set {λu }u∈U .
(2) The choice of the isometric action Ku y Zλu and the Ku equivariant map hRu : Ru → Zλu .
(3) The choice of the elements gλ ’s.
6. Quasi-actions on RAAG’s
In this section we will apply the construction in Section 5.6 to study
quasi-actions on RAAG’s.
We assume G(Γ) 6= Z throughout Section 6.
6.1. The cubulation. Throughout Subsection 6.1 we assume G(Γ) 6'
Z is a RAAG with | Out(G(Γ))| < ∞, and ρ : H y G(Γ) is an (L, A)quasi-action.
Recall that G(Γ) acts on X(Γ) by deck transformations, and this
action is simply transitive on the vertex set of X(Γ). By picking a base
point in X(Γ), we identify G(Γ) with the 0-skeleton of X(Γ).
Definition 6.1. A quasi-isometry φ : G(Γ) → G(Γ) is flat-preserving if
it is a bijection and for every standard flat F ⊂ X(Γ) there is a standard
flat F 0 ⊂ X(Γ) such that φ maps the 0-skeleton of F bijectively onto
the 0-skeleton of F 0 . The standard flat F 0 is uniquely determined, and
we denote it by φ∗ (F ). Note that if φ is flat-preserving, then φ−1 is
also flat-preserving.
50
JINGYIN HUANG AND BRUCE KLEINER
By Theorem 1.3, without loss of generality we can assume ρ : H y
G(Γ) is an action by flat-preserving bijections which are also (L, A)quasi-isometries.
On the one hand, we want to think G(Γ) as a metric space with
the word metric with respect to its standard generating set, or equivalently, with the induced l1 -metric from X(Γ); on the other hand, we
want to treat G(Γ) as a right-angled building (see Section 5.1), more
precisely, we want to identify points in G(Γ) with chambers in the associated right-angled building of G(Γ). Then the ρ preserves the spherical
residues in G(Γ), thus there is an induced ρ|B| : H y |B| on the Davis
realization |B| of G(Γ).
Let Λ be the collection of parallel classes of standard geodesic lines
in X(Γ), in other words, Λ is the collection of parallel classes of rank 1
residues in G(Γ), and let T be the type map defined in the beginning
of Section 5.3. There is another induced action ρΛ : H y Λ. For each
λ ∈ Λ, let Hλ be the stabilizer of λ. Pick a residue R in the parallel
class λ, and let ρλ : Hλ y R be the factor action in Definition 5.35.
Note that R is an isometrically embedded copy of Z with respect to the
metric on G(Γ); moreover, ρλ is an action by (L0 , A0 )-quasi-isometries.
Here we can choose L0 and A0 such that they depend only on L and A,
so in particular they do not depend on λ and R.
For the action ρΛ : H y Λ, we pick a representative from each Horbit and form the set {λu }u∈U . By the construction in Section 5.6, it
remains to choose an isometric action Gu y Zλu and a Gu -equivariant
map hRu : Ru → Zλu for each u ∈ U (Ru is a residue in the parallel
class λu ). The choice is provided by the following result, whose proof
is postponed to Section 7.
Proposition 6.2. If a group K has an action on Z by (L, A)-quasiisometries, then there exists another action K y Z by isometries which
is related to the original action by a surjective equivariant (L0 , A0 )quasi-isometry f : Z → Z where L0 , A0 depend on L and A.
From the above data, we produce H-equivariant blow-up data hR :
R → ZT (R) for each rank 1 residue R ⊂ G(Γ) as in Section 5.6. Note
that each hR is an (L00 , A00 )-quasi-isometry with constants depending
only on L and A.
Let q : Y → |B| be the restriction quotient associated with the
above blow-up data. Then there is an induced action H y Y by
isomorphisms, and q is H-equivariant. It follows from Lemma 5.20
that Y is uniformly locally finite.
GROUPS QUASI-ISOMETRIC TO RAAG’S
51
Claim. There exists an (L1 , A1 )-quasi-isometry G(Γ) → Y with
L1 , A1 depending only on L and A.
Proof of claim. Let h0R : R → ZT (R) be another blow-up data such
that each h0R is an isometry (such blow-up data always exists), and let
q 0 : Y 0 → |B| be the associated restriction quotient. By Theorem 5.21,
Y 0 is isomorphic to Xe , which is the universal cover of the exploded
Salvetti complex introduced in Section 5.1. For any λ ∈ Λ, we define
fλ = hR ◦ (h0R )−1 , here R is a residue such that T (R) = λ and the
definition of fλ does not depend on R. Each fλ is an (L00 , A00 )-quasiisometry and the collection of all fλ ’s induces a quasi-isomorphism
between the blow-up data {h0R } and {hR }. It follows from Lemma 5.31
that there exists a quasi-isometry between ϕ : Y 0 ∼
= Xe → Y , and the
claim follows.
Let B0 be the set of vertices of rank 0 in |B|. There is a natural
identification of B0 with G(Γ). Letting Y0 = q −1 (B0 ), we get that
q induces a bijection between Y0 and B0 . We define Y00 similarly. It
follows from (2) of Lemma 5.20 that Y00 and Y0 are D-dense in Y 0 and Y
respectively for D depending on L and A. Note that q −1 : G(Γ) → Y0
is H-equivariant, and if the action ρ : H y G(Γ) is cobounded, then
H y Y is cocompact.
The above quasi-isometry ϕ can be represented by q −1 ◦ q 0 : Y00 → Y0
(Remark 5.32). By Lemma 5.34, (q 0 )−1 : B0 = G(Γ) → Y00 is also a
quasi-isometry, and thus q −1 : G(Γ) → Y0 is a quasi-isometry. This
map is H-equivariant, so if ρ : H y G(Γ) is proper, then H y Y is
also proper.
Remark 6.3. Here we discuss a refinement of the above construction.
Instead of requiring each h0R to be an isometry, it is possible to choose
each h0R such that:
(1) h0R is a bijection.
(2) h0R is an (L2 , A2 )-quasi-isometry.
(3) fλ : Zλ → Zλ is a surjective map which respects the order of the
Z, hence can be extended to a surjective cubical map Rλ → Rλ .
The surjectivity in (3) follows from our choice in Proposition 6.2. In
this case, the space Y 0 is still isomorphic to Xe (Theorem 5.21). Let
Ψ and Ψ0 be the fiber functors associated with the blow-up data {hR }
and {h0R }. As in the proof of Lemma 5.27, the fλ ’s induce a natural transformation from Ψ0 to Ψ which is made of a collection of
surjective cubical maps from objects in Ψ0 to objects in Ψ; moreover,
52
JINGYIN HUANG AND BRUCE KLEINER
these maps are quasi-isometries with uniform quasi-isometry constants.
Recall that we can describe Y as the quotient of the disjoint collection {σ × Ψ(σ)}σ∈Face(|B|) (see the proof of Theorem 4.15), and a similar description holds for Y 0 . Thus there is a surjective cubical map
φ : Y 0 → Y induced by the natural transformation. Actually φ is a
restriction quotient, since the inverse image of each hyperplane is a
hyperplane. We also know φ is a quasi-isometry by Lemma 4.19.
The following theorem summarizes the above discussion.
Theorem 6.4. If the outer automorphism group Out(G(Γ)) is finite
and G(Γ) 6' Z, then any quasi-action ρ : H y X(Γ) is quasiconjugate
to an action ρ̂ of H by cubical isometries on a uniformly locally finite
CAT (0) cube complex Y . Moreover:
(1) If ρ is cobounded, then ρ̂ is cocompact.
(2) If ρ is proper, then ρ̂ is proper.
(3) Let |B| be the Davis realization of the right-angled building associated with G(Γ), let H y |B| be the induced action, and
let Xe = Xe (Γ) be the universal cover of the exploded Salvetti
complex for G(Γ). Then Y fits into the following commutative
diagram:
φ
Xe
<
q
>
q0
>Y
|B|
Here q 0 , q and φ are restriction quotients. The map φ is a
quasi-isometry whose constants depend on the constants of the
quasi-action ρ, and q is H-equivariant.
Corollary 6.5. Suppose the outer automorphism group Out(G(Γ)) is
finite. Then H is quasi-isometric to G(Γ) if and only if there exists an
H-equivariant restriction quotient map q : Y → |B| such that:
(1) |B| is the Davis realization of some right-angled Γ-building.
(2) The action H y Y is geometric.
(3) If v ∈ |B| is a vertex of rank k, then q −1 (v) = Ek .
Proof. The only if direction follows from Theorem 6.4. For the if direction, it suffices to show Y is quasi-isometric to G(Γ). Let Φ be the
fiber functor associated with q.
GROUPS QUASI-ISOMETRIC TO RAAG’S
53
Pick a vertex v ∈ |B| of rank k and let Fv = q −1 (v). We claim Stab(v)
acts cocompactly on Fv . By a standard argument, to prove this it suffices to show that {h(Fv )}h∈H is a locally finite family in Y . Suppose
there exists an R-ball N ⊂ Y such that there are infinitely many distinct elements in {h(Fv )}h∈H which have nontrivial intersection of N .
Since Y admits a geometric action, it is locally finite, and thus there
exists a vertex x ∈ |B| which is contained in infinitely many distinct elements in {h(Fv )}h∈H . This is impossible, since if h(Fv ) 6= h0 (Fv ), then
q(h(Fv )) and q(h0 (Fv )) are distinct vertices in |B| by the H-equivariance
of q.
Consider a cube σ ⊂ |B| and let v be its vertex of minimal rank.
We claim Φ(σ) → Φ(v) is surjective, hence is an isometry. By (3),
the action H y |B| preserves the rank of the vertices, thus Stab(σ) ⊂
Stab(v). We know that Stab(v) acts cocompactly on q −1 (v); since the
poset {w ≥ v} is finite, Stab(σ) has finite index in Stab(v), and so
Stab(σ) also acts cocompactly on q −1 (v). Now the image of Φ(σ) →
Φ(v) is a convex subcomplex of q −1 (v) that is Stab(σ)-invariant, so it
coincides with q −1 (v).
By Corollary 5.15, we can assume q is the restriction quotient of a
set of blow-up data {hR }. Pick a vertex v ∈ |B| of rank 1 and let D(v)
be the downward complex of v (see Section 5.3). Let Rv ⊂ B be the
associated residue and let Rv → RT (R) be the map induced by hRv .
Then q −1 (Dv ) is isomorphic to the mapping cylinder of this map. Since
the Stab(v) acts cocompactly on q −1 (Dv ), there are only finite many
orbits of vertices of rank 1, and the assumptions of Corollary 5.33 are
satisfied. It follows that Y is quasi-isometric to G(Γ).
It is possible to drop the H-equivariant assumption on q under the
following conditions. Here we do not put any assumption on Γ.
Theorem 6.6. Let B be a right-angled Γ-building. Suppose q : Y → |B|
is a restriction quotient such that for every cube σ ⊂ |B|, and every
interior point x ∈ σ, the point inverse q −1 (x) is a copy of Erank(v) ,
where v ∈ σ is the vertex of minimal rank.
If H acts geometrically on Y by automorphisms, then H is quasiisometric to G(Γ).
Proof. First we assume Γ satisfies the assumption of Theorem 5.24.
Then the above result is a consequence of Corollary 5.15, Theorem 5.24
and the argument in Corollary 6.5.
54
JINGYIN HUANG AND BRUCE KLEINER
For arbitrary Γ, we make a join decomposition Γ = Γ1 ◦Γ2 ◦· · · Γk ◦Γ0
where Γ0 satisfies the assumption of Theorem 5.24, and all Γi ’s are
discrete graphs with more than one vertex. By Corollary 5.16, there
are induced cubical product decomposition Y = Y1 × Y2 × · · · Yk × Y 0
and restriction quotients qi : Yi → |Bi | and q 0 : Y 0 → |B 0 | which
satisfy the assumption of the theorem. By [CS11, Proposition 2.6], we
assume H respects this product decomposition by passing to a finite
index subgroup. Since Y 0 is locally finite and cocompact, the same
argument in Corollary 6.5 implies Y 0 is quasi-isometric to G(Γ0 ). Each
Yi is a locally finite and cocompact tree which is not quasi-isometric to
a line. So Yi is quasi-isometric to G(Γ0 ). Thus Y is quasi-isometric to
G(Γ).
Corollary 6.7. Suppose Out(G(Γ)) is finite and G(Γ) 6' Z. Let B be
the right-angled building of G(Γ). Then H is quasi-isometric to G(Γ)
if and only if H acts geometrically on a blow-up of B in the sense of
Section 5.3 by automorphisms.
6.2. Reduction to nicer actions. Though every action ρ : H y
G(Γ) by flat-preserving bijections which are also (L, A)-quasi-isometries
is quasiconjugate to an isometric action H y Y as in Theorem 6.4, it is
in general impossible to take Y = X(Γ), even if the action ρ is proper
and cobounded.
Definition 6.8. Let H = Z/2 ⊕ Z with the generator of Z/2 and Z
ρ0
denoted by a and b respectively. Let H y Z be the action where
ρ0 (b)(n) = n + 2, and ρ0 (a) acts on Z by flipping 2n and 2n + 1 for all
n ∈ Z. An action K y Z is 2-flipping if it factors through the action
ρ0
H y Z via an epimorphism K → H.
Lemma 6.9. Let ρK : K y Z be a 2-flipping action. Then ρK is not
conjugate to an action by isometries on Z (with respect to the word
metric on Z).
Proof. Suppose there exists a permutation p : Z → Z which conjugates
ρK to an isometric action. Let h : K → G be the epimorphism in
Definition 6.8. Pick k1 , k2 ∈ K such that h(k1 ) is of order 2 and
h(k2 ) is of order infinity. Then pk1 p−1 is a reflection of Z and pk2 p−1
is a translation. However, this is impossible since h(k1 ) and h(k2 )
commute.
Lemma 6.10. There does not exists an action ρK : K y Z by (L, A)quasi-isometries with the following property. K has two subgroup K1
GROUPS QUASI-ISOMETRIC TO RAAG’S
55
and K2 such that ρK |K1 is conjugate to a 2-flipping action and ρK |K2
is conjugate to a transitive action on Z by translations.
Proof. By Proposition 6.2, there exists an isometric action ρ0K : K y Z
ρK
ρ0K
and a K-equivariant surjective map f : K y Z −→ K y Z. We claim
f is also injective. Given this claim, we can deduce a contradiction to
Lemma 6.9 by restricting the action to K1 . To see the claim, we restrict
the action to K2 . Thus we can assume without loss of generality that
ρK is a transitive action by translations. Suppose f (a1 ) = f (a1 + k)
for a1 , k ∈ Z and k 6= 0. Then the equivariance of f implies f (a1 ) =
f (a1 + nk) for any integer n ∈ Z, which contradicts that f is a quasiisometry.
Theorem 6.11. Suppose G(Γ) is a RAAG with | Out(G(Γ))| < ∞
and G(Γ) 6' Z. Then there is a pair H, H 0 of finitely generated groups
quasi-isometric to G(Γ) that does not admit discrete, virtually faithful cocompact representations into the same locally compact topological
group.
Recall that a discrete, virtually faithful cocompact representation
from H to a locally compact group Ĝ is a homomorphism h : H → Ĝ
such that its kernel is finite, and its image is a cocompact lattice.
Proof. Pick a vertex u ∈ Γ and let Γ0 be a graph obtained by taking
two copies of Γ and gluing them along the closed star of u. There is
a graph automorphism α : Γ0 → Γ0 which fixes the closed star of u
pointwise and flips the two copies of Γ. Then α induces an involution
α : G(Γ0 ) → G(Γ0 ), which gives rise to a semi-product H = G(Γ0 )oZ/2.
Note that G(Γ0 ) is a subgroup of index 2 in both G(Γ) and H. Therefore this induces a quasi-isometry q : H → G(Γ), an also a quasi-action
ρH : H y G(Γ). By the previous discussion, we can assume ρH is an
action by flat-preserving quasi-isometries. We look at the associated
collection of factor actions {Hλ y Z}λ∈Λ (see Definition 5.35), recall
that Λ is the collection of parallel classes of rank 1 residues in G(Γ),
and a rank 1 residue in some class λ can be identified with a copy of
Z. Up to conjugacy by bijective quasi-isometries, these factor actions
are either transitive actions on Z or 2-flipping actions.
We claim that G(Γ) and H do not admit discrete, virtually faithful
cocompact representations into the same locally compact topological
group. Suppose such a topological group Ĝ exists. Then by [MSW03,
Chapter 6], Ĝ has a quasi-action on G(Γ). We assume Ĝ acts on G(Γ)
56
JINGYIN HUANG AND BRUCE KLEINER
by flat-preserving quasi-isometries as before. Then there are restriction
actions ρ0G(Γ) : G(Γ) y G(Γ) and ρ0H : H y G(Γ) which are discrete
and cobounded. Since any two discrete and cobounded quasi-actions
H y G(Γ) are quasi-conjugate, it follows from Theorem 1.3 that ρH
and ρ0H are conjugate by a flat-preserving quasi-isometry. Thus factor
actions of ρ0H is conjugate to factor actions of ρH by bijective quasiisometries. Similarly, we deduce that the factor actions of ρ0G(Γ) are
conjugate to transitive actions by left translations on Z via bijective
quasi-isometries. Note that the factor actions of ρ0G(Γ) and the factor
actions of ρ0H are both restrictions of factor actions of Ĝ y G(Γ),
however, this is impossible by Lemma 6.10.
Corollary 6.12. The group H = G(Γ0 ) o Z/2 cannot act geometrically
on X(Γ).
We now give a criterion for when one can quasi-conjugate a quasiaction on X(Γ) to an isometric action H y X(Γ).
Theorem 6.13. Let ρ : H y G(Γ) be an action by flat-preserving
bijections. If for each λ ∈ Λ, the factor action ρλ : Hλ y Z can be
conjugate to an action by isometries with respect to the word metric
of Z, then there is an flat-preserving bijection g : G(Γ) → G(Γ) which
conjugates ρ : H y G(Γ) to an action ρ0 : H y X(Γ) by flat-preserving
isometries. If ρ is also an action by (L, A)-quasi-isometries, then g can
be taken to be a quasi-isometry.
Proof. We repeat the construction in Section 6.1 and assume each hR :
R → ZT (R) is a bijection. Let q : Y → |B|, Y0 , q −1 : G(Γ) → Y0 and
the action ρ̂ : H y Y by automorphisms be as in Section 6.1. Recall
that q −1 is H-equivariant. There is an isomorphism i : Y → Xe by
Theorem 5.21, moreover, i(Y0 ) is exactly the collection of 0-dimensional
standard flats X0 in Xe . We deduce from the construction of i that
the isometric action H y Xe induced by ρ̂ preserves standard flats
in Xe . By the construction of Xe , there exists a natural identification
f : X0 → G(Γ) such that any automorphism of Xe which preserves its
standard flats induces a flat-preserving isometry of G(Γ) (with respect
to the word metric) via f . It suffices to take g = f ◦ i ◦ q −1 .
Suppose we have already conjugated the flat-preserving action ρ :
H y G(Γ) to an action ρ0 : H y X(Γ) (or H y G(Γ)) by flatpreserving isometries. We ask whether it is possible to further conjugate ρ0 to an action by left translations.
GROUPS QUASI-ISOMETRIC TO RAAG’S
57
We can oriented each 1-cell in S(Γ) and label it by the associated
generator. This lifts to orientations and labels of edges of X(Γ). If H
preserves this orientation and labelling, then ρ0 is already an action by
left translations. In general, it suffices to require H preserves a possibly
different orientation and labelling which satisfy several compatibility
conditions. Now we recall the following definitions from [Hua14a].
Definition 6.14 (Coherent ordering). A coherent ordering for G(Γ)
is a blow-up data for G(Γ) such that each map hR is a bijection. Two
coherent orderings are equivalent if the their maps agree up to translations.
Let P(Γ) be the extension complex defined in Section 3.3. Note that
we can identify ΛG(Γ) with the 0-skeleton of P(Γ). Any flat-preserving
action H y G(Γ) induces an action H y P(Γ) by simplicial isomorphisms. Let F (Γ) be the flag complex of Γ.
Definition 6.15 (Coherent labelling). Recall that for each vertex x ∈
X(Γ), there is a natural simplicial embedding ix : F (Γ) → P(Γ) by
considering the standard flats passing through x. A coherent labelling
of G(Γ) is a simplicial map L : P(Γ) → F (Γ) such that L ◦ ix : F (Γ) →
F (Γ) is a simplicial isomorphism for every vertex x ∈ X(Γ).
The next result follows from [Hua14a, Lemma 5.7].
Lemma 6.16. Let ρ0 : H y G(Γ) be an action by flat-preserving
bijections and let H y P(Γ) be the induced action. If there exists an
H-invariant coherent ordering and an H-invariant coherent labelling,
then ρ0 is conjugate to an action by left translations.
Since each vertex of P(Γ) corresponds to a parallel class of v-residues
for vertex v ∈ Γ, this gives a labelling of vertices of P(Γ) by vertices of
Γ. We can extend this labelling map to a simplicial map L : P(Γ) →
F (Γ), which gives rise to a coherent labelling.
Corollary 6.17. Let ρ : H y G(Γ) be an action by flat-preserving
bijections. Suppose:
(1) The induced action H y P(Γ) preserves the vertex labelling of
P(Γ) as above.
(2) For each vertex v ∈ P(Γ), the action ρv : Hv y Z is conjugate
to an action by translations.
Then ρ is conjugate to an action H y G(Γ) by left translations.
Note that condition (2) is equivalent to the existence of an H-invariant
coherent ordering.
58
JINGYIN HUANG AND BRUCE KLEINER
7. Actions by quasi-isometries on Z
In this section we prove Proposition 6.2.
7.1. Tracks. Tracks were introduced in [Dun85]. They are hypersurfacelike objects in 2-dimensional simplicial complexes.
Definition 7.1 (Tracks). Let K be 2-dimensional simplicial complex.
A track τ ⊂ K is a connected embedded finite simplicial graph such
that:
(1) For each 2-simplex ∆ ⊂ K, τ ∩ ∆ is a finite disjoint union of
curves such that the end points of each curve are in the interior
of edges of ∆.
(2) For each edge e ∈ K, τ ∩ e is a discrete set in the interior of e.
Let {∆λ }λ∈Λ be the collection of 2-simplices that contains e. If
v ∈ τ ∩ e, then for each λ, τ ∩ ∆λ contains a curve that contains
v.
Given a track τ ⊂ K, we defined the support of τ , denoted Spt(τ ),
to be the minimal subcomplex of K which contains τ .
We can view hyperplanes defined in Section 3.2 as analogue of tracks
in the cubical setting. Each track τ ⊂ K has a regular neighbourhood
which fibres over τ . When K is simply-connected, K \ τ has two
connected components, moreover, the regular neighbourhood of τ is
homeomorphic to τ × (−, ).
Two tracks τ1 and τ2 are parallel if Spt(τ1 ) = Spt(τ2 ) and there is
a region homeomorphic to τ1 × (0, ) bounded by τ1 and τ2 . A track
τ ⊂ K is essential if the components of K \ τ are unbounded. The
following result follows from [Dun85, Proposition 3.1]:
Lemma 7.2. If K is simply-connected and has more than one end,
then there exists an essential track τ ⊂ K.
Next we look at essential tracks which are “minimal”; these turn out
to behave like minimal surfaces. First we metrize K as in [SS96].
Let ∆ = ∆(ξ1 ξ2 ξ3 ) be an ideal triangle in the hyperbolic plane. We
mark a point in each side of the triangle as follows. Let φ be the unique
isometry which fixes ξ3 and flips ξ1 and ξ2 , we mark the unique point
in ξ1 ξ2 which is fixed by φ. Other sides of ∆ are marked similarly. This
is called a marked ideal triangle.
We identify each 2-simplex of K with a marked ideal triangle in the
hyperbolic plane and glue these triangles by isometries which identify
GROUPS QUASI-ISOMETRIC TO RAAG’S
59
the marked points. This gives a collection of complete metrics on each
connected component of K − K (0) which is not an interval. We denote
this collection of metrics by dH . If a group G acts on K by simplicial
isomorphisms, then G also acts by isometries on (K, dH ). The original
definition in [SS96] does not required these marked points, see Remark
7.4 for why we add them.
Each track τ has a well-defined length under dH , which we denote by
l(τ ). We also define the weight of τ , denoted by w(τ ), to be cardinality
of τ ∩ K (1) . The complexity c(τ ) is defined to be the ordered pair
(w(τ ), l(τ )). We order the complexity lexicographically, namely c(τ1 ) <
c(τ2 ) if and only if w(τ1 ) < w(τ2 ) or w(τ1 ) = w(τ2 ) and l(τ1 ) < l(τ2 ).
The following result follows from [SS96, Lemma 2.11 and Lemma
2.14]:
Lemma 7.3. Suppose K is a uniformly locally finite and simply-connected
simplicial 2-complex with at least two ends. Suppose K does not contain separating vertices. Then there exists an essential track τ ⊂ K
which has the least complexity with respect to dH among all essential
tracks in K.
Remark 7.4. Let {τi }∞
i=1 be a minimizing sequence. Since K is uniformly locally finite, there are only finitely many combinatorial possibilities for {τi }∞
i=1 . Thus we can assume all the τi ’s are inside a finite
subcomplex L. Moreover, we can construct a hyperbolic metric dH on
L as above and it suffices to work in the space (L, dH ). However, if we
do not use marked points in the construction of the hyperbolic metric
on K, then each τi may sit inside a copy of L with different shears
along the edges of L.
In [SS96], K is assumed to be cocompact, so one does not need to
worry about the above issue.
Remark 7.5. If we metrize each simplex in K with the Euclidean metric, then Lemma 7.3 and Lemma 7.6 may not be true. For example,
one can take the following picture, where the dotted line is part of
some track τ . Once we shorten τ , it may hit the central vertex of the
hexagon. However, this cannot happen if we have hyperbolic metrics
on each simplex. Once τ gets too close to some vertex, then it takes a
large amount of length for τ to escape that vertex since dH is complete
(actually it does not matter if dH is not complete, since we also have a
upper bound on the weight of τ ).
60
JINGYIN HUANG AND BRUCE KLEINER
The next result can be proved in a similar fashion:
Lemma 7.6. Let K be a simply-connected simplicial 2-complex. Let
A ⊂ K be a uniformly locally finite subcomplex such that
(1) A contains an essential track of K.
(2) A does not contain any separating vertex of K.
Then there exists an essential track τ of K which has the least complexity among all essential tracks of K with support in A.
Lemma 7.7. [SS96, Lemma 2.7] Let τ1 and τ2 be two essential tracks
of K which are minimal in the sense of Lemma 7.3 or Lemma 7.6, then
either τ1 = τ2 , or τ1 ∩ τ2 = ∅.
7.2. The proof of Proposition 6.2. First we briefly recall the notion
of Rips complex. See [BH99, Chapter III.Γ.3] for more detail. Let
(X, d) be a metric space and pick R > 0. The Rips complex PR (X, d)
is the geometric realization of the simplicial complex with vertex set X
whose n-simplices are the (n + 1)-element subsets {x0 , · · · , xn } ⊂ X of
diameter at most R.
Let d be the usual metric on Z. Define a new metric d¯ on Z by
¯ y) = sup d(g(x), g(y))
d(x,
g∈G
¯ is quasi-isometric to (Z, d), and G acts on (Z, d)
¯
Note that (Z, d)
¯
by isometries. Since (Z, d) is Gromov-hyperbolic, the Rips complex
¯ is contractible for some R = R(L, A) (see [BH99, Proposition
PR (Z, d)
¯ Then K is simplyIII.Γ.3.23]). Let K be the 2-skeleton of PR (Z, d).
connected, uniformly locally finite and 2-ended.
We make K a piecewise Euclidean complex by identifying each 2face with an equilateral triangle and identifying each edge with [0, 1].
Let dE be the resulting length metric. There is an inclusion map
i : (Z, d) → (K, dE ) which is a quasi-isometry with quasi-isometry
constants depending only on L and A.
GROUPS QUASI-ISOMETRIC TO RAAG’S
61
Claim 7.8. There exist D1 = D1 (L, A) and a collection of disjoint
essential tracks {τi }i∈I of K such that
(1) {τi }i∈I is G-invariant.
(2) The diameter of each τi with respect to dE is ≤ D1 .
(3) Each connected component of K \ (∪i∈I τi ) has diameter ≤ D1 .
In the following proof, we denote the ball of radius D centered at x
in K with respect to dE by BE (x, D). Let diamE be the diameter with
respect to dE .
Proof of Claim 7.8. First we assume K does not have separating vertices. Since K is quasi-isometric to Z, there exists D = D(L, A) such
that K \ BE (x, D) has at least two unbounded components for each
x ∈ K. Thus every (D + 1)-ball contains an essential track with weight
bounded above by D0 = D0 (L, A). We put a G-invariant hyperbolic
metric dH on K as in Section 7.1. By Lemma 7.3, there exists an essential track τ ⊂ K of least complexity. Note that diamE (τ ) ≤ D0 since
the weight w(τ ) ≤ D0 . Lemma 7.7 implies the G-orbits of τ give rise
to collection of disjoint essential tracks in K.
A collection of tracks {τi }i∈I of K is admissible if
(1) Each track in {τi }i∈I is essential and different tracks have empty
intersection.
(2) No two tracks in {τi }i∈I are parallel.
(3) The collection {τi }i∈I is G-invariant.
(4) diamE (τi ) ≤ D0 for each i ∈ I.
There exists a non-empty admissible collection of tracks by previous
discussion.
Let {τi }i∈I be a maximal admissible collection of tracks. Then this
collection satisfies the above claim with D1 = 2D0 + 5D. To see this,
let C be one connected component of K \ (∪i∈I τi ). Since each track
is essential and K is 2-ended, either diamE (C) < ∞ and C̄ \ C (C̄ is
the closure of C) is made of two tracks τ1 and τ2 , or diamE (C) = ∞
and C̄ \ C is made of one track. Let us assume the former case is true.
The latter case can be dealt in a similar way. Let A be the maximal
subcomplex of K which is contained in C. Then A is uniformly locally
finite and C \ A is contained in the 1-neighbourhood of τ1 ∪ τ2 .
Suppose diamE (C) ≥ 2D0 + 5D. Since diamE (τi ) ≤ D0 for i = 1, 2,
there exists x ∈ A such that BE (x, 2D) ⊂ A. Thus A contains an
essential track of X with its weight bounded above by D0 . Let η ⊂ A
62
JINGYIN HUANG AND BRUCE KLEINER
be an essential track of K which has the least complexity in the sense
of Lemma 7.6, then w(η) ≤ D0 , hence diamE (η) ≤ D0 . Moreover,
by Lemma 7.7, for each g ∈ Stab(A) = Stab(C), either g · η = η or
g · η ∩ η = ∅. Thus we can enlarge the original admissible collection of
tracks by adding the G-orbits of η, which yields a contradiction.
The case when K has separating vertices is actually easier, since one
can find essential tracks on the -sphere of each separating vertices.
The rest of the proof is identical.
We now continue with the proof of Proposition 6.2.
Pick a regular neighbourhood N (τi ) for each τi such that collection
{N (τi )}i∈I is disjoint and G-invariant. Then we define a map φ from
K to a tree T by collapsing each component of Y \ ∪i∈I N (τi ) to a
vertex and collapsing each N (τi ), which is homeomorphic to τi × (0, 1),
to the (0, 1) factor. It is easy to make φ equivariant under G, and by
the above claim, φ is a quasi-isometry with quasi-isometry constants
depending only on L and A. Note that T is actually a line since τ is
essential and K is 2-ended. Then Proposition 6.2 follows by considering
the G-equivariant map φ ◦ i : (Z, d) → T .
Remark 7.9. If the action G y Z by (L, A)-quasi-isometries in Proposition 6.2 is not cobounded, then the resulting isometric action Λ : G y
Z is also not cobounded, hence there are two possibilities:
(1) if G coarsely preserve the orientation of Z, then Λ is trivial;
(2) otherwise Λ factors through a Z/2-action by reflection.
Part II: An alternate blow-up construction for buildings
8. Blowing-up buildings by metric spaces
In Section 5 we discussed a construction for blowing-up a right-angled
building. Here we give a somewhat different approach, which allows
us to deal with more general situations. It unifies several other objects discussed in this paper and has interesting applications to graph
products.
8.1. Construction. We are motivated by the fact that the universal
cover of the Salvetti complex can obtained by gluing a collection of
standard flats in a way determined by the associated building. Similarly, we can construct the universal cover of exploded Salvetti complex
by gluing a collection of branched flats. Here we specify the gluing
GROUPS QUASI-ISOMETRIC TO RAAG’S
63
pattern, and replace standard flats or branched flats by other product structures, which allows us to generalize the construction of these
objects.
Let Γ be a finite simplicial graph, and let B be a building modelled
on the right-angled Coxeter group W (Γ) with finite defining graph Γ.
Step 1: For every vertex v of Γ, and every v-residue rv ⊂ B, we
associate a metric space Zrv , and a map frv which maps chambers in
rv to Zrv . If another v-residue rv0 is parallel to rv , then we associate rv0
with the same space Zrv , and a map frv0 which is induced by frv and
the parallelism between rv and rv0 .
Let Mrv be the mapping cylinder of frv . Mrv is obtained by attaching
edges to Zrv , and the endpoints of these edges which correspond to the
domain of frv are called tips of Mrv . There is a 1-1 correspondence
between the tips of Mrv and chambers in rv . Each edge in Mrv has
length = 1, and Mrv is endowed with the quotient metric (see [BH99,
Chapter I.5.19]).
In summary, for each parallel class of v-residues, we have associated
a space Mrv , whose tips are in 1-1 correspondence with chambers in
any chosen v-residue in this parallel class. The spaces Zrv ’s and the
maps frv ’s is called the blow-up data on building B.
Step 2:
Let Sr be the poset of spherical residues in B. We define a functor
Φ from Sr to the category of nonempty metric spaces and isometric
embeddings as follows.
Q
Let rJ = j∈J rj be the product decomposition of a spherical JQ
residue into its rank 1 residues. Define Φ(rJ ) = j∈J Mrj , which is the
Cartesian product of the metric spaces Mrj ’s.
Suppose rJ 0 is a spherical J 0 -residue such that rJ 0Q⊂ rJ . In Q
this case
J ⊂ J. Moreover, rJ 0 can be expressed as a product j ∈J
u
×
/ 0 j
j∈J 0 rj ,
here uj ∈ rj is a chamber. By step 1, each uj corresponds to a tip
tj ∈ Mrj . Then we Q
define the Q
morphism Φ(r
QJ 0 ) → Φ(rJ ) to be the
isometric embedding j ∈J
j∈J Mrj .
/ 0 tj ×
j∈J 0 Mrj →
0
Step 3:
F
We begin with the disjoint union rJ ∈Sr Φ(rJ ), and for every inclusion rJ 0 ⊂ rJ , we glue Φ(rJ 0 ) with a subset of Φ(rJ ) by using the map
Φ(rJ 0 ) → Φ(rJ ) defined in step 2.
64
JINGYIN HUANG AND BRUCE KLEINER
We denote the resulting space with the quotient metric ([BH99, Definition I.5.19]) by Π. Π is called the blow-up of B with respect to the
blow-up data Zrv ’s and frv ’s.
The following lemma gives an alternative description of our gluing
process.
Lemma 8.1. Two points x1 ∈ Φ(rJ1 ) and x2 ∈ Φ(rJ2 ) are glued together if and only if rJ = rJ1 ∩rJ2 6= ∅ and there exists a point x ∈ Φ(rJ )
such that for i = 1, 2, its image under Φ(rJ ) → Φ(rJi ) is xi .
Proof. The if direction is clear. For the only if direction, define x1 ∼ x2
if the condition in the lemma is satisfied, it suffices to show ∼ is an
equivalence relation. However, this follows from our construction.
We can collapse each mapping cylinder in the construction to its
range, namely Mrv → Zrv . This induces a collapsing Π → Π̄. Π̄ is
called the reduced blow-up of B with respect to Zrv ’s and frv ’s.
Remark 8.2. There is some flexibility in this construction; for example, we can also collapse certain collections of mapping cylinders and
keep other mapping cylinders to obtain a “semi-reduced” blow-up. Also
instead of requiring Φ(rJ ) to be the Cartesian product of the metric
spaces Mrj ’s, we can use lp -product for 1 ≤ p ≤ ∞. We will see an
example later where it is more natural to use l1 -product.
Let B1 ⊂ B be a residue. We restrict the blow-up data on B to B1
and obtain a blow-up Π1 of B1 . There is a natural map Π1 → Π, which
is injective by Lemma 8.1. If B1 is a spherical residue, then the image
of Π1 → Π is a product of mapping cylinders. It is called a standard
product, and there is a 1-1 correspondence between spherical residues
in B and standard products in Π.
8.2. Properties.
Proposition 8.3. If each Zrv in the above construction is a point, then
the resulting blow-up Π is isometric to the Davis realization of B.
Proof. Let |B| be the Davis realization of B. Let rJ ⊂ B be a spherical residue and let D(rJ ) be the subcomplex of |B| spanned by those
vertices corresponding to residues inside rJ . Note that D(rJ ) is a convex subcomplex, and there is a naturally defined isomorphism hrJ :
D(rJ ) → Φ(rJ ). One readily verifies the following:
(1) The collection of all such D(rJ )’s covers |B|.
GROUPS QUASI-ISOMETRIC TO RAAG’S
65
(2) There is a 1-1 correspondence between D(rJ )’s and Φ(rJ )’s in
step 2 of Section 8.1.
(3) The gluing pattern of these D(rJ )’s is compatible with step 3
of Section 8.1 via the maps hrJ ’s.
It follows that hrJ ’s induce a cubical isomorphism |B| → Π.
Remark 8.4. If each Zrv is a point, then the corresponding reduced
blow-up is a point. However, later we will see cases where the reduced
blow-up is more interesting than the blow-up.
It follows that for any blow-up Π of B, there is a surjective projection
map from Π to the Davis realization of B which is induced by collapsing
each Zrv to a point. This is called the canonical projection. This map
is a special case of the following situation.
Let Π1 be the blow-up of B1 with respect to Zrv ’s and frv ’s, and let
Π2 be the blow-up of B2 with respect to Yrv ’s and grv ’s. Let Π̄1 and Π̄2
be the corresponding reduced blow-ups. Suppose:
(1) There exists a bijection η : B1 → B2 such that both η and η −1
preserve spherical residues.
(2) For each rv , there exists a map hrv : Zrv → Yη(rv ) such that the
following diagram commutes:
rv −−−→ η(rv )
η
gη(r )
f
y v
y rv
Zrv −−−→ Yη(rv )
hrv
Then there are induced maps h : Π1 → Π2 and h̄ : Π̄1 → Π̄2 . Let
η∗ : |B1 | → |B2 | be the map between Davis realizations induced by η.
Then h fits into the following commuting diagram:
Π1 −−−→
h
y
Π2
y
|B1 | −−−→ |B2 |
η∗
Let B 0 ⊂ B be a residue. Let projB0 : B → B 0 be the projection
map defined in Section 3.4. Pick a spherical residue r ∈ B. Suppose
r0 = projr (B 0 ) and B00 = projB0 (r). Let r = r0 × r0⊥ be the product
decomposition of r. It follows from [AB08, Lemma 5.36] that the map
projB0 : r → B00 is a composition of the factor projection r → r0 and
66
JINGYIN HUANG AND BRUCE KLEINER
projB0 : r0 → B00 . Recall that r0 and B00 are parallel and the second
map induces the parallelism map between r0 and B00 .
Let Π be a blow-up of B. We restrict the blow-up data on B to
B , and let Π0 be the resulting blow-up of B 0 . Since there is a 1-1
correspondence between spherical residues in B or B 0 with standard
products in Π or Π0 , then above discussion implies that we can assign
to each standard product of Π a standard product in Π0 , together with
a map between them. This assignment is compatible with the gluing
pattern of these standard products, thus induces a map ρ : Π → Π0 .
This map ρ is a 1-Lipschitz retraction map, and is called a residue
retraction map. The following is an consequence of the existence of
such map.
0
Corollary 8.5. The inclusion Π0 → Π is an isometric embedding.
Theorem 8.6. If each Zrv is CAT (0), then the blow-up Π of a rightangled building B with the quotient metric is also CAT (0).
Proof. We start with the following observation. If R ⊂ B is a residue,
then R is also a building. We can restrict any blow-up data of B to a
blow-up data of R. Let ΠB and ΠR be the resulting blow-ups. And let
|B| and |R| be the Davis realization of B and R. Then the following
diagram commutes:
ΠR −−−→
i1
y
ΠB
y
|R| −−−→ |B|
i2
The two vertical maps are canonical projections, and i1 and i2 are
isometric embeddings (Corollary 8.5). Moreover, the image of i1 is
exactly the inverse image of |R| under the canonical projection (here
|R| is viewed as a subspace of |B|).
We first show ΠB is locally CAT (0). Define the rank of B to be the
maximum possible number of vertices in a clique of Γ. We induct on
the rank of B. The rank 1 case is clear. Now assume B is of rank n,
and all blow-ups of buildings of rank ≤ n − 1 are locally CAT (0).
Pick vertex x ∈ |B| and suppose x corresponds to a J-residue r ⊂ B.
Let π : ΠB → |B| be the canonical projection and let st(x) be the open
star of x in |B|. It suffices to show π −1 (st(x)) is locally CAT (0).
GROUPS QUASI-ISOMETRIC TO RAAG’S
67
Case 1: J 6= ∅. Let J ⊥ be the vertices in Γ which are adjacent to
every vertex in J. Let R be the J ∪ J ⊥ -residue that contains r. We
restrict the blow-up data on B to R, and let ΠR be resulting metric
space. Let R = R1 ×R2 be the product decomposition (see Section 3.4)
induced by J ∪ J ⊥ . Then we also have |R| = |R1 | × |R2 | and ΠR =
ΠR1 × ΠRi where ΠRi is some blow-up of Ri for i = 1, 2. By the
induction assumption, ΠR2 is locally CAT (0). Moreover, R1 can be
further decomposed into a product of buildings of rank 1, thus ΠR1 is
locally CAT (0). It follows that ΠR is locally CAT (0). Since St(x) ⊂
|R|, π −1 (st(x)) ⊂ ΠR . Since ΠR → ΠB is a local isometry, it follows
that π −1 (st(x)) is locally CAT (0).
Case 2: J = ∅. In this case, π induces a simplicial isomorphism
between π −1 (St(x)) and St(x). It follows from the CAT (0) property of
|B| and [BH99, Lemma I.5.27] that π −1 (st(x)) is locally CAT (0).
It remains to prove ΠB is simply connected. Actually we show the
canonical projection π : ΠB → |B| is a homotopy equivalence. We
construct a homotopy inverse φ : |B| → ΠB as follows.
Recall that |B| be also be viewed as a blow-up of B where the associated metric spaces Zr0 v ’s are single points. Let Mr0v be the mapping cylinder of the constant map rv → Zr0 v . There is a natural map
hrv : Mrv → Mr0v induced by collapsing each Zrv to Zr0 v .
For each rv , we choose a map h0rv : Zr0 v → Zrv such that if rv and ru
are parallel residues of rank 1, then h0rv and h0ru have the same image.
We extend h0rv to a map Mr0v → Mrv as follows. Pick a chamber c ∈ rv ,
and let ec and e0c be the edges in Mrv and Mr0v corresponding to the
chamber c respectively. We identify these two edges with [0, 1] such
that the 0-endpoints are the tips. We map the [1/2, 1] half of e0c to the
constant speed geodesic joining the midpoint of ec and the image of
h0rv in Zrv (such geodesic is unique since Mrv is CAT (0)), and map the
[0, 1/2] half of e0c identically to the [0, 1/2] half of ec .
One readily verify that one can take suitable products of the maps
: Mr0v → Mrv and glue them together to obtain a map φ : |B| → ΠB .
Moreover, for each rv , there is a geodesic homotopy between h0rv ◦ hrv
to identity, and these homotopies induces a homotopy between φ ◦ π
and identity. Similarly we can produce a homotopy between π ◦ φ and
identity.
h0rv
8.3. An equivariant construction. The construction in this section
is similar to Section 5.6.
68
JINGYIN HUANG AND BRUCE KLEINER
Definition 8.7. Let B be a right-angled building. A bijection f :
B → B is flat-preserving if both f and f −1 map spherical residues to
spherical residues.
This definition is motivated by Definition 6.1.
Suppose H acts on B by flat-preserving bijections. Let Λ be the
collection of all parallel classes of rank 1 residues, then there is an
induced action H y Λ. We pick a representative from each orbit of
this action and denote the resulting set by {λu }u∈U .
Let Hu be the stabilizer of λu . Pick rank 1 residue ru in the parallel
class Λu and let ρu : Hu y ru be the induced factor action (see Section 5.6). We pick a metric space Zru , an isometric action Hu y Zru
and a Hu -equivariant map fru : ru → Zru . And we deal with residues
which are parallel to ru as in step 1 of Section 8.1.
We repeat this process for each element in {λu }u∈U . If λ ∈
/ {λu }u∈U ,
then we fix an element gλ ∈ H such that gλ (λ) ∈ {λu }u∈U . For each
rank 1 residue rv in the parallel class λ, we associate the metric space
Zgλ (rv ) and the map frv = fgλ (rv ) ◦ gλ , and we deal with residues which
are parallel to rv as before.
Let Π be the blow-up of B with respect to the above choice of spaces
and maps and let B̄ be the corresponding reduced blow-up. Then there
are induced actions H y Π and H y Π̄ by isometries. Moreover, the
canonical projection Π → |B| is H-equivariant.
Next we apply this construction to graph products of groups. Pick
a finite simplicial graph Γ with its vertex set denoted by I, and pick
a collection of groups {Hi }i∈I , one for each vertex of Γ. Let H be the
graph product of the Hi ’s with respect to Γ and let B = B(Γ) be the
right-angled building associated with this graph product (see [Dav98,
Section 5]). We can identify H with the set of chambers in B. Under
this identification, the rank 1 residues in B correspond to leftQcosets of
Hi ’s and spherical residues in B correspond to left cosets of j∈J Hj ’s
where J is the vertex set of some clique in Γ.
In the following discussion, we will slightly abuse notation. When
we say a residue in H, we actually mean the corresponding object in
B under the above identification.
Let ΛH be the collection of all parallel classes of rank 1 residues in
H, and let λi ∈ ΛH be the parallel class represented by Hi . Then
there is a 1-1 correspondence between λi ’s and H-orbits of the induced
action H y ΛH . Thus we take {λu }u∈U in the above construction to
GROUPS QUASI-ISOMETRIC TO RAAG’S
69
be {λi }i∈I . For each Hi , pick a metric space Zi , an isometric action
Hi y Zi , and a Hi -equivariant map fi : Hi → Zi . Given such data, we
can produce a blow-up ΠH of the building B.
Theorem 8.8. Let H be the graph product of {Hi }i∈I with respect to
a finite simplicial graph Γ. Suppose one of the following properties is
satisfied by all of Hi ’s:
(1)
(2)
(3)
(4)
It
It
It
It
acts
acts
acts
acts
geometrically on a CAT (0) space;
properly on a CAT (0) space;
geometrically on a CAT (0) cube complex;
properly on a CAT (0) cube complex.
Then H also has the same property.
Proof. We prove (1). The other assertions are similar. Suppose each Hi
acts geometrically on a CAT (0) space Zi . By Theorem 8.6, it suffices
to show the action H y ΠH is geometric.
We first show that H y ΠH is cocompact. Note that ΠH is locally
compact, since each mapping cylinder in the construction of ΠH is
locally compact. We deduce that ΠH is proper since ΠH is CAT (0).
Thus it suffices to show H y ΠH is cobounded. Let |H| be the Davis
realization of B and let π : ΠH → |H| be the H-equivariant canonical
projection. Let V0 ⊂ |H| be the collection of rank 0 vertices in |H|.
Note that π −1 (v) is exactly one point for any v ∈ V0 . The H-action
on V0 , hence on π −1 (V0 ) is transitive. Since each action Hi y Zi is
cobounded, the set of tips in the mapping cylinder of fi : Hi → Zi
is cobounded. Thus π −1 (V0 ) is cobounded in ΠH since it arises from
tips of mapping cylinders in the construction of ΠH . It follows that
H y ΠH is cobounded, hence cocompact.
Now we prove H y ΠH is proper. It suffices to show the intersection
of π −1 (V0 ) with any bounded metric ball in ΠH is finite. For each
λ ∈ ΛH , pick a residue rv in the parallel class λ. Let ξλ : ΠH → Mrv
be the residue projection. Note that ξλ does not depend on the choice
of rv in the parallel class. We claim there exists constant L > 0 such
that
L−1 d(x, y) ≤ Σλ∈ΛH d(ξλ (x), ξλ (y)) ≤ Ld(x, y)
for any x, y ∈ π −1 (V0 ).
To see the second inequality, pick a geodesic line ` ∈ ΠH connecting
x and y. Since each standard product in ΠH is convex (Corollary 8.5),
the intersection of ` with each standard product is a (possibly empty)
70
JINGYIN HUANG AND BRUCE KLEINER
segment. Thus we can assume ` is an concatenation of geodesic segments {`i }ki=1 such that each `i is contained in a standard product Pi .
Recall that each Pi is a product of mapping cylinders. For each i, we
construct another piecewise geodesic `¯i such that:
`¯i and `i have the same endpoints.
i
`¯i is a concatenation of geodesic segments {`¯ij }kj=1
such that the
projection of `¯ij to all but one of the product factors of Pi are
trivial.
(3) There exists D which is independent of x, y and `i such that
length(`¯i ) ≤ D length(`i ).
(1)
(2)
It follows from the construction of residue projection that for each λ,
ξλ (`¯ij ) is either a point, or a segment which has the same length as `¯ij ,
and there is exactly one λ such that ξλ (`¯ij ) is non-trivial. Thus the
second inequality follows.
To see the first inequality, let cx and cy be chambers in B that correspond to x and y respectively. Let G ⊂ B be a shortest gallery
connecting cx and cy . Every two adjacent elements in the gallery are
contained in a rank 1 residue, and different pairs of adjacent elements
give rise to rank 1 residues in different parallel class (otherwise we can
shorten the gallery). Note that G corresponds to a chain of points in
π −1 (V0 ). Connecting adjacent points in this chain by geodesics induces
a piecewise geodesic ` ⊂ ΠH connecting x and y, moreover, the length
of ` is equal to Σλ∈ΛH d(ξλ (x), ξλ (y)), thus the first inequality holds.
Given x ∈ π −1 (V0 ), any other point y ∈ π −1 (V0 ) in the R-ball can
be connected to x via a chain of points in π −1 (V0 ) which is induced by
a shortest gallery as above. The above claim implies that the size of
the chain is bounded above in terms of R. Two adjacent points have
distance in ΠH bounded above in terms of R, and they correspond
to chambers in the same rank 1 residue. There are only finite many
such chains since our assumption implies there are only finitely many
choices for each successive point of such a chain. Thus the intersection
of π −1 (V0 ) with any bounded metric ball in ΠH is finite.
Remark 8.9. Suppose we use l1 -product instead of Cartesian product
in step 2 of Section 8.1. Let dl1 be the resulting metric on ΠH . Then
dl1 (x, y) = Σλ∈ΛH d(ξλ (x), ξλ (y)) for any x, y ∈ π −1 (V0 ).
The construction in this section unifies several other constructions
as follows:
(1) The Davis realization of right-angled building (Proposition 8.3).
GROUPS QUASI-ISOMETRIC TO RAAG’S
71
(2) If B is the right-angled building associated with a graph product
of Z’s, each space is R and each map is a bijection between
chambers and integer points in R, then the blow-up of B is the
universal covering of the exploded Salvetti complex with the
correct group action.
(3) The corresponding reduced blow-up in (2) is the universal covering of Salvetti complex.
(4) Let B be as in (2). If each space is R, and each map arises from
Proposition 6.2, then the blow-up of B is the geometric model
we construct for group quasi-isometric to RAAG’s in Section 6.
(5) Suppose H is a graph product of a collection of groups {Hi }i∈I
and B is the associated building. Suppose each Hi acts on a
CAT (0) cube complex Zi with a free orbit. And we pick a Hi equivariant map from Hi to the free orbit. Then the action of
the H on the associated reduced blow-up of B is equivariantly
isomorphic to the graph product of group actions defined on
[Hag08, Section 4.2]
Part III: A wallspace approach to cubulating RAAG’s
In the remainder of the paper, we use wallspaces to give a different
approach to cubulating quasi-actions on RAAG’s. This gives a shorter
path to the cubulation, but gives less precise information about its
structure.
We assume G(Γ) 6= Z in Section 9 and Section 10.
9. Construction of the wallspace
9.1. Background on wallspaces. We will be following [HW14] in
this section.
Definition 9.1. (Wallspaces) Let Z be a nonempty set. A wall of Z
is a partition of Z into two subsets {U, V }, each of which is called a
halfspace. Two points x, y ∈ Z are separated by a wall if x and y are
in different halfspaces. (Z, W) is a wallspace if Z is endowed with a
collection of walls W such that every two distinct points x, y ∈ Z are
separated by finitely many walls. Here we allow the situation that two
points are not separated by any wall.
Two distinct walls W = {U1 , U2 } and W 0 = {V1 , V2 } are transverse
if Ui ∩ Vj 6= ∅ for all 1 ≤ i, j ≤ 2.
72
JINGYIN HUANG AND BRUCE KLEINER
An orientation of a wall is a choice of one of its halfspaces. An
orientation σ of the wallspace (Z, W) is a choice of orientation σ(W )
for each W ∈ W.
Definition 9.2. A 0-cube of (Z, W) is an orientation σ such that
(1) σ(W ) ∩ σ(W 0 ) 6= ∅ for all W, W 0 ∈ W.
(2) For each x ∈ Z, we have x ∈ σ(W ) for all but finitely many
W ∈ W.
Each CAT(0) cube complex C gives rise to a wallspace (Z, W) where
Z is the vertex set of C and the walls are induced by hyperplanes of
C. Conversely, each wallspace (Z, W) naturally give rises to a dual
CAT(0) cube complex C(Z, W). The vertices of C(Z, W) are in 1-1
correspondence with 0-cubes of (Z, W). Two 0-cubes are joined by an
edge if and only if their orientations are different on exactly one wall.
Lemma 9.3. ([HW14, Corollary 3.13]) C(Z, W) is finite dimensional
if and only if there is a finite upper bound on the size of collections of
pairwise transverse walls. In this case, dim(C(Z, W)) is equal to the
largest possible number of pairwise transverse walls.
There is a 1-1 correspondence between walls in (Z, W) and hyperplanes in C(Z, W). We can define a map which associates each finite
dimensional cube in C(Z, W) a collection of pairwise transverse walls
in (Z, W) by considering the hyperplanes which have non-trivial intersection with the cube.
Lemma 9.4. ([HW14, Proposistion 3.14]) The above map induces a 11 correspondence between finite dimensional maximal cubes in C(Z, W)
and finite maximal collections of pairwise transverse walls in (Z, W).
Two 0-cubes σx and σy of (Z, W) are separated by a wall W ∈ W
if and only if σx (W ) and σy (W ) are different halfspaces of W . In this
case, the hyperplane in C(Z, W) corresponding to W separates the
vertices associated with σx and σy .
9.2. Preservation of levels. In the rest of this section, we assume
Out(G(Γ)) is finite and ρ : H y G(Γ) is an (L, A)-quasi-action. We
refer to Definition 6.1 for flat preserving bijections.
By Theorem 1.3, we can assume ρ is an action by flat-preserving
bijections which are also (L, A)-quasi-isometries.
Let P(Γ) be the extension complex, and for every standard flat F ⊂
X(Γ), let ∆(F ) ∈ P(Γ) be the parallel class; see Section 3.3. Note that
GROUPS QUASI-ISOMETRIC TO RAAG’S
73
every flat preserving bijection h : G(Γ) → G(Γ) induces a simplicial
isomorphism ∂h : P(Γ) → P(Γ). Thus ρ induces an action H y P(Γ)
by simplicial isomorphisms.
Definition 9.5 (Levels). Pick a vertex v ∈ P(Γ) and let l ⊂ X(Γ) be a
standard geodesics such that ∆(l) = v. For every vertex x ∈ l, a v-level
(of height x) is defined to be {z ∈ G(Γ) | πl (z) = x}, here πl is the
CAT(0) projection onto lv . Note that πl−1 (x) is a convex subcomplex
of X(Γ) and the v-level of height x is exactly the vertex set of this
convex subcomplex. The definition of v-level does not depend on the
choice of the standard geodesic l in the parallel class.
Lemma 9.6. Let h : G(Γ) → G(Γ) be a flat-preserving projection.
Then for any v ∈ P(Γ), h sends v-levels to ∂h(v)-levels.
Proof. Pick standard geodesic l such that ∆(l) = v and pick vertex
x ∈ l. Suppose l0 is the standard geodesic such that α(v(l)) = v(l0 ). Let
y ∈ G(Γ) be such that πl (y) = x. It suffices to show πl0 (h(y)) = h(x).
Let ω be a combinatorial geodesic joining y and x and let {xi }ni=0
be vertices in ω such that (1) x0 = y and xn = x; (2) for 1 ≤ i ≤ n,
there exists a standard geodesic li such that xi ∈ li and xi−1 ∈ li .
Note that ω ⊂ πl−1 (x) since πl−1 (x) is a convex subcomplex. Thus
li ⊂ πl−1 (x) for all i, hence ∆(li ) 6= v for all i. Suppose x0i = h(xi )
and let li0 be the standard geodesic which contains x0i and x0i−1 . Then
∂h(∆(li )) = ∆(li0 ), thus ∆(li0 ) 6= ∆(l0 ). It follows that πl0 (li0 ) is a point,
and hence πl0 (x0i ) = πl0 (x0i−1 ) for all i.
Lemma 9.7. Let K ⊂ X(Γ) be a standard subcomplex and define ∆(K)
to be the union of all ∆(F ) with F ranging over all standard flats in
K. Pick vertex v ∈ P(Γ). If v ∈
/ ∆(K), then K (0) is contained in a
v-level.
Proof. Let l ∈ X(Γ) be a standard geodesic such that ∆(l) = v. It
suffices to prove for each edge e ∈ K, πl (e) is a point. Suppose the
contrary is true. Then πl (e) is an edge of l by Lemma 3.3, and hence the
standard geodesic le that contains e is parallel to l. However, le ⊂ K
since K is standard, which implies v ∈ ∆(K).
9.3. A key observation. The flat-preserving action H y G(Γ) induces an action H y P(Γ). For each vertex v ∈ P(Γ), let Hv ⊂ H
be the subgroup which stabilizes v. In other words, if l is a standard
geodesic such that ∆(l) = v, then Hv is the stabilizer of the parallel
set of l.
74
JINGYIN HUANG AND BRUCE KLEINER
By Lemma 9.6, Hv permutes v-levels. Recall that the CAT(0) projection πl induces a 1-1 correspondence between v-levels and vertices of l;
moreover, for any two v-levels L1 and L2 , d(L1 , L2 ) = d(πl (L1 ), πl (L2 ))
(d is the word metric). Thus the collections of v-levels can be identified
with Z (endowed with the standard metric), and we have an induced
action ρv : Hv y Z by (L0 , A0 )-quasi-isometries; here L0 , A0 can be
chosen to be independent of v.
We recall the following result which is proved in Section 7.
Proposition 9.8. If a group G has an action on Z by (L, A)-quasiisometries, then there exists another action G y Z by isometries which
is related to the original action by a surjective equivariant (L0 , A0 )quasi-isometry f : Z → Z with L0 , A0 depending on L and A.
Definition 9.9 (Branched lines and flats). A branched line is constructed from a copy of R by attaching finitely many edges of length 1
to each integer point. We also require the valence of each vertex in a
branched line is bounded above by a uniform constant. This space has
a natural simplicial structure and is endowed with the path metric. A
branched flat is a product of finitely many branched lines.
Definition 9.10 (Branching number). Let β be a branched line. We
define the branching number of β, denoted by b(β), to be the maximum
valence of vertices in β.
The following picture is a branched line with branching number = 5:
Remark 9.11. A branched line can be roughly thought as a “mapping
cylinder” of the map f in Proposition 9.8.
Definition 9.12 (Tips and wall structure on tips). Let β be a branched
line. We call the vertices of valence ≤ 2 in β the tips of β, and the
collection of all tips is denoted by t(β). The set of hyperplanes in β
(namely the midpoints of edges) induces a natural wall structure on
t(β).
For a branched flat F , we define t(F ) to be the product of the tips
of its branched line factors.
Corollary 9.13. If a group G has an action on Z by (L, A)-quasiisometries, then there exist a branched line β and an isometric action
GROUPS QUASI-ISOMETRIC TO RAAG’S
75
G y β which is related to the original action by a bijective equivariant
map q : Z → t(β). Moreover, there exists a constant M depending only
on L and A such that q is an M -bi-Lipschitz map and each vertex in
β has valance ≤ M .
Proof. Let f : Z → Z be the map in Proposition 9.8. We identify
the range of f as integer points of R. Pick x ∈ Z, if the cardinality
|f −1 (x)| ≥ 2, we attach |f −1 (x)| many edges of length 1 to R along x.
When |f −1 (x)| = 1, no edge will be attached. Let β be the resulting
branched line. Then there is a natural bijective equivariant map q :
Z → t(β) and the corollary follows from Proposition 9.8.
In the proof of Corollary 9.13, the CAT(0) cube complex dual to t(β)
with the wall structure described in Definition 9.12 is isomorphic to β.
Note that this is not true for general branched lines.
9.4. An invariant wallspace. G(Γ) has a natural wallspace structure
which is induced from X(Γ); however, H may not act on this wallspace.
We want to find an alternative wallspace structure on G(Γ) which is
consistent with the action of H.
Definition 9.14 (v-walls and v-halfspaces). Pick a vertex v ∈ P(Γ).
It follows from Corollary 9.13 that there exist a branched line βv , an
isometric action H y βv , and an Hv -equivariant surjective map ηv :
G(Γ) → t(βv ) such that the inverse image of each point in t(βv ) is
a v-level. A v-wall (or a v-halfspace) of G(Γ) is the pullback of some
wall (or halfspace) of t(βv ) (see Definition 9.12) under the map ηv . The
collection of all v-walls is denoted by Wv .
Now for each vertex v ∈ P(Γ), we want to choose a collection of
v-walls, βv and ηv in an H-equivariant way. Recall that H y G(Γ)
induces an action H y P(Γ). We pick one representative from each
vertex orbit of H y P(Γ), which gives rise to a collection {vi }i∈I . We
choose Wvi as above and let W be the union of all walls in Wvi for
i ∈ I, together with the H-orbits of these walls.
Now we record several consequences of the above construction.
Lemma 9.15.
(1) Each v-halfspace is a union of v-levels, thus a v-wall and a v 0 wall do not induce the same partition of G(Γ) if v 6= v 0 .
(2) Distinct v-walls in W are never transverse.
(3) C(G(Γ), Wv ) is isomorphic to βv .
76
JINGYIN HUANG AND BRUCE KLEINER
(4) Pick a standard geodesic line l such that ∆(l) = v, and identify
the vertex set v(l) with Z. Then there exists N > 0 independent
of v ∈ P(Γ) such that the map ξv : Z ∼
= v(l) → C(G(Γ), Wv ) ∼
=
βv is an N -bi-Lipschitz bijection.
(5) There exists M > 0 independent of v ∈ P(Γ) such that the
branching number b(βv ) is at most M .
(4) and (5) follow from Corollary 9.13.
Lemma 9.16. (G(Γ), W) is a wallspace.
Proof. For any pair of points x, y ∈ G(Γ), we need to show there are
finitely many walls separating x and y. If a v-wall separates x from
y, then x and y are in different v-level. There are only finitely many
such vertices in P(Γ). Given such v, there are finitely many v-walls
separating x and y.
By construction, H acts on the wallspace (G(Γ), W). Let C be the
CAT(0) cube complex dual to (G(Γ), W), then there is an induced
action H y C. Now we look at several properties of C.
10. Properties of the cubulation
10.1. Dimension. We need the following lemmas before we compute
the dimension of the dual complex C.
Lemma 10.1. Let v1 , v2 ∈ P(Γ) be distinct vertices. Then a v1 -wall
and a v2 -wall are transverse if and only if v1 and v2 are adjacent.
Proof. The if direction is clear (one can consider a standard 2-flat corresponding to the vertices v1 and v2 ). The other direction follows from
Lemma 10.2 below.
Lemma 10.2. Let v1 , v2 ∈ P(Γ) be non-adjacent vertices. For i = 1, 2,
let li be standard geodesic such that ∆(li ) = vi . Let x2 ∈ l2 be the image
πl2 (l1 ) of the CAT(0) projection πl2 and let x1 = πl1 (l2 ). Then any v2 level which is not of height x2 is contained in the v1 -level of height x1
(see the picture below).
GROUPS QUASI-ISOMETRIC TO RAAG’S
77
Proof. Suppose the contrary is true. Then there exists a vertex x02 ∈ l2
with x02 6= x2 , a v2 -level K of height x02 and a vertex x ∈ K such that x
is not inside the v1 -level of height x1 . Let ω ⊂ X(Γ) be a combinatorial
geodesic connecting x and x02 . Since πl1 (x02 ) = x1 6= πl1 (x), there exists
an edge e ⊂ ω and edge e1 ⊂ l1 such that e and e1 are parallel. Thus
πl2 (e) = πl2 (e1 ) = x2 by Lemma 3.6. On the other hand, since πl−1
(x02 )
2
0
is convex, πl2 (e) ⊂ πl2 (ω) = x2 , which is a contradiction.
Lemma 10.3. dim(C) = dim(G(Γ)).
Proof. Let {Wi }li=1 ⊂ W be a collection of pairwise intersecting walls
and suppose Wi is a ui -wall. Then ui 6= uj for i 6= j. Thus {ui }li=1 are
vertices of a simplex in P(Γ) by Lemma 10.1. It follows from Lemma
9.3 that dim(C) = dim(P(Γ)) + 1 = dim(G(Γ)).
10.2. Standard branched flats. Pick a maximal simplex ∆ ∈ P(Γ).
Let W∆ be the collection of u-walls in W with u ranging over all vertices
of ∆. Let C(G(Γ), W∆ ) be the CAT(0) cube complex dual to the
wallspace (G(Γ), W∆ ).
Lemma 10.4. There exists a natural isometric embedding
C(G(Γ), W∆ ) ,→ C(G(Γ), W) = C.
Proof. First we want to assign an orientation to each wall W ∈ W \W∆ .
If W is a u-wall, then by the maximality of ∆, there exists vertex
v ∈ ∆ which is not adjacent to u. Pick standard geodesics lu such that
∆(lu ) = u. Let F be the maximal standard flat such that ∆(F ) = ∆.
Since u ∈
/ ∆, F (0) is contained in a u-level by Lemma 9.7. We orient
W such that
(10.5)
v(F ) ⊂ σ(W ).
Pick a 0-cube σ of (G(Γ), W∆ ). We claim one can obtain a 0-cube
of (G(Γ), W) by adding the orientations of walls in W \ W∆ defined as
78
JINGYIN HUANG AND BRUCE KLEINER
above. Condition (1) of Definition 9.2 follows from (10.5). To see (2),
suppose W ∈ W \ W∆ such that z ∈
/ σ(W ); then W separates z from
v(F ) by (10.5) and there are only finitely many such W by Lemma
9.16. Now we have a map from 0-cubes of (G(Γ), W∆ ) to 0-cubes of
(G(Γ), W), which extends to higher dimensional cells.
This embedding is isometric since two walls are transverse in the
wallspace (G(Γ), W∆ ) if and only if they are transverse in (G(Γ), W).
The image of C(G(Γ), W∆ ) under the above embedding is called a
maximal standard branched flat. The following properties are immediate:
(1) Each maximal standard branched flat is a convex subcomplex
of C, since embedding in Lemma 10.4 is isometric.
(2) Each maximal standard branched flat splits as a product of
branched lines C(G(Γ), W∆ ) ∼
= Πv∈∆ C(G(Γ), Wv ) by Lemma
10.1.
Lemma 10.6. Every maximal cube of C is contained in a maximal
standard branched flat, and hence every point of C is contained in a
maximal standard branched flat.
Proof. By Lemma 9.4, every maximal cube is determined by a maximal
collection of pairwise transverse walls. By Lemma 10.1, these walls
correspond to a maximal simplex in ∆ ⊂ P(Γ). Thus the lemma follows
from Lemma 10.4 since the image of the embedding C(G(Γ), W∆ ) →
C contains a maximal cube whose dual hyperplanes are the required
collection.
There is a bijective map Φ: {maximal standard flats in X(Γ)} →
{maximal standard branched flats in C}, since both sides of Φ are in
1-1 correspondence with maximal simplexes of P(Γ).
10.3. An equivariant quasi-isometry. Every point x ∈ G(Γ) gives
rise to a 0-cube of (X, W) by considering the halfspaces containing x.
This induces an H-equivariant map φ : G(Γ) → C. Let F ⊂ X(Γ) be a
maximal standard flat, so φ(v(F )) ⊂ Φ(F ) by (10.5). Actually φ(v(F ))
is exactly the collection of tips of Φ(F ). To see this, note that φ|v(F ) can
be written as a composition: v(F ) → C(G(Γ), W∆F ) → C(G(Γ, W)).
Thus there is a natural splitting φ|v(F ) = Πni=1 ξvi , here {vi }ni=1 are the
vertices of ∆(F ) and ξvi : Z → C(G(Γ), Wvi ) is the map in Lemma
9.15.
GROUPS QUASI-ISOMETRIC TO RAAG’S
79
Lemma 10.7. The map φ is coarsely surjective.
Proof. By Lemma 10.6, it suffices show there exists a constant D which
does not depend on the maximal flat F such that φ(v(F )) is D-dense
in Φ(F ). This follows from Lemma 9.15.
Proposition 10.8. The map φ : G(Γ) → C is an H-equivariant injective quasi-isometry.
Proof. We prove φ is a bi-Lipschitz embedding, then the proposition
follows from Lemma 10.7. Pick x, y ∈ G(Γ) and pick vertex v ∈ P(Γ).
Suppose l is a standard geodesic with ∆(l) = v. Let dv (x, y) be the
number of v-hyperplanes in X(Γ) which separate x from y (h is a vhyperplane if and only if l ∩ h is one point), and let dv (φ(x), φ(y)) be
the number of walls in Wv that separate x from y. It suffices to show
there exists L > 0 which does not depend on x, y and v such that
(10.9)
L−1 dv (x, y) ≤ dv (φ(x), φ(y)) ≤ Ldv (x, y).
Let x0 = πl (x) and y0 = πl (y). Then dv (x, y) = d(x0 , y0 ). Recall that
in Lemma 9.15 we define a map ξv : v(l) → C(G(Γ), Wv ) ∼
= βv , which
is L-bi-Lipschitz with L independent of v. Moreover, dv (φ(x), φ(y)) =
d(ξ(x0 ), ξ(y0 )). Hence (10.9) follows from Lemma 9.15.
Lemma 10.10. For any two maximal standard flats F1 , F2 ⊂ X(Γ),
F1 ∩ F2 6= ∅ if and only if Φ(F1 ) ∩ Φ(F2 ) 6= ∅.
Proof. The only if direction follows from the previous discussion. If
F1 ∩ F2 = ∅, then there is a hyperplane h separating F1 from F2 . Let l
be a standard geodesic dual to h and let v = ∆(l). By the maximality
of ∆(F1 ), there exists a vertex v1 ∈ ∆(F1 ) which is not adjacent to v,
and thus v(F1 ) is contained is some v-level V1 by Lemma 3.6. Similarly,
v(F2 ) ⊂ V2 for some v-level V2 6= V1 . If W ∈ Wv separates V1 from V2 ,
then W also separates Φ(F1 ) from Φ(F2 ) by (10.5).
If x ∈ C is a vertex, since each edge that contains x is inside a
maximal standard branched flat (Lemma 10.7), and these maximal
standard flats intersect each other, Lemma 10.10 and (5) of Lemma
9.15 imply:
Corollary 10.11. C is uniformly locally finite.
We have actually proved the following result since the condition on
Out(G(Γ)) has not been used after Section 9.2.
80
JINGYIN HUANG AND BRUCE KLEINER
Theorem 10.12. Let H y G(Γ) be an H-action by flat-preserving
bijections which are also (L, A)-quasi-isometries. Then there exists
a uniformly locally finite CAT(0) cube complex C with dim(G(Γ)) =
dim(C) such that the above H-action is conjugate to an isometric action H y C. If the H-action is proper or cobounded, then the resulting
isometric action is proper or cocompact respectively.
Corollary 10.13. Suppose Out(G(Γ)) is finite and let H y G(Γ) be a
quasi-action. Then there exists a uniformly locally finite CAT(0) cube
complex C with dim(G(Γ)) = dim(C) such that the above quasi-action
is quasiconjugate to an isometric action H y C. If the quasi-action
is proper or cobounded, then the resulting isometric action is proper or
cocompact respectively.
Roughly speaking, the cube complex C is obtained by replacing each
standard geodesic in X(Γ) by a suitable branched line.
Remark 10.14. The cubulation in Corollary 10.13 is slightly different
from the one in Theorem 6.4. However, if we modify Definition 9.12
such that tips are the vertices of valence 1 and repeat the whole construction, then cubulation in Corollary 10.13 coincides with the one in
Theorem 6.4.
Since we will not be using it, we will not give the argument. However, it is instructive to think about the following example. Given an
isometric action H y X(Γ), the output cube complex is X(Γ) in Corollary 10.13 and is Xe (Γ) in Theorem 6.4. Here Xe (Γ) is the universal
cover of exploded Salvetti complex defined in Section 5.1.
Corollary 10.15. Suppose Out(G(Γ)) is finite. If G is a finitely generated group quasi-isometric to G(Γ), then G acts geometrically on a
CAT(0) cube complex.
10.4. Preservation of standard subcomplex. We look at how standard subcomplexes behave under the map φ : G(Γ) → C in Proposition
10.8.
Let K ⊂ X(Γ) be a standard subcomplex and let WK be the collection of walls in W that separate two vertices of K. Let ∆(K) be the
same as in Lemma 9.7.
Lemma 10.16. Let v ∈ P(Γ) be a vertex.
(1) If v ∈ ∆(K), then Wv ⊂ WK .
(2) If v ∈
/ ∆(K), then Wv ∩ WK = ∅.
(3) There is a natural isometric embedding C(G(Γ), WK ) ,→ C.
GROUPS QUASI-ISOMETRIC TO RAAG’S
81
Proof. If v ∈ ∆(K), then there exists a standard geodesic l ⊂ K with
∆(l) = v. Thus Wv ⊂ WK . If v ∈
/ ∆(K), it follows from Lemma 9.7
(0)
that K is contained in a v-level, thus (2) is true. It follows from
(1) and (2) that for each wall W ∈ W \ WK , we can orient W such
that K (0) ⊂ σ(W ). The rest of the proof for (3) is identical to Lemma
10.4.
Theorem 10.17. Let φ : G(Γ) → C be as in Proposition 10.8. Then
there exists D > 0 which only depends on the constants of the quasiaction such that for any standard subcomplex K ⊂ X(Γ), there exists
a convex subcomplex K 0 ⊂ C such that φ(K (0) ) is a D-dense subset of
K 0.
Proof. By construction, the isometric embedding C(G(Γ), WK ) ,→ C
fits into the following commuting diagram:
K (0)
φ0
y
−−−→ G(Γ)
φ
y
C(G(Γ), WK ) −−−→
C
Here φ0 sends a vertex of K to the 0-cube of (G(Γ), WK ) which is consist
of halfspaces containing this vertex. It suffices to show the image of φ0
is D-dense.
This can be proved by the same arguments in Section 10.1. Namely,
for each simplex ∆ which is maximal in ∆(K), there is a natural isometric embedding C(G(Γ), W∆ ) → C(G(Γ), WK ), which gives rise to
standard branched flats which are maximal in C(G(Γ), WK ). These
branched flats cover C(G(Γ), WK ) and they are in 1-1 correspondence
with standard flats which are maximal in K. Moreover, given a standard flat F which is maximal in K, φ0 (F (0) ) is exactly the set of tips of
the corresponding branched flat in C(G(Γ), WK ). Now we can conclude
in the same way as Lemma 10.7.
References
[AB08]
[AGM13]
[Ahl02]
P. Abramenko and K. S Brown. Buildings: theory and applications.
Springer Science & Business Media, 2008.
I. Agol, D. Groves, and J. Manning. The virtual Haken conjecture.
Documenta Mathematica, 18:1045–1087, 2013.
A. R. Ahlin. The large scale geometry of products of trees. Geometriae
Dedicata, 92(1):179–184, 2002.
82
[Bas72]
JINGYIN HUANG AND BRUCE KLEINER
H. Bass. The degree of polynomial growth of finitely generated nilpotent
groups. Proceedings of the London Mathematical Society, 3(4):603–614,
1972.
[BB97]
M. Bestvina and N. Brady. Morse theory and finiteness properties of
groups. Inventiones mathematicae, 129(3):445–470, 1997.
[BH99]
M. Bridson and A. Haefliger. Metric spaces of non-positive curvature,
volume 319 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin,
1999.
[BJN10]
J.. Behrstock, T. Januszkiewicz, and W. D. Neumann. Quasi-isometric
classification of some high dimensional right-angled Artin groups.
Groups Geom. Dyn., 4(4):681–692, 2010.
[BKS08a] M. Bestvina, B. Kleiner, and M. Sageev. The asymptotic geometry of
right-angled Artin groups. I. Geom. Topol., 12(3):1653–1699, 2008.
[BKS08b] M. Bestvina, B. Kleiner, and M. Sageev. Quasiflats in CAT (0) complexes. arXiv preprint arXiv:0804.2619, 2008.
[BM00]
M. Burger and S. Mozes. Lattices in product of trees. Inst. Hautes
Études Sci. Publ. Math., (92):151–194 (2001), 2000.
[BM01]
N. Brady and J. Meier. Connectivity at infinity for right angled
Artin groups. Transactions of the American Mathematical Society,
353(1):117–132, 2001.
[BN08]
J. Behrstock and W. D. Neumann. Quasi-isometric classification of
graph manifold groups. Duke Mathematical Journal, 141(2):217–240,
2008.
[BW12]
N. Bergeron and D. T. Wise. A boundary criterion for cubulation. Amer.
J. Math., 134(3):843–859, 2012.
[CCV07] R. Charney, J. Crisp, and K. Vogtmann. Automorphisms of 2dimensional right-angled Artin groups. Geom. Topol., 11:2227–2264,
2007.
[CD95a]
R. Charney and M. W. Davis. Finite K(π, 1)s for Artin groups. In
Prospects in topology (Princeton, NJ, 1994), volume 138 of Ann. of
Math. Stud., pages 110–124. Princeton Univ. Press, Princeton, NJ, 1995.
[CD95b] R. Charney and M. W. Davis. The K(π, 1)-problem for hyperplane complements associated to infinite reflection groups. J. Amer. Math. Soc.,
8(3):597–627, 1995.
[CF12]
R. Charney and M. Farber. Random groups arising as graph products.
Algebr. Geom. Topol, 12(2):979–995, 2012.
[Cha]
R. Charney. Problems related to Artin groups. Preprint, http://
people.brandeis.edu/~charney/papers/Artin_probs.pdf.
[Cha07]
R. Charney. An introduction to right-angled Artin groups. Geometriae
Dedicata, 125(1):141–158, 2007.
[CK00]
C. B. Croke and B. Kleiner. Spaces with nonpositive curvature and their
ideal boundaries. Topology, 39(3):549–556, 2000.
[CS11]
P. Caprace and M. Sageev. Rank rigidity for CAT (0) cube complexes.
Geom. Funct. Anal., 21(4):851–891, 2011.
[Dav98]
M. W. Davis. Buildings are CAT(0). In Geometry and cohomology in
group theory (Durham, 1994), volume 252 of London Math. Soc. Lecture
Note Ser., pages 108–123. 1998.
GROUPS QUASI-ISOMETRIC TO RAAG’S
[Day12]
83
M. B. Day. Finiteness of outer automorphism groups of random rightangled Artin groups. Algebr. Geom. Topol., 12(3):1553–1583, 2012.
[Dro87]
C. Droms. Isomorphisms of graph groups. Proceedings of the American
Mathematical Society, 100(3):407–408, 1987.
[Dun85]
M. J. Dunwoody. The accessibility of finitely presented groups. Inventiones mathematicae, 81(3):449–457, 1985.
[GMRS98] R. Gitik, M. Mitra, E. Rips, and M. Sageev. Widths of subgroups.
Transactions of the American Mathematical Society, 350(1):321–329,
1998.
[Gro81a] M. Gromov. Groups of polynomial growth and expanding maps (with
an appendix by jacques Tits). Publications Mathématiques de l’IHÉS,
53:53–78, 1981.
[Gro81b] M. Gromov. Hyperbolic manifolds, groups and actions. In Riemann
surfaces and related topics: Proceedings of the 1978 Stony Brook Conference (State Univ. New York, Stony Brook, N.Y., 1978), volume 97 of
Ann. of Math. Stud., pages 183–213. Princeton Univ. Press, Princeton,
N.J., 1981.
[Gro87]
M. Gromov. Hyperbolic groups. In Essays in group theory, volume 8 of
Math. Sci. Res. Inst. Publ., pages 75–263. Springer, New York, 1987.
[Hag08]
F. Haglund. Finite index subgroups of graph products. Geometriae Dedicata, 135(1):167–209, 2008.
[Hag14]
M. F. Hagen. Cocompactly cubulated crystallographic groups. J. Lond.
Math. Soc. (2), 90(1):140–166, 2014.
[HP]
M. F. Hagen and P. Przytycki. Cocompactly cubulated graph manifolds.
arXiv:1310.1309.
[HP15]
Mark F Hagen and Piotr Przytycki. Cocompactly cubulated graph manifolds. Israel Journal of Mathematics, 207(1):377–394, 2015.
[Hua14a] J. Huang. Quasi-isometry rigidity of right-angled Artin groups I: the
finite out case. arXiv preprint arXiv:1410.8512, 2014.
[Hua14b] J. Huang. Top dimensional quasiflats in CAT(0) cube complexes.
arXiv:1410.8195, 2014.
[HW08]
F. Haglund and D. T. Wise. Special cube complexes. Geom. Funct.
Anal., 17(5):1551–1620, 2008.
[HW14]
G. C. Hruska and D. T. Wise. Finiteness properties of cubulated groups.
Compositio Mathematica, 150(03):453–506, 2014.
[JŚ01]
T. Januszkiewicz and J. Światkowski. Commensurability of graph products. Algebr. Geom. Topol., 1:587–603 (electronic), 2001.
[KK13]
Sang-hyun Kim and T. Koberda. Embedability between right-angled
Artin groups. Geometry & Topology, 17(1):493–530, 2013.
[KKL98] M. Kapovich, B. Kleiner, and B. Leeb. Quasi-isometries and the de
Rham decomposition. Topology, 37(6):1193–1211, 1998.
[KL97a]
M. Kapovich and B. Leeb. Quasi-isometries preserve the geometric decomposition of Haken manifolds. Inventiones mathematicae,
128(2):393–416, 1997.
[KL97b]
B. Kleiner and B. Leeb. Rigidity of quasi-isometries for symmetric
spaces and Euclidean buildings. Comptes Rendus de l’Académie des
Sciences-Series I-Mathematics, 324(6):639–643, 1997.
84
[KL01]
JINGYIN HUANG AND BRUCE KLEINER
B. Kleiner and B. Leeb. Groups quasi-isometric to symmetric spaces.
Communications in analysis and geometry, 9(2):239–260, 2001.
[KM98]
M. Kapovich and J. J Millson. On representation varieties of Artin
groups, projective arrangements and the fundamental groups of smooth
complex algebraic varieties. Publications Mathématiques de l’Institut
des Hautes Études Scientifiques, 88(1):5–95, 1998.
[KM12]
J. Kahn and V. Markovic. Immersing almost geodesic surfaces in a
closed hyperbolic three manifold. Ann. of Math. (2), 175(3):1127–1190,
2012.
[Lau95]
M. R Laurence. A generating set for the automorphism group of a
graph group. Journal of the London Mathematical Society, 52(2):318–
334, 1995.
[Lee95]
B. Leeb. 3-manifolds with(out) metrics of nonpositive curvature. Invent.
Math., 122(2):277–289, 1995.
[MSW03] L. Mosher, M. Sageev, and K. Whyte. Quasi-actions on trees I. Bounded
valence. Annals of mathematics, pages 115–164, 2003.
[Pan89]
P. Pansu. Métriques de Carnot-Carathéodory et quasiisométries des espaces symétriques de rang un. Annals of Mathematics, pages 1–60, 1989.
[PW02]
P. Papasoglu and K. Whyte. Quasi-isometries between groups with infinitely many ends. Commentarii Mathematici Helvetici, 77(1):133–144,
2002.
[Ron09]
M. Ronan. Lectures on Buildings: Updated and Revised. University of
Chicago Press, 2009.
[Sag95]
M. Sageev. Ends of group pairs and non-positively curved cube complexes. Proc. London Math. Soc. (3), 71(3):585–617, 1995.
[Sag12]
M. Sageev. CAT(0) cube complexes and groups. IAS/Park City Mathematics Series, 2012.
[Ser89]
H. Servatius. Automorphisms of graph groups. Journal of Algebra,
126(1):34–60, 1989.
[SS96]
P. Scott and G. A Swarup. An algebraic annulus theorem. Mathematical
Sciences Research Inst., 1996.
[Sta68]
J. R. Stallings. On torsion-free groups with infinitely many ends. Annals
of Mathematics, pages 312–334, 1968.
[Sul81]
D. Sullivan. On the ergodic theory at infinity of an arbitrary discrete
group of hyperbolic motions. In Riemann surfaces and related topics:
Proceedings of the 1978 Stony Brook Conference (State Univ. New York,
Stony Brook, N.Y., 1978), volume 97 of Ann. of Math. Stud., pages 465–
496. Princeton Univ. Press, Princeton, N.J., 1981.
[Tuk86]
P. Tukia. On quasiconformal groups. J. Analyse Math., 46:318–346,
1986.
[Why10] K. Whyte. Coarse bundles. arXiv preprint arXiv:1006.3347, 2010.
[Wis96]
D. T. Wise. Non-positively curved squared complexes aperiodic tilings
and non-residually finite groups. Princeton University, 1996.
[Wis11]
D. T. Wise. The structure of groups with a quasiconvex hierarchy, 2011.
| 4math.GR
|
An information-theoretic approach for selecting arms
in clinical trials
arXiv:1708.02426v2 [math.ST] 14 Mar 2018
Pavel Mozgunov and Thomas Jaki
Department of Mathematics and Statistics,
Lancaster University, Lancaster, UK
Abstract
The question of selecting the “best” amongst different choices is a common
problem in statistics. In drug development, our motivating setting, the question
becomes, for example: which treatment gives the best response rate. Motivated
by a recent development in the theory of context-dependent information measures,
we propose an experimental design based on a simple and intuitive criterion to
govern arm selection in an experiment with multinomial outcomes. The criterion
leads to reliable selection of the correct arm without any parametric or monotonicity assumptions. The asymptotic properties of the design are studied for different
allocation rules and the small sample size behaviour is evaluated in simulations in
the context of Phase I and Phase II clinical trials. We compare the proposed design
to currently used alternatives and discuss its practical implementation.
Keywords: Multinomial Outcomes; Dose Finding; Experimental Design; Information Gain; Weighted Differential Entropy
1
Introduction
Over the past decades a variety of different methods for clinical trials aiming to select the
“optimal” arm (e.g. dose, treatment, combination of treatments,...) have been proposed
in the literature (O’Quigley and others, 1990; Wages and others, 2011; Magirr and others,
2012; Villar and others, 2015; Lee and others, 2016). Given m arms, the aims of Phase
I and Phase II clinical trials are often to identify the target arm (TA) corresponding to
the toxicity probability closest to the target 0 < γt < 1 and/or the efficacy probability
closest to the target 0 < γe < 1. Despite the similar problem formulation for Phase I
(evaluating toxicity) and Phase II (evaluating efficacy) trials, quite different approaches
are generally utilized.
In Phase I dose-escalation trials, model-based designs assuming a monotonic dosetoxicity relationship are shown to have good operating characteristics in the setting of a
single cytotoxic drug (Iasonos and others, 2016). The ability to find the TA using these
1
methods is, however, rather limited if the assumption of monotonicity is not satisfied
(Shen and O’Quigley, 1996). While this is not common for cytotoxic drugs, the uncertainty about toxicity and efficacy orderings holds for drugs combinations, dose-schedules
and molecularly targeted drugs (Wages and others, 2011; Lagarde and others, 2015).
To overcome the issue of an unknown ordering in the context of Phase I trials, some
specialised approaches for combination and dose-schedule trials have been proposed (e.g.
Thall and others, 2003; Wages and others, 2011; Guo and others, 2016). The common
features of the majority of novel Phase I methods relaxing the monotonicity assumption
is either relying on a parametric model or on explicit order of toxicity/efficacy. While
such methods allow to borrow information across treatment arms, they might fail to
find the TA if the model or ordering is misspecified. There is also growing interest in
advanced trials with a large number of potential orderings where specifying all of them
(or a corresponding parametric model) is not feasible (see e.g. Mozgunov and Jaki, 2018,
for an example). In addition, more complex outcomes than a simple binary endpoint
are becoming more frequent in dose-finding trials (see e.g. Yuan and others, 2007; Lee
and others, 2010, 2017) as they can carry more information about the drug’s mechanism
of action. Despite this, methods for studies with complex outcomes that do not require
monotonicity or a complex model are sparse to date.
Thinking more broadly about selecting one or more arms during a trial (the main
objective of many Phase II studies), different methods that consider arms being independent have been proposed (see e.g. Stallard and Todd, 2003; Koenig and others, 2008;
Whitehead and Jaki, 2009; Magirr and others, 2012). Williamson and others (2016) have
recently advocated designs maximising the expected number of responses in small populations trials. As a result, adaptive randomisation methods and optimal multi-arm Bandit
(MAB) approaches are starting to be considered more commonly in Phase II clinical
trials. Although, MAB designs outperform other well-established methods of randomization (e.g. fixed randomization) in terms of expected number of successes, they can suffer
low statistical power for testing comparative hypotheses. This problem corresponds to
the ‘exploration vs exploitation‘ trade-off (Azriel and others, 2011) and some rule-based
modifications have been proposed archive a better balance of the two objectives (see e.g.
Villar and others, 2015; Williamson and others, 2016). However, the majority of MAB
approaches consider univariate binary response only and hone in on the arm with the
largest effect by default and cannot be applied to clinical trials which aim to select the
TA corresponding to the target probability γe where γ is often between (0.7, 1). On the
other hand, model-based alternatives suffer from the problem of model misspecification.
Although, some of these challenges can be overcome by methods such as MCPmod (Bretz
and others, 2005), we believe that flexible alternatives that can be potentially used in the
described settings are required.
Overall, the research problems described above can be considered as the general issue
of correct identification of the TA whose response probability is closest to the percentile
0 < γ < 1 or equivalently in the multidimensional case, whose characteristics are closest
to the vector γ ∈ Sd where Sd is a d-dimensional unit simplex defined in (2.1). We
propose a general experimental design for studies with multinomial outcomes to solve this
generic problem. Based on developments in the information theory of context-dependent
2
measures of information (Belis and Guiasu, 1968; Kelbert and Mozgunov, 2015; Kelbert
and others, 2016), we derive a criterion which governs arm selection in the experiment.
The criterion is based on the maximisation of the information gain when considering
an experiment with a particular interest in arms whose response probabilities are in
the neighbourhood of γ. Recently other designs using the information gain principle
have been proposed (see e.g. Barrett, 2016; Kim and Gillen, 2016). In contrast to these
methods, the proposed approach allows incorporation of the context of the outcomes
(e.g. avoid high toxicity or low efficacy) in the information measures themselves. This
is achieved by assigning a greater ”weight” to the information obtained about arms
with characteristics close to the desired level. Another difference to the majority of
information-theoretic approaches is that the design is based on the so-called ”patients’
gain” which allocates each new patient to the treatment that is considered best while
taking into account the uncertainty about the estimates for each arm. This leads to
fulfilling of statistical goals of the experiment under the ethical constraints.
The proposed criterion is not restricted to a particular model and can be used, for
examples, to govern selection within traditional parametric designs. However, motivated
by relaxing parametric and monotonicity assumptions, we demonstrate that good operating characteristics of the design can be achieved without employing these assumptions.
For the special case of a complex combination-schedule clinical trial the proposed design
has already been shown to be superior compared to currently employed methods and to
be practically applicable to an ongoing clinical trial (Mozgunov and Jaki, 2018). In this
work, we generalise the approach for an arbitrary number of outcomes and general class
of ”weight” functions, study the asymptotic behaviour of the design and compare the
performance to the currently used method in contexts of Phase I and Phase II trials.
The remainder of the paper is organized as follows: derivations of the criterion and
assignment rules are given in Section 2. The asymptotic behaviour is studied in Section
3. Section 4 presents illustrative examples of the design together with a comparisons to
alternative methods. We conclude with a discussion in Section 5.
2
Methods
The novel design is based on the maximisation of the information gain in the experiment
with an interval of the specific interest (neighbourhood of γ). Below, we derive an explicit
formula for the information gain in the context of a trial with multinomial outcomes.
2.1
Information-theoretic concepts
Consider a discrete random variable taking one of d values and a corresponding random
probability vector Z = Z (1) , Z (2) , . . . , Z (d) ∈ Sd defined on a unit simplex
d
S
= {Z : Z
(1)
> 0, Z
(2)
> 0, . . . , Z
(d)
> 0;
d
X
Z (i) = 1}.
(2.1)
i=1
T
Assume that Z has a prior Dirichlet distribution Dir(v + J) where v = v (1) , . . . , v (d) ∈
P
Rd+ , di=1 v (i) = β and J is a d-dimensional unit vector. After n realizations of a discrete
3
random variable in which x(i) outcomes of i are observed, i = 1, . . . , d, the random vector
Zn has a Dirichlet posterior distribution with density function
Qd
d
(i)
Y
(i) (i)
+ v (i) + 1)
1
(i) x +v
i=1 Γ(x
fn (p|x) =
, B(x + v + J) =
p
Pd
B(x + v + J) i=1
(i)
(i)
Γ
+ v + 1)
i=1 (x
(2.2)
(1)
(1)
Pd
Pd (i)
(d) T
(d)
(i)
(i)
where p = p , . . . , p
, x = x , . . . , x , i=1 x = n, 0 < p < 1, i=1 p = 1
and B(x + v + J) is the Beta-function and Γ(x) is the Gamma-function.
T
Let α = α(1) , . . . , α(d) ∈ Sd be the vector in the neighbourhood of which fn concentrates as n → ∞. A classic question of interest in this setting is to estimate the
probability vector, α. The information required to answer the estimation question can
be measured by the Shannon differential entropy of fn (Cover and Thomas, 2012)
Z
h(fn ) = −
fn (p|x)logfn (p|x)dp
(2.3)
Sd
with convention 0 log 0 = 0. The classic formulation of the estimation question, however,
does not take into account the fact that an investigator would like to find the target
arm (TA) having pre-specified characteristics γ = γ (1) , . . . , γ (d) ∈ Sd . It does not also
reflect that one would like to have more precise estimation about the vector α for those
arms only which have characteristics close to γ. This is a consequence of the fact that
the classic information measures do not depend on the nature of the outcomes p, but
on the probability of the corresponding event f (p) and therefore are called context-free.
While it gives the notion of information great flexibility which explains its successful
application in various fields, the context-free nature might be considered as a drawback
in many application areas as it would be demonstrated below.
To take into account the context of the experiment and the nature of the outcomes p,
one can consider an estimation experiment with ”sensitive” area (i.e. the neighbourhood
of γ). The information required in such an experiment can be measured by the weighted
Shannon differential entropy (Belis and Guiasu, 1968; Clim, 2008; Kelbert and Mozgunov,
2015; Kelbert and others, 2016) of fn with a positive weight function φn (p)
Z
φn
h (fn ) = −
φn (p)fn (p|x)logfn (p|x)dp.
(2.4)
Sd
The crucial difference between the information measures given in Equation (2.3) and
Equation (2.4) is the weight function, φn (p), which emphasizes the interest in the neighbourhood of γ rather than on the whole Sd . It reflects that the information about the
probability vector which lies in the neighbourhood of γ is more valuable in the experiment.
Due to the limited sample size in an actual studies, an investigator is typically interested in answering the question: Which arm has an associated probability vector closest
to γ while ensuring accurate estimation of the probability vector for the TA only. For
this question, the information gain from considering the experiment with sensitive area
equals to
∆n = h(fn ) − hφn (fn ).
(2.5)
4
Following the information gain approach, the first term in the equation above is the
information in a classic experiment using context-free measure, while the second (novel)
term is the information when the context of events is taken into account. Alternatively, ∆n
can be considered as an average amount of the additional statistical information required
when considering the context-dependent estimation problem instead of the traditional
one.
The information gain in Equation (2.5) requires specification of a weight function
which defines the ”value” of the information in different areas of the simplex Sd . To
track the influence of the weight function explicitly we consider a weight function in the
Dirichlet form:
d
Y
γ (i) nκ
φn (p) = C(x, γ, n)
p(i)
(2.6)
i=1
where κ ∈ (0, 1) is a parameter and C(x, γ, n) is a constant which is chosen to satisfy the
R
normalization condition Sd φn (p)fn (p|x)dp = 1. The parameter κ is restricted to the unit
R
interval to ensure asymptotically unbiased estimates of the vector α: limn→∞ Sd pφn (p)fn (p)dp =
α. This emphasises the interest in the identification of the TA for the small and moderate
sample sizes typical for many applications. The asymptotic behaviour of the information
gain, ∆n , for the family of weight functions (2.6) is studied in Theorem 1.
Theorem 1. Let h(fn ) and hφn (fn ) be the standard and weighted differential entropies
(i)
of (2.2) with weight function (2.6). Let limn→∞ x n(n) = α(i) for i = 1, 2, . . . , d and
Pd
(i)
= n, then
i=1 x
1
1
∆n = O
as n → ∞ if κ < ;
1−2κ
n
2
!
2
d
1
1 X γ (i)
1
2κ−1
∆n = −
−1 n
+ ω(α, γ, κ, n) + O
as n → ∞ if κ ≥
(i)
η(1−κ)−κ
2 i=1 α
n
2
where
ω(α, γ, κ, n) =
η
X
(−1)j−1
j=3
j
njκ−j+1
j
d
X
γ (i)
i=1
j−1
(α(i) )
!
−1
and η = b(1 − κ)−1 c
Proof. The proof is given in the Appendix.
The information gain, ∆n , tends to 0 for κ < 1/2 which implies that assigning a
value of information with rate less than 1/2 is insufficient to emphasize the importance
of the context of the study. However, the limit is non-zero for κ ≥ 1/2. Following
the conventional information gain approach, one would like to make a decision which
maximises the statistical information in the experiment. The information gain ∆n is
always non-positive and for any fixed n its asymptotics achieves the maximum value 0
at the point α(i) = γ (i) , i = 1, . . . , d (all constants are cancelled out). Indeed, Theorem
1 implies that when maximising the information gain ∆n , one tends to collect more
information about the arm which has characteristics α close to the target γ. To keep the
trackable solution which can be easily interpreted in applications, we construct the arm
5
selection criterion using the leading term of the asymptotic expression for ∆n in Theorem
1:
!
d
(i) 2
X
γ
1
δ (κ) (α, γ) :=
− 1 n2κ−1 .
(2.7)
2 i=1 α(i)
Note that maximising the leading term of the information gain asymptotics is equivalent to minimising δ (κ) (α, γ). Equation (2.7) can be considered as the measure of the
divergence between α and γ and the criterion which governs the selection such that the
information gain is maximised. The criterion (2.7) is intuitive as it reflects explicitly the
fact that an investigator tends to collect more information about the arm with probability vector close to γ, and also shares some desirable properties. Clearly, δ (κ) (·) ≥ 0
and δ (κ) (·) = 0 iff α = γ for all κ and n. The boundary values α(i) = 0, i = 1, . . . , d
correspond to infinite values of δ (κ) (α, γ) which is advocated by Aitchison (1992) as one
of the important properties for functions defined on simplex Sd . We construct the design
based on the selection criterion (2.7) below.
2.2
Selection criterion
Consider a discrete set of m arms, A1 , . . . , Am , associated with probability vectors α1 , . . . , αm
and n1 , . . . , nm observations. Arm Aj is optimal if it satisfies δ (κ) (αj , γ) = inf i=1,...,m δ (κ) (αi , γ).
(κ)
To estimate δ (κ) (αi , γ) a random variable δ̃ni ≡ δ (κ) (Zni , γ) is introduced where Zni has
(κ)
(κ)
the Dirichlet distribution given in (2.2). Let us fix an arm Zni ≡ Zn and denote δ̃ni ≡ δ̃n .
It is a known that a Dirichlet random variable (after appropriate normalization) weakly
converges to a multivariate normal distribution. In fact, a stronger result can be shown
R
(x)
using the Kullback-Leibler distance D(f || g) = R f (x)log fg(x)
dx where g and f are
probability density functions.
e n = Σ−1/2 (Zn − α) be a random variable with pdf fen where pdf of Zn
Theorem 2. Let Z
P
(i)
is given in (2.2) with limn→∞ x n(n) = α(i) for i = 1, 2, . . . , d, di=1 x(i) = n and where
(i)
(i)
)
Σ is a d-dimensional square matrix with elements Σ[ij] = α (1−α
if i = j and Σ[ij] =
n
α(i) α(j)
− n
if i 6= j. Let Z be the multivariate Gaussian random variable MN (0, Id−1 )
(Id−1 is the (d − 1)-dimensional
unit matrix) with pdf ϕ and the differential entropy
d−1
1
h(ϕ) = 2 log (2πe)
. Then the Kullback-Leibler divergence of ϕ from f˜n tends to 0
e n weakly converges to Z.
as n → ∞ which implies that Z
Proof. The proof in given in the Appendix.
Using Theorem 2 the following result can be obtained for the proposed criterion.
(κ)
Theorem 3. Under the assumptions of Theorem 2, let δ̃n = δ (κ) (Zn , γ), ∇δ (κ) (z, γ) =
h (κ)
iT
(κ)
∂δ (z,γ)
∂δ (κ) (z,γ)
,
.
.
.
,
, δ̄n = Σ̄−1/2 δ (κ) (Zn , γ) − δ (κ) (α, γ) where Σ̄ = ∇T
α Σ∇α
∂z (1)
∂z (d)
and ∇α ≡ ∇δ (κ) (z, γ) evaluated at z = α. Let Z̄ be a standard Gaussian RV. Then,
(κ)
(κ)
(κ)
limn→∞ Eδ̃n = δ (κ) (α, γ), limn→∞ Vδ̃n = 0, and δ̄n weakly convergences to Z̄.
A single summary statistics for δ (κ) (Zn , γ) is needed to select the most promising
arm in the sequential experiment. While a Bayesian estimator can be used, we will
6
(κ)
focus on the intuitively clear and simple ‘plug-in‘ estimator, δ̂ (κ) (p̂n , γ) ≡ δ̂n with p̂n =
(i) +v (i)
(1)
(i)
(d)
(i)
[p̂n , . . . , p̂n , . . . , p̂n ] and p̂n = xn+β
(i) , i = 1, . . . , d, the mode of the posterior Dirichlet
distribution. The estimator for the arm Aj takes the form
!
d
(i) 2
X
γ
1
− 1 nj2κ−1 , j = 1, 2, . . . , m.
δ̂n(κ)
= δ (κ) (p̂nj , γ) =
(2.8)
j
2 i=1 p̂(i)
nj
(κ)
(κ)
(κ)
Note that by Theorem 3 for any ε > 0 limnj →∞ P δ̃nj ∈ [δ̂nj − ε, δ̂nj + ε] = 1. The
statistics (2.8) is used to govern the selection among arms during the experiment. Note
that the estimator above requires a vector of prior parameters vj , j = 1, . . . , m to start
the experiment. This choice implies an initial ordering in which an investigator would
like to test the arms before the data is available.
2.3
Specific Assignment rules
The criterion (2.7) summarizes the arm’s characteristics and can be applied to different
types of sequential experiments. We consider two assignment rules: Rule I which randomizes between arms and Rule II which selects the “best” arm. These rules follow the
setting of the motivating clinical trials: Rule II is widely used in Phase I trials where the
randomization to all doses is not ethical or in the typical MAB setting where the primary
goal is to maximize the number of successes. Note that randomization (when is ethical)
allows to decrease the probability of identifying a suboptimal arm (Thall and Wathen,
2007).
2.3.1
Rule I: Randomization
Under Rule I, the arm used next in the experiment is randomised with probabilities w̃j ≡
(κ)
1/δ̃nj
Pm
(κ) ,
i=1 1/δ̃ni
(κ)
j = 1, . . . , m and from Theorem 3 wj = limn1 ,n2 ,...,nm →∞ E (w̃j ) =
1/δnj
Pm
(κ) .
i=1 1/δni
When no observations have yet been collected, the procedure randomizes according to
(κ)
the criterion based on the prior distribution alone, δ̂βj , j = 1, . . . , m. Then, given nj
observations, xj outcomes for arm Aj , j = 1, . . . , m and using the ‘plug-in‘ estimator (2.8),
(κ)
arm Aj is selected with probability ŵj = 1 if δ̂nj = 0 and with probability
(κ)
1/δ̂nj
ŵj = Pm
(κ)
i=1 1/δ̂ni
if δ̂n(κ)
> 0, i = 1, . . . , m.
i
(2.9)
The method proceeds until N observations are attained. The arm Aj satisfying
(1/2)
δ̂Nj
=
inf
i=1,...,m
(0.5)
δ̂Ni .
(2.10)
is adopted for the final recommendation, where Ni is a total number of observation
on an arm Ai , i = 1, 2, . . . , m. The value κ = 0.5 in (2.10) is used so that the final
recommendation is not penalized by the sample size.
7
2.3.2
Rule II: Select the best
Let N be a total sample size and begin with the experiment with the arm that minimizes
(κ)
δ̂βj , j = 1, . . . , m. Given nj observations, xj outcomes for the arm Aj , j = 1, . . . , m and
(κ)
(κ)
using the ‘plug-in‘ estimator, an arm Aj is selected if it satisfies δ̂nj = inf i=1,...,m δ̂ni .
The method proceeds until the total number of N is attained. Again, we adopt Aj as in
(2.10) for the final recommendation.
2.4
Criterion in the context of clinical trials
Further in the examples (Section 4) we apply the novel selection criterion to Phase I
and Phase II clinical trials. In this case the arms are the different treatments (doses,
combinations, schedules,...) and the goal is to find the treatment corresponding to specific
toxicity (efficacy) characteristics. As the proposed information gain and corresponding
selection criterion tends to assign the next patients to the best estimated TA during the
trial, the criterion based on ∆n is a patient’s gain (also known as ‘best intention‘ ) criterion
as classified by Whitehead and Williamson (1998). The balance in the ‘exploration vs
exploitation‘ trade-off is tuned by the term nj2κ−1 reflecting the penalty on the number of
observations on the same arm (i.e. many observations on one arm would favour selection
of other arms). This implies that the design will keep selecting a specific arm only, if
the corresponding estimate, α is close γ. As the trial progresses the design requires
an increasing level of confidence that the selected arm is the TA. Clearly, κ = 1/2
corresponds to no penalty and is of particular interest in trials with small sample sizes
while larger values of κ > 1/2 correspond to a greater interest in the statistical power of
the experiment. Importantly, the first term in Equation (2.7) guarantees that the vast
majority of patients will be assigned to the TA even for κ > 1/2.
As many Phase I and Phase II clinical trials consider a binary endpoint d = 2 (toxicity:yes/no or response:yes/no), we focus on this case in the examples. Then, (2.2) reduces
to the Beta-distribution and the proposed criterion takes the form
(i)
δ̂n(κ)
(α, γ) =
j
1 (p̂nj − γ)2 2κ−1
n
2 p̂n(i)j (1 − p̂n(i)j ) j
(i)
(2.11)
which is a normalized distance between p̂nj and γ. Note that this is not equivalent to the
(i)
Euclidean distance (p̂nj − γ)2 which is a commonly used criterion for selection (see e.g.
Shen and O’Quigley, 1996). As the criterion tends to assign patients to the TA, it still
has the Euclidean distance term in the nominator. However, it also takes into account
the uncertainty in the denominator which is a variance of the probability of a binary
event. The denominator can be also considered as a penalty term which ”drives away”
(i)
(i)
the allocation from the bounds (p̂nj = 0 or p̂nj = 1) as the boundary values correspond
to the infinite value of the criterion. Note that, as the maximum variance of the binary
(i)
(i)
probability is achieved at p̂nj = 0.5, so the criterion favours greater values of p̂nj which
can be unethical if the objective of a study is to control the risk of toxicity. We will study
whether the construction of the criterion creates any practical limitation in the context
of Phase I clinical trial in Section 4.2.
8
3
Asymptotic behaviour
Considering the asymptotic behaviour of a procedure ensures that the experimental design
becomes more accurate as a sample size grows (Azriel and others, 2011). Recall that the
goal of the sequential experiment is to find an arm j which corresponds to the minimum
(κ)
value δj ≡ δ(αj , γ) among all arms using random variables δ̃i = δ (κ) (Zni , γ), i =
(κ)
1, . . . , m. Denote the arm to be selected by ν = arg mini=1,...,m δ̃i . Below, we consider the
risk-adjusted average approach (Polley and Cheung, 2008) and the probability of correct
P
selection (PCS) (Cheung, 2013) AN = m1 m
j=1 Pπj (ν = j) where Pπj is the probability
computed under the vector πj = [α1,j , . . . , αm,j ]T , 1 ≤ j ≤ m, which assumes that αj,j
corresponds to the TA. The design is consistent if limN →∞ AN = ∞. The main result of
this section is formulated in the following Theorem.
Theorem 4. Let us consider the experimental design with a selection criteria based on
(κ)
δ̃ni , m arms and corresponding true probabilities vectors αi , i = 1, . . . , m. Then,
(a) The design is consistent under Rule I (2.3.1) for κ ≥ 0.5.
(b) The design is consistent under Rule II (2.3.2) for κ > 0.5.
Proof. The proof is given in the Appendix.
4
Examples
In this section the performance of the proposed experimental design is studied in the
context of Phase I and Phase II clinical trials under different assignment rules. We will
refer to our proposal as the Weighted Entropy (WE) design (WEI under Rule I and
WEII under Rule II) and compare its performance to several well-established alternative
approaches. All computations have been conducted using R (R Core Team, 2015).
4.1
4.1.1
Phase II clinical trial
Setting
Consider a Phase II clinical trial whose primary endpoint is a binary measure of efficacy
(e.g. response to treatment). The goals of the study are (i) to find the most effective
treatment and (ii) to treat as many patients as possible on the optimal treatment. Clearly,
Rule I is preferable for the first goal and Rule II for the second one. We consider two
hypothetical trials, each with m = 4 treatments, investigated by Villar and others (2015)
for Multi-Arm Bandit models (MAB). We compare the performance of the proposed
approach to the MAB approach based on the Gittins index (Gittins and Jones, 1979),
which is the optimal design in terms of maximising expected number of successes (ENS),
and to fixed and equal randomization (FR) which is best in terms of the statistical power.
Trial 1 investigates N1 = 423 and the true efficacy probabilities are (0.3, 0.3, 0.3, 0.5)
while Trial 2 considers N2 = 80 and the scenario (0.3, 0.4, 0.5, 0.6). Following the original
application we consider the hypothesis H0 : p0 ≥ pi for i = 1, 2, 3 with the family-wise
error rate calculated at p0 = . . . = p3 = 0.3, where p0 corresponds to the control treatment
9
efficacy probability. The Dunnett test (Dunnett, 1984) is used for the hypothesis testing
in the FR setting. The hypothesis testing for MAB and WE design is performed using
an adjusted Fisher’s exact test (Agresti, 1992). The adjustment chooses the cutoff values
to achieve the same type-I error as the FR. The Bonferroni correction is used for MAB
and WE designs to correct for the multiple testing and the family-wise error rate is set
to be less or equal to 5%. Characteristics of interest are (i) the type-I error rate (α),
(ii) statistical power (1 − η), (iii) the expected number of successes (ENS) and (iv) the
average proportion of patients on the optimal treatment (p∗ ).
The WE design requires specification of the target value which can be specified in
many clinical trial by clinicians. This can be defined as the maximum efficacy that they
expect to see for the particular diseases. Below we consider the most challenging setting
in which no target value is specified and the goal is to simply maximise the number of
successes (as in MAB). To achieve this we set the target probability to a value close
to 1. The target γ = 0.999 is used which corresponds to the aim ”to find the arm
with the highest efficacy probability”. The vector of the prior mode probabilities p(0) =
[0.99, 0.99, 0.99, 0.99]T is chosen to reflect no prior knowledge about which arm has the
highest success probability and that each treatment is considered as highly efficacious until
data suggests otherwise. This choice of prior reflects the equipoise principle (Djulbegovic
and others, 2000). We choose β0 = 5 to ensure enough observations on the control arm
and β1 = β2 = β3 = 2 to reflect no prior knowledge for competing arms. We fix κ = 0.5
for WEI and use different values of κ for WEII .
4.1.2
Results
The trade-off between the expected number of successes (ENS) and the statistical power
for different values of the penalty parameter κ under Rule II is illustrated in Figure 1. In
both trials, greater values of κ correspond to greater power and lower ENS as the increase
in penalty tends to more diverse allocations. The exception is κ ∈ (0.5, 0.55) in Trial 1
where the inconsistency for κ = 0.5 leads to locking-in on the suboptimal treatment.
We choose two values of κ for the subsequent comparison. These choices correspond to
(i) high ENS, but unacceptable power (dashed line) and (ii) slightly reduced ENS, but
higher power (dotted line).
The operating characteristics of considered designs in Trial 1 are given in Table 1.
Under the null hypothesis, the performance of all methods is similar and the type-I error
is controlled. Under the alternative hypotheses, the WEII design with κ = 0.55 performs
comparably to the MAB in terms of the ENS, but yields almost 10% points increase in
power. Nevertheless, it has unacceptable low statistical power which can be increased by
using higher values of the penalty parameter (κ = 0.65). It leads to an increase in the
power from 0.53 to 0.86 at the cost of the slight (≈ 4%) decrease in the ENS. In fact
WEII then has comparable power to the FR, while treating almost 40 more patients on
the superior treatment. Another way to increase the statistical power is to use WEI for
which both the associated power and the ENS is higher than for the FR.
The operating characteristics of the designs for Trial 2 with fewer patients and a
linear increasing trend is given in the Table 2. Under the null hypothesis, all designs
10
Trial 2
38
175
39
40
ENS
185
ENS
41
195
42
Trial 1
0.60
0.70
0.80
0.50
0.60
0.70
0.80
0.5
0.6
0.7
0.5
0.6
0.7
0.8
0.9
1.0
0.2
Power
0.6
0.0
0.2
0.4
Power
0.8
0.4
1.0
0.50
κ
0.8
0.9
κ
Figure 1: ENS and power (fixed cutoff value) for the WE design under the Rule II for different
κ. Dashed lines correspond to the values chosen for the subsequent study. Results are based on
104 replications
Table 1: Operating characteristics of the WE design under the Rule I (WEI ), under the
Rule II (WEII ) for different κ (in brackets), MAB design and FR in Trial 1 with N = 423
under the null and alternative hypothesises. Results are based on 104 replication.
H0 : p0 = p1 = p2 = p3 = 0.3
H1 : p0 = p1 = p2 = 0.3, p3 = 0.5
Method
∗
α
p (s.e)
ENS(s.e.)
(1 − η)
p∗ (s.e.)
ENS (s.e.)
MAB
0.05 0.25 (0.18) 126.68 (9.4)
0.43
0.83 (0.10) 198.25 (13.7)
FR
0.05 0.25 (0.02) 126.91 (9.4)
0.82
0.25 (0.02) 147.91 (9.6)
WEI (0.50) 0.05 0.24 (0.05) 126.84 (9.5)
0.88
0.39 (0.06) 159.90 (11.0)
WEII (0.55) 0.05 0.21 (0.20) 126.89 (9.4)
0.55
0.83 (0.18) 197.13 (17.8)
WEII (0.65) 0.05 0.23 (0.13) 126.86 (9.4)
0.87
0.74 (0.10) 189.26 (13.7)
perform similarly and type-I errors are controlled at the 5% level. Under the alternative
hypothesis, the MAB and WEII with κ = 0.55, again, yield the highest (and similar) ENS
among all alternatives, but also low statistical power. The WEI or increased κ for WEII
result in a considerable power increase. Both designs have a greater (or similar) power
and result in more ENS than the FR.
11
Table 2: Operating characteristics of the WE design under the Rule I (WEI ), under the
Rule II (WEII ) for different κ (in brackets), MAB design and FR in Trial 2 with N = 80
under the null and alternative hypothesises. Results are based on 104 replication.
H0 : p0 = p1 = p2 = p3 = 0.3
H1 : pi = 0.3 + 0.1i, i = 0, 1, 2, 3
Method
∗
α
p (s.e)
ENS(s.e.)
(1 − η)
p∗ (s.e.)
ENS (s.e.)
MAB
0.00 0.25 (0.13) 23.97 (4.10)
0.01
0.49 (0.21) 41.60 (5.4)
FR
0.05 0.25 (0.04) 24.02 (4.10)
0.50
0.25 (0.04) 35.98 (4.3)
WEI (0.50) 0.05 0.23 (0.07) 23.92 (4.11)
0.59
0.33 (0.10) 37.55 (4.7)
WEII (0.55) 0.01 0.20 (0.15) 24.01 (4.10)
0.11
0.50 (0.27) 40.72 (5.9)
WEII (0.65) 0.05 0.22 (0.12) 23.96 (4.08)
0.52
0.47 (0.21) 40.19 (5.4)
Overall, WE designs can perform comparably to the optimal MAB design in terms of
the ENS, but with greater statistical power for both large and small sample sizes. They
have similar statistical power to the FR, but with the considerably greater ENS. The ENS
and power trade-off can be tuned via the built-in parameter κ. Although, some modification to the MAB designs were proposed (e.g. see Villar and others, 2015) to prevent
the low statistical power, the majority of those are ruled-based. The proposed approach
allows to avoid any algorithm-based rules and keeps the procedure fully adaptive. Additionally, the computation of the Gittens index for the MAB design is not trivial, requires
special attention and is widely discussed in the literature (e.g. see Villar and others, 2015,
and reference there in). Some of them require calibration and can be computationally
intensive. In contrast, the proposed criterion is extremely simple and easy to compute.
While the proposed designs are compared for the target γ = 0.999, similar performance is
obtained for the problem of seeking an arm associated with a given response probability
(γ ∈ (0.7, 1)). This is, for example, of interest when seeking the effective dose 80 (ED80),
the dose for which 80% of subjects respond to the treatment.
4.2
4.2.1
Phase I clinical trial
Setting
To study the WE design in the context of Phase I clinical trials, let us consider m = 7
arms, N = 20 patients and the arm selection allowed after each patient. The goal is
to find the arm (which could be combination, schedule or combination-schedule) with a
toxicity probability closest to γ = 0.25. In these studies randomization to all arms is
not ethical for safety reasons and therefore, Rule II and κ = 0.5 are used. We would
like to emphasize that we do not consider the classic dose-escalation problem in which
the doses can be put according to increasing toxicity and focus on the setting in which
clinicians cannot put arms according to increasing toxicity (as e.g. often in scheduling
trials). While clinicians will be able to provide a presumed ordering of the arms, this
order might be misspecified.
We consider scenarios in which the prior order chosen by clinicians is either correct
or misspecified. The scenarios with correctly specified ordering have a monotonic armtoxicity relationship and the scenarios with misspecified ordering have a non-monotonic
12
relationship. The investigated scenarios are shown in Figure 2 and include a variety of
monotonic and non-monotonic shapes as well as one setting with highly toxic arms only.
3
4
5
6
7
0.8
0.6
0.2
0.0
1
2
3
4
5
6
7
1
4
5
Scenario 5
Scenario 6
4
5
6
7
7
6
7
0.6
0.0
0.2
0.4
Toxicity
0.6
0.0
0.2
0.4
Toxicity
0.6
0.4
3
6
0.8
Scenario 4
0.8
Arm
0.2
2
3
Arm
0.0
1
2
Arm
0.8
2
0.4
Toxicity
0.6
0.0
0.2
0.4
Toxicity
0.6
0.4
0.0
0.2
Toxicity
1
Toxicity
Scenario 3
0.8
Scenario 2
0.8
Scenario 1
1
2
3
Arm
4
Arm
5
6
7
1
2
3
4
5
Arm
Figure 2: Toxicity scenarios. The optimal arm is marked by a triangle and the maximum
tolerated toxicity γ = 0.25 is marked by dashed horizontal line. The monotonic scenarios
(1-3, 6) correspond to correctly prespecified ordering of arms according to increasing
toxicity and non-monotonic scenarios (4-5) to misspecified ordering of arms.
It is assumed that limited information about treatments is available and a linear
increase in the toxicity probabilities is expected such that p̂β1 < p̂β2 < . . . < p̂β7 . For
safety reasons, the trial is required to start at d1 . An ‘operational‘ prior, that is a prior
that gives good operating characteristics under different scenarios, is calibrated. Details
on the calibration are given in the Appendix. The resulting prior uses β = 1 and prior
toxicity risk modes of p̂ = [0.25, 0.3, 0.35, 0.4, 0.45, 0.5, 0.55]T . For the simulation study
below, the penalty parameter is fixed at κ = 0.5 due to the small sample size.
The WE design is compared to common Phase I designs. Specifically the Bayesian
Continual Reassessment Method (CRM) (O’Quigley and others, 1990) and Escalation
with Overdose Control (EWOC) (Babb and others, 1998) are considered as methods
that assume a monotonic toxicity relationship. Additionally, the partial ordering CRM
(POCRM, Wages and others, 2011) which relaxes the monotonicity assumption is also
considered. It uses the original CRM design with additional randomization among a
pre-specified set of orderings. In our comparison we consider only correct orderings
(Table 3) to allow for the best possible performance of the POCRM under the evaluated
scenarios. The same prior toxicity probabilities p̂ and a rough prior distribution of the
13
Table 3: Orderings for POCRM.
Order
1 (1,2,3,4,5,6,7)
2 (1,2,7,6,5,4,3)
3 (7,6,5,4,1,2,3)
model parameters were chosen for the model-based alternatives. Finally, we include the
non-parametric optimal benchmark (O’Quigley and others, 2002) which provides the best
theoretical performance if the patients’ complete toxicity profiles are known.
The main characteristics to consider are: (i) the proportional of correct selections
and (ii) the average number of toxic responses. The bcrm package (Sweeting and others,
2013) is used for CRM and EWOC and the pocrm package is used for POCRM. For the
proposed method and the non-parametric optimal design one-million-fold simulations
are used while 100,000-fold simulations are used for the model-based methods due to
computational constraints.
4.2.2
Safety constraint
For ethical reasons an escalation procedure should be planned so that only few patients
are assigned to highly toxic treatments which is typically achieved by the use of a safety
constraint. The majority of existing safety constraints are based on the assumption of
monotonicity and hence are not suitable for the proposed design. We adopt the following
R1
safety constraint instead. The treatment Aj is safe if after n patients γ ∗ fnj (p)dp ≤ θnj
where γ ∗ is an upper toxicity threshold, θnj controls the overdosing probability and fnj
is the posterior Beta distribution for the toxicity probability. Note that the overdosing
threshold θnj changes as the trial progresses. It should be a decreasing function of n
with θ0 = 1 to give a possibility to test all the treatments (if data suggests so) and
θf inal ≤ 0.3 to ensure that the final recommendation is safe. As an illustration, the
linear non-increasing θn = max(1 − rn, θf inal ) is used with r > 0. We have calibrated
the parameters of the safety constraints (details in the Appendix) and used γ ∗ = 0.45
and r = 0.035 in the simulations. Similar safety constraints were incorporated in the
model-based methods.
4.2.3
Results
The simulation results in monotonic scenarios 1-3 are given in Table 4. The WE design
performs comparably to the CRM and POCRM designs and recommends the correct
treatment with the probability nearly 0.25 and 0.30 in scenario 1 and 2. In scenario 1 the
WE design underestimates the target treatment and recommends a less toxic treatment
more often due to the safety constraint. Despite that, the performance of all methods is
not far from the non-parametric optimal benchmark which shows that the detection of
the target treatment is quite challenging. Proportions of terminations are close to 0 and
the average number of toxic responses is largely the same. For the EWOC, the level of
the target treatment is underestimated in both scenarios.
14
Table 4: The operating characteristics of the WE, CRM, POCRM and EWOC designs.
‘Term‘, ‘T‘ and N̄ correspond to the termination proportion, the average number of toxic
responses and the average number of patients, respectively. The most likely recommendation is in bold, the actual target regimen is in italics.
d1
Scenario
Optimal
WE
CRM
POCRM
EWOC
0.06
0.92
7.03
2.66
2.69
8.06
Scenario
Optimal
WE
CRM
POCRM
EWOC
0.10
6.05
16.78
17.24
14.98
28.72
Scenario
Optimal
WE
CRM
POCRM
EWOC
0.15
29.87
38.07
37.47
33.57
51.00
d2
d3
d4
d5
d6
Scenario 1. Linear response
0.12
0.15
0.18
0.24
0.36
9.12
10.60 14.44 31.54 27.50
14.72 23.33 30.11 23.34
1.39
7.21
14.17 20.58 26.62 15.95
11.25 22.30 15.73 22.60 20.62
13.80 20.30 23.70 20.20
9.62
Scenario 2. Logistic shape
0.18
0.25
0.32
0.50
0.68
29.03 30.12 28.27
6.48
0.05
26.43 29.54 22.51
3.76
0.11
25.88 28.70 19.37
6.24
0.56
27.32 27.89 18.50
6.70
1.04
27.66 24.32 13.65
4.21
0.00
Scenario 3. J shape
0.20
0.50
0.55
0.60
0.65
58.31 10.0 2
1.69
0.11
0.00
44.65 6.59
3.44
1.48
0.28
37.85 17.41
2.92
0.36
0.07
37.76 13.27
2.55
0.54
1.33
26.11 11.01
0.88
0.13
0.00
d7
Term T
N̄
0.40
5.87
0.05
12.53
4.60
3.88
0.1
0.3
0.2
0.4
3.36
4.38
4.94
3.76
20.0
19.9
20.0
18.8
0.82
0.00
0.00
0.04
1.86
0.00
0.9
1.9
1.7
1.4
5.23
4.84
5.54
3.27
20.0
19.7
20.0
18.0
0.70
0.00
0.02
0.00
6.04
0.00
5.5
3.9
4.9
10.9
5.94
5.10
6.06
3.60
19.8
19.4
19.8
16.8
In scenario 3, the WE design shows a better performance than the model-based alternatives with nearly 45% of correct recommendations against about 40% for CRM and
POCRM. The safety constraint allows to prevent the recommendation of highly toxic
treatments and controls the total number of toxic responses. Again, the EWOC underestimates the target therapy, but results only in 3 toxic responses compared to 5 for the
CRM and 6 toxicities for the WE and the POCRM. As expected, methods that relax
monotonic assumption result in more toxic responses that monotonicity based designs.
The results for non-monotonic (Scenarios 4-5) and unsafe (Scenario 6) cases are given
in Table 5. As expected, the designs based on the monotonicity assumption are not able
to find the target treatment in non-monotonic settings. Comparing other designs, the
WE design has a substantial advantage. It finds the correct arm with the probability
nearly 0.28 compared to 0.20 for the POCRM while exposing nearly the same number of
patients to toxic treatments. The safety constraint allows recommendation of the optimal
treatment even in non-monotonic scenarios where the target treatment lies beyond the
toxic treatments (d3 − d4 ).
Considering Scenario 6, the WE design terminates earlier with probability 0.8 and
performs similar to the POCRM and EWOC. It outperforms the CRM which recommends
15
Table 5: The operating characteristics of the WE, CRM, POCRM and EWOC designs.
‘Term‘, ‘T‘ and N̄ correspond to the termination proportion, the average number of toxic
responses and the average number of patients, respectively. The most likely recommendation is in bold, the actual target regimen is in intalics.
d1
d2
0.05
0.88
14.11
4.26
2.87
7.18
0.10
7.36
19.13
19.90
11.39
24.90
Scenario 0.35
Optimal 16.18
WE
15.57
CRM
47.41
POCRM 16.81
EWOC 30.75
0.40
3.01
12.65
2.51
5.98
1.26
Scenario
Optimal
WE
CRM
POCRM
EWOC
0.55
16.35
5.53
0.32
0.57
0.00
Scenario
Optimal
WE
CRM
POCRM
EWOC
0.50
80.53
13.63
32.24
13.18
16.17
d3
d4
d5
d6
Scenario 4. Inverted-U shape
0.40
0.35
0.25
0.15
19.12 18.96 38.47 13.64
11.77 18.27 27.90 8.50
17.70
6.31
2.84
3.00
11.75
9.32
19.11 33.94
18.60
3.79
2.52
3.79
Scenario 5. Inverted-U shape
0.40
0.35
0.25
0.15
3.01
16.18 39.46 18.65
13.31 18.27 27.92 8.90
0.97
0.48
0.72
0.40
5.66
12.42 20.10 23.13
0.78
0.47
0.47
0.31
Scenario 6. Unsafe
0.60
0.65
0.70
0.75
3.10
0.02
0.00
0.00
2.45
0.88
0.27
0.06
0.08
0.00
0.00
0.00
0.12
0.04
0.01
2.06
0.12
0.00
0.00
0.00
Term T
N̄
0.12
1.57
0.23
46.10
11.62
30.60
0.1
0.3
0.2
6.6
4.26
3.26
4.29
2.73
20.0
19.9
20.0
18.9
0.10
3.51
0.58
30.10
10.23
9.78
9.9
27.3
9.7
56.2
5.81
4.27
5.14
3.30
19.7
16.0
19.5
11.0
0.80
0.00
0.00
0.00
0.08
0.00
77.2
67.4
83.9
83.7
8.02
5.33
7.12
3.07
14.2
10.3
12.5
6.1
d7
a highly toxic arm with a larger probability (32.56% against 19.16%). However, methods
that relax the monotonicity assumption result in more toxic responses and require more
patients on average to come to the termination conclusion. While the CRM and the
EWOC require 5 and 6 patients only, it takes nearly 14 and 13 patients for the WE
and POCRM as they explore all arms before concluding none is safe. This, however,
would not be considered as a severe drawbacks of these methods as in many application
clinicians would stop the trial for safety reasons based on extensive patients profile.
Summarizing, the proposed design performs comparably to the model-based approaches
in monotonic settings and clearly outperforms them in non-monotonic ones. Importantly,
despite the denominator of the proposed criterion which might favour slightly more toxic
arms, the design does not cause any practical concerns as soon as the allocation is restricted to the safety set. In other words, the criterion allocates the patients to the best
estimated TA taking into account information about the uncertainty in the estimates and
about the safety set. Importantly, the time-varying safety constraint achieves the goals
motivated by the ethical concerns while not preventing the target treatment selection in
safe and non-monotonic scenarios. One can conclude that the design is ethical and can
be applied in practice.
16
5
Discussion
In this work, we propose a family of criteria for selecting the best arm in experiments with
multinomial outcomes. The novel criterion leads to accurate selection without the need
for parametric or monotonicity assumptions. The fundamental property of the criterion
is the infinite penalization of the bounds which was argued to be a crucial property for
a parameter defined on the restricted space (Aitchison, 1992). This property drives the
allocation away from the bounds of a restricted space to the neighbourhood of the target
value. The consistency conditions of the proposed design and exact rate for the special
case of binary outcomes are obtained. It is shown how one can benefit from the proposed
design in Phase I and Phase II clinical trials. The proposal was demonstrated to have a
comparable or better characteristics than other alternatives. It preserves flexibility and
allows to tailor the design parameters in light of the investigation goal. Additionally,
the design is computationally simple and a large set of simulation can be performed in a
feasible time.
Despite clinical trials being used as the main motivation throughout, the design can
be applied to a wide range of problems of a similar nature. For example, applications
where the MAB approach has found application: online advertising, portfolio design,
queuing and communication networks, etc. (see Gittins and others, 2011, and references
there in). On top of that, the proposed design can be used in more general problems
of percentile estimation rather than the identification of the highest success probability.
It is important to emphasize that the derived selection criterion can be also applied in
conjunction with a parametric models which also expands its possible applications. In
fact, the parameters can be estimated by any desirable method and then ‘pluged-in‘ in
the criterion which preserves its properties.
Acknowledgments
This project has received funding from the European Union’s Horizon 2020 research and
innovation programme under the Marie Sklodowska-Curie grant agreement No 633567
and, in part, from from Prof Jaki’s Senior Research Fellowship (NIHR-SRF-2015-08-001)
supported by the National Institute for Health Research. The views expressed in this
publication are those of the authors and not necessarily those of the NHS, the National
Institute for Health Research or the Department of Health.
References
Agresti, Alan. (1992). A survey of exact inference for contingency tables. Stat. Sc.,
131–153.
Aitchison, John. (1992). On criteria for measures of compositional difference. Mathematical Geology 24(4), 365–379.
Azriel, David, Mandel, Micha and Rinott, Yosef. (2011). The treatment versus
17
experimentation dilemma in dose finding studies. Journal of Statistical Planning and
Inference 141(8), 2759–2768.
Babb, James, Rogatko, André and Zacks, Shelemyahu. (1998). Cancer
phase I clinical trials: efficient dose escalation with overdose control. Statistics in
Medicine 17(10), 1103–1120.
Barrett, James E. (2016). Information-adaptive clinical trials: a selective recruitment
design. Journal of the Royal Statistical Society: Series C 65(5), 797–808.
Belis, Mariana and Guiasu, Silviu. (1968). A quantitative-qualitative measure
of information in cybernetic systems (corresp.). IEEE Transactions on Information
Theory 14(4), 593–594.
Bretz, Frank, Pinheiro, José C and Branson, Michael. (2005). Combining multiple comparisons and modeling techniques in dose-response studies. Biometrics 61(3),
738–748.
Cheung, Ying Kuen. (2013). Sample size formulae for the bayesian continual reassessment method. Clinical Trials 10(6), 852–861.
Clim, A. (2008). Weighted entropy with application. Anal. Bucuresti, Mat., Anul 57,
223–231.
Cover, Thomas M and Thomas, Joy A. (2012). Elements of information theory.
John Wiley.
Csiszar, Imre and Körner, János. (2011). Information theory: coding theorems for
discrete memoryless systems. Cambridge University Press.
Djulbegovic, Benjamin, Lacevic, Mensura, Cantor, Alan, Fields, Karen K,
Bennett, Charles L, Adams, Jared R, Kuderer, Nicole M and Lyman,
Gary H. (2000). The uncertainty principle and industry-sponsored research. The
Lancet 356(9230), 635–638.
Dunnett, Charles W. (1984). Selection of the best treatment in comparison to a
control with an application to a medical trial. Design of experiments: Ranking and
selection, 47–66.
Gittins, John, Glazebrook, Kevin and Weber, Richard. (2011). Multi-armed
bandit allocation indices. John Wiley & Sons.
Gittins, John C and Jones, David M. (1979). A dynamic allocation index for the
discounted multiarmed bandit problem. Biometrika, 561–565.
Gradshteyn, Izrail Solomonovich and Ryzhik, Iosif Moiseevich. (2014). Table
of integrals, series, and products. Academic press.
18
Guo, Beibei, Li, Yisheng and Yuan, Ying. (2016). A dose–schedule finding design
for phase I–II clinical trials. Journal of the Royal Statistical Society: Series C 65(2),
259–272.
Iasonos, Alexia, Wages, Nolan A, Conaway, Mark R, Cheung, Ken, Yuan,
Ying and O’Quigley, John. (2016). Dimension of model parameter space and
operating characteristics in adaptive dose-finding studies. Statistics in Medicine 35(21),
3760–3775.
Kelbert, Mark and Mozgunov, Pavel. (2015). Asymptotic behaviour of the
weighted renyi, tsallis and fisher entropies in a bayesian problem. Eurasian Mathematical Journal 6(2), 6–17.
Kelbert, Mark, Yurii, Suhov, Izabella, Stuhl and Yasaei, Sekeh Salimeh.
(2016). Basic inequalities for weighted entropies. Aequationes Mathematicae, 1–32.
Kim, Steven B and Gillen, Daniel L. (2016). A bayesian adaptive dose-finding
algorithm for balancing individual-and population-level ethics in phase i clinical trials.
Sequential Analysis 35(4), 423–439.
Koenig, Franz, Brannath, Werner, Bretz, Frank and Posch, Martin.
(2008). Adaptive dunnett tests for treatment selection. Statistics in Medicine 27(10),
1612–1625.
Lagarde, Fabien, Beausoleil, Claire, Belcher, Scott M, Belzunces, Luc P,
Emond, Claude, Guerbet, Michel and Rousselle, Christophe. (2015). Nonmonotonic dose-response relationships and endocrine disruptors: a qualitative method
of assessment. Environmental Health 14(1), 13.
Lee, Juhee, Thall, Peter F, Ji, Yuan and Müller, Peter. (2016). A decisiontheoretic phase i–ii design for ordinal outcomes in two cycles. Biostatistics 17(2),
304–319.
Lee, Shing M, Cheng, Bin and Cheung, Ying Kuen. (2010). Continual reassessment method with multiple toxicity constraints. Biostatistics 12(2), 386–398.
Lee, Shing M, Ursino, Moreno, Cheung, Ying Kuen and Zohar, Sarah.
(2017). Dose-finding designs for cumulative toxicities using multiple constraints. Biostatistics.
Magirr, Dominic, Jaki, Thomas, Whitehead, John and others. (2012). A
generalized dunnett test for multi-arm clinical studies with treatment selection.
Biometrika 99(2), 494.
Mozgunov, Pavel and Jaki, Thomas. (2018). An information-theoretic phase
i/ii design for molecularly targeted agents that does not require an assumption of
monotonicity. Major Revision in Journal of the Royal Statistical Society: Series C ,
arXiv:1803.04397.
19
O’Quigley, John, Paoletti, Xavier and Maccario, Jean. (2002).
parametric optimal design in dose finding studies. Biostatistics 3(1), 51–56.
Non-
O’Quigley, John, Pepe, Margaret and Fisher, Lloyd. (1990). Continual reassessment method: a practical design for phase 1 clinical trials in cancer. Biometrics,
33–48.
Polley, Mei-Yin and Cheung, Ying Kuen. (2008). Two-stage designs for dosefinding trials with a biologic endpoint using stepwise tests. Biometrics 64(1), 232–241.
R Core Team. (2015). R: A Language and Environment for Statistical Computing. R
Foundation for Statistical Computing, Vienna, Austria.
Resnick, Sidney I. (2013). A probability path. Springer Science & Business Media.
Shen, Larry Z and O’Quigley, John. (1996). Consistency of continual reassessment
method under model misspecification. Biometrika 83(2), 395–405.
Stallard, Nigel and Todd, Susan. (2003). Sequential designs for phase III clinical
trials incorporating treatment selection. Statistics in Medicine 22(5), 689–703.
Sweeting, Michael, Mander, Adrian, Sabin, Tony and others. (2013). Bcrm:
Bayesian continual reassessment method designs for phase I dose-finding trials. Journal
of Statistical Software 54(1), 1–26.
Thall, Peter F, Millikan, Randall E, Mueller, Peter and Lee, Sang-Joon.
(2003). Dose-Finding with Two Agents in Phase I Oncology Trials. Biometrics 59(3),
487–496.
Thall, Peter F and Wathen, J Kyle. (2007). Practical bayesian adaptive randomisation in clinical trials. European Journal of Cancer 43(5), 859–866.
Villar, Sofı́a S, Bowden, Jack and Wason, James. (2015). Multi-armed bandit
models for the optimal design of clinical trials: benefits and challenges. Statistical
science: a review journal of the Institute of Mathematical Statistics 30(2), 199.
Wages, Nolan A, Conaway, Mark R and O’Quigley, John. (2011). Continual
reassessment method for partial ordering. Biometrics 67(4), 1555–1563.
Whitehead, John and Jaki, Thomas. (2009). One-and two-stage design proposals
for a phase II trial comparing three active treatments with control using an ordered
categorical endpoint. Statistics in Medicine 28(5), 828–847.
Whitehead, John and Williamson, David. (1998). Bayesian decision procedures
based on logistic regression models for dose-finding studies. J. of Biopharm. Stat. 8(3),
445–467.
Williamson, S Faye, Jacko, Peter, Villar, Sofı́a S and Jaki, Thomas. (2016).
A bayesian adaptive design for clinical trials in rare diseases. Comp. Statistics & Data
Analysis.
20
Yuan, Z, Chappell, R and Bailey, H. (2007). The continual reassessment method
for multiple toxicity grades: A bayesian quasi-likelihood approach. Biometrics 63(1),
173–179.
21
Appendices
A
Proofs
Proof of Theorem 2.1
Proof. The problem reduces to computing integrals of the following forms
Z
x(i)
fn (p)dp = x(i) ψ x(i) + 1 − ψ (n + d) , i = 1, 2, . . . , d,
log p(i)
Sd
Z
log p(i)
x(i)
φn (p)fn (p)dp = x(i) ψ x(i) + γ (i) nκ + 1 − ψ (n + nκ + d) , i = 1, 2, . . . , d
Sd
d
where ψ(x) = dx
logΓ(x) is the digamma function. Using the asymptotics of the digamma
function (Gradshteyn and Ryzhik, 2014), Taylor series expansion and simplifying terms,
the results immediately follows.
Proof of Theorem 2.2
Proof. Following the proof of Theorem 2.1, one can obtain that h (fn ) = log B (x + v + J)+
P
nψ (n + d) − di=1 x(i) ψ x(i) + 1 . As n → ∞
1
1
h (fn ) = log (2πe)d−1 + log
2
2
Qd
i=1 α
nd−1
(i)
1
+O
.
n
Using that h fen = h (fn ) + log |det Σ−1/2 | (Cover and Thomas, 2012) where det(A) is
Qd
α(i)
1
a determinant of the matrix A and log |det Σ−1/2 | = − 21 log i=1
+
O
as n → ∞.
d−1
n
n
Then,
Z
e
D(fn || ϕ) = −h fen −
fen (p) log ϕ(p)dp
Sd
1
1
1
= − log (2πe)d−1 + log (2π)d−1 +
2
2
2
1
= O
,
n
Z X
d−1
Sd i=1
p
(i) 2
fen (p)dp + O
1
n
R Pd−1 (i) 2
as Sd i=1
p
fen (p)dp = d − 1 + O n1 is the sum of the second moments. Using Pinsker’s inequality (Csiszar and Körner, 2011) it implies the convergence in total
variation (Cover and Thomas, 2012) which implies the weak convergence.
Proof of Theorem 3.1
Proof. (a) Rule I. Under Rule I the proportion of observations on each arm converges
to a constant (Theorem 2.3). Therefore, it is initially assumed that it is fixed and the
probability measures below are conditional on the allocation proportion.
22
(1/2)
We start from from κ = 1/2 and adopt notation δ̃ni
δ̃k+1 } and Ck,k+1 ≡ {δ̃k > δ̃k+1 } we find that
≡ δ̃i . Denoting C̄k,k+1 ≡ {δ̃k <
{ν = k} ⇔ {C̄k,1 ∩ C̄k,2 ∩ . . . ∩ C̄k,k−1 ∩ C̄k,k+1 ∩ . . . ∩ C̄k,m } ⇔ {∩m
i=1,i6=k C̄k,i }.
Using DeMorgan’s law and Boole’s inequality (Resnick, 2013), one can obtain
m
m
X
X
P(ν = k) = 1 − P ∪m
C
≥
1
−
P
(C
)
=
1
−
P
δ̃
>
δ̃
k,i
k
i
i=1,i6=k k,i
i=1,i6=k
i=1,i6=k
where
P δ̃k > δ̃i = P
(A.1)
δk − δi
δ̃i − δi − δ̃k + δk
<
Σk,i
Σk,i
!
≈Φ
δk − δi
Σk,i
(A.2)
with Φ(.) denoting the distribution function of a standard normal random variable,
Σk,i = (Σk + Σi )1/2 and Σk the variance corresponding to arm k as in the Theorem
2.3. As arms k and i are independent, there are two independent random variables in
the left-hand side of the second term in (A.2) and each of them converges to a Gaussian
random variables (Theorem 2.3). Therefore, the sum converges to a standard Gaussian
random variable after an appropriate normalization. Consequently, for j = 1, . . . , m
Pπj (ν = j) ≥ 1 −
m
X
i=1,i6=j
Φ
δj,j − δi,j
Σj,i
(A.3)
By the construction of πj , δj,j − δi,j < 0. The number of observations on each arm Ni is
proportional to the total sample size N under Rule I: Nj ' wj N and Nj ' wj−1 N where
an ' bn means that limn→∞ abnn = 1. Thus,
√
δm,m − δi,j
'c N
Σm,i
(A.4)
where c is a negative constant. Plugging-in terms in the accuracy formula, we obtain
that limN →∞ AN ≥ 1.
For κ > 1/2, the probability of the final selection in the experiment is still given by
(A.1) for κ = 1/2 as the penalty term is not taken into account for the final recommendation. Then, the only difference is the number of observation on each arm, proportional
to the total number of patients Nj ' lj (N )N with lj depending
√ on N . This results in a
different constant c < 0 in (A.4), but in the unchanged rate N due to the same rate in
both nominator and denominator in Equation (2.8) with respect to N .
Binary outcomes
While the asymptotic result (A.2) is given in terms of Σk,i , it can be writtenq
explicitly.
p
αj (1−αj )
0
In the special case of binary outcomes d = 2, Σk,i = σk2 + σi2 and σj = |δi |
,
nj
j = i, k with
(κ)
∂δi (z, γ)
(γ − α)(γ(2α − 1) − α)
(κ) 0
= δi =
.
∂z
α2 (1 − α)2
z=α
23
Therefore,
Φ
δk − δi
σk,i
√
nk ni (δk − δi )
=Φ
!
p
nk αi (1 − αi )(δi 0 )2 + ni αk (1 − αk )(δk 0 )2
.
From the expression above, the rate obtained in (A.4) is explicit.
(b) Rule II. Consider κ = 12 . The design based on this measure and on its point estimate
is inconsistent as it does not guarantee an infinite number of patients on all arms. We use
an example with two arms only with the arm A1 being the optimal. Suppose that prior
parameters are specified such that δ̂β2 δ̂β1 and δ2 δ̂β1 , so A2 is selected initially.
While the number of observations on the arm 2 increases and the estimate δ̂2 approaches
the true value δ2 (Theorem 2.3), the estimate δ̂1 remains unchanged. One can find prior
values δ̂β1 δ2 such that A1 is never selected, because the point estimate δ̂2 would not go
below δ̂β1 . Consequently, the selection would get stuck at the suboptimal arm regardless
of further outcomes. So, the number of patients on both arms does not tend to infinity
as N → ∞.
For 21 < κ < 1, let Njk be the indicator function such that
(
(κ)
(κ)
1 with probability P(δ̃j (t) = mini δ̃i (t))
Nj (t) =
(κ)
(κ)
0 with probability 1 − P(δ̃j (t) = mini δ̃i (t))
(κ)
where δ̃j (t) is a random variable corresponding to the posterior density function after
P
t observations in the experiment. Let nj (t) = ti=1 Nj (i) be the number of observations
on arm j up to the moment t. We then obtain
E(nj (t)) =
t
X
ENj (u) =
t
X
i
u=1
u=1
(κ)
(κ)
P(δ̃j (u) = min δ̃i (u)).
(κ)
(κ)
(κ)
Note, that P(δ̃j (u) = mini δ̃i (u)) has already been studied in (a). The mean of δ̃j (u)
associated with an arm Aj to be selected is an increasing polynomial with respect to N .
The probability to be the minimum decreases for j and increases for i = 1, . . . , j − 1, j +
1, . . . , m. It follows that the probability of being selected is not a monotonic function and
lim
t→∞
t
X
(κ)
(κ)
P(δ̃j (u) = min δ̃i (u)) = ∞.
i
u=1
The final selection is the arm satisfying (2.9). Consequently, the number of observations
on each arm tends to infinity and we obtain that limN →∞ AN = 1 using the arguments
of (a).
B
B.1
Parameters calibration for the Phase I clinical trial
Operational prior
A prior for treatment dj can be specified through the mode of the prior distribution,
ν
p̂βj = βjj . To guarantee the procedure to start from d1 we set p̂β1 = γ. As an investigator
24
has the same amount of knowledge about each treatment we set β1 = . . . = β7 = β. Larger
values of β and the rate of increase correspond to the more conservative escalation scheme
as an investigator needs more observations on each particular treatment to escalate.
Similarly, a greater differences in prior toxicity probabilities would correspond to the
more conservative scheme as well, because it would require more evidence to escalate.
Therefore, one can expect a set prior parameters that would lead to a similar PCS.
The investigation of these parameter influence on the operational characteristics of the
proposed approach is given below.
We consider six different scenarios with different location of the target (Figure 3). For
simplicity, we would assume a priori that toxicity increases linearly between treatments.
Given m = 7, we set the difference between prior toxicities on the d1 and d7 and, then,
interpolate the linear curve for the rest. We would define step = p̂β7 − p̂β1 . Then, we
vary values of step and β for each scenarios. The PCS for different combination of step
and β is given in the Figure 3.
Figure 3: The PCS for the WE design using N = 20 and different combinations of β
(vertical axis) and step (horizontal axis). Results are based on 106 simulations.
Brighter colours correspond to higher values of the PCS. A conservative prior (top
right corner on the grid) prevents the WE design from the correct recommendation in
upper line graphs scenarios as higher doses can be hardly reached with N = 20. At the
same time, it leads to an accurate selection in scenarios with highly toxic doses (lower
line scenarios). In contrast, less conservative prior results in higher proportion of correct
recommendation in upper line scenarios and worse in lower line ones. Therefore, there
is a trade-off between the ability to investigate higher doses and the desire to prevent
the high number of toxic responses. Therefore, the geometric mean of the PCS over all
scenarios is chosen as the criterion for the operational prior choice. The geometric mean
for different set of parameters is given in the Figure 4.
There is a set of the prior parameter that lead to the same geometric mean of the
25
2.0
35
1.5
30
25
1.0
Beta
20
15
0.5
10
5
0
0.1
0.2
0.3
0.4
0.5
Step
Figure 4: The geometric mean of the proportion of correct recommendations by the
proposed WE method using different set of prior parameters: β (vertical axis) and the
difference between the risk of toxicity on the lowest and highest dose (horizontal axis) in
six scenarios: with the TD at the bottom, in the middle and at the top of the investigated
dose range. 106 simulations are used.
PCS across all scenarios. We choose a prior that carries limited information β = 1 and
fix the rate to maximize the geometric mean among all scenarios. Thus, the following
vector of modes p̂ is chosen
p̂ = [0.25, 0.3, 0.35, 0.4, 0.45, 0.5, 0.55]T .
We also consider the case of N = 25 thought the same set of scenarios to illustrate the
influence of the prior parameter for a larger sample size (Figure 5). There is a a similar
dependence pattern on β and step. However, the set of the equivalent operational prior
parameters is now wider that means the importance of the prior distribution decreases
with the sample size as one would expect.
B.2
Safety constraint
Parameters r and γ ∗ determine the strictness of the safety constraint. Greater values of
r and smaller values of γ ∗ would lead to more conservative escalation. It helps to avoid
high risk in unsafe scenarios, but also to prevent the correct recommendation in flat safe
scenarios. There is a clear trade-off in the choice of these parameters that is precisely
studied below.
Let us consider two extreme scenarios: a linear flat dose response shape with the
target treatment far from the bottom (d5 ) and the scenario with no safe treatment at
all. The operating characteristics of the proposed method in these two scenarios with
different parameters of the safety constraint γ ∗ and r are given in Table 6.
26
Figure 5: The geometric mean of the proportion of correct recommendations by the
proposed WE method using different set of prior parameters: β (vertical axis) and the
difference between the risk of toxicity on the lowest and highest dose (horizontal axis) in
six scenarios. Total sample size N = 25. 106 simulations are used.
Table 6: The operating characteristics of the proposed design in a linear and an unsafe
scenario for different parameters of the safety constraint. The figures in the upper line
of each cell corresponds to termination proportion in the unsafe scenario. The lower cell
entries corresponds to the PCS in the flat linear scenario. The bold figures correspond to
γ ∗ = 0.45 and r = 0.035 which were used in simulations. Results based on 106 replications.
r
0.010 0.015 0.020 0.025 0.030 0.035 0.040 0.045
0.00 0.32 4.32 18.47 36.15 49.06 61.49 75.70
γ ∗ = 0.55
26.47 26.65 26.40 26.05 26.85 25.03 24.10 20.23
0.15 2.50 17.76 38.75 52.74 63.06 74.94 87.22
γ ∗ = 0.50
26.27 26.22 26.53 27.24 25.46 23.30 20.35 17.10
1.13 12.72 35.72 56.49 67.16 77.55 86.53 93.49
γ ∗ = 0.45
26.15 26.02 26.81 25.18 24.26 23.15 18.16 11.05
7.47 37.95 59.49 70.52 80.53 88.32 94.18 97.63
γ ∗ = 0.40
26.04 25.91 24.90 21.98 17.66 17.47 11.05 3.51
33.98 58.22 74.42 84.14 90.52 94.86 97.90 99.20
γ ∗ = 0.35
25.65 24.54 20.45 15.55 13.77 9.21
6.25 0.70
55.51 77.02 87.21 92.99 96.50 98.55 99.37 99.83
γ ∗ = 0.30
24.21 18.09 14.40 11.42 7.13
0.95
0.08 0.04
27
The upper line in each cell corresponds to the proportion of times the proposed method
declares that there is no safe dose when there is actually no safe dose. The lower line
corresponds to the proportion of trials the actual TD25 was recommended in a linear
scenario. The most relaxed safety constraint corresponds to the left upper corner. In this
case no trials are terminated in a highly toxic scenario and the proportion of times the
TD25 is recommended in the linear scenario is high. The right lower corner corresponds
to the strictest safety constraint. In this case near all trials will be terminated when
there is no safe dose, but the method will often not find the TD25 in the linear scenario.
Therefore, the trade-off is to sacrifice the accuracy of the method when the TD25 is far
from the bottom in order to prevent the recommendation of highly toxic dose in unsafe
scenario.
28
| 10math.ST
|
ON THE RESOLUTION OF FAN ALGEBRAS OF PRINCIPAL IDEALS
OVER A NOETHERIAN RING
arXiv:1612.02939v4 [math.AC] 11 Jan 2017
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
Abstract. We construct explicitly a resolution of a fan algebra of principal ideals over a
Noetherian ring for the case when the fan is a proper rational cone in the plane. Under
some mild conditions on the initial data, we show that this resolution is minimal.
1. Introduction
Let R be a commutative ring, m, n ∈ N, p1 , . . . , pn ∈ R, F a fan in Rm such that its
support |F | is a convex set, and f1 , . . . , fn fan linear maps on F (see Definition 1.4 for more
details). With all this data, one can consider the following algebra
X
f (v)
±1
±1
(1)
Bf ,F ,p =
hp11 · . . . · pfnn (v) ixv ⊂ R[x±1
1 , x2 , . . . , xm ],
v∈|F |∩Zm
where x1 , . . . , xm are indeterminates, f = (f1 , . . . , fn ), p = (p1 , . . . , pn ), xv = xv11 · · · xvmm and
|F | ⊂ Rm is the support of the fan. We will call Bf ,F ,p the fan algebra associated to the
data (f, F , p). It is a natural combinatorial object associated to a fan. In this paper, we will
describe the free resolution of Bf ,F ,p as an R-algebra, when m = 2, R is Noetherian, and |F |
is a proper rational cone.
Fan algebras have been introduced in [Mal15] (see also [Mal13, EM15]), but here we extend
the definition further. In the aforementioned papers, it has been noted that they generalize
intersection algebras in the case of principal ideals, an interesting class of algebras in its own
right. The finite generation of intersection algebras has applications to the lengths of Tors
of quotients by powers of ideals, as shown in [F02], and to the asymptotic growth of powers
of ideals, as shown in [CES07].
Before we state the main result of this paper, we will introduce the necessary definitions
and put the fan algebra in context.
Definition 1.1. A convex polyhedral cone in Rm is a convex cone generated by a finite
subset S ⊂ Rm , i.e. a set of the form
(
)
X
σ=
λu · u : λu ≥ 0 .
u∈S
Date: January 12, 2017.
2010 Mathematics Subject Classification. Primary: 13A30 ; Secondary: 05E40,13P10, 13P20 .
Key words and phrases. Fan algebra, convex polyhedral cones, fan linear functions, intersection algebra,
minimal resolutions.
Cortadellas and D’Andrea are supported by the Spanish MEC research project MTM2013-40775-P.
1
2
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
A supporting hyperplane for σ is a hyperplane H in Rm containing the origin, and such
that σ is contained in one of the half-spaces determined by H. The intersection between a
supporting hyperplane H and σ is by definition a face of σ. The convex polyhedral cone σ is
called rational if S can be taken from Zm . A strongly convex polyhedral cone is a convex
polyhedral cone σ with σ ∩ −σ = 0. In what follows, a convex rational polyhedral cone which
is not either the whole space nor contained in a line , will simply be called a proper rational
cone.
We now recall the notion of a fan, commonly used in toric geometry, that is of central
importance in this paper.
Definition 1.2. A fan F in Rm is a finite collection of strongly rational cones σ such that
(1) for all σ ∈ F , each face of σ is also in F .
(2) for all σ1 , σ2 in F , the intersection σ1 ∩ σ2 is a face of each.
The support of F is the union of all σ in F . We will denote this by |F |. In this paper, we
make the additional assumption that the support |F | is also a convex cone.
Example 1.3. A simple example of a fan is {σ1 , σ2 , τ0 , τ1 , τ2 , τ3 } where
σ1 = {(x, y) ∈ (R≥0 )2 : 3x ≥ 2y}, σ2 = {(x, y) ∈ (R≥0 )2 : 3x ≤ 2y},
τ0 = {(x, 0) ∈ (R≥0 )2 }, τ1 = {(x, y) ∈ (R≥0 )2 : 3x = 2y}, τ2 = {(0, y) ∈ (R≥0 )2 }, τ3 = {(0, 0)}.
Its support is (R≥0 )2 .
Definition 1.4. Let F be a fan in Rm . A function f : |F | ∩ Zm → N is called fan linear if
f is N-linear on each face of F and subadditive on the support of the fan, i.e. f (u + v) ≤
f (u) + f (v) for all u, v ∈ |F | ∩ Zm .
Example 1.5. For the fan F from Example 1.3, let f (x, y) = max(3x, 2y), for all (x, y) ∈
N2 = |F | ∩ Z2 . This obviously defines a fan linear map on F .
With these definitions in place, we can define the concept that is central to our paper.
Definition 1.6. Let n, m ≥ 1 be integers, F a fan in Rm , f = {f1 , . . . , fn } a collection of
fan linear maps on F , and I = (I1 , . . . , In ) ideals in a commutative ring R.
The fan algebra associated to the data (f, F , I) is defined as
(2)
Bf ,F ,I :=
X
f (v)
I1 1
±1
±1
. . . Infn (v) xv ⊂ R[x±1
1 , x2 , . . . , xm ],
v∈|F |∩Zm
where x = (x1 , . . . , xm ) are indeterminates.
In the case Ij is the principal ideal hpj i, j = 1, . . . , n, we have that Bf ,F ,I is the fan algebra
Bf ,F ,p defined in (1).
Remark 1.7.
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
3
(1) If |F | = (R≥0 )m and fi = 0, i = 1, . . . , n, it is easy to see that Bf ,F ,I = R[x1 , . . . , xm ],
the ring of polynomials in several variables. If instead we have that |F | = Rm and
±1
the fi′ s identically zero as before, we obtain R[x±1
1 , . . . , xm ], the ring of Laurent
polynomials in m variables.
(2) P
Let I1 , . . . , In be ideals of R. Their intersection algebra is defined as BR (I1 , . . . , In ) :=
r1
r1
rn
rn
r1 ,...,rn ∈N (I1 ∩ . . . ∩ In )x1 · · · xn ⊂ R[x1 , x2 , . . . , xn ]. Intersection algebras of principal ideals in an UFD are fan algebras, as it was shown in [Mal15]. A particularly
interesting case is that of two principal monomial ideals in k[p1 , . . . , pn ], where k
is a field and p1 , . . . , pn indeterminates. More specifically let a = (a1 , . . . , an ), b =
(b1 , . . . , bn ) two nonnegative integer vectors and I1 = hpa i, I2 = hpb i. Then I1r ∩ I2s =
max(a1 r,b1 s)
max(an r,bn s)
hp1
· · · pn
i and
X max(a r,b s)
1
1
n r,bn s)
BR (I1 , I2 ) =
hp1
· · · pmax(a
ixr1 xs2 .
n
r≥0,s≥0
We will denote this algebra as B(a, b). This algebra is toric, normal, Cohen-Macaulay
and of dimension n + 2, as shown in [Mal13, Mal15, EM15].
(3) Set m = n, |F | = (R≥0 )n , and I = (I1 , . . . , In ) ideals in a commutative ring R and
fi : (R≥0 )n ∩ Zn → N the projection on the ith coordinate, i = 1, . . . , n. Then
X
Bf ,F ,I =
I1k1 · · · Inkn xk11 · · · xknn
(k1 ,...,kn )∈Nn
is the multi-Rees algebra of the ideals I1 , . . . , In in R.
(4) Let I be an ideal in R. Then, for the fan and the fan linear map in Example 1.3 and,
respectively, Example 1.5 we have
X
I max(3r,2s) xr1 xs2 ,
(3)
Bf ,F ,I =
(r,s)∈N2
which is the same as BR ((p3 ), (p2 )), when R is an UFD and p a prime element of R.
We will write BR,p (3, 2) for this R-algebra.
Let Bf ,F ,I be a fan algebra. When R is Noetherian, this R-algebra is finitely generated.
Indeed, a set of generators of this algebra over R was given in [Mal15, Theorem 2.3.7] (see
also [Mal13, EM15]) for the case R being a domain and |F | = (R≥0 )m . It is easy to see that
such a proof can be easily extended to the case of a fan F with the properties given above,
and any Noetherian ring R, as we recall here: denote with C1 , . . . , Cℓ the maximal cones of
F , and for i = 1, . . . , ℓ, set Qi = Ci ∩ Zm . This pointed monoid has a unique Hilbert Basis,
which we denote with:
HQi = {vi1 , . . . , vigi } ⊂ Zm .
(4)
f (vij )
As R is Noetherian, for each vij , the ideal I1 1
f (vij )
I1 1
f (vij )
· . . . · Inn
is finitely generated, i.e.
· . . . · Infn (vij ) = hrij1 , . . . , rijhij i
Next result follows straightforwardly from Theorem 2.3.7 in [Mal15].
4
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
Theorem 1.8. With notations and assumptions as above, Bf ,F ,I is generated as an algebra
over R by the set
(5)
{rijh xvij ; i = 1, . . . , ℓ, j = 1, . . . , gi , h = 1, . . . , hij }.
In general (5) is far from being a minimal set of generators of the fan algebra, although
in some cases -like in the intersection algebra of principal ideals, see [Mal13]- it has been
shown to be the case. It is also of interest computing a whole resolution of Bf ,F ,I as an
R-algebra, but very little seems to be known at the present about this ring in general. In
this text, we describe completely the case m = 2, with all the ideals being principal, i.e.
when Ij = hpj i for a suitable pj ∈ R, and |F | =
6 R2 (and hence contained in a half-plane,
since the support of the fan is assumed to be convex by definition). This situation includes
the intersection algebras of principal ideals, hence we generalize and extend the results in
[Mal13]. Our main result given in Theorem 1.9 describes a resolution of this R-algebra. If R
is ∗-local, under some mild conditions on the functions {fi }1≤i≤n and the elements {pi }1≤i≤n ,
we show that the resolution is minimal graded with the induced Z-grading. If in addition
R is a ∗-local ring and all the pi ’s are nonzero divisors in the maximal ideal, then we can
describe combinatorially the Z-graded Betti numbers in terms of the elements of H. Along
the way, we disprove Conjecture 3.2.3 in [Mal13].
1.1. Statement of the main result. From now on, we will work with principal ideals,
and fans F of cones in R2 such that |F | is also convex, and contained in a half-plane. In
R2 we have the advantage that there is a standard orientation for both cones and vectors
which we will be use. So assume w.l.o.g. that C1 , . . . , Cℓ , the maximal cones in F , are sorted
clockwise. In addition, we will also assume that for each i = 1, . . . , ℓ, the elements in the set
HQi = {vi1 , . . . , vigi } which was defined in (4), are also sorted clockwise.
We will say that the family of functions f is strict with respect to the fan F if for each
1 ≤ i < ℓ, there is ki ∈ {1, . . . , n} such that fki is not linear in Ci ∪ Ci+1 . Note that if f
is not strict with respect to F one can take a coarser fan F ′ such that f is also a family of
piece-wise linear functions compatible with F ′ and strict with respect to this new fan, and
moreover from the definition we get
Bf ,F ,p = Bf ,F ′ ,p ,
so we can always assume w.l.o.g. that the set of functions is strict with respect to the fan.
Set H := ∪ℓi=1 HQi , and denote with M the cardinality of this set. Let p1 , . . . , pn ∈ R. In
f (v)
f (v)
the sequel, we will set pf (v) = p11 . . . pnn , for short. For each v ∈ H, let Yv be a new
variable, and set Y = {Yv , v ∈ H}. In light of Theorem 1.8, we have an epimorphism
ϕ0
(6)
R[Y] → Bf ,F ,p
Yv 7→ pf (v) xv
Note that ϕ0 is a Z2 -graded map if we declare deg(Yv ) = v, and deg(r) = (0, 0) for all r ∈ R.
If R is ∗-local with maximal ideal m, then we can see that R[Y] is also ∗-local with maximal
ideal equal to (m, Y) with the degree of a monomial given by the sum of its exponents, while
the elements of R have degree 0. We refer to this grading as the induced Z-grading.
The following is the main result of this paper.
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
5
Theorem 1.9. If |F | is a proper rational cone contained in R2 , then the following is a
Z2 -graded resolution of Bf ,F ,p over R[Y]:
(7)
M −1
M −1
M −1
ϕM −2
ϕ1
ϕ0
ϕ2
ϕ3
0 → R[Y]M −2 → . . . → R[Y]3( 4 ) → R[Y]2( 3 ) → R[Y]( 2 ) → R[Y] → Bf ,F ,p → 0,
M −1
M −1
the maps ϕj : R[Y]j ( j+1 ) → R[Y](j−1)( j ) , j = 1, 2, . . . are defined in (36) for j = 1, (55)
for j = 2 and their recursive construction is shown at in Section 5.
The resolution is minimal graded over R[Y] if R is ∗- local with maximal ideal m, f is
strict with respect to F , none of the pi ’s is a zero divisor in R, and all of them belong to m.
Here, we regard R[Y] as graded with the induced Z-grading.
Recall from Definition 1.1 that a proper rational cone is not either the whole space nor
contained in a line, so in particular we have M ≥ 2. When |F | is a half-line, then clearly we
have that the map ϕ0 of (6) is an isomorphism, and hence the resolution stops there. If |F |
is a line, then Bf ,F ,p is isomorphic to R[Y1 , Y−1]/hY1 Y−1 − pf (1)+f (−1) i, so it resolution is also
not a hard task.
The situation where |F | = R2 seems to be more tricky, and a resolution such as (7) does
not hold anymore in general, see the discussion in Section 6.
Theorem 1.9 follows directly from Theorems 3.4, 4.10, 5.5 and 5.6. The maps ϕj , j =
1, 2, . . . are constructed recursively starting from (36), and (55). We can actually make
explicit the description of all the Betti numbers of this ring when the resolution (7) is
minimal.
We can state an interesting consequence of our main result to intersection algebras. To that
end, let k be a field, p1 , . . . , pn indeterminates, R = k[p1 , . . . , pn ] and a = (a1 , . . . , an ), b =
(b1 , . . . , bn ) two integer vector with positive entries. Consider the intersection algebra B(a, b)
as introduced in Remark 1.7.
As specified in Remark 1.7, the R-algebra B(a, b) is Cohen-Macaulay of dimension n + 2.
To see B(a, b) as a fan algebra, consider the nonnegative quadrant in Z2 and the fan generated
by the rays of slope ai /bi , i = 1, . . . , n. Let us call this fan F . The fan functions are given by
fi (r, s) = max(ai r, bi s), i = 1, . . . , n and these functions are strict with respect to the fan F .
If M equals the cardinality of H, then we know that M is the minimal number of generators
of Bf ,F ,p = B(a, b), as in Theorem 1.9.
Corollary 1.10. Using the notation and assumptions above, B(a, b) is Gorenstein if and
only if M = 3. In this case, a1 = b1 , . . . , an = bn and
B(a, b) =
k[p1 , . . . , pn , A, B, C]
.
(AB − pa11 · · · pann C)
Proof. Note that B(a, b) is a quotient of R[Y] where Y = {Y1 , . . . , YM }. But B(a, b) is
Cohen-Macaulay, so it is a Cohen-Macaulay R[Y]-module of finite projective dimension. By
the Auslander-Buchsbaum theorem, we have that its projective dimension over R[Y] equals
n+M −(n+2) = M −2. Our Theorem 1.9, together with Proposition 3.2 in [H99], shows that
the type of B(a, b) is given the (M − 2)-th Betti number in our resolution,
after localizing
M −1
at (m, Y). The formula for the Betti numbers is β0 = 1, βi = i · i+1 , for i ≥ 1. So, either
6
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
−1
M=
6 2 since the a, b have positive entries, so the type equals βM −2 = (M − 3 + 1) M
=
M −1
M − 2.
Finally, B(a, b) is Gorenstein if and only if its type is 1. So, in this case, M = 3. It is
easy to see that this is equivalent to ai = bi for all i = 1, . . . , n. Moreover, by applying our
techniques or by direct computation, one sees that
k[p1 , . . . , pn , A, B, C]
.
B(a, a) =
(AB − pa11 · · · pann C)
Remark 1.11. This result generalizes Corollary 2.23 in [EM15] and Proposition 2.10 in [ES16].
The paper is organized as follows. Section 2 is expository in nature: it describes an
algorithmic procedure to construct the Hilbert basis for an integer cone in the plane and
outlines the fundamental facts about Gröbner bases needed later. Section 3 is dedicated to
the construction of the presentation ideal of the fan algebra. In Section 4 we study with
detail the map ϕ1 of the resolution, the analysis of the remaining maps and the proof of
the main Theorem being completed in Section 5. The paper concludes with some examples
which illustrate that the situation for when the fan is the whole plane is more complex than
the results presented here, and hence new ideas are needed in order to tackle with this and
more general scenarios.
2. Hilbert and Gröbner Bases
We present here a number of results on Hilbert and Gröbner bases that are essential in
our subsequent sections. This material is generally known to experts, but since we do not
know a good reference for the statements that we need, we include them here for the benefit
of the reader.
2.1. Hilbert bases of two dimensional strongly rational cones. We need a criteria
and an algorithm to construct a Hilbert basis of the monoid of integer points lying in the
cone generated by {(b, a), (b′ , a′ )} ⊂ R2 . Some of the facts below can be found without proof
in [MS], Example 7.19.
First we start with the following easy lemma, whose proof is left to the reader.
Lemma 2.1. Let v, w be two integer vectors such that det(v, w) = 1. Then the Hilbert basis
of the monoid of integer points lying in the cone generated by v, w is {v, w}.
Proposition 2.2. Suppose that the sequence {v1 , . . . , vN } ⊂ Z2 lies in an open half-plane
(i.e. there exists v such that hv, vi i > 0 ∀i = 1, . . . , N), and is such that
a) det(vi+1 , vi ) = 1, for i = 1, . . . , N − 1.
b) For every i = 1, . . . , N −2, vi+1 lies in the interior of the parallelogram whose vertices
are 0, vi , vN , vi + vN .
Then
(1) det(vj , vi ) ≥ 1 for all j > i.
(2) For all j = 1, . . . , N, if we write vj = αj v1 + βj vN with αj , βj ∈ R, then
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
7
• αj , βj ≥ 0 for all j = 1, . . . , N,
• the sequence {αj }j=1,...,N is strictly decreasing.
(3) The family {v1 , . . . , vN } is a Hilbert basis of the monoid generated by the integer
points lying in the convex hull of the positive rays generated by v1 and vN .
(4) There is a unique Hilbert basis of this monoid.
Remark 2.3. Note that condition a) plus the fact that {v1 , . . . , vN } is contained in a halfplane implies that it is sorted clockwise.
Proof. The fact that det(v2 , v1 ) > 0 implies that {v1 , v2 } is a basis of the R−vector space
R2 . Write
v3 = λ1 v1 + λ2 v2 , λ1 , λ2 ∈ R.
We then have 0 > det(v2 , v3 ) = λ1 det(v2 , v1 ). As det(v2 , v1 ) > 0, we deduce that λ1 < 0.
This implies that λ2 > 0 as otherwise we would get hv3 , vi < 0, a contradiction.
We compute then det(v3 , v1 ) = λ2 det(v2 , v1 ) > 0, and inductively from here one can
prove det(vj , v1 ) > 0 for j > 1. More generally, by starting with any index i ≥ 1, one gets
(1).
To prove (2), note that we have -thanks to (1)- det(vN , v1 ) > 0, so {v1 , vN } is also a basis
of R2 . We will prove our claim by induction on j = 1, . . . , N − 1.
• For j = 1, we have αj = 1, βj = 0, so the claim follows.
• For a general j < N, condition (b) implies that there exist α, β ∈ (0, 1) such that
vj+1 = α vj + β vN .
By using the inductive hypothesis, we have that
vj+1 = α αj v1 + βj vN + β vN = α αj v1 + (α βj + β)vN ,
(8)
and we deduce that 0 < αj+1 = α αj < αj , 0 < βj+1 = α βj + β, which completes
with the proof of (2).
For (3), denote with Mv1 ,vN the monoid in Z2 generated by the integer points
lying in the positive rays generated by v1 , vN . Condition (a) implies that every
vector in Mv1 ,vN lies in one cone generated by the positive rays of vj , vj+1 , for some
j = 1, . . . , N − 1. As det(vj+1 , vj ) = 1, this implies that any element in this smaller
cone with integer coordinates is a linear combination of vj , vj+1 with integer (and
nonegative) coefficients, by Lemma 2.1. This proves that the family {v1 , . . . , vN }
generates Mv1 ,vN .
To show that the generation is minimal, suppose then that for some j ∈ {1, . . . , N},
we have
X
vj =
ak vk , ak ∈ N.
k6=j
For i = 1, . . . , N, we write vi = αi v1 + βi vN , and hence,
X
X
ak βk vN .
ak αk v1 +
vj = αj v1 + βj vN =
k6=j
k6=j
8
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
As the sequence {αk }k=1,...,N is positive strictly decreasing, we must have ak = 0 for
1 ≤ k < j. So, (8) turns into
X
vj =
ak vk ,
k>j
which implies then that
0 = det(vj , vj ) =
X
ak det(vj , vk ) ≥
k>j
X
aj ,
k>j
i.e. aj = 0 for all j > k, which would imply that vj is the zero vector, a contradiction.
This shows that (8) is not possible, and hence the family is a minimal set of generators.
′
To prove (4), suppose that {v1′ , . . . , vM
} is another set of minimal generators of
M. We have then,
!
!
M
M
M
N
N
X
X
X
X
X
aij bjk vk i = 1, . . . , N,
aij
vi =
aij vj′ =
bjk vk =
j=1
j=1
(9)
j=1
P
with aij , blk ∈ N. Note that for all i = 1, . . . , N, we must have M
k=1 aij bji ≥ 1, as
otherwise {v1 , . . . , vN } would not be minimal. We deduce then that
!
!
N
M
M
X
X
X
aij bjk vk , i = 1, . . . , N.
0=
aij bji − 1 vi +
j=1
(10)
k=1
k=1
k=1, k6=i
j=1
By performing the inner product of the two sides of the equality above with the
vector v, we deduce that
M
X
1 if i = k
aij bjk = δik =
i = 1, . . . , N.
0 if i 6= k,
j=1
Set A = aij 1≤i≤N, 1≤j≤M , B = bjk 1≤j≤M, 1≤k≤N . Identity (10) imply that A · B is
the identity matrix of N × N, which implies that N ≤ M. A symmetric argument
shows that M ≤ N, and hence the equality N = M holds. So, we have that A and
B are square matrices being inverses of each other.
From here the conclusion is straightforward, as it is easy to see that two square
matrices having coefficients in N being inverse of each other must be permutations of
the rows (or columns) of the identity matrix. We deduce then straightforwardly that
′
there is a permutation of the indices of {v1′ , . . . , vN
} which converts this sequence
into {v1 , . . . , vN }. This concludes with the proof of the proposition.
The following algorithm would help us compute a Hilbert basis of a given cone. It will be
also of use later in the sequel to obtain properties from such bases.
Algorithm 2.4 (Computation of a Hilbert basis of a cone).
Input: Two linearly independent lattice vectors (b, a), (b′ , a′ ) ∈ Z2 .
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
9
Output: The Hilbert Basis of the monoid generated by the integer points lying in the strongly
rational cone generated by (b, a) and (b′ , a′ ).
Procedure:
(1) Assume (b, a) and (b′ , a′ ) have coprime coordinates, otherwise, replace them
′ with
b a′
1
1
′
′
(b, a) and/or gcd (a′ ,b′ ) (b , a ) respectively. Assume also w.l.o.g. that det
>
gcd(a,b)
b a
0.
(2) Set S0 :=S1 := {(b′ , a′ ), (b, a)}.
(3) If det S1 = 1, return S0 and terminate the algorithm. Else
(a) Let S1 = {(b′ , a′ ), (b∗ , a∗ )}.Using the Euclidean Algorithm, compute (u, v) ∈ Z2
such that
u v
(11)
det
= 1.
b∗ a∗
(b) Write (b′ , a′ ) = α(b∗ , a∗ ) + β(u, v), with α, β ∈ Z.
(c) S0 ∪ {(u, v) + ⌈ αβ ⌉(b∗ , a∗ )} 7→ S0 and S1 := {(b′ , a′ ), (u, v) + ⌈ αβ ⌉(b∗ , a∗ )}. Go to
step 3.
Proposition 2.5. Algorithm 2.4 is correct.
Proof. First, it is straightforward to check recursively that det(S1 ) 6= 0 at every step of the
algorithm, which implies that every β ∈ Z computed is not identically zero, so step (c)
always has sense. Note also that every time we pass through step 3, due to (11) we have
that {(u, v), (b∗ , a∗ )} is a Z-basis of Z2 with positive orientation. At the beginning of each
step, if the algorithm is not going to terminate there, we have that
′
b a′
= β.
1 < det(S1 ) = det
b∗ a∗
In the end, we will have vN = (b′ , a′ ), and also at each step, if vi = (b∗ , a∗ ) then, following
the notation of the algorithm,
α
vi+1 = (u, v) +
(b∗ , a∗ ),
β
with (b′ , a′ ) = α(b∗ , a∗ ) + β(u, v). This implies straightforwardly that
α α
1
vi+1 =
(b∗ , a∗ ) + (b′ , a′ ) = λ vi + λ′ vN ,
−
β
β
β
with 0 < λ, λ′ < 1.
This shows that for each i = 1, . . . , N − 2, vi+1 lies in the interior of the parallelogram
with vertices 0, vi , vN , vi + vN , so we can apply Proposition 2.2.
At the end of step 3, we now have that - after changing S1 with the formula given in (c)
′
α
β α
b
a′
u v
= β⌈ ⌉ − α < β.
= det
· det
det
∗
∗
1 ⌈ αβ ⌉
u + ⌈ αβ ⌉b∗ v + ⌈ αβ ⌉a∗
b a
β
Note that also we have that β⌈ αβ ⌉−α cannot be equal to zero as otherwise this will imply that
(b′ , a′ ) is β has entries that are not relatively prime. So, we now have that the determinat of
10
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
the “new” S1 is strictly smaller than β and larger than zero. This shows that the algorithm
eventually terminates.
To conclude, we must show that the output is the Hilbert basis of the set of lattice points
lying in the cone generated by (b, a), (b′ , a′ ). Write S0 = {v1 , . . . , vN } sorted clockwise. Note
that all these vectors lie in N2 . By construction, we have already det(vi+1 , vi ) = 1, i =
1, . . . , N − 1. Now apply Proposition 2.2. This concludes the proof.
Corollary 2.6. Suppose that the cone generated by v1 and vN is strongly rational, and that
the clockwise-sorted family {v1 , . . . , vN } ⊂ Z2 is a Hilbert basis of this monoid. Then, for
j = 1, . . . , N − 2, vj+1 lies in the interior of the parallelogram whose vertices are 0, vj , vN
and vj + vN .
Proof. Apply Algorithm 2.4 with input v1 , vN . Thanks to Proposition 2.5, its ouput gives a
Hilbert basis of the aforementioned monoid, which satisfies that every vector in the family
-when properly sorted clockwise- lies in the interior of the parallelogram determined by the
immediately previous in the list and the last one. This implies the claim for this family. By
Proposition 2 (4), the Hilbert basis is unique, so it must coincide with {v1 , . . . , vN } ⊂ Z2 .
This concludes with the proof of the claim.
The following result will be useful in the sequel.
Lemma 2.7. Let {v1 , . . . , vN } ⊂ Z2 be the clockwise sorted Hilbert basis of the monoid
generated by this set. For each v such that hv, vj i > 0, 1 ≤ j ≤ N, we have that
hv, vj i ≤ max{hv, v1 i, hv, vN i}.
Proof. Suppose w.l.o.g. that j = 2, and write v2 = λv1 + µvN , with 0 < λ, µ < 1, the strict
inequalities are due to the fact that v2 lies in the interior of the parallelogram determined
by v1 and vN . If we show that λ + µ ≤ 1, then the claim will follow as
hv, v2i = λhv, v1 i + µhv, vN i ≤ (λ + µ) max{hv, v1 i, hv, vN i}
To do this, denote with D the determinant of {vN , v1 }. According to the Pick’s Theorem
applied to the parallelogram with vertices 0, v1 , vN , v1 + vN , we get that D = I + B/2 − 1,
where I is the number of interior points, while B is the number of boundary points, and
here B = 4. Note that D ≥ 2 as otherwise there would no be v2 as an interior point. From
1 = det(v2 , v1 ) = µ det(vN , v1 ) = µ D, we deduce that µ = D1 . On the other hand, by solving
the linear system v2 = λv1 + µvN using Cramer’s rule, we deduce that λ ≤ D−1
and hence
D
λ+µ≤
1
D−1
+
= 1,
D
D
which proves the claim.
We conclude with an example of how the algorithm works.
Example 2.8. Consider the fan from Example 1.5. The positive quadrant is split in two
strongly rational cones delimited by the line passing through the origin and (2, 3). We will
compute Hilbert bases for the monoids of integer points from each ofthese cones. First
2 3
consider the cone generated by {(0, 1), (2, 3)}. Note that det
= 2 > 0. Start
0 1
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
11
with the set S1 = {(2, 3), (0, 1)}. We need to find u, v integers such that u · 1 − v · 0 = 1,
and any (1, a) with a integer satisfies this condition. Write (2, 3) = α(0, 1) + β(1, a) and
note that ⌈ αβ ⌉ = 2 − a, which produces the vector (1, a) + (2 − a)(0, 1) = (1, 2). The new
2 3
S1 = {(2, 3), (1, 2)} and the algorithm ends, since det
= 1. So this monoid has the
1 2
Hilbert basis {(2, 3), (1, 2), (0, 1)}.
Repeating the algorithm for the monoid of integer points in the cone generated by (1, 0), (2, 3),
we get the Hilbert basis {(1, 0), (1, 1), (2, 3)}.
2.2. Gröbner bases. The theory of Gröbner bases on a ring R[Y1 , . . . , Ys ] of polynomials
over a Noetherian ring R is a bit more subtle than the one developed for fields, but it is
well-known and their properties well stablished. We will follow the presentation given in
[AL94, Chapter 4].
Set Y = Y1 , . . . , Ys for short. The ring R[Y] is also Noetherian, and given a monomial
order ≺ on the set of monomials of this ring, there are well-defined notions of leading monomial, leading term and leading coefficient of nonzero elements in R[Y] which will be denoted
with lm, lt, and lc respectively. We will assume all along this subsection that the monomial
order ≺ has been fixed. The main subtlety here comes with the concept of reduction of a
polynomial with respect to a family, which is defined as follows:
Definition 2.9. Given g, h ∈ R[Y] and a set of non-zero polynomials G = {g1 , . . . , gt } ⊂
R[Y], we say that g reduces to h modulo G in one step, denoted
G
g → h,
if and only if
h = g − (c1 M1 g1 + . . . + ct Mt gt )
for c1 , . . . , ct ∈ R and power products M1 , . . . , Mt satisfying lm(g) = Mi lm(gi ) for all i such
that ci 6= 0, and lt(g) = c1 M1 lt(g1 ) + . . . + ct Mt lt(gt ).
We say that g reduces to h modulo G, denoted
G
g →+ h,
if and only exist there exist polynomials h1 , . . . , hp ∈ R[Y] such that
G
G
G
G
G
g → h1 → h2 → . . . → hp → h.
Remark 2.10. If all the gi above satisfy lc(gi ) = 1, 1 ≤ i ≤ t, then one can develop a (first
step of the) standard division algorithm by setting ci0 = lc(g), i0 being the maximal index
such that lm(g) = Mi0 lm(gi0 ), and ci = 0 for all i 6= i0 , if there is some lm(gi ) dividing
lm(g), or removing this leading term and adding it to the remainder otherwise.
Theorem 2.11 (Theorem 4.1.10 in [AL94]). Let g, g1 , . . . , gt ∈ R[Y] with gi 6= 0, i =
1, . . . , t, and set G = {g1 , . . . , gt }. Then there exists r ∈ R[Y], minimal with respect to G (in
12
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
G
the sense that it cannot be reduced further modulo G) such that g →+ r. Moreover, there are
h1 , . . . , ht ∈ R[Y] such that
g = h1 g1 + . . . + ht gt + r,
with lm(g) = max (max1≤i≤t (lm(hi )lm(gi )), lm(r)) .
As usual, for a subset W ⊂ R[Y], the leading term ideal of W is denoted as
(12)
Lt(W ) = hlt(w), w ∈ W i.
Theorem 2.12 (Theorem 4.1.12 in [AL94]). Let I be an ideal of R[Y], and G = {g1 , . . . , gt }
a subset of non-zero polynomials in I. The following are equivalent:
(1) Lt(G) = Lt(I).
(2) For any polynomial g ∈ R[Y] we have
G
g ∈ I if and only if g →+ 0.
(3) For any g ∈ I, g = h1 g1 + . . . + ht gt for some polynomials h1 , . . . ht ∈ R[Y] such that
lt(g) = max1≤i≤t (lt(hi )lt(gi )) .
Definition 2.13. A set G of non-zero polynomials contained in an ideal I is called a Gröbner
basis for I if G satisfies one of the three equivalent conditions of Theorem 2.12
The following claim resumes the properties of Gröbner bases in Noetherian rings. Proofs
of these facts can be found in [AL94]
Theorem 2.14. Let I ⊂ R[Y] be a non-zero ideal. Then I has a Gröbner basis. Let G be
a Gröbner basis of I.
• I = hGi.
G
• If g ∈ I and g →+ r, where r is minimal (meaning that r cannot be reduced further
modulo G), then r = 0.
• If R = k[x1 , . . . xℓ ], with k a field, and G is a Gröbner basis of an ideal in the ring
of polynomials k[x1 , . . . , xℓ , Y1 , . . . , Ys ] with respect to an elimination order with the x
variables larger than the y variables, then G is a Gröbner basis of the ideal generated
by its elements in R[Y] with the induced order in this ring.
The following result will be of use in the sequel.
Lemma 2.15. Let p ∈ R not a unit in this ring, g, h ∈ R[Y] and G = {g1 , . . . , gt } ⊂ R[Y]
G
with lc(gj ) = 1 for all j, such that g →+ r, with r minimal with respect to G, and the
reduction was done by using the Division Algorithm outlined in Remark 2.10. Denote with
g, G, r the classes of g, G, and r modulo p. Then
G
g →+ r
in R/hpi, and r is minimal with respect to G.
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
13
Proof. As the leading coefficients of the gi ’s are 1, then it is clear that r is minimal with
respect to G as otherwise a reduction could have been done in R.
To prove the other part of the claim, note that if we look at each step of the division
algorithm as outlined in Remark 2.10, we either have lc(g) ∈ hpi or not. In the first case,
then modulo p this leading coefficient vanishes, but then we will have that ci0 in the algorithm
will be an element of hpi and hence equal to zero in R/hpi, so modulo p we would not have
seen any reduction at this step, as the leading coefficient vanishes. In the other case, we have
ci0 = lc(f ) 6= 0, and hence the reduction happens also modulo p. Applying this reasoning
recursively, we get the claim.
Definition 2.16. Let I be a non-zero ideal in R[Y] and G a Gröbner basis for I. We say
that G is a minimal Gröbner basis provided that for all g ∈ G, g is minimal with respect to
the set G \ {g} (as before, “minimal” here means that g cannot be reduced further modulo
G \ {g}).
Proposition 2.17 (Exercise 4.1.9 in [AL94]). Every Gröbner basis for I contains a minimal
Gröbner basis for I.
Definition 2.18. A minimal Gröbner basis G is called reduced if, for all g ∈ G, lc(g) = 1,
and moreover no non-zero term in g is minimal with respect to G \ {g}.
In contrast with the case of fields, not necessarily every ideal I has a reduced Gröbner
basis (for instance if all the coefficients of the elements in I belong to some proper ideal of
R). But many of the ideals which will appear in this paper will do so.
We will consider now Gröbner bases for submodules of R[Y]m , for m ≥ 1. Denote with
{e1 , . . . , em } the canonical basis of this R[Y]-module. Recall that a monomial in R[Y]m is
a vector of the type Y α ei , 1 ≤ i ≤ m, with Y α being a classical monomial in R[Y]. A
term (or monomial) order on the monomials of R[Y]m is a total order ≺ on these monomials
satisfying:
(1) U ≺ Y α U for every monomial U ∈ R[Y]m , and classical monomial Y α ∈ R[Y];
(2) if U, V are monomials in R[Y]m with U ≺ V, then Y α U ≺ Y α V for every classical
monomial Y α ∈ R[Y].
Fixed a monomial order, for an element f ∈ R[Y]m , one defines the leading monomial,
the leading coefficient and the leading term of f in the usual way, and denotes them with
lm(f), lc(f) and lt(f) respectively.
The initial module Lt(M) is defined in the obvious way, generalizing (12). Given a submodule M ⊂ R[Y]m , a set G = {g1 , . . . , gt } ⊂ M is called a Gröbner basis for M with
respect to ≺ if and only if
Lt(M) = Lt(G).
As R and R[Y] are Noetherian rings, then every submodule has a Gröbner basis. The
following definitions generalize those given above.
Definition 2.19. For a given term order in R[Y], a Gröbner basis G of M is called minimal
if for all g ∈ G, g is minimal with respect to the set G\{g} (as before, “minimal” here means
that g cannot be reduced further modulo G \ {g}).
14
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
A minimal Gröbner basis G of M is called reduced if, for all g ∈ G, lc(g) = 1, and
moreover no non-zero term in g is minimal with respect to the set G \ {g}.
Let Y α ei , Y β ej be two monomials
in R[Y]m , the least common multiple of these two
monomials (denoted lcm Y α ei , Y β ej ) is equal to either 0 if i 6= j, or Y max{α, β} ei otherwise.
Let g, g′ ∈ R[Y]m are such that lc(g) = lc(g′ ) = 1. The vector
lcm lm(g), lm(g′ ) ′
lcm lm(g), lm(g′ )
′
g−
g ∈ R[Y]m
(13)
S(g, g ) :=
′
lm(g)
lm(g )
′
is called the S-polynomial of g and g . Note that the S-polynomial is actually a vector of
polynomials.
Theorem 2.20. Let G = {g1 , . . . , gt } be a set of non-zero vectors in R[Y]m , with lc(gj ) =
1, j = 1, . . . , t. Then G is a Gröbner basis for the submodule M = hg1 , . . . , gt i ⊂ R[Y]m if
and only if for all i 6= j, we can write
t
X
(14)
S(gi , gj ) =
Fijν (Y)gν ,
ν=1
with Fijν (Y) ∈ R[Y], such that
(15)
max {lm Fijν (Y)lm(gν ) } = lm (S(gi , gj )) .
1≤ν≤t
Proof. The proof given in [AL94, Theorem 3.5.19] for the case of R equals a field works
straightforwardly in this case. Note that we need to use that lc(gj ) = 1 (which can be easily
assumed in the field case).
For a sequence g1 , . . . , gt ∈ R[Y]m , the syzygy module of this sequence is the submodule
of R[Y]t defined as
(16)
t
syz(g1 , . . . , gt ) = {(h1 , . . . , ht ) ∈ R[Y] :
t
X
hj gj = 0}.
j=1
Suppose further that G = {g1 , . . . , gt } is a Gröbner basis of a submodule M of R[Y]m , with
t
lc(gj ) = 1, j = 1, . . . , t. Let {ẽ1 , . . . , ẽt } be the canonical basis
of R[Y] , and write S(gi , gj )
with 1 ≤ i < j ≤ t as in (14). Set Xi,j := lcm lm(gi ), lm(gj ) and
t
(17)
si,j :=
X
Xi,j
Xi,j
ẽi −
ẽj −
Fijν (Y )e˜ν ∈ R[Y]t .
lm(gi )
lm(gj )
ν=1
By (13) and (14), we easily see that si,j ∈ syz(g1 , . . . , gt ).
Let G = {g1 , . . . , gt } be a sequence non-zero vectors in R[Y]m . We define an order ≺G on
the monomials of R[Y]t as follows
lm(Y α gi ) ≺ lm(Y β gj ) or
α
β
(18)
Y ẽi ≺F Y ẽj ⇐⇒
lm(Y α gi ) = lm(Y β gj ) and j < i.
We call ≺G the order induced by G. We have the following theorem which is essentially due
to Schreyer, see [E96].
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
15
Theorem 2.21. If G = {g1 , . . . , gt } is a Gröbner basis of a submodule M ⊂ R[Y]m , with
lc(gj ) = 1 for all j = 1, . . . , t, then {si,j }1≤i<j≤t is a Gröbner basis of syz(g1 , . . . , gt ) ⊂ R[Y]t
with respect to ≺G . Moreover, if si,j 6= 0, then
(19)
lt(si,j ) =
Xi,j
ẽi ,
lt(gi )
1 ≤ i < j ≤ t.
Proof. The proofs of [AL94, Lemma 3.7.9 & Theorem 3.7.13] for the case when R is a field
hold straightforwardly to this case.
Remark 2.22. From (19) we get immediately that for si,j 6= 0, lc(si,j ) = 1.
3. ker(ϕ0 )
In this section, we will compute generators of the kernel of ϕ0 . From now on, unless we
give more precisions about the ring R, a minimal set of generators of an R-module will be
a set of generators of this module with the property that if we remove one of the elements
of the set, it does not generate the same module anymore. Recall from (6) that we have a
presentation
ϕ0
R[Y] → Bf ,F ,p
Yv 7→ pf (v) xv ,
where ϕ0 is a Z2 -graded map if we declare deg(Yv ) = v, and deg(r) = (0, 0) for all r ∈ R.
Let ≺ be the lexicographic monomial order in R[Y] which satisfies Yv ≺ Yv′ if and only if v
is “to the left” of v′ . To be more precise, if we sort H = {v1 , . . . , vM } clockwise, then our
monomial order is lexicographic order in R[Yv1 , . . . , YvM ] with Yvi Yvj if and only if i ≤ j.
If we use this notation, we will denote with Yj the variable Yvj to simplify notation.
We will say that two distinct vectors v and v′ ∈ H are adjacent if in the open strongly
rational cone generated by these two vectors there are no other elements of H. Note that
being adjacent implies that all the lattice vectors lying in the cone generated by v and v′
can be expressed as a nonegative integer combination of these two vectors. Moreover, two
adjacent vectors lie simultaneously in one and only one of the cones Ci , which immediately
implies that
(20)
fu (αv + βv′) = αfu (v) + βfu (v′ ) ∀α, β ∈ R≥0 , u = 1, . . . , n.
Proposition 3.1. Suppose that |F | is a proper rational cone contained in R2 . If v and v′
are distinct, non-adjacent elements of H, then there exist adjacent w, w′ ∈ H, α, α′ ∈ N
such that
(21)
′
Sv,v′ := Yv Yv′ − Ywα Ywα′ pγvv′ ∈ ker(ϕ0 ),
with
(22)
γvv′ = f1 (v) + f1 (v′ ) − f1 (v + v′ ), . . . , fn (v) + fn (v′ ) − fn (v + v′ ) .
The expression defining Sv,v′ in (20) is unique. Its leading term with respect to ≺ is Yv Yv′ .
Note that the degree of Sv,v′ with the Z2 -graduation, is equal to v + v′ .
16
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
Proof. The vector v + v′ must belong to one of the strongly rational cones generated by two
consecutive elements of H. Let w and w′ be two adjacent vectors whose cone contains the
former sum. Then, there exist α, α′ ∈ N with
v + v′ = αw + α′ w′ .
(23)
Note that we must have either
(24)
v ≺ w, w′ ≺ v′ or v′ ≺ w, w′ ≺ v,
due to the way we labeled the vectors in the cone, and moreover the expression (23) is
unique. By the definition of ϕ0 given in (6), and using (23), (20) and (22), it is easy to see
that Sv,v′ ∈ ker(ϕ0 ). The fact that its leading term is Yv Yv′ follows straightforwardly from
(24).
Definition 3.2. Sf ,F ,p := {Sv, v′ | v, v′ ∈ H, v ≺ v′ , non-adjacents}.
Recall that the cardinality of H was denoted by M in the introduction. Therefore,
−2)
#Sf ,F ,p = (M −1)(M
= M2−1 .
2
Lemma 3.3. With the lexicographic monomial order ≺ defined above, we have that
(1) Lt Sf ,F ,p = hYv Yv′ : v, v′ ∈ H, v ≺ v,′ non-adjacentsi.
(2) Y c ∈
/ Lt Sf ,F ,p if and only if there exist adjacent vectors w, w′ ∈ H and A, A′ ∈ N
′
such that Y c = YwA YwA′ .
Proof. The first part of the claim follows straightforwardly from the definition of Sf ,F ,p and
′
Proposition 3.1. To prove the rest, it is clear first that any monomial of the form YwA YwA′ ,
′
′
with adjacent
w, w ∈ HQi for some i, and A, A ∈ Z≥0 are not divisible by any monomial
in Lt Sf ,F ,p , which proves one of the implications.
Q
c
cv
To prove
the converse, suppose that Y = v∈H Yv is not divisible by any element in
Lt Sf ,F ,p . Due to the first part of the claim, we must have then that cw > 0 if and only if
cw′ = 0 for any w′ non adjacent with w, otherwise, Y c will be divisible by Yw Yw′ . We have
the following scenarios:
• All cw = 0, then Y c = 1, and the claim holds with A = A′ = 0 in this case.
• There exists w0 such that cw0 > 0. Then, all the cw′ with w′ non adjacent with w0 will
be zero. There are at most two adjacent vectors to w0 ; namely, the one immediately
“above” w0 , and the other which is immediately “below” it. Name them w0+ and w0−
respectively. We may have c+
w0 > 0 or cw0− > 0, but not the two of them as these two
are non-adjacent (they have w0 “between” them) and hence the binomial Yw0+ Yw0−
cw+ cw−
belongs to Lt Sf ,p,a,b , and would divide Yw+0 Yw0−0 if the two exponents are strictly
0
c
positive. Hence, we have that Y =
claim.
cw+
c
Yww0 0 Yw+0 ,
0
c
cw−
or Y c = Yww0 0 Yw−0 , which proves the
0
Now we are ready for one of the main results of this text.
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
17
Theorem 3.4. If |F | is a proper rational cone contained in R2 , no pi , i = 1, . . . , n, is a zero
divisor in R, then Sf ,F ,p is a reduced Gröbner basis of ker(ϕ0 ).
Proof. Suppose we have a nonzero polynomial P (Y) ∈ ker(ϕ0 ) whose leading term is not
divisible by any of the leading terms of elements in Sf ,F ,p . Thanks to Lemma 3.3, we will
have then
′
(25)
Lt P = r YvA YvA′ ,
with r ∈ R \ {0} v, v′ ∈ H adjacent vectors, and A, A′ ∈ N. If A = A′ = 0, then P = r 6= 0
and cannot be in the kernel of ϕ0 as ϕ0 (r) = r. So at least one of the exponents must be
different from zero. Suppose also w.l.o.g. that v′ ≺ v.
′
Write P = r YvA YvA′ + P̃ . As ϕ0 (P (Y)) = P (ϕ0 (Y)) = 0, we must have
X
′
′
′ ′
qw xw
(26)
r pAf (v)+A f (v ) xAv+A v = −P̃ |Yz 7→pf (z) xz =
w
with qw ∈ R, and all w in HQ lying “above” v, i.e.
X
(27)
w=
cz z, cz ∈ N, and cv ≤ A.
zv
This is due to the fact that all the exponents appearing in the polynomial expansion of P̃
′
have lexicographic degree strictly smaller than YvA YvA′ , i.e. they must be of the form
Y
(28)
Yzcz ,
zv
with cz ∈ N, and cv ≤ A. By applying ϕ0 to such a monomial, we get one of the exponents
in x appearing in (27).
As no pi is a zero divisor in R, we then have that the left hand side of (26) is nonzero.
Then, there must be some xw appearing on the right-hand-side with nonzero coefficient, i.e.
there must be a family of nonegative integers cz ∈ N, with z v and cv ≤ A, such that
X
(29)
Av + A′ v′ =
cz z.
zv
Hence, we must have that
P
det (Av + A′ v′ , v′ ) = det ( cz z, v′ )P
A det(v, v′ )
= cv det(v, v′ ) + z≺v cz det(z, v′ ),
and from here we deduce that
0 ≤ (A − cv ) det(v, v′ ) =
X
cz det(z, v′ ) ≤ 0,
z≺v
which implies thatQA = cv , z = 0 if z ∈
/ {v, v′ }, and hence cv′ = A′ . From (28) and (29) we
A A′
cz
deduce then that zv Yz = Yv Yv′ , which is impossible because the monomial in the left
hand side should of this equality be strictly smaller than the one in the right hand side in
the lexicographic order. This contradiction shows that
Lt Sf ,F ,p = Lt(ker(ϕ0 )),
18
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
and hence, Sf ,F ,p is a Gröbner basis of ker(ϕ0 ). The fact that it is reduced follows straightforwardly from Definition 2.18, and the shape of the generators defined in (21).
Even though Sf ,F ,p is a reduced Gröbner basis, it is not necessarily a set of minimal
generators of ker(ϕ0 ), see cautionary Example 3.12.
Proposition 3.5. Suppose that for i = 1, . . . , n, pi ∈ R is neither a unit nor a zero divisor.
Then, the family {pf (v) xv ; v ∈ H} is a minimal set of generators of Bf ,F ,p (in the sense that
no generator is a product of others) if and only if there are no elements in Sf ,F ,p of the form
Yv Yv ′ − Yw .
Proof of Proposition 3.5. Clearly, if there is an element of the form Yv Yv′ −Yw in Sf ,F ,p, then
the generators are not minimal. Conversely, suppose that pf (w0 ) xw0 is a product of other
elements of the family. This would be equivalent to the existence of an element of the form
Y
Ywαw − Yw0
(30)
w
lying in the kernel of ϕ0 . Due to Theorem 3.4, we then have that if we divide this element
by the Gröbner basis Sf ,F ,p the remainder will be zero. As Yw0 is not divisible by any of the
(quadratic) leadingQ
monomials of the elements of Sf ,F ,p , we must have that the leading
P term
αw
of (30) is actually w Yw , and hence its total degree in the Y variables, namely w αw ,
must
be at least equal to 2. Moreover, (30) implies that the remainder of the division of
Q
αw
w Yw with respect to Sf ,F ,p is equal to Yw0 .
In what follows the total degree of a monomial refers to the sum of the exponents appearing
in the monomial. Due to the particular shape
of the elements of Sf ,F ,p given in (21), every
Q
time we divide a monomial of the form w Ywαw with one of the elements of the Gröbner
′
basis, such as Yv Yv′ − Ywα Ywα′ pγvv′ , one of the following two possible outcomes happen:
i) the total degree of the remainder does not decrease (this is the case when α + α′ ≥ 2
in (21)),
ii) the total degree decreases by one when the division is performed over an element of
the form Yv Yv′ − Yz pγvv′ (here, α + α′ ≤ 1 and z is either w or w′ ).
Q
α′w
γ
′
In any case, the remainder of this division
is
of
the
form
p
Y
− Yw0 , for some γ, αw
.
w
w
Q
αw
At the beginning of the division of w Yw with respect to the Gröbner basis Sf ,F ,p , the
aforementioned monomial has total degree at least two 2. After a finite number of steps
within the algorithm, we must obtain the remainder Yw0 , which has total degree one. This
implies that step ii) must happen at least once.
If there are no elements in Sf ,F ,p of the form Yv Yv′ − Yw , then for Yv Yv′ − Yz pγvv′ ∈ Sf ,F ,p ,
we will have that pγww′ ∈ (pi0 ) 6= R for some i0 ∈ {1, . . . , n}. In R/(pi0 ) 6= (0), the binomial
Yv Yv′ − Yz pγvv′ actually becomes Yv Yv′ , and the division algorithm stops here, which implies
that the remainder in this quotient ring is equal to zero. By Lemma 2.15, this contradicts
the fact that Yw0 mod pi0 6= 0. This proves that an expression like (30) can never exists in
ker(ϕ0 ), i.e. the family {pf (v) xv ; v ∈ H} is a minimal set of generators of Bf ,F ,p.
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
19
Recall from Section 1.1 that we say that f is strict with respect to the fan F if for each
1 ≤ i < ℓ, there is ki ∈ {1, . . . , n} such that fki is not linear in Ci ∪ Ci+1 . The fan F can
always be chosen in such a way that this condition is accomplished.
Definition 3.6. Let w, w′ ∈ H and f strict with respect to the fan F . We say that fk
separates w and w′ if fk (w) + fk (w′ ) − fk (w + w′) > 0. Similarly, if fk is not linear in
Ci ∪ Ci+1 , then we say that fk separates Ci and Ci+1 .
We leave it to the reader to check that if fk separates Ci and Ci+1 , then fk separates any
vectors w, w′, where w ∈ Cj , j ≤ i and w′ ∈ Cl , l ≥ i + 1.
Proposition 3.7. If f is strict with respect to F , and no pi is a zero divisor or a unit, then
{pf (v) xv ; v ∈ H} is a minimal set of generators of Bf ,F ,p .
Proof. Thanks to Proposition 3.5, we only need to know that there are no elements of the
form Yv Yv′ −Yw in Sf ,F ,p . But if this is the case, then we must have v+v′ = w, which implies
straightforwardly that the vectors v and v′ belong to different cones as otherwise the three
vectors v, v′ , w would be part of a Hilbert basis, which would not be minimal because of the
relation among them. Hence, the element appearing in Sf ,p,a,b is actually Yv Yv′ − Yw pγvv′ .
The fact that f is strict combined with the hypothesis that none of the pi ’s is a unit implies
that pγvv′ 6= 1. This concludes the proof.
Now we turn to the problem of finding conditions to ensure that Sf ,F ,p is a minimal set
of generators of ker(ϕ0 ). The following claim will be used to prove the main result in this
direction.
Proposition 3.8. Suppose that f is strict with respect to F , and that |F | is a strongly convex
rational cone. For v ∈ H, set
Dv := {(w, w′) ∈ H2 , w ≺ w′ , w + w′ = v}.
If Dv 6= ∅, then there exists k ∈ {1, . . . , n} such fk (w) + fk (w′ ) − fk (w + w′ ) > 0 for all
(w, w′) ∈ Dv .
Proof. As before, if we have w + w′ = v, as these three vectors belong to H, we must have
then that w and w′ are in different cones, and v “between” them, i.e. w ≺ v ≺ w′. In
particular, we have that w and w′ are not adjacents. Suppose that v ∈ Ci . We have two
possible scenarios for w and w′ :
a) w ∈ Cj , j < i and w′ ∈ Cl , l ≥ i,
b) w ∈ Ci , and w′ ∈ Cl , l > i.
As f is strict with respect to F , if we pick k as the index of the function which separates
Ci−1 and Ci , all the cases considered in a) are covered. On the other hand, to deal with the
cases appearing in b) we should pick as k the index of the function which separates Ci with
Ci+1 . Note that this index also works for those cases in a) where w′ ∈ Cj with j > i.
To conclude the proof we must show that we cannot have simultaneously
(31)
w + w′ = z + z′ = v,
with w, w′ , z, z′ ∈ H, w ∈ Cj , j < i, w′ ∈ Ci , z ∈ Ci , z′ ∈ Cj ′ , j ′ > i, i.e. we only have to
consider one of the two possible k’s above for all the cases. Suppose then that (31) holds,
20
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
and also w.l.o.g. that the sequence {z, v, z′ } is sorted clockwise. We straightforwardly get
that {w, z, v, w′, z′ } is sorted clockwise as the three vectors in the middle are in Ci . So we
must have that v is an element of the (unique) Hilbert basis of the monoid of lattice points
in the cone generated by z and w′ .
As the support of the fan |F | is a strongly rational cone, there exists v0 ∈ Z2 such that
hv0 , v′ i > 0 for all v′ ∈ |F |. From (31) we get then that
max{hv0 , w′ i, hv0 , zi} < hv0 , vi,
which is a contradiction with Lemma 2.7. This concludes the proof of the Proposition.
The hypothesis that |F | is strongly convex is necessary as the following example shows.
Example 3.9. Consider F given by its three maximal cones C1 , C2 , and C3 with
• C1 being generated by (−1, 0) and (−1, 1),
• C2 being generated by (−1, 1) and (1, 1),
• C3 being generated by (1, 1) and (1, 0).
In this case, we have that H = {v1 := (−1, 0), v2 := (−1, 1), v3 := (0, 1), v4 := (1, 1), v5 :=
(1, 0)}, and these vectors satisfy
(−1, 0) + (1, 1) = (−1, 1) + (1, 0) = (0, 1),
which is a situation like (31). Now set
f1 (v) =
f2 (v) =
0
if v ∈ C1 ,
v1 + v2 if v ∈ C2 ∪ C3
v2 − v1 if v ∈ C1 ∪ C2 ,
0
if v ∈ C3
As f1 separates C1 from C2 and f2 separates C2 from C3 , we get that f = (f1 , f2 ) is strict,
but
f1 (−1, 1) + f1 (1, 0) − f1 (0, 1) = 0 = f2 (−1, 0) + f2 (1, 1) − f2 (0, 1).
Remark 3.10. If we do not assume |F | being strongly convex but only contained in a hyperplane, the proof of Proposition 3.8 can be applied also to this situation, but we would
conclude that, given Dv 6= ∅, there exists a set of indices S with exactly two elements, such
that for every (w, w′) ∈ Dv , there exists k ∈ S such that fk separates w and w′ ; namely, we
let S consists of the index separating Ci−1 from Ci , and the index which separates Ci from
Ci+1 .
Recall that if f is strict, then for each i = 1, . . . , ℓ − 1, there is an index ki which satisfies
that fki is non linear in Ci ∪ Ci+1 . If this is the case, we will assume that we are given already
a distinguished list {k1 , . . . , kℓ−1 } of such indices.
Theorem 3.11. With notations as above, if |F | is a proper convex rational cone, f strict with
respect to F , no pki is a zero divisor, i = 1, . . . , ℓ, and for 1 ≤ i ≤ j ≤ ℓ − 1, hpki , pkj i =
6 R,
then Sf ,F ,p is a set of minimal generators of ker(ϕ0 ).
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
21
Proof. Note that the condition in the hypothesis also ensure that no pki is a unit, 1 ≤ i ≤ ℓ−1.
Suppose first that |F | is strongly rational convex, and that we have a pair of non-adjacent
vectors v, v′ ∈ H such that Sv,v′ is in the ideal generated by all the others elements in Sf ,F ,p ,
i.e.
X
′
α′
(32)
Yv Yv′ − Ywα Ywα′ pγvv′ =
Qz,z′ (Y) Yz Yz′ − Yuαzz′ Yu′zz′ pγzz′ ,
z,z′
the sum being over all the pairs z, z′ of non-adjacent vectors in H except v, v′. If we set all
the variables Yz to zero except Yv and Yv′ in (32), we get
X
α′
(33)
Yv Yv ′ = −
Qz,z′ (0, . . . , Yv , Yv′ , 0, . . . , 0)Yuαzz′ Yu′zz′ pγzz′ ,
z,z′
where the sum to the right is now indexed over all the pairs of no adjacent z, z′ ∈ H such
that either
(34)
z + z′ = v or z + z′ = v′ .
α
α′
This is due to the fact that no monomial among those of the form Yu zz′ Yu′zz′ appearing in
(32) can be a multiple of Yv Yv′ (as v and v′ are no adjacent), but we may have
(35)
′
′
z + z′ = αzz′ u + αzz
′u ,
′
′
′
′
with αzz′ = 1 (resp. 0) and αzz
′ = 0 (resp. 1), and u ∈ {v, v } or u ∈ {v, v }, which
implies (34).
By Proposition 3.8, all the pγzz′ are multiples of some pki (resp. pkj ) for those z + z′ = v
(resp. z + z′ = v′ ). So, the right hand side of (33) belongs to the ideal hpki , pkj i =
6 R[Y]. By
extracting the coefficient of Yv Yv′ on the right hand side of (33), we get then an element in
hpki , pkj i =
6 R, so the equality in (33) cannot hold. This concludes the proof of the Theorem
for the case |F | being strongly rational.
For |F | being a closed half-plane, we argue as follows: Recall that we have H = {v1 , . . . , vM },
where the sequence is sorted clockwise. So, in this case, we have that v1 = −vM . Assume
again that we have a situation like (32). If both v and v′ lie in the same cone Ci , then due
to Remark 3.10, we will conclude straightforwardly that all the pγzz′ in (33) are multiples of
either pki−1 or pki . From here, the claim follows straightforwardly.
Suppose now that v ∈ Ci and v′ ∈ Ci′ with i < i′ . If we remove vM from H and apply
Proposition 3.8 to the subfan generated by H \ {vM }, we would get that all the elements in
Dv except possibly the pair (vM , z) have an index ki∗ such that fki∗ separates them. Due to
Remark 3.10, we deduce that we can choose ki∗ ∈ {ki−1, ki }.
If in addition the pair (vM , z) belongs to Dv , then we must have with vM + z = v, which
implies that z is “to the left” of v. So, the exponent pγvM z will be a multiple of the product
γzz′
of all the pkj j ≥ i.
coming from (33) with z + z′ = v, belong
QSo, we have that all the p
to the ideal hpki∗ , ℓ≥i pkℓ i.
Reasoning symmetrically with v′ as before (we remove now v1 = −vM from H), we
conclude that all the elements in Dv′ except possibly (z, v1 ) induce an element pγzz′ which is
multiple of only one between pki′ −1 , pki′ , which we denote with pki∗′ , and the remaining power
22
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
Q
γzz′
pγv1 z must be a multiple
of
coming from (33) with z + z′ = v′ ,
ℓ≤i′ −1 pkℓ . So, all the p
Q
belong to the ideal h ℓ≤i′ −1 pkℓ , pki∗′ i.
Q
Q
As hpki∗ , ℓ≥i pkℓ i ∩ h ℓ≤i′ −1 pkℓ , pki∗′ i ⊂ hpki∗ , pki∗′ i, we can argue as before and conclude
with the proof.
The hypothesis on hpi , pj i =
6 R cannot be avoided. See Example 3.12. We illustrate our
results with some examples.
Example 3.12. Let n = 2, p1 , p2 ∈ R not zero divisors, and F being the fan whose maximal
cones are the following:
• C1 being the angular region determined by (0, 1) and (1, 3);
• C2 being the angular region determined by (1, 3) and (3, 1);
• C3 being the angular region determined by (3, 1) and (1, 0).
The set H for this case is the following:
H = {v1 := (0, 1), v2 := (1, 3), v3 := (1, 2), v4 := (1, 1), v5 := (2, 1), v6 := (3, 1), v7 := (1, 0)}.
We choose as fan-functions
f1 (v1 , v2 ) := max{3v1 , v2 }
f2 (v1 , v2 ) := max{v1 , 3v2 }.
Note that the family f is strict. The map ϕ0 is described by:
Y1 → p1 p32 x2
Y5 → p61 p32 x21 x2
Y2 → p31 p92 x1 x32
Y6 → p91 p32 x31 x2
Y3 → p31 p62 x1 x22
Y7 → p31 p2 x1 .
Y4 → p31 p32 x1 x2
Theorem 3.4 states that Sf ,F ,p , the minimal Gröbner basis of this ideal has 15 =
whose leading terms must be the following 15 monomials:
6×5
2
elements
{Y1 Y7 , Y1 Y4 , Y7 Y3 , Y1 Y3 , Y7 Y2 , Y4 Y2 , Y7 Y4 , Y1 Y5 , Y3 Y5 , Y2 Y5 , Y7 Y5 , Y1 Y6 , Y4 Y6 , Y3 Y6 , Y2 Y6 } .
By computing explicitly this Gröbner basis with Mathematica, we get the following set:
{Y1 Y7 − p1 p2 Y4 , Y1 Y4 − p2 Y3 , Y7 Y3 − p1 Y42 , Y1 Y3 − p2 Y2 , Y7 Y2 − p1 Y4 Y3 , Y4 Y2 − Y32 ,
Y7 Y4 − p1 Y5 , Y1 Y5 − p2 Y42 , Y3 Y5 − Y43 , Y2 Y5 − Y42 Y3 , Y7 Y5 − p1 Y6 , Y1 Y6 − p2 Y4 Y5 ,
Y4 Y6 − Y52 , Y3 Y6 − Y42 Y5 , Y2 Y6 − Y44 } ,
which can easily be seen to be equal to Sf ,F ,p . If the ideal hp1 , p2 i is not the whole ring, then
Sf ,F ,p is also a minimal set of generators of ker(ϕ0 ), and the presentation of ker(ϕ0 ) given
in (6) is minimal.
Suppose now that we have p1 = p2 + 1. We then have that hp1 , p2 i = R, and we easily
verify that
Y5 Y3 − Y43 = Y4 (Y7 Y3 − p1 Y42 ) − Y1,2 (Y7 Y4 − p1 Y5 ) − Y4 (Y1 Y5 − p2 Y42 ) + Y5 (Y1 Y4 − p2 Y3 ),
i.e. the system of generators is not minimal.
If we set now p2 = 1 and impose no further conditions on p1 except that it is not a zero
divisor, we can verify straightforwardly that
Y5 Y3 − Y43 = Y4 (Y1 Y5 − Y42 ) − Y5 (Y1 Y4 − Y3 ),
i.e. the presentation (6) is not minimal.
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
23
Example 3.13. Let a, b be positive integers, and define F as the fan whose maximal cones
are
• C1 being the convex hull of the rays R≥0 · (0, 1) and R≥0 · (ab + 1, a);
• C2 being the convex hull of the rays R≥0 · (ab + 1, a) and R≥0 · (1, 0).
Set also p1 = x1 . The case a = 1 and b = 1 have been studied in [Mal13]. The set H in this
case, sorted already in the lexicographic order, has the following a + b + 2 elements:
H = {(0, 1), (1, 1), (2, 1), . . . , (b, 1), (ab + 1, a), ((a − 1)b + 1, a − 1), . . . , (b, 1)(1, 0)}.
Theorem 3.4 above states that Sf ,F ,p has (a+b+1)(a+b)
elements.
2
If a = 1, we recover Theorem 3.2.1 in [Mal13], as it is easy to see that Sf ,F ,x1 is the ideal
generated by the 2 × 2 minors of the 2 × (b + 2) matrix
Y(0,1) Y(1,1) . . . Y(j,1)
. . . Y(b+1,1)
.
Y(1,1) Y(2,1) . . . Y(j+1,1) . . . Y(1,0)
, which is the number
Note that the number of 2 × 2 minors of this matrix is equal to (b+2)(b+1)
2
of elements in the minimal set of generators. However, we do not produce the same set but
it is very easy to see that the ideal generated is the same.
The other extremal case of this situation -also treated in [Mal13]- was when b = 1. Thanks
to Theorem 3.4 we can show that Conjecture 3.2.3 in [Mal13] does not hold.
Indeed, we have in this case a1 = a, b1 = a + 1. So, the cones here are the following:
C1 = h(0, 1), (a + 1, a)i
C2 = h(a + 1, a), (1, 0)i;
and then we have
HQ0 = {(0, 1), (1, 1), (a + 1, a)}
HQ1 = {(a + 1, a), (a, a − 1), . . . , (2, 1), (1, 0)}.
The cardinality of H is then a + 3. Theorem 3.4 gives then a list of (minimal) generators
of Bf ,F ,x1 . Denote a = n ≥ 2 to faciliate the comparison to Conjecture 3.2.3 in [Mal13].
2
One of them is Y(0,1) Y(2,1) − Y(1,2)
x1n−1 . We claim that this element does not belong to the
ideal proposed in [Mal13, Conjecture 3.2.3]. Indeed, with the notation there, this element
n−1
translates as x1 x4 − x23 xn+4
, which can be easily seen to be part of the kernel of ϕ0 . But if
we try to write it as a polynomial combination of the prospective generators of the ideal, and
then set all of the variables equal to zero except x1 and x4 , we would have
x1 x4 = p(x1 , x4 ) · x1 xn4 + q(x1 , x4 ) · x2n
4 ,
which is impossible.
P
max(3r,2s) r s
Example 3.14. Let us produce a presentation for BR,p (3, 2) =
)x1 x2 ⊂
r,s (p
R[x1 , x2 ], where p ∈ R is a nonzerodivisor, see Remark 1.7, (4). As in Example 2.8,
H = {v1 := (0, 1), v2 := (1, 2), v3 := (2, 3), v4 := (1, 1), v5 := (1, 0)} and M = 5. So,
BR,p (3, 2) = R[p3 x1 , p2 x2 , p4 x1 x22 , p3 x1 x2 , p6 x21 x32 ]. The map
ϕ0 : R[Y] → B
24
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
sends
Y1 → p2 x2 , Y2 → p4 x1 x22 , Y3 → p6 x21 x32 ,
Y4 → p3 x1 x2 , Y5 → p3 x1 .
Recall that our lexicographic order is of the form Y1 ≺ Y2 ≺ Y3 ≺ Y4 ≺ Y5 . We have
relations in the kernel of this map. They are
5−1
2
S1,3 = Y1 Y3 − Y22 , S1,4 = Y1 Y4 − pY2 , S1,5 = Y1 Y5 − p2 Y4
=6
S2,4 = Y2 Y4 − pY3 , S2,5 = Y2 Y5 − pY42 , S3,5 = Y3 Y5 − Y43 .
In this example, we have v1 + v5 = v4 and
max(3, 0) + max(0, 2) − max(3, 2) = 5 − 3 = 2,
which gives the relation Y1 Y5 − p2 Y4 . We denote this relation S1,5 instead of Sv1 ,v5 . The rest
are produced similarly. Therefore, BR,p (3, 2) = R[Y1 , . . . , Y5 ]/I, where
I = hY1 Y3 − Y22 , Y1 Y4 − pY2 , Y1Y5 − p2 Y4 , Y2Y4 − pY3 , Y2 Y5 − pY42 , Y3 Y5 − Y43 i.
4. The map ϕ1
With the notation and terminology of Section 2, we are going to study the syzygy module
M −1
M −1
syz Sf ,F ,p ⊂ R[Y]( 2 ) , where R[Y]( 2 ) is the free module of rank M2−1 over R[Y] with
basis {e(v, v′ ) , v ≺ v′ ∈ H, non-adjacents}. So, we have the map
M −1
ϕ1
syz Sf ,F ,p ⊂ R[Y]( 2 ) → R[Y]
(36)
e(v, v′ )
7→ Sv, v′ .
This syzygy submodule is also N2 -homogeneous if we declare
(37)
deg(e(v, v′ ) ) := v + v′ .
For v, v′ , w, w′ ∈ H such that neither v, v′ nor w, w′ are adjacents, we denote with
sv,v′ ,w,w′ the S-polynomial between Sv,v′ and Sw,w′ . As sv,v′ ,w,w′ ∈ ker(ϕ0 ), the division
algorithm of this polynomial against Sf ,F ,p gives
X
qz,z′ (Y)Sz,z′ ,
(38)
sv,v′ ,w,w′ =
z,z′
′
with qz,z′ (Y) ∈ R[Y], the sum being over all pair of non-adjacent z ≺ z′ ∈ H. Let Y α , Y α
′
be such that Y α Sv,v′ − Y α Sw,w′ = sv,v′ ,w,w′ . We then easily have that
X
′
(39)
Sv,v′ ,w,w′ := Y α e(v,v′ ) − Y α e(w,w′ ) −
qz,z′ (Y)e(z,z′ ) ∈ syz Sf ,F ,p .
z,z′
Note that
(40)
Y α ∈ {Yw , Yw′ , Yw Yw′ },
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
25
′
and similarly Y α ∈ {Yv , Yv′ , Yv Yv′ }. So, we have that
v + v′ + w + w′ if {v, v′ } ∩ {w, w′ } = ∅
v + v′ + w
if
w′ ∈ {v, v′ }
deg (Sv,v′ ,w,w′ ) =
v + v′ + w′
if
w ∈ {v, v′ }.
The following result will be useful in the sequel. It follows straightforwardly by noticing the
structure of the polynomials Sv, v′ defined in (21).
Lemma 4.1. Write Sv,v′ ,w,w′ as in (39). Then, all the vectors z, z′ appearing in the
expansion of this syzygy belong to the cone generated by {v, v′ , w, w′}. If v = w and
v ≺ v′ , w ≺ w′ , then the pairs z, z′ in (39) actually belong to the cone generated by
{z, v′, w′ }, with v ≺ z and v, z adjacent.
We will apply the results of Section 2 to this situation. To do this, we need to sort
the elements of the set Sf ,F ,p. As they are indexed by pairs of non-adjacent {vi , vj } with
1 ≤ i < j + 1 ≤ N, we will sort them by using the standard lexicographic order in N2 , i.e.
(i, j) ≺ (i′ , j ′ ) if and only if i < i′ , or i = i′ and j < j ′ .
Theorem 2.21 implies then that the collection {Svi ,vj ,vi′ ,vj′ } over all pairs of non-adjacent
(i, j) ≺ (i′ , j ′ ), generates syz Sf ,F ,p as an R[Y]-module, and it is a Gröbner basis of this
submodule for the induced order by this set, which will be denoted by ≺S . Note that this
induced order depends on the way we have sorted the elements of Sf ,F ,p (the lexicographic
M −1
order above) but it is not any kind of lexicographic order in R[Y]( 2 ) . To simplify notation,
we will also denote with e(i,j) the element e(vi , vj ) . Recall that we will also denote with Yj the
variable Yvj .
Definition 4.2. For i, j ∈ {1, . . . , M} with i+1 < j, and k ∈ {i+1, i+2, . . . , M}\{j −1, j},
we define
Svi , vj , vi , vk if i < j < k
(41)
∫i,j,k =
Svi , vj , vk , vj if i < k < j.
A triplet (i, j, k) for which ∫i,j,k is defined will be called “admissible”. For such an admissible triplet, by computing explicitly ∫i,j,k , we get that it is equal to either
(42)
X
X
qi,j,k,u,u′ (Y)e(u, u′ ) ,
qi,j,k,u,u′ (Y)e(u, u′ ) , or Yk e(i, j) − Yi e(k, j) +
Yk e(i, j) − Yj e(i, k) +
u<u′ −1
u<u′ −1
with
(43)
i ≤ u < u′ ≤ max{j, k}.
Lemma 4.3. With notation as above, if a term of the form rYt with r ∈ R \ {0} appears in
qi,j,k,u,u′ (Y), then the triplet (u, u′, t) is not admissible.
Proof. The presence of such a term implies that, in the process of doing the division algorithm
(38), we would have ran into a nonzero term of the form rYt Yu Yu′ . From here, to produce
the element rYt e(u,u′ ) , we should made the division of this term with the binomial Svu ,vu′
whose leading term is Yu Yu′ . But if the triplet (u, u′, t) is admissible, we would have that
26
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
u < u′ , u < t, and t ∈ {u + 1, u + 2, . . . , M} \ {u′ − 1, u′ }. Due to the way we have indexed the
set Sf ,F ,p above, and the fact that our Division Algorithm defined in Remark 2.10 chooses
the maximal index among those dividing the leading term, we have that the division of this
term must be done by Svu′ ,vt , except in the case when t = u′ + 1 (as t = u′ − 1 cannot happen
due to our choice of t), in which case the division should be done by Svu ,vt . In none of the
cases we get a term with coordinate e(u,u′ ) . This concludes the proof of the claim.
Lemma 4.4. For an admissible triplet (i, j, k), if vi , vj and vk are all contained in one
single cone of the fan, then ∫i,j,k ∈ hY1 , . . . , YM isyz Sf ,F ,p .
Proof. Suppose w.l.o.g. i < j < k, the other case being treated analogously. We have
′
β
α
Svi ,vj = Yi Yj − Ytα Yt+1
and Svi ,vk = Yi Yk − Yuβ Yu+1
,
′
with i ≤ t < j, i ≤ u < k, α + α′ ≥ 2 and β + β ′ ≥ 2, as all these identities involve vectors of
the Hilbert basis of the semigroup which generates the common cone in R2 . When computing
the S-polynomial between these two binomials, we will end up with another binomial each
of its terms has now total degree at least 3. Thanks to Lemma 4.1, we know that the whole
division algorithm which will produce the expression (39) will involve only pairs of nonadjacent vectors vt , vt′ in the common cone. This implies that at each step of the Division
Algorithm, a monomial of total degree κ is replaced by another of degree at least κ as the
γ′
with γ + γ ′ ≥ 2. We
division is always done against expressions of the form Yt Yt′ − Ymγ Ym+1
conclude then that the polynomial qi,j,k,u,u′ (Y) from (42) lies in hY1 , . . . , YM isyz Sf ,F ,p .
We recall again that for f is strict, we can choose a set of indices {k1 , . . . , kℓ−1 } such that
fki is non linear in Ci ∪ Ci+1 , i = 1, . . . , ℓ − 1.
Lemma 4.5. With notation as above, if a nonzero constant term of the form r ∈ R appears
in the support of one of the polynomials qi,j,k,u,u′ (Y) from (42), and f is strict, then there is
an index s ∈ {2, . . . , ℓ − 1} depending on u such that r ∈ hpks−1 , pks i.
Proof. As before, we can suppose w.l.o.g. i < j < k, the other case being treated analogously.
Choose s as the index such that vu ∈ Cs . As f is strict, each of the (u, u′) coming from (42)
which produce a nontrivial qi,j,k,u,u′ (Y) must satisfy (43). Moreover, due to Lemma 4.4, we
cannot have the three vectors vi , vj , vk lying in the same cone. This implies that there are
two possible scenarios:
(1) vi ∈ HQs∗ , with s∗ < s,
(2) vi ∈ HQs , and vk ∈ HQs∗ , with s < s∗ .
In the first case, by unravelling the Division Algorithm to produce ∫i,j,k , one concludes that
r ∈ hpks−1 i. In the second case, via the same analysis and using also Lemma 4.4 we conclude
that r ∈ hpks i. This concludes with the proof of the Lemma.
(2)
Denote with Sf ,F ,p := {∫i,j,k }, where the indices (i, j, k) run over all 1 ≤ i < j − 1 ≤
M, k ∈ {i + 1, . . . , n} \ {j − 1, j}.
PN −u j+u
+1
j u = (u + 1) Nu+2
Lemma 4.6. For u, N ∈ N such that 1 ≤ u < N, we have j=1
.
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
27
Proof. By global induction on N. If N = 2, the only possible value is u = 1, and the
claim follows straightforwardly. In the general case, if u < N, then by using the induction
hypothesis we compute
PN −u j+u
PN +1−u j+u
N +1
j
=
(N
+
1
−
u)
+ j=1 j u
j=1
u
u
(N +1)!
N +1
+1
1
1
+ u+2
= (N + 1 − u) u + (u + 1) Nu+2
= u!(N
−u−1)!
N
−u
+2
= (u + 1) Nu+2
.
The case u = N < N + 1 follows by a direct computation.
The following set will be useful in computing syzygies.
Definition 4.7. For 1 ≤ u ≤ M − 3, let Su be the set of those ordered (u + 2)-tuples
(i, j, k1 , . . . , ku ) such that i < j − 1 ≤ M − 1, 1 ≤ i < k1 < k2 < . . . < ku ≤ M and
km ∈
/ {j − 1, j} for all m = 1, . . . , u.
−1
Lemma 4.8. The cardinality of Su is equal to (u + 1) M
.
u+2
Proof. To count the elements of Su we proceed as follows: as the indices j − 1 and j must be
disjoint from the whole sequence i < k1 < . . . < ku , we first count the number of sequences
k1 < k2 < . . . < ku contained in the set {i + 1, . . . , M − 1}. This number is clearly M −1−i
.
u
For each such distribution, we can “choose” as the value of j − 1 any of the remaining
numbers not taken yet, and “expand” the sequence by adding j to the right (and hence
expanding the cardinality of the set up to M . The number of positions available for j − 1
is clearly equal to (M − 1 − u − i). So, the cardinality of Su is
MX
−u−2
MX
−u−2
M −1
j+u
M −1−i
,
= (u + 1)
(M − 1 − u − i) =
j
u+2
u
u
j=1
i=1
the last equality follows thanks to Lemma 4.6
(2)
Corollary 4.9. The cardinality of Sf ,F ,p
is equal to 2 M3−1 .
Proof. The claim follows straightforwardly from Lemma 4.8 with u = 1.
(2)
Theorem 4.10. If |F | is a proper convex polyhedral cone, Sf ,F ,p is a minimal Gröbner basis
of syz Sf ,F ,p for the induced order ≺S . If in addition, f is strict with respect to F , none of
the pki ’s is a zero divisor, and for 1 ≤ i ≤ j ≤ ℓ − 1, hpki , pkj i =
6 R, then it is also a minimal
set of generators of this submodule.
Proof. From (41) and (19), we deduce straightforwardly that for the monomial order ≺S ,
(44)
lt(∫i,j,k ) = Yk e(i, j) .
Theorem 2.21 implies that the set {Svi , vj , vk , vu } indexed by pairs of pairs of non-adjacent
(i, j) ≺ (k, u) in H, generate syz Sf ,F ,p . In addition, from (40) we get that the leading
term of each of these elements is a monomial of total degree one or two in the variables Y’s
times e(i, j) . To show that it is a Gröbner basis, it will be enough to show that any of these
monomials is a multiple of Yk , with k ∈ {i + 1, . . . , M} \ {j − 1, j}.
28
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
We straightforwardly verify that, for (i, j) ≺ (k, u),
lcm(Yi Yj , Yk Yu )
(45)
= lt Svi , vj , vk , vu ∈ {Yk , Yu , Yk Yu }.
Yi Yj
It is also easy
that neither Yi nor Yj can appear in (45). So, the piece in degree
to conclude
(ei , ej ) of Lt syz Sf ,F ,p is contained in
hYi+1 , . . . , Yj−1, Yj+1, . . . , YM i e(i,j) .
From (44), we only need to show that Yj−1 e(i, j) ∈
/ Lt syz Sf ,F ,p . But if this were the case,
then it should come from either Svi , vj , vi , vj−1 - which cannot happen as in our order we have
(i, j −1) ≺ (i, j), or from Svi ,vj ,vj−1 ,vj . But vj−1 and vj are adjacents, so the latter expression
(2)
has no sense. These arguments imply that Sf ,F ,p is a Gröbner basis of syz Sf ,F ,p . The fact
that it is minimal follows straightforwardly as if we remove one of these elements, the leading
module generated by the remaining family will not contain the leading term of the removed
syzygy.
To prove now that it is also a minimal set of generators of this submodule given the extra
hypothesis, we proceed as follows: assume that one of them is a polynomial combination of
the others. So, we must have
X
(46)
∫i,j,k =
qi′ ,j ′,k′ (Y)∫i′,j ′,k′
i′ ,j ′ ,k ′
for admissible triplets (i′ , j ′ , k ′ ), with qi′ ,j ′ ,k′ (Y) ∈ R[Y] of Z2 -degree vi + vj + vk − vi′ −
vj ′ − vk′ . Thanks to Lemma 4.1, we get that the e(i,j) coordinates in (46) produces
X
qi′1 ,j1′ ,k1′ (Y)qi′1 ,j1′ ,k1′ ,i,j (Y)
(47)
Yk = −qi,k,j (Y)Yk +
i′1 ,j1′ ,k1′
for some triplets (i′1 , j1′ , k1′ ) which must verify i′1 ≤ i < j ≤ max{j1′ , k1′ } due to (43). Note
that −qi,k,j (Y)Yk appears if and only if (i, k, j) is also an admissible triplet.
Now we evaluate Yu 7→ 0 for all u 6= k in (47) to get
(48)
X
qi′1 ,j1′ ,k1′ (0, . . . , Yk , . . . , 0)qi′1 ,j1′ ,k1′ ,i,j (0, . . . , Yk , . . . , 0)
Yk = −qi,k,j (0, . . . , Yk , . . . , 0)Yk +
i′1 ,j1′ ,k1′
which we write
(49)
Yk = −Qi,k,j (Yk )Yk +
X
Qi′1 ,j1′ ,k1′ (Yk )Qi′1 ,j1′ ,k1′ ,i,j (Yk ) ∈ R[Yk ].
i′1 ,j1′ ,k1′
As the polynomials qi,k,j (Y), qi′1 ,j1′ ,k1′ (Y), qi′1 ,j1′ ,k1′ ,i,j (Y) ∈ R[Y] are Z2 -graded, this applies
also to Qi,k,j (Yk ), Qi′1 ,j1′ ,k1′ (Yk ), Qi′1 ,j1′ ,k1′ ,i,j (Yk ) ∈ R[Yk ], from which we deduce that they are
all monomials. Moreover, as the Z2 -degree of qi,k,j (Y) is (0, 0), we have then that (49)
becomes
X
Ai′ ,j ′ ,k′
Bi′ ,j ′ ,k′ ,i,j
ri′1 ,j1′ ,k1′ Yk 1 1 1 ri′1 ,j1′ ,k1′ ,i,j Yk 1 1 1
(50)
Yk = −ri,k,j Yk +
i′1 ,j1′ ,k1′
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
29
for suitable ri,k,j , ri′1 ,j1′ ,k1′ , ri′1 ,j1′ ,k1′ ,i,j ∈ R, Ai′1 ,j1′ ,k1′ , Bi′1 ,j1′ ,k1′ ,i,j ∈ N. Note that ri,k,j = qi,k,j (0).
If 1+ri,k,j 6= 0, then from (50) we deduce that there must be an admissible triplet (i′1 , j1′ , k1′ )
such that ri′1 ,j1′ ,k1′ · ri′1 ,j1′ ,k1′ ,i,j 6= 0, and Ai′1 ,j1′ ,k1′ + Bi′1 ,j1′ ,k1′ ,i,j = 1. From Lemma 4.3, we deduce
Bi′ ,j ′ ,k′ ,i,j
that the case Bi′1 ,j1′ ,k1′ ,i,j = 1 cannot happen, as ri′1 ,j1′ ,k1′ ,i,j Yk 1 1 1 is one of the terms of
qi′1 ,j1′ ,k1′ ,i,j (Y), and the triplet (i, j, k) is admissible.
So, we must have Bi′1 ,j1′ ,k1′ ,i,j = 0, which by Lemma 4.4 can only happen if not all of
{vi′1 , vj1′ , vk1′ } lie in the same cone. By Lemma 4.5, we deduce that ri′1 ,j1′ ,k1′ ,i,j ∈ hpks−1 , pks i,
s being the index of the cone containing vi . This implies that 1+ri,k,j ∈ hpks−1 , pks i. Note that
the latter also holds trivially if 1 + ri,k,j = 0, so we can remove from now on our assumption
on the value of this element. In addition, we must have ri,k,j 6= 0 as otherwise we would
have 1 ∈ hpks−1 , pks i, contradicting the hypothesis. So, the Theorem would have been proven
already if the triplet (i, k, j) is not admissible.
To finish with our argument, we consider instead the coordinate e(i,k) in (46) and evaluate
Yu → 0 for u 6= j to get
X
qi′1 ,j1′ ,k1′ (0, . . . , Yj , . . . , 0)qi′1 ,j1′ ,k1′ ,i,k (0, . . . , Yj , . . . , 0)
(51)
0 = ri,k,j Yj +
i′1 ,j1′ ,k1′
instead of (48). The 0 at the left hand side of (51) is due to the fact that if (i, k, j) is
admissible and has as one of its terms Yk e(i,j) , then we must have
(52)
Yj e(i,k) = lt(∫i,k,j ) ≻ Yk e(i,j) = lt(∫i,j,k ),
both terms appearing in ∫i,k,j . If qi,j,k,i,k (Y), which appears in the expansion of ∫i,j,k in the
coordinate e(i,k) were nonzero, note that its Z2 -degree would be equal to vj and after the
evaluation Yu → 0 for u 6= j it would become rYj which thanks to Lemma 4.3 would imply
that r = 0. This justifies the 0 in the left hand side of (51), and from there we deduce as
before
X
Bi′ ,j ′ ,k′ ,i,k
Ai′ ,j ′ ,k′
ri′1 ,j1′ ,k1′ YYv 1 1 1 ri′1 ,j1′ ,k1′ ,i,k Yj 1 1 1 ,
(53)
− ri,k,j Yj =
j
i′1 ,j1′ ,k1′
for suitable ri,k,j , ri′1 ,j1′ ,k1′ , ri′1 ,j1′ ,k1′ ,i,j ∈ R, Ai′1 ,j1′ ,k1′ , Bi′1 ,j1′ ,k1′ ,i,j ∈ N. Arguing as before, we deduce that −ri,j,k ∈ hpks−1 , pks i, (the index s is the same as in the previous case, as it indexes
the cone Cs containing vi ) which, combined with 1 + ri,j,k ∈ hpks−1 , pks i, contradicts the
hypothesis of the Theorem. This concludes the proof of the Theorem.
Example 4.11. We will continue our running example with the computation of the kernel
of ϕ1 for the fan algebra BR,p (3, 2) from Remark 1.5, (4). The map
ϕ1 : R[Y]5 → R[Y]
sends ei,j → Si,j , and we expect 2 43 = 8 generating relations in the kernel, due to the
cardinality of S1 , as in Lemma 4.8. Sorted using the order induced by {e1,3 , . . . , e3,5 } with
30
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
the lexicographic order Y1 ≺ . . . ≺ Y5 in mind, these are
∫1,3,4
∫1,3,5
∫1,4,2
∫1,4,5
∫1,5,2
∫1,5,3
∫2,4,5
∫2,5,3
(54)
=
=
=
=
=
=
=
=
Y4 e1,3 − Y3 e1,4 + Y2 e2,4
Y5 e1,3 − Y3 e1,5 + pY4 e2,4 + Y2 e2,5
Y2 e1,4 − Y1 e2,4 − pe1,3
Y5 e1,4 − Y4 e1,5 + pe2,5
Y2 e1,5 − Y1 e2,5 − pY4 e1,4
Y3 e1,5 − Y1 e3,5 − pY4 e2,4 − Y42 e1,4
Y5 e2,4 − Y4 e2,5 + pe3,5
Y3 e2,5 − Y2 e3,5 − Y42 e2,4 .
For the first relation, we have the triplet (1, 3, 4) ∈ S1 and hence we need to consider the
elements e1,3 → S1,3 = Y1 Y3 − Y22 , e1,4 → S1,4 = Y1 Y4 − pY2 . The S-polynomial of S1,3 , S1,4 is
Y4 (Y1 Y3 − Y22 ) − Y3 (Y1 Y4 − pY2 ) = −Y2 (pY3 − Y2 Y4 ) = −Y2 S2,4 , so Y4 e1,3 − Y3 e1,4 + Y2 e2,4 is
in the kernel of ϕ1 , which is our first relation.
5. Higher Syzygies
From Theorem 4.10 we deduce now the following exact sequence of R[Y]-modules:
M −1
M −1
ϕ2
ϕ1
ϕ0
R[Y]2( 3 ) → R[Y]( 2 ) → R[Y] → Bf ,F ,p → 0
(55)
e(i,j,k)
7→
∫i,j,k
e(i, j)
7→
Si, j ,
L
M −1
where we have made the identification R[Y]2( 3 ) ≃ i,j,k R[Y] · e(i,j,k) , the sum being over
the admissible triplets (i, j, k).
To continue with a full resolution of (55), we do a process similar to (41) recursively: we
(2)
sort the elements in Sf ,F ,p with the lexicographic order in N3 , and given 1 ≤ i < j − 1 ≤ n
and k < k ′ with k, k ′ ∈ {i + 1, . . . , n} \ {j − 1, j} we compute the syzygy between ∫i,j,k and
∫i,j,k′ , which we denote by Si,j,k,k′ . Thanks to (44), we know that
X
qi,j,k,k′,i0 ,j0,k0 (Y)∫i0 ,j0 ,k0 ,
(56)
Si,j,k,k′ = Yk′ ∫i,j,k − Yk ∫i,j,k′ =
i0 ,j0 ,k0
where the expression in the right hand side comes from applying the division algorithm to
(2)
Yk′ ∫i,j,k − Yk ∫i,j,k′ with respect to Sf ,F ,p. We then define
X
qi,j,k,k′,i0 ,j0 ,k0 (Y)e(i0 ,j0,k0 ) ∈ ker(ϕ2 ),
(57)
∫i,j,k,k′ = Yk′ e(i,j,k) − Yk e(i,j,k′) −
i0 ,j0 ,k0
(3)
and set Sf ,F ,p the set of all these elements. As before, we will say that such a triplet (i, j, k, k ′ )
(3)
is admissible. It is easy to verify that the cardinality of Sf ,F ,p is equal to the number of
ordered 4-tuples (i, j, k,
k ′ ) with i < j − 1, k < k ′ with k, k ′ ∈ {i + 1, . . . , M} \ {j − 1, j}.
This number is 3 M4−1 , thanks to Lemma 4.8 with ℓ = 2.
The following is be the generalization of (43) to this case, its proof being straightforward
by using Lemma 4.1 and (43).
Lemma 5.1. With notations as above, we have that i ≤ i0 , j0 , k0 ≤ max{j, k, k ′ }.
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
31
Lemmas 4.3, 4.4, and 4.5 also have a straightforward generalization, which are given next.
Lemma 5.2. With notation as in (57), if a term of the form rYt with r ∈ R \ {0} appears
in qi,j,k,k′i0 ,j0 ,k0 (Y), then the triplet (i0 , j0 , k0 , t) is not admissible.
Proof. The presence of a term like that implies that in the process of doing the division
algorithm (38) we would have ran into a nonzero term of the form rYt Yk0 e(i0 ,j0 ) . From here,
to produce the element rYt e(i0 ,j0 ,k0 ) , we should divide this term with ∫i0 ,j0,k0 . It (i0 , j0 , k0 , t)
were admissible, we could also divide it with ∫i0 ,j0 ,t . As k0 < t, and due to the way we have
designed the Division Algorithm in Remark 2.10 by choosing the maximal index in the family
(2)
Sf ,F ,p , the division should be done with the latter, leading to a contradiction. This proves
the claim.
Lemma 5.3. For an admissible triplet (i, j, k, k ′ ), if vi , vj , vk and vk′ are all contained in
M −1
one single cone of the fan, then ∫i,j,k,k′ ∈ hY1 , . . . , YM iR[Y]2( 3 ) .
Proof. In the conditions of the hypothesis, we have that both ∫i,j,k and ∫i,j,k′ have all their
coefficients in hY1 , . . . , YM i thanks to Lemma 4.4. So, we actually have that the coefficients
of Yk′ ∫i,j,k − Yk ∫i,j,k′ belong to hY1 , . . . , YM i2 . The claim now follows straightforwardly as in
(2)
the division algorithm over elements of Sf ,F ,p , all the leading terms of these elements are one
of the variables Yu , u = 1, . . . , M.
Lemma 5.4. With notation as above, if a nonzero constant term of the form r ∈ R appears
in the Taylor expansion of one of the polynomials qi,j,k,k′,i0 ,j0 ,k0 (Y) from (57), and f is strict,
then there is an index s ∈ {2, . . . , ℓ − 1} depending on i0 such that r ∈ hpks−1 , pks i.
Proof. The proof follows again as in Lemma (4.5), by unravelling the Division Algorithm to
produce the polynomial Si,j,k,k′ defined in (56). It is easy to see that all the leading terms
during this Division Algorithm either belong to the ideal hY1 , . . . , YM i (and hence are not
under the conditions of this hypothesis), or are multiple of some pks for s ∈ {1, . . . , ℓ − 1}.
By using now Lemma 4.5, it is easy to conclude that the two possible indices corresponding
to qi,j,k,k′,i0 ,j0 ,k0 (Y) are those determined by the cone where vi0 lies. This concludes with the
proof of the Lemma.
By applying now all these claims, we can reproduce mutatis mutandis the proof of Theorem
4.10.
(3)
(2)
Theorem 5.5. The set Sf ,F ,p is a minimal Gröbner basis of syz Sf ,F ,p for the induced
order ≺S (2) . If |F | is a proper convex polyhedral cone, f is strict with respect to F , none of
the pki ’s is a zero divisor, and for 1 ≤ i ≤ j ≤ ℓ − 1, hpki , pkj i =
6 R, then it is also a minimal
set of generators of this submodule.
Proof. The fact that this set is a minimal Gröbner basis follow directly from Theorem 2.21
(3)
and comparing the leading terms of all possible syzygies and those coming in Sf ,F ,p .
To prove minimal generation,
we proceed as in the proof of Theorem 4.10: from an exP
pression like ∫i,j,k,k′ = i1 ,j1,k1 ,k′ qi1 ,j1 ,k1,k1′ (Y)∫i1,j1 ,k1 ,k1′ , we pick the coordinate ei,j,k in both
1
32
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
sides to get
(58)
Yk ′ =
X
qi1 ,j1 ,k1 ,k1′ (Y)qi1 ,j1 ,k1 ,k1′ ,i,j,k (Y)
i1 ,j1 ,k1 ,k1′
(note that in this case there cannot be a term of the form −qi,j,k′,k (Y)Yk′ as in (47) due to
the fact that (i, j, k ′ , k) cannot be an admissible element as k < k ′ , so the proof gets actually
simpler in this case). Now we set Yu 7→ 0 for u 6= k ′ , and proceed mutatis mutandis the
proof of Theorem 4.10 by using now Lemmas 5.1, 5.2, 5.3, and 5.4.
M −1
M −1
In general, for s ≥ 3, given the map ϕs : R[Y]s( s+1 ) → R[Y](s−1)( s ) where the canonical
M −1
basis of R[Y]s( s+1 ) is denoted with ei,j,k1,...,ks−1 with i < j − 1, i < k1 < k2 < . . . <
ks−1 , and no kt ∈ {j − 1, j}, and we have ei,j,k1 ,...,ks−1 7→ ∫i,j,k1,...,ks−1 with lt(∫i,j,k1,...,ks−1 ) =
M −1
such that k ∈
/ {j − 1, j}, we can
Y
e
∈ R[Y](s−1)( s ) . For a given k > k
ks−1 i,j,k1 ,...,ks−2
s
s−1
s
proceed analogously as done in (56) and (57), to produce ∫i,j,k1,...,ks ∈ ker(ϕs ) such that
(s+1)
lt(∫i,j,k1,...,ks ) = Yks ei,j,k1,...,ks−1 . Denote with Sf ,F ,p the set of all such elements ∫i,j,k1,...,ks ,
−1
. We sort as usual this set by using
which by Lemma 4.8 will have cardinality (s + 1) Ms+2
the lexicographic order on (i, j, k1 , . . . , ks−1 ).
Straightforward generalizations of Lemmas 5.1, 5.2, 5.3, and 5.4 imply the following higher
dimensional result.
(s+1)
(s)
Theorem 5.6. The set Sf ,F ,p is a minimal Gröbner basis of syz Sf ,F ,p for the induced
order ≺S (s) . If |F | is a proper convex polyhedral cone, f is strict with respect to F , none of
the pki ’s is a zero divisor, and for 1 ≤ i ≤ j ≤ ℓ − 1, hpki , pkj i =
6 R, then it is also a minimal
set of generators of this submodule.
Now we are in position to prove our main Theorem stated in the Introduction.
Proof of Theorem 1.9. The ring R[Y] is *-local with respect to the ideal (m, Y), because
(m, Y) is a maximal ideal which is Z2 -graded, and therefore it is also Z-graded under the
induced grading. According to the comments following Example 1.5.14 in [BH], to have a
minimal graded free resolution over R[Y], we need to check that the kernel of ϕ0 is contained
M −1
in (m, Y), and also that for each j ≥ 1, ker(ϕj ) ⊂ (m, Y)R[Y]j ( j+1 ) . All these follow from
Theorems 3.4, 4.10, 5.5 and 5.6.
An immediate corollary of this result is
Corollary 5.7. Under the notations of Theorem 1.9, the projective dimension of Bf ,F ,p over
R[Y] is M − 2, if (R, m) is local, f is strict with respect to F , none of the pi ’s is a zero
divisor in R, and all of them belong m.
Example 5.8. We return to the Example 1.5 and continue Example 4.11 with the computation of the kernel of ϕ2 : R[Y]8 → R[Y]6 , given by ei,j,k → ∫i,j,k , where (i, j, k) ∈ S1 . We
need to find the set S2 = {(i, j, k, k ′ ) : 1 ≤ i < j − 1 ≤ 5, k < k ′ ∈ {i + 1, . . . , 5} \ {j − 1, j}},
see Definition 4.7. One gets S2 = {(1, 3, 4, 5), (1, 4, 2, 5), (1, 5, 2, 3)}.
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
33
To compute S-polynomials here, it is better to write down explicitly “matrix” of the syzygies
∫i,j,k : we will encode in the following 8 × 6 matrix the coordinates of the the syzygies in (54)
in the basis {e1,3 , e1,4 , e1,5 , e2,4 , e2,5 , e3,5 } :
Y4 −Y3
0
−Y2
0
0
Y5
0
−Y3 pY4
Y2
0
0
−Y1
0
0
−p Y2
Y5
−Y4
0
p
0
0
.
0
−Y1
0
0 −pY4 Y2
0 −Y 2 Y
−pY4
0 −Y1
3
4
0
0
0
Y5
−Y4 p
0
0
0
−Y42 Y3 −Y2
This matrix has rank 5, and computing explicitly a basis of its kernel, we get a basis of 3
elements. From here it is easy to deduce the higher syzigies:
∫1,3,4,5 = Y5 e1,3,4 − Y4 e1,3,5 + Y3 e1,4,5 − Y2 e2,4,5 − pe2,5,3
∫1,4,2,5 = Y5 e1,4,2 − Y2 e1,4,5 + pe1,3,5 − Y4 e1,5,2 + pe1,5,3 + Y1 e2,4,5
∫1,5,2,3 = Y3 e1,5,2 − Y2 e1,5,3 + Y1 e2,5,3 − pY4 e1,3,4 − Y42 e1,4,2 .
Alternatively, we can compute these relations by hand. For (1, 3, 4, 5) ∈ S2 we need to
consider the S-polynomial between ∫1,3,4 and ∫1,3,5 which is Y5 ∫1,3,4 − Y4 ∫1,3,5 = −Y3 Y5 e1,4 +
Y2 Y5 e2,4 + Y3 Y4 e1,5 − py42e2,4 − Y2 Y4 e2,5 . The leading term in this expression is −Y3 Y5 e1,4 so
we can write Y5 ∫1,3,4 − Y4 ∫1,3,5 = −Y3 ∫1,4,5 + pY3 e2,5 + Y2 Y5 e2,4 − pY42 e2,4 − Y2 Y4 e2,5 . The next
leading term is now −Y2 Y4 e2,5 , so we need to use ∫2,4,5 which leads to Y5 ∫1,3,4 − Y4 ∫1,3,5 =
Y3 ∫1,4,5 − Y2 ∫2,4,5 − p∫2,5,3 . This in turn gives the relation:
∫1,3,4,5 = Y5 e1,3,4 − Y4 e1,3,5 + Y3 e1,4,5 − Y2 e2,4,5 − pe2,5,3 .
The rest can be obtained similarly.
We can now write the minimal resolution of BR,p (3, 2) over R[Y1 , . . . , Y5 ] = R[Y] as
ϕ3
ϕ2
ϕ1
ϕ0
0 → R[Y]3 → R[Y]8 → R[Y]6 → R[Y] → BR,p (3, 2) → 0
where the matrix corresponding to ϕ2 with respect to the bases {e1,3 , e1,4 , e1,5 , e2,4 , e2,5 , e3,5 }
and {∫1,3,4 , ∫1,3,5, ∫1,4,2 , ∫1,4,5, ∫1,5,2 , ∫1,5,3, ∫2,4,5, ∫2,5,3 } is
Y4 −Y3
0
−Y2
0
0
Y5
0
−Y3 pY4
Y2
0
−p Y2
0
−Y1
0
0
0
Y5
−Y4
0
p
0
.
0 −pY4 Y2
0
−Y1
0
0 −Y42 Y3 −pY4
0 −Y1
0
0
0
Y5
−Y4 p
0
0
0
−Y42 Y3 −Y2
and the matrix corresponding to ϕ3 with respect to the bases {∫1,3,4, ∫1,3,5 , ∫1,4,2, ∫1,4,5 , ∫1,5,2, ∫1,5,3 , ∫2,4,5}
and {∫1,3,4,5 , ∫1,4,2,5, ∫1,5,2,3} is
34
TERESA CORTADELLAS BENÍTEZ, CARLOS D’ANDREA, AND FLORIAN ENESCU
Y5
−Y4
0
−Y3
0
0 −Y2 p
0
p
Y5 −Y2 −Y4
p
Y1 0 .
2
−pY4
0 −Y4
0
Y3 −Y2
0 Y1
6. The case |F | = R2
We conclude this paper by showing some examples where the support of the fan is the
whole plane, and our results do not hold in general in this case.
We start with the simplest situation which is v1 = (1, 0), v2 = (0, 1), v3 = (−1, −1). Every
pair of these vectors are adjacent, and there are three cones in this fan. So, the set Sf ,F ,p from
Definition 3.2 is empty, and hence according to Theorem 3.4, ϕ0 should be an isomorphism.
However, it is straightforward to verify that Y1 Y2 Y3 − pf (v1 )+f (v2 )+f (v3 ) ∈ ker(ϕ0 ), so that
statement cannot be true. Note also that in this example the only generator of the kernel
does not have total degree 2 as in the other case.
Still for this example we have a resolution like (7), so one may wonder whether a statement
like Theorem 1.9 holds, but with different maps ϕj than those defined in the text. The
following example shows that this is not the case anymore: set
v1 = (1, 0), v2 = (1, 1), v3 = (0, 1), v4 = (−1, 0), v5 = (−1, −1), v6 = (0, −1),
and the map Yj 7→ pj , j = 1, . . . , 6. If we set R to be Q[p1, p2, p3 , p4 , p5 , p6 ] where each pi is
an indeterminate, by computing explicitly the kernel of ϕ0 we get the following minimal set
of 11 generators:
{Y3 Y3 − p3 p6 , Y2 Y5 − p2 p5 , Y1 Y4 − p1 p4 , Y2 Y4 Y6 − p2 p4 p6 , Y1 Y3 Y5 − p1 p3 p5 , p6 Y1 Y5 − p1 p5 Y6 ,
p5 Y 4 Y 5 − p4 p6 Y 5 , p4 Y 3 Y 5 − p3 p5 Y 4 , p3 Y 2 Y 4 − p2 p4 Y 3 , p2 Y 1 Y 3 − p1 p3 Y 2 , p2 p6 Y 1 − p1 Y 2 Y 6 } .
As in the previous case, we note that there are elements of degree 3 among the generators, and
also elements which belong to the ideal generated by the pi ’s, so all our previous methods do
not apply here and new ideas are needed in order to deal with this situation. The resolution
of this fan algebra can be computed explicitly, and is equal to
0 → R[Y]7 → R[Y]24 → R[Y]27 → R[Y]11 → R[Y] → Bf ,F ,I → 0.
Acknowledgements
All our computations and experiments were done with the aid of the softwares Macaulay
2([Mac]) and Mathematica ([Wol10]). The last author thanks Sandra Spiroff for discussions
related to the second section of the paper. Part of this work was done while the second
author and third author were visiting Georgia State University and, respectively, Universitat
de Barcelona. They are grateful for the support and working atmosphere provided by these
institutions.
RESOLUTIONS OF FAN ALGEBRAS OF PRINCIPAL IDEALS
35
References
[AL94]
Adams, William W.; Loustaunau, Philippe. An introduction to Gröbner bases. Graduate Studies
in Mathematics, 3. American Mathematical Society, Providence, RI, 1994.
[BH]
Bruns, Winfred; Herzog, Jurgen. Cohen-Macaulay rings. Cambridge University Press.
[CES07] Ciupercă, Cătălin; Enescu, Florian; Spiroff, Sandra. Asymptotic growth of powers of ideals. Illinois
J. Math. 51 (2007), no. 1, 29–39.
[E96]
Eisenbud, David. Commutative algebra with a view towards algebraic geometry. Springer, GTM
150, ISBN 0-387-94269-6, second edition, 1996.
[EM15] Enescu, Florian; Malec, Sara. Intersection algebras for principal monomial ideals in polynomial
rings. J. Algebra Appl. 14 (2015), no. 7, 1550108, 23 pp.
[ES16] Enescu, Florian; Spiroff, Sandra. Computing the invariants of the intersection algebras for principal
monomial ideals. preprint, 2016.
[F02]
Fields, J. Bruce. Lengths of Tors determined by killing powers of ideals in a local ring. J. Algebra
247 (2002), no. 1, 104–133.
[H99]
Huneke, Craig. Hyman Bass and ubiquity: Gorenstein rings. Algebra, K-theory, groups, and education (New York, 1997), 55-78, Contemp. Math., 243, Amer. Math. Soc., Providence, RI, 1999.
[Mac]
Grayson, Daniel R.; Stillman, Michael E. Macaulay 2, a software system for research in algebraic
geometry. Avabilable at http://www.math.uiuc.edu/Macaulay2/
[Mal13] Malec, Sara. Intersection Algebras and pointed rational cones. Ph.D. Thesis, 2013.
[Mal15] Malec, Sara. On the Intersection Algebra of Principal Ideals. Comm. Algebra 43 (2015), no. 2,
623–635.
[MS]
Miller, Ezra; Sturmfels, Bernd Combinatorial Commutative Algebra. Springer, GTM 227, ISBN
0-387-22356-8, 2005.
[Wol10] Wolfram Research, Inc. Mathematica, Version 8.0, Champaign, IL (2010).
Facultat de Educació, Universitat de Barcelona. Passeig de la Vall d’Hebron 171, 08035
Barcelona, Spain
E-mail address: [email protected]
Facultat de Matemàtiques i Informàtica, Universitat de Barcelona. Gran Via 585, 08007
Barcelona, Spain
E-mail address: [email protected]
URL: http://atlas.mat.ub.es/personals/dandrea
Department of Mathematics and Statistics, Georgia State University, Atlanta, GA 30303
USA
E-mail address: [email protected]
URL: http://www2.gsu.edu/~matfxe/
| 0math.AC
|
1
Reconstruction of Hidden Representation
for Robust Feature Extraction
arXiv:1710.02844v1 [cs.LG] 8 Oct 2017
Zeng Yu, Tianrui Li, Senior Member, IEEE, Ning Yu, Yi Pan, Senior Member, IEEE,
Hongmei Chen, Member, IEEE, and Bing Liu, Fellow, IEEE
Abstract—This paper aims to develop a new and robust approach to feature representation. Motivated by the success of AutoEncoders, we first theoretical summarize the general properties of all algorithms that are based on traditional Auto-Encoders: 1) The
reconstruction error of the input or corrupted input can not be lower than a lower bound, which can be viewed as a guiding principle for
reconstructing the input or corrupted input. 2) The reconstruction of a hidden representation achieving its ideal situation is the necessary
condition for the reconstruction of the input to reach the ideal state. 3) Minimizing the Frobenius norm of the Jacobian matrix has a
deficiency and may result in a much worse local optimum value. 4) Minimizing the reconstruction error of the hidden representation is
more robust than minimizing the Frobenius norm of the Jacobian matrix. Based on the above analysis, we propose a new model termed
Double Denoising Auto-Encoders (DDAEs), which uses corruption and reconstruction on both the input and the hidden representation.
We demonstrate that the proposed model is highly flexible and extensible. We also show that for handling inessential features, our
model is more robust than Denoising Auto-Encoders (DAEs). Comparative experiments illustrate that our model is significantly better
for representation learning than the state-of-the-art models.
Index Terms—Deep architectures, auto-encoders, unsupervised learning, feature representation, reconstruction of hidden representation.
F
1
I NTRODUCTION
R
EPRESENTATION learning via deep neural networks
has developed into an important area of machine
learning research in recent years. This development has
also witnessed a wide range of successful applications in
the fields of computer vision [1], speech recognition [2],
and natural language processing [3]. Reviews of recent
progresses can be found in [4], [5], [6], [7].
A deep neural network usually has a deep architecture
that uses at least one layer to learn the feature representation of the given data. A representation learning
procedure is applied to discover multiple levels of representation: the higher the level, the more abstract the
representation. It has been shown that the performance
of deep neural networks is heavily dependent on the
multilevel representation of the data [7]. In the past
few years, researchers have endeavored to design a
variety of efficient deep learning algorithms that may
capture some characteristics of the data-generating distribution. Among these algorithms, the traditional AutoEncoders (AEs) [8] perhaps received the most research
attention due to their conceptual simplicity, ease of
• Z. Yu, T.R. Li (corresponding author) and H.M. Chen are with the
School of Information Science and Technology, Southwest Jiaotong University, Chengdu 611756, China. E-mail: [email protected], {trli, hmchen}@swjtu.edu.cn
• N. Yu is with Department of Informatics, University of South Carolina
Upstate, Spartanburg, SC 29303, USA. E-mail: [email protected]
• Y. Pan is with the Department of Computer Science, Georgia State
University, Atlanta, GA 30303, USA. E-mail: [email protected]
• B. Liu is with the Department of Computer Science, University of Illinois
at Chicago, Chicago, IL 60607, USA. E-mail: [email protected]
training, and inference and training efficiency. They are
used to learn the data-generating distribution of the
input dataPby minimizing the reconstruction error of
the input x∈X L(x, g(f (x))), where f (.) is the encoder
function, g(.) is the decoder function and L(.) is the reconstruction error. Recently, they have become one of the
most promising approaches to representation learning
for estimating the data-generating distribution. Since the
appearance of Auto-Encoders, many variants of representation learning algorithms based on Auto-Encoders
have been proposed, e.g., Sparse Auto-Encoders [9],
[10], Denoising Auto-Encoders (DAEs) [11], Higher Order Contractive Auto-Encoders [12], Variational AutoEncoders [13], Marginalized Denoising Auto-Encoders
[14], Generalized Denoising Auto-Encoders [15], Generative Stochastic Networks [16], MADE [17], Laplacian
Auto-Encoders [18], Adversarial Auto-Encoders [19],
Ladder Variational Auto-Encoders [20] and so on.
In an Auto-Encoder-based algorithm, minimizing the
reconstruction error of the input with the encoder and
decoder function is a common practice for feature learning. The learned features are usually applied in subsequent tasks such as supervised classification [21]. In
the past few years, many research works have shown
that the reconstruction of the input with the encoder
and decoder function is not only an efficient way for
learning feature representation, but its resulting representations also substantially help the performance of
the subsequent tasks. In general, a lower value of the
reconstruction error of the input has a better feature
representation of the input. In an ideal situation, the
value of this reconstruction error is equal to 0, i.e., the
2
input can be completely reconstructed. In this paper,
we show that the reconstruction error of the input from
every traditional Auto-Encoders based algorithm has a
lower bound, which is greater than or equal to 0.
As an important method of representation learning,
minimizing the Frobenius norm of the Jacobian matrix
of the hidden representation has been widely used in
deep learning models. The first application is the CAEs
[22], which try to learn locally invariant features by
minimizing the Frobenius norm of the Jacobian matrix
of hidden representation. After that, many frameworks
based on minimizing the Frobenius norm of the Jacobian
matrix of hidden representation have been developed in
computer vision tasks. Specifically, Liu et al. [23] developed a multimodal feature learning model with stacked
CAEs for video classification. To find stable features,
Schulz et al. [24] designed a two-layer encoder which is
regularized by an extension of a previous work on CAEs.
Geng et al. [25] proposed a novel deep supervised and
contractive neural network for SAR image classification
by using the idea of minimizing the Frobenius norm of
the Jacobian matrix of hidden representation. Shao et
al. [26] introduced an enhancement deep feature fusion
method for rotating machinery fault diagnosis through
a combination of DAEs and CAEs. However, we will
demonstrate that minimizing the Frobenius norm of the
Jacobian matrix of hidden representation has a deficiency
in learning feature representation.
To learn robust feature representation, minimizing the
reconstruction error of hidden representation is also
important and efficient. This idea has been emphasized
by popular deep learning algorithms such as Ladder
Networks [27], [28] and Target Propagation Networks
[29]. In order to reconstruct the hidden representation,
Ladder Networks need two streams of information to
reconstruct the hidden representation: one is used to
generate a clean hidden representation with an encoder
function; the other is utilized to reconstruct the clean
hidden representation with a combinator function [27],
[28]. The final objective function is the sum of all the
reconstruction errors of hidden representation. It should
be noted that reconstructing the hidden representation
in each layer needs to use information of two layers,
which makes Ladder Networks difficult to be trained
with a layer-wise pre-training strategy. Training a deep
learning model in a layer-wise manner, as it is known,
is an unsupervised learning approach, which may have
many potential advantages. To reconstruct the hidden
representation, Target Propagation Networks [29] can
be trained in a layer-wise manner. Nevertheless, in
the Target Propagation Networks, reconstructing hidden
representation is decomposed into two separate targets,
which may be trapped into a local optimum. To the best
of our knowledge, reconstructing hidden representation
as a whole and training it in a layer-wise manner has
not yet been investigated so far.
In this paper, we first study the general properties of
all algorithms based on the traditional Auto-Encoders.
We aim to design a robust approach for feature representation based on these properties. We follow the framework of layer-wise pre-training and consider the idea
of reconstruction of hidden representation. We propose
a new deep learning model that takes the advantages
of corruption and reconstruction. Our model consists of
two separate parts: constraints on the input (Constraints
Part) and reconstruction on the hidden representation
(Reconstruction Part). Constraints Part can be viewed as
a traditional deep learning model such as auto-encoder
and its variants. Reconstruction Part can be viewed
as explicitly regularizing the hidden representation or
adding additional feedback to the pre-training stage.
For simplicity and convenience, we use a DAE as the
Constraints Part to build our model. Because the best
results are obtained by utilizing the corruption in both
input and hidden representation, we refer it as Double
Denoising Auto-Encoders (DDAEs).
The contributions of this paper are summarized as
follows:
•
•
•
We prove that for all algorithms based on traditional
Auto-Encoders, the reconstruction error of the input
can not be lower than a lower bound, which can
sever as a guiding principle for the reconstruction of
the input. We also show that the necessary condition
for the reconstruction of the input to reach the ideal
state is that the reconstruction of hidden representation achieves its ideal condition. When the input
is corrupted with noises, we also demonstrate that
the reconstruction error of the corrupted input also
can not be lower than a lower bound.
We validate that minimizing the Frobenius norm
of the Jacobian matrix of the hidden representation
has a deficiency and may result in a much worse
local optimum value. We also show that minimizing
reconstruction error of the hidden representation
for feature representation is more robust than minimizing the Frobenius norm of the Jacobian matrix,
which may be the main reason why the proposed
DDAEs always outperform CAEs.
We propose a new approach to learning robust feature representations of the input based on the above
evidences. Compared with the existing methods,
DDAEs have the following advantages: 1) DDAEs
are flexible and extensible and have a potentially
better capability of learning invariant and robust
feature representations. 2) For dealing with noises
or some inessential features, DDAEs are more robust than DAEs. 3) DDAEs can be trained with
two different pre-training strategies by optimizing
the objective function in a combining or separate
manner, respectively.
The rest of this paper is organized as follows. Section 2
introduces the basic DAEs and CAEs. Section 3 presents
the lower bound of the reconstruction error of the input
and the necessary condition for the reconstruction of
the input to reach its ideal state. Section 4 illustrates
3
the defect of CAEs and gives a theoretical proof on
why DDAEs can outstrip CAEs. Section 5 describes
the proposed DDAEs framework. Section 6 compares
the performance of DDAEs with other relevant state-ofthe-art representation learning algorithms using various
testing datasets. Conclusions together with some further
studies are summarized in the last section.
2
P RELIMINARIES
DDAEs are designed according to the traditional AutoEncoders [8] that learn feature representation by minimizing the reconstruction error. For ease of understanding, we reveal DDAEs by starting to describe some
conventional auto-encoder variants and notations.
2.1 Denoising Auto-encoders (DAEs): Extracting
Robust Features of Reconstruction
Similar to traditional Auto-Encoders [8], the Denoising
Auto-Encoders (DAEs) [11] firstly use the encoder and
decoder procedure to train one-layer neural network by
minimizing the reconstruction error, and then stack a
deep neural network with the trained layers. The only
difference between traditional Auto-Encoders and DAEs
is that DAEs train the neural network with corrupted input while the traditional Auto-Encoders use the original
input. The corrupted input x̃ ∈ <Dx is usually obtained
from a conditional distribution q(x̃|x) by injecting some
noises into the original input x ∈ <Dx . Typically, the
most widely-used noises in the simulations are Gaussian
noise x̃ = x + , ∼ N (0, σ 2 I) and masking noise, where
ν% (ν is given by researchers) of the input components
are set to 0.
To extract robust features, a DAE firstly maps the
corrupted input x̃ to a hidden representation h ∈ <Dh
by the encoder function f :
h = f (x̃) = Sf (Wx̃ + bh ),
(1)
where W ∈ <Dh ×Dx is a connection weight matrix, bh ∈
<Dh is a bias vector of hidden representation and Sf is
an activation function, typically a logistic sigmoid(τ ) =
1
1+e−τ . After that, the DAE reversely maps the hidden
representation h back to a reconstruction input x∗ ∈ <Dx
through the decoder function g:
x∗ = g(h) = Sg (W0 h + bx ),
(2)
where W0 ∈ <Dx ×Dh is a tied weight matrix, i.e., W0 =
WT , bx ∈ <Dx is a bias vector and Sg is an activation
function, typically either the identity (yielding linear
reconstruction) or a sigmoid. Finally, the DAE learns the
robust features by minimizing the reconstruction error
on a training set X = {x1 , x2 , · · · , xN }.
X
JDAE (θ) =
E [L(x, g(f (x̃)))] ,
(3)
x∈X
where θ = {W, bx , bh }, E(δ) is the mathematical expectation of δ, L is the reconstruction error. Typically the
squared error L(x, y) = kx − yk2 is used when Sg is the
identity
i loss L(x, y) =
hP function and the cross-entropy
Dx
x
log(x
)
+
(1
−
x
)log(1
−
x
)
is selected when
−
i
i
i
i=1 i
Sg is the sigmoid function.
It has been shown that DAEs can extract robust features by injecting some noises into the original input
and implicitly capture the data-generating distribution
of input in the conditions that the reconstruction error
is the squared error and the data are continuous-valued
with Gaussian corruption noise [30], [15], [31].
2.2 Contractive Auto-encoders (CAEs): Extracting
Locally Invariant Features of Hidden Representation
To extract locally invariant features, the CAEs [22] penalize the sensitivity by adding an analytic contractive
penalty to the traditional Auto-Encoders. The contractive penalty is the Frobenius norm of first derivatives
kJf (x)k2F of the encoder function f (x) with respect to
the input x.
Formally, the objective optimized by a CAE is
X
JCAE (θ) =
L(x, g(f (x))) + αkJf (x)k2F , (4)
x∈X
where α is a hyper parameter that controls the strength
of the regularization. For a sigmoid encoder, the contractive penalty is simply computed:
kJf (x)k2F =
Dh
X
(hj (1 − hj ))2
j=1
Dx
X
Wij2 .
(5)
i=1
Compared with DAEs, the CAEs have at least two differences: 1) The penalty is analytic rather than stochastic;
2) A hyper parameter α allows to control the tradeoff
between reconstruction and robustness. Actually, in an
optimizing searching algorithm, it seems more likely that
CAEs try to find invariant features by restricting step
lengths to small numbers (that is, numbers close to zero)
in each search.
3 L OWER B OUND OF
E RROR OF THE I NPUT
THE
R ECONSTRUCTION
Generally, in an algorithm based on traditional AutoEncoders, the smaller the reconstruction error of the
input, the better the algorithm. Ideally, the value of
reconstruction error of the input is equal to 0. It means
that the algorithm can completely reconstruct the input.
However, in this paper, we prove that the reconstruction
error of the input has a lower bound, which can be
viewed as a criterion for the reconstruction of the input.
We also illustrate that the reconstruction of hidden representation achieves its ideal condition is the necessary
condition for the reconstruction of the input to reach
the ideal state. When the input is corrupted with noises,
we also demonstrate that the reconstruction error of the
corrupted input has a lower bound, too.
4
Lemma 1. Let H = (~ij )m×n be a real-valued random
matrix, where each elementP~ij of
H is a real-valued random
m Pn
2
variable. Let kHk2F =
the square of
i=1
j=1 ~ij beP
m
Frobenius norm on H, L(x, y) = kx − yk2 = i=1 (xi − yi )2
m
be the squared error and x, y ∈ < be real-valued random
vectors. Then we get that
Ek(x − y)T Hk2 ≤ Ekx − yk2 · EkHk2F .
(6)
Proof: See supplementary material.
Pm Pn
2
Lemma 2. Let kAk2F =
i=1
j=1 aij be the square of
Frobenius norm on real-valued
matrix A = (aij )m×n and
Pm
L(x − y) = kx − yk2 = i=1 (xi − yi )2 be the squared error.
Then we have that
k(x − y)T Ak2 ≤ kx − yk2 · kAk2F .
(7)
Proof: See supplementary material.
3.1
Lower Bound and Necessary Condition
We present the lower bound of reconstruction error of
the input and a rigorous theoretical analysis below. We
also reveal the necessary condition for the reconstruction
of the input to reach the ideal state.
Theorem 1. Let L(x, y) = kx − yk2 be the squared error. If
we use the clean input x and clean hidden representation hc
to reconstruct themselves, then as x∗c → x, we have
L(x, g(f (x))) ≥
L(hc , f (g(hc )))/kJf (x)k2F ,
(8)
where hc is the corresponding hidden representation of the
clean input x, i.e., hc = f (x), x∗c = g(hc ) = g(f (x)) is
the reconstructed input and kJf (x)k2F = 0 iff the encoder
function f (x) is a constant.
Furthermore, we can get that
X
JCAE (θ) =
L(x, g(f (x))) + αkJf (x)k2F
x∈X
≥
X
p
λ L(hc , f (g(hc ))).
(9)
hc ∈Hc
where JCAE (θ) is the objective function of the CAE, X =
{x1 , x2 , · · · , xN } is a training set.
Proof: For a clean input x ∈ X, the corresponding
clean hidden representation and reconstructed input are
hc and x∗c , respectively. Let h∗c = f (x∗c ) be the reconstructed hidden representation. Then we can approximate the encoder function f (x∗c ) by its Taylor expansion
around x with Lagrange remainder term
f (x∗c ) = f (x) + (x∗c − x)T ∇f [x + ρ(x∗c − x)],
ρ(x∗c
where ∇f [x +
− x)] is the first-order derivative of
encoder function f (·) with respect to x + ρ(x∗c − x) and
ρ ∈ (0, 1) is a constant.
Using Lemma 2, we have that
L(hc , f (g(hc ))) =
kh∗c
− hc k
2
= kf (x∗c ) − f (x)k2
= k(x∗c − x)T ∇f [x + ρ(x∗c − x)]k2
≤ kx∗c − xk2 · k∇f [x + ρ(x∗c − x)]k2F ,
where k · k2 is the squared error and kAk2F is square of
Frobenius norm on matrix A.
When x∗c → x, i.e., the reconstructed input x∗c infinitely
approaches x, then as an ideal state, we have
lim k∇f [x + ρ(x∗c − x)]k2F = kJf (x)k2F ,
x∗
c →x
(10)
and
L(hc , f (g(hc ))) ≤ kx∗c − xk2 · kJf (x)k2F
= L(x, g(f (x))) · kJf (x)k2F ,
(11)
Hence, we get
L(x, g(f (x))) ≥ L(hc , f (g(hc )))/kJf (x)k2F ,
where kJf (x)k2F = 0 if and only if the encoder function
f (x) is a constant.
√
Moreover, from (11) and basic inequality 2 ab ≤ a +
b, a, b ≥ 0, we have
p
λ2
λ L(hc , f (g(hc ))) ≤ kx∗c − xk2 + kJf (x)k2F
4
λ2
= L(x, g(f (x))) + kJf (x)k2F ,
4
Hence, let α =
λ2
4 ,
we can get that
X
JCAE (θ) =
L(x, g(f (x))) + αkJf (x)k2F
x∈X
≥
X
λ
p
L(hc , f (g(hc ))).
hc ∈Hc
Remark 1. Theorem 1 summarizes the general rule of the
reconstruction of the input for all the algorithms based on
traditional Auto-Encoders. As we can see, the reconstruction
error of the input can not be lower than a lower bound, which
gives a guiding principle for reconstructing the input.
The traditional view for reconstructing the input is
that the smaller the reconstruction error of the input,
the better the algorithm. The ideal situation is that
the value of the reconstruction error of the input is 0,
i.e., the algorithm can completely reconstruct the input.
However, Theorem 1 demonstrates that the ideal value
of the reconstruction error of the input is a lower bound,
which is greater than or equal to 0. Hence, compared
with traditional view, Theorem 1 gives a more accurate
quantitative description of the reconstruction error of the
input.
Remark 2. Theorem 1 also provides a necessary condition for
the reconstruction of the input to reach the ideal state, namely,
the reconstruction of hidden representation achieves its ideal
condition.
Because if the reconstruction of hidden representation
does not achieve its ideal condition, the reconstruction
of the input also can not reach the ideal state. Nevertheless, when the reconstruction of hidden representation achieves the ideal state, this theorem does not
guarantee the reconstruction of the input also obtains
5
the ideal state. Therefore, we develop our algorithm
to learn robust feature representation by minimizing
the reconstruction error of both the input and hidden
representation.
Remark 3. It presents the relationship between the reconstruction error of the input and the reconstruction error of
hidden representation as well as the reconstruction error of
hidden representation and the objective function of CAE.
Remark 4. This theorem is also the main evidence that minimizing reconstruction error of hidden representation is more
robust for feature learning than minimizing the Frobenius
norm of Jacobia matrix of hidden representation.
Since DDAEs use the reconstruction of hidden representation as the objective function and CAEs learn
features by minimizing the Frobenius norm of Jacobia
matrix of hidden representation, we can conclude that
DDAEs are more robust for feature representation than
CAEs. This may also be the main reason why DDAEs
always outperform CAEs.
and
EL(h, f (g(h̃))) ≤ EL(x̃, g(h̃)) · EkJf (x̃)k2F ,
Hence, we have
EL(x̃, g(h̃)) ≥ EL(h, f (g(h̃)))/EkJf (x̃)k2F .
Remark 5. Theorem 2 summarizes that even though the
input is corrupted with noises, the reconstruction error of
the corrupted input also can not be lower than a lower
bound, which is the guiding principle for reconstructing the
corrupted input. However, the lower bound of this situation
is an expectation.
Remark 6. This theorem is also the main evidence why
minimizing reconstruction error of hidden representation is
more robust for feature representation than minimizing the
Frobenius norm of Jacobia matrix of hidden representation
when confronted with corrupted input.
4
3.2
R OBUSTNESS OF H IDDEN R EPRESENTA TION R ECONSTRUCTION
Lower Bound with Corrupted Input
We now show that when the input is corrupted with
noises, the reconstruction error of the corrupted input
also has a lower bound.
Theorem 2. Let L(x, y) = kx − yk2 be the squared error. If
some noises are added to the original input x, then as x̄ → x̃,
we have
EL(x̃, g(h̃)) ≥ EL(h, f (g(h̃)))/EkJf (x̃)k2F ,
(12)
where x̃ is the corrupted input, h = f (x̃) is the hidden
representation, h̃ is the corrupted hidden representation and
x̄ = g(h̃) is the intermediate reconstructed input.
Proof: Let h∗ = f (x̄) be the reconstructed hidden
representation. Then we can approximate the encoder
function f (x̄) by its Taylor expansion around x̃ with
Lagrange remainder term
f (x̄) = f (x̃) + (x̄ − x̃)T ∇f [x̃ + ρ(x̄ − x̃)],
In this section, we theoretically prove that minimizing
the Frobenius norm of the Jacobian matrix of the hidden
representation has a deficiency and may result in a much
worse local optimum value. We also show that minimizing reconstruction error of hidden representation for
feature representation is more robust than minimizing
the Frobenius norm of Jacobia matrix of hidden representation. In the supplementary material, we present
three examples to illustrate them. In fact, the same defect
will appear on the situation of maximizing the Frobenius
norm of Jacobia matrix. However, we only demonstrate
the deficiency of minimizing the Frobenius norm of
Jacobia matrix. The similar results can be acquired on
the case of maximizing the Frobenius norm of Jacobia
matrix.
4.1
Theoretical Explanation
where ∇f [x̃ + ρ(x̄ − x̃)] is the first-order derivative of
function f (·) with respect to x̃ + ρ(x̄ − x̃) and ρ ∈ (0, 1) is
a constant. Here ∇f [x̃+ρ(x̄− x̃)] is a real-valued random
matrix and x̃ + ρ(x̄ − x̃) is a real-valued random vector.
According to Lemma 1, we get that
The main theoretical contribution of this paper is that
we show when x∗c → x,
EL(h, f (g(h̃))) = Ekh∗ − hk2
where hc is the corresponding hidden representation of
the clean input x, i.e., hc = f (x), x∗c = g(hc ) = g(f (x))
is the reconstructed input, f (·) is the encoder function,
g(·) is the decoder function, L(hc , f (g(hc ))) is the reconstruction error of hidden representation and kJf (x)k2F =
P ∂hj (x) 2
is the Frobenius norm of Jacobia matrix of
ij
∂xi
hidden representation hc with respect to input x. We give
the proof of the Inequation (13) in Section 3. Now we
theoretically show that minimizing the Frobenius norm
of Jacobia matrix of hidden representation is invalid in
some situations. Meanwhile, we also demonstrate that in
= Ekf (x̄) − f (x̃)k
2
= Ek(x̄ − x̃)T ∇f [x̃ + ρ(x̄ − x̃)]k2
≤ Ekx̄ − x̃k2 · Ek∇f [x̃ + ρ(x̄ − x̃)]k2F ,
where k · k2 is the squared error and kHk2F is the square
of Frobenius norm on random matrix H.
When the intermediate reconstructed input x̄ infinitely
approaches x̃, we get
lim k∇f [x̃ + ρ(x̄ − x̃)]k2F = kJf (x̃)k2F ,
x̄→x̃
L(hc , f (g(hc ))) = k(x∗c − x)T Jf (x)k2
≤ kx∗c − xk2 · kJf (x)k2F ,
(13)
6
these situations, reconstruction of hidden representation
is more robust than minimizing the Frobenius norm of
Jacobia matrix.
Let us consider three special optimization problems:
1) When the algorithm reaches such areas, all of the
first derivatives are equal to constants; 2) Some of the
first derivatives are equal to constants; 3) Any one of
the first derivatives is not equal to a constant, but the
Frobenius norm of Jacobia matrix is a constant. We
firstly consider the simple situation that all of the first
∂hj (x)
= cij is
derivatives are equal to constants, i.e., ∂x
i
a constant for all i = 1, 2, ..., Dx and j = 1, 2, .., Dh .
In this situation, the Frobenius norm of Jacobia matrix
of hidden representation is a constant. It means that
once the algorithm reaches these areas, minimizing the
Frobenius norm of Jacobia matrix losts its ability to
find the optimum value. The algorithm stops early or
searches in a random direction that even includes a
much worse direction, far away from the optimum value.
Therefore, minimizing this Frobenius norm is invalid for
such a situation. However, when x∗c → x, because of Inequation (13), the value of reconstruction error of hidden
representation continues to decrease. Hence, when all of
the first derivatives are equal to constants, minimizing
reconstruction error of hidden representation works and
continues to find the optimum value. It also means that
reconstruction of hidden representation is more robust
than minimizing the Frobenius norm of Jacobia matrix.
For the second situation, when some of the first derivatives are equal to constants, minimizing the Frobenius
norm of Jacobia matrix of hidden representation seems
to be working. However, we theoretically show that
minimizing the Frobenius norm of its Jacobia matrix may
encourage to obtain a much worse local optimum value.
For this case, we only show the results that one of the
first derivatives is equal to a constant, the similar results
can be obtained with the multiple constants of the first
derivatives.
Theorem 3. Let (xu )T = (xu1 , xu2 , ..., xuDx ) be the current
value and the clean input, xT = (x1 , x2 , ..., xDx ), be the optimum value. If one of the first derivatives is equal to constant,
w
w
then there exists a next value (xw )T = (xw
1 , x2 , ..., xDx ),
obtained by minimizing the Frobenius norm of Jacobia matrix
of hidden representation with respect to x, such that
kxw − xk22 kxu − xk22 ,
(14)
kJf (xw )k2F < kJf (xu )k2F .
(15)
and
Proof: See supplementary material.
Remark 7. Theorem 3 demonstrates that if one of the first
derivatives is equal to a constant, minimizing the Frobenius
norm of Jacobia matrix encourages to obtain a much worse
local optimum value. However, because of Inequation (13),
minimizing the reconstruction error of hidden representation
does work and may find the optimum value.
For the third situation, when all of the first derivatives
are varying but the Frobenius norm of Jacobia matrix is
a constant, it is similar to the first situation.
More specifiq
cally, let hj (x) = f (x1 , x2 , ..., Dx ) = x21 + x22 + ... + x2Dx
be the encoder function for all j = 1, 2, .., Dh . Then
we have that the Frobenius norm of Jacobia
matrix is
2
P
∂hj (x)
2
= Dx · Dh .
a constant, i.e., kJf (x)kF =
ij
∂xi
For such case, we also can prove that minimizing this
Frobenius norm is invalid and reconstruction of hidden
representation is more robust than minimizing the Frobenius norm of Jacobia matrix.
From the discussion above, we can conclude that when
some or all of the first derivatives are equal to constants
or all of the first derivatives are varying while the Frobenius norm of Jacobia matrix is a constant, minimizing
the reconstruction error of hidden representation for
feature representation is more robust than minimizing
the Frobenius norm of Jacobia matrix. This may be the
main reason why DDAEs always outperform CAEs in
our experiments.
5
D OUBLE D ENOISING AUTO -E NCODERS
We have shown that the necessary condition for the
reconstruction of the input to reach the ideal state is that
the reconstruction of hidden representation achieves its
ideal condition in Section 3. We also show that minimizing (maximizing) the Frobenius norm of Jacobia matrix
may get a much worse local optimum value and minimizing reconstruction error of hidden representation for
feature representation is more robust than minimizing
the Frobenius norm of Jacobia matrixas as illustrated in
Section 4. Therefore, in this paper, we consider how to
decrease the reconstruction error of hidden representation, which may get a better feature representation. We
add the idea of reconstruction of hidden representation
to the DAEs and propose a new deep learning model that
takes the advantages of corruption and reconstruction.
We anticipate that our proposed model has the capability
to learn invariant and robust feature representation.
5.1
DDAEs Architecture
As previously stated, a DDAE usually has two separate
parts: constraints on the input (Constraints Part) and
reconstruction on the hidden representation (Reconstruction Part). We use the DAE as the Constraints Part in a
DDAE. In fact, one can replace Constraints Part by any
other auto-encoder variant. For example, we can replace
the DAE with a Sparse Auto-Encoder or a CAE. It means
that DDAE is flexibility. The Reconstruction Part is done
by first corrupting the hidden representation h ∈ <Dh
into h̃ ∈ <Dh according to a conditional distribution
q(h̃|h), and then mapping the corrupted hidden representation h̃ into an intermediate reconstructed input x̄ =
g(h̃) = Sg (W0 h̃ + bx ) ∈ <Dx from which we reconstruct
the hidden representation h∗ = f (x̄) = Sf (Wx̄ + bh ).
Fig. 1 (a) illustrates a schematic representation of the
7
Supervised
fine-tuning
Classifier layer
Decoding
Corrupting
Encoding
Decoding
~
x
x
x
Encoding
En
cod
h*
ing
Corrupting
in
g
Corrupting
Pre-training
in
g
~
h
h
ing
De
co
d
cod
Encoding
En
Second-layer DDAE
pre-training
Corrupting
in
g
Representation
of samples
ing
De
co
d
x*
Well-trained
affinity graph
cod
De
co
d
En
Decoding
Corrupting
L(h,h*)
Encoding
L(x,x*) λ
First-layer DDAE
pre-training
Corrupting
Encoding
Training
samples
(b)
(a)
Fig. 1. (a) The DDAE architecture. A sample x is stochastically corrupted to x̃. The auto-encoder then maps it to
hidden representation h (via Encoding) and attempts to reconstruct x via Decoding, producing reconstruction x∗ .
Reconstruction error is measured by the loss L(x, x∗ ). Meanwhile, the hidden representation h is also stochastically
corrupted to h̃, and then, h̃ is mapped to an intermediate reconstructed input x̄ (via Decoding) and attempts
to
p
reconstruct h via Encoding, producing reconstruction h∗ . Reconstruction error is measured by the loss L(h, h∗ ).
(b) An example of two-layer DDAEs. Hidden representation of the first-layer DDAE is taken as the input of the secondlayer DDAE. A classifier layer is added on the top of the stacked two-layer DDAEs to form a multilayer perceptron
(MLP) classifier. For training a MLP classifier, the stacked two-layer DDAEs are firstly pre-trained in a greedy, layerwise manner. After that, the MLP classifier, which is initialized by the pre-trained parameters, is fine-tuned by utilizing
back-propagation.
procedure. Note that we use the reconstruction error
of hidden representation L(h, h∗ ) = L(h, f (g(h̃))) instead of the error between intermediate reconstructed
input and original input, L(x, x̄), or more complicated
expressions, such as the combination of L(x, x̄) and
L(h, h∗ ). It is because the intermediate reconstructed
input x̄ = g(h̃) = Sg (W0 h̃ + bx ) is almost equal to the
reconstruction input x∗ = g(h) = Sg (W0 h + bx ). As a
result, L(x, x̄) has the similar effect on the Constraints
Part L(x, x∗ ) = L(x, g(f (x̃))).
Fig. 1 (b) demonstrates an example of two-layer
DDAEs. Usually, a DDAE is used to stack multiple layers
to form a deep DDAEs architecture: output of a DDAE
is used as input of the next DDAE. A classifier layer is
built on the top of the stacked deep DDAEs architecture
to form a multi-layer classifier. For training a multilayer classifier, the stacked deep DDAEs architecture
are firstly pre-trained in a greedy, layer-wise manner.
Subsequently, the multi-layer classifier is initialized by
the pre-trained parameters and fine-tuned by utilizing
back-propagation.
5.2
Training DDAEs
To train a DDAE, there are two ways: one is to optimize
a combination of Constraints Part and Reconstruction
Part (DDAE-COM); the other is to optimize them separately (DDAE-SEP). For convenience, we use a linear
combination of Constraints Part and Reconstruction Part
as the objective function of DDAE-COM. Parameters
θ = {W, bx , bh } are trained to minimize the reconstruction error over a training set X = {x1 , x2 , · · · , xN }.
The objective function optimized by stochastic gradient
descent becomes:
q
X
JCOM (θ) =
E L(x, g(f (x̃))) + λ L(h, f (g(h̃))) ,
x∈X
(16)
where L(x, g(f (x̃))) is q
the reconstruction error of the
DAE (Constraints Part), L(h, f (g(h̃))) is the reconstruction error of hidden representation (Reconstruction Part),
E(δ) is the mathematical expectation of δ, x̃ ∈ <Dx is
obtained from a conditional distribution q(x̃|x), h = f (x̃)
and λ is a hyper parameter that controls the tradeoff
between Constraints Part and Reconstruction Part.
From Equation (16), we can conclude that a DDAE
can be regarded as a general expression that extends the
DAE. If the hyper parameter λ in (16) is set to be 0, a
DDAE is the same as that of the DAE. That also means
the DAE is a special case of our proposed method, i.e.,
DDAE is a generalization of the basic DAE algorithm.
8
Algorithm 1 The DDAE-COM algorithm
Input: Training data X, learning rate η, mini-batch size
m, tradeoff coefficient λ.
Initialize the parameters θ = {W, bx , bh }.
repeat
0
Select m samples X
from X.
q
P
Let J (θ) =
E L(x, g(f (x̃))) + λ L(h, f (g(h̃)))
x∈X 0
Update parameters θ by
θ ←− θ − η ∂J∂θ(θ)
until Stopping criteria are met
q
It should be pointed out that we utilize L(h, f (g(h̃)))
instead of L(h, f (g(h̃))) to calculate the reconstruction
error of hidden representation. It is because the value of
L(h, f (g(h̃))) is large at the beginning of training and we
need to normalize it.
For the way of optimizing separately, DDAE-SEP
firstly minimizes the following objective function over
a mini-batch X 0 = {x1 , x2 , · · · , xm }:
X
O1 : JSEP −1 (θ) =
E [L(x, g(f (x̃)))] ,
(17)
dealing with noises or some inessential features such
as backgrounds is more robust than DAEs. In DAEs,
corrupting the input and then reconstructing it makes
DAEs can learn robust features. Our proposed model not
only corrupts and reconstructs the input, but also does
the same thing on hidden representation. For feature representation, corrupting hidden representation and then
reconstructing it can partially reduce the negative effects
such as the noises propagated by the input or some
inessential features such as backgrounds, while DAEs
do not deal with such noises or inessential features.
Therefore, compared with DAEs, our proposed model is
more robust to deal with the noises and some inessential
features.
Algorithm 2 The DDAE-SEP algorithm
Input: Training data X, learning rate η1 and η2 , minibatch size m.
Initialize the parameters θ = {W, bx , bh }.
repeat
0
Select m samples X
P from X
Let JSEP −1 (θ) =
E [L(x, g(f (x̃)))]
x∈X 0
Update parameters θ by
−1 (θ)
θ ←− θ − η1 ∂JSEP
∂θ
Recalculate the hidden representation H 0 of the
selected m samples withhupdated θ i
P
Let JSEP −2 (θ) =
E L(h, f (g(h̃))
x∈X 0
where x is a training sample selected from a mini-batch
X 0 and L(x, g(f (x̃))) is the reconstruction error of the
DAE (Constraints Part) on the selected sample. Subsequently, DDAE-SEP updates the parameters θ optimized
by the first objective function O1 and minimizes the
second objective function:
i
X h
O2 : JSEP −2 (θ) =
E L(h, f (g(h̃))) .
(18)
h∈H 0
Update parameters θ by
−2 (θ)
θ ←− θ − η2 ∂JSEP
∂θ
until Stopping criteria are met
h∈H 0
where h is the corresponding hidden representation of
x with updated parameters θ, H 0 is the corresponding
hidden representation of X 0 and L(h, f (g(h̃))) is the
reconstruction error of hidden representation (Reconstruction Part). Once the parameters θ are updated by
objective functions O1 and O2 , DDAE-SEP will train on
the next mini-batch and repeat the same procedure until
stopping criteria are met. For more details about how to
implement DDAE-COM and DDAE-SEP, please refer to
Algorithms 1 and 2.
5.3
Properties of DDAEs
Please note that we use corrupted hidden representation h̃ instead of a clean hidden representation h to
reconstruct a clean hidden representation in Equations
(16) and (18). There are two main reasons: 1) Although
DDAEs use the manifold learning to extract robust features, we can not guarantee all the noises have been
eliminated, they may propagate to hidden representation. 2) Even if all the noises have been eliminated,
DDAEs may learn some inessential features such as
backgrounds. The two are also the reasons why corrupting and reconstructing hidden representation for
6
E XPERIMENTS
We evaluate DDAEs on twelve UCI datasets, nine image
recognition datasets and two human genome sequence
datasets and compare the performance with some competitive state-of-the-art models. Several important parameters will also be experimentally evaluated.
6.1
A Description of Dataset
The twelve UCI datasets are selected from the UCI machine learning repository to evaluate the performance of
DDAEs with other algorithms. For all the UCI datasets,
we utilize the 10-fold cross validation to evaluate the
competing algorithms and give the average error rates
with 10 runs. Note that most of the UCI datasets are
tested in the recent work, Deep Support Vector Machine
(DeepSVM) [32]. Table 1 summarizes the basic information of twelve UCI datasets.
The nine image recognition datasets consist of the
well-known MNIST digits classification problem and
other eight benchmark datasets. The MNIST digits come
from the 28×28 gray-scale images of handwritten digits.
The other eight benchmark datasets consist of five tenclass problems modified from MNIST digits and three
9
Dataset
sonar
ionosphere
ILPD
breast cancer
australian
diabetes
vehicle
vowel
german numer
cardiotocography
segment
splice
Samples
208
351
583
683
690
768
846
990
1000
2126
2310
3175
Features
60
34
10
10
14
8
18
10
24
21
19
60
Classes
2
2
2
2
2
2
4
11
2
10
7
2
two-class problems with shape classification. The five
ten-class problems are variants of MNIST digits: smaller
subset of MNIST (basic), digits with random angle rotation (rot), digits with random noise background (bgrand), digits with random image background (bg-img)
and digits with rotation and image background (bg-imgrot). The three two-class problems are shape classification
tasks: white tall and wide rectangles on black background (rect), tall and wide rectangular image overlayed
on different background images (rect-img), convex and
concave shape (convex). All these data sets are also used
in the works of Larochelle et al. [33], Rifai et al. [22] and
Vincent et al. [34] and divided into three parts: a training
set for pre-training and fine-tuning, a validation set for
the choice of hyper-parameters and a testing set for the
report result. Details on image recognition datasets are
listed in Table 2.
TABLE 2
Digit image recognition used in the experiments
Dataset
rect
rect-img
convex
MNIST
basic
rot
bg-rand
bg-img
bg-img-rot
Train
1000
10000
7000
50000
10000
10000
10000
10000
10000
Valid.
200
2000
1000
10000
2000
2000
2000
2000
2000
Test
50000
50000
50000
10000
50000
50000
50000
50000
50000
Classes
2
2
2
10
10
10
10
10
10
The two human genome sequence datasets are the
standard benchmark from fruitfly.org for predicting gene
splicing sites on human genome sequences. The first
dataset is the Acceptor locations containing 6,877 sequences with 90 features. The second data set is the
Donor locations including 6,246 sequences with 15 features. The Acceptor dataset has 70bp in the intron (ending with AG) and 20bp of the following exon. The Donor
dataset has 7bp of the exon and 8bp of the following
intron (starting with GT). All these sequences consist of
four letters (A, T, C and G). To use these two datasets,
we need firstly to transform the four letters with four
real numbers and then use these datasets to classify each
sequence.
6.2 Experimental Verification
In Section 3, we have proved that if the reconstruction of
a hidden representation does not reach its ideal situation,
the reconstruction of the input can not obtain the ideal
value. In Section 4, we also have shown that minimizing
the reconstruction error of hidden representation for
feature representation is more robust than minimizing
the Frobenius norm of Jacobia matrix of hidden representation. However, one question still needs to be solved:
how to experimentally validate these points?
6.2.1 Reconstruction of Hidden Representation vs. Reconstruction of the Input
To show the robustness of the reconstruction of hidden
representation against the reconstruction of the input, we
evaluate the classification performance by only minimizing the Reconstruction Error of Hidden Representation
(REHR) and compare the results with only minimizing
the reconstruction error of the input, i.e., stacked AEs
(SAE) [22]. Fig. 2 shows the classification error rates of
REHR on nine image recognition datasets with different
layers. For the sake of fairness, the results of SAE are
derived from the work of Vincent et al. [34]. As shown
in Fig. 2, REHR with 3 layers almost always gets the
better results on all the nine image recognition datasets.
Fig. 3 presents some example images with corresponding filters learned by the models of stacked AEs (SAEs)
[22], stacked DAEs [34], stacked CAEs [22], stacked
REHR and stacked DDAEs. This figure shows features
learned by the first layer of all the models on the rect and
bg-img-rot datasets. Compared with SAE, REHR seems to
learn ”much more” on the test data.
Test classification error (%)
TABLE 1
UCI datasets used in the experiments
60
REHR-1
REHR-2
REHR-3
SAE-3
50
40
30
20
10
0
rect
rect-img
convex
MNIST
basic
rot
bg-rand
bg-img
bg-img-rot
Dataset
Fig. 2. Classification error rates on nine benchmark classification tasks. The results are based on only minimizing
the reconstruction error of hidden representation (REHR)
with different layers. The results of SAE-3 are based on
only minimizing the reconstruction error of the input and
its result comes from Vincent et al. [34].
6.2.2 Reconstruction of Hidden Representation vs. Minimizing Frobenius Norm of Jacobia Matrix
We conduct two comparison experiments to show the robustness of hidden representation reconstruction against
10
(a)
(b)
(d)
(c)
(e)
(f)
Fig. 3. Example images with corresponding filters learned by different models on rect (top) and bg-img-rot (bottom)
datasets. (a) Example images; (b) Filters learned by SAE; (c) Filters learned by CAE; (d) Filters learned by DAE; (e)
Our results of REHR; (f) Our results of DDAE.
minimizing the Frobenius norm of Jacobia matrix: 1)
We compare the results of only using the reconstruction
of hidden representation for feature representation with
only minimizing the Frobenius norm of Jacobia matrix.
2) We also show the results of adding reconstruction of
the input on both reconstruction of hidden representation and minimizing the Frobenius norm of Jacobia
matrix. In the second situation, we only illustrate the
comparison results of DDAEs and CAEs in practice.
In the first comparison experiment, we use MNIST
as the testing dataset. As shown in Table 4, we get
the classification error rate of about 1.53% with only
minimizing the reconstruction error of hidden representation for feature representation. However, we can not
get the classification error rate with only minimizing the
Frobenius norm of Jacobia matrix of hidden representation. It is because the non-convergence problem will
appear when we only use minimizing Frobenius norm
of Jacobia matrix of hidden representation for feature
representation. For the second comparison experiment,
we just show the comparison results of DDAEs and
CAEs. As we can see from the Tables 3 and 5, DDAEs
always outperform CAEs. Therefore, we can conclude
that not only minimizing the reconstruction error of
hidden representation for feature representation is more
robust than minimizing the Frobenius norm of Jacobia
matrix of hidden representation, but also DDAEs are
more robust for feature representation than CAEs.
6.3
Parameter Evaluation
In order to illustrate the effectiveness of hidden representation reconstruction, we evaluate the influence with
the varying range of hyper parameters (the number of
hidden layers, the number of units per hidden layer, the
learning rate for unsupervised pre-training, the learning
rate for supervised fine-tuning, etc.). In fact, it is difficult
to find the optimal combination of the hyper parameters
in a deep network. Fortunately, many researchers have
proposed various rules for choosing hyper-parameters
in the deep networks [35], [36], [37]. In our experiments,
we refer to the strategies used in [33]. We initialize all
the parameters with random values, and then fix other
hyper parameters and perform a grid search over the
range of one hyper parameter by utilizing mini-batch
stochastic gradient descent. To show the influence of
hyper parameter λ in DDAE-COM that controls the
tradeoff between Constraints Part and Reconstruction
Part, we compare DDAEs and DAEs with the adjustment
of λ. For comparison, we use the bg-img-rot as the testing
dataset. We fix all other hyper parameters for both
models and present the classification error rates on the
bg-img-rot dataset as shown in Fig. 4 (a). Clearly, DDAEs
perform better than DAEs when the hyper parameter λ
is located in a proper scope.
We also contrast DDAEs to DAEs with increasing the
number of hidden layers and the number of hidden units
per layer to show the influence of double corruption. Fig.
4 (b) shows the comparative classification error rates on
bg-img-rot dataset. The results of DDAEs marked blue
in Fig. 4 (b) illustrate that as we increase the number
of hidden layers from 1 to 4, the classification error
rates gradually descend. It could be that DDAEs have
the capability to capture the underlying data-generating
distributions of both input and hidden representation
while DAEs just capture the distribution of input. Fig. 4
(b) also shows that DDAEs outperform DAEs, especially
when the number of hidden units per layer is low. It
may be much easy for DDAEs to capture the underlying
data-generating distribution of the hidden representation
when the number of hidden representation per layer is
low.
6.4
Comparisons with State-of-the-art Results
To further show the robustness of hidden representation
reconstruction for feature representation, we compare
the performance with some state-of-the-art models on
11
44.5
58
DAE−b−3
DDAE−COM−3
DDAE−COM−4
44.1
DAE−b−1
DDAE−COM−1
DAE−b−2
DDAE−COM−2
DAE−b−3
DDAE−COM−3
DAE−b−4
DDAE−COM−4
55
Test classification error (%)
Test classification error (%)
44.3
43.9
43.7
50
43.5
43.3
43.1
45
42.9
42.7
42.5
42.3
0 0.01
0.03
0.05
0.07
λ
(a)
0.09
0.11
0.13
0.15
40
300
500
1000
Number of hidden units per layer
2000
(b)
Fig. 4. Experimental results on bg-img-rot dataset. (a) Classification error rates on the varying hyper-parameter λ in
DDAE-COM. DAE-b-3 is a 3 hidden layers stacked DAEs with masking noise and its result is obtained from Vincent et
al. [34]. (b) A comparison of DDAEs and DAEs with increasing the number of hidden layers and the number of hidden
units per layer.
twelve UCI datasets, nine image recognition datasets and
two human genome sequence datasets.
6.4.1
UCI Dataset Classification
We first test the classification performance of DDAEs
on twelve UCI datasets. By utilizing DDAE-COM algorithm, we compare DDAEs with SVM model (SVM),
stacked Deep Neural Networks (DNNs) [32], stacked
DeepSVMs [32], stacked CAEs [22] and stacked DAEs
[34]. We use tied weights, sigmoid activation function
and squared error of reconstruction loss for the networks
of CAEs, DAEs and DDAEs. In DDAEs model, we
employ a 2-layer (200-150) architecture as the most frequently used structure. Because the number of features
is small for a UCI dataset, we only use the corruption
with first and second layers of DAEs and DDAEs, not
with the input data.
Table 3 presents the classification error rates of 2 hidden layers stacked DDAEs with masking noise (DDAEb-2), compared with SVM model (SVM), stacked Deep
Neural Networks (DNNs), a 4 layers stacked DeepSVMs
(DeepSVM-4), a 2 hidden layers stacked CAEs (CAE2) and a 2 hidden layers stacked DAEs with masking
noise (DAE-b-2). In general, networks with minimizing
reconstruction error of hidden representation (DDAEs)
perform better than other networks without this constraint. It is possible that DDAEs use the reconstruction
of input and hidden representation, which may learn the
underlying data-generating distributions of both input
and hidden representation.
6.4.2
Digit Image Recognition
After testing on the small UCI dataset classification
problem, we compare DDAEs against the several stateof-the-art models for unsupervised feature extraction:
SVM models with RBF kernel (SVMrbf ), stacked Deep
Belief Networks (DBNs), stacked Deep Boltzmann Machines (DBMs) [38], stacked AEs (SAEs) [22], stacked
DAEs [34], stacked CAEs [22], stacked Rectified Factor Networks (RFNs) [39] and Ladder Networks [27],
[28]. All these models also adopt tied weights, sigmoid
activation function for both encoder and decoder, and
cross-entropy reconstruction loss except DBNs, DBMs
and RFNs. DBNs and DBMs optimize the parameters
by using contrastive divergence, while RFNs use the
expectation-maximization algorithm. Stochastic gradient
descent is applied as the optimization method for all
these models.
The classification results of DDAEs with other models
on MNIST dataset are listed in Table 4. By using zeromasking corruption noises (MN) and DDAE-COM algorithm, DDAEs with 3 layers can achieve an error rate of
about 1.35%, while the traditional DAEs is about 1.57%.
When using Gaussian corruption noises (GS), the test
error of 3-layer DDAEs reduces to 1.12% with DDAECOM algorithm and 1.08% with DDAE-SEP algorithm.
With the well-known trick technique, dropout [40], the
test error of 3-layer DDAEs can further reduce to about
0.69% when training with DDAE-COM algorithm and
about 0.66% when training with DDAE-SEP algorithm.
In the experiments, a 3-layer (1000-1000-2000) and a 4layer (1000-1000-2000-1000) architecture are tested. All
the hyperparameters are selected according to the performance on the validation set. As for dropout, we use
12
TABLE 3
Classification error rates on UCI datasets with 10-fold cross validations. The best results obtained by all considered
models are marked in bold.
SVM
14.67 ±1.86
12.70±1.24
35.26 ±1.79
2.93±1.08
13.06±1.56
18.68 ±1.22
13.36 ±1.25
1.83 ±1.14
22.40±1.21
21.84 ±1.58
3.54 ±1.26
16.66 ±0.76
DNNs
13.31 ±1.42
12.14±2.87
34.26 ±1.67
0.79±1.07
12.22±1.75
13.04 ±1.78
12.85 ±1.64
1.52 ±1.46
15.88±0.99
20.25 ±1.51
3.16 ±1.34
7.57 ±1.92
DeepSVM-4
9.35±1.44
0.15±3.11
11.02±2.09
12.48 ±2.30
16.30±1.33
6.91 ±2.33
a fixed dropout rate 20% for all the input layer and the
hidden layers. A momentum, which increases from 0.5
to 0.9, is adopted to speed up learning. A fixed learning
rate of 4.0 is used and no weight decay is utilized.
TABLE 4
Comparison of different models on MNIST.
Methods
SAE [22]
DAE + MN [22]
REHR + 3 layers
SVMrbf [34]
DDAE-COM + MN + 3 layers
RFNs [39]
DAE + GS [22]
DBN [40]
CAE [22]
DDAE-COM + GS + 3 layers
DDAE-SEP + GS + 4 layers
DDAE-SEP + GS + 3 layers
DBM [40]
DBN + dropout finetuning [40]
DBM + dropout finetuning [40]
DDAE-COM + GS + 3 layers + dropout finetuning
DDAE-SEP + GS + 3 layers + dropout finetuning
Error (%)
1.78
1.57
1.53
1.40
1.35
1.27
1.18
1.18
1.14
1.12
1.09
1.08
0.96
0.92
0.79
0.69
0.66
By using DDAE-COM algorithm, we also test our new
model on the eight deep learning benchmark datasets
with 3 and 4 layers (DDAE-3 and DDAE-4). Table 5
reports the resulting classification performance for our
model (DDAE-3 and DDAE-4), together with the performance of SVMs with RBF kernel, a 3 hidden layers
stacked Deep Belief Networks (DBN-3), a 3 hidden layers
stacked DAEs with masking noise (DAE-b-3), a 2 hidden
layers stacked CAEs (CAE-2) and a stacked Rectified
Factor Networks (RFNs). As we can see from the table,
our new model works remarkably well on all datasets. It
is better than or equivalent to the state-of-the-art models
in 6 out of 8 data sets with three layers and 7 out of 8
datasets with four layers. It should be pointed out that
we give the same corrupting noise level to both input
and hidden representation in each experiment.
CAE-2
12.91±1.46
8.87±1.74
32.76±1.68
0.12±1.05
10.35±1.48
11.96±1.75
12.12±1.65
1.07±1.57
16.35±1.15
19.82±1.47
3.07±1.13
6.48 ±1.52
DAE-b-2
11.68±1.41
8.25±1.72
31.94±1.58
0.08±1.02
10.43±1.54
11.25±1.63
11.83±1.53
0.84±1.52
15.71±1.12
18.62±1.35
2.59±1.05
5.97±1.42
DDAE-b-2
11.23±1.35
7.86±1.68
31.28±1.54
0.05±1.02
10.16±1.43
10.84±1.58
11.81±1.47
0.47±1.42
15.28±1.06
18.26±1.31
2.17±1.03
5.32±1.40
6.4.3 Human Genome Sequence Classification
In order to further demonstrate the effectiveness of
DDAEs, we evaluate the performance on two human
genome sequence datasets. For these two sequence
datasets, we also use the 10-fold cross validation and
present the average error rates with 10 runs. Fig. 5 illustrates the performance of stacked SAEs (SAE-2), stacked
REHR (REHR-2), stacked CAEs (CAE-2), stacked DDAs
with masking noise (DDA-b-2) and stacked DDAEs with
masking noise (DDAE-b-2). Note that all the models use
2 hidden layers. With DDAE-b-2 model, we obtain the
classification error rates of 8.36% for Acceptor dataset
and 9.85% for Donor dataset, the best results of five
compared models in these experiments.
Test classification error (%)
Dataset
sonar
ionosphere
ILPD
breast cancer
australian
diabetes
vehicle
vowel
german numer
cardiotocography
segment
splice
14
13
12
11
SAE-2
REHR-2
CAE-2
DAE-b-2
DDAE-b-2
10
9
8
Acceptor
Donor
Dataset
Fig. 5.
dataset.
7
Experimental results on genome sequence
C ONCLUSION
AND
F UTURE W ORK
In this paper, we demonstrated that the reconstruction
error of the input has a lower bound and minimizing the
Frobenius norm of Jacobia matrix of hidden representation has a deficiency and may encourage to get a much
worse local optimum value. Based on these evidences,
a new deep neural network, DDAEs, for unsupervised
representation learning was proposed by using the idea
of learning invariant and robust features for the small
change on both input and hidden representation. The
13
TABLE 5
Comparison of stacked DDAEs with other models. Test error rate on all considered datasets is reported together with
a 95% confidence interval. The best results obtained by all these models are marked in bold. With DDAE-COM
algorithm, DDAEs appear to achieve superior or equivalent to the best other model in six out of eight data sets with 3
layers (DDAE-3) and seven out of eight data sets with 4 layers (DDAE-4).
Dataset
rect
rect-img
convex
basic
rot
bg-rand
bg-img
bg-img-rot
SVMrbf
2.15±0.13
24.04±0.37
19.13±0.34
3.03±0.15
11.11±0.28
14.58±0.31
22.61±0.37
55.18±0.44
DBN-3
2.60±0.14
22.50±0.37
18.63±0.34
3.11±0.15
10.30±0.27
6.73±0.22
16.31±0.32
47.39±0.44
DAE-b-3
1.99±0.12
21.59±0.36
19.06±0.34
2.84±0.15
9.53±0.26
10.30±0.27
16.68±0.33
43.76±0.43
idea was implemented by minimizing the reconstruction
error after injecting noises into both input and hidden
representation. It is shown that our model is very flexible
and extendible. It is also demonstrated that minimizing
the reconstruction error of hidden representation for
feature representation is more robust than minimizing
the Frobenius norm of Jacobia matrix of hidden representation. The comprehensive experiments indicate that
DDAEs can achieve the performance superior to the
existing state-of-the-art models on twelve UCI datasets
and two human genome sequence datasets. For the
nine image recognition datasets, DDAEs are better than
or equivalent to most state-of-the-art models. Since the
reconstruction of the hidden representation always helps
an auto-encoder to perform better, and competes or improves upon the representations learning, how to design
a useful constraint or any other operations on the hidden
representation for the development of a more efficient
representation learning model would be an interesting
extension in further studies.
CAE-2
1.21±0.10
21.54±0.36
2.48±0.14
9.66±0.26
10.90±0.27
15.50±0.32
45.23±0.44
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in
Neural Information Processing Systems, pp. 1097–1105, 2012.
L. Deng, M. L. Seltzer, D. Yu, A. Acero, A.-r. Mohamed, and
G. E. Hinton, “Binary coding of speech spectrograms using a deep
auto-encoder.,” in Interspeech, pp. 1692–1695, 2010.
R. Socher, J. Pennington, E. H. Huang, A. Y. Ng, and C. D.
Manning, “Semi-supervised recursive autoencoders for predicting
sentiment distributions,” in Proceedings of the Conference on Empirical Methods in Natural Language Processing, pp. 151–161, 2011.
Y. Bengio et al., “Learning deep architectures for ai,” Foundations
and Trends R in Machine Learning, vol. 2, no. 1, pp. 1–127, 2009.
Y. Bengio, A. Courville, and P. Vincent, “Representation learning:
A review and new perspectives,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 35, no. 8, pp. 1798–1828,
2013.
Y. Bengio, “Deep learning of representations: Looking forward,”
in International Conference on Statistical Language and Speech Processing, pp. 1–37, 2013.
Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature,
vol. 521, no. 7553, pp. 436–444, 2015.
Y. Bengio, P. Lamblin, D. Popovici, H. Larochelle, et al., “Greedy
layer-wise training of deep networks,” in Advances in Neural
Information Processing Systems, pp. 153–160, 2007.
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
RFNs
0.63±0.06
20.77±0.36
16.41±0.32
2.66±0.14
7.94±0.24
15.66±0.32
-
DDAE-3
0.65±0.06
20.68±0.36
16.24±0.32
2.43±0.14
9.15±0.25
10.18±0.27
14.49±0.31
43.41±0.43
DDAE-4
0.56±0.06
20.56±0.36
15.78±0.31
2.45±0.14
9.08±0.25
10.19±0.27
14.51±0.31
42.82±0.43
K. Kavukcuoglu, R. Fergus, Y. LeCun, et al., “Learning invariant
features through topographic filter maps,” in IEEE Conference on
Computer Vision and Pattern Recognition, pp. 1605–1612, 2009.
J. Xu, L. Xiang, Q. Liu, H. Gilmore, J. Wu, J. Tang, and A. Madabhushi, “Stacked sparse autoencoder (ssae) for nuclei detection on
breast cancer histopathology images,” IEEE Transactions on Medical
Imaging, vol. 35, no. 1, pp. 119–130, 2016.
P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol, “Extracting and composing robust features with denoising autoencoders,”
in International Conference on Machine Learning, pp. 1096–1103,
2008.
S. Rifai, G. Mesnil, P. Vincent, X. Muller, Y. Bengio, Y. Dauphin,
and X. Glorot, “Higher order contractive auto-encoder,” in European Conference Machine Learning and Knowledge Discovery in
Databases, pp. 645–660, 2011.
D. P. Kingma and M. Welling, “Auto-encoding variational bayes,”
arXiv preprint arXiv:1312.6114, 2013.
M. Chen, K. Q. Weinberger, F. Sha, and Y. Bengio, “Marginalized
denoising auto-encoders for nonlinear representations.,” in International Conference on Machine Learning, pp. 1476–1484, 2014.
Y. Bengio, L. Yao, G. Alain, and P. Vincent, “Generalized denoising auto-encoders as generative models,” in Advances in Neural
Information Processing Systems, pp. 899–907, 2013.
Y. Bengio, E. Laufer, G. Alain, and J. Yosinski, “Deep generative
stochastic networks trainable by backprop,” in International Conference on Machine Learning, pp. 226–234, 2014.
M. Germain, K. Gregor, I. Murray, and H. Larochelle, “Made:
Masked autoencoder for distribution estimation.,” in International
Conference on Machine Learning, pp. 881–889, 2015.
K. Jia, L. Sun, S. Gao, Z. Song, and B. E. Shi, “Laplacian autoencoders: An explicit learning of nonlinear data manifold,” Neurocomputing, vol. 160, pp. 250–260, 2015.
A. Makhzani, J. Shlens, N. Jaitly, I. Goodfellow, and B. Frey,
“Adversarial autoencoders,” arXiv preprint arXiv:1511.05644, 2015.
C. K. Sønderby, T. Raiko, L. Maaløe, S. K. Sønderby, and
O. Winther, “Ladder variational autoencoders,” in Advances in
Neural Information Processing Systems, pp. 3738–3746, 2016.
H. Kamyshanska and R. Memisevic, “The potential energy of an
autoencoder,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 37, no. 6, pp. 1261–1273, 2015.
S. Rifai, P. Vincent, X. Muller, X. Glorot, and Y. Bengio, “Contractive auto-encoders: Explicit invariance during feature extraction,”
in International Conference on Machine Learning, pp. 833–840, 2011.
Y. Liu, X. Feng, and Z. Zhou, “Multimodal video classification
with stacked contractive autoencoders,” Signal Processing, vol. 120,
pp. 761–766, 2016.
H. Schulz, K. Cho, T. Raiko, and S. Behnke, “Two-layer contractive
encodings for learning stable nonlinear features,” Neural Networks,
vol. 64, pp. 4–11, 2015.
J. Geng, H. Wang, J. Fan, and X. Ma, “Deep supervised and
contractive neural network for sar image classification,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 55, no. 4,
pp. 2442–2459, 2017.
H. Shao, H. Jiang, F. Wang, and H. Zhao, “An enhancement deep
14
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
feature fusion method for rotating machinery fault diagnosis,”
Knowledge-Based Systems, vol. 119, pp. 200–220, 2017.
A. Rasmus, M. Berglund, M. Honkala, H. Valpola, and T. Raiko,
“Semi-supervised learning with ladder networks,” in Advances in
Neural Information Processing Systems, pp. 3546–3554, 2015.
M. Pezeshki, L. Fan, P. Brakel, A. Courville, and Y. Bengio,
“Deconstructing the ladder network architecture,” in International
Conference on Machine Learning, pp. 2368–2376, 2016.
Y. Bengio, “How auto-encoders could provide credit assignment in deep networks via target propagation,” arXiv preprint
arXiv:1407.7906, 2014.
G. Alain and Y. Bengio, “What regularized auto-encoders learn
from the data-generating distribution.,” Journal of Machine Learning Research, vol. 15, no. 1, pp. 3563–3593, 2014.
P. Vincent, “A connection between score matching and denoising
autoencoders,” Neural Computation, vol. 23, no. 7, pp. 1661–1674,
2011.
Z. Qi, B. Wang, Y. Tian, and P. Zhang, “When ensemble learning
meets deep learning: a new deep support vector machine for
classification,” Knowledge-Based Systems, vol. 107, pp. 54–60, 2016.
H. Larochelle, D. Erhan, A. Courville, J. Bergstra, and Y. Bengio,
“An empirical evaluation of deep architectures on problems with
many factors of variation,” in International Conference on Machine
Learning, pp. 473–480, 2007.
P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.-A. Manzagol,
“Stacked denoising autoencoders: Learning useful representations
in a deep network with a local denoising criterion,” Journal of
Machine Learning Research, vol. 11, no. Dec, pp. 3371–3408, 2010.
J. S. Bergstra, R. Bardenet, Y. Bengio, and B. Kégl, “Algorithms for
hyper-parameter optimization,” in Advances in Neural Information
Processing Systems, pp. 2546–2554, 2011.
J. Bergstra and Y. Bengio, “Random search for hyper-parameter
optimization,” Journal of Machine Learning Research, vol. 13,
pp. 281–305, 2012.
J. Snoek, H. Larochelle, and R. P. Adams, “Practical bayesian
optimization of machine learning algorithms,” in Advances in
Neural Information Processing Systems, pp. 2951–2959, 2012.
R. Salakhutdinov and G. Hinton, “Deep boltzmann machines,” in
Artificial Intelligence and Statistics, pp. 448–455, 2009.
D.-A. Clevert, A. Mayr, T. Unterthiner, and S. Hochreiter, “Rectified factor networks,” in Advances in Neural Information Processing
Systems, pp. 1855–1863, 2015.
N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov, “Dropout: a simple way to prevent neural
networks from overfitting.,” Journal of Machine Learning Research,
vol. 15, no. 1, pp. 1929–1958, 2014.
| 1cs.CV
|
A Networked Swarm Model for UAV Deployment
in the Assessment of Forest Environments
Matthias R. Brust∗†
∗ Computer
arXiv:1607.04439v1 [cs.RO] 15 Jul 2016
Science Division
Technological Institute of Aeronautics, Brazil
Louisiana Tech University, Ruston, USA
Email: [email protected]
Abstract—Autonomous Unmanned Aerial Vehicles (UAVs) have
gained popularity due to their many potential application fields.
Alongside sophisticated sensors, UAVs can be equipped with
communication adaptors aimed for inter-UAV communication.
Inter-communication of UAVs to form a UAV swarm raises
questions on how to manage its communication structure and
mobility. In this paper, we consider therefore the problem of
establishing an efficient swarm movement model and a network
topology between a collection of UAVs, which are specifically
deployed for the scenario of high-quality forest-mapping.
The forest environment with its highly heterogeneous distribution of trees and obstacles represents an extreme challenge for a
UAV swarm. It requires the swarm to constantly avoid possible
collisions with trees, to change autonomously the trajectory,
which can lead to disconnection to the swarm, and to reconnect
to the swarm after passing the obstacle, while continue collecting
environmental data that needs to be fused and assessed efficiently.
In this paper, we propose a novel solution to the formation
flight problem for UAV swarms. The proposed method provides
an adaptive and reliable network structure, which maintains
swarm connectivity and communicability. These characteristics
are needed to achieve a detailed and accurate description of the
environment from the data acquired by the UAV swarm.
The main characteristics of our approach are high scalability
regarding the number of UAVs in the swarm and the adaptive
network topology within the swarm.
I. I NTRODUCTION
Unmanned Aerial Vehicles (UAVs) have gained popularity
in many industries due to their vast potential applications, such
as survey, search and rescue, agriculture, or forestry [1]. In
the last decade, a significant amount of research was carried
on using and processing the information supplied by a single
UAV. However, the reduced costs of commercial-grade UAV
facilitates usage of interconnect multiple UAVs in an adaptive
and autonomous acting system [2].
Such a UAV swarm is capable of accomplishing tasks which
one UAV either fulfills with difficulty, such as accurate determination of the location for an object, or fails to accomplish
altogether, such as mapping of inaccessible caves or dense
rain forest, assessment of real-time environmental processes,
or wildlife monitoring [3]. Furthermore, compared to a one
UAV, a UAV swarm is able not only to solve more tasks,
but also to reduce the time of executing various activities and
to increase the quality of collected data. Additionally, if the
Bogdan M. Strimbu†‡
† School
of Forestry
Louisiana Tech University, Ruston, USA
‡ National Forest Inventory, ICAS Bucharest, Romania
Email: [email protected]
task requires navigational autonomy within an unknown or
difficult environment, a UAV swarm offers robustness through
redundancy and self-organization, which cannot be achieved
by deploying one UAV [3].
A particular case of a demanding flying environment is
represented by the interior of a forest. Data collection using
sensors carried above the forest canopy was executed for more
than one century [4], but accurate resource assessment still
eludes foresters, as little useful information can be obtained
from sources located outside forest [5]. The limited success
in describing forest from afar is induced by the lack of
algorithms that accurately classify the information remotely
sensed acquired in elementary components (e.g., trees, shrubs,
stem).
A UAV swarm can significantly increase the productivity
and accuracy of data describing the forest as a set of three
dimensional objects, each having multiple attributes attached
(such as taper, infestation with damaging biotic agents, or
thermal radiance) [6]. An illustration of a deployment of
a UAV swarm in a forest area is shown in Fig. 1. Each
UAV maintains communication capabilities with neighboring
UAVs which are in its transmission range. The UAVs adapt
to the unknown environment, circumvent obstacles, and map
the forest with a variety of sensors. Afterwards, the swarm
leader UAV transmits the collected and aggregated data to the
base station or operator for data analysis and advanced postprocessing.
However, the advantages promised by an autonomous UAV
swarm face challenges that come with the efficient swarm
formation and communication preservation. The set of UAVs
should exhibit a swarm-like behavior that provides an adaptive
and reliable network structure while fulfilling the required
tasks for assessment of environment (e.g., tree dimensions,
amount of light, or spatial variation of humidity inside the
forest) [7], [8].
In this paper, we propose a solution for establishing swarm
behavior and a network topology between a collection of
UAVs, which are specifically deployed for the scenario of
high-quality forest-mapping. Our approach applies a leader
election algorithm to a set of autonomous micro UAVs. The
leader will have tasks such as gathering information from the
swarm collective and leading the swarm to the destination. The
leader UAV is additionally in charge to communicate to the
base station (cf. Fig. 1). The UAVs in the formation sense and
observe the events in the environment and the more powerful
leader collects and processes information from them and reacts
in case of obstacle avoidance and acts as it the cause for route
planning and maneuvers.
II. R ELATED W ORK
Although there is high interest in applications for UAVs in
environmental monitoring, such as forest-mapping, airborne
surveillance or space exploration, most approaches are focused
on the usage of one UAV. One of the reasons is that formation,
coordination and control of a UAV swarm system containing
multiple UAVs induces additional challenges, such as three
dimensional movements, which can be N P -hard.
Existing approaches for optimizing formation acquisition
and maintenance mostly focus on coverage problems [9], [10].
A variant of the coverage problem is the maximal coverage
problem, where packing of a maximum number of circles is
required. For the two dimensional case the problem has a
polynomial time solution [11].
When considering three dimensions, the question for optimal node positioning is called the sphere packing problem
[12]. A related problem in geometry is the kissing number
problem [13], which is the number of non-overlapping unit
spheres arranged such that each sphere touches another.
The properties of network topologies resulting from random deployment of nodes in a three-dimensional area are
studied by Ravelomanana [14]. Ravelomanana considers the
k-connectivity problem, which looks for the lower bound of
the transmission range r, so that every node has at least k
direct neighbors.
Schleich et al. [10] propose a decentralized and localized
approach for UAV mobility control, which optimizes the
network connectivity. The approach maintains the connectivity
via a tree-based overlay network, whereby the root is the base
station. Their empirical results show that the maintenance of
the connectivity can have a negative impact on the coverage
while the overall connectivity can improve.
The main feature of our approach is that a multi-path
communication structure is maintained while using only local
information, thus its preserving locality. Consequently, the
formation can effectively span over a significantly wider area.
Furthermore, we also introduce a leader navigator in the swarm
movement, which distinguishes our approach from classical
swarm modeling approaches.
The traditional remote sensing techniques employed in
forest resources assessment and monitoring rely on imaginary
data often operating in the visible spectrum, which fail to
provide useful information on cloudy days. Although remedies
have been found, such as combining visible and infrared light
[15] a more direct approach is the usage of UAVs, specifically
when flying under the canopy. A significant reduction in
data acquisition costs while increasing accuracy occurs when
instead of one flying entity multiple UAVs are used [16], [17],
Figure 1. A UAV swarm is deployed in a forest environment to collect and
fuse data for transmission to ground controller.
which conducted studies on land use in the Congo Basin,
where the ground was often obscured by clouds and, therefore,
developed ways to create composite pictures of sources with
visible and infrared light.
While UAV swarms approaches are popular in military,
communication, and marine applications [18], [19], only few
applications are focused on forestry, mainly in forest fire
surveillance [20]. The scenario considered throughout this
paper aims forest-mapping of healthy trees, and introduces
a set of particular challenges to the UAV swarm movement,
such as a continuous change of the communication topology
due to the high occurrence of trees in a relatively reduced area,
which differentiates it to existing approaches.
III. S WARM S YSTEM M ODEL
In this section, we define a basic swarm system model and
the formal notions used to describe it.
A. UAV Swarm System Model
The communication network of a UAV swarm S is represented by a symmetric Euclidean graph G = (V, E) constructed such that V ∈ R3 is a set of UAV nodes in a three
dimensional bounded region with side length l. The UAV
nodes are deployed according to a deployment model D. The
links in set E of the graph G fulfill the condition that for
any pair u, v ∈ V of nodes, dist (u,v) ≤ r =⇒ {u, v} ∈ E
and dist (u,v) >r =⇒ {u, v} ∈
/ E, where r is the effective
transmission range for each v ∈ V . Each UAV v ∈ V can
relocate to any position in V ∈ R3 . An exemplary swarm
network topology, which results from our swarm system model
is shown in Fig. 2.
Our work addresses the connections only and does not
measure energy efficiency from a signal-point of view, but
from a topological point of view; wherefore, we do not
describe a radio propagation model here.
For each UAV node v ∈ V let there be a neighboring list
N eigh(v) ⊂ V , which is the set of UAVs directly connected
to UAV v, such that ∀u ∈ N eigh(v), d(v, u) ≤ r. The
neighboring list N eigh(v) is created initially and is updated
with frequency f , since the neighborhood N eigh(v) of a node
v is subject to change.We further assume that each UAV node
v ∈ V can communicate exclusively with its current direct
neighbors N eigh(v) (1-hop neighbors).
Since only localized communication is used, there is no
need to establish multi-hop control communication. This
specification guarantees complexity efficiency in our localitypreservation approach.
The swarm leader uses different ways to deliver data,
which can be the submission of data if satellite or cellular
communication is available at the base station. For this, the
leader can stop periodically the swarm activity to fly above the
tree crowns for data submission. The data can also be delivered
after returning to the spatial proximity to the base station.
However, in this paper we exclusively deal with the swarm
behavior and networking aspects within the swarm. Swarmto-base station communication is subject of future work.
One crucial objective for the deployment of a UAV swarm
is to obtain accurate location information of the UAVs and the
objects under investigations such as trees, the trees’ crown and
canopy.
However, the inherent density of a forrest imposes an
extreme challenge on on-site data collection: GPS is not
a viable option in terms of availability and efficiency for
UAVs. Additionally, thick forest structure may block satellite
communication from time to time as the UAVs fly below the
crowns and canopy.
Therefore, an autonomous UAV swarm needs to rely on a
relative positioning system, where only a small subset of the
UAVs relies on GPS, for planning or maneuvering.
Therefore, our model assumes that for each swarm Si only
the leader UAV Si (L) is in charge of absolute positions (e.g.
GPS data or preinstalled maps). Therefore, the swarm leaders
have a priori knowledge of their positions.
IV. A PPROACH : A N ETWORKED S WARM M ODEL
Various requirements and challenges exist to assess unpredictable environments such as a forest using a UAV swarm.
Differentiating features such as autonomous control or the
continuous motion during the flight require the swarm to have
a high resistance threshold concerning changes in positions,
and the swarm must reorganize obstacles and overcome them.
The algorithm for describing swarm movement and maintaining the network topology between the UAVs in the swarm is
localized and fully distributed. Additionally, our approach does
not require communication with a base station or the global
positioning system.
A. Leader UAV and swarm communication network
The UAV swarm formation process needs to consider that in
restricted settings, the leader UAV will be predetermined due
to higher-energy resources and/or communication capabilities.
Moreover, we aim at providing a scalable approach, in which
the leader election is triggered in the initialization phase
and re-triggered periodically during the flight. The possibility
of changing the leader UAV allows the swarm to readjust
Figure 2. A networked UAV swarm. The transmission range is illustrated
by the shaded spheres around each UAV. UAVs in mutual transmission range
are able to communicate and transfer collected data (continuous line).
to environmental conditions, and to maximize its operation
efficiency.
In the first part of Algorithm 1, we deal with the communication network in the UAV swarm. The leader UAV
and, subsequently, the communication structure in the swarm
is determined. Initially, weights are assigned to each UAV,
which change according to the conditions in Algorithm 1 to
determine its function in the swarm communication network.
The described procedure explicitly supports changes of UAVs
in the network position, therefore, providing an adaptive and
self-organizing approach.
B. Swarm movement and formation control
Our approach requires knowledge of destination only by the
leader UAV. The additional swarm members follow the leader
in a collision-free manner by adjusting their velocity according
to the conditions that describe the behavior of the swarm.
To achieve this leader-follower model, we significantly
altered the leaderless Boid model [21], and build a distributed
formation-featured model that fulfills the swarm requirements
for our application scenario. We propose the following extended conditions for the UAV swarm with a leader:
1) The leader UAV receives its current position and destination.
2) Each UAV (including the leader UAV) aligns with its
direct neighbors.
3) Each UAV (including the leader UAV) maintains direct
connectivity with its neighbors.
4) Each UAV (including the leader UAV) avoids collision
with any UAV.
5) The leader UAV approaches the destination.
These five conditions represent a complete description of the
UAV movement. The first condition ensures that coordinates of
the current position and the destination are known by the designated leader UAV. The leader respects all conditions, while
Algorithm 1: A networked swarm model
Input : A set of UAVs U with positions Pos and
velocities Vel, weightLimit, velleader ,
leaderID, d1 , d2
Output: A set of UAVs U with adjusted positions and
velocities
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
foreach UAV u ∈ U do
uw ← Weight(u);
N (u) ← FindDirectNeighbors(u);
/* Communication network
*/
R1 ← Max({∀u ∈ N (u) : Weight(u)});
R2 ← Min({∀u ∈ N (u) : Weight(u)});
R3 ← n ∈ N (u) if Weight(n) = leaderID;
if R2 < uw then
uw ← R2 + 1;
if R1 = weightLimit then
uw ← leaderID;
if R2 ≥ uw ∧ uw 6= leaderID then
uw ← R2 + 1;
if uw = leaderID ∧ R3 ∈
/ ∅ then
uw ← uw + 1;
end
end
end
end
/* Swarm behavior
*/
newVel() = {};
P
centreOf M ass ← ni ∈N (u) Pos(ni );
newVel(1) ← (centreOf M ass−Position(u)) ∗ 0.2;
if There is a neighbor n ∈ N (u) with
|Pos(n) − Pos(u)| < 100 then
newVel(2) ← newVel(2) - (Pos(n) - Pos(u));
end
P
newVel(3) ← n∈N (u) Vel(n) |N 1|−1 ;
newVel(3) ← (newVel(3)−Vel(u)) ∗ 0.2;
if Weight(u) = leaderID then
dif f ← d2 - Pos(u);
if dif f < 5 then
newVel(4) ← {0, 0, 0};
else
newVel(4) ← dif f /Norm[dif f ] ∗ velleader ;
end
end
P
Vel(u) ← Vel(u) + v∈newV el v;
Pos(u) ← Move(Pos(u),Vel(u));
end
moving towards destination. The second condition describes
the behavior of a UAV in relationship with its neighbors
and is responsible for alignment maintenance of the local
group movement behavior. This condition is important to
spatially organize the swarm and, in particular, to offer an
enhanced coverage of the object under investigation (e.g. tree).
Conditions three and four describe two opposing behaviors; the
former reinforces UAV movements that preserve the swarm
(i.e. relationships to its momentarily direct neighbors), while
the latter ensures that swarm’s members are at a collide
avoidance distance (i.e. preset minimum distance). The fifth
condition serves primarily to direct the swarm to its destination. The procedure implementing the five conditions is
formally described in the second part of Algorithm 1.
Local rules or conditions limit the increase in complexity
induced by scenarios with large numbers of UAVs. Therefore,
the proposed approach is scalable, as it can be applied to
swarms with either few or multiple UAVs.
V. S IMULATIONS AND R ESULTS
Settings: We refer to the notions of the basic swarm
system model in III for the descriptions of the simulation
settings.
We define the UAV swarm S as a graph G = (V, E)
constructed such that V ∈ R3 is a set of UAV nodes in a
three dimensional bounded region with side length l = 200 m.
The number of UAVs used is |V | = 8 and the transmission
range r = 55 m. The set E is determined by the condition
described in III, which states that for any pair u, v ∈ V of
nodes dist (u,v) ≤ r =⇒ {u, v} ∈ E and dist (u,v) >r =⇒
{u, v} ∈
/ E.
The UAV nodes are positioned according to the deployment
model D with coordinates d1 = {25, 25, 25} (i.e., deployment)
and d2 = (100, 100, 100) (i.e., destination) whereby d1 , d2 ∈
R3 . The leader UAV velocity is velleader = 0.7 m
s .
The neighboring list N eigh(v) ⊂ V will be created initially
so that ∀u ∈ N eigh(v), d(v, u) ≤ r and updated with
frequency f = 10s−1 .
The UAVs start in the simulation from virtually the same
position. The leader UAV moves from the deployment coordinates d1 to the destination coordinates d2 with velocity
velleader . Algorithm 1 (re-)calculates the velocity vectors of
each UAV such that the swarm maintains connectivity, but
avoids collisions.
Simulation behavior: We implemented a UAV swarm
simulator to visually and empirically verify the usability and
performance of our proposed approach described in Algorithm
1. Simulation results for the simulation settings described in
Section V are shown in Fig. 3.
Figure 3a shows that the UAVs are positioned randomly
around the deployment coordinates d1 = {25, 25, 25}. Noteworthy, the algorithm supports random position for each UAV
in the simulation area.
Figure 3b shows that the leader election has been executed
and that the leader UAV is directly connected to each of the
(a) Initial deployment and leader election
(b) Establishing communication network and navigation
(c) Navigation and maintenance of communication
(d) Positioning of the UAV swarm
Figure 3. Simulation of a networked UAV swarm with self-organizing communication network as described in Algorithm 1. The leader UAV is dark colored
and the established communication network is shown in bold continuous lines.
UAVs in the swarm. Additionally, the leader UAV starts moving towards the destination coordinates d2 = {100, 100, 100}.
Figure 3c shows that the leader UAV pushes the swarm
towards the destination d2 . Observe that only the leader knows
the destination coordinates and the rest of the swarm follows
the leader, while maintaining the swarm behavior.
For initiation of the coverage procedure, we assume that the
swarm’s movement to the destination has been realized with
a lower transmission range, which keeps the UAVs closer and
the swarm compact. Arriving at the destination coordinates d2
the UAVs transmission is extended to its full range. Figure 3d
shows the leader UAV arriving at the tree, which represents the
destination. Subsequently, the UAV swarm forms a coverage
around the tree, while maintaining full connectivity to the
entire swarm. Each UAV is connected directly or over several
hops to the leader UAV.
Time assessment: Although, the leader UAV pushes the
UAV swarm towards the destination d2 , the swarm movement
procedure in Algorithm 1 induces a delay in the collective’s
movement to avoid disconnection but at the same time to
avoid collision between UAVs. Therefore, we are interested
in the movement delay caused by our approach compared to
a straight flight of solely the leader UAV from d1 to d2 .
For this, we conducted five runs of the simulation, each
with a different number of UAVs (n = 4, n = 8, n = 12). We
measured the time needed for the swarm to arrive at the tree
(d2 ) and compared this to the theoretical time, which is the
time needed by only one UAV to fly directly with velocity
velleader from d1 to d2 . This quantitative measure reflects
applicability and scalability of the proposed approach, since
the swarm conditions impose influence on the overall task.
Results in Table I and Fig. 4 show that the flying time
from d1 to d2 (i.e. simulation time in discrete steps) does
Parameter
Euclidean distance to destination
Velocity vector of the leader UAV
Theoretical time to destination
n=4
190
195
192
196
193
193.2
Run
1
2
3
4
5
Average
n=8
190
199
199
194
201
196.6
Value
129.9
0.7
185.57
(e.g. localization and mapping) will be addressed in future
work.
R EFERENCES
n = 12
263
283
225
234
218
244.6
Table I
T EMPORAL PERFORMANCE OF THE UAV SWARM .
not differ significantly when 4 or 8 UAVs are deployed (i.e.,
1.8% from direct flight time), which is 4.1% time increase for
n = 4 and 5.9% for n = 8. However, for the case n = 12,
we observe a considerable increase in time (i.e., 31%). For
the envisioned scenario (i.e., the forest environment where we
focus on assessing one or several trees) the number of UAVs
below 10 is realistic and practical.
300
4 UAVs
8 UAVs
12 UAVs
250
Time
200
150
100
50
0
1
2
3
4
5
Run
Figure 4.
Presentation of the UAV swarm’s temporal performance.
VI. C ONCLUSIONS
We propose an approach for the formation flight problem for
UAV swarms, which maintain connectivity and communicability. The two characteristics are important for detailed, precise,
and accurate data collection and aggregation, and play a crucial
role in assessing the forest environment. Our approach is based
on local decision-making, and, therefore, assures low message
and computation complexity, while providing scalability for
a large number of UAVs in the swarm. Results indicate that
the proposed swarm movement is almost as fast to reach the
destination location as the theoretical approach comprising of
just one UAV.
However, melding the aspects of swarm movement and
communication network is more than a matter of concatenating
two isolated aspects. Additional issues of optimal swarm
positioning, such as ad hoc data fusion and object assessment
[1] F. Kendoul, “Survey of advances in guidance, navigation, and control of
unmanned rotorcraft systems,” Journal of Field Robotics, vol. 29, no. 2,
pp. 315–378, 2012.
[2] A. Bürkle, F. Segor, and M. Kollmann, “Towards autonomous micro
uav swarms,” Journal of intelligent & robotic systems, vol. 61, no. 1-4,
pp. 339–353, 2011.
[3] G. Danoy, M. R. Brust, and P. Bouvry, “Connectivity stability in
autonomous multi-level uav swarms for wide area monitoring,” in
ACM Symposium on Development and Analysis of Intelligent Vehicular
Networks and Applications (ACM DIVANet), pp. 1–8, ACM, 2015.
[4] B. Campbell James and H. Wynne Randolph, “Introduction to remote
sensing,” 2011.
[5] R. Shrestha and R. H. Wynne, “Estimating biophysical parameters of
individual trees in an urban environment using small footprint discretereturn imaging lidar,” Remote Sensing, vol. 4, no. 2, pp. 484–508, 2012.
[6] B. Strimbu and J. Innes, “An analytical platform for cumulative impact assessment based on multiple futures,” Journal of environmental
management, vol. 92, no. 7, pp. 1740–1752, 2011.
[7] M. R. Brust, D. Turgut, C. H. Ribeiro, and M. Kaiser, “Is the clustering
coefficient a measure for fault tolerance in wireless sensor networks?,”
in IEEE International Conference on Communications (ICC), pp. 183–
187, IEEE, 2012.
[8] M. R. Brust, M. İ. Akbaş, and D. Turgut, “Multi-hop localization
system for environmental monitoring in wireless sensor and actor
networks,” Wiley Journal on Concurrency and Computation: Practice
and Experience, vol. 25, no. 5, pp. 701–717, 2011.
[9] M. R. Brust, C. H. C. Ribeiro, J. A. B. Filho, J. Mesit, and R. K.
Guha, “Effective Coverage Area and Network Density in Simulations
of Ad hoc and Sensor Networks,” International Journal of Simulation Systems, Science and Technology, vol. 10, no. 4, 2009.
[10] J. Schleich, A. Panchapakesan, G. Danoy, and P. Bouvry, “Uav fleet
area coverage with network connectivity constraint,” in Proceedings of
the 11th ACM international symposium on Mobility management and
wireless access, pp. 131–138, ACM, 2013.
[11] C.-F. Huang and Y.-C. Tseng, “The coverage problem in a wireless
sensor network,” Mobile Networks and Applications, vol. 10, no. 4,
pp. 519–528, 2005.
[12] T. C. Hales, “The sphere packing problem,” Journal of Computational
and Applied Mathematics, vol. 44, no. 1, pp. 41–76, 1992.
[13] O. R. Musin, “The kissing problem in three dimensions,” Discrete &
Computational Geometry, vol. 35, no. 3, pp. 375–384, 2006.
[14] V. Ravelomanana, “Extremal properties of three-dimensional sensor
networks with applications,” Mobile Computing, IEEE Transactions on,
vol. 3, no. 3, pp. 246–257, 2004.
[15] M. C. Hansen, D. P. Roy, E. Lindquist, B. Adusei, C. O. Justice,
and A. Altstatt, “A method for integrating modis and landsat data for
systematic monitoring of forest cover and change in the congo basin,”
Remote Sensing of Environment, vol. 112, no. 5, pp. 2495–2513, 2008.
[16] C. Frey, D. Zarzhitsky, W. M. Spears, D. F. Spears, C. Karlsson,
B. Ramos, J. C. Hamann, and E. A. Widder, “A physicomimetics control
framework for swarms of autonomous surface vehicles,” in OCEANS
2008, pp. 1–6, IEEE, 2008.
[17] J. L. Sanchez-Lopez, J. Pestana, P. de la Puente, R. Suarez-Fernandez,
and P. Campoy, “A system for the design and development of visionbased multi-robot quadrotor swarms,” in Unmanned Aircraft Systems
(ICUAS), 2014 International Conference on, pp. 640–648, IEEE, 2014.
[18] Y. Wei, M. B. Blake, and G. R. Madey, “An operation-time simulation
framework for uav swarm configuration and mission planning,” Procedia
Computer Science, vol. 18, pp. 1949–1958, 2013.
[19] S. Hauert, J.-C. Zufferey, and D. Floreano, “Evolved swarming without
positioning information: an application in aerial communication relay,”
Autonomous Robots, vol. 26, no. 1, pp. 21–32, 2009.
[20] D. W. Casbeer, D. B. Kingston, R. W. Beard, and T. W. McLain,
“Cooperative forest fire surveillance using a team of small unmanned
air vehicles,” International Journal of Systems Science, vol. 37, no. 6,
pp. 351–360, 2006.
[21] C. W. Reynolds, “Flocks, herds and schools: A distributed behavioral
model,” ACM SIGGRAPH Computer Graphics, vol. 21, no. 4, pp. 25–34,
1987.
| 3cs.SY
|
arXiv:1611.02471v1 [math.AC] 8 Nov 2016
EXISTENCE OF UNIMODULAR ELEMENTS IN A PROJECTIVE MODULE
MANOJ K. KESHARI AND MD. ALI ZINNA
A BSTRACT. Let R be an affine algebra over an algebraically closed field of characteristic 0 with
dim(R) = n. Let P be a projective A = R[T1 , · · · , Tk ]-module of rank n with determinant
L. Suppose I is an ideal of A of height n such that there are two surjections α : P →
→ I and
φ : L ⊕ An−1 →
→ I. Assume that either (a) k = 1 and n ≥ 3 or (b) k is arbitrary but n ≥ 4 is
even. Then P has a unimodular element (see 4.1, 4.3).
1. I NTRODUCTION
Let R be a commutative Noetherian ring of dimension n. A classical result of Serre [Se]
asserts that if P is a projective R-module of rank > n, then P has a unimodular element.
However, as is shown by the example of projective module corresponding to the tangent
bundle of an even dimensional real sphere, this result is best possible in general. Therefore,
it is natural to ask under what conditions P has a unimodular element when rank(P ) = n. In
[Ra 1], Raja Sridharan asked the following question.
Question 1.1. Let R be a ring of dimension n and P be a projective R-module of rank n with trivial
determinant. Suppose there is a surjection α : P →
→ I, where I ⊂ R is an ideal of height n such that I
is generated by n elements. Does P has a unimodular element?
Raja Sridharan proved that the answer to Question 1.1 is affirmative in certain cases (see
[Ra 1, Theorems, 3, 5]) and “negative” in general.
Plumstead [P] generalized Serre’s result and proved that if P is a projective R[T ]-module of
rank > n, then P has a unimodular element. Bhatwadekar and Roy [B-R 2] extended Plumstead’s result and proved that projective R[T1 , . . . , Tr ]-modules of rank > n have a unimodular
element. Mandal [Ma] proved analogue of Plumstead that projective R[T, T −1 ]-modules of
rank > n have a unimodular element. In another direction, Bhatwadekar and Roy [B-R 1]
proved that projective modules over D = R[T1 , T2 ]/(T1 T2 ) of rank > n have a unimodular
element. Later Wiemers [Wi] extended this result and proved that if D = R[T1 , . . . , Tr ]/I is a
discrete Hodge algebra over R (here I is a monomial ideal), then projective D-modules of rank
> n have a unimodular element.
In view of results mentioned above, it is natural to ask the following question. Let A be
either a polynomial ring over R or a Laurent polynomial ring over R or a discrete Hodge
Date: December 31, 2017.
2000 Mathematics Subject Classification. 13C10.
Key words and phrases. Projective modules, unimodular elements, polynomial extensions.
1
2
MANOJ K. KESHARI AND MD. ALI ZINNA
algebra over R. Let P be a projective A-module of rank n. Under what conditions P has a
unimodular element? We will mention two such results.
Bhatwadekar and Raja Sridharan [B-RS 2, Theorem 3.4] proved: Let R be a ring of dimension n
containing an infinite field. Let P be a projective R[T ]-module of rank n. Assume Pf has a unimodular
element for some monic polynomial f ∈ R[T ]. Then P has a unimodular element.
Das and Raja Sridharan [D-RS, Theorem 3.4] proved: Let R be a ring of even dimension n
containing Q. Let P be a projective R[T ]-module of rank n with trivial determinant. Suppose there is a
surjection α : P →
→ I, where I is an ideal of R[T ] of height n such that I is generated by n elements.
Assume further that P/T P has a unimodular element. Then P has a unimodular element.
Note that when n is odd, the above result is not known. Further, the requirement in the
hypothesis that P/T P has a unimodular element is indeed necessary, in view of negative
answer of Question 1.1. Motivated by Bhatwadekar-Sridharan and Das-Sridharan, we prove
the following results.
Theorem 1.2. (see 3.1) Let R be a ring of dimension n containing an infinite field. Let P be a projective
A = R[T1 , · · · , Tk ]-module of rank n. Assume Pf (Tk ) has a unimodular element for some monic
polynomial f (Tk ) ∈ A. Then P has a unimodular element.
Theorem 1.3. (see 3.3) Let R be a ring of even dimension n containing Q. Let P be a projective
A = R[T1 , · · · , Tk ]-module of rank n with determinant L. Suppose there is a surjection α : P →
→ I,
where I is an ideal of A of height n such that I is a surjective image of L⊕An−1 . Further assume that
P/(T1 , · · · , Tk )P has a unimodular element. Then P has a unimodular element.
Theorem 1.4. (see 3.4) Let R be a ring of even dimension n containing Q. Assume that height of the
Jacobson radical of R is ≥ 2. Let P be a projective R[T, T −1]-module of rank n with trivial determinant.
Suppose there is a surjection α : P →
→ I, where I is an ideal of R[T, T −1] of height n such that I is
generated by n elements. Then P has a unimodular element.
Theorem 1.5. (see 3.6) Let R be a ring of even dimension n containing Q and P be a projective
D = R[T1 , T2 ]/(T1 T2 )-module of rank n with determinant L. Suppose there is a surjection α : P →
→ I,
where I is an ideal of D of height n such that I is a surjective image of L⊕Dn−1 . Further, assume that
P/(T1 , T2 )P has a unimodular element. Then P has a unimodular element.
In view of above results, we end this section with the following question.
Question 1.6. Let R be a ring of even dimension n containing Q.
(1) Let P be a projective A = R[T, T −1]-module of rank n with trivial determinant. Suppose there
is a surjection α : P →
→ I, where I ⊂ A is an ideal of height n such that I is generated
by n elements. Assume further that P/(T − 1)P has a unimodular element. Does P has a
unimodular element?
(2) Let D = R[T1 , . . . , Tk ]/I be a discrete Hodge algebra over R and P be a projective D-module of
rank n with determinant L. Suppose there is a surjection α : P →
→ I, where I is an ideal of D
n−1
of height n such that I is a surjective image of L⊕D
. Further, assume that P/(T1 , . . . , Tk )P
has a unimodular element. Does P has a unimodular element?
EXISTENCE OF UNIMODULAR ELEMENTS IN A PROJECTIVE MODULE
3
(3) Generalize question (1) replacing A by R[X1 , . . . , Xr , Y1±1 , . . . , Ys±1 ] and P/(T − 1)P by
P/(X1 , . . . , Xr , Y1 − 1, . . . , Ys − 1)P .
2. P RELIMINARIES
Assumptions. Throughout this paper, rings are assumed to be commutative Noetherian and
projective modules are finitely generated and of constant rank. For a ring A, dim(A) and J (A)
will denote the Krull dimension of A and the Jacobson radical of A respectively. In this section
we state some results for later use.
Definition 2.1. Let R be a ring and P be a projective R-module. An element p ∈ P is called
unimodular if there is a surjective R-linear map φ : P →
→ R such that φ(p) = 1. In particular,
n
a row (a1 , · · · , an ) ∈ R is unimodular (of length n) if there exist b1 , · · · , bn in R such that
a1 b1 + · · · + an bn = 1. We write U m(P ) for the set of unimodular elements of P .
Theorem 2.2. [K-Z 1, Corollary 3.1] Let R be a ring of dimension n, A = R[X1 , · · · , Xm ] a polynomial ring over R and P be a projective A[T ]-module of rank n. Assume that P/T P and Pf both contain
a unimodular element for some monic polynomial f (T ) ∈ A[T ]. Then P has a unimodular element.
The following result is a consequence of a result of Ravi Rao [R, Corollary 2.5] and Quillen’s
local-global principle [Q, Theorem 1].
Proposition 2.3. Let R be a ring of dimension n. Suppose n! is invertible in R. Then all stably free
R[T ]-modules of rank n are extended from R.
Lemma 2.4. [D-RS, Lemma 3.3] Let R be a ring and I = (a1 , · · · , an ) be an ideal of R, where
n is even. Let u, v ∈ R be such that uv = 1 modulo I. Assume further that the unimodular row
(v, a1 , · · · , an ) is completable. Then there exists σ ∈ Mn (R) with det(σ) = u modulo I such that if
(a1 , · · · , an )σ = (b1 , · · · , bn ), then b1 , · · · , bn generate I.
Proposition 2.5. [D-RS, Proposition 4.2] Let R be a ring containing Q of dimension n with ht(J (R)) ≥
1. Then any stably free R(T )-module of rank n is free, where A(T ) is obtained from A[T ] by inverting
all monic polynomials in T .
Lemma 2.6. [Bh 2, Lemma 2.2] Let R be a ring and I ⊂ R be an ideal of height r. Let P and Q be
two projective R/I-modules of rank r and let α : P →
→ I/I 2 and β : Q →
→ I/I 2 be surjections. Let
ψ : P −→ Q be a homomorphism such that β ◦ ψ = α. Then ψ is an isomorphism.
Definition 2.7. Let R be a ring and A = R[T, T −1]. We say f (T ) ∈ R[T ] is special monic if f (T )
is a monic polynomial with f (0) = 1. Write A for the ring obtained from A by inverting all
special monic polynomials of R[T ]. Then it is easy to see that dim(A) = dim(R).
Theorem 2.8. [K, 4.4, 4.6, 4.8, 4.9] Let R be a ring containing Q of dimension n ≥ 3 and A =
R[T, T −1]. Assume that ht(J (R)) ≥ 2. Let I ⊂ A be an ideal of height n such that I/I 2 is generated
by n elements and ωI : (A/I)n →
→ I/I 2 be a local orientation of I. Let P be a projective A-module of
∼
rank n and χ : A → ∧n (P ) an isomorphism. Then following holds.
4
MANOJ K. KESHARI AND MD. ALI ZINNA
(1) Suppose that the image of (I, ωI ) is zero in E(A). Then ωI is a global orientation of I, i.e., ωI
can be lifted to a surjection from An to I.
(2) Suppose that e(P, χ) = (I, ωI ) in E(A). Then there exists a surjection α : P →
→ I such that
ωI is induced from (α, χ).
(3) P has a unimodular element if and only if e(P, χ) = 0 in E(A).
(4) The canonical map E(A) −→ E(A) is injective.
Remark 2.9. Let R be a ring containing Q of dimension n with ht(J (R)) ≥ 2. Let I ⊂ A =
R[T, T −1] be an ideal of height n ≥ 3 and let ωI : (A/I)n →
→ I/I 2 be a local orientation of I.
Let θ ∈ GLn (A/I) be such that det(θ) = f . Then ωI ◦ θ is another local orientation of I, which
we denote by f ωI . On the other hand, if ωI and ω
eI are two local orientations of I, then by (2.6),
it is easy to see that ω
eI = f ωI for some unit f ∈ A/I.
The following result is due to Bhatwadekar and Raja Sridharan [B-RS 2, Theorem 4.5] in
domain case. Same proof works in general (See [D, Proposition 5.3]).
Theorem 2.10. Let R be an affine algebra over an algebraically closed field of characteristic 0 with
dim(R) = n. Let P be a projective R[T ]-module of rank n with trivial determinant. Suppose there is a
surjection α : P →
→ I, where I is an ideal of R[T ] of height n such that I is generated by n elements.
Then P has a unimodular element.
One can obtain the following result from [Su 2] and [Ra 2, Lemma 2.4],
Proposition 2.11. Let R be an affine algebra over an algebraically closed field of characteristic 0 with
dim(R) = n ≥ 3. Let L be a projective R-module of rank one. Then the canonical map E(R, L) −→
E0 (R, L) is an isomorphism.
The following result is a consequence of local-global principle for Euler class groups [D-Z 2,
Theorem 4.17].
Proposition 2.12. Let R be a ring containing Q with dim(R) = n ≥ 3. Let P be a projective R[T ]module of rank n with determinant L[T ]. Suppose that P/T P as well as P ⊗ Rm [T ] have a unimodular
element for every maximal ideal m of R. Then P has a unimodular element.
∼
Proof. Let χ : L[T ] → ∧n (P ) be an isomorphism and α : P →
→ I be a surjection, where I ⊂ R[T ]
is an ideal of height n. Let e(P, χ) = (I, ωI ) in E(R[T ], L[T ]), where (I, ωI ) is obtained from the
pair (α, χ). As P ⊗ Rm [T ] has a unimodular element for every maximal ideal m of R, the image
Q
of e(P, χ) in m E(Rm [T ], Lm [T ]) is trivial. By [D-Z 2, Theorem 4.17] the following sequence of
groups
Y
0 −→ E(R, L) −→ E(R[T ], L[T ]) −→
E(Rm [T ], Lm [T ])
m
is exact. Therefore, there exists (J, ωJ ) ∈ E(R, L) such that (JR[T ], ωJ ⊗ R[T ]) = e(P, χ) in
E(R[T ], L[T ]). Then we have e(P/T P, χ ⊗ R[T ]/(T )) = (J, ωJ ) in E(R, L). Since P/T P has
a unimodular element, by [B-RS 1, Corollary 4.4], (J, ωJ ) = 0 in E(R, L). Consequently ,
EXISTENCE OF UNIMODULAR ELEMENTS IN A PROJECTIVE MODULE
5
e(P, χ) = 0 in E(R[T ], L[T ]). Hence, by [D-Z 2, Corollary 4.15], P has a unimodular element.
The following result is due to Swan [Sw, Lemma 1.3]. We give a proof due to Murthy.
Lemma 2.13. Let R be a ring and P be a projective R[T, T −1]-module. Let f (T ) ∈ R[T ] be monic such
that Pf (T ) is extended from R. Then P is extended from R[T −1 ].
Proof. If degree of f (T ) is m, then f (T )T −m = 1 + T −1 g1 (T −1 ) := g(T −1 ). Consider the
following Cartesian diagram.
R[T −1 ]
/ R[T −1 ]T −1 (= R[T, T −1 ])
/ R[T, T −1]g(T −1 ) = R[T, T −1]f (T )
R[T
−1
]g(T −1 )
Since Pf (T ) is extended from R, Pg(T −1 ) is extended from R[T −1 ]g(T −1 ) . Using a standard
patching argument, there exists a projective R[T −1 ]-module Q such that Q ⊗ R[T −1 ] R[T, T −1] ≃
P . This completes the proof.
3. M AIN RESULTS
In this section, we shall prove our main results stated in the introduction.
Theorem 3.1. Let R be a ring of dimension n containing an infinite field. Let P be a projective
R[T1 , · · · , Tk ]-module of rank n. Assume Pf (Tk ) has a unimodular element for some monic polynomial
f (Tk ) in the variable Tk . Then P has a unimodular element.
Proof. When k = 1, we are done by [B-RS 2, Theorem 3.4]. Assume k ≥ 2 and use induction
on k. Since (P/T1 P )f has a unimodular element, by induction on k, P/T1 P has a unimodular
element. Also (P ⊗ R(T1 )[T2 , . . . , Tk ])f has a unimodular element and dim (R(T1 )) = n. So
again by induction on k, P ⊗ R(T1 )[T2 , . . . , Tk ] has a unimodular element. Since P is finitely
generated, there exists a monic polynomial g ∈ R[T1 ] such that Pg has a unimodular element.
Applying (2.2), we get that P has a unimodular element.
The next result is due to Das-Sridharan [D-RS, Theorem 3.4] when L = R.
Proposition 3.2. Let R be a ring of even dimension n containing Q. Let P be a projective R[T ]-module
of rank n with determinant L. Suppose there are surjections α : P →
→ I and φ : L⊕R[T ]n−1 →
→ I,
where I is an ideal of R[T ] of height n. Assume further that P/T P has a unimodular element. Then P
has a unimodular element.
Proof. Let Rred = R/n(R), where n(R) is the nil radical of R. It is easy to derive that P has a
unimodular element if and only if P ⊗ Rred has a unimodular element. Therefore, without loss
of generality we may assume that R is reduced. We divide the proof into two steps.
Step 1: Assume that L is extended from R. Since P/T P has a unimodular element, in view of
(2.12), it is enough to prove that if m is a maximal ideal of R, then P ⊗ Rm [T ] has a unimodular
6
MANOJ K. KESHARI AND MD. ALI ZINNA
element. Note that P ⊗ Rm [T ] has trivial determinant and α ⊗ Rm [T ] : P ⊗ Rm [T ] →
→ IRm [T ]
is a surjection. Using surjection φ ⊗ Rm [T ], we get that IRm [T ] is generated by n elements.
Applying [D-RS, Theorem 3.4] to the ring Rm [T ], it follows that P ⊗ Rm [T ] has a unimodular
element. Hence we are done.
Step 2: Assume that L is not necessarily extended from R. Since R is reduced, we can find a
ring R ֒→ S ֒→ Q(R) such that
(1)
(2)
(3)
(4)
The projective S[T ]-module L ⊗ R[T ] S[T ] is extended from S,
S is a finite R-module,
the canonical map Spec(S) ֒→ Spec(R) is bijective, and
for every P ∈ Spec(S), the inclusion R/(P ∩ R) −→ S/P is birational.
Since L ⊗ R[T ] S[T ] is extended from S, by Step 1, P ⊗ R[T ] S[T ] has a unimodular element.
Applying [Bh 1, Lemma 3.2], we conclude that P has a unimodular element.
The following result extends (3.2) to polynomial ring R[T1 , · · · , Tk ].
Theorem 3.3. Let R be a ring of even dimension n containing Q. Let P be a projective A =
R[T1 , · · · , Tk ]-module of rank n with determinant L. Suppose there are surjection α : P →
→ I and
n−1
φ : L⊕A
→
→ I, where I ⊂ A is an ideal of height n. Assume further that P/(T1 , · · · , Tk )P has a
unimodular element. Then P has a unimodular element.
Proof. When k = 1, we are done by (3.2). Assume k ≥ 2 and use induction on k. We use “bar”
when we move modulo (Tk ). Since R contains Q, replacing Xk by Xk − λ for some λ ∈ Q, we
may assume that ht (I) ≥ n.
Consider the surjection α : P →
→ I induced from α. If ht(I) > n, then I = R[T1 , · · · , Tk−1 ]
and hence P contains a unimodular element. Assume ht(I) = n. Since P /(T1 , · · · , Tk−1 )P =
→ I, by
P/(T1 , · · · , Tk )P has a unimodular element and φ induces a surjection φ : L⊕An−1 →
induction hypothesis, P has a unimodular element. By similar arguments, we get that P/Tk−1 P
has a unimodular element.
Write A = R(Tk )[T1 , · · · , Tk−1 ] and P ⊗ A = Pb . We have surjections α ⊗ A : Pb →
→ IA and
n−1
b
φ ⊗ A : (L ⊗ A)⊕A
→
→ IA. If IA = A, then P has a unimodular element. Assume ht(IA) =
n. Since Pb/Tk−1 Pb = P/Tk−1 P ⊗ A, we get that Pb /Tk−1 Pb and hence Pb /(T1 , · · · , Tk−1 )Pb has a
unimodular element. By induction on k, Pb has a unimodular element. So, there exists a monic
polynomial f ∈ R[Tk ] such that Pf has a unimodular element. Applying (2.2), we get that P
has a unimodular element.
Now we prove (1.4) mentioned in the introduction.
Theorem 3.4. Let R be a ring containing Q of even dimension n with ht(J (R)) ≥ 2. Let P be
a projective R[T, T −1]-module of rank n with trivial determinant. Assume there exists a surjection
α:P →
→ I, where I ⊂ R[T, T −1] is an ideal of height n such that I is generated by n elements. Then
P has a unimodular element.
EXISTENCE OF UNIMODULAR ELEMENTS IN A PROJECTIVE MODULE
7
∼
Proof. Let χ : R[T, T −1] → ∧n (P ) be an isomorphism. Let e(P, χ) = (I, ωI ) ∈ E(R[T, T −1]) be
obtained from the pair (α, χ). By (2.8 (3)), it is enough to prove that e(P, χ) = (I, wI ) = 0 in
E(R[T, T −1 ]).
Suppose ωI is given by I = (g1 , · · · , gn ) + I 2 . Also I is generated by n elements, say, I =
(f1 , · · · , fn ). By (2.9), there exists τ ∈ GLn (R[T, T −1 ]/I) such that (f 1 , · · · , f n ) = (g 1 , · · · , g n )τ ,
where “bar” denotes reduction modulo I. Let U, V ∈ R[T, T −1] be such that det(τ ) = U and
U V = 1 modulo I.
Claim: The unimodular row (U, f1 , · · · , fn ) over R[T, T −1]) is completable.
First we show that the theorem follows from the claim. Since n is even, by (2.4), there
exists η ∈ Mn (R[T, T −1 ]) with det(η) = V such that if (f1 , · · · , fn )η = (h1 , · · · , hn ), then
I = (h1 , · · · , hn ). Further, (g1 , · · · , g n )τ ◦ η = (h1 , · · · , hn ). Note τ ◦ η ∈ SLn (R[T, T −1]/I).
Let A be the ring obtained from R[T, T −1] by inverting all special monic polynomials. Since
dim(A/IA) = 0, we have SLn (A/IA) = En (A/IA). Let Θ ∈ SLn (A) be a lift of (τ ◦η)⊗A. Then
IA = (h1 , · · · , hn )Θ−1 and (h1 , · · · , hn )Θ−1 = (g 1 , · · · , g n ). In other words, (IA, ωI ⊗ A) = 0 in
E(A). Since ht(J (R)) ≥ 2, by (2.8 (4)), the canonical map E(R[T, T −1 ]) −→ E(A) is injective.
Therefore, (I, ωI ) = 0 in E(R[T, T −1 ]). This completes the proof. So we just need to prove the
claim.
Proof of the claim: Let Q be the stably free R[T, T −1]-module associated to the unimodular
row (U, f1 , · · · , fn ). If S is the set of all monic polynomials in R[T ], then S −1 R[T, T −1] = R(T ).
Applying (2.5), we have PS is free and hence there exists a monic polynomial f ∈ R[T ] such that
Pf is free. In particular, Pf is extended from R. Hence by (2.13), there exists a projective R[T −1 ]module Q1 such that Q ≃ Q1 ⊗ R[T −1 ] R[T, T −1]. Since (Q1 ⊕R[T −1 ])T −1 ≃ Q⊕R[T, T −1] ≃
R[T, T −1]n+1 , by Quillen and Suslin [Q, Su 1], we get Q1 ⊕R[T −1 ] ≃ R[T −1 ]n+1 . By (2.3), Q1 is
extended from R, say, Q1 ≃ Q2 ⊗ R R[T −1 ]. Since ht(J (R)) ≥ 1, by Bass [B], Q2 is free. Hence
Q is free, i.e. (V, f1 , · · · , fn ) is completable. This proves the claim.
Next result is the converse of (3.4).
Theorem 3.5. Let R be a ring containing Q of even dimension n with ht(J (R) ≥ 2. Let P be a
projective R[T, T −1]-module of rank n with trivial determinant. Let I ⊂ R[T, T −1] be an ideal of
height n which is generated by n elements. Suppose that P has a unimodular element. Then there exists
a surjection α : P →
→ I.
∼
Proof. Let χ : R[T, T −1] → ∧n (P ) be an isomorphism and e(P, χ) ∈ E(R[T, T −1]) be obtained
from the pair (P, χ). Since P has a unimodular element, by (2.8 (3)), e(P, χ) = 0 in E(R[T, T −1 ]).
Let I = (f1 , · · · , fn ) and ωI be the orientation of I induced by f1 , · · · , fn . Then (I, ωI ) = 0 =
e(P, χ) in E(R[T, T −1 ]). By (2.8 (2)), there exists a surjection α : P →
→ I such that (I, ωI ) is
induced from (α, χ).
Theorem 3.6. Let R be a ring of even dimension n containing Q and D = R[X, Y ]/(XY ). Let P
be a projective D-module of rank n with determinant L. Suppose there are surjections α : P →
→ I
8
MANOJ K. KESHARI AND MD. ALI ZINNA
and φ : L ⊕ Dn−1 →
→ I, where I is an ideal of D of height n. Assume further that P/(X, Y )P has a
unimodular element. Then P has a unimodular element.
Proof. Without loss of generality, we may assume that R is reduced. We continue to denote the
images of X and Y in D by X and Y . We give proof in two steps.
Step 1: Assume L is extended from R. Let “bar” and “tilde” denote reductions modulo (Y )
and (X) respectively. We get surjections α : P →
→ I and α
e : Pe →
→ Ie induced from α. Note that
e
ht(I) ≥ n and ht(I) ≥ n. If ht(I) > n, then I = R[X] and P has a unimodular element. Assume
ht(I) = n. Since D = R[X] and P /XP = P/(X, Y )P has a unimodular element, by (3.2), P has
a unimodular element. Similarly, Pe has a unimodular element.
e ). Since det(Pe ) = L
e is extended from R, it follows
Let p1 ∈ U m(P ). Then pe1 ∈ U m(P
e ) is surjective.
from [B-L-Ra, Theorem 5.2, Remark 5.3] that the natural map U m(Pe ) ։ U m(P
Therefore there exists p2 ∈ U m(Pe ) such that p2 = pe1 . Since the following square of rings
D
/ R[X]
R[Y ]
/R
is Cartesian with vertical maps surjective, p1 ∈ U m(P ) and p2 ∈ U m(Pe ) will patch up to give
a unimodular element of P .
Step 2: Assume L is not necessarily extended from R. Since R is reduced, by [K-Z 2, Lemma
3.7], there exists a ring S such that
(1)
(2)
(3)
(4)
R ֒→ S ֒→ Q(R),
S is a finite R-module,
R ֒→ S is subintegral and
L ⊗ R S is extended from S.
Note that R ֒→ S is actually a finite subintegral extension. Hence, by [K-Z 2, Lemma 2.11],
R[X, Y ]/(XY ) ֒→ S[X, Y ]/(XY ) is also subintegral. Since L ⊗ R S is extended from S, by Step
1, P ⊗ S[X, Y ]/(XY ) has a unimodular element. Since R contains Q, by [D-Z 1, Corollary 3.14],
P has a unimodular element.
Using [B-L-Ra, Theorem 5.2, Remark 5.3] and following the proof of (3.6), we can prove the
following result.
Theorem 3.7. Let R be a ring of even dimension n containing Q and D = R[Y1 , . . . , Ym ]/Y1 (Y2 , . . . , Ym ).
Let P be a projective D-module of rank n with determinant L. Suppose there are surjections α : P →
→I
and φ : L ⊕ Dn−1 →
→ I, where I is an ideal of D of height n. Assume further that P/(Y1 , . . . , Ym )P
has a unimodular element. Then P has a unimodular element.
4. A PPLICATIONS
In this section, we give some applications of results proved earlier. When L = R, the
following result is proved in [B-RS 2, Theorem 4.5].
EXISTENCE OF UNIMODULAR ELEMENTS IN A PROJECTIVE MODULE
9
Theorem 4.1. Let R be an affine algebra over an algebraically closed field of characteristic 0 with
dim(R) = n ≥ 3. Let P be a projective R[T ]-module of rank n with determinant L. Suppose I is an
ideal of R[T ] of height n such that there are two surjections α : P →
→ I and φ : L ⊕ R[T ]n−1 →
→ I.
Then P has a unimodular element.
→ I(0). We
Proof. Let “bar” denote reduction modulo (T ). Then α induces a surjection α : P →
can assume ht(I(0)) ≥ n. If ht(I(0)) > n, then I(0) = R and P has a unimodular element.
∼
Assume ht(I(0)) = n and fix an isomorphism χ : L → ∧n (P ). Let e(P , χ) = (I(0), ωI(0) ) in
→ I(0) is a surjection, by (2.11), e(P , χ) = 0
E(R, L) be induced from (α, χ). Since φ : L⊕Rn−1 →
in E(R, L). By [B-RS 1, Corollary 4.4], P/T P has a unimodular element. Rest of the proof is
exactly as in (3.2) using (2.10).
Corollary 4.2. Let R be an affine algebra over an algebraically closed field of characteristic 0 with
dim(R) = n ≥ 3 and D = R[X, Y1 . . . , Ym ]/X(Y1 , . . . , Ym ). Let P be a projective D-module of
rank n with determinant L. Suppose I is an ideal of D of height n such that there are two surjections
α:P →
→ I and φ : L ⊕ Dn−1 →
→ I. Then P has a unimodular element.
Proof. Follow the proof of (3.6) and use (4.1).
Theorem 4.3. Let R be an affine algebra over an algebraically closed field of characteristic 0 such
that dim(R) = n ≥ 4 is even. Let P be a projective A = R[T1 , · · · , Tk ]-module of rank n with
determinant L. Suppose I is an ideal of A of height n such that there are two surjections α : P →
→I
n−1
and φ : L ⊕ A
→
→ I. Then P has a unimodular element.
Proof. By (2.11) and [B-RS 1, Corollary 4.4], we get P/(T1 , · · · , Tk )P has a unimodular element.
By (3.3), we are done.
In the following special case, we can remove the restriction on dimension in (3.3). This result
improves [B-RS 2, Theorem 4.4].
Proposition 4.4. Let R be a ring of dimension n ≥ 2 containing Q and P be a projective A =
R[T1 , · · · , Tk ]-module of rank n with determinant L. Suppose M ⊂ A is a maximal ideal of height
n such that there are two surjections α : P →
→ M and φ : L ⊕ An−1 →
→ M. Then P has a
unimodular element.
Proof. Assume k = 1. If M contains a monic polynomial f ∈ R[T1 ], then Pf has a unimodular
element. Using [B-RS 2, Theorem 3.4], P has a unimodular element. Assume M does not
contain a monic polynomial. Since T1 ∈
/ M, ideal M + (T1 ) = R[T1 ]. By surjection α, we get
P/T1 P has a unimodular element. Follow the proof of (3.2) and use [B-RS 2, Theorem 4.4] to
conclude that P has a unimodular element. We are done when k = 1.
Assume k ≥ 2 and use induction on k. For rest of the proof, follow the proof of (3.3).
10
MANOJ K. KESHARI AND MD. ALI ZINNA
R EFERENCES
[B] H. Bass, K-theory and stable algebra, Publ. Math. Inst. Hautes Études Sci. 22 (1964), 5-60.
[Bh 1] S. M. Bhatwadekar, Inversion of monic polynomials and existence of unimodular elements (II), Math. Z. 200
(1989), 233-238.
[Bh 2] S. M. Bhatwadekar, Cancellation theorems for projective modules over a two dimensional ring and its
polynomial extensions, Compositio Math. 128 (2001), 339-359.
[B-R 1] S. M. Bhatwadekar, Amit Roy, Stability theorems for overrings of polynomial rings, Invent.Math 68 (1982),
117-127.
[B-R 2] S.M. Bhatwadekar and A. Roy, Some theorems about projective modules over polynomial rings. J. Algebra 86
(1984), 150-158.
[B-RS 1] S. M. Bhatwadekar and Raja Sridharan, Euler class group of a Noetherian ring, Compositio Math. 122 (2000),
183-222.
[B-RS 2] S. M. Bhatwadekar and Raja Sridharan, On a question of Roitman, J. Ramanujan Math. Soc. 16 (2001), 45-61.
[B-L-Ra] S. M. Bhatwadekar, H. Lindel and Ravi A. Rao, The Bass-Murthy question: Serre dimension of Laurent
polynomial extensions, Invent. Math. 81 (1985), no. 1, 189-203.
[D] M. K. Das, The Euler class group of a polynomial algebra II, J. Algebra 299 (2006), 94-114.
[D-RS] M. K. Das, Raja Sridharan, The Euler class groups of polynomial rings and unimodular elements in projective
modules, J. Pure Applied Algebra 185 (2003), 73-86.
[D-Z 1] M. K. Das and Md. Ali Zinna, On invariance of the Euler class group under a subintegral base change, J.
Algebra 398 (2014), 131-155.
[D-Z 2] M. K. Das and Md. Ali Zinna, The Euler class group of a polynomial algebra with coefficients in a line bundle,
Math. Z. 276 (2014), 757-783.
[K] M. K. Keshari, Euler class group of a Laurent polynomial ring: Local case, J. Algebra 308 (2007), 666-685.
[K-Z 1] M. K. Keshari and Md. Ali Zinna, Unimodular elements in projective modules and an analogue of a result of
Mandal, J. Commutative Algebra (to appear).
[K-Z 2] M. K. Keshari and Md. Ali Zinna, Efficient generation of ideals in a discrete Hodge algebra, J. Pure Applied
Algebra (to appear).
[Ma] Satya Mandal, Basic Elements and Cancellation over Laurent Polynomial Rings, J. Algebra 79 (1982), 251-257.
[P] B. Plumstead, The conjecture of Eisenbud and Evans, Amer. J. Math 105 (1983), 1417-1433.
[Q] D. Quillen, Projective modules over polynomial rings, Invent. Math. 36 (1976), 167-171.
[R] Ravi A. Rao, The Bass-Quillen conjecture in dimension 3 but characteristic 6= 2, 3 via a question of A. Suslin, Invent.
Math. 93 (1988), 609-618.
[Ra 1] Raja Sridharan, Non-vanishing sections of algebraic vector bundles, J. Algebra 176 (1995), 947-958.
[Ra 2] Raja Sridharan, Projective modules and complete intersections, K-Theory 13 (1998), 269-278.
[Se] J.P. Serre, Sur les modules projectifs, Semin. Dubreil-Pisot 14 (1960-61).
[Su 1] A. A. Suslin, Projective modules over polynomial ring are free, Soviet Math. Dokl. 17 (1976), 1160-1164.
[Su 2] A.A. Suslin, A cancellation theorem for projective modules over affine algebras, Sov. Math. Dokl. 18 (1977), 12811284.
[Sw] R. G. Swan, Projective Modules over Laurent Polynomial Rings, Transactions of the AMS Vol. 237 (1978), 111-120.
[Wi] A. Wiemers, Some Properties of Projective Modules over Discrete Hodge Algebras, J. Algebra 150 (1992), 402-426.
D EPARTMENT
OF
M ATHEMATICS , I NDIAN I NSTITUTE OF T ECHNOLOGY B OMBAY, P OWAI , M UMBAI 400076, I N -
DIA .
E-mail address: [email protected]
D EPARTMENT
OF
M ATHEMATICS , I NDIAN I NSTITUTE OF T ECHNOLOGY B OMBAY, P OWAI , M UMBAI 400076, I N -
DIA .
E-mail address: [email protected]
| 0math.AC
|
Reasoning in a Hierarchical System with Missing Group Size Information
Subhash Kak
Abstract.
Reasoning under uncertainty is basic to any artificial intelligence system designed to emulate
human decision making. This paper analyzes the problem of judgments or preferences
subsequent to initial analysis by autonomous agents in a hierarchical system. We first examine
a method where comparisons between two alternatives are made across different groups, and
we show that this method reduces instances of preference reversal of the kind encountered in
Simpson’s paradox. It is also shown that comparing across all permutations gives better results
than exhaustive pairwise comparison. We next propose a method where the preference
weights are recomputed based on relative group size values and show that it works well. This
problem is also of interest for preventing fraud where the data that was obtained all together is
presented in groups in a manner that supports a hypothesis that is reverse of the one valid for
the entire group. We also discuss judgments where more than two choices are involved.
Keywords: hierarchical reasoning systems, group choice, probability reversal, decision theory
Introduction
In artificial intelligence systems, decisions must be made after initial processing of the raw data
that reduces its dimensions through appropriate methods of representation [1][2]. The lookahead human reasoning is accomplished under conditions of uncertainty and that needs to be
emulated by artificial agents [3], where the logic is so designed that it mitigates the effects of
the uncertainty.
Here we consider the situation where the uncertainty is specifically concerning the sizes of the
groups and the decision is to choose between alternatives. Such decision making is also a
characteristic of responses to surveys and voting systems where the respondents’ decision is
made based on prior bias or incomplete information. Probability computations based on
frequency of an outcome on a small group can change when data from more than one cohort
are aggregated, as is true of the Simpson’s paradox [4][5].
This problem is significant in human decision making and responses to surveys because of
uncertainty of knowledge and aspects related to attention [6][7]. Cognitive processing may also
lead to preference reversal [8][9]. Human judgments may be modeled to have a superpositional
component that is quite like a quantum system [10][11][12] that provides insight into human
decision making as in the disjunction effect. It should be noted that this comes with associated
limitations [13][14].
1
This paper considers agents in a hierarchical system where the higher level must choose
between alternatives where it does not possess all the data that went in the alternatives picked
by the lower level agents. It is shown how a strategy of mixing inputs from different agents and
having the lower agents work in the size discrepancies in their preferences can improve the
decision by the higher agent. A method to recompute the preferences is also presented and it is
shown that it improves the higher level agent results.
The Hierarchical Decision System
Consider the hierarchical information system of Figure 1, in which the decision by D is made on
partially processed data by agents A1 through An.
Figure 1. A hierarchical decision-system
Let the raw data arriving at the agents A1 through An be called Mi. Agent i performs Oi so that it
obtains Ui:
Oi(Mi) = Ui
and, doing so, reduces the dimensions and size of the data. The higher level decision agent, D,
operates on the Ui and obtains Jk, k=1, 2, for a binary preference system:
Jk = S(U1,U2, …, Un), k=1, 2
To provide an example, assume that there are 3 agents that receive information on
performance in trials related to two treatments, that we call Treatment 1 and Treatment 2,
respectively. Table 1 gives the data on the results:
Table 1. An experiment with three agents
Treatment 1
Treatment 2
Agent 1
Agent 2
Agent 3
Totals
0.00 (0/1)
0.20 (1/5)
0.75 (3/4)
1.00 (1/1)
0.60 (3/5)
0.75 (3/4)
0.60 (6/10)
0.50 (5/10)
2
Each agent judges Treatment 2 to be superior. The frequency differences are substantial and
equal to 0.20, 0.25, and 0.15, respectively. However, when the preferences of the three agents
are aggregated by the higher decision agent, Treatment 1 scores over Treatment 2 by 0.10
points.
The preference reversal is due to the that if
“pathological” situations so that and
𝑐𝑐1 +𝑐𝑐2 +𝑐𝑐3
𝑑𝑑1 +𝑑𝑑2 +𝑑𝑑3
𝑎𝑎1
>
>
𝑐𝑐1
,
𝑎𝑎2
𝑏𝑏1
𝑑𝑑1 𝑏𝑏2
𝑎𝑎1 +𝑎𝑎2 +𝑎𝑎3
𝑏𝑏1 +𝑏𝑏2 +𝑏𝑏3
>
𝑐𝑐2
𝑑𝑑2
and
𝑎𝑎3
𝑏𝑏3
>
𝑐𝑐3
𝑑𝑑3
there will be
. In general, there would be such
“pathological” cases irrespective of the number of such inequalities.
The numbers involved for each of the agents is kept small just for convenience and there could
be larger numbers of the kind encountered in testing for efficacy of one treatment over
another, or comparison with a placebo [15]. Typically, the cohorts in such testing do not have
identical size because many volunteers drop out before the treatment has run its course or due
to the difficulty of getting the volunteers to enlist.
For a larger example, see Table 2, where the data is in terms of six sets of groups (dealt with by
different agents). For each of the agents Treatment 2 is superior to Treatment 1 by margins that
range from 0.04 to 0.13, but when the data is aggregated by the higher level decision agent,
Treatment 1 scores over Treatment 2 by a margin of 0.06
.
Table 2. An experiment with six agents
Treatment 1
Treatment 2
Agent 1
Agent 2
Agent 3
Agent 4
Agent 5
Agent 6
Totals
0.62
(5/8)
0.73
(19/26)
0.85
(23/27)
0.89
(8/9)
0.60
(15/24)
0.73
(57/78)
0.84
(68/81)
0.85
(23/27)
0.69
(55/80)
0.73
(192/263)
0.87
(234/270)
0.93
(81/87)
0.82
(400/490)
0.76
(380/490)
The frequency of occurrence of Simpson’s paradox is known for different assumptions made
about the data. In in a random 2 × 2 × 2 table with uniform distribution, the paradox occurs
with a probability of 1/60 [16], whereas in random path models under certain conditions it is a
bit higher than 1/8 [17]. Further issues about preference reversal arise that are discussed in
[18],[19].
In general, if the decision agent D has data in categories provided as in Table 3. Let si be the
outcomes out of a total of ei for Treatment 1 and pi be the corresponding outcomes out of a
total of fi for Treatment 2, where the outcome is to be interpreted as the proportion of subjects
who improved within the cohort.
3
Table 3. Agent preferences represented directly
Treatment 1
Treatment 2
Agent 1
Agent 2
……
Agent n
s1/g1
p1/h1
s2/g2
p2/h2
……
……
sn/gn
pn/hn
It is natural to have the following decision system:
∑𝑖𝑖 𝑠𝑠𝑖𝑖 /𝑔𝑔𝑖𝑖 > ∑𝑖𝑖 𝑝𝑝𝑖𝑖 /ℎ𝑖𝑖 , then Treatment 1 > Treatment 2
When n>2, the decision may also be made by majority logic.
These ideas are also relevant to voting schemes {20][21][22], where si/gi and pi/hi represent
preference values for the two different outcomes.
Comparisons across Sets
We first consider comparison in pairs, allowing switching of the responses. We see that
between agents 1 and 2 the data becomes
0.00 (0/1)
0.20 (1/5)
0.75 (3/4)
1.00 (1/1)
which when switched yields a result of 1 -1. Likewise, for 2 and 3, it yields
0.75 (3/4)
1.00 (1/1)
0.60 (3/5)
0.75 (3/4)
This is one tie and 2 better in one case. For 1 and 3, we have
0.00 (0/1)
0.20 (1/5)
0.60 (3/5)
0.75 (3/4)
This again yields a tie.
So switching data in pairs shows that preferences for Treatment 2 and Treatment 1 are not that
different in their efficacy.
In comparison across sets, the decision agent D at the higher level uses not just the preference
values that have been provided to it for that specific group. Therefore, if the outputs of the
Agents 1, 2, and 3 are called A1, B1, C1, respectively in favor of Treatment 1, and A2, B2, C2,
4
respectively, in favor of Treatment 2, then they should be compared with each other in all
possible arrangements, of which there will be 6 in this example. In other words A1, B1, C1 should
be compared not only with A2, B2, C2 but also with
A2, C2, B2
B2, C2, A2
B2, A2, C2
C2, A2, B2
C2, B2, A2
The full table for our example will be as follows in Table 4, where the first row represents
frequency values for Treatment 1 for the three cohorts and the remaining rows represent
the frequency values for Treatment 2 in different permutations with respect to Treatment
1.
Table 4. Comparison across groups
0.36
0
0.75
0.60
0.20
1.00
0.20 1.00
0.60
1.00 0.20
0.60
1.00 0.60
0.20
0.60 1.00
0.20
0.20
0.60
1.00
We see that when the comparisons are made in all arrangements, Treatment 2 scores 12 out of
18 times (the recomputed values are shown in bold). Therefore, its favorability is not as
absolute as when shown in Table 1.
More than two alternatives
We now consider multiple alternatives. Consider 3 alternatives as below in Table 5 where the
total number of subjects is not the same in the two groups; however the number of subjects for
testing each alternative is the same.
Table 5. Example with three alternatives
Group 1
Group 2
Totals
A 1/10 (0.10)
69/90 (0.77) 70/100 (0.70)
B 10/50 (0.20) 40/50 (0.80) 50/100 (0.50)
C 22/80 (0.27) 18/20 (0.90) 40/100 (0.40)
5
Even though the best performance for data aggregated across the groups is for A, followed by B
and C (see third column of Table 5), the results are reversed within each group, so that for each
of them separately, C scores the highest, followed by B and A, in that order.
Accounting for difference in group samples
We seek a function that leaves the preferences unchanged if the sample size is equal to the
mean or is larger, and penalizes preferences obtained from samples that are smaller than the
mean, with the penalty increasing as the departure from the mean becomes larger. Many
different functions may be used; here we propose one that corresponds to Figure 1.
Multiplies for frequency
1
mean
Sample size
Figure 1. Recomputation function shape
If the preference weight s is obtained from the group of size g (where the mean of all the
groups is gm), then its value should be adjusted for the differences in group sizes in this general
form:
1,
𝑤𝑤ℎ𝑒𝑒𝑒𝑒 𝑔𝑔 ≥ 𝑔𝑔𝑚𝑚
𝑠𝑠𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 = �
𝑘𝑘𝑘𝑘𝑒𝑒 −𝑊𝑊(𝑔𝑔−𝑔𝑔𝑚𝑚 ) , 𝑤𝑤ℎ𝑒𝑒𝑒𝑒 𝑔𝑔 < 𝑔𝑔𝑚𝑚
(1)
where W(g-gm) is some appropriate function and where k is some suitably chosen constant. An
attractive choice of this function is in equation (2) for it is quadratic and it gives us the value of
1 at the edge of g = gm:
𝑊𝑊(𝑔𝑔 − 𝑔𝑔𝑚𝑚 ) =
𝛿𝛿(𝑔𝑔 − 𝑔𝑔𝑚𝑚 )2
𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉 𝑜𝑜𝑜𝑜 𝑔𝑔
But here, for convenience, we use a δ= ½, k=1, so that equation (1) transforms into
1,
𝑠𝑠𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 = � −(𝑔𝑔−𝑔𝑔𝑚𝑚 )2 /2𝑔𝑔𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑛𝑛𝑐𝑐𝑐𝑐
𝑠𝑠𝑒𝑒
,
𝑤𝑤ℎ𝑒𝑒𝑒𝑒 𝑔𝑔 ≥ 𝑔𝑔𝑚𝑚
𝑤𝑤ℎ𝑒𝑒𝑒𝑒 𝑔𝑔 < 𝑔𝑔𝑚𝑚
2
(2)
(3)
Thus we see that the recomputation weight is 𝑒𝑒 −0.5𝑣𝑣 if the g is v standard deviations away from
the mean. So if the test size is one standard deviation away, the preference value should be
6
reduced by 0.60; if it is 2 standard deviations away, it should be reduced by 0.14; for 3 standard
deviations away, by 0.011, and so on.
In general, the choice of the function may be based on the nature of the data so that the
comparison between the preferences is intuitively satisfactory. But the problem of alternative
recomputation functions shall not be considered here.
Given equation (1), we now recompute the preference frequencies of Tables 1, 2 and 4. In Table
1, the values of Agent 1 require no change since the first is 0 and the second is for a group
whose size exceeds the mean of 3. For Agent 2, the first row is fine as 4 is larger than the mean
of 2.5. The variance is 8.5-6.25 = 2.25 and so the second entry will become exp(-1.52/2 ×2.25 =
0.60. The second entry for Agent 3, likewise becomes 0.75× exp(-0.52/2×0.25) = 0.45 (since the
variance for 4 and 5 is 0.25).
Table 6. Recomputed preference frequencies using Equation (3)
Treatment 1
Treatment 2
Agent 1
Agent 2
Agent 3
0.00 (0/1)
0.20 (1/5)
0.75 (3/4)
0.36 (1/1)
0.60 (3/5)
0.45 (3/4)
The recomputed values are shown in bold. We find that for Agent 1, Treatment 2 remains
superior; Treatment 1 is much better for Agents 2 and 3. This is satisfactory analysis and quite
in agreement with our intuition.
Table 7. Recomputing of Table 2 preferences (recomputed values are bold)
Treatment 1
Treatment 2
Agent 1
Agent 2
Agent 3
Agent 4
Agent 5
Agent 6
0.38
(5/8)
0.73
(19/26)
0.85
(23/27)
0.54
(8/9)
0.36
(15/24)
0.73
(57/78)
0.84
(68/81)
0.52
(23/27)
0.42
(55/80)
0.73
(192/263)
0.87
(234/270)
0.56
(81/87)
Unlike for the original data where Treatment 2 scored over Treatment 1 for all the agents, now
Treatment 1 and Treatment 2 are tied 3-3, which is much more reasonable and not likely to be
misinterpreted.
Now let us consider the example with three alternatives of Table 5. The data of Group 1 has a
mean of 46.66, therefore the recomputation is required to be done only for the entry for A,
whereas the data for Group 2 has a mean of 53.33, so we must recompute values for B and C;
note the variances are 2177 and 8156, respectively. With this information, we can use Equation
3 to revise the figures and these are in Table 8.
7
Table 8. Recomputing Table 5 preference values
Group 1
Group 2
A 1/10 (0.098) 69/90 (0.77)
B 10/50 (0.20) 40/50 (0.68)
C 22/80 (0.27) 18/20 (0.88)
Once again, it provides preferences results that are superior to those where the raw data is
used, as done earlier.
Lastly, we take a new example of choice between 3 alternatives, where the raw figures are 2/4,
3/7 and 4/10, with corresponding preference values of 0.5, 0.43, and 0.40. The mean of the
group sizes is 7 and the variance is 6. We can now obtain the corrections to the figures for
group that is smaller than the mean, which is the first one. After correction, its value drops to
0.416 and now we find it is not the best of the three.
Conclusions
This paper analyzes the problem of arriving at probability judgments subsequent to initial
analysis by autonomous agents in a hierarchical system. From the perspective of the higher
agent, this represents the problem of decision making under a degree of uncertainty, which is
typical in big data with lower-level processing farmed out to different contractors. We propose
a method where the comparisons are made across different cohorts, and we show that this
reduces instances of preference reversal of the kind encountered in Simpson’s paradox. It is
also shown that comparing across all permutations gives better results than exhaustive pairwise
comparison. We also propose a method where the preference weights are recomputed based
on relative group size values.
The problem of preference reversal is also of interest for preventing fraud where the data that
was obtained all together is manipulated [20] and presented in groups in a manner that
supports a hypothesis that is reverse of the one valid for the entire group. It is also of interest in
economic systems where one is interested in why agents behave in a particular fashion
[21][22]. It is also of interest in voting systems where such voting is cast independently in
different groups, as would be the case in surveys.
References
1. Newell, A. (1990), Unified Theories of Cognition, Harvard University Press.
2. Newell, A. and Simon, H.A. (1972), Human Problem Solving, Prentice-Hall, Englewood Cliffs.
3. Pomerol, J.-C. (1997) Artificial intelligence and human decision making. European Journal of
Operational Research 99: 3-25
4. Kak, S. (2017) Probability reversal and the disjunction effect in reasoning systems.
arXiv:1709.04029
8
5. Pearl, J. (2014) Understanding Simpson’s paradox. The American Statistician 88: 8-13
6. Kak, S. (2008) Foreword. In G. Feuerstein, The Yoga Tradition. Hohm Press, Prescott.
7. Pribram, K.H. and King, J. (1996) Learning as Self-Organization. Lawrence Erlbaum Associates,
Mahwah.
8. Kim, B.E., Seligman, D., and Kable, J.W. (2012). Preference reversals in decision making under
risk are accompanied by changes in attention to different attributes. Front. Neurosci. 6: 109.
9. Schkade, D.A., and Johnson, E.J. (1989). Cognitive processes in preference reversals. Org.
Behav. Human Perform. 44: 203–231.
10. Kak, S. (2000) Active agents, intelligence, and quantum computing. Information Sciences 128: 117
11. Kak, S. (1996) The three languages of the brain: quantum, reorganizational, and associative. In
Learning as Self-Organization, K. Pribram and J. King (editors). Lawrence Erlbaum Associates,
Mahwah, NJ, 185-219
12. Aerts, D. (2009) Quantum structure in cognition. Journal of Mathematical Psychology 53: 314348
13. Kak, S. (1999) The initialization problem in quantum computing. Foundations of Physics 29: 267279
14. Kak, S. (2007) Quantum information and entropy. Int. Journal of Theoretical Physics 46: 860-876
15. Kaptchuk, T.J. and Miller, F.G. (2015) Placebo effect in medicine. N Engl J Med 373:8-9
16. Pavlides, M.G. & Perlman, M.D. (2009). How likely is Simpson's paradox? The American
Statistician 63 (3): 226–233
17. Kock, N. (2015). How likely is Simpson’s paradox in path models? International Journal of eCollaboration, 11(1), 1–7.
18. Kock, N., & Gaskins, L. (2016). Simpson's paradox, moderation and the emergence of quadratic
relationships in path models: An information systems illustration. International Journal of
Applied Nonlinear Science, 2(3), 200-234.
19. Holt, G. B. (2016). Potential Simpson's paradox in multicenter study of intraperitoneal
chemotherapy for ovarian cancer. Journal of Clinical Oncology, 34(9), 1016-1016.
20. Yukalov, V.I., and Sornette, D. (2014). Manipulating decision making of typical agents. IEEE
Trans. Syst. Man Cybern. Syst. 44, 1155–1168.
21. Hunt, E. (2007). The Mathematics of Behavior. Cambridge University Press.
22. Holt, C.A. (1986). Preference reversal and the independence axiom. Am. Econ. Rev. 76,
508–515.
9
| 2cs.AI
|
Elementary equivalence vs commensurability for hyperbolic
groups
Vincent Guirardel, Gilbert Levitt, Rizos Sklinos
arXiv:1701.08853v1 [math.GR] 30 Jan 2017
Abstract
We study to what extent torsion-free (Gromov)-hyperbolic groups are elementarily
equivalent to their finite index subgroups. In particular, we prove that a hyperbolic
limit group either is a free product of cyclic groups and surface groups, or admits infinitely many subgroups of finite index which are pairwise non elementarily equivalent.
1
Introduction
There are many ways in which one may try to classify finitely generated (or finitely
presented) groups. The most obvious one is up to isomorphism, or to commensurability (having isomorphic subgroups of finite index). More generally, Gromov suggested
the looser notions of quasi-isometry and measure equivalence (commensurable groups are
quasi-isometric and measure equivalent, though the converse is far from true).
First-order logic provides us with a coarser relation than the relation of isomorphism,
where one tries to classify groups up to elementary equivalence. Two groups are elementarily equivalent if and only if they satisfy the same first-order sentences. As usual, one
must restrict the class of groups under consideration. A representative example is the work
of Szmielew [Szm55] that characterised all abelian groups (possibly infinitely generated)
up to elementary equivalence.
Our concern here is to investigate the relation between elementary equivalence and
commensurability for the special class of torsion-free hyperbolic groups, a class of groups
almost orthogonal to the class of abelian groups (hyperbolic will always mean Gromovhyperbolic, also called word-hyperbolic or δ-hyperbolic).
The solution of the Tarski problem by Kharlampovich-Myasnikov [KM06] and Sela
[Sel06] says that all non-abelian free groups are elementarily equivalent. Even more, both
works characterised the finitely generated groups that are elementarily free (i.e. elementarily equivalent to a non-abelian free group). Specifically, there exist finitely generated
groups which are elementarily free but not free. The simplest examples are fundamental
groups of closed surfaces with Euler characteristic at most −2, and more generally nonabelian free products of such surface groups and cyclic groups. This particular family is
closed under taking finite index subgroups; in particular, these groups are elementarily
equivalent to their finite index subgroups.
On the other hand, if G, G0 are two torsion-free hyperbolic groups with no non-trivial
cyclic splitting (fundamental groups of closed hyperbolic 3-manifolds, for instance), the
situation is rigid: if G and G0 are elementarily equivalent (or just have the same universal
theory), then they are isomorphic (see Proposition 7.1 of [Sel09]). In particular, being
co-Hopfian [Sel97], the group G is not elementarily equivalent to any of its proper finite
index subgroups.
All finitely generated elementarily free groups are torsion-free and hyperbolic (see
[Sel06, page 713] and [Sel01]). It is therefore natural to ask whether all finitely generated
elementarily free groups are elementarily equivalent to their finite index subgroups, and
1
more generally which torsion-free hyperbolic groups are elementarily equivalent to their
finite index subgroups.
We shall see that the answer to the first question is very much negative. Since it is
conceivable that there exist hyperbolic groups with no proper finite index subgroups, we
focus on classes of hyperbolic groups which have many finite index subgroups: limit groups
and cubulable groups.
Limit groups were first introduced as finitely generated ω-residually free groups. They
may be viewed as limits of free groups in the space of marked groups [CG05]. In model
theoretic terms, they may be characterized as finitely generated groups having the same
universal theory as free groups [Rem89]. Some of them, for instance non-abelian free
products of cyclic groups and surface groups as above, are elementarily free, but most
of them are not. A group is cubulable (or cubulated) if it acts properly discontinuously
and cocompactly on a CAT(0) cubical complex. Cubulable hyperbolic groups played a
prominent role in the proof by Agol and Wise of the virtually Haken conjecture about
3-dimensional manifolds [Ago13].
Our main results are the following:
Theorem 1.1. Let G be a torsion-free hyperbolic group. If G is a limit group or is
cubulable, then either G is a free product of surface groups and cyclic groups, or it has
infinitely many normal subgroups of finite index which are all different up to elementary
equivalence.
Corollary 1.2. A finitely generated elementarily free group is elementarily equivalent to
all its finite index subgroups if and only if it is a free product of surface groups and cyclic
groups.
We also prove:
Theorem 1.3. There exists a one-ended hyperbolic limit group G which is not elementarily
free, but has an elementarily free finite index subgroup.
Among torsion-free hyperbolic groups, Sela [Sel09] singled out a special family, which
he called elementary prototypes, with the property that two elementary prototypes are
elementarily equivalent if and only if they are isomorphic. Our way of proving Theorem
1.1 is to construct non-isomorphic finite index subgroups of G which are elementary prototypes. In Theorem 1.3 we construct G as the fundamental group of a surface with a socket.
Let us now describe the contents of the paper in a more detailed way.
In order to make our main results as independent as possible of the deep theory needed
to solve Tarski’s problem and establish the results of [Sel09], we view one-ended prototypes
as groups having no non-injective preretraction (in the sense of [Per11]), and in Section 5
we give a direct argument to show that for such groups elementary equivalence is the same
as isomorphism. As the name suggests, a non-injective preretraction yields a retraction
and thus a hyperbolic floor [Per11].
A one-ended torsion-free hyperbolic group has a canonical splitting over cyclic groups,
its JSJ splitting [Sel97, Bow98, GL]. A special role is played by the quadratically hanging
(QH) vertex groups: they are isomorphic to the fundamental groups of compact surfaces,
and incident edge groups are boundary subgroups.
Using a key technical lemma proved in Section 3, we show (Proposition 4.7) that the
existence of a non-injective preretraction is equivalent to that of a surface Σ associated to
a QH vertex group of the JSJ decomposition, and a map p : π1 (Σ) → G with the following
properties: p is a conjugation on every boundary subgroup, is not an isomorphism of
π1 (Σ) onto a conjugate, and has non-abelian image. We also show that p cannot exist if
the genus of Σ is too small (Corollary 3.5).
2
The main new results are contained in the last two sections. We first consider fundamental groups of surfaces with sockets, obtained by enlarging the fundamental group
of a surface with boundary by adding roots of generators of boundary subgroups. We
determine which of these groups are limit groups, which of them are elementarily free, and
which of them are elementary prototypes (i.e. do not have non-injective preretractions). In
particular, we prove Theorem 1.3 by showing that the fundamental group of an orientable
surface of genus 2 with a socket of order 3 is a limit group which is an elementary prototype but has an elementarily free subgroup of index 3. We also study a related family of
groups, for which all sockets are identified.
In Section 7 we prove Theorem 1.1, using the fact that a one-ended torsion-free hyperbolic group G is an elementary prototype if its cyclic JSJ decomposition satisfies the
following conditions: all surfaces are orientable, the girth of the graph of groups is large
compared to the complexity of the surfaces which appear, and edge groups map injectively
to the abelianization of G. This condition about edge groups is the main reason why we
restrict to limit groups and cubulable groups.
Acknowledgements This work was mainly conducted during the 2016 Oberwolfach
workshop on Model Theory, for which the authors acknowledge support and hospitality. The first author acknowledges support from the Institut universitaire de France.
The second author is grateful to the organizers of the 2015 Workshop on Model Theory
and Groups in Istanbul for making him think about the topic studied here. The third
author was supported by the LABEX MILYON (ANR-10-LABX-0070) of Université de
Lyon, within the program ”Investissements d’Avenir” (ANR-11-IDEX-0007) operated by
the French National Research Agency (ANR).
2
Preliminaries
Graphs of groups. Graphs of groups are a generalization of free products with amalgamation and HNN-extensions. Recall that a graph of groups G consists of a graph Γ, the
assignment of a group Gv or Ge to each vertex v or non-oriented edge e, and injections
Ge → Gv whenever v is the origin of an oriented edge (see [Ser77] for precise definitions);
the image of Ge is called an incident edge group of Gv . This data yields a group G(G),
the fundamental group of G, with an action of G(G) on a simplicial tree T (the Bass-Serre
tree of G). We sometimes call the graph of groups Γ rather than G.
A graph of groups decomposition (or splitting) of a group G is an isomorphism of G
with the fundamental group of a graph of groups. The vertex and edge groups Gv and
Ge are then naturally viewed as subgroups of G. We always assume that G is finitely
generated and that the graph of groups is minimal (no proper connected subgraph of Γ
carries the whole of G); it follows that Γ is a finite graph. We allow the trivial splitting,
with Γ a single vertex v with Gv = G.
Killing all vertex groups defines an epimorphism from G to the topological fundamental
group of the graph Γ (a free group, possibly cyclic or trivial).
The (closed) star of a vertex v of Γ is the union of all edges containing v. The open
star of v is obtained by removing all vertices w 6= v from the closed star.
Grushko decompositions. A finitely generated group has a Grushko decomposition:
it is isomorphic to a free product A1 ∗ . . . ∗ An ∗ Fm , where each Ai for i ≤ n is nontrivial, freely indecomposable and not infinite cyclic, and Fm is a free group of rank m.
Moreover, the numbers n and m are unique, as well as the Ai ’s up to conjugacy (in G)
and permutation. We call the Ai ’s the Grushko factors of G. When G is torsion-free, they
are precisely the one-ended free factors of G by a classical theorem of Stallings.
3
Decompositions as free products correspond to graph of groups decompositions with
trivial edge groups. A one-ended group does not split over a trivial group, but it may split
over Z. We discuss this when G is a torsion-free hyperbolic group.
Hyperbolic groups and their cyclic JSJ decompositions. Recall that a finitely
generated group is hyperbolic if its Cayley graph (with respect to some, hence any, finite
generating set) is a hyperbolic metric space: there exists δ > 0 such that any point on one
side of any geodesic triangle is δ-close to one of the other two sides. Small cancellation
groups and fundamental groups of negatively curved closed manifolds are hyperbolic. In
particular free groups, fundamental groups of closed surfaces with negative Euler characteristic, and free products of such groups, are hyperbolic. A one-ended torsion-free
hyperbolic group G is co-Hopfian [Sel97]: it cannot be isomorphic to a proper subgroup.
A one-ended torsion-free hyperbolic group G has a canonical JSJ decomposition Γcan
over cyclic groups [Sel97, Bow98, GL]. We mention the properties that will be important
for this paper. The graph Γcan is bipartite, with every edge joining a vertex carrying a
cyclic group to a vertex carrying a non-cyclic group. The action of G on the associated
Bass-Serre tree T is invariant under automorphisms of G and acylindrical in the following
strong sense: if a non-trivial element g ∈ G fixes a segment of length ≥ 2 in T , then this
segment has length exactly 2 and its midpoint has cyclic stabilizer.
Surface-type vertices. There are two kinds of vertices of Γcan carrying a non-cyclic
group: rigid ones, and quadratically hanging (QH) ones. We will be concerned mostly
with QH vertices.
If v is a QH vertex of Γcan , the group Gv is isomorphic to the fundamental group of
a compact (possibly non-orientable) surface Σ. Incident edge groups Ge are boundary
subgroups of π1 (Σ): there exists a component C of ∂Σ such that the image of Ge is equal
to π1 (C). We do not specify base points, so we should really say that Ge is conjugate to
π1 (C) in π1 (Σ); similarly, any non null-homotopic simple closed curve in Σ determines an
infinite cyclic subgroup of π1 (Σ), well-defined up to conjugacy.
Moreover, given any component C of ∂Σ, there is a unique incident edge e such that
Ge equals π1 (C). These properties are not true for general QH vertices (see [GL]), so as in
[Per11] we recall them by calling QH vertices v of Γcan surface-type vertices (see Definition
3.1).
Surfaces. The genus g = g(Σ) of a compact surface Σ is the largest number of nonintersecting (possibly one-sided) simple closed curves (other than components of ∂Σ) that
can be drawn on the surface without
P disconnecting it. If Σ1 , . . . , Σk are disjoint compact
subsurfaces of Σ, one clearly has k g(Σi ) ≤ g(Σ). Surfaces of genus 0 are obtained by
removing open discs from a sphere, and they are called planar.
If Σ has b boundary components, its Euler characteristic is χ(Σ) = 2 − 2g − b if Σ is
orientable, χ(Σ) = 2 − g − b otherwise. Surfaces appearing in surface-type vertices will
always have non-abelian fundamental group; equivalently, χ(Σ) will be negative. Two
surfaces are homeomorphic if and only if they are both orientable or non-orientable, and
g and χ are the same.
Given a compact surface Σ, there is an upper bound for the cardinality of a family of
disjoint non-parallel simple closed curves which are not null-homotopic (a curve is nullhomotopic if it bounds a disc, two curves are parallel if they bound an annulus); for
instance, the bound is 3g − 3 if Σ is a closed orientable surface of genus g ≥ 2.
There are 5 surfaces with χ(Σ) = −1. The orientable ones are the pair of pants
(g = 0, b = 3) and the once-punctured torus (g = 1, b = 1). The non-orientable ones
are the twice punctured projective plane (g = 1, b = 2), the once-punctured Klein bottle
(g = 2, b = 1), and the closed surface of genus 3. With the exception of the punctured
4
torus, these surfaces do not carry pseudo-Anosov diffeomorphisms and are not allowed to
appear in hyperbolic towers (in the sense of [Sel01, Per11]). They should be considered as
exceptional (see Definition 4.3).
Actions of surface groups on trees. Let Σ be a compact surface. We describe a
standard construction associating a finite family C of disjoint simple closed curves on Σ
to an action of π1 (Σ) on a tree T such that boundary subgroups are elliptic (they fix a
vertex in T ).
The group π1 (Σ) acts freely on the universal covering Σ̃. It also acts on T , and we
construct an equivariant continuous map f˜ : Σ̃ → T .
First suppose that Σ is closed. Fix a triangulation of Σ and lift it to the universal
covering Σ̃. We define f˜ on the set of vertices, making sure that vertices Σ̃ are mapped to
vertices of T . We then extend it to edges in a linear way, and to triangles (2-simplices).
There is a lot of freedom in this construction, but we make sure that f˜ is in general
position with respect to midpoints of edges: the preimage C˜ of the set of midpoints of
edges of T intersects each triangle in a finite collection of disjoint arcs joining one side of
the triangle to another side. The construction is the same if Σ has a boundary, but we
require that each line in ∂ Σ̃ be mapped to a single vertex of T (this is possible because
boundary subgroups act elliptically on T ).
The subset C˜ ⊂ Σ̃ is π1 (Σ)-invariant and its projection to Σ is a finite family C of
disjoint simple closed curves.
Limit groups. A group G is residually free if, for every nontrivial element g ∈ G, there
exists a morphism h : G → F for some free group F such that h(g) is nontrivial.
A group G is ω-residually free if, for every finite set {g1 , . . . , gn } ⊂ G \ {1}, there exists
a morphism h : G → F for some free group F such that h(gi ) is nontrivial for all i ≤ n.
Remeslenikov [Rem89] proved that the class of ω-residually free groups coincides with
the class of ∀-free groups, i.e. the class of groups that have the same universal theory as a
free group. In his work on the Tarski problem, Sela viewed finitely generated ω-residually
free groups in a more geometric way and called them limit groups because they arise from
limiting processes (see [Sel01, CG05]). Limit groups have the same universal theory as a
free group, but they are not necessarily elementarily free.
Free abelian groups and free groups are limit groups. Fundamental groups of orientable
closed surfaces, and of non-orientable surfaces with g ≥ 4, are limit groups (they are even
elementarily free). A finitely generated subgroup of a limit group, and a free product of
limit groups, are limit groups. A limit group is hyperbolic if and only if it does not contain
Z2 ([Sel01]).
3
Pinched curves on surfaces
Definition 3.1 (Surface-type vertex). Let Γ be a graph of groups decomposition of a group
G. A vertex v of Γ is called a surface-type vertex if the following conditions hold:
• the group Gv carried by v is the fundamental group of a compact surface Σ (usually
with boundary), with π1 (Σ) non-abelian;
• incident edge groups are maximal boundary subgroups of π1 (Σ), and this induces a
bijection between the set of boundary components of Σ and the set of incident edges.
We say that Gv is a surface-type vertex group. If u is any lift of v to the Bass-Serre
tree T of Γ, we say that u is a surface-type vertex, and its stabilizer Gu (which is conjugate
to Gv ) is a surface-type vertex stabilizer.
5
Surface-type vertices are QH (quadratically hanging) vertices in the sense of [GL].
Definition 3.2 (Pinching). Let Σ be a compact surface. Given a homomorphism p :
π1 (Σ) → G, a family of pinched curves is a collection C of disjoint, non-parallel, two-sided
simple closed curves Ci ⊂ Σ, none of which is null-homotopic, such that the fundamental
group of each Ci is contained in ker p (the curves may be parallel to a boundary component).
The map p is non-pinching if there is no pinched curve: p is injective in restriction to
the fundamental group of any simple closed curve which is not null-homotopic.
Let S be a component of the surface Σ̂ obtained by cutting Σ along C. Its fundamental
group is naturally identified with a subgroup of π1 (Σ), so p restricts to a map from π1 (S)
to G, which we also denote by p.
Lemma 3.3. Let Γ be a graph of groups decomposition of a group G, with Bass-Serre
tree T . Assume that Γ has a single surface-type vertex v with Gv = π1 (Σ), together with
vertices v1 , . . . , vn (with n ≥ 1), and every edge of Γ joins v to some vi (in particular, edge
groups of Γ are infinite cyclic).
Let p : π1 (Σ) → G be a homomorphism such that the image of every boundary subgroup
of π1 (Σ) fixes an edge of T , and p is not an isomorphism onto some subgroup of G conjugate
to π1 (Σ).
Let C be a maximal family of pinched curves on Σ, and let S be a component of the
surface obtained by cutting Σ along C.
Then the image of π1 (S) by p is contained in a conjugate of some Gvi (i.e. it fixes a
vertex of T which is not a lift of v).
Proof. The boundary of S consists of components of ∂Σ and curves coming from C. Define
S0 by gluing a disc to S along each curve coming from C. There is an induced map
π : π1 (S0 ) → G, hence an action of π1 (S0 ) on the Bass-Serre tree T , and we must show
that π1 (S0 ) fixes a lift of some vi . We may assume that the image of π is non-trivial.
As in the preliminaries, we consider the universal covering S̃0 , and π1 (S0 )-equivariant
maps f˜ : S̃0 → T such that every line in ∂ S̃0 is mapped to a vertex (such maps exist
because boundary subgroups of π1 (Σ) fix a point in T ). Assuming that f˜ is in general
position, we consider preimages of midpoints of edges and we project to S0 . This yields
a finite family C0 of simple closed curves on S0 , and we choose f˜ so as to minimize the
number of these curves (in particular, no curve in C0 is null-homotopic).
The map f˜ induces a map f : S0 → Γ sending each component Y of S0 \ C0 into the
open star of some vertex vY of Γ. We shall show that no vY may be equal to v. Assuming
this, we deduce that C0 has to be empty, since every curve in C0 separates a component
mapped to the star of v from a component mapped to the star of some vi , and the lemma
follows.
Let therefore Y be a component with vY = v, and Z its closure. The components of ∂Z
come from either C0 or ∂Σ. The restriction πZ of π to π1 (Z) has an image contained in a
conjugate of π1 (Σ), which we may assume to be π1 (Σ) itself. It is non-pinching (its kernel
cannot contain the fundamental group of an essential simple closed curve) by maximality
of C, and it sends any boundary subgroup H of π1 (Z) into a boundary subgroup of π1 (Σ)
because πZ (H) fixes an edge of T . By Lemma 6.2 of [Per11], there are three possibilities:
either ∂Z = ∅, or the image of πZ is contained in a boundary subgroup of π1 (Σ), or it has
finite index in π1 (Σ). We show that each possibility contradicts one of our assumptions.
If ∂Z is empty (so, in particular, C0 = ∅ and Z = S0 ), we note that π1 (Σ) is a free
group because n ≥ 1, and we represent πZ : π1 (Z) → π1 (Σ) by a map from Z to a graph.
Maximality of C implies that Z = S0 is a sphere or a projective plane, so the image of π
is trivial, a contradiction.
Now suppose that the image of πZ is contained in a boundary subgroup H of π1 (Σ).
Denote by ṽ the vertex of T fixed by π1 (Σ), and by ṽṽi the unique edge incident on ṽ
6
fixed by H (a lift of some edge vvi of Γ). Consider the component Ỹ of the preimage of
Y in S̃0 mapping to the star of ṽ. The union of all regions adjacent to Ỹ is mapped by
f˜ into the open star of ṽi , and we may redefine f˜ so as to remove C0 ∩ ∂Z from C0 , thus
contradicting the original choice of f˜ (if C0 ∩ ∂Z = ∅, then Z = S0 and we may replace f˜
by the constant map equal to ṽi ).
If the image of πZ has finite index d in π1 (Σ), then d has to be 1 because χ(Σ) ≥
χ(S0 ) ≥ χ(Z) = dχ(Σ), and we conclude from Lemma 3.12 of [Per11] that Σ = S0 = Z
and p is an isomorphism from π1 (Σ) to itself, contradicting our hypothesis.
Remark 3.4. The proof shows the following more general result, which will be useful
in [GLS]: Let Γ and Σ be as in Lemma 3.3. Let Σ0 be a compact surface, and let p :
π1 (Σ0 ) → G be a homomorphism such that the image of every boundary subgroup of π1 (Σ0 )
is contained in a conjugate of some Gvi . Let C be a maximal family of pinched curves on
Σ0 , and let S be a component of the surface obtained by cutting Σ0 along C. Then, up to
conjugacy in G, the image of π1 (S) by p is contained in some Gvi , or there is a subsurface
Z ⊂ S such that p(π1 (Z)) is a finite index subgroup of π1 (Σ).
Corollary 3.5. Let Γ be as in Lemma 3.3. Denote by n the total number of vertices vi
of Γ, by n1 the number of vi ’s which have valence 1, and by b the number of boundary
components of Σ.
If p : π1 (Σ) → G maps every boundary subgroup of π1 (Σ) injectively into a G-conjugate,
and p is not an isomorphism onto a G-conjugate of π1 (Σ), then the genus g of Σ satisfies
g ≥ n1 and g + b ≥ 2n.
Recall that g = 21 (2 − χ(Σ) + b) if Σ is orientable, and g = 2 − χ(Σ) + b otherwise. By
a G-conjugate of H, we mean a subgroup of G conjugate to H.
Proof. We fix a maximal family of pinched curves C ⊂ Σ, and we let Σ̂ be the (possibly
disconnected) surface obtained by cutting Σ along C.
Consider a vertex vi of Γ of valence 1. Let C be the boundary component of Σ
associated to the edge vvi , and let S be the component of Σ̂ containing C. By Lemma 3.3,
the image of π1 (S) by the restriction of p fixes a vertex of T which is not a lift of v. Because
of our assumption on p(π1 (C)), this vertex must be in the orbit of vi . It follows that C
is the only boundary component of Σ contained in S: since vi has valence 1, the group
Gvi cannot intersect non-trivially a conjugate of π1 (C 0 ) if C 0 6= C is another component
of ∂Σ. The boundary of S thus consists of C and curves from C. Since π1 (C) is not
contained in ker p but curves in C are pinched, S cannot be planar: it cannot be a sphere
with holes since a generator of π1 (C) would then be a product of elements representing
pinched curves, contradicting injectivity of p on π1 (C).
We conclude that Σ̂ has at least n1 non-planar components. This implies g ≥ n1 . The
second inequality follows since n1 + b ≥ 2n.
4
Preretractions
Definition 4.1 (JSJ-like decomposition, [Per11, Definition 5.8]). Let Γ be a graph of
groups decomposition of a group G. It is JSJ-like if:
• edge groups are infinite cyclic;
• at most one vertex of any given edge is a surface-type vertex (in the sense of Definition 3.1), and at most one carries a cyclic group;
• (Acylindricity) if a non-trivial element of G fixes two distinct edges of T , then they
are adjacent and their common vertex has cyclic stabilizer.
7
The basic example of a JSJ-like decomposition is the canonical cyclic JSJ decomposition of a torsion-free one-ended hyperbolic group G. More generally, the tree of cylinders
(in the sense of [GL11]) of any splitting of a torsion-free CSA group G over infinite cyclic
groups is JSJ-like.
Proposition 4.2 ([Per11, Proposition 6.1]). Let Γ be a JSJ-like graph of groups decomposition of a group G. If r : G → G sends each vertex group and each edge group isomorphically
to a conjugate of itself, then r is an automorphism.
This statement is stronger than Proposition 6.1 of [Per11], but the proof is identical.
If G is torsion-free hyperbolic, Γ may be any graph of groups decomposition with infinite
cyclic edge groups (one applies the proposition to its tree of cylinders).
Definition 4.3 (Exceptional surfaces). Pairs of pants, once-punctured Klein bottles, twicepunctured projective planes, and closed non-orientable surfaces of genus 3, are surfaces
with Euler characteristic -1 which do not carry pseudo-Anosov diffeomorphisms. We call
them, as well as surface-type vertices associated to them, exceptional.
Definition 4.4 (Preretraction, [Per11, Definition 5.9]). Let Γ be a JSJ-like decomposition
of a group G. A preretraction associated to Γ is a homomorphism r : G → G such that,
for each vertex group Gv :
1. if Gv is not surface-type, the restriction of r to Gv is conjugation by some gv ∈ G;
2. if Gv is an exceptional surface-type vertex group, the restriction of r to Gv is conjugation by some gv ∈ G;
3. if Gv is a non-exceptional surface-type vertex group, then the image of Gv is nonabelian.
Note that this implies that the restriction of r to any edge group of Γ is a conjugation.
Example 4.5. Let G be the fundamental group of a non-exceptional closed surface. Any
map from G onto a non-abelian free subgroup is a non-injective preretraction (with Γ the
trivial splitting).
Conditions 2 and 3 are important in order to draw conclusions from the existence of
a non-injective preretraction (as in [Per11]). In the present paper, however, we often do
not need these assumptions.
Definition 4.6 (Weak preretraction). A map r : G → G satisfying Condition 1 of Definition 4.4 will be called a weak preretraction.
Proposition 4.2 implies that a preretraction r is an isomorphism if there is no nonexceptional surface-type vertex. More generally:
Proposition 4.7. Let Γ be a JSJ-like decomposition of a group G. The following are
equivalent:
• there is a non-injective preretraction r associated to Γ;
• Γ has a non-exceptional surface-type vertex group Gv = π1 (Σ), and there is a map
p : π1 (Σ) → G which is a conjugation on each boundary subgroup, has non-abelian
image, and is not an isomorphism of π1 (Σ) onto a conjugate.
Similarly, there exists a non-injective weak preretraction r if and only if there is an
arbitrary surface-type Gv with a map p which is a conjugation on each boundary subgroup
and is not an isomorphism of π1 (Σ) onto a conjugate.
8
Proof. The existence of r implies that of p by Proposition 4.2. To prove the converse direction, we may collapse the edges of Γ not containing the surface-type vertex v associated
to Σ and thus assume that Γ is as in Lemma 3.3 (unless G is a non-exceptional closed
surface group, in which case p itself is a preretraction which is non-injective because G is
co-Hopfian). As in Lemma 3.3, we denote by n the number of vertices adjacent to v.
Given p which is a conjugation on each boundary subgroup, we may extend it “by the
identity” to a (weak) preretraction r. We first illustrate this when Σ has two boundary
components (so n ≤ 2).
If n = 2, we have G = A1 ∗C1 π1 (Σ) ∗C2 A2 , and p equals conjugation by some gi on
the cyclic group Ci ; we then define r as being conjugation by gi on Ai . If n = 1, then
G has presentation hπ1 (Σ), A, t | c1 = a1 , tc2 t−1 = a2 i, with ai ∈ A and ci generators of
boundary subgroups Ci of π1 (Σ). If pi is conjugation by gi on Ci , we define r as being
conjugation by g1 on A and mapping t to g1 tg2 −1 .
The general case is similar; if Γ is not a tree, one chooses a maximal subtree in order
to get a presentation with stable letters as in the case n = 1 above.
There remains to check that r is non-injective. We may assume that p is injective.
Then there is no pinching, so by Lemma 3.3 the image of p is contained in a conjugate of
some Gvi . Let ẽ = ṽṽi ⊂ T be a lift of the edge vvi . We may assume that π1 (Σ) is the
stabilizer of ṽ, and p is the identity on the stabilizer of ẽ. The extension r then preserves
the stabilizer of ṽi , and up to postcomposing r with the conjugation by some element of
Gẽ , we may assume that r is the identity on the stabilizer of ṽi . The image of p fixes a
vertex in the orbit of ṽi , and by acylindricity this vertex must be ṽi . It follows that r
cannot be injective.
Remark 4.8. In particular, if there is a non-injective preretraction associated to an arbitrary non-trivial JSJ-like decomposition, there is one associated with a Γ as in Lemma
3.3.
5
Preretractions and elementary equivalence
Let G be a finitely generated group. It may be written as A1 ∗ · · · ∗ An ∗ F , with Ai
non-trivial, not isomorphic to Z, freely indecomposable, and F free. We call the Ai ’s the
Grushko factors of G.
The following is folklore, but we give a proof for completeness.
Proposition 5.1. Let G and G0 be torsion-free hyperbolic groups. Assume that G and G0
are elementarily equivalent. If A is a Grushko factor of G, and A has no non-injective
preretraction associated to its cyclic JSJ decomposition Γcan , then A embeds into G0 .
Such an A is an example of an elementary prototype [Sel09, Definition 7.3].
Proof. We assume that A does not embed into G0 , and we construct a non-injective preretraction. By Sela’s shortening argument (see [Per11, Proposition 4.3]), there exist finitely
many non-trivial elements ri ∈ A such that, given any homomorphism f : A → G0 (necessarily non-injective), there is a modular automorphism σ of A such that the kernel of
f 0 = f ◦ σ contains one of the ri ’s.
Let Γcan be the canonical cyclic JSJ decomposition of A. By definition, modular
automorphisms act on non-surface vertex groups Gv of Γcan by conjugation, so f and f 0
differ by an inner automorphism on such vertex groups Gv . If Γ has an exceptional surfacetype vertex group Gu that has a Dehn twist of infinite order (i.e. if Σ is a once-punctured
Klein bottle or a closed non-orientable surface of genus 3), this property does not hold for
Gu (this detail was overlooked in [Per11, Lemma 5.6]).
To remedy this, one considers Γ0can obtained from Γcan by splitting Gu along the
fundamental group of a suitable 2-sided simple closed curve which is invariant under the
9
mapping class group (see Remark 9.32 in [GL]); in Γ0can , all exceptional surfaces have
finite mapping class group, and modular automorphisms of A act by conjugation on all
exceptional surface groups of Γ0can .
Thus the following statement, which is expressible in first-order logic, holds (compare
Section 5.4 of [Per11]): given any f : A → G0 such that non-exceptional surface-type vertex
groups Gv of Γ0can have non-abelian image, there is f 0 : A → G0 such that ker f 0 contains
one of the ri ’s, non-exceptional surface-type vertex groups Gv have non-abelian image by
f 0 , and for each other vertex group Gv of Γ0can there is g ∈ G0 such that f 0 agrees with
ig ◦ f on Gv (with ig denoting conjugation by g).
Since G and G0 are elementarily equivalent, this statement holds with f the inclusion
from A to G, and yields a non-injective f 0 : A → G. Composing with a projection from
G onto A yields a non-injective weak preretraction from A to A associated to Γ0can . In
order to get a preretraction (i.e. to ensure that non-exceptional surface-type vertex groups
have non-abelian image), we apply Proposition 5.12 of [Per11]. By Proposition 4.7, this
preretraction also gives a non-injective preretraction associated to Γcan .
Corollary 5.2. If a finitely generated group G is elementarily free, all its Grushko factors
have non-injective preretractions associated to their cyclic JSJ decompositions.
Recall that every finitely generated elementarily free group is torsion-free and hyperbolic ([Sel01], [Sel06]).
Corollary 5.3. Let G and G0 be one-ended torsion-free hyperbolic groups. Suppose that
G, G0 do not have non-injective preretractions associated to their cyclic JSJ decompositions. If G and G0 are elementarily equivalent, they are isomorphic.
This follows from the proposition and the fact that G and G0 are co-Hopfian [Sel97].
6
Surfaces with sockets
The fundamental group G = π1 (Σ) of a closed hyperbolic surface Σ is a limit group which
is elementarily equivalent to a non-abelian free group, with one exception: if Σ is the
non-orientable surface of genus 3, then G is not even a limit group.
In this section we consider fundamental groups of surfaces with sockets (which we
simply call socket groups). They are obtained from π1 (Σ), with Σ a compact hyperbolic
surface of genus g with b ≥ 1 boundary components Bi , by adding a root to the element
hi representing Bi (see Figure 1). More precisely, G has presentation
hh1 , . . . , hb , z1 , . . . , zb , a1 , b1 ), . . . , ag , bg | h1 · · · hb = [a1 , b1 ] · · · [ag , bg ], zini = hi i
or
hh1 , . . . , hb , z1 , . . . , zb , a1 , . . . , ag | h1 · · · hb = a21 · · · a2g , zini = hi i,
depending on whether Σ is orientable or not. The exponent ni , the order of the i-th socket,
is a positive integer, and we always assume ni ≥ 3: if ni = 2, we can remove the socket
by attaching a Möbius band to Bi .
Socket groups are torsion-free one-ended hyperbolic groups. They were introduced in
[Sel97] to study splittings over maximal cyclic subgroups.
We shall show:
Proposition 6.1. The group G is a limit group if and only if one of the following holds:
• the surface Σ is orientable with at least four boundary components (i.e. b ≥ 4);
• the surface Σ is orientable with two or three boundary components and genus g ≥ 1;
10
Figure 1: A socket group can be seen topologically as the fundamental group of a surface
for which every boundary component is glued onto a circle under an n-cover map for some
n.
• the surface Σ is orientable with one boundary component and genus g ≥
n is the order of the unique socket;
n+1
2 ,
where
• the surface Σ is non-orientable and b + g ≥ 4.
Proposition 6.2. The following assertions are equivalent:
1. the group G is elementarily equivalent to a free group;
2. the groups G has a non-injective preretraction associated to its cyclic JSJ decomposition;
3. the surface Σ is non-orientable, b + g ≥ 4, every ni is even, and g ≥ b.
The group G is naturally the fundamental group of a graph of groups Γ with a central
vertex carrying π1 (Σ), and b terminal vertices carrying hzi i (the edge groups are generated
by the hi ’s). It is a torsion-free one-ended hyperbolic group, and the assumption ni ≥ 3
ensures that Γ is its canonical cyclic JSJ decomposition (see Proposition 4.24 in [DG11]).
Proof of Proposition 6.1. By Proposition 4.21 of [CG05], G is a limit group if and only if
the equation
z1n1 · · · zbnb = [a1 , b1 ] · · · [ag , bg ]
or
z1n1 · · · zbnb = a21 · · · a2g ,
with unknowns zi , ai , bi in a free group F (x, y), has a solution which is not contained in a
cyclic subgroup, and with every zi non-trivial.
k
Recall that, given integers ki ≥ 2, a relation xk11 · · · xpp = 1 between elements of F (x, y)
implies that hx1 , . . . , xp i is cyclic when p ≤ 3 [LS62]. Thus G is not a limit group if Σ is
non-orientable and b + g ≤ 3. On the other hand, if Σ is non-orientable and b + g ≥ 4, it
is easy to find a suitable solution of the equation above, by setting all unknowns equal to
powers of x or y.
Now suppose that Σ is orientable and there is a single socket (b = 1). If g ≥ n+1
2 , it
follows from [Cul81] that [x, y]n is a product of g commutators, so G is a limit group. If
g < n+1
2 , no non-trivial n-th power is a product of g commutators by [CE95], so G is not
a limit group.
Finally, assume that Σ is orientable and b ≥ 2. As suggested by Jim Howie, the
equation z1n1 z2n2 = [a1 , b1 ] is solved by setting z1 = xn2 , z2 = yx−n1 y −1 , a1 = xn1 n2 ,
b1 = y. It follows that G is a limit group when g ≥ 1. If g = 0, the fact recalled above
implies that G is a limit group if and only if b ≥ 4.
Proof of Proposition 6.2. We first prove (2) =⇒ (3). Assuming that there is a preretraction, we apply Lemma 3.3 to the map p provided by Proposition 4.7. Cut Σ open along
C, and call Si the component of the surface thus obtained which contains Bi . For each i,
up to conjugation, we get a map from π1 (Si ) to hzi i sending hi to itself and killing the
fundamental group of every other boundary component of Si . This is possible only if Si
is non-orientable and ni is even.
11
The inequality g ≥ b follows from Corollary 3.5. By Proposition 4.2, the existence of
a non-injective preretraction forces that of a non-exceptional surface. This rules out the
once-punctured Klein bottle (g = 2, b = 1) and thus implies b + g ≥ 4.
Conversely, we assume that Σ and the ni ’s are as in (3), and we construct a noninjective preretraction r fixing every zi (hence every hi ). Write ni = 2mi . If b = 1, then
g ≥ 3; we choose z ∈ G not commuting with z1 , and we map a1 to z1m1 , a2 to z, a3 to z −1 ,
and ai to 1 for i > 3. If 2 ≤ b ≤ g, we map ai to zimi for i ≤ b, to 1 for i > b. This proves
(3) =⇒ (2).
(1) =⇒ (2) follows from Corollary 5.2. Conversely, the map r constructed above expresses G as an extended hyperbolic tower over a free group, showing that G is elementarily
free [Sel06].
Remark 6.3. The proof shows that G has a non-injective weak preretraction (in the sense of
Definition 4.6) associated to its cyclic JSJ decomposition if and only if Σ is non-orientable,
every ni is even, and g ≥ b.
Corollary 6.4. Let G be a socket group, with Σ a once-punctured orientable surface of
genus 2, and a single socket of order 3 (see Figure 2). Then G is a one-ended hyperbolic
limit group G which is not elementarily free but contains an elementarily free subgroup of
finite index.
Remark 6.5. In fact, G has no non-injective weak preretraction by Remark 6.3.
Proof. By Propositions 6.1 and 6.2, G is a limit group that is not elementarily free. It
has a subgroup G0 of index 3, obtained as the kernel of a map from G to Z/3Z killing
a1 , b1 , a2 , b2 , which is the fundamental group of the space obtained by gluing three oncepunctured surfaces of genus 2 along their boundary. Identifying two of these surfaces yields
a retraction from G0 onto the fundamental group of the closed orientable surface of genus
4, so G0 is elementarily free by [Sel06].
Figure 2: A covering of degree 3.
We now consider a slightly different family of groups, where all sockets are identified.
More precisely, G is the fundamental group of a graph of groups with one surface-type
vertex group Gv = π1 (Σ), where Σ has at least two boundary components, and one vertex
carrying a cyclic group hzi. It is presented as
hh1 , . . . , hb , z, t1 , . . . , tb , a1 , b1 , . . . , ag , bg | h1 · · · hb = [a1 , b1 ] · · · [ag , bg ], ti z ni ti −1 = hi , t1 = 1i
or
hh1 , . . . , hb , z, t1 , . . . , tb , a1 , . . . , ag | h1 · · · hb = a21 · · · a2g , ti z ni ti −1 = hi , t1 = 1i,
12
Figure 3: A socket group with identified sockets.
depending on whether Σ is orientable or not (see Figure 3).
When Σ is orientable one may orient the components of ∂Σ in a consistent way, and
the sign of the ni ’s matters. In the non-orientable case the components may be oriented
separately and only |ni | has a meaning.
Proposition 6.6. Let G be a socket group with identified sockets as above. There is a noninjective weak preretraction associated to the cyclic JSJ decomposition of G if and only if
b
b
X
X
the sum
ni is 0 when Σ is orientable, the sum
ni is even when Σ is non-orientable.
i=1
i=1
Proof. Assume that there is p : π1 (Σ) → G which is a conjugation on each boundary
subgroup and is not an isomorphism of π1 (Σ) onto a conjugate (see Proposition 4.7). We
apply Lemma 3.3. First suppose there is no pinching. All hi ’s are mapped into the same
ni
conjugate of hzi, so we may assume that hi is mapped to a power
P of z, which must be z
p
q
n
(note that z conjugate to z implies p = q). We deduce that z i is a product of squares
or commutators in hzi. If there is pinching, apply the argument to each pinched surface.
Conversely, we define a non-injective preretraction r as follows:
we map z to z, every
P
n
(
n
)/2
i
i
hi to z , every ti , ai , bi to 1, except that we map a1 to z
in the non-orientable
case.
Deciding whether there is a non-injective (true) preretraction r (with Σ non-exceptional
and the image of π1 (Σ) non-abelian) is more subtle.
First suppose that Σ is orientable. If its genus is ≥ 1, one may make r(π1 (Σ)) nonabelian by sending a1 and b1 to some x ∈ G which does not commute with z.
If Σ is a 4-punctured sphere and r exists, Lemma 3.3 implies that there is a pinched
curve. It must divide Σ into two pairs of pants, each containing two boundary components
of Σ, say B1 and B2 on one side, B3 and B4 on the other. One must then have n1 + n2 =
n3 + n4 = 0. Conversely, if this holds, one defines r on the generating set {z, t2 , t3 , t4 } by
sending z to itself, t2 to the trivial element, and by sending t3 and t4 to some x ∈ G such
that z and xzx−1 do not commute. This argument shows:
Proposition 6.7. Let G be a socket group with identified sockets as above, with Σ an
orientable surface of genus g. There is a non-injective preretraction (i.e. G is elementarily
b
X
free) if and only if the sum
ni is 0 and either:
i=1
(i) g ≥ 1, or
(ii) g = 0 and there is a proper nonempty subset I ⊂ {1, . . . , b} with
13
P
i∈I
ni = 0.
When Σ is non-orientable, the easy case is when g ≥ 3 since one may use a2 and a3
to make r(π1 (Σ)) non-abelian. When g is 1 or 2, one has to consider the ways in which a
maximal family of pinched curves may divide Σ. A case by case analysis yields:
Proposition 6.8. Let G be a socket group with identified sockets as above, with Σ a
non-orientable surface of genus g. Then there is a non-injective preretraction (i.e. G is
b
X
elementarily free) if and only if the sum
ni is even and one of the following holds:
i=1
(i) g ≥ 3;
(ii) g = 2 and Σ has at least three boundary components (i.e. b ≥ 3);
(iii) g = 2, the surface Σ has two boundary components (i.e. Σ is a twice punctured Klein
bottle), and |n1 | = |n2 | or both n1 , n2 are even;
(iv) g = 1, the surface Σ has at least three boundary componentsP(i.e. b ≥ 3), and there
exist a proper subset I ⊂ {1, . . . , b} and signs εi = ±1 with i∈I εi ni = 0.
7
Finite index subgroups
The most obvious examples of limit groups, namely free abelian groups, free groups, surface
groups, are elementarily equivalent to their finite index subgroups. The goal of this section
is to show that these are basically the only examples of limit groups with this property,
at least among hyperbolic limit groups.
Theorem 7.1. Let G be a hyperbolic limit group, or a torsion-free cubulable hyperbolic
group. If G is not a free product of cyclic groups and of surface groups, it has infinitely
many normal finite index subgroups which are all different up to elementary equivalence.
Corollary 7.2. Let G be a finitely generated elementarily free group. If G is not a free
product of cyclic groups and surface groups, it has infinitely many finite index subgroups
which are not elementarily free.
On the other hand, if G is a free product of cyclic groups and fundamental groups
of closed surfaces with χ(Σ) ≤ −2, it is elementarily free by [Sel06], and so are all its
non-cyclic finitely generated subgroups.
The corollary is clear, because a finitely generated elementarily free group is a hyperbolic limit group by [Sel01] and [Sel06]. We shall deduce the theorem from a more
technical statement, which requires a definition.
Definition 7.3. We say that G has enough abelian virtual quotients if, given any infinite
cyclic subgroup C, there is a finite index subgroup GC of G such that C ∩ GC has infinite
image in the abelianization of GC ; equivalently, C has a finite index subgroup which is a
retract of a finite index subgroup of G.
Note that having enough virtual abelian quotients is inherited by subgroups. This
property is slightly weaker than G being LR over cyclic groups in the sense of [LR08]
(they require that C itself be a retract of a finite index subgroup). It implies the existence
of infinitely many finite index subgroups.
Theorem 7.4. Let G be a one-ended torsion-free hyperbolic group. If G has enough virtual
abelian quotients and is not a surface group, then G has infinitely many characteristic
subgroups of finite index which have no non-injective weak preretraction associated to their
cyclic JSJ decomposition.
14
The theorem readily extends to groups which are only virtually torsion-free and are
not virtual surface groups.
Having enough virtual abelian quotients is a technical assumption (which holds for
limit groups and cubulable hyperbolic groups, as we shall see). It is not optimal, but note
that some assumption is needed: if there is a hyperbolic group which is not residually
finite, there is one which has no proper subgroup of finite index [KW00], and such a
group obviously does not satisfy Theorem 7.4. Also note that groups with property (T)
do not have enough virtual abelian quotients, but they cannot have non-injective weak
preretractions because they have no splittings.
We first explain how to deduce Theorem 7.1 from Theorem 7.4.
Lemma 7.5. Residually free groups (in particular limit groups) and cubulable hyperbolic
groups have enough abelian virtual quotients.
Proof. If G is residually free, there is a map p from G to a free group which is injective on
C. By Hall’s theorem, the image of C under p is a free factor of a finite index subgroup,
and we let GC be its preimage.
If G is hyperbolic and cubulable, then by [Ago13] it is virtually special: there is some
finite index subgroup G1 < G acting freely and cocompactly on a CAT(0) cube complex
X such that X/G1 is special. Since quasiconvex subgroups are separable by [HW08,
Corollary 7.4], up to passing to a further finite index subgroup, we can assume that the 2neighbourhoods of hyperplanes are embedded in X/G1 so that X/G1 is fully clean. Denote
C1 = C ∩ G 1 .
By [HW08, Proposition 7.2], there is a convex cube complex Y ⊂ X which is C1 invariant and C1 -cocompact and defines a local isometry f : Y /C1 → X/G1 which is fully
special by [HW08, Lemma 6.3]. By [HW08, Proposition 6.5], there is a finite cover X/G2
of X/G1 and an embedding Y /C1 → X/G2 with a retraction X/G2 → Y /C1 . Grouptheoretically, we get a retraction from the finite index subgroup G2 of G to C1 . This
proves the lemma.
Note that any residually free hyperbolic group is a limit group [Bau67].
Proof of Theorem 7.1. Given G as in the theorem, we write its Grushko decomposition as
G = A1 ∗ · · · ∗ Ap ∗ F ∗ S1 ∗ · · · ∗ Sq where F is free, each Si is a closed surface group, and
each Ai is a one-ended group which is not a surface group (the Ai ’s are not necessarily
distinct up to isomorphism). By assumption, we have p ≥ 1.
By Theorem 7.4 and Lemma 7.5, we can find in each Ai a decreasing sequence Ai (1) %
Ai (2) % Ai (3) % . . . of characteristic finite index subgroups which have no non-injective
preretraction associated to theirQcyclic JSJ decomposition. Let Gn be the kernel of the
natural projection from G onto pi=1 Ai /Ai (n). It is a finite index subgroup of G whose
non-surface Grushko factors are all isomorphic to some Ai (n).
We claim that there are infinitely many distinct Gn ’s up to elementary equivalence. If
not, we may assume that they are all equivalent. Applying Proposition 5.1 to G1 and G2 ,
we find that A1 (1) embeds into G2 , hence into some Grushko factor Aj (2). We cannot
have j = 1 because A1 (1) properly contains A1 (2) and torsion-free one-ended hyperbolic
groups are co-Hopfian [Sel97]. We may therefore assume j = 2. We then find that A2 (2)
embeds in some A3 (k), and k > 2 by co-Hopfianity. Iterating this argument leads to a
contradiction.
The remainder of this section is devoted to the proof of Theorem 7.4.
We consider the canonical cyclic JSJ decomposition Γcan (G), or simply Γcan , and the
associated Bass-Serre tree Tcan . Its non-rigid vertices are of surface-type.
Recall that the girth of a graph is the smallest length of an embedded circle (∞ if the
graph is a tree). We say that Γcan has large girth if its girth N is large compared to the
15
complexity of the surfaces which appear in Γcan ; precisely, (N − 2)/2 should be larger than
the maximal cardinality of a family of non-parallel disjoint simple closed curves on a given
surface (the curves may be boundary parallel, but should not be null-homotopic).
If G0 is a finite index subgroup, its JSJ tree is obtained by restricting the action of
G on Tcan to G0 . This is because by [Bow98] one may construct Tcan purely from the
topology of the boundary of G, and ∂G0 = ∂G. If v is a surface-type vertex for the action
of G on Tcan , it is one also for the action of G0 , but in general the surface is replaced by
a finite cover.
The proof requires two lemmas.
Lemma 7.6. There is a characteristic subgroup of finite index G0 ⊂ G whose JSJ decomposition has the following properties:
1. the surfaces which appear in Γcan (G0 ) are all orientable;
2. each edge group maps injectively to the abelianization of G0 ;
3. the graph Γcan (G0 ) has large girth (as defined above).
Note that the first two properties are inherited by finite index subgroups.
Lemma 7.7. Let G be a torsion-free one-ended hyperbolic group whose JSJ decomposition
satisfies the properties of the previous lemma. If G is not a surface group, then G has no
non-injective weak preretraction associated to its cyclic JSJ decomposition.
Before proving these lemmas, let us explain how they imply Theorem 7.4.
First suppose that there is G0 as in Lemma 7.6 such that Γcan (G0 ) is not a tree. Then
choose a decreasing sequence of characteristic finite index subgroups H1 ⊃ H2 ⊃ · · · of
the topological fundamental group of the graph Γcan (G0 ), and define Gn as the preimage
of Hn under the natural epimorphism from G0 to the fundamental group of the graph
Γcan (G0 ).
Since Γcan (G0 ) is invariant under automorphisms of G0 , each Gn is characteristic in G.
Its JSJ decomposition is the covering of Γcan (G0 ) associated to Hn , with the lifted graph
of groups structure. The group Gn satisfies the first two properties of Lemma 7.6, and also
the third one because the surfaces appearing in Γcan (Gn ) are the same as in Γcan (G0 ), so
Lemma 7.7 applies.
If Γcan (G0 ) is a tree for every subgroup of finite index G0 < G as in Lemma 7.6, we fix
any such G0 and we let Gn be any sequence of distinct characteristic subgroups of finite
index (it exists because there are enough abelian virtual quotients). We cannot control
the complexity of surfaces, but Lemma 7.7 applies to Gn because the girth is infinite.
We now prove the lemmas.
Proof of Lemma 7.6. Since the first two properties are inherited by finite index subgroups,
it suffices to construct G0 having one given property.
To achieve (1), view the JSJ decomposition Γcan as expressing G as the fundamental
group of a graph of spaces X, with a surface Σ for each surface-type vertex. If there is
a non-orientable surface, consider a 2-sheeted covering of X which is trivial (a product)
above the complement of the non-orientable surfaces, and is the orientation covering Σ̂
over each non-orientable surface (note that a boundary component of Σ lifts to two curves
in Σ̂). Though Γcan is canonical, the 2-sheeted covering is not, so we define G0 as the
intersection of all subgroups of G of index 2.
For (2), let Ci be the edge groups of Γcan , and let G0 be any characteristic subgroup
of finite index contained in every group GCi provided by Definition 7.3. Each edge group
of Γcan (G0 ) is the image of some Ci ∩ G0 by an automorphism of G0 , so maps injectively
to the abelianization.
16
If the girth of Γcan is too small, we choose a characteristic finite index subgroup H0 of
the topological fundamental group π1 (Γcan ) such that the associated finite cover of Γcan
has large girth, and we define G0 as the preimage of H0 under the epimorphism from
G to the topological fundamental group of Γcan . As in the proof of the theorem given
above, invariance of Γcan under automorphisms implies that G0 is characteristic, and the
complexity of surfaces does not change.
Proof of Lemma 7.7. We consider the canonical cyclic JSJ decomposition Γ = Γcan . Recall
that it is bipartite: each edge joins a vertex carrying an infinite cyclic group to a vertex
carrying a non-cyclic group.
We argue by way of contradiction, assuming that there is a non-injective weak preretraction. By Proposition 4.7, there is a surface-type vertex group Gv = π1 (Σ) and a
map p : π1 (Σ) → G which is a conjugation on each boundary subgroup and is not an
isomorphism of π1 (Σ) onto a conjugate.
Let C be a maximal family of pinched curves on Σ (see Definition 3.2), and let S be
a component of the compact surface obtained by cutting Σ along C. Since G is not a
surface group, we may assume that S contains a boundary component C of Σ. It must
contain another boundary component C 0 of Σ: otherwise, since Σ is orientable and curves
in C are pinched, the image of a generator of π1 (C) would be a product of commutators,
contradicting the second item in Lemma 7.6.
We apply Lemma 3.3 to the graph of groups obtained from Γ by collapsing all edges
which do not contain v. We find that the image of π1 (S) by p is contained (up to conjugacy)
in the fundamental group Ri of a subgraph of groups Γi of Γ, which is a component of the
complement of the open star of v. The image of the fundamental groups of C and C 0 are
contained in the groups carried by distinct vertices w and w0 of Γi which are adjacent to
v in Γ, but far away from each other in Γi because the girth of Γ is large.
As in the proof of Lemma 3.3, we construct a surface S0 by attaching discs to boundary
curves of S coming from C, and we represent the induced map π : π1 (S0 ) → Ri by an
equivariant map from the universal covering S̃0 to the Bass-Serre tree of Γi . We consider
preimages of midpoints of edges, and we project to S0 . We obtain a finite family C0 of
disjoint simple closed curves on S0 .
View Γi as covered by stars of vertices u carrying a cyclic group. If C0 is a curve in
C0 , the image of π1 (C0 ) by π is contained (up to conjugacy) in a unique Gu , and curves
associated to different vertices cannot be parallel. Since any path joining w to w0 in Γi
must go through at least (N −2)/2 stars, with N the girth of Γ, the family C0 must contain
at least (N − 2)/2 non-parallel curves. We get a contradiction if N is large enough.
References
[Ago13] Ian Agol. The virtual Haken conjecture. Doc. Math., 18:1045–1087, 2013. With
an appendix by Agol, Daniel Groves, and Jason Manning.
[Bau67] Benjamin Baumslag. Residually free groups. Proc. London Math. Soc. (3),
17:402–418, 1967.
[Bow98] Brian H. Bowditch. Cut points and canonical splittings of hyperbolic groups.
Acta Math., 180(2):145–186, 1998.
[CE95]
Leo P. Comerford, Jr. and Charles C. Edmunds. Genus of powers in a free group.
In Geometric group theory (Columbus, OH, 1992), volume 3 of Ohio State Univ.
Math. Res. Inst. Publ., pages 67–71. de Gruyter, Berlin, 1995.
[CG05]
Christophe Champetier and Vincent Guirardel. Limit groups as limits of free
groups. Israel J. Math., 146:1–75, 2005.
17
[Cul81] Marc Culler. Using surfaces to solve equations in free groups.
20(2):133–145, 1981.
Topology,
[DG11] François Dahmani and Vincent Guirardel. The isomorphism problem for all
hyperbolic groups. Geom. Funct. Anal., 21(2):223–300, 2011.
[GL]
Vincent Guirardel and Gilbert Levitt.
arXiv:1602.05139 [math.GR].
JSJ decompositions of groups.
[GL11]
Vincent Guirardel and Gilbert Levitt. Trees of cylinders and canonical splittings.
Geom. Topol., 15(2):977–1012, 2011.
[GLS]
Vincent Guirardel, Gilbert Levitt, and Rizos Sklinos. in preparation.
[HW08] Frédéric Haglund and Daniel T. Wise. Special cube complexes. Geom. Funct.
Anal., 17(5):1551–1620, 2008.
[KM06] Olga Kharlampovich and Alexei Myasnikov. Elementary theory of free nonabelian groups. J. Algebra, 302(2):451–552, 2006.
[KW00] Ilya Kapovich and Daniel T. Wise. The equivalence of some residual properties
of word-hyperbolic groups. J. Algebra, 223(2):562–583, 2000.
[LR08]
D. D. Long and A. W. Reid. Subgroup separability and virtual retractions of
groups. Topology, 47(3):137–159, 2008.
[LS62]
R. C. Lyndon and M. P. Schützenberger. The equation aM = bN cP in a free
group. Michigan Math. J., 9:289–298, 1962.
[Per11]
Chloé Perin. Elementary embeddings in torsion-free hyperbolic groups. Ann.
Sci. Éc. Norm. Supér. (4), 44(4):631–681, 2011.
[Rem89] V. N. Remeslennikov. ∃-free groups. Siberian Math. J., 30(6):998–1001, 1989.
translation (1990) from Sibirsk. Mat. Zh. 30 (1989), 193–197.
[Sel97]
Z. Sela. Structure and rigidity in (Gromov) hyperbolic groups and discrete groups
in rank 1 Lie groups. II. Geom. Funct. Anal., 7(3):561–593, 1997.
[Sel01]
Zlil Sela. Diophantine geometry over groups. I. Makanin-Razborov diagrams.
Publ. Math. Inst. Hautes Études Sci., 93:31–105, 2001.
[Sel06]
Z. Sela. Diophantine geometry over groups. VI. The elementary theory of a free
group. Geom. Funct. Anal., 16(3):707–730, 2006.
[Sel09]
Z. Sela. Diophantine geometry over groups. VII. The elementary theory of a
hyperbolic group. Proc. Lond. Math. Soc. (3), 99(1):217–273, 2009.
[Ser77]
Jean-Pierre Serre. Arbres, amalgames, SL2 . Société Mathématique de France,
Paris, 1977. Rédigé avec la collaboration de Hyman Bass, Astérisque, No. 46.
[Szm55] W. Szmielew. Elementary properties of Abelian groups. Fund. Math., 41:203–
271, 1955.
18
Vincent Guirardel
Institut de Recherche Mathématique de Rennes
Université de Rennes 1 et CNRS (UMR 6625)
263 avenue du Général Leclerc, CS 74205
F-35042 RENNES Cédex
e-mail: [email protected]
Gilbert Levitt
Laboratoire de Mathématiques Nicolas Oresme
Université de Caen et CNRS (UMR 6139)
(Pour Shanghai : Normandie Univ, UNICAEN, CNRS, LMNO, 14000 Caen, France)
e-mail: [email protected]
Rizos Sklinos
Institut Camille Jordan
Université Claude Bernard Lyon 1 (UMR 5208)
43 boulevard du 11 novembre 1918
69622, Villeurbanne cedex, France
e-mail: [email protected]
19
| 4math.GR
|
ON THE SUBGROUPS AND NORMAL SUBGROUPS OF THE GROUP
REPRESENTATION OF THE CAYLEY TREE.
arXiv:1601.03291v1 [math.GR] 12 Jan 2016
F. H. HAYDAROV
Abstract. In this paper, we give a characterization of the normal subgroups of index
2s (2n+1), s ∈ {1, 2}, n ∈ N and of the subgroups of index three of the group representation
of the Cayley tree.
Key words. Gk - group, subgroup, normal subgroup, homomorphism, epimorphism.
AMS Subject Classification: 20B07, 20E06.
1. Introduction
In the theory of groups there are very important unsolved problems, most of which
arise in solving of problems of natural sciences as physics, biology etc. In particular, if
configuration of a physical system is located on a lattice (in our case on the graph of a
group) then the configuration can be considered as a function defined on the lattice. Usually, more important configuration (functions) are the periodic ones. It is well-known that
if the lattice has a group representation then periodicity of a function can be defined by a
given subgroup of the representation. More precisely, if a subgroup, say H, is given, then
one can define a H- periodic function as a function, which has a constant value (depending
only on the coset) on each (right or left) coset of H. So the periodicity is related to a
special partition of the group (that presents the lattice on which our physical system is
located). There are many works devoted to several kinds of partitions of groups (lattices)
(see e.g. [1], [3], [5], [7]).
One of the central problems in the theory of Gibbs measures is to study periodic
Gibbs measures corresponding to a given Hamiltonian of model. For any subgroup H of
the group Gk we define H-periodic Gibbs measures. To find new periodic and weakly periodic Gibbs measures one usually needs to find new subgroups of the group representation
of the Cayley tree. In Chapter 1 of [5] it is given a one to one correspondence between the
set of vertices V of the Cayley tree Γk and the group Gk and it is given a full description
of subgroups of index two and there are also constructed several normal subgroups of the
group Gk . But there wasn’t a full description of normal subgroups of finite index (without
index two) and to the best of our knowledge until now there wasn’t any description of a not
normal subgroup of finite index of the group representation of the Cayley tree. In [4] it is
given a full description of normal subgroups of indices 4 and 6 for the group representation
of the Cayley tree. In this paper we continue this investigation and construct all normal
subgroups of index 2s (2n + 1), s ∈ {1, 2}, n ∈ N and all subgroups of index three for the
1
2
F. H. HAYDAROV
group representation of the Cayley tree.
Cayley tree and its group representation. A Cayley tree (Bethe lattice) Γk of order k ≥ 1 is an infinite homogeneous tree, i.e., a graph without cycles, such that exactly
k + 1 edges originate from each vertex. Let Γk = (V, L) where V is the set of vertices
and L that of edges (arcs). Let Gk be a free product of k + 1 cyclic groups of the second order with generators a1 , a2 , ...ak+1 , respectively. It is known that there exists a one
to one correspondence between the set of vertices V of the Cayley tree Γk and the group
Gk . To give this correspondence we fix an arbitrary element x0 ∈ V and let it correspond
to the unit element e of the group Gk . Using a1 , ..., ak+1 we label the nearest-neighbors
of element e, moving in positive direction. Now we’ll label the nearest-neighbors of each
ai , i = 1, ..., k + 1 by ai aj , j = 1, ..., k + 1. Since all ai have the common neighbor e we
have ai ai = a2i = e. Other neighbors are labeled starting from ai ai in positive direction.
We label the set of all the nearest-neighbors of each ai aj by words ai aj aq , q = 1, ..., k + 1
starting from ai aj aj = ai by the positive direction. Iterating this argument one gets a oneto-one correspondence between the set of vertices V of the Cayley tree Γk and the group Gk .
Any(minimal represented) element x ∈ Gk has the following form: x = ai1 ai2 ...ain ,
where 1 ≤ im ≤ k + 1, m = 1, ..., n. The number n is called the length of the word x and is
denoted by l(x). The number of letters ai , i = 1, ..., k + 1, that enter the non-contractible
representation of the word x is denoted by wx (ai ).
The following result is well-known in group theory. If ϕ is a homomorphism of a
group G with the kernel H then H is a normal subgroup of the group G and ϕ(G) ≃ G/H,
(where G/H is the quotient group) so the index |G : H| coincides with the order |ϕ(G)| of
the group ϕ(G).
Usually we define natural homomorphism g from G onto the quotient group G/H by
the formula g(a) = aH for all a ∈ G. Then Kerϕ = H.
Definition 1. Let M1 , M2 , ..., Mm be some sets and Mi 6= Mj , for i 6= j. We call the
intersection ∩m
i=1 Mi contractible if there exists i0 (1 ≤ i0 ≤ m) such that
i0 −1
m
∩m
i=1 Mi = ∩i=1 Mi ∩ ∩i=i0 +1 Mi .
Let Nk = {1, ..., k + 1}. The following Proposition describes several normal subgroups
of Gk .
Put
(
)
X
H A = x ∈ Gk |
ωx (ai ) is even , A ⊂ Nk .
(1.1)
i∈A
Proposition 1. [5] For any ∅ =
6 A ⊆ Nk , the set HA ⊂ Gk satisfies the following properties:
(a) HA is a normal subgroup and |Gk : HA | = 2;
(b) HA 6= HB , for all A 6= B ⊆ Nk ;
(c) Let A1 , A2 , ..., Am ⊆ Nk . If ∩m
i=1 HAi is non-contractible, then it is a normal subgroup of
index 2m .
Theorem 1. [5]
SUBGROUPS AND NORMAL SUBGROUPS
3
1. The group Gk does not have normal subgroups of odd index (6= 1).
2. The group Gk has normal subgroups of arbitrary even index.
2. Subgroups and normal subgroups of finite index.
2.1. Normal subgroups of index 2s (2n + 1), s ∈ {1, 2}, n ∈ N.
Definition 2. A commutative group G of order 2n is called K2n − group if G is generated
by free product of n elements ci , i = 1, n satisfying the relations o(ci ) = 2, i = 1, n,
Proposition 2. Let ϕ be a homomorphism of the group Gk onto a finite commutative group
G. Then ϕ(Gk ) is isomorphic to K2i for some i ∈ N.
Proof. Let (G, ∗) be a commutative group of order n and ϕ : Gk → G be an epimorphism.
We will first show that n ∈ {2i | i ∈ N}. Suppose n does not belong to {2i | i ∈ N}.
Then there exist m > 1, m, s ∈ N ∪ {0} such that n = 2s m, where m is odd. Since
ϕ : Gk → G is an epimorphism there exist distinct elements ϕ(ajp ), p = 1, s. Clearly,
< ϕ(aj1 ), ϕ(aj2 )..., ϕ(ajs ) > is a subgroup of G and | < ϕ(aj1 ), ϕ(aj2 )..., ϕ(ajs ) > | = 2s .
Since m > 1 there exist at least one element aj0 ∈ Gk such that ϕ(aj0 ) 6= ϕ(aji ), i ∈
{1, 2, ..., s}. Hence < ϕ(aj0 ), ϕ(aj1 )..., ϕ(ajs ) > is a subgroup of G. Then by Lagrange’s
theorem |G| : 2s+1 which contradicts our assumption that m is odd. Thus n ∈ {2i | i ∈ N}.
If n = 2q then G is equal to < ϕ(aj1 ), ϕ(aj2 )..., ϕ(ajq ) > We define the following mapping
γ(aji ) = ci ∈ Kq for all i ∈ {1, 2, ..., q}. It is easy to check that this mapping is a one to one
correspondence from G to Kq . This completes the proof.
Let A1 , A2 , ..., Am ⊂ Nk , m ∈ N and ∩m
i=1 HAi is non-contractible. Then we denote by
ℜ the following set
ℜ = {∩m
i=1 HAi | A1 , A2 , ..., Am ⊂ Nk , m ∈ N}.
Theorem 2. Let ϕ be a homomorphism from Gk to a finite commutative group. Then there
exists an element H of ℜ such that Kerϕ ≃ H and conversely.
Proof. We give a one to one correspondence between {Kerϕ| ϕ(Gk ) ≃ G ∈ {K2i | i ∈ N}}
and ℜ. Let ϕ be a homomorphism from Gk to a finite commutative group of order p. Then by
Proposition 2 the number p belongs to the set {2i | i ∈ N} and if p = 2n then ϕ : Gk → K2n
is an isomorphism. For any nonempty sets A1 , A2 , ..., An ⊂ Nk and ai ∈ Gk , i ∈ Nk we
define the following homomorphism
4
φA1 A2 ...An (ai ) =
F. H. HAYDAROV
cj , if i ∈ Aj \ (A2 ∪ A3 ∪ ... ∪ An ), j = 1, n
........
cj1 cj2 , if i ∈ (Aj1 ∩ Aj2 ) \ (A1 ∪ ...Aj1 −1 ∪ Aj1 +1 ∪ ...
... ∪ Aj2 −1 ∪ Aj2 +1 ∪ ... ∪ An ), 1 ≤ j1 < j2 ≤ n
........
c1 ...cj−1 cj+1...cn , if i ∈ (A1 ∩ ... ∩ Aj−1 ∩ Aj+1 ... ∩ An ) \ Aj , j = 1, n
c1 c2 ...cn , if i ∈ (A1 ∩ A2 ∩ ... ∩ An )
e, if i ∈ Nk \ (A1 ∪ A2 ∪ ... ∪ An ).
If i ∈ ∅ then we’ll accept that there is no index i ∈ Nk satisfying the required condition.
For x ∈ Gk if ϕA1 A2 ...An (x) = e the number of ai , i = 1, n of ai appearing in the word x
must be even. Therefore
KerϕA1 A2 ...An = HA1 ∩ HA2 ∩ ... ∩ HAn .
Thus the following equality holds (up to isomorphism)
{Kerϕ| ϕ(Gk ) ≃ G ∈ {K2i | i ∈ N}} = ℜ.
The group G has finitely generators of the order two. Assume that r is the minimal
number of such generators of the group G and without loss of generality we can take these
generators to be b1 , b2 , ...br . Let e1 be the identity element of the group G. We define a
homomorphism from Gk onto G. Let Ξn = {A1 , A2 , ..., An } be a partition of Nk \A0 , 0 ≤
|A0 | ≤ k +1−n. Then we consider the homomorphism un : {a1 , a2 , ..., ak+1} → {e1 , b1 ..., bm }
given by
(
e1 , if x = ai , i ∈ A0
un (x) =
(2.1)
bj , if x = ai , i ∈ Aj , j = 1, n.
For b ∈ G we denote by Rb [b1 , b2 , ..., bm ] a representation of the word b by generators
b1 , b2 , ..., br , r ≤ m. Define the homomorphism γn : G → G by the formula
e , if x = e1
1
(2.2)
γn (x) =
bi , if x = bi , i = 1, r
Rbi [b1 , ..., br ], if x = bi , i ∈
/ {1, ..., r}.
Put
(p)
HΞn (G) = {x ∈ Gk | l(γn (un (x))) : 2p}, 2 ≤ n ≤ k − 1.
(2.3)
Let γn (un (x))) = x̃. We introduce the following equivalence relation on the set Gk :
x ∼ y if x̃ = ỹ. It’s easy to check this relation is reflexive, symmetric and transitive.
SUBGROUPS AND NORMAL SUBGROUPS
5
Let (G, ∗) be a group of order n, generated by two elements of order 2, and let ℑn be
the set of all such groups.
Proposition 3. For the group Gk the following equality holds
{Kerϕ| ϕ : Gk → G ∈ ℑ2n is an epimorphism} =
(n)
= {HB0 B1 B2 (G)| B1 , B2 is a partition of the set Nk \ B0 , 0 ≤ |B0 | ≤ k − 1}.
Proof. For an arbitrary group G ∈ ℑ2n we give a one to one correspondence between the
two given sets. Let e1 be the identity element of G and B1 , B2 be a partition of Nk \ B0 ,
where B0 ⊂ Nk , 0 ≤ |B0 | ≤ k − 1. Then we define the homomorphism ϕB0 B1 B2 : Gk → G
by the formula
(
b1 , if i ∈ B1
ϕB0 B1 B2 (ai ) =
(2.4)
b2 , if i ∈ B2 .
There is only one such a homomorphism (corresponding to B0 , B1 , B2 ). It’s clearly x ∈
KerϕB0 B1 B2 if and only if x̃ is equal to e1 . Hence it is sufficient to show that if y ∈
(n)
HB0 B1 B2 (G) then ỹ = e1 . Suppose that there exist y ∈ Gk such that l(ỹ) ≥ 2n. Let
ỹ = bi1 bi2 ...biq , q ≥ 2n and S = {bi1 , bi1 bi2 , ..., bi1 bi2 ...biq }. Since S ⊆ G there exist x1 , x2 ∈
S such that x1 = x2 which contradicts the fact that ỹ is a non-contractible. Thus we
(n)
have proved that l(ỹ) < 2n. Since y ∈ HB0 B1 B2 (G) the number l(ỹ) must be divisible
(n)
by 2n. Thus for any y ∈ HB0 B1 B2 (G) we have ỹ = e1 . Hence for the group G we get
(n)
KerϕB0 B1 B2 = HB0 B1 B2 (G).
Let us denote by ℵn the following set
(n)
{HB0 B1 B2 (G)| B1 , B2 is a partition of the set Nk \ B0 , 0 ≤ |B0 | ≤ k − 1, |G| = 2n}.
Theorem 3. Any normal subgroup of index 2n with n = 2i (2s + 1), i = {0, 1}, s ∈ N has
(n)
the form HB0 B1 B2 (G), |G| = 2n i.e.,
ℵn = {H| H ⊳ Gk , |Gk : H| = 2n}
Proof. We will first prove that ℵn ⊆ {H| H ⊳ Gk , |Gk : H| = 2n}. Let G be a group with
|G| = 2n and B1 , B2 a partition of the set Nk \B0 , 0 ≤ |B0 | ≤ k−1. For x = ai1 ai2 ...ain ∈ Gk
(n)
(n)
it’s sufficient to show that x−1 HB0 B1 B2 (G) x ⊆ HB0 B1 B2 (G). We have (as in the proof of
(n)
Proposition 3) if y ∈ HB0 B1 B2 (G) then ỹ = e1 , where e1 is the identity element of G.
(n)
Now we take any element z from the set x−1 HB0 B1 B2 (G) x. Then z = x−1 h x for some
(n)
h ∈ HB0 B1 B2 (G). We have
= γn
z̃ = γn (vn (z)) = γn vn (x−1 h x) = γn vn (x−1 )vn (h)vn (x) =
vn (x−1 ) γn (vn (h)) γn (vn (x)) = (γn (vn (x)))−1 γn (vn (h)) γn (vn (x)) .
6
F. H. HAYDAROV
(n)
From γn (vn (h)) = e1 we get z̃ = e1 i.e., z ∈ HB0 B1 B2 (G). Hence
(n)
HB0 B1 B2 (G) ∈ {H| H ⊳ Gk , |Gk : H| = 2n}.
Now we’ll prove that {H| H ⊳ Gk , |Gk : H| = 2n} ⊆ ℵn . Let H ⊳ Gk , |Gk : H| = 2n.
We consider the natural homomorphism φ : Gk → Gk : H i.e. the homomorphism given by
φ(x) = xH, x ∈ Gk . There exist e, b1 , b2 , ...b2n−1 such that φ : Gk → {H, b1 H, ..., b2n−1 H} is
an epimorphism. Let ({H, b1 H, ..., b2n−1 H}, ∗) = ℘, i.e., ℘ is the factor group. If we show
that ℘ ∈ ℑ2n then the theorem will be proved. Assuming that ℘ ∈
/ ℑ2n then there are
c1 , c2 , ..., cq ∈ ℘, q ≥ 3 such that ℘ =< c1 , ..., cq > . Clearly < c1 , c2 > is a subgroup of ℘
and | < c1 , c2 > | is greater than three.
Case n=2s+1. By Lagrange’s theorem | < c1 , c2 > | ∈ m| 2n
∈ N . If e2 is the
m
2
identity element of ℘ then from c1 = e2 we take | < c1 , c2 > | = 2n. Hence < c1 , c2 >= ℘
but c3 ∈<
/ c1 , c2 > . Thus ℘ ∈ ℑ2n .
Case n=2(2s+1). From Lagrange’s theorem | < c1 , c2 > | ∈ {4, 2(2s + 1), 4(2s + 1)} .
Let | < c1 , c2 > | = 4. If the number four isn’t equal to one of these numbers | < c1 , c3 > | or
| < c2 , c3 > | then we’ll choose that pairs. If | < c1 , c2 > | = | < c1 , c3 > | = | < c2 , c3 > | = 4
then < c1 , c2 , c3 > is a group of order eight. Again by Lagrange’s theorem |℘| = 2n must
be divisible by 8 which is impossible.
Let | < c1 , c2 > | = 2(2s + 1). Then
< c1 , c2 >= {e, c1 , c2 , c1 c2 , c1 c2 c1 , . . . , c1 c2 ...c1 } = A.
| {z }
4s
It’s easy to check that
c3 A ∪ A ⊆ ℘,
c3 A ∩ A = ∅,
|c3 A ∪ A| = |c3 A| + |A| = 2n = |℘|.
We then deduce that c3 A ∪ A = ℘ but we’ll show c3 c1 c3 ∈ ℘ does not belong to
c3 A ∪ A. Since c1 , c2 , c3 are generators we have c3 c1 c3 ∈<
/ c1 , c2 > so c3 c1 c3 = xc3 with
x ∈< c1 , c2 >. But x = c3 c1 ∈<
/ c1 , c2 > .
If | < c1 , c2 > | = 4(2l + 1) then < c1 , c2 >= ℘ but c3 ∈<
/ c1 , c2 > . Hence ℘ ∈ ℑ2n .
This completes the proof.
2.2. Subgroups of index three. In this section we will characterize the subgroups of
index three for the group representation of the Cayley tree.
Let (A1 , A2 ) be a partition of the set Nk \A0 , 0 ≤ |A0 | ≤ k − 1. Put mj be a minimal
element of Aj , j ∈ {1, 2}. Then we consider the function uA1 A2 : {a1 , a2 , ..., ak+1 } →
{e, a1 ..., ak+1 } given by
(
e, if x = ai , i ∈ Nk \ (A1 ∪ A2 )
uA1 A2 (x) =
amj , if x = ai , i ∈ Aj , j = 1, 2.
SUBGROUPS AND NORMAL SUBGROUPS
7
Define γ :< e, am1 , am2 > → {e, am1 , am2 } by the formula
e if x = e
am1 if x ∈ {am1 , am2 am1 }
am2 if x ∈ {am2 , am1 am2 }
γ(x) =
γ
a
a
a
...γ(a
a
)
if x = ami am3−i ami ...am3−i , l(x) ≥ 3, i = 1, 2
m
m
m
m
m
i
3−i
i
i
3−i
γ ami am3−i ami ...γ(am3−i ami ) if x = ami am3−i ami ...ami , l(x) ≥ 3, i = 1, 2
For a partition (A1 , A2 ) of the set Nk \A0 , 0 ≤ |A0 | ≤ k − 1 we consider the following set.
ΣA1 A2 (Gk ) = {x ∈ Gk | γ(uA1A2 (x)) = e}
Lemma 1. Let (A1 , A2 ) be a partition of the set Nk \A0 , 0 ≤ |A0 | ≤ k − 1. Then x ∈
P
A1 A2 (Gk ) if and only if the number l(uA1 A2 (x)) is divisible by 3.
Proof. Let x = ai1 ai2 ...ain ∈ Gk and l(uA1 A2 (x)) be odd (the even case is similar). Then we
can write uA1 A2 (x) = ami am3−i ...ami am3−i ami . γ(uA1 A2 (x)) is equal to
γ(ami am3−i ...ami am3−i ami ) = γ ami am3−i ...ami γ(am3−i ami ) = γ(ami am3−i ...am3−i ami ) = ...
|
{z
}
l(uA1 A2 (x))−3
We continue this process until the length of the word γ(uA1A2 (x)) will be less than three.
From γ(uA1 A2 (x)) = e we deduce that l(uA1 A2 (x)) is divisible by 3.
Conversely if l(uA1 A2 (x)) is divisible by 3 then x is generated by the elements
am1 am2 am1 and am2 am1 am2 . Since γ(am1 am2 am1 ) = γ(am2 am1 am2 ) = e we get x ∈
P
A1 A2 (Gk ).
Proposition 4. For the group Gk the following equality holds
{K| K is a subgroup of Gk of index 3} =
= {ΣA1 A2 (Gk )| A1 , A2 is a partition of Nk \ A0 }.
Proof. Let K be a subgroup of the group Gk with |Gk : K| = 3. Then there exist p, q ∈ Nk
such that |{K, ap K, aq K}| = 3. Put
A0 = {i ∈ Nk | ai ∈ K},
A1 = {i ∈ Nk | ai ap ∈ K},
A2 = {i ∈ Nk | ai aq ∈ K}
From |Gk : K| = 3 we conclude that {A1 , A2 }P
is a partition of Nk \ A0 . Let mi be a
minimal element of Ai , i = 1, 2. If we show
A1 A2 (Gk ) is a subgroup of Gk (corresponding to K) then
it’ll
be
given
one
to
one
correspondence
between given sets. For
P
P
x = ai1 ai2 ...ain ∈ A1 A2 (Gk ), y = aj1 aj2 ...ajm ∈ A1 A2 (Gk ) it is sufficient to show that
P
xy −1 ∈
A1 A2 (Gk ). Let uA1 A2 (x) = ami am3−i ....ams , uA1 A2 (y) = amj am3−j ....amt , where
8
F. H. HAYDAROV
(i, j, s, t) ∈ {1, 2}4. Since uA1 A2 is a homomorphism we have
(
|l (uA1 A2 (x)) − l(uA1 A2 (y −1)) |, if s = t
−1
l uA1 A2 (xy ) =
l (uA1 A2 (x)) + l(uA1 A2 (y −1)) , if s 6= t.
By Lemma
1 we see that l (uA1 A2 (x)) is divisible by 3 and l (uA1 A2 (y −1 )) =
l (uA1 A2 (y))−1
is also divisible by 3, so l (uA1 A2 (xy −1 )) is a multiple of 3, which shows
P
that xy −1 ∈ A1 A2 (Gk ). This completes the proof.
References
[1] Cohen,D.E, Lyndon,R.C., (1963), Free bases for normal subgroups of free groups,
Trans.Amer.Math.Soc., 108, pp. 526-537.
[2] D.S, Malik, J.N, Mordeson, M.K, Sen., (1997), Fundamentals of Abstract Algebra, McGraw-Hill Com.
[3] Ganikhodjaev, N.N, Rozikov, U.A., (1997), Description of periodic extreme Gibbs measures of some
lattice model on the Cayley tree, Theor.Math.Phys. 111, pp. 480-486.
[4] U.A, Rozikov, F.H, Haydarov., (2014), Normal subgroups of finite index for the group represantation
of the Cayley tree, TWMS Jour.Pure.Appl.Math. 5, pp. 234-240.
[5] U.A, Rozikov., (2013) Gibbs measures on a Cayley tree, World Sci. Pub, Singapore.
[6] Kurosh, A.G., (1953) Group theory, Akademic Verlag, Berlin.
[7] Young, J.W., (1927), On the partitions of a group and the resulting classification, Bull.Amer.Math.Soc.,
33, pp.453-461.
F.H.Haydarov, National University of Uzbekistan, Tashkent, Uzbekistan.
E-mail address: haydarov [email protected].
| 4math.GR
|
Consensus analysis of large-scale nonlinear homogeneous multi-agent
formations with polynomial dynamics
arXiv:1612.01375v1 [cs.SY] 5 Dec 2016
Paolo Massioni, Gérard Scorletti
Abstract—Drawing inspiration from the theory of linear “decomposable systems”, we provide a method, based on linear
matrix inequalities (LMIs), which makes it possible to prove the
convergence (or consensus) of a set of interacting agents with
polynomial dynamic. We also show that the use of a generalised
version of the famous Kalman-Yakubovic-Popov lemma allows
the development of an LMI test whose size does not depend on
the number of agents. The method is validated experimentally
on two academic examples.
Index Terms—Multi-agent systems, nonlinear systems, consensus, polynomial dynamic, sum of squares.
I. I NTRODUCTION
Large-scale systems are an emerging topic in the system
and control community, which is devoting significant efforts
on the development of analysis and control synthesis methods
for them. This deep interest can clearly be seen from the large
number of works published in the field in the last 40 years
[1], [2], [3], [23], [8], [12], [18], [20], [14], [17].
One of the main objectives of the studies is the development
and validation of “distributed control laws” for obtaining a
certain specified goal for a system of this kind. By “distributed
control”, opposed to “centralized control”, we mean a control
action that is computed locally according to the physical
spatial extension of the system, which is seen as an interconnection of simpler subsystems. One of the main problems of
large-scale systems is the “curse of dimensionality” that goes
with them, i.e. the analysis and synthesis problems related to
dynamical systems grow with the size, and for system of very
high order, such problems becomes computationally infeasible.
In the literature, if we restrict to linear systems, we can find a
few solutions [18], [8], [1] that can effectively overcome the
curse of dimensionality for a class of systems with a certain
regularity, namely for what we call “homogeneous systems”
which are made of the interconnection of a huge number of
identical subunits (also sometimes called “agents”).
In this paper we focus on this same problem, more specifically on the stability analysis of large-scale dynamical systems,
for the nonlinear case. Namely, we will consider a formation or
a multi-agent system made of a high number of identical units
interacting with one another through a symmetric graph-like
interconnection. Assuming that the dynamical equation of each
subunit is described by a polynomial in the state vector, we can
show that a linear matrix inequality (LMI) test can be devised
in order to verify the relative stability of such a formation. We
P. Massioni is with Laboratoire Ampère, UMR CNRS 5005,
INSA de Lyon, Université de Lyon, F-69621 Villeurbanne, France
[email protected]
G. Scorletti is with Laboratoire Ampère, UMR CNRS 5005,
Ecole Centrale de Lyon, Université de Lyon, F-69134 Ecully, France
[email protected]
will also be able to formulate such a test in a form which is not
strictly depending on the formation size, making it possible to
check the stability of formations with virtually any number of
agents, basically extending the analysis results of [18], [9] to
the nonlinear (polynomial) case.
II. P RELIMINARIES
A. Notation
We denote by N the set of natural numbers, by R the set
of real numbers and by Rn×m the set of real n × m matrices.
A⊤ indicates the transpose of a matrix A, In is the identity
matrix of size n, 0n×m is a matrix of zeros of size n × m
and 1n ∈ Rn a column vector that contains 1 in all of its
entries. The notation A 0 (resp. A 0) indicates that all
the eigenvalues of the symmetric matrix A are positive (resp.
negative) or equal to zero, whereas A ≻ 0 (resp. A ≺ 0)
indicates that all such eigenvalues
are strictly positive (resp.
n
negative). The symbol
indicates the binomial coefficient,
k
for which we have
n!
n
=
.
k
k!(n − k)!
The symbol ⊗ indicates the Kronecker product, for which we
remind the basic properties (A ⊗ B)(C ⊗ D) = (AC ⊗ BD),
(A ⊗ B)⊤ = (A⊤ ⊗ B ⊤ ), (A ⊗ B)−1 = (A−1 ⊗ B −1 ) (with
matrices of compatible sizes). We employ the symbol ∗ to
complete symmetric matrix expressions avoiding repetitions.
B. Agent dynamic
We consider a set of N ∈ N identical agents or subsystems
of order n, which interact with one another. Each agent, if
taken alone, is supposed to be described by a polynomial
dynamic, of the kind
ẋi = fd (xi ) = Aa χi
(1)
⊤
n
where i = 1, ..., N , xi = [xi,1 , xi,2 , ..., xi,n ] ∈ R is the
state of the ith agent, fd is a polynomial function of degree
d ∈ N, Aa ∈ Rn×ρ and χi ∈ Rρ is the vector containing all the
monomials in xi up to degree d (for example, if n = 2, d = 2,
then χi = [1, xi,1 , xi,2 , x2i,1 , xi,1 xi,2 , x2i,2 ]⊤ ). The value of ρ
is given by
n+d
ρ=
.
(2)
n
This approach is based on the sum of squares (SOS) literature
[19], which basically allows the relaxation of polynomial
problems into linear algebra’s. In this context, it is possible to
express polynomials p up to degree 2d as quadratic forms with
⊤
respect to χi , i.e. p(xi ) = χ⊤
∈ Rρ×ρ .
i X χi , with X = X
2
This quadratic expression is not unique, due to the fact that
different products of monomials in χi can yield the same
result, for example x2i,1 is either x2i,1 times 1 or xi,1 times
xi,1 . This implies that there exist linearly independent slack
ρ×ρ
matrices Qk = Q⊤
, with k = 1, . . . , ι such that
k ∈ R
⊤
χi Qk χi = 0. The number of such matrices is [19]
2
!
1
d+n
d+n
n + 2d
ι=
+
−
.
(3)
2
d
d
2d
C. Formations
Moving from one single agent to a whole formation, we
employ a matrix P ∈ RN ×N to describe the interactions
among the agents. Basically P is a sparse matrix whose entries
in the ith row and j th column indicate whether the ith agent is
influenced by the state of the j th , according to the definition
that follows.
Definition 1 (Formation). We call a formation (of non-linear
agents with polynomial dynamics) a dynamical system of order
nN , with n, N ∈ N, described by the following dynamical
equation
ẋ = (IN ⊗ Aa + P ⊗ Ab )χ
(4)
⊤
⊤ ⊤
where x = [x⊤
∈
RnN , χ
=
1 , x2 , ..., xN ]
⊤
⊤
⊤ ⊤
ρN
N ×N
[χ1 , χ2 , ..., χN ] ∈ R , P ∈ R
and Aa , Ab ∈ Rn×ρ .
This definition extends and adapts the definition of “decomposable systems” given in [18] to polynomial dynamics.
In the linear case, a formation defined above boils down to
the dynamical equation
ẋ = (IN ⊗ Aa + P ⊗ Ab )x.
(5)
In [18] it has been shown that if P is diagonalisable, then
this system (of order nN ) is equivalent to a set of parameterdependent linear systems of order n. This is obtained with
the change of variables x = (S ⊗ In )x̂, where x̂ =
⊤
⊤ ⊤
nN
[x̂⊤
and S is the matrix diagonalising
1 , x̂2 , ..., x̂N ] ∈ R
−1
P, i.e. S PS = Λ, with Λ diagonal. This turns (5) into
x̂˙ = (IN ⊗ Aa + Λ ⊗ Ab )x̂, which is a block-diagonal system
equivalent to the set
x̂˙ i = (Aa + λi Ab )x̂i
(6)
for i = 1, ..., N , with λi the ith eigenvalue of P. This idea
of decomposing a distributed system into a set of parametervarying systems is very practical and it has inspired several
works in the domain of consensus and distributed control [13],
[26], [7], [4]. In this paper, to our knowledge, the idea is
adapted to nonlinear dynamics for the first time.
D. Problem formulation
The topic of this paper is to find a proof of convergence of
the state of the agents under a given dynamics expressed as in
(4). We do not require that each agent by itself converges to a
point, but that they all converge eventually to the same state,
which could be either an equilibrium point or a trajectory. In
order to do so, we formulate first an assumption on the pattern
matrix P.
Assumption 2. The pattern matrix P ∈ RN ⊗N in (4) is
symmetric and it has one and only one eigenvalue equal to
0, associated to the eigenvector 1N , i.e P1N = 0.
This assumption is very common in the literature, it basically ensures that the interconnection matrix is a (generalised)
graph Laplacian of a symmetric connected graph [5]. Such
matrices have real eigenvalues and eigenvectors. We can then
formulate the problem on which this paper focuses.
Problem 3. We consider (4) with initial conditions x(0) ∈
RnN . Prove that limt→∞ ||xi − xj || = 0, ∀i, j ∈ {1, ..., N }.
III. F ORMATION LYAPUNOV FUNCTION
In order to be able to prove the convergence of all the agents
to the same trajectory, we define what we call a “formation
Lyapunov function”, which will have the property of tending
to zero when the agents are converging. We summarise these
notions in a definition and a lemma.
Definition 4 (Formation Lyapunov function candidate). We
define as “formation Lyapunov function candidate” a function
!
l
X
(7)
P i ⊗ Li x = x⊤ Lx
V (x) = x⊤
i=1
L⊤
i
n×n
∈R
, l ∈ N, l 6 N . The reason for this
with Li =
special structure will be clear later on, in fact it allows the
block-diagonalisation of the Lyapunov matrix in the same way
as P can be diagonalised.
Lemma 5. Consider (4) and a formation Lyapunov function
N ×(N −1)
candidate V (x) = x⊤ Lx as in (7). Let 1⊥
be
N ∈ R
the orthogonal complement of 1N , i.e. [1N 1⊥
]
is
full
rank
N
⊥
and 1⊤
N 1N = 0.
⊤
⊥
If (1⊥
N ⊗ In ) L (1N ⊗ In ) > 0, then we have that xi = xj
∀i, j ∈ {1, ..., N } if and only if V (x) = 0.
Proof: Necessity is almost obvious: if xi = xj ∀i, j ∈
{1, ..., N }, then x = 1N ⊗ xi ; the fact that P1N = 0 implies
that V (x) = 0.
We prove the sufficiency by contradiction, i.e. we suppose
that there exist i and j for which xi 6= xj and V (x) = 0.
The vector x with the complete state must then have at
least one component which is orthogonal to the columns of
(1N ⊗ In ), because (1N ⊗ In ) contains columns with all the
corresponding agent states equal. So, based on the fact that
⊤
⊥
(1⊥
N ⊗ In ) L (1N ⊗ In ) > 0, then V (x) > 0 contradicting
the hypothesis.
IV. M AIN
RESULT
We are now ready to formulate our main result. A preliminary lemma is given first, which allows diagonalising the
Lyapunov function in the same way as a linear system is
decomposed in [18].
Lemma 6. If Assumption 2 holds, then 1) there exist a matrix
S ∈ RN ×N such that S ⊤ S = SS ⊤ = IN , and S ⊤ PS = Λ,
with Λ diagonal. Moreover, we have that 2) S ⊤ 1N = T =
[t1 t2 ... tN ], with ti ∈ RN = 0 if λi = Λi,i 6= 0.
3
Proof: The first part of the lemma is proven by the
fact that all symmetric matrices are diagonalisable by an
orthonormal matrix S (i.e. S −1 = S ⊤ ) [10]. For the second
part, consider that due to Assumption 2, 1N is an eigenvector
of P with eigenvalue 0; the matrix S contains the normalised
eigenvectors of P in its columns, and all of these eigenvectors
are orthogonal to one another because S ⊤ S = IN . So each
ti is the dot product between 1N and the ith eigenvector, and
it is non zero if and only if λi = 0.
Theorem 7. Consider (4) with given N , Aa , Ab and P
statisfying Assumption 2; moreover, we order the eigenvalues
of P so that the first eigenvalue is the one equal to zero, i.e.
λ1 = 0. If for a chosen l ∈ N, there exist τj ∈ R, and matrices
n×n
Lj = L⊤
such that
j ∈R
Pl
j
(8)
j=1 λi Lj ≻ 0
Pl
Pι
⊤
Π( j=1 τj Qj + j=1 (λji (Γ⊤ Lj Aa + A⊤
a Lj Γ + Γ Lj Γ)+
j+1
⊤
⊤
⊤
λi (Γ Lj Ab + Ab Lj Γ))Π 0
(9)
for i = 2, ... N , where Γ = [0n,1 In 0n,ρ−n−2 ] (i.e. Γxi =
χi ), Π = [0ρ−1,1 Iρ−1 ], then limt→∞ ||xi − xj || = 0, ∀i, j ∈
{1, ..., N }.
Proof: In order to assure the convergence of the agents,
we need to assure the conditions stated in Lemma 5, namely
that a function V (x) = x⊤ Lx exists, with and (1⊥
N ⊗
In )⊤ L (1⊥
⊗
I
)
>
0,
and
that
V̇
(x)
<
0
for
V
(x)
>
0.
n
N
⊤
⊥
For the condition (1⊥
⊗
I
)
L
(1
⊗
I
)
>
0,
consider
that
n
n
N
N
S contains a scaled version of 1N in its first column and 1⊥
N in
P
the rest of the matrix. Knowing that L = (S ⊗ In )( li=1 Λi ⊗
Li )(S ⊤ ⊗ In ) thanks to Lemma 6, this condition is equivalent
to (8).
For what concerns the condition V̇ (x) < 0 for V (x) > 0,
it is satisfied if V̇ (x) 6 −ǫx⊤ Lx, i.e.
⊤
χ⊤ (QN + Γ⊤
N L(IN ⊗ Aa ) + ΓN L(P ⊗ Ab )+
⊤
(IN ⊗ A⊤
)LΓ
+
(P
⊗
A
)LΓ
+
ǫΓ⊤
N
N
a
b
N LΓN )χ 6 0
(10)
where
Γ
=
(I
⊗
Γ)
(so
Γ
χ
=
x)
and
Q
=
I
⊗
N
N
N
N
N
Pι
⊤
τ
Q
(for
which,
by
definition
of
Q
,
χ
Q
χ
=
0
j
j
j
N
j=1
for all values of the τi ). By the fact that P = SΛS ⊤ and
IN = SS ⊤ (Assumption 2 and Lemma 6), using the properties
of Kronecker product (10) is equivalent to
⊤
χ̂⊤ (QN + Γ⊤
N L̂(IN ⊗ Aa ) + ΓN L̂(Λ ⊗ Ab )+
⊤
⊤
(IN ⊗ Aa )L̂ΓN + (Λ ⊗ Ab )L̂ΓN + ǫΓ⊤
N L̂ΓN )χ̂ 6 0
(11)
Pl
j
⊤
with L̂ =
j=1 Λ ⊗ Lj and χ̂ = (S ⊗ Iρ )χ. Notice in
this last inequality that the term between χ̂⊤ and χ̂ is blockdiagonal, as it is the sum of terms of the kind IN ⊗ X
or Λi ⊗ X (i ∈ N). If we define χ̂i ∈ Rρ such that
⊤
⊤ ⊤
χ̂ = [χ̂⊤
1 , χ̂2 , ..., χ̂N ] , then (11) is equivalent to
PN
Pι
Pl
j
⊤
⊤
⊤
i=1 χ̂i (
j=1 τj Qj +
j=1 λi (Γ Lj Aa + Aa Lj Γ)+
Pl
Pl
j
j+1 ⊤
⊤
⊤
j=1 λi (Γ Lj Γ))χ̂i 6 0.
j=1 λi (Γ Lj Ab + Ab Lj Γ)+ǫ
(12)
The term of the sum for i = 1 is always 0 (as we chose
λ1 = 0), so there is no contribution from it and it can be
discarded. Concerning the vectors χi , remember that they all
⊤
ρ−1
contain 1 in their first entry, i.e. χi = [1 χ̃⊤
.
i ] , χ̃i ∈ R
th
For each χ̂i , the first entry is by its definition the i entry of
the vector e = S ⊤ 1, which contains zeros in all of its entries
but the first (due to Lemma 6). So for i = 2, ..., N , we have
that χ̂i = Π⊤ Πχ̂i . So (12) is equivalent to
PN
Pι
⊤ ⊤
i=2 χ̂i Π Π(
j=1 τj Qj +
Pl
j
⊤
λ
(Γ
L
A
+
A⊤
j a
a Lj Γ)+
i
Pl j=1 j+1
(13)
⊤
(Γ Lj Ab + A⊤
b Lj Γ)+
j=1 λi
P
ǫ lj=1 λji (Γ⊤ Lj Γ))Π⊤ Πχ̂i 6 0
The set of LMIs in (9) imply (13), which concludes the proof.
This theorem allows proving the convergence of N agents
with two sets of N − 1 parameter-dependent LMIs, whose
matrix size is respectively n (i.e. the order of each agent taken
alone) and ρ − 1. This result is already interesting as it avoids
using LMIs scaling with N n, which is the global system order.
In the next section, we explore whether it is possible to further
reduce the computational complexity.
V. VARIATION ON THE MAIN RESULT
We explore the possibility of using a generalised version
of the famous Kalman-Yakubovich-Popov (KYP) lemma [22].
This lemma allows turning a parameter-depending LMI into a
parameter-independent one.
A. The Kalman-Yakubovic-Popov lemma
The Kalman-Yakubovic-Popov lemma or KYP [22] is
a widely celebrated result for dynamical systems that allows turning frequency-dependent inequalities into frequencyindependent ones, by exploiting a state-space formulation. It
turns out that such a result can be adapted and generalised to
inequalities depending on any scalar parameter. Namely, we
will use the following generalised version of the KYP.
Lemma 8 (Generalized KYP [6]). Consider
η
X
ξi Mi ,
M (ξ) = M0 +
(14)
i=1
with ξ ∈ Rl a vector of decision variables and Mi = Mi⊤ ∈
RnM ×nM , i = 1, ..., η. The quadratic constraint
⊤
φ(θ) M (ξ) φ(θ) ≺ 0 for θ ∈ [θ, θ]
(15)
⊤
is verified if and only if there exist D = D ≻ 0 and G = −G ⊤
such that
⊤
C̃
M (ξ) C̃ D̃ +
⊤
D̃
⊤
I 0
−2D
(θ + θ)D + G I 0
≺0
à B̃
à B̃
(θ + θ)D − G
−2θθD
(16)
with à ,B̃, C̃ and D̃ such that
à B̃
φ(θ) = D̃ + C̃θI(I − ÃθI)−1 B̃ = θI ⋆
, (17)
C̃ D̃
where the operator ⋆ implicitly defined above is known as the
Redheffer product [27].
The lemma applies as well if the sign ≺ in (15) is replaced
by : in this case replace ≺ with in (16) as well.
4
B. Second main result
Let us define
λ = max {λi }, λ = min {λi }.
26i6N
(18)
26i6N
Then, for θ ∈ [λ, λ], the following set of LMIs
Pl
j
(19)
j=1 θ Lj ≻ 0
Pl
Pι
⊤
Π( j=1 τj Qj + j=1 (θj (Γ⊤ Lj Aa + A⊤
a Lj Γ + Γ Lj Γ)+
j+1
⊤
⊤
⊤
θ (Γ Lj Ab + Ab Lj Γ))Π 0
(20)
“embeds” the set of LMIs in (8) and (9) (notice that we have
moved from a discrete set of values to a continuous interval
which includes them all). Subsequently, Lemma 8 can be used
to turn the θ-dependent LMIs in (19) and (20) into parameterindependent ones. The dependence of the terms in (19) and
(20) from θ (which is ultimately λi ) is polynomial, so we need
to define
i⊤
h
(21)
φν (θ) = θceil((l+1)/2) Iν , θceil((l+1)/2)−1 Iν , ..., Iν
which requires
Ãν = Uν ⊗ In ,
C̃ν =
Iν
01×ν
⊗ In ,
B̃ν =
D̃ν =
0(ν−1)×1
1
0ν×1
1
⊗ In ,
⊗ In .
Theorem 9. Consider (4) with given N , Aa , Ab and P statisfying Assumption 2; excluding the first eigenvalue of P, which
is equal to 0, we have that λ 6 λi 6 λ, with i = 2, ... N .
If for a chosen l ∈ N, there exist τi ∈ R, and matrices
n×n
Li = L⊤
, and there exist Dνk , Gνk ∈ Rνk ×νk ,
i ∈ R
⊤
Dνk = Dνk ≻ 0 and Gνk = −Gν⊤k such that
[∗]⊤ Mk C̃νk D̃νk +
Iνk
0
(λ + λ)Dνk + Gνk
⊤ −2Dνk
≺0
[∗]
Ãνk B̃νk
∗
−2λλDνk
(23)
for k = 1, 2, with ν1 = n and ν2 = ρ − 1, where Γ =
[0n,1 In 0n,ρ−n−2 ] (i.e. Γxi = χi ), Π = [0ρ−1,1 Iρ−1 ], and
Ãνk , B̃νk , C̃νk , D̃νk , φνk are defined in (22) and (21), with
l
X
λji Lj
(24)
j=1
Pι
φν (λi )⊤ M2 φν (λi ) = Π( j=1 τj Qj +
Pl 2 j ⊤ 2
⊤
⊤
j=1 (λi (Γ Lj Aa + Aa Lj Γ + Γ Lj Γ)+
j+1
⊤
λi (Γ⊤ Lj Ab + A⊤
b Lj Γ))Π ,
VI. E XAMPLES
In order to provide a few challenging examples of application of the proposed method, we focus on a problem that is
widely studied in the nonlinear dynamics community, namely
the synchronisation of oscillators [21], [25]. The approach here
is of course numerical and different (or complementary) with
respect to the ones found in such a literature, where the objective is usually to find a control law and then analytically prove
stability. Our approach is just to propose a control law and test
numerically whether it will make the subsystems converge or
not. We consider two famous examples of nonlinear systems,
namely the Van der Pol oscillator [11] and the Lorenz attractor
[16].
A. Van der Pol oscillator
(22)
where Uν ∈ Riν×ν is a matrix containing 1’s in the first upper
diagonal and 0’s elsewhere, and ν = n for (19) and ν = ρ − 1
for (20). We are now ready to formulate our second main
result.
φν1 (λi )⊤ M1 φν1 (λi ) = −
With this second theorem, we replace the two sets of N − 1
LMIs of size n and ρ − 1, with only two LMIs of matrix size
n ceil((l + 3)/2) and ρ ceil((l + 3)/2). This is an interesting
result because the computational complexity is no longer
depending on N , i.e. the number of agents. On the other hand,
the choice of a bigger l will improve the chances of solving
the LMIs for high values of N .
(25)
then limt→∞ ||xi − xj || = 0, ∀i, j ∈ {1, ..., N }.
Proof: A direct application of Lemma 8 for M1 and M2
implies that the hypotheses of Theorem 7 are satisfied if the
hypotheses here are.
We consider a system of N agents of equation
ẋi = yi
ẏi = µ(1 − xi )yi − xi − cri
(26)
for which n = 2, with ri = −(xi−1 + yi−1 ) + 2(xi + yi ) −
(xi+1 + yi+1 ) (where the index is to be considered as modulo
N , i.e. 0 → N , N + 1 → 1). The interconnection between
each oscillator is given by the term c, a proportional feedback
gain. This feedback law has just been guessed, and we use
Theorem 9 to prove whether it works. We have coded the
related LMI problem in Matlab using Yalmip [15], choosing
µ = 0.5, N = 10, l = 6 and c = 15 (all arbitrary values). By
using SeDuMi [24] as solver, we managed to find a feasible
solution, which yields a valid formation Lyapunov function.
Figure 1 and Figure 2 show the evolution of the system during
a simulation, with the individual states shown as well as the
value of the Lyapunov function over time.
B. Lorenz attractor
We consider now system of N agents of equation
ẋi = σl (yi − xi ) − crx,i
ẏi = xi (ρl − zi ) − yi − cry,i
żi = xi yi − βl zi − crz,i
(27)
with ri,• = − •i−1 +2 •i −•i+1 (again the index is taken
modulo N ). We set arbitarily ρl = 28, σl = 10, βl = 8/3,
N = 8, l = 6 and c = 50. This time we used Theorem 7, successfully obtaining a formation Lyapunov function. Figure 3,
Figure 4 and Figure 5 again show the evolution of the system
during a simulation, with individual states and the value of the
Lyapunov function. Notice that the Lorenz oscillator does not
converge to a limit cycle but to a chaotic trajectory.
5
5
xi
4
yi
3
50
2
40
30
0
z
states
1
-1
20
-2
10
-3
starting points
0
-4
-5
0
2
4
6
8
10
-10
-20
12
40
t
20
-10
0
Fig. 1.
0
10
x
Evolution of the state of the 10 coupled Van der Pol oscillators.
20
-20
y
Fig. 4. Tridimensional visualisation of the state of some of the coupled
Lorenz systems of the example (the trajectories eventually converge to the
consensus trajectory).
10 4
10 2
10 4
10 -2
10 2
10 -4
10 0
10 -6
10 -2
V
V
10 0
10 -8
10 -10
10 -4
0
2
4
6
8
10
12
10 -6
t
10 -8
Fig. 2. Value of the formation Lyapunov function V for the coupled Van
der Pol oscillators.
10 -10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
t
50
Fig. 5. Value of the formation Lyapunov function V for the coupled Lorenz
systems.
xi
y
40
z
i
i
VII. C ONCLUSION
states
30
We have introduced a new method for proving convergence
or consensus of multi-agent system with polynomial dynamic.
This method is the generalisation of the analysis methods in
[18] and it has proven effective in test cases featuring dynamical oscillators. Further research will investigate if convex
controller synthesis results can be obtained with a similar
approach.
20
10
0
-10
-20
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
t
Fig. 3.
Evolution of the state of the 8 coupled Lorenz systems.
0.8
R EFERENCES
[1] B. Bamieh, F. Paganini, and M.A. Dahleh. Distributed control of
spatially invariant systems. IEEE Trans. Aut. Control, 47(7):1091–1107,
2002.
[2] R. D’Andrea and G.E. Dullerud. Distributed control design for spatially
interconnected systems. IEEE Trans. Aut. Control, 48(9):1478–1495,
2003.
6
[3] E.J. Davison and W. Gesing. Sequential stability and optimization of
large scale decentralized systems. Automatica, 15(3):307–324, 1979.
[4] O. Demir and J. Lunze. A decomposition approach to decentralized
and distributed control of spatially interconnected systems. In 18th
IFAC World Congress, volume 44, pages 9109–9114, Milan, Italy, 2011.
Elsevier.
[5] R. Diestel. Graph Theory. Springer, 1996.
[6] M. Dinh, G. Scorletti, V. Fromion, and E. Magarotto. Parameter dependent H∞ control by finite dimensional LMI optimization: application
to trade-off dependent control. International Journal of Robust and
Nonlinear Control, 15(9):383–406, 2005.
[7] A. Eichler, C. Hoffmann, and H. Werner. Robust control of decomposable LPV systems. Automatica, 50(12):3239–3245, 2014.
[8] J.A. Fax and R.M. Murray. Information flow and cooperative control of
vehicle formations. IEEE Trans. Aut. Control, 49(9), 2004.
[9] R. Ghadami and B. Shafai. Decomposition-based distributed control for
continuous-time multi-agent systems. IEEE Transactions on Automatic
Control, 58(1):258–264, 2013.
[10] G.H. Golub and C.F. Van Loan. Matrix Computations. John Hopkins
University Press, 3rd edition, 1996.
[11] R. Grimshaw. Nonlinear ordinary differential equations, volume 2. CRC
Press, 1991.
[12] C. Langbort, R.S. Chandra, and R. D’Andrea. Distributed control design
for systems interconnected over an arbitrary graph. IEEE Trans. Aut.
Control, 49(9):1502–1519, 2004.
[13] Z. Li, Z. Duan, and G. Chen. On H∞ and H2 performance regions of
multi-agent systems. Automatica, 47(4):797–803, 2011.
[14] Z. Li, Z. Duan, G. Chen, and L. Huang. Consensus of multiagent systems
and synchronization of complex networks: A unified viewpoint. IEEE
Trans. Circuits Syst. Regul. Pap., 57(1):213–224, 2010.
[15] J. Löfberg. Yalmip: a toolbox for modeling and optimization in
MATLAB. In Proc. of the CACSD Conference, Taipei, Taiwan, 2004.
[16] E.N. Lorenz. Deterministic nonperiodic flow. Journal of the Atmospheric
Sciences, 20(2):130–141, 1963.
[17] P. Massioni. Distributed control for alpha-heterogeneous dynamically
coupled systems. Systems & Control Letters, 72:30–35, 2014.
[18] P. Massioni and M. Verhaegen. Distributed control of vehicle formations:
a decomposition approach. In Proc. of the 47th IEEE Conference on
Decision and Control, pages 2906–2912. IEEE, 2008.
[19] P.A. Parrilo. Semidefinite programming relaxations for semialgebraic
problems. Mathematical programming, 96(2):293–320, 2003.
[20] A.P. Popov and H. Werner. A robust control approach to formation
control. In Proc. of the 10th European Control Conference, Budapest,
Hungary, August 2009.
[21] M. Pourmahmood, S. Khanmohammadi, and G. Alizadeh. Synchronization of two different uncertain chaotic systems with unknown parameters
using a robust adaptive sliding mode controller. Communications in
Nonlinear Science and Numerical Simulation, 16(7):2853 – 2868, 2011.
[22] A. Rantzer. On the Kalman-Yakubovich-Popov lemma. Systems &
Control Letters, 28(1):7–10, 1996.
[23] G. Scorletti and G. Duc. An LMI approach to decentralized H∞ control.
International Journal of Control, 74(3):211–224, 2001.
[24] J.F. Sturm. Using SeDuMi 1.02, a MATLAB toolbox for optimization
over symmetric cones. Optimization Methods and Software, 11–12:625–
653, 1999.
[25] L. Torres, G. Besançon, D. Georges, and C. Verde. Exponential nonlinear
observer for parametric identification and synchronization of chaotic
systems. Mathematics and Computers in Simulation, 82(5):836–846,
2012.
[26] M. Zakwan, Z. Binnit-e-Rauf, and M. Ali. Polynomial based fixedstructure controller for decomposable systems. In 13th International
Bhurban Conference on Applied Sciences and Technology (IBCAST),
pages 140–144. IEEE, 2016.
[27] K. Zhou, J.C. Doyle, and K. Glover. Robust and optimal control,
volume 40. Prentice Hall New Jersey, 1996.
| 3cs.SY
|
An Evolutionary Approach to Drug-Design Using a Novel
Neighbourhood Based Genetic Algorithm
Arnab Ghosh1, Avishek Ghosh1, Arkabandhu Chowdhury 1 , Amit Konar1
1
Department of Electronics and Tele-communication Engineering
Jadavpur University, Kolkata-700032, India
Email:[email protected],[email protected],
Abstract: The present work provides a new approach to evolve ligand structures which represent possible drug to
be docked to the active site of the target protein. The structure is represented as a tree where each non-empty node
represents a functional group. It is assumed that the active site configuration of the target protein is known with
position of the essential residues. In this paper the interaction energy of the ligands with the protein target is
minimized. Moreover, the size of the tree is difficult to obtain and it will be different for different active sites. To
overcome the difficulty, a variable tree size configuration is used for designing ligands. The optimization is done
using a novel Neighbourhood Based Genetic Algorithm (NBGA) which uses dynamic neighbourhood topology. To
get variable tree size, a variable-length version of the above algorithm is devised. To judge the merit of the
algorithm, it is initially applied on the well known Travelling Salesman Problem (TSP).
1. Introduction:
A strategy in drug design is to find compounds that bind to protein targets that constitute active sites which sustain
viral proliferation. The challenge is to predict accurately structures of the compounds (ligands) when the active site
configuration of the protein is known [1]. The literature addresses the challenge using a novel Genetic Algorithm
that uses ring parent topology to generate offspring. It is found that the algorithm gives better candidate solution
than traditional Genetic Algorithm many existing variation of it.
Evolutionary computation is used to place functional groups in appropriate leaves of the tree structured ligand. The
objective is to minimize the interaction energy between the target protein and the evolved ligand, thus leading to the
most stable solution. In [1] a fixed tree structure of the ligand is assumed. However it is difficult to get a prior
knowledge of the structure and for a given geometry, no unique solution is the best solution. So variable length
structure is used in the paper. Depending upon the geometry of the active site, a ligand can have a maximum or a
minimum length (denoted by
). The length of the ligand lies in between these two values.
2. Genetic Algorithm:
Genetic Algorithms (GAs) [2-6] are search algorithms based on the mechanics of the natural selection process
(biological evolution). The most basic concept is that optimization is based on evolution, and the "Survival of the
fittest" concept. GAs have the ability to create an initial population of feasible solutions, and then recombine them in
a way to guide their search to only the most promising areas of the state space. Each feasible solution is encoded as a
chromosome (string) also called a genotype, and each chromosome is given a measure of fitness via a fitness
(evaluation or objective) function. The fitness of a chromosome determines its ability to survive and produce
offspring. A finite population of chromosomes is maintained. GAs use probabilistic rules to evolve a population
from one generation to the next. The generations of the new solutions are developed by genetic recombination
operators:
Biased Reproduction: selecting the fittest to reproduce
Crossover: combining parent chromosomes to produce children chromosomes
Mutation: altering some genes in a chromosome.
Crossover combines the "fittest" chromosomes and passes superior genes to the next generation.
Mutation ensures the entire state-space will be searched, (given enough time) and can lead the population out of
a local minima.
2.1. Neighbourhood Based Approach (NBGA) :
Firstly we create a random sequence pool. Parents are selected randomly from the sequence pool and a ring parent
topology is developed (shown in figure 1, 2). Consecutive two parents in the ring go under crossover process and
two offspring are generated. After that trio selection procedure is applied (figure 3).
Pseudo code of selection procedure:
Population t {git }; i [1, max_ pop]
Mutant {mi } mutation ({git });
git select ( git , mi );
Parent { pi } rand _ select ({git })
Son {si };
(si1, si 2 ) crossover( pi , pi 1 );
g it 1 select ( pi , si1 , si 2 );
Population t 1 {g it 1};
Select is a function that selects a sequence on the basis of cost function. Sequence with minimum cost function is
selected.
g1
p1
g2
p2
g3
p3
gn-1
pn-1
gn
pn
Figure 1. Generation of parents after shuffling the population
Figure 2. Ring Parent Topology
P1
s1
p2
s2
s1
p3
s2
s1
p4
s2
s1
s2
Figure 3. Trio Selection
2.2. Some Modified Mutation Schemes:
Multiple Exchange mutation: Here we select more than two dimensions at a time and exchange their position (i.e.
randomly placed them at their place). This mutation ensures higher degree of convergence but accuracy becomes
less. So we apply this scheme at the beginning of the algorithm and gradually drop out. Generally a random integer
(ri) generator decides how many positions would be exchanged. Here we use highest number of position (hi) to be
roughly one sixth of total no of dimensions and gradually decrease this number with generation (2<ri ≤hi, hi<n/6,
n=no of cities, gen=1). After a certain number of generation we use simple Exchange mutation technique (hi=2).
Illustrative example:
(B C D A E G F H J I K L M N O P Q)
(B E D A M G F H J I C L K N O P Q)
(Offspring)
Figure 4. Multiple Exchange Mutation
ri=4
Multilevel mutation: Actually two or more number of mutations are mixed up to generate offspring. There is no
exact procedure to work out but we can apply same or another type of mutation in the offspring and thus increase the
degree of mutation. This scheme will be effective based on the problem or situation. So very small probability is
assigned and uses it after certain no of generation. We here use two multilevel mutations: exchange+displacement
mutation and simple inversion+displacement mutation.
Second type is known as inversion mutation and can be very effective at the time of convergence.
Illustrative example:
(B C A | E J I | D F H G)
|IJE|
(inverted segment)
(B | I J E | C A D F H G)
(Offspring)
Figure 5. Multilevel Mutation
2.3.Pseudo Code:
Begin Neighbourhood Based GA
Create initial population
whilegeneration_count<k do
/* k = max. number of generations. */
Begin
Mutation
Select parent randomly
Construct ring parent topology
Crossover
Apply trio selection procedure
Increment generation_count
end
Output the best individual found
end Neighbourhood Based GA
3. Travelling Salesman Problem:
The Travelling Salesman Problem (TSP) is a well-known problem in the area of network and combinatorial
optimization. Its importance stems from the fact that there is a plethora of fields in which it finds potential
applications such as DNA fragment assembly and VLSI design. Formally, the TSP may be defined as a permutation
problem with the objective of finding the path of the shortest length (or the minimum cost) on an undirected graph
that represents cities or nodes to be visited. The traveling salesman starts at one node, visits all other nodes
successively only one time each, and finally returns to the starting node. i.e., given n cities, named {c1, c2,....,cn},
and permutations, π1, ..., πn!, the objective is to choose πi such that the sum of all Euclidean distances between each
node and its successor is minimized. The successor of the last node in the permutation is the first one. The Euclidean
distance d, between any two cities with coordinate (x1, y1) and (x2, y2) is calculated by
d ( x1 x2) 2 ( y1 y 2) 2
(1)
More mathematically we may define the TSP as follows:
Given an integer n ≥ 3 and an n × n matrix C = (cij), where each cij isa nonnegative integer, which cyclic
permutation π of the integers from 1 to n minimizes the sum
n
c
i 1 i ( i )
?
The results are compared with those obtained using SWAP_GATSP [7], MMGA [8], IGA [9], OX SIM, (standard
GA with order crossover and simple inversion mutation) [10] MOC SIM (Modified order crossover and SIM), and
self organizing map (SOM) [11] for solving TSP. Table 1 summarizes the results obtained over 30 runs by running
the NBGA, SWAP_GATSP, OX SIM and MOC SIM on the aforesaid nine different TSP instances
Algo
Statistics
NBGA
SWAP_GATSP
OX_SIM
MOC_SIM
Problems
Grtschels24
24
1272
Bayg29
29
1610
Grtschels48
48
5046
eil51
51
426
St70
70
675
Best
Average
Error
Best
Average
Error
Best
Average
Error
Best
Average
Error
Best
Average
Error
1272
1272
0.0000
1610
1610
0.0000
5046
5084
0.7531
429
432
1.4084
682
684
1.3333
1272
1272
0.0000
1610
1615
0.3106
5046
5110
1.2683
439
442
3.7559
685
701
3.8519
1272
1272
0.0000
1610
1690
4.9689
5097
5410
7.2136
493
540
26.7606
823
920
36.2963
1272
1272
0.0000
1610
1622
0.7453
5057
5184
2.7348
444
453
6.3380
698
748
10.8148
Table 1: Comparison of NBGA with other Algorithms
Experimental results (Table 1) using NBGA are found to be superior in terms of quality of solution (best result,
average result and error percentage) with less number of generations when compared with those of other existing
GAs.
4. Evolving Ligand Molecules:
Protein with known active site configuration is used for evolving ligand structures. Our specific target is the known
antiviral binding site of the Human Rhinovirus strain 14. This active site is known as the VP1 barrel for its
resemblance with a barrel. The molecule which can easily be fit in the structure having minimum interaction energy
will be the evolved drug (ligand). For simplification, a 2-dimensional structure is chosen [1].
Figure 6.Active site of Human Rhinovirus strain 14
Figure 6 illustrates the binding site of Human Rhinovirus strain 14 and a typical structure of ligand is illustrated. For
designing the ligand, the co-ordinates of the residues of the protein must be known. A ligand molecule is assumed to
have a tree like structure on both sides of a fixed pharmacophore illustrated in figure 8. The structure has a left hand
side known as left tree and a right hand side called right tree. For this configuration the right hand side contains 10
leaves and the left hand side contains 7 leaves. Each leaf denoted as a rounded number represents a functional group
among the 8 functional groups listed in figure 7 along with their bond length projection on the x axis (table 2). It is
to be noted that when group 8 occupies a position, the length of the tree is reduced enabling a variable length tree
structure. For fixed length tree, group 8 is absent. The job is to find appropriate functional groups for the leaves such
that the protein ligand docking energy (interaction energy) is minimized. Variable length NBGA is used to design
variable size tree where fixed length structures are optimized with NBGA.
OH
OH
1.Alkyl-1C
2.Alkyl-3C
3.Alkyl-1C-Polar
4.Polar
OH
8.Absence of group
5.Alkyl-3C-polar
6.Aromatic
OH
7.Aromatic Polar
Figure 7. Functional Groups
Functional group
Alkyl-1C
Alkyl-3C
Alkyl-1C-Polar
Alkyl-3C-Polar
Polar
Aromatic
Aromatic-Polar
NUL (no group)
Bond length along x axis (Å)
0.65
1.75
1.1
2.2
0.01
1.9
2.7
Table 2. Bond lengths of Functional Groups
Code
1
2
3
4
5
6
7
8
Figure 8. Active site co-ordinates of Human Rhinovirus strain 14
After computing the co-ordinates of each functional group, the eucledian distances between the group and all the
residues of the active site are computed to find the Van der Waals energy.
5. Variable-Length Version of NBGA:
The size of representing tree is made variable. The length of the chromosome can vary between certain range
denoted by
, where these two are defined as:
is calculated 7 for right tree and 2 for left tree. 8 is used to represent the absence of a functional group. One
extra function named “Correct” is used to check correctness of the newly formed chromosome. Some constrains
should be maintained. The 1st, 3rd and 6th position of right tree and 1st and 3rd position of left tree cannot contain any
type of polar group unless they are the terminals of the tree. If any of the positions among 8 th, 9th, 10th of the right
tree contain any group then 6th position cannot contain 8. Similarly for left tree if any of the positions among 5th, 6th,
and 7th contain any group then 3rd position cannot contain 8. Crossover, Mutation operators are implemented
similarly as described above. We can look for a simple example: say right tree chromosome for two parents be:
[1 5 2 8| 8 1 5 |1 4 6]; [2 7 18 3 2 |5 8 6| 2]
Then corresponding two sons will be:
[1 5 2 8 5 8 6 1 4 6]; [2 7 1 8 3 2 8 1 5 2]
Now we can see that the right tree of the first son contain 8 at 6 th position though its 8th, 9th and 10th places are not
empty. So by using “Correct” function we place a 2 at 6th position so that the two sons will be:
[1 5 2 8 5 2 6 1 4 6]; [2 7 1 8 3 2 8 1 5 2]
These two chromosomes are valid and accepted.
6. Fitness Evaluation:
The computation of fitness was based on the interaction energy of the residues with the closest functional group and
the chemical properties of these pairs. The distance between residues and functional groups should not be more than
2.7Å and less than 0.7Å. If the functional group and the closest residue are of different polarity, a penalty is
imposed. The interaction energy of the ligand with the protein is the sum of Van der Waals potential energy between
the groups and the amino acid residues present in the active site of the target protein. The Van der Waals potential
energy is computed as :
)
Where
and
are integers and
and
)
are constants [12]. Finally fitness
is computed as
Where is a constant (typically 100) and E is the total interaction energy in Kcal/mol. Therefore, maximizing
fitness leads to minimizing interaction energy.
7. Results:
The initial population size is taken to be 100 and the algorithm is run for 100 generations. The interaction energy
using Binary PSO (BPSO), Classical GA, Neighbourhood Based GA (NBGA), Variable length BPSO (VBPSO),
Variable length NBGA (VNBGA) is tabulated in table 3. The configuration of the best right tree and the best left
tree for both fixed length (using NBGA) and variable length (using VNBGA) is drawn in the figure 9 and figure 10.
It is found that variable length structures lead to more stable and fit candidate solution than the fixed length
structure. The fitness of best candidate solution is plotted against generation ( is taken 100) in figure 11.
OH
OH
OH
OH
OH
a.
Best left tree
OH
b. Best right tree
OH
OH
OH
OH
OH
c. Ligand structure
Figure 9. Fixed length tree
OH
Variable Length Tree:
Best left tree:
Best right tree:
OH
OH
OH
OH
OH
a.
Best left tree
b. Best right tree
OH
OH
OH
OH
OH
OH
c. Ligand structure
Figure 10. Variable length structure
Algorithm
Binary PSO (fixed length tree)
GA (fixed length tree)
NBGA (fixed length tree)
VBPSO (variable length tree)
VNBGA (variable length tree)
Interaction energy (Kcal/mol)
13.76375
11.64542
11.57486
8.57756
8.10467
Table 3. Interaction energy values corresponding to Human Rhinovirus strain 14
OH
14
12
Best fitness
10
8
6
BPSO
VL-BPSO
GA
NBGA
VL-NBGA
4
2
0
0
10
20
30
40
50
60
70
80
90
100
Generations/Iterations
Figure 11.Computation of fitness over generations
8. Conclusions and Further research scopes:
We can conclude that Neighbourhood Based GA approach gives better result both for TSP and ligand design
problem. It can be found that using VNBGA we can obtain more stable structure for ligand molecule. The paper
uses a 2-dimensional approach which is quite unrealistic. A 3-dimensional approach using more complex tree
structure will be our future research goal. Although protein-ligand interaction energy is minimized in the current
paper, it may be required in some cases to minimize ligand energy to find stable solution. A multi-objective
approach to minimize protein ligand interaction energy as well as ligand energy will be our future endeavor.
9. References:
1.
Goh G, Foster JA, Evolving Molecules for Drug Design Using Genetic Algorithm, Proc. Int. Conf. on
Genetic & Evol. Computing, Morgan Kaufmann, 27 – 33, 2000
2. Goldberg DE, Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley, New
York, 1989
3. Davis L, Handbook of Genetic Algorithms, New York:Van Nostrand Reinhold, 1991
4. De Jong K, Learning with Genetic Algorithms: An Overview, Machine Learning 3. Norwell, MA: Kluwer,
pp. 121 – 138, 1988
5. Michalewicz Z, Genetic Algorithms + Data Structure = Evolution Programs, Berlin, Germany: SpringerVerlag, 1992
6. Filho JLR, Treleavan PC, Genetic Algorithm Programming Environments, IEEE Comput, 28-43, 1994
7. Ray SS, Bandyopadhyay S, Pal SK (2004) Newoperators of genetic algorithms for traveling salesman
problem. Cambridge, UK, ICPR- 04 2:497–500
8. Tsai CF, Tsai CW, Yang T (2002) A modified multiple-searching method to genetic algorithms for solving
traveling salesman problem. In: IEEE int conf systems, Man and cybernetics, vol. 3, pp6–9
9. Jiao L, Wang L (2000) A novel genetic algorithm based on immunity. IEEE Transactions on Systems, Man
and Cybernetics, Part A 30(5):552–561
10. Larranaga P, Kuijpers C, Murga R, Inza I, Dizdarevic S (1999) Genetic algorithms for the traveling
salesman problem: a review of representations and operators. Artificial Intell Rev 13:129–170
11. Bai Y, Zhang W, Jin Z (2006) An new self-organizing maps strategy for solving the traveling salesman
problem. Chaos, Solitons & Fractals 28(4):1082–1089
12. Leach AR, Molecular Modeling Principles and Applications, Pearson, Prentice Hall, 2001
| 5cs.CE
|
Think Globally, Embed Locally — Locally Linear Meta-embedding of Words
Danushka Bollegala∗§ Kohei Hayashi†§ Ken-ichi Kawarabayashi‡§
arXiv:1709.06671v1 [cs.CL] 19 Sep 2017
The University of Liverpool, Liverpool, United Kingdom.∗
Japan Advanced Institute of Science and Technology, Tokyo, Japan.†
National Institute of Informatics, Tokyo, Japan. ‡
Japan Science and Technology Agency, ERATO, Kawarabayashi Large Graph Project.§
Abstract
Distributed word embeddings have shown superior performances in numerous Natural Language Processing (NLP)
tasks. However, their performances vary significantly across
different tasks, implying that the word embeddings learnt by
those methods capture complementary aspects of lexical semantics. Therefore, we believe that it is important to combine the existing word embeddings to produce more accurate
and complete meta-embeddings of words. For this purpose,
we propose an unsupervised locally linear meta-embedding
learning method that takes pre-trained word embeddings as
the input, and produces more accurate meta embeddings. Unlike previously proposed meta-embedding learning methods
that learn a global projection over all words in a vocabulary,
our proposed method is sensitive to the differences in local
neighbourhoods of the individual source word embeddings.
Moreover, we show that vector concatenation, a previously
proposed highly competitive baseline approach for integrating word embeddings, can be derived as a special case of the
proposed method. Experimental results on semantic similarity, word analogy, relation classification, and short-text classification tasks show that our meta-embeddings to significantly
outperform prior methods in several benchmark datasets, establishing a new state of the art for meta-embeddings.
1
Introduction
Representing the meanings of words is a fundamental task in
Natural Language Processing (NLP). One popular approach
to represent the meaning of a word is to embed it in some
fixed-dimensional vector space (Turney and Pantel, 2010).
In contrast to sparse and high-dimensional counting-based
distributional word representation methods that use cooccurring contexts of a word as its representation (Baroni,
Dinu, and Kruszewski, 2014), dense and low-dimensional
prediction-based distributed word representations have obtained impressive performances in numerous NLP tasks
such as sentiment classification (Socher et al., 2013), and
machine translation (Zou et al., 2013). Several distributed
word embedding learning methods based on different learning strategies have been proposed (Pennington, Socher, and
Manning, 2014; Mikolov, Chen, and Dean, 2013; Huang et
al., 2012; Collobert and Weston, 2008; Mnih and Hinton,
2009).
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
Previous works studying the differences in word embedding learning methods (Chen et al., 2013; Yin and Schütze,
2016) have shown that word embeddings learnt using different methods and from different resources have significant variation in quality and characteristics of the semantics
captured. For example, Hill et al. (2014, 2015) showed that
the word embeddings trained from monolingual vs. bilingual corpora capture different local neighbourhoods. Bansal,
Gimpel, and Livescu (2014) showed that an ensemble of
different word representations improves the accuracy of dependency parsing, implying the complementarity of the different word embeddings. This suggests the importance of
meta-embedding – creating a new embedding by combining
different existing embeddings. We refer to the input word
embeddings to the meta-embedding process as the source
embeddings. Yin and Schütze (2016) showed that by metaembedding five different pre-trained word embeddings, we
can overcome the out-of-vocabulary problem, and improve
the accuracy of cross-domain part-of-speech (POS) tagging.
Encouraged by the above-mentioned prior results, we expect an ensemble containing multiple word embeddings to
produce better performances than the constituent individual
embeddings in NLP tasks.
There are three main challenges a meta-embedding learning method must overcome.
First, the vocabularies covered by the source embeddings
might be different because they have been trained on different text corpora. Therefore, not all words will be equally
represented by all the source embeddings. Even in situations
where the implementations of the word embedding learning methods are publicly available, it might not be possible to retrain those embeddings because the text corpora
on which those methods were originally trained might not
be publicly available. Moreover, it is desirable if the metaembedding method does not require the original resources
upon which they were trained such as corpora or lexicons,
and can directly work with the pre-trained word embeddings. This is particularly attractive from a computational
point of view because re-training source embedding methods on large corpora might require significant processing
times and resources.
Second, the vector spaces and their dimensionalities of the
source embeddings might be different. In most predictionbased word embedding learning methods the word vec-
tors are randomly initialised. Therefore, there is no obvious
correspondence between the dimensions in two word embeddings learnt even from two different runs of the same
method, let alone from different methods (Tian et al., 2016).
Moreover, the pre-trained word embeddings might have different dimensionalities, which is often a hyperparameter set
experimentally. This becomes a challenging task when incorporating multiple source embeddings to learn a single
meta-embedding because the alignment between the dimensionalities of the source embeddings is unknown.
Third, the local neighbourhoods of a particular word under different word embeddings show a significant diversity.
For example, as the nearest neighbours of the word bank,
GloVe (Pennington, Socher, and Manning, 2014), a word
sense insensitive embedding, lists credit, financial, cash,
whereas word sense sensitive embeddings created by Huang
et al. (2012) lists river, valley, marsh when trained on the
same corpus. We see that the nearest neighbours for the different senses of the word bank (i.e. financial institution vs.
river bank) are captured by the different word embeddings.
Meta-embedding learning methods that learn a single global
projection over the entire vocabulary are insensitive to such
local variations in the neighbourhoods (Yin and Schütze,
2016).
To overcome the above-mentioned challenges, we propose a locally-linear meta-embedding learning method that
(a) requires only the words in the vocabulary of each source
embedding, without having to predict embeddings for missing words, (b) can meta-embed source embeddings with different dimensionalities, (c) is sensitive to the diversity of the
neighbourhoods of the source embeddings.
Our proposed method comprises of two steps: a neighbourhood reconstruction step (Section 3.2), and a projection
step (Section 3.3). In the reconstruction step, we represent
the embedding of a word by the linearly weighted combination of the embeddings of its nearest neighbours in each
source embedding space. Although the number of words in
the vocabulary of a particular source embedding can be potentially large, the consideration of nearest neighbours enables us to limit the representation to a handful of parameters per each word, not exceeding the neighbourhood size.
The weights we learn are shared across different source embeddings, thereby incorporating the information from different source embeddings in the meta-embedding. Interestingly, vector concatenation, which has found to be an accurate meta-embedding method, can be derived as a special
case of this reconstruction step.
Next, the projection step computes the meta-embedding
of each word such that the nearest neighbours in the source
embedding spaces are embedded closely to each other in
the meta-embedding space. The reconstruction weights can
be efficiently computed using stochastic gradient descent,
whereas the projection can be efficiently computed using a
truncated eigensolver.
It is noteworthy that we do not directly compare different
source embeddings for the same word in the reconstruction
step nor in the projection step. This is important because
the dimensions in source word embeddings learnt using different word embedding learning methods are not aligned.
Moreover, a particular word might not be represented by all
source embeddings. This property of the proposed method is
attractive because it obviates the need to align source embeddings, or predict missing source word embeddings prior to
meta-embedding. Therefore, all three challenges described
above are solved by the proposed method.
The above-mentioned properties of the proposed method
enables us to compute meta-embeddings for five different source embeddings covering 2.7 million unique words.
We evaluate the meta-embeddings learnt by the proposed
method on semantic similarity prediction, analogy detection, relation classification, and short-text classification
tasks. The proposed method significantly outperforms several competitive baselines and previously proposed metaembedding learning methods (Yin and Schütze, 2016) on
multiple benchmark datasets.
2
Related Work
Yin and Schütze (2016) proposed a meta-embedding learning method (1TON) that projects a meta-embedding of a
word into the source embeddings using separate projection matrices. The projection matrices are learnt by minimising the sum of squared Euclidean distance between the
projected source embeddings and the corresponding original source embeddings for all the words in the vocabulary. They propose an extension (1TON+) to their metaembedding learning method that first predicts the source
word embeddings for out-of-vocabulary words in a particular source embedding, using the known word embeddings.
Next, 1TON method is applied to learn the meta-embeddings
for the union of the vocabularies covered by all of the source
embeddings.
Experimental results in semantic similarity prediction,
word analogy detection, and cross-domain POS tagging
tasks show the effectiveness of both 1TON and 1TON+.
In contrast to our proposed method which learns locallylinear projections that are sensitive to the variations in the
local neighbourhoods in the source embeddings, 1TON and
1TON+ can be seen as globally linear projections between
meta and source embedding spaces. As we see later in Section 4.4, our proposed method outperforms both of those
methods consistently in all benchmark tasks demonstrating the importance of neighbourhood information when
learning meta-embeddings. Moreover, our proposed metaembedding method does not directly compare different
source embeddings, thereby obviating the need to predict
source embeddings for out-of-vocabulary words. Locallylinear embeddings are attractive from a computational pointof-view as well because during optimisation we require information from only the local neighbourhood of each word.
Although not learning any meta-embeddings, several
prior work have shown that by incorporating multiple word
embeddings learnt using different methods improve performance in various NLP tasks. For example, Tsuboi (2014)
showed that by using both word2vec and GloVe embeddings
together in a POS tagging task, it is possible to improve the
tagging accuracy, if we had used only one of those embeddings. Similarly, Turian, Ratinov, and Bengio (2010) collectively used Brown clusters, CW and HLBL embeddings, to
improve the performance of named entity recognition and
chucking tasks.
Luo et al. (2014) proposed a multi-view word embedding
learning method that uses a two-sided neural network. They
adapt pre-trained CBOW (Mikolov et al., 2013) embeddings from Wikipedia and click-through data from a search
engine. Their problem setting is different from ours because their source embeddings are trained using the same
word embedding learning method but on different resources
whereas, we consider source embeddings trained using different word embedding learning methods and resources. Although their method could be potentially extended to metaembed different source embeddings, the unavailability of
their implementation prevented us from exploring this possibility.
Goikoetxea, Agirre, and Soroa (2016) showed that concatenation of word embeddings learnt separately from a
corpus and the WordNet to produce superior word embeddings. Moreover, performing Principal Component Analysis
(PCA) on the concatenated embeddings slightly improved
the performance on word similarity tasks. In Section 4.3, we
discuss the relationship between the proposed method and
vector concatenation.
3
3.1
Locally Linear Meta-Embeddings
Problem Settings
To explain the proposed meta-embedding learning method,
let us consider two source word embeddings, denoted by
S1 and S2 . Although we limit our discussion here to two
source embeddings for the simplicity of the description, the
proposed meta-embedding learning method can be applied
to any number of source embeddings. Indeed in our experiments we consider five different source embeddings. Moreover, the proposed method is not limited to meta-embedding
unigrams, and can be used for n-grams of any length n, provided that we have source embeddings for those n-grams.
We denote the dimensionalities of S1 and S2 respectively
by d1 and d2 (in general, d1 6= d2 ). The sets of words covered by each source embedding (i.e. vocabulary) are denoted
by V1 and V2 . The source embedding of a word v ∈ V1 is
represented by a vector v (1) ∈ Rd1 , whereas the same for
a word v ∈ V2 by a vector v (2) ∈ Rd2 . Let the set union
of V1 and V2 be V = {v1 , . . . , vn } containing n words. In
particular, note that our proposed method does not require a
word v to be represented by all source embeddings, and can
operate on the union of the vocabularies of the source embeddings. The meta-embedding learning problem is then to
learn an embedding v (P) ∈ RdP in a meta-embedding space
P with dimensionality dP for each word v ∈ V.
For a word v, we denote its k-nearest neighbour set in embedding spaces S1 and S2 respectively by N1 (v) and N2 (v)
(in general, |N1 (v)| 6= |N2 (v)|). As discussed already in
Section 1, different word embedding methods encode different aspects of lexical semantics, and are likely to have
different local neighbourhoods. Therefore, by requiring the
meta embedding to consider different neighbourhood constraints in the source embedding spaces we hope to exploit
the complementarity in the source embeddings.
3.2
Nearest Neighbour Reconstruction
The first-step in learning a locally linear meta-embedding is
to reconstruct each source word embedding using a linearly
weighted combination of its k-nearest neighbours. Specifically, we construct each word v ∈ V separately from its
k-nearest neighbours N1 (v), and N2 (v). The reconstruction
weight wvu assigned to a neighbour u ∈ N1 (v) ∪ N2 (v) is
found by minimising the reconstruction error Φ(W) defined
by (1), which is the sum of local distortions in the two source
embedding spaces.
Φ(W) =
2 X
X
2
v (i) −
i=1 v∈V
X
u∈Ni (v)
wvu u(i)
(1)
2
Words that are not k-nearest neighbours of v in either of the
source embedding spaces will have their weights set to zero
(i.e. wvu = 0, ∀u ∈
/ N1 (v) ∪ N2 (v)). Moreover, we require
the sum of
reconstruction
weights for each v to be equal to
P
one (i.e. u∈V wuv = 1).
To compute the weights wvu that minimise (1), we compute its error gradient ∂Φ(W)
∂wvu as follows:
−2
2
X
i=1
>
v (i) −
X
wvx x(i) u(i) I[u ∈ Ni (v)]
x∈Ni (v)
Here, the indicator function, I[x], returns 1 if x is true and
0 otherwise. We uniformly randomly initialise the weights
wvu for each neighbour u of v, and use stochastic gradient descent (SGD) with the learning rate scheduled by AdaGrad (Duchi, Hazan, and Singer, 2011) to compute the optimal values of the weights. The initial learning rate is set to
0.01 and the maximum number of iterations to 100 in our
experiments. Empirically we found that these settings to be
adequate for convergence. Finally, we normalise
P the weights
wuv for each v such that they sum to 1 (i.e. u∈V wvu = 1).
Exact computation of k nearest neighbours for a given
data point in a set of n points requires all pairwise similarity computations. Because we must repeat this process
for each data point in the set, this operation would require a
time complexity of O(n3 ). This is prohibitively large for the
vocabularies we consider in NLP where typically n > 103 .
Therefore, we resort to approximate methods for computing
k nearest neighbours. Specifically, we use the BallTree algorithm (Kibriya and Frank, 2007) to efficiently compute the
approximate k-nearest neighbours, for which the time complexity of tree construction is O(n log n) for n data points.
The solution to the least square problem given by (1) subjected to the summation constraints can be found by solving a set of linear equations. Time complexity of this step
is O(N (d1 |N1 |3 + d2 |N2 |3 )), which is cubic in the neighbourhood size and linear in both the dimensionalities of the
embeddings and vocabulary size. However, we found that
the iterative estimation process using SGD described above
to be more efficient in practice. Because k is significantly
smaller than the number of words in the vocabulary, and often the word being reconstructed is contained in the neighbourhood, the reconstruction weight computation converges
after a small number (less than 5 in our experiments) of iterations.
3.3
Projection to Meta-Embedding Space
In the second step of the proposed method, we compute the
meta-embeddings v (P) , u(P) ∈ RdP for words v, u ∈ V
using the reconstruction weights wvu we computed in Section 3.2. Specifically, the meta-embeddings must minimise
the projection cost, Ψ(P), defined by (2).
Ψ(P) =
X
v (P) −
2
X
X
2
wvu u(P)
i=1 u∈Ni (v)
v∈V
(2)
2
By finding a P space that minimises (2), we hope to preserve
the rich neighbourhood diversity in all source embeddings
within the meta-embedding. The two summations in (2) over
N1 (v) and N2 (v) can be combined to re-write (2) as follows:
2
Ψ(P) =
X
v∈V
X
v (P) −
0
wvu
u(P)
u∈N1 (v)∪N2 (v)
(3)
2
0
Here, wuv
is computed using (4).
0
wvu
= wvu
2
X
I[u ∈ Ni (v)]
(4)
i=1
The dP dimensional meta-embeddings are given by the
eigenvectors corresponding to the smallest (dP + 1) eigenvectors of the matrix M given by (5).
M = (I − W0 )> (I − W0 )
(5)
0
Here, W0 is a matrix with the (v, u) element set to wvu
.
The smallest eigenvalue of M is zero and the corresponding
eigenvector is discarded from the projection. The eigenvectors corresponding to the next smallest dP eigenvalues of
the symmetric matrix M can be found without performing
a full matrix diagonalisation (Bai, 2000). Operations involving M such as the left multiplication by M, which is required
by most sparse eigensolvers, can exploit the fact that M is
expressed in (5) as the product between two sparse matrices. Moreover, truncated randomised methods (Halko, Martinsson, and Tropp, 2010) can be used to find the smallest
eigenvectors, without performing full eigen decompositions.
In our experiments, we set the neighbourhood sizes for all
words in all source embeddings equal to n (i.e ∀i |Ni (v)| =
N, ∀v ∈ V), and project to a dP (< N ) dimensional metaembedding space.
4
4.1
Experiments and Results
Source Word Embeddings
We use five previously proposed pre-trained word embedding sets as the source embeddings in our experiments:
(a) HLBL – hierarchical log-bilinear (Mnih and Hinton,
2009) embeddings released by Turian, Ratinov, and Bengio (2010) (246,122 word embeddings, 100 dimensions,
trained on Reuters Newswire (RCV1) corpus),
(b) Huang – Huang et al. (2012) used global contexts to
train multi-prototype word embeddings that are sensitive
to word senses (100,232 word embeddings, 50 dimensions, trained on April 2010 snapshot of Wikipedia),
(c) GloVe – Pennington, Socher, and Manning (2014) used
global co-occurrences of words over a corpus to learn
word embeddings (1,193,514 word embeddings, 300 dimensions, trained on 42 billion corpus of web crawled
texts),
(d) CW – Collobert and Weston (2008) learnt word embeddings following a multitask learning approach covering multiple NLP tasks (we used the version released by
(Turian, Ratinov, and Bengio, 2010) trained on the same
corpus as HLBL containing 268,810 word embeddings,
200 dimensions),
(e) CBOW – Mikolov et al. (2013) proposed the continuous bag-of-words method to train word embeddings (we
discarded phrase embeddings and selected 929,922 word
embeddings, 300 dimensions, trained on the Google News
corpus containing ca. 100 billion words).
The intersection of the five vocabularies is 35,965 words,
whereas their union is 2,788,636. Although any word embedding can be used as a source we select the abovementioned word embeddings because (a) our goal in this paper is not to compare the differences in performance of the
source embeddings, and (b) by using the same source embeddings as in prior work (Yin and Schütze, 2016), we can
perform a fair evaluation.1 In particular, we could use word
embeddings trained by the same algorithm but on different
resources, or different algorithms on the same resources as
the source embeddings. We defer such evaluations to an extended version of this conference submission.
4.2
Evaluation Tasks
The standard protocol for evaluating word embeddings is to
use the embeddings in some NLP task and to measure the
relative increase (or decrease) in performance in that task.
We use four such extrinsic evaluation tasks:
Semantic similarity measurement: We measure the similarity between two words as the cosine similarity
between the corresponding embeddings, and measure
the Spearman correlation coefficient against the human similarity ratings. We use Rubenstein and Goodenough’s dataset (Rubenstein and Goodenough, 1965)
(RG, 65 word-pairs), rare words dataset (RW, 2034
word-pairs) (Luong, Socher, and Manning, 2013), Stanford’s contextual word similarities (SCWS, 2023 wordpairs) (Huang et al., 2012), the MEN dataset (3000 wordpairs) (Bruni et al., 2012), and the SimLex dataset (Hill,
Reichart, and Korhonen, 2015) (SL 999 word-pairs).
In addition, we use the Miller and Charles’ dataset (Miller
and Charles, 1998) (MC, 30 word-pairs) as a validation dataset to tune various hyperparameters such as the
1
Although skip-gram embeddings are shown to outperform
most other embeddings, they were not used as a source by Yin and
Schütze (2016). Therefore, to be consistent in comparisons against
prior work, we decided not to include skip-gram as a source.
neighbourhood size, and the dimensionality of the metaembeddings for the proposed method and baselines.
Word analogy detection: Using the CosAdd method, we
solve word-analogy questions in the Google dataset
(GL) (Mikolov et al., 2013) (19544 questions), and in the
SemEval (SE) dataset (Jurgens et al., 2012). Specifically,
for three given words a, b and c, we find a fourth word
d that correctly answers the question a to b is c to what?
such that the cosine similarity between the two vectors
(b − a + c) and d is maximised.
Relation classification: We use the DiffVec (DV) (Vylomova et al., 2016) dataset containing 12,458 triples of the
form (relation, word1 , word2 ) covering 15 relation types.
We train a 1-nearest neighbour classifer where for each
target tuple we measure the cosine similarity between the
vector offset for its two word embeddings, and those of
the remaining tuples in the dataset. If the top ranked tuple
has the same relation as the target tuple, then it is considered to be a correct match. We compute the (microaveraged) classification accuracy over the entire dataset
as the evaluation measure.
Short-text classification: We use two binary short-text
classification datasets: Stanford sentiment treebank (TR)2
(903 positive test instances and 903 negative test instances), and the movie reviews dataset (MR) (Pang and
Lee, 2005) (5331 positive instances and 5331 negative instances). Each review is represented as a bag-of-words
and we compute the centroid of the embeddings of the
words in each bag to represent that review. Next, we
train a binary logistic regression classifier with a crossvalidated `2 regulariser using the train portion of each
dataset, and evaluate the classification accuracy using the
test portion of the dataset.
4.3
Baselines
Concatenation (CONC): A simple baseline method for
combining pre-trained word embeddings is to concatenate
the embedding vectors for a word w to produce a metaembedding for w. Each source embedding of w is `2 normalised prior to concatenation such that each source embedding contributes equally (a value in [−1, 1]) when measuring the word similarity using the dot product. As also observed by Yin and Schütze (2016) we found that CONC performs poorly without emphasising GloVe and CBOW by a
constant factor (which is set to 8 using MC as a validation
dataset) when used in conjunction with HLBL, Huang, and
CW source embeddings.
Interestingly, concatenation can be seen as a special case
in the reconstruction step described in Section 3.2. To see
this, let us denote the concatenation of column vectors v (1)
and v (2) by x = (v (1) ; v (2) ), and u(1) and u(2) by y =
(u(1) ; u(2) ), where x, y ∈ Rd1 +d2 . Then, the reconstruction
error defined by (1) can be written as follows:
2
Φ(W) =
X
v∈V
2
x−
X
wvu y
u∈N (v)
http://nlp.stanford.edu/sentiment/treebank.html
(6)
2
Here, the vocabulary V is constrained to the intersection
VS ∩ VT because concatenation is not defined for missing words in a source embedding. Alternatively, one could
use zero-vectors for missing words or (better) predict the
word embeddings for missing words prior to concatenation. However, we consider such extensions to be beyond
the simple concatenation baseline we consider here.3 On the
other hand, the common neighbourhood N (v) in (6) can
be obtained by either limiting N (v) to N1 (v) ∩ N2 (v) or,
by extending the neighbourhoods to the entire vocabulary
(N (v) = V). (6) shows that under those neighbourhood constraints, the first step in our proposed method can be seen as
reconstructing the neighbourhood of the concatenated space.
The second step would then find meta-embeddings that preserve the locally linear structure in the concatenated space.
One drawback of concatenation is that it increases the
dimensionality of the meta-embeddings compared to the
source-embeddings, which might be problematic when storing or processing the meta-embeddings (for example, for the
five source embeddings we use here dP = 100 + 50 + 300 +
200 + 300 = 950).
Singular Value Decomposition (SVD): We create an N ×
950 matrix C by arranging the CONC vectors for the union
of all source embedding vocabularies. For words that are
missing in a particular source embedding, we assign zero
vectors of that source embedding’s dimensionality. Next, we
perform SVD on C = UDV> , where U and V are unitary
matrices and the diagonal matrix D contains the singular values of C. We then select the d largest left singular vectors
from U to create a dP dimensional embeddings for the N
words. Using the MC validation dataset, we set dP = 300.
Multiplying U by the singular values, a technique used to
weight the latent dimensions considering the salience of the
singular values, did not result in any notable improvements
in our experiments.
4.4
Meta-Embedding Results
Using the MC dataset, we find the best values for the neighbourhood size n = 1200 and dimensionality dP = 300 for
the Proposed method. We plan to publicly release our metaembeddings on acceptance of the paper.
We summarise the experimental results for different methods on different tasks/datasets in Table 1. In Table 1, rows
1-5 show the performance of the individual source embeddings. Next, we perform ablation tests (rows 6-20) where we
hold-out one source embedding and use the other four with
each meta-embedding method. We evaluate statistical significance against best performing individual source embedding
on each dataset. For the semantic similarity benchmarks we
use Fisher transformation to compute p < 0.05 confidence
intervals for Spearman correlation coefficients. In all other
(classification) datasets, we used Clopper-Pearson binomial
exact confidence intervals at p < 0.05.
3
Missing words does not affect the performance of CONC because all words in the benchmark datasets we use in our experiments are covered by all source embeddings.
RG
WS
RW
SCWS
MEN
SL
RG
0.85
RW
SCWS
MEN
SL
0.8
0.75
Spearman Correlation
Spearman Correlation
0.80
0.70
0.65
0.60
0.55
0.7
0.6
0.5
0.4
0.50
0.45
100
WS
0.9
200
300
400
500
600
700
800
0.3
900
0
500
Dimensionality
1000
1500
2000
2500
Neighbourhood size
(a)
(b)
ensemble
ablation
sources
Figure 1: Performance vs. dimensionality (neighbourhood size fixed at 1200) shown in (a), and vs. neighbourhood size (dimensionality fixed at 300) shown in (b) for meta embedding learning.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
Model
GloVe
CBOW
HLBL
Huang
CW
CONC (-GloVe)
CONC (-CBOW)
CONC (-HLBL)
CONC (-Huang)
CONC (-CW)
SVD (-GloVe)
SVD (-CBOW)
SVD (-HLBL)
SVD (-Huang)
SVD (-CW)
Proposed (-GloVe)
Proposed (-CBOW)
Proposed (-HLBL)
Proposed (-Huang)
Proposed (-CW)
CONC
SVD
1TON
1TON+
Proposed
RG
81.7
76.0
35.3
51.3
29.9
75.0
80.8
83.0
83.0
82.9
78.6
80.5
82.7
82.5
82.5
79.8
80.9
82.1
81.2
83.1
82.9
82.7
80.7
82.7
83.4
MC
80.8
82.2
49.3
58.8
34.3
79.0
81.0
84.0
84.0
84.0
79.9
81.2
83.6
85.0
83.9
79.7
82.1
86.1
85.2
84.8
84.1
83.9
80.7
85.0
86.2
WS
64.3
69.8
35.7
58.0
28.4
70.0
65.2
71.9
71.6
71.9
68.4
64.4
70.3
70.3
70.4
71.1
67.4
71.3
73.1
72.5
71.9
70.4
74.5
75.3
75.7∗
RW
38.4
53.4
19.1
36.4
15.3
55.3
46.0
53.4
48.8
53.3
53.9
45.3
52.6
48.6
52.5
54.7
58.7∗
58.3∗
55.1
58.5
53.3
52.6
60.1∗
61.6∗
62.8∗
SCWS MEN
54.0 74.3
53.4 78.2
47.7 30.7
63.7 56.1
39.8 25.7
62.9 77.7
56.3 74.9
61.4 80.1∗
60.8 80.1∗
61.6 80.2∗
61.6 77.5
55.3 74.2
60.1 79.9∗
59.8 79.9∗
60.1 80.0∗
62.3 78.2
58.7 75.7
62.1 81.9∗
63.7 81.4∗
62.3 81.1∗
61.5 80.2∗
60.0 79.9∗
61.6 73.5
60.2 74.1
63.8 82.2∗
SL
37.4
44.2
22.1
21.7
15.6
41.5
37.3
41.6
41.9
41.6
40.1
35.7
39.6
39.9
39.7
46.1
45.2
34.8
42.3
43.5
41.6
39.7
46.4
46.3
48.7
GL
70.5
75.2
16.6
8.3
4.7
64.0
70.0
72.7
72.8
72.6
61.7
70.9
73.5
73.7
73.3
84.2∗
85.2∗
86.3∗
82.6∗
88.4∗
72.9
73.4
76.8
77.0
89.9∗
SE
39.9
39.1
34.8
35.2
34.6
38.7
38.8
39.5
40.0
39.6
38.5
38.7
39.8
40.0
39.8
39.8
40.1
40.3
41.1
41.9
39.6
39.7
42.3∗
40.1
43.1∗
DV
87.7
87.4
72.0
76.0
75.6
82.9
86.0
84.9
86.7
84.9
84.1
86.7
87.3
87.3
87.2
85.4
87.1
87.7
87.5
87.8
84.9
87.2
87.6
83.9
88.7∗
SA
73.4
73.6
62.6
64.8
62.7
72.1
71.6
71.0
71.2
72.3
71.6
73.4
73.2
73.5
73.1
72.2
73.8
73.7
73.9
71.6
72.4
73.4
73.8
73.9
74.0
MR
70.0
71.0
61.6
60.9
61.4
69.1
69.9
69.4
69.1
69.9
69.8
69.1
70.4
70.8
70.7
70.2
70.1
71.1
71.2
71.1
69.9
70.7
70.3
69.2
71.3
Table 1: Results on word similarity, analogy, relation and short-text classification tasks. For each task, the best performing
method is shown in bold. Statistically significant improvements over the best individual source embedding are indicated by an
asterisk.
Among the individual source embeddings, we see that
GloVe and CBOW stand out as the two best embeddings.
This observation is further confirmed from ablation results,
where the removal of GloVe or CBOW often results in a
decrease in performance. Performing SVD (rows 11-15) after concatenating, does not always result in an improvement.
SVD is a global projection that reduces the dimensionality
of the meta-embeddings created via concatenation. This result indicates that different source embeddings might require
different levels of dimensionality reductions, and applying a
single global projection does not always guarantee improvements. Ensemble methods that use all five source embed-
dings are shown in rows 21-25. 1TON and 1TON+ are proposed by Yin and Schütze (2016), and were detailed in Section 2. Because they did not evaluate on all tasks that we
do here, to conduct a fair and consistent evaluation we used
their publicly available meta-embeddings4 without retraining by ourselves.
Overall, from Table 1, we see that the Proposed method
(row 25) obtains the best performance in all tasks/datasets.
In 6 out of 12 benchmarks, this improvement is statistically
significant over the best single source embedding. Moreover,
4
http://cistern.cis.lmu.de/meta-emb/
4.5
Complementarity of Resources
We have shown empirically in Section 4.4 that using the
proposed method it is possible to obtain superior metaembeddings from a diverse set of source embeddings. One
important scenario where meta-embedding could be potentially useful is when the source embeddings are trained
on different complementary resources, where each resource
share little common vocabulary. For example, one source
embedding might have been trained on Wikipedia whereas
a second source embedding might have been trained on
tweets.
To evaluate the effectiveness of the proposed metaembedding learning method under such settings, we design the following experiment. We select MEN dataset, the
largest among all semantic similarity benchmarks, which
contains 751 unique words in 3000 human-rated word-pairs
for semantic similarity. Next, we randomly split the set of
words into two sets with different overlap ratios. We then
S1
S2
CONC
0.2
0.4
0.6
Meta
0.9
0.8
0.7
Spearman Correlation
in the MEN dataset (the largest among the semantic similarity benchmarks compared in Table 1 with 3000 word-pairs),
and the Google dataset, the improvements of the Proposed
method over the previously proposed 1TON and 1TON+ are
statistically significant.
The ablation results for the Proposed method show that,
although different source embeddings are important to different degrees, by using all source embeddings we can obtain the best results. Different source embeddings are trained
from different resources and by optimising different objectives. Therefore, for different words, the local neighbours predicted by different source embeddings will be
complementary. Unlike the other methods, the Proposed
method never compares different source embeddings’ vectors directly, but only via the neighbourhood reconstruction
weights. Consequently, the Proposed method is unaffected
by relative weighting of source embeddings. In contrast, the
CONC is highly sensitive against the weighting. In fact, we
confirmed that the performance scores of the CONC method
were decreased by 3–10 points when we did not do the
weight tuning described in Section 4.2. The unnecessity of
the weight tuning is thus a clear advantage of the Proposed
method.
To investigate the effect of the dimensionality dP on the
meta-embeddings learnt by the proposed method, in Figure 1a, we fix the neighbourhood size N = 1200 and measure the performance on semantic similarity measurement
tasks when varying dP . Overall, we see that the performance
peaks around dP = 300. Such behaviour can be explained
by the fact that smaller dP dimensions are unable to preserve
information contained in the source embeddings, whereas
increasing dP beyond the rank of the weight matrix W is
likely to generate noisy eigenvectors.
In Figure 1b, we study the effect of increasing the neighbourhood size n equally for all words in all source embeddings, while fixing the dimensionality of the metaembedding dP = 300. Initially, performance increases with
the neighbourhood size and then saturates. This implies that
in practice a small local neighbourhood is adequate to capture the differences in source embeddings.
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.0
0.8
1.0
Overlap Ratio
Figure 2: Two word embedding learning algorithms trained
on different but overlapping corpora to produce two source
embeddings S1 and S2 . and their meta-embedding.
select sentences from 2017 January dump of Wikipedia that
contains words from only one of the two sets. We create two corpora of roughly equal number of sentences via
this procedure for different overlap ratios. We train skipgram with negative sampling (SGNS) (Mikolov, Chen, and
Dean, 2013) on one corpus to create source embedding S1
and GloVe (Pennington, Socher, and Manning, 2014) on the
other corpus to create source embedding S2 . Finally, we use
the proposed method to meta-embed S1 and S2 .
Figure 2 shows the Spearman correlation between the human similarity ratings and cosine similarities computed using the word embeddings on the MEN dataset for S1 , S2 and
their meta-embeddings created using the proposed method
(Meta) and concatenation baseline (CONC). From Figure 2,
we see that the meta embeddings obtain the best performance across all overlap ratios. The improvements are larger
when the overlap between the corpora is smaller, and diminishes when the two corpora becomes identical. This result
shows that our proposed meta-embedding learning method
captures the complementary information available in different source embeddings to create more accurate word embeddings. Moreover, it shows that by considering the local neighbourhoods in each of the source embeddings separately, we can obviate the need to predict embeddings for
missing words in a particular source embedding, which was
a limitation in the method proposed by Yin and Schütze
(2016).
5
Conclusion
We proposed an unsupervised locally linear method for
learning meta-embeddings from a given set of pre-trained
source embeddings. Experiments on several NLP tasks show
the accuracy of the proposed method, which outperforms
previously proposed meta-embedding learning methods on
multiple benchmark datasets. In future, we plan to extend the
proposed method to learn cross-lingual meta-embeddings by
incorporating both cross-lingual as well as monolingual information.
References
2000. Templates for the Solution of Algebraic Eigenvalue
Problems: A Practical Guide. Society for the Industrial
and Applied Mathematics.
Bansal, M.; Gimpel, K.; and Livescu, K. 2014. Tailoring
continouos word representations for dependency parsing.
In Proc. of ACL.
Baroni, M.; Dinu, G.; and Kruszewski, G. 2014. Don’t
count, predict! a systematic comparison of contextcounting vs. context-predicting semantic vectors. In Proc.
of ACL, 238–247.
Bruni, E.; Boleda, G.; Baroni, M.; and Tran, N. K. 2012.
Distributional semantics in technicolor. In Proc. of ACL,
136–145.
Chen, Y.; Perozzi, B.; Al-Rfou, R.; and Skiena, S. 2013. The
expressive power of word embeddings. In Proc. of ICML
Workshop.
Collobert, R., and Weston, J. 2008. A unified architecture for
natural language processing: Deep neural networks with
multitask learning. In Proc. of ICML, 160 – 167.
Duchi, J.; Hazan, E.; and Singer, Y. 2011. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research 12:2121–
2159.
Goikoetxea, J.; Agirre, E.; and Soroa, A. 2016. Single or
multiple? combining word representations independently
learned from text and wordnet. In Proc. of AAAI, 2608–
2614.
Halko, N.; Martinsson, P. G.; and Tropp, J. A. 2010. Finding
structure with randomness: Probabilistic algorithms for
constructung approximate matrix decompositions. SIAM
REVIEW 53(2):217 – 288.
Hill, F.; Cho, K.; Jean, S.; Devin, C.; and Bengio, Y. 2014.
Not all neural embeddings are born equal. In NIPS workshop.
Hill, F.; Cho, K.; Jean, S.; Devin, C.; and Bengio, Y. 2015.
Embedding word similarity with neural machine translation. In ICLR Workshop.
Hill, F.; Reichart, R.; and Korhonen, A. 2015. Simlex-999:
Evaluating semantic models with (genuine) similarity estimation. Computational Linguistics 41(4):665–695.
Huang, E. H.; Socher, R.; Manning, C. D.; and Ng, A. Y.
2012. Improving word representations via global context
and multiple word prototypes. In Proc. of ACL, 873–882.
Jurgens, D. A.; Mohammad, S.; Turney, P. D.; and Holyoak,
K. J. 2012. Measuring degrees of relational similarity. In
Proc. of SemEval.
Kibriya, A. M., and Frank, E. 2007. An empirical comparison of exact nearest neighbout algorithms. In Proc. of
PKDD, 140–151.
Luo, Y.; Tang, J.; Yan, J.; Xu, C.; and Chen, Z. 2014. Pretrained multi-view word embedding using two-side neural
network. In Proc. of AAAI, 1982–1988.
Luong, M.-T.; Socher, R.; and Manning, C. D. 2013. Better
word representations with recursive neural networks for
morphology. In CoNLL.
Mikolov, T.; Sutskever, I.; Chen, K.; Corrado, G. S.; and
Dean, J. 2013. Distributed representations of words and
phrases and their compositionality. In Proc. of NIPS,
3111–3119.
Mikolov, T.; Chen, K.; and Dean, J. 2013. Efficient estimation of word representation in vector space. In Proc. of
ICLR.
Miller, G., and Charles, W. 1998. Contextual correlates of
semantic similarity. Language and Cognitive Processes
6(1):1–28.
Mnih, A., and Hinton, G. E. 2009. A scalable hierarchical
distributed language model. In Koller, D.; Schuurmans,
D.; Bengio, Y.; and Bottou, L., eds., Proc. of NIPS. 1081–
1088.
Pang, B., and Lee, L. 2005. Seeing stars: Exploiting class
relationships for sentiment categorization with respect to
rating scales. In Proc. of ACL, 115–124.
Pennington, J.; Socher, R.; and Manning, C. D. 2014.
Glove: global vectors for word representation. In Proc.
of EMNLP, 1532–1543.
Rubenstein, H., and Goodenough, J. 1965. Contextual correlates of synonymy. Communications of the ACM 8:627–
633.
Socher, R.; Perelygin, A.; Wu, J.; Chuang, J.; Manning,
C. D.; Ng, A.; and Potts, C. 2013. Recursive deep models
for semantic compositionality over a sentiment treebank.
In Proc. of EMNLP, 1631–1642.
Tian, Y.; Kullkarni, V.; Perozzi, B.; and Skiena, S. 2016. On
the convergent properties of word embedding methods.
arXiv.
Tsuboi, Y. 2014. Neural networks leverage corpus-wide information for part-of-speech tagging. In Proc. of EMNLP,
938–950.
Turian, J.; Ratinov, L.; and Bengio, Y. 2010. Word representations: A simple and general method for semi-supervised
learning. In Proc. of ACL, 384 – 394.
Turney, P. D., and Pantel, P. 2010. From frequency to meaning: Vector space models of semantics. Journal of Aritificial Intelligence Research 37:141 – 188.
Vylomova, E.; Rimell, L.; Cohn, T.; and Baldwin, T. 2016.
Take and took, gaggle and goose, book and read: Evaluating the utility of vector differences for lexical relational
learning. In ACL, 1671–1682.
Yin, W., and Schütze, H. 2016. Learning meta-embeddings
by using ensembles of embedding sets. In Proc. of ACL,
1351–1360.
Zou, W. Y.; Socher, R.; Cer, D.; and Manning, C. D. 2013.
Bilingual word embeddings for phrase-based machine
translation. In Proc. of EMNLP, 1393–1398.
| 9cs.NE
|
HTM-MAT: An online prediction software toolbox based on cortical machine learning algorithm
a
V.I.E Anireh , E.N. Osegi
a.
b,*
Department of Computer Science, Rivers State University of Science and Technology, Nigeria
b. Department of Information and Communication Technology, National Open University of Nigeria
Abstract.
HTM-MAT is a MATLAB® - based toolbox for implementing cortical learning algorithms (CLA) including related
cortical-like algorithms that possesses spatiotemporal properties. CLA is a suite of predictive machine learning
algorithms developed by Numenta Inc. and is based on the hierarchical temporal memory (HTM). This paper
presents an implementation of HTM-MAT with several illustrative examples including several toy datasets and
compared with two sequence learning applications employing state-of-the-art algorithms - the recurrentjs based on
the Long Short-Term Memory (LSTM) algorithm and OS-ELM which is based on an online sequential version of the
Extreme Learning Machine. The performance of HTM-MAT using two historical benchmark datasets and one real
world dataset is also compared with one of the existing sequence learning applications, the OS-ELM. The results
indicate that HTM-MAT predictions are indeed competitive and can outperform OS-ELM in sequential prediction
tasks.
Keywords:
Machine learning; predictions; cortical learning algorithms; hierarchical temporal memory
*Corresponding author.
E-mail: [email protected]
1
1.
Motivation and Significance
Predictive systems are very vital in the monitoring and operation of modern day businesses, infrastructure and
equipment’s. For instance in the power industry, utility operators need to determine the expected load profile in
advance in order to make proper load schedule for the day; bank operators need to determine the likelihood that a
fraudulent bank transaction will occur; web system operators need to detect rogue like behaviour in order to avert
security breaches within a given web portal. In order to gain insight into the behaviour of the aforementioned tasks,
such systems require predictive coding techniques which in turn play a major role in predicting the hidden causes of
incoming sensory information [1]. In recent times, cortical learning algorithms (CLA) have been proposed in [2] and
also applied to the solution of several real challenging problems including retina analysis [3], sign language and
gaze gesture recognition [4-5] and weld flaw detection [6] have been investigated with very promising results. This
suite of algorithms has very interesting properties and represents a major step in the quest for a more biological
plausible neural network for machine learning tasks. However, until this present moment, an open source cortical
learning toolbox that can facilitate research and understanding of CLA-like algorithms in a variety of machine
learning predictive tasks is lacking in the technical computing community.
In this paper we present HTM-MAT, a software tool which allows such predictive tasks to be accomplished. HTMMAT is a barebones minimalist version of the cortical learning algorithm proposed in [2] that allows regression and
classification functions to be implemented in a predictive way. The theory is based on the hypothesis of the use of
spatiotemporal sparse memory predictive neural architecture for online machine learning tasks. The primary
purpose of developing HTM-MAT is to make readily available a domain friendly, reusable and comprehensive
cortical-learning toolbox for machine learning experiments to be performed on both synthetic and real world data
in a technical computing environment.
It is intended to serve as a reusable, straightforward and easy to re-invent/reuse machine learning tool for
academic researchers, students and machine learning experts. HTM-MAT depends on the Hierarchical Temporal
Memory (HTM) which is the underlying theory that defines its core operations.
1.1. HTM theoretical background
Hierarchical Temporal Memory (HTM) is originally a memory-prediction framework, and an emerging Artificial
Intelligence (AI) approach that is based on the notion that time plays a crucial role in the operation of neural
networks [2,7-9]. In its current implementation, it uses the cortical learning algorithms - a suite of algorithms that
favors the use of a more biological plausible paradigm. HTM advocates that the neocortex is the primary entity and
the main seat of intelligence in the brain.
The core operations that make HTM interestingly unique are centered on the properties of sparse distributed
representations (SDRs) being the fundamental data structure in HTM systems [10]. The primary operations in HTM
include [9]:
The activation and prediction of cell states/columns based on the notion of permanence
Computation of a similarity metric called the overlap
A learning rule for updating permanence’s
The prediction and activation states at time step, t, are typically computed using Eq. (1) and Eq. (2) respectively
as:
~d
t
1 if d || Di j ||1
0 otherwise
t
ij
(1)
2
1
t
1
a
ij
0
if
j W t and
it j 1 1
j W t and t 1 0
i j
i
otherwise
(2)
if
where,
~
Didj = an M×N binary matrix representing the permanence of a connected synapse belonging to a dendrite
segment
d
= a cortical segment
i, j = cell and column states
t = an M×N binary matrix denoting the activation state of the cortical network
= the cortical segment activation threshold or permanence
N = number of cortical columns
M = number of neurons (cortical cells) per column
W t = percentage of columns with most synaptic inputs
~d
The operation || Di j ||1 , is fundamentally referred to as the overlap .
t
In order to improve on the HTM predictions, a Hebbian-like learning rule is applied to the dendrite segment:
Didj p Didj At 1 p Didj
+
(3)
-
where, p and p are permanence increments and decrements respectively.
The percentage of active columns typically follows a 2% rule for maximum fixed sparse distribution. The assumption
that the 2% rule holds and can be determined apriori remains a fundamental issue and yet to be resolved by HTM
researchers. More so, it is still not clear if the brain uses this rule for learning sparsely. Nonetheless, the HTM model
still works reasonably well in certain sequence learning tasks such as in time series related problems.
3
1.2. Concept of an online sequential cortical learning algorithm
Online sequential learning emphasizes the predictions of patterns that show some interesting regularities
or properties in data e.g. the prediction of the likelihood of an anomaly in equipment given a history of
equipment conditions. HTM-MAT is a sequence learning framework based on cortical learning principles and is
generically designed to operate in an online manner.
A typical online learning process for cortical-like algorithms consists of four key stages. First, a dataset
containing the source file (sensory input data) is encoded into a suitable (mixed-integer) representation. This is
achieved by using a scalar encoder. Second, using the concepts introduced earlier in Section 1.1, a sparse
distributed representation (SDR) of the encoder representations is generated using a spatial pooler algorithm.
Third, the SDR formed in the spatial pooler stage is further processed to generate a hierarchy of sparse
representations using a temporal pooler algorithm. Fourth, a greedy prediction search is performed to extract
the possible sparse pattern meeting a pre-specified overlap threshold. Full details of the spatiotemporal
processing inherent in cortical-like applications including a comprehensive mathematical treatment of the
formulation of sparse distributed structures can be found in [2, 9 and 10]. Note, that in HTM-MAT, we do not
consider the expensive method of combinatorics proposed in [10], but rather use a Monte Carlo simulation to
evolve a sparse distributed representation. An implementation of HTM-MAT is provided in [11].
1.2.1 Related sequential learning algorithms
The online sequential extreme learning machine (OS-ELM) is also an emerging sequence learning
framework that is based on the very popular Extreme Learning Machines (ELM).OS-ELM has been shown to be
very fast and highly accurate [12]. However, this model does not fully account for the notion of time and
hierarchy in space. It also does not account for sparse distributed representations which are core features of
truly biological-plausible machine learning applications. In addition OS-ELM does not have any inherent
mechanism for handling categorical inputs and requires supervised operation. Implementations of OS-ELM can
be found in [13].
Another sequence learning algorithm, Long Short-Term Memory (LSTM) neural network originally developed in
[14] represent the state-of-the-art in predictive mining tasks specifically in the area of language modelling and
grammar recognition/prediction. It can operate both in supervised, semi-supervised and unsupervised modes. In
particular, LSTM networks can be used as a reference benchmark for evaluating the predictive power of other
machine learning algorithms or neural networks where the task is to predict in advance, the possible output
response state given a history of previous states. LSTM model neural networks have been used in a variety of
tasks including language modelling [15-17], grammar/word learning tasks [18] and speech recognition [19, 20].
An implementation of LSTM can be found in [21].
1.3. HTM-MAT
HTM-MAT is a MATLAB ® toolbox for performing machine learning experiments on streaming or sequential
data. The primary feature of HTM-MAT is that it uses a sparse mixed-integer representation for encoding data
during training and learning. Using a sparse mixed integer representation gives HTM-MAT the advantage of a
character-level or symbolic explanation facility i.e. HTM-MAT can generate more meaningful and expressive
results that still follow a sparse distributed representation using the core HTM principles.
Since HTM-MAT uses an unsupervised machine learning algorithm, it can also derive a solution to the
problem in a more compact and adaptive manner. Though, HTM-MAT is specifically intended for researchers
and students in academic environments, it may be possible to adapt it to industry or commercial
applications.HTM-MAT can also be helpful to users new to the use of cortical-like machine learning algorithms
for predictive classification tasks.
4
2. Software description
HTM-MAT exploits the functional programming capabilities of MATLAB ®. HTM-MAT interacts with the user
through a set of functional routines typically instantiated in a main functional class. The architectural concept is as
shown in Fig.1. Basically, the user interacts with HTM-MAT (top-layer) through a set of functions (middle-layer).
These functions include routines for handling spatial and temporal pooling operations as well as for classifying the
recognition units. The lower layer (DATASET FILE) is used to define the dataset needed for the prediction or
classification task.
HTM-MAT
HTM-MAT FUNCTIONS
DATASET FILE
Fig1.Architecture of HTM-MAT application
Full details of the core functions used in HTM-MAT can be found in the user guide that comes with the software
package [11].
3. Illustrative Examples
A sample implementation of a prediction task using a synthetic or benchmark dataset is as shown in Table 1. In
this task, the goal is to implement a code routine to predict the target value of the last exemplar given a history of
previous inputs. This routine consists of the following blocks of code:
HTM-System Parameters: First the data source (dataset) name is defined followed by a definition of
several core system parameters such as the number of Monte Carlo (MC) iterations (iters), minimum
overlap (min_overlap), permanence threshold (perms_th), the number of considered sequences (seq_size)
and the percent overlap adjustment for greedy-max predictions.
Spatiotemporal processing: This requires a function call to ST_HTM.m functional class using most of the
parameters defined in the HTM-System Parameters block. This operation returns the full mean absolute
percentage error (MAPE), a mixed-integer representation of the original input data sequence
(original_data) and sparse generated data (sparsed_data), the overlap (overlap), temporal sparse data
(sparsed_data_T), minimum values obtained meeting the overlap threshold (min_val) and the
permanence.
Error-plots: Plots the MAPE errors generated by the system after the pre-specified number of MC
iterations.
Char-Integer transformation: This transformation converts the mixed-integer character representations of
original and sparse generated data into an equivalent numeric (integer-based) representation.This is very
important for extracting the predictions (using a prefix-search) and performing the necessary
5
classifications.
Argmax-predictions: This block allows greedy character-wise predictions to be performed on the
recognition units. The predictions are stored in a variable called predicted_out.
Performance metrics: This block computes the performance metrics such as the percentage classification
accuracy, mean absolute percentage error (MAPE), and the root-mean squared error (RMSE) given the
original (in_p) and sparse recognition (out_s) integer units.
Table1. A sample HTM-MAT implementation
data_name = 'heart_data'; % dataset
%% HTM system parameters:
iters = 50;
min_overlap = 2;
perms_th = 0.21;
desired_localActivity = 2;
seq_size = 700;
per_adjust = 99;
%% Spatiotemporal processing:
[MAPE_n,original_data,sparsed_data,overlap,sparsed_data_T,min_val,permanence] = ST_HTM(iters,...
min_overlap,perms_th,desired_localActivity,data_name,seq_size);
%% Error Plots:
plot(MAPE_n);
%% Char-Integer Transformation:
in_p = original_data.*1;
out_s = sparsed_data.*1;
%% Argmax-predictions:
predicted_out(pred_idx_n,:) = char(out_s(pred_idx_n,:));
%% Performance metrics:
%Compute the errors:
RMSE = RMSE_func(in_p,out_s)
In Table 2 is shown a comparative prediction of the HTM-MAT with two applications employing state-of-the-art
algorithms using four toy datasets; these applications, recurrentjs developed in [21] and OS-ELM applications using
three different activation functions [13] have been earlier introduced in sub-section 1.2.1. The toy datasets include
a multiplication operation obtained from the very popular times-table used in high-school mathematics, and three
fictitious datasets of word representations (applicable to only HTM-MAT and recurrentjs since they both can handle
character-level input data). These datasets are included in the HTM-MAT toolbox [11].
In the first toy dataset (times_trainv1), the target value of the last exemplar was intentionally set to zero in order to
find out if the algorithm can accurately compute the correct answer.
Table 3 also shows comparative experimental results of HTM-MAT with the OS-ELM applications using two
historical benchmark datasets and one field data for the predictive classification task. The historical datasets can be
obtained from [23] and consists of a heart dataset (heart_data) for categorizing a heart condition as either present
(1) or absent (0) and a fraud dataset (australian_data) for detecting the presence (1) or absence of a fraud (0) . The
field data represent two-hourly pressure readings obtained from a sensor attached to an oil pipeline where a 1
indicates a likely threat on the facility. These results are interpreted in terms of the root-mean-squared error
6
(RMSE). Simulations were performed for a maximum of 50 MC iterations for HTM-MAT. The results in Table 2
indicate that HTM-MAT is indeed capable of numeric and word predictions that are indeed a reflection of the
pattern in the dataset. The results in Table 3 show that HTM-MAT outperforms the OS-ELM on all tasks presented
and achieves 100% accuracy on the real world dataset.
It is important to note that the OS-ELM parameters was especially hand-tuned to obtain the best possible set of
predictions which was found to be achievable when the number of hidden neurons is greater than 100 and the
number of initial data and block size is set to 3 and 0 respectively.
Table2. Comparative word/numeric predictions of HTM-MAT, recurrentjs and OS-ELM using toy datasets
Dataset
recurrentjs
OS-ELMrbf
236
2 3 6.1
word3a
Fishing,
Fish-feed,
Fish
word3b
word3c
times_trainv1
a
OS-ELMsig
b
OS-ELMsin
c
2 3 5.5
2 3 5.8
236
NA
NA
NA
Fast,
F sh
Football,
Fans
NA
NA
NA
Foot
VideoPlayer,
Video
NA
NA
NA
Video
OS-ELMrbf – OS-ELM using radial basis activation function
OS-ELMsig – OS-ELM using sigmoid activation function
c
OS-ELMsin – OS-ELM using sine activation function
a
b
Table 3
Root-mean squared errors of HTM-MAT and OS-ELM variants using two historical and one real dataset
Dataset
HTMMAT
OS-ELMrbf
(RMSE)
OS-ELMsig
(RMSE)
OS-ELMsin
(RMSE)
HTM-MAT
(RMSE)
heart_data
1.5275
0.6734
1.5816
0.2582
australian_data
0.8236
0.7164
0.8249
0.0381
pressure_data
0.5075
0.5075
0.5075
0.0000
7
4. Impact
HTM-MAT is an open source MATLAB toolbox that allows cortical learning algorithms to be explored in a
technical computing environment. HTM-MAT to the best of our knowledge presents the first attempt to implement
a comprehensive toolbox of HTM cortical like algorithms in a technical computing environment such as MATLAB ®.
HTM-MAT is a reusable framework that can easily be adapted to include novel neural-like algorithms that are
biologically plausible. This implies that HTM-MAT is indeed a framework for exploring any kind of neural network or
technology that is deeply rooted in cortical-like principles. However, this remains an open challenge for neural
network experts but a world of possibilities for software developers. It is expected that HTM-MAT will facilitate
research in this direction and help advance the state-of-the-art in predictive applications.
5. Conclusions
This paper presents HTM-MAT, an open source toolbox implementation of the cortical learning algorithm (CLA)
which is based on the hierarchical temporal memory (HTM) theory. HTM-MAT is inspired by how the neocortex
operates and proposes an online spatiotemporal sparse-distributed representational neural learning structure. This
paper also presents illustrative examples using simple tasks on how to use HTM-MAT. It also compares its
prediction performance with results reported using similar applications employing state-of-the-art algorithms on
three historical benchmark datasets. The results of simulations show that HTM-MAT is indeed capable of precise
word/numeric predictions and can give better results than a popular sequence learning application.
References
[1]. Huang, Yanping, Rajesh PN Rao. "Predictive coding." Wiley Interdisciplinary Reviews: Cognitive Science 2, 2011;
5: 580-593.
[2]. Hawkins, J., Ahmad, S., Dubinsky, D. Hierarchical temporal memory including HTM cortical learning algorithms,
Technical Report, v0.2.1; 2011.
https://www.numenta.org/resources/HTM_CorticalLearningAlgorithms.pdf
[3]. Boone, A. R., Karnowski, T. P., Chaum, E., Giancardo, L., Li, Y., Tobin, K. W. Image processing and hierarchical
temporal memories for automated retina analysis. In Biomedical Sciences and Engineering Conference (BSEC),
2010. IEEE.
[4]. Rozado, D., Rodriguez, F. B., Varona, P. Optimizing hierarchical temporal memory for multivariable time series.
In International Conference on Artificial Neural Networks, Springer Berlin Heidelberg, 2010, September; 506518.
[5]. Rozado, D., Rodriguez, F. B., Varona, P. Gaze gesture recognition with hierarchical temporal memory networks.
In International Work-Conference on Artificial Neural Networks Springer Berlin Heidelberg, 2011, June; 1-8.
[6]. Rodriguez-Cobo, L., Ruiz-Lombera, R., Conde, O. M., López-Higuera, J. M., Cobo, A., Mirapeix, J. Feasibility study
of Hierarchical Temporal Memories applied to welding diagnostics. Sensors and Actuators A: Physical, 2013;
204: 58-66.
[7]. Hawkins, J., Blakeslee, S. On intelligence. Macmillan; 2007.
[8]. Hawkins, J., Ahmad, S. Why neurons have thousands of synapses, a theory of sequence memory in neocortex.
Frontiers in neural circuits, 2016; 10(23):1-13.
[9]. Cui, Y., Ahmad, S., Hawkins, J. Continuous online sequence learning with an unsupervised neural network
model. Neural Computation, 2016; 28(11): 2474-2504.
[10].
Ahmad, S., Hawkins, J. Properties of Sparse Distributed Representations and their Application to
Hierarchical Temporal Memory. Neurons and Cognition, arXiv: 150307469 [q-bio.NC], 2015.
[11].
Osegi, E.N., Anireh, V.I. HTM-MAT, https://www.mathworks.com/matlabcentral/fileexchange/51968
(April. 2016).
[12].
Liang, N. Y., Huang, G. B., Saratchandran, P., Sundararajan, N. A fast and accurate online sequential
learning algorithm for feedforward networks. IEEE Transactions on neural networks, 2006; 17(6):1411-1423.
8
[13].
Zhu, Q.Y., Huang, G.B. OS-ELM, http://www.ntu.edu.sg/home/egbhuang/elm_codes.html (April. 2004).
[14].
Hochreiter, S., Schmidhuber, J. Long short-term memory. Neural computation, 1997; 9(8): 1735-1780.
[15].
Palangi, H., Deng, L., Shen, Y., Gao, J., He, X., Chen, J., Ward, R. Deep sentence embedding using long
short-term memory networks: Analysis and application to information retrieval. IEEE/ACM Transactions on
Audio, Speech and Language Processing (TASLP), 2016; 24(4): 694-707.
[16].
Sundermeyer, M., Schlüter, R., Ney, H. LSTM Neural Networks for Language Modeling. In Interspeech,
2012, September; 194-197.
[17].
Kim, Y., Jernite, Y., Sontag, D., Rush, A. M. Character-Aware Neural Language Models. In Thirtieth AAAI
Conference on Artificial Intelligence, 2016, March; 2741-2749
[18].
Sutskever, I., Vinyals, O., Le, Q. V. Sequence to sequence learning with neural networks. In Advances in
neural information processing systems, 2014; 3104-3112.
[19].
Graves, A., Jaitly, N., Mohamed, A. R. Hybrid speech recognition with deep bidirectional LSTM. In
Automatic Speech Recognition and Understanding (ASRU), 2013 IEEE Workshop on, 2013, December; 273-278.
[20].
Graves, A., Mohamed, A. R., Hinton, G. Speech recognition with deep recurrent neural networks. In 2013
IEEE international conference on acoustics, speech and signal processing, 2013, May; 6645-6649.
[21].
Karpathy A. Recurrentjs, https://github.com/karpathy/recurrentjs (May. 2015).
[22].
Huang, G. B., Zhu, Q. Y., Siew, C. K. Extreme learning machine: a new learning scheme of feedforward
neural networks. In Neural Networks, 2004. Proceedings. 2004 IEEE International Joint Conference on, 2004,
July; 2: 985-990.
[23].
[dataset] UCI machine learning repository; http://archive.ics.uci.edu/ml/. [accessed 04:04:2017].
9
| 9cs.NE
|
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
1
Push and Pull Search for Solving Constrained
Multi-objective Optimization Problems
arXiv:1709.05915v1 [cs.NE] 15 Sep 2017
Zhun Fan, Senior Member, IEEE, Wenji Li, Xinye Cai, Hui Li, Caimin Wei, Qingfu Zhang, Fellow, IEEE,
Kalyanmoy Deb, Fellow, IEEE, and Erik D. Goodman
Abstract—This paper proposes a push and pull search (PPS)
framework for solving constrained multi-objective optimization
problems (CMOPs). To be more specific, the proposed PPS
divides the search process into two different stages, including
the push and pull search stages. In the push stage, a multiobjective evolutionary algorithm (MOEA) is adopted to explore
the search space without considering any constraints, which can
help to get across infeasible regions very fast and approach
the unconstrained Pareto front. Furthermore, the landscape of
CMOPs with constraints can be probed and estimated in the push
stage, which can be utilized to conduct the parameters setting for
constraint-handling approaches applied in the pull stage. Then,
a constrained multi-objective evolutionary algorithm (CMOEA)
equipped with an improved epsilon constraint-handling is applied
to pull the infeasible individuals achieved in the push stage to
the feasible and non-dominated regions. Compared with other
CMOEAs, the proposed PPS method can more efficiently get
across infeasible regions and converge to the feasible and nondominated regions by applying push and pull search strategies
at different stages. To evaluate the performance regarding convergence and diversity, a set of benchmark CMOPs is used to
test the proposed PPS and compare with other five CMOEAs,
including MOEA/D-CDP, MOEA/D-SR, C-MOEA/D, MOEA/DEpsilon and MOEA/D-IEpsilon. The comprehensive experimental
results demonstrate that the proposed PPS achieves significantly
better or competitive performance than the other five CMOEAs
on most of the benchmark set.
Index Terms—Push and Pull Search, Constraint-handling
Mechanisms, Constrained Multi-objective Evolutionary Algorithms.
I. I NTRODUCTION
ANY real-world optimization problems can be summarized as optimizing a number of conflicting objectives simultaneously with a set of equality and/or inequality
constraints, which are denoted as constrained multi-objective
optimization problems (CMOPs). Without lose of generality,
a CMOP considered in this paper can be defined as follows
[1]:
M
minimize F(x) = ( f1 (x), . . . , fm (x))T
subject to gi (x) ≥ 0, i = 1, . . . , q
h j (x) = 0, j = 1, . . . , p
x
∈ Rn
(1)
where F(x) = ( f1 (x), f2 (x), . . . , fm (x))T is an m-dimensional
objective vector, and F(x) ∈ Rm . gi (x) ≥ 0 is an inequality
constraint, and q is the number of inequality constraints.
h j (x) = 0 is an equality constraint, and p represents the number
of equality constraints. x ∈ Rn is an n-dimensional decision
vector.
When solving CMOPs with both inequality and equality
constraints, we usually convert the equality constraints into
inequality ones by introducing an extremely small positive
number . The detailed transformation is given as follows:
h j (x)0 ≡ − |h j (x)| ≥ 0
(2)
To deal with a set of constraints in CMOPs, the overall constraint violation is a widely used approach, which summarizes
them into a single scalar, as follows:
φ(x) =
q
Õ
i=1
| min(gi (x), 0)| +
p
Õ
| min(h j (x)0, 0)|
(3)
j=1
Given a solution xk ∈ Rn , if φ(xk ) = 0, xk is a feasible
solution. All the feasible solutions constitute a feasible region
S, which is defined as S = {x|φ(x) = 0, x ∈ Rn }. For any two
solutions xa, xb ∈ S, xa is said to dominate xb if fi (xa ) ≤
fi (xb ) for each i ∈ {1, ..., m} and F(xa ) , F(xb ), denoted as
xa xb . If there is no other solution in S dominating solution
x∗ , then x∗ is called a Pareto optimal solution. All of the Pareto
optimal solutions constitute a Pareto optimal set (PS). The
mapping of the PS in the objective space is called a Pareto
optimal front (PF), which is defined as PF = {F(x)|x ∈ PS}.
A key issue in CMOEAs is to maintain a balance the
objectives and constraints. In fact, most constraint-handling
mechanisms in evolutionary computation are applied for this
purpose.
For example, the penalty function approach adopts a penalty
factor λ to maintain the balance between the objectives and
constraints. It converts a CMOP into an unconstrained MOP
by adding the overall constraint violation multiplied by a
predefined penalty factor λ to each objective [2]. In the case of
λ = ∞, it is called a death penalty approach [3], which means
that infeasible solutions are totally unacceptable. If λ is a static
value during the search process, it is called a static penalty
approach [4]. If λ is changing during the search process, it
is called a dynamic penalty approach [5]. In the case of λ
is changing according to the information collected from the
search process, it is called an adaptive penalty approach [6]–
[9].
In oder to avoid tuning the penalty factors, another type of
constraint-handling methods is suggested, which compares the
objectives and constraints separately. Representative examples
include constraint dominance principle (CDP) [10], epsilon
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
constraint-handling method (EC) [11], stochastic ranking approach (SR) [12], and so on.
In CDP [10], three basic rules are adopted to compare any
two solutions. In the first rule, given two solutions xi, x j ∈ Rn ,
if xi is feasible and x j is infeasible, xi is better than x j .
In the case of xi and x j are both infeasible, the one with
a smaller constraint violation is better. In the last rule, xi
and x j are both feasible, and the one dominating the other
is better. CDP is a popular constraint-handling method, as
it is simple and has no extra parameters. However, it is
not suitable for solving CMOPs with very small and narrow
feasible regions [13]. For many generations, most or even
all solutions in the working population are infeasible when
solving CMOPs with this property. In addition, the diversity of
the working population can hardly be well maintained, because
the selection of solutions is only based on the constraint
violations according to the second rule of CDP.
In order to solve CMOPs with small and narrow feasible
regions, the epsilon constraint-handling (EC) [11] approach
is suggested. It is similar to CDP except for the relaxation
of the constraints. In EC, the relaxation of the constraints is
controlled by the epsilon level ε, which can help to maintain
the diversity of the working population in the case when most
solutions are infeasible. To be more specific, if the overall
constraint violation of a solution is less than ε, this solution is
deemed as feasible. The epsilon level ε is a critical parameter
in EC. In the case of ε = 0, EC is the same as CDP. Although
EC can be used to solve CMOPs with small feasible regions,
controlling the value of ε properly is non-trivial at all.
Both CDP [10] and EC [11] first compare the constrains,
then compare the objectives. SR [12] is different from CDP
and EC in terms of the order of comparison. It adopts a
probability parameter p f ∈ [0, 1] to decide if the comparison
is based on objectives or constraints. For any two solutions,
if a random number is less than p f , the one with the nondominated objectives is better. The comparison is based on
objectives. On the contrary, if the random number is greater
than p f , the comparison is first based on the constraints, then
on the objectives, which is the same as that of CDP. In the
case of p f = 0, SR is equivalent to CDP.
In recent years, many work has been done in the field
of many-objective evolutionary algorithms (MaOEAs) [14],
which gives us a new way to solve CMOPs. In order to balance
the constraints and the objectives, some researchers adopt
multi-objective evolutionary algorithms (MOEAs) or MaOEAs
(in the case of the number of objectives is greater than three)
to deal with constraints [15]. For an M-objective CMOP, its
constraints can be converted into one or k extra objectives.
Then the M-objective CMOP is transformed into an (M + 1)or (M + k)-objective unconstrained MOP, which can be solved
by MOEAs or MaOEAs. Representative examples include Cai
and WangâĂŹs Method (CW) [16] and the in-feasibility driven
evolutionary algorithms (IDEA) [17].
To maintain a good balance between the constraints and
the objectives, some researchers combine several constrainthandling mechanisms, which can be further divided into two
categories, including adopting different constraint-handling
mechanisms in different evolutionary stages or in different sub-
2
problems. For example, the adaptive trade-off model (ATM)
[18] uses two different constraint-handling mechanisms, including multi-objective approach and the adaptive penalty
functions in different evolutionary stages. The ensemble of
constraint-handling methods (ECHM) [19] uses three different
constraint-handling techniques, including epsilon constrainthandling (EC) [11], self-adaptive penalty functions (SP) [9]
and superiority of feasible solutions (SF) [20]. Three subpopulations are generated in ECHM, and each sub-population
uses a different constraint-handling method.
In this paper, we propose a biphasic CMOEA, namely
push and pull search (PPS), to balance the objectives and the
constraints. Unlike the above-mentioned constraint-handling
methods, the PPS divide the search process into two different
stages. In the first stage, only the objectives are optimized,
which means the working population is pushed toward the
unconstrained PF without considering any constraints. Furthermore, the landscape of constraints in CMOPs can be estimated
in the push stage, which can be applied to conduct the
parameters setting of constraint-handling approaches applied
in the pull stage. In the pull stage, an improved EC [11]
constraint-handling approach is adopted to pull the working
population to the constrained PF. In summary, it provides a
new framework and has the following potential advantages.
1) It has the ability to get across large infeasible regions
in front of the constrained PF. Since the constraints
are ignored in the push stage, the infeasible regions in
front of the real PF present no barriers for the working
population.
2) It facilitates the parameters setting in the constrainthandling methods. Since the landscape of constraints
has already been explored by the push process, a lots
of information can be discovered and gathered to guide
the search and parameter setting in the pull stage.
The rest of the paper is organized as follows. Section II
introduces the general idea of PPS. Section III gives an instantiation of PPS in the framework of MOEA/D, called PPSMOEA/D. Section IV designs a set of experiments to compare the proposed PPS-MOEA/D with five other CMOEAs,
including MOEA/D-IEpsilon, MOEA/D-Epsilon, MOEA/DSR, MOEA/D-CDP and C-MOEA/D. Finally, conclusions are
made in section V.
II. T HE G ENERAL F RAMEWORK OF P USH AND P ULL
S EARCH
Constraints define the infeasible regions in the decision
space, which may also affect the PFs of CMOPs in the
objective space. The influence of the infeasible regions on the
PFs can be generally classified into four different situations.
For each situation, the search behavior of PPS is illustrated by
Fig. 1-4 respectively, which can be summarized as follows.
1) Infeasible regions block the way towards the PF, as
illustrated by Fig. 1(a). In this circumstance, the unconstrained PF is the same as the constrained PF, and
PPS has significant advantages compared with other
CMOEAs. Since the constraints are ignored in the push
stage of PPS, the infeasible regions have no effect
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
3
on the searching of PPS. Fig. 1(a)-(e) show the push
process in different stages, where the working population
crosses the infeasible regions in this case without any
extra efforts. As the constrained PF is the same as
the unconstrained PF, the true PF has already been
approximated by the working population in the push
process, the pull search has no affect on the working
population, as shown in Fig. 1(f).
2) The unconstrained PF is covered by infeasible regions
and becomes no more feasible. Every constrained Pareto
optimal point lies on some constraint boundaries, as
illustrated by Fig. 2(a). In this circumstance, PPS first
approaches the unconstrained PF by using the push
strategy as illustrated by Fig. 2(a)-(c). After the working
population approaches the unconstrained PF, the pull
strategy is applied to pull the working population towards the real PF, as illustrated by Fig. 2(d)-(f).
3) Infeasible regions make the original unconstrained PF
partially feasible, as illustrated by Fig. 3(a). In this
situation, some parts of the true PF have already been
achieved during the push search as illustrated by Fig.
3(c). In the pull stage, infeasible solutions are pulled to
the feasible and non-dominated regions as illustrated by
Fig. 3(d)-(f). Finally, the whole real PF has been found
by PPS.
4) Infeasible regions may reduce the dimensionality of the
PF of a CMOP, as illustrated by Fig. 4(a). In the push
stage, a set of non-dominated solutions without considering constraints are achieved as illustrated by Fig. 4(a)-(c)
. Then, infeasible solutions in the working population are
pulled toward the feasible and non-dominated regions in
the pull search process, as shown in Fig. 4(d)-(f).
From the above analysis, it can be observed that PPS
can deal with CMOPs with all situations of PFs affected by
constraints.
The main steps of PPS includes the push and pull search
processes. However, when to apply the push or pull search
is very critical. A strategy to switch the search behavior is
suggested as follows.
rk ≡ max{r zk , rnk } ≤
(4)
where rk represents the max rate of change between the ideal
and nadir points during the last l-th generation. is a userdefined parameter, here we set = 1e − 3. The rates of change
of the ideal and nadir points during the last l-th generation are
defined in Eq. 5 and Eq. 6 respectively.
r zk = max {
i=1,...,m
rnk = max {
|zik − zik−l |
max{|zik−l |, 1e − 6}
|nik
−
nik−l |
i=1,...,m max{|nk−l |, 1e
i
− 6}
}
k−l ), nk−l = (nk−l, . . . , nk−l ) are the ideal and nadir
(z1k−l, . . . , zm
m
1
points in the (k − l)-th generation.
At the beginning of the search, rk is initialized to 1.0. At
each generation, rk is updated according to Eq. 4. If rk is less
or equal then the predefined threshold , the search behavior
is switched to the pull search.
To summarize, PPS divides its search process into two
different stages, including the push search and pull search. At
the first stage, the push search without considering constraints
is adopted to approximate the unconstrained PF. Once Eq.
4 is satisfied, the pull search is used to pull the infeasible
solutions to the feasible and non-dominated regions. In this
situation, constraints are considered in the pull search process.
PPS terminates when a predefined halting condition is met. In
the following section, we will describe the instantiation of the
push and pull strategy in a MOEA/D framework.
III. A N I NSTANTIATION OF PPS IN MOEA/D
In this section, the details of an instantiation of the push
search method, as well as the method adopted in the pull
search, and the proposed PPS in the framework of MOEA/D
are described.
A. The push search
In the push search stage, an unconstrained MOEA/D is used
to search for non-dominated solutions without considering any
constraints. When solving a CMOP by using MOEA/D, we
can decompose the CMOP into a set of single constrained
optimization subproblems, and optimize them simultaneously
in a collaborative way. Each subproblem is associated with
a decomposition function by using a weight vector λi . In
the decomposition-based selection approach, an individual
is selected into next generation based on the value of the
decomposition function.
There are three popular decomposition approaches, including weighted sum [21], Tchebycheff [21] and boundary intersection approaches [22]. In this paper, we adopt the Tchebycheff decomposition method, with the detailed definition given
as follows.
1
(| f j (x) − z ∗j |)
(7)
j=1,...,m λ i
j
Í
where λi is a weight vector, and j=1,...,m λij = 1, λij ≥ 0. z ∗
is the ideal points as described in section II.
In the push search stage, a newly generated solution x is kept
to the next generation based on the value of g te as described
in Algorithm 1.
g te (x, λi, z∗ ) = max
(5)
B. The pull search
}
(6)
k ), nk
k ) are the
where z k = (z1k , . . . , zm
= (n1k , . . . , nm
ideal and nadir points in the k-th generation, and
zik = min j=1,..., N fi (x j ), nik = max j=1,..., N fi (x j ). z k−l =
In this process, some infeasible solutions are pulled to
the feasible and non-dominated regions. To achieve this, a
constraint-handling mechanism is adopted to punish the infeasible solutions in the pull search stage. An improved epsilon
constraint-handling to deal with constraints is proposed with
the detailed formulation given as follows.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
f2
4
f2
Individuals
Infeasible regions
PF
f2
Individuals
Infeasible regions
PF
PF
PF
PF
PF
f1
f1
(a)
f2
Individuals
Infeasible regions
f1
(b)
f2
Individuals
Infeasible regions
PF
(c)
f2
Individuals
Infeasible regions
PF
PF
Individuals
Infeasible regions
PF
PF
PF
f1
f1
(d)
f1
(e)
(f)
Fig. 1. Infeasible regions block the way towards the PF, and the unconstrained PF is the same as the constrained PF. (a)-(e) show the different stages of the
push search process, and the working population can get across the infeasible regions without any extra efforts dealing with constraints. (f) shows the pull
search process, which is the same as (e) in this particular case, since the real PF is the same as the unconstrained PF, and has already been achieved by the
working population in the push search process.
PF with constraints
Individuals
Infeasible regions
PF without constraints
f2
PF with constraints
Individuals
Infeasible regions
PF without constraints
f2
f1
f2
(c)
PF with constraints
Individuals
Infeasible regions
PF without constraints
f2
PF with constraints
Individuals
Infeasible regions
PF without constraints
f1
f1
f1
(d)
f1
(b)
PF with constraints
Individuals
Infeasible regions
PF without constraints
PF with constraints
Individuals
Infeasible regions
PF without constraints
f1
(a)
f2
f2
(e)
(f)
Fig. 2. The unconstrained PF is covered by infeasible regions and becomes no more feasible. The real PF lies on some boundaries of constraints. (a)-(c)
show the push search process, and the working population can get across the infeasible regions without any barriers. (e)-(f) show the pull search process, the
infeasible solutions in the working population are pulled to the feasible and non-dominated regions gradually.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
PF with constraints
f2
5
PF with constraints
f2
Individuals
Infeasible regions
PF without constraints
Individuals
Infeasible regions
PF without constraints
PF
PF without constraints
PF
PF
f1
f1
(a)
f1
(b)
PF with constraints
f2
PF with constraints
f2
Individuals
Infeasible regions
PF with constraints
f2
Individuals
Infeasible regions
(c)
PF without constraints
PF with constraints
f2
Individuals
Infeasible regions
Individuals
Infeasible regions
PF without constraints
PF
PF without constraints
PF
PF
f1
f1
(d)
f1
(e)
(f)
Fig. 3. Infeasible regions make the original unconstrained PF partially feasible. (a)-(c) show the push search process, and some parts of the real PF have
been found in this process. In the pull stage, infeasible solutions have been pulled to the feasible and non-dominated regions gradually as shown in (d)-(f).
f2
f2
PF
Individuals
Feasible regions
PF without constraints
f1
PF without constraints
f1
f1
(b)
f2
PF
Individuals
Feasible regions
(c)
PF
PF
f2
Individuals
Feasible regions
PF without constraints
Individuals
Feasible regions
PF without constraints
f1
(d)
Individuals
Feasible regions
PF without constraints
(a)
f2
PF
f2
PF
Individuals
Feasible regions
PF without constraints
f1
f1
(e)
(f)
Fig. 4. Infeasible regions reduce the dimensionality of the PF. (a)-(c) show the push search process of PPS, and a set of non-dominated solutions are achieved
without considering any constraints. (d)-(f) show the pull search process of PPS, many infeasible solutions are pulled to the feasible and non-dominated
regions, which reside in a reduced dimensionality of the space.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
6
Algorithm 1: Push Subproblem
1
2
3
4
5
6
7
8
Algorithm 2: Pull Subproblem
PushSubproblems(x j ,y j )
Function result =
result = f alse
if g te (yi |λ j , z∗ ) ≤ g te (x j |λ j , z ∗ ) then
xj = yj
result = ture
end
return result
end
1
2
3
4
5
6
7
8
9
10
11
(
ε(k) =
12
(1 − τ)ε(k − 1), if r fk < α
ε(0)(1 − Tkc )cp, if r fk ≥ α
(8)
where r fk is the ratio of feasible solutions in the k-th generation. τ is to control the speed of reducing the relaxation
of constraints in the case of r fk < α, and τ ∈ [0, 1]. α is
to control the searching preference between the feasible and
infeasible regions, and α ∈ [0, 1]. cp is to control the speed of
reducing relaxation of constraints in the case of r fk ≥ α. ε(k)
is updated until the generation counter k reaches the control
generation Tc . ε(0) is set to the maximum overall constraint
violation of the working population at the end of the push
search.
The above proposed epsilon setting method in Eq. 8 is
different from our previous work [23] and [24]. In this paper,
ε(0) is set to the maximum overall constraint violation of
the working population at the end of the push search, while
ε(0) is set to the overall constraint violation of the top θ-th
individual in the initial population in [23] and [24]. In the case
of r f0 == 0, ε(0) = 0 according to the epsilon setting in [23]
and [24], which can not help to explore the infeasible regions
and get across large infeasible regions. To avoid this situation,
ε(k) is increased if r fk is greater than a predefined threshold
in [23] and [24]. However, increasing ε(k) properly is still an
issue that needs to be further investigated. In this paper, ε(k)
does not need to be increased to get across large infeasible
regions, because the working population has already gotten
across large infeasible regions in the push stage.
In the pull stage, a newly generated solution x is selected
into the next generation based on the value of g te and the
overall constraint violation φ(x) as illustrated by Algorithm 2.
C. PPS Embedded in MOEA/D
Algorithm 3 outlines the psuecode of PPS-MOEA/D, a
PPS embedded in MOEA/D. A CMOP is decomposed into
N single objective subproblems, and these subproblems are
initialized at line 1. The max rate of change of ideal and
nadir points is initialized to 1.0, and the flag of search stage
PushStage = true. Then, the algorithm repeats lines 3-27
until Tmax generations have been reached. The value of ε(k)
and the search strategy are set at lines 4-14 based on Eq.
4 and Eq. 8. ε(k) and ε(0) are initialized at line 7, and
the function MaxViolation() calculates the maximum overall
constraint violation in the current working population. Lines
13
14
15
16
17
18
Function result = PullSubproblems(x j ,y j ,ε(k))
result = f alse
if φ(y j ) ≤ ε(k) and φ(x j ) ≤ ε(k) then
if g te (yi |λ j , z ∗ ) ≤ g te (x j |λ j , z ∗ ) then
xj = yj
result = ture
end
else if φ(y j ) == φ(x j ) then
if g te (y j |λ j , z ∗ ) ≤ g te (x j |λ j , z ∗ ) then
xj = yj
result = ture
end
else if φ(y j ) < φ(x j ) then
xj = yj
result = ture
end
return result
end
15-19 show the process of generating a new solution. The
updating of the ideal points are illustrated by lines 20-22.
Lines 23-33 show the updating of subproblems, and different
search strategies are adopted according to the state of the
searching. In the case of PushStage == true, the push search
is adopted, otherwise, the pull search is used. The generation
count k is updated at line 35. The set of feasible and nondominated solutions N S is updated according to the nondominated ranking in NSGA-II [10] at line 36.
Remark: The proposed PPS is a general framework for
solving CMOPs. It can be instantiated in many different
MOEAs, even though only PPS-MOEA/D is realized in this
paper. At each search stage, many information can be gathered
to extract useful knowledge which can be used to conduct the
searching in both search stages. In fact, knowledge discovery
is a critical step in the PPS framework. In this paper, we only
adopts some statistic information. For example, the maximum
overall constraint violation at the end of the push search stage
is adopted to set the value of ε(0). The ratio of feasible
solutions at the pull search stage is used to control the value of
ε(k). In fact, many data mining methods and machine learning
approaches can be integrated in the PPS framework for solving
CMOPs more efficiently.
IV. E XPERIMENTAL STUDY
A. Experimental Settings
To evaluate the performance of the proposed PPS
method, the other five CMOEAs, including MOEA/DIEpsilon, MOEA/D-Epsilon, MOEA/D-SR, MOEA/D-CDP
and C-MOEA/D, are tested on LIR-CMOP1-14 [24]. The
detailed parameters are listed as follows:
1) The mutation probability Pm = 1/n (n denotes the
dimension of a decision vector). The distribution index
in the polynomial mutation is set to 20.
2) DE parameters: CR = 1.0, f = 0.5.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
Algorithm 3: PPS-MOEA/D
Input: N: the number of subproblems. Tmax : the
maximum generation. N weight vectors:
λ1, . . . , λ N . T: the size of the neighborhood. δ:
the selecting probability from neighbors. nr : the
maximal number of solutions replaced by a child.
Output: N S : a set of feasible non-dominated solutions
1 Decompose a CMOP into N subproblems associated with
λ1, . . . , λ N . Generate an initial population
P = {x1, . . . , x N }. For each i = 1, . . . , N, set
B(i) = {i1, . . . , iT }, where λi1 , . . . , λiT are the T closest
weight vectors to λi .
2 Initialize the rate of change of ideal and nadir points
(maxChange = 1.0) over the previous L generations.
Set PushStage = true, k = 1.
3 while k ≤ Tmax do
4
if k >= L then maxChange = CalcMaxChange(k)
according to Eq. 4 ;
5
if k < Tc then
6
if maxChange ≤ α And PushStage == true then
7
PushStage = f alse; ε(k) = ε(0) =
MaxViolation();
8
end
9
if PushStage == f alse then
10
Update ε(k) according to Eq. 8;
11
end
12
else
13
ε(k) = 0;
14
end
15
Generate a random permutation r p from {1, . . . , N }.
16
for i ← 1 to N do
17
Generate a random number r ∈ [0, 1], j = r p(i).
18
if r < δ then S = B( j) else S = {1, . . . , N } ;
19
Generate y j through the DE operator, and
perform polynomial mutation on y j .
20
for t ← 1 to m do
21
if zt∗ > ft (y j ) then zt∗ = ft (y j ) ;
22
end
23
Set c = 0.
24
while c , nr or S , do
25
select an index j from S randomly.
26
if PushStage == true then
27
result = PushSubproblems(x j , y j );//
Algorithm 1;
28
else
29
result = PullSubproblems(x j , y j , ε(k)); //
Algorithm 2;
30
end
31
if result == true then c = c + 1;
32
S = S\{ j};
33
end
34
end
35
k = k + 1;
Ð
36
N S = NDSelect(N S P); // According to NSGA-II
[10];
37 end
7
3) Population size: N = 300. Neighborhood size: T = 30.
4) Halting condition: each algorithm runs for 30 times independently, and stops until 300,000 function evaluations
are reached.
5) Probability of selecting individuals from its neighborhood: δ = 0.9.
6) The max number of solutions updated by a child: nr = 2.
7) Parameters setting in PPS-MOEA/D: Tc = 800, α =
0.95, τ = 0.1, cp = 2, l = 20.
8) Parameters setting in MOEA/D-IEpsilon: Tc = 800, α =
0.95, τ = 0.1, and θ = 0.05N.
9) Parameters setting in MOEA/D-Epsilon: Tc = 800, cp =
2, and θ = 0.05N.
10) Parameter setting in MOEA/D-SR: Sr = 0.05.
B. Performance Metric
To measure the performance of PPS-MOEA/D, MOEA/DIEpsilon, MOEA/D-Epsilon, MOEA/D-CDP, MOEA/D-SR
and C-MOEA/D, two popular metrics–the inverted generation
distance (IGD) [25] and the hypervolume [26] are adopted.
•
Inverted Generational Distance (IGD):
The IGD metric reflects the performance regarding convergence and diversity simultaneously. The detailed definition is
given as follows:
Í
IGD(P∗, A) =
y∗ ∈ P∗
d(y ∗, A)
| P∗ |
(9)
qÍ
m
∗
2
d(y ∗, A) = min{
i=1 (yi − yi ) }
y∈A
where P∗ denotes a set of representative solutions in the
real PF, A is an approximate PF achieved by a CMOEA. m
denotes the number of objectives. For LIR-CMOPs with two
objectives, 1000 points are sampled uniformly from the true PF
to construct P∗ . For LIR-CMOPs with three objectives, 10000
points are sampled uniformly from the PF to constitute P∗ . It
is worth noting that a smaller value of IGD may indicate better
performance with regards to diversity and/or convergence.
•
Hypervolume (HV):
HV reflects the closeness of the set of non-dominated solutions
achieved by a CMOEA to the real PF. The larger HV means
that the corresponding non-dominated set is closer to the true
PF.
HV(S) = VOL(
Ø
r
[ f1 (x), z1r ] × ...[ fm (x), zm
])
(10)
x ∈S
where VOL(·) is the Lebesgue measure, m denotes the number
r )T is a user-defined reference
of objectives, zr = (z1r , ..., zm
point in the objective space. For each LIR-CMOP, the reference point is placed at 1.2 times the distance to the nadir
point of the true PF. It is worth noting that a larger value of
HV may indicate better performance regarding diversity and/or
convergence.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
C. Discussion of Experiments
1) A Comparison between PPS-MOEA/D and MOEA/DIEpsilon: The difference between PPS-MOEA/D and
MOEA/D-IEpsilon is that MOEA/D-IEpsilon does not apply
the push search without considering any constraints, and they
adopt different epsilon setting methods. To illustrate the effect
of the PPS framework, the performance regarding IGD and
HV metrics of PPS-MOEA/D and MOEA/D-IEpsilon on LIRCMOP1-14 is compared.
The statistical results of the IGD values on LIR-CMOP1-14
achieved by PPS-MOEA/D and MOEA/D-IEpsilon in 30 independent runs are listed in Table I. For LIR-CMOP1-2, LIRCMOP5-8, LIR-CMOP11 and LIR-CMOP14, PPS-MOEA/D
performs significantly better than MOEA/D-IEpsilon. For LIRCMOP9, MOEA/D-IEpsilon is significantly better than PPSMOEA/D. For the rest of test instances, MOEA/D-IEpsilon
and PPS-MOEA/D have no significant difference.
The statistical results of the HV values on LIR-CMOP114 achieved by PPS-MOEA/D and MOEA/D-IEpsilon in 30
independent runs are listed in Table II. For LIR-CMOP1-3,
LIR-CMOP5-8 and LIR-CMOP14, PPS-MOEA/D is significantly better than MOEA/D-IEpsilon. There is no significantly
difference between PPS-MOEA/D and MOEA/D-IEpsilon on
LIR-CMOP4 and LIR-CMOP10-13. For LIR-CMOP9, PPS is
significantly worse than MOEA/D-IEpsilon.
Since the real PF of LIR-CMOP9 locates on its unconstrained PF and has several disconnected parts, as illustrated
by Fig. 5(a), PPS-MOEA/D should performs well on this
problems. However, the statistic results of both IGD and
HV indicate that PPS-MOEA/D is significantly worse than
MOEA/D-IEpsilon. One possible reason is that PPS-MOEA/D
has not converged to the whole unconstrained PF in the push
search stage.
To verify this hypothesis, we adjust the value of l in Eq.
(5) and Eq. (6), which is a critical parameter to switch the
search behavior. The mean values of IGD and HV values of
MOEA/D-IEpsilon and PPS-MOEA/D with different values of
l are shown in Fig. 5(b) and Fig. 5(c), respectively. We can
observe that, in cases of l ≥ 70, PPS-MOEA/D is better than
MOEA/D-IEpsilon.
To summarize, PPS-MOEA/D is superior to MOEA/DIEpsilon on eight test problems, is not significantly different
from MOEA/D-IEpsilon on five test problems, is inferior
to MOEA/D-IEpsilon on only one test instance. It can be
concluded that the PPS framework can help to enhance the
performance of a CMOEA in terms of the IGD and HV
metrics.
2) A Comparison of PPS-MOEA/D and the other four
decomposition-based CMOEAs: The statistical results of the
IGD values on LIR-CMOP1-14 achieved by PPS-MOEA/D
and the other four decomposition-based CMOEAs in 30 independent runs are listed in Table I. According to the WilcoxonTest in this table, it is clear that PPS-MOEA/D is significantly
better than MOEA/D-Epsilon, MOEA/D-CDP, MOEA/D-SR,
C-MOEA/D on all of the fourteen tested problems in terms of
the IGD metric.
The statistical results of the HV values on LIR-CMOP1-14
achieved by PPS-MOEA/D and the other four decomposition-
8
based CMOEAs in 30 independent runs are listed in Table II.
It can be observed that PPS-MOEA/D is significantly better
than MOEA/D-Epsilon, MOEA/D-CDP, MOEA/D-SR and CMOEA/D on all the test instances in terms of the HV metric.
It is clear that the proposed PPS-MOEA/D performs better
than the other four decomposition-based CMOEAs on all of
the test instances.
3) The advantages of PPS-MOEA/D: The box plots of IGD
values for the six tested CMOEAs on LIR-CMOP6, LIRCMOP7 and LIR-CMOP11 are shown in Fig. 6. It is clear that
the mean IGD value of PPS-MOEA/D is the smallest among
the six tested CMOEAs. To further illustrate the advantages of
the proposed PPS-MOEA/D, the populations achieved by each
tested CMOEAs on LIR-CMOP6, LIR-CMOP7, and LIRCMOP11 during the 30 independent runs are plotted in Fig.
8, Fig. 9 and Fig. 10, respectively.
For LIR-CMOP6, there are two large infeasible regions in
front of the PF, and the unconstrained PF is the same as the
constrained PF, as illustrated by Fig. 7(a). In Fig. 8, we can
observe that PPS-MOEA/D can get across the large infeasible
regions at each running time, while the other five CMOEAs
are sometimes trapped into local optimal.
For LIR-CMOP7, there are three large infeasible regions,
and the unconstrained PF is covered by infeasible regions and
becomes no more feasible as illustrated by Fig. 7(b). In Fig.
9, we can observe that PPS-MOEA/D can converge to the real
PF at each running time, while the other five CMOEAs can
not always converge to the real PF.
For LIR-CMOP11, infeasible regions make the original
unconstrained PF partially feasible, as illustrated by Fig. 7(c).
The PF of LIR-CMOP11 is disconnected and has seven Pareto
optimal solutions. PPS-MOEA/D can find the seven discrete
Pareto optimal solutions at each running time, while the other
five CMOEAs can not find all the seven Pareto optimal
solutions sometimes, as shown in Fig. 10. In Fig. 10(b), it
seems that MOEA/D-IEpsilon has found all the Pareto optimal
solutions. However, MOEA/D-IEpsilon can not find all the
Pareto optimal solutions more than fifteen out of 30 runs.
For MOEA/D-Epsilon, MOEA/D-CDP, MOEA/D-SR and CMOEA/D, some Pareto optimal solutions are not found during
the thirty independent runs, as illustrated by 10(c)-(f).
V. C ONCLUSION
This paper proposes a general PPS framework to deal
with CMOPs. More specifically, the search process of PPS
is divided into two stages, including the push and pull search
processes. At the push stage, constraints are ignored, which
can help PPS to get across infeasible regions in front of
the unconstrained PFs. Moreover, the landscape affected by
constraints can be estimated at the push stage, and this
information, such as the ration of feasible solutions and
the maximum overall constraint violation, can be applied to
conduct the settings of parameters coming from the constrainthandling mechanisms in the pull stage. When the max rate of
change between ideal and nadir points is less or equal than
a predefined threshold, PPS is switched to the pull search
process. The infeasible solutions achieved in the push stage
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
9
TABLE I
IGD RESULTS OF PPS AND THE OTHER FIVE CMOEA S ON LIR-CMOP1-14.
Test Instances
mean
LIR-CMOP1
std
mean
LIR-CMOP2
std
mean
LIR-CMOP3
std
mean
LIR-CMOP4
std
mean
LIR-CMOP5
std
mean
LIR-CMOP6
std
mean
LIR-CMOP7
std
mean
LIR-CMOP8
std
mean
LIR-CMOP9
std
mean
LIR-CMOP10
std
mean
LIR-CMOP11
std
mean
LIR-CMOP12
std
mean
LIR-CMOP13
std
mean
LIR-CMOP14
std
Wilcoxon-Test (S-D-I)
PPS-MOEA/D
6.413E-03
1.938E-03
4.673E-03
7.844E-04
8.6.0E-03
5.184E-03
4.677E-03
1.116E-03
1.837E-03
9.263E-05
2.490E-03
3.399E-04
2.797E-03
9.854E-05
2.778E-03
7.558E-05
9.940E-02
1.519E-01
2.108E-03
7.754E-05
2.832E-03
1.359E-03
2.704E-02
5.002E-02
6.455E-02
2.177E-03
6.419E-02
1.690E-03
–
MOEA/D-IEpsilon
7.972E-03†
3.551E-03
5.234E-03†
1.009E-03
1.131E-02
6.422E-03
4.848E-03
2.051E-03
2.132E-03†
3.791E-04
2.334E-01†
5.062E-01
3.735E-02†
5.414E-02
2.749E-02†
5.917E-02
4.977E-03‡
1.367E-02
2.111E-03
7.111E-05
5.809E-02†
5.793E-02
3.358E-02
5.178E-02
6.460E-02
1.639E-03
6.540E-02†
2.037E-03
8-5-1
MOEA/D-Epsilon
5.745E-02†
2.887E-02
5.395E-02†
2.128E-02
8.806E-02†
4.356E-02
6.506E-02†
3.011E-02
1.150E+00†
1.979E-01
1.269E+00†
2.947E-01
1.509E+00†
5.094E-01
1.620E+00†
3.054E-01
4.902E-01†
4.221E-02
2.132E-01†
5.315E-02
3.468E-01†
9.285E-02
2.520E-01†
8.980E-02
1.200E+00†
3.055E-01
1.021E+00†
4.855E-01
14-0-0
MOEA/D-CDP
1.107E-01†
5.039E-02
1.434E-01†
6.055E-02
2.608E-01†
4.335E-02
2.527E-01†
4.340E-02
1.045E+00†
3.632E-01
1.090E+00†
5.198E-01
1.460E+00†
6.082E-01
1.375E+00†
6.149E-01
4.810E-01†
5.244E-02
2.159E-01†
6.810E-02
3.418E-01†
9.216E-02
2.690E-01†
9.058E-02
1.210E+00†
3.172E-01
1.107E+00†
3.976E-01
14-0-0
MOEA/D-SR
1.811E-02†
1.665E-02
9.634E-03†
7.232E-03
1.775E-01†
7.196E-02
1.949E-01†
6.401E-02
1.040E+00†
3.655E-01
9.427E-01†
5.899E-01
1.079E+00†
7.581E-01
1.012E+00†
7.244E-01
4.854E-01†
4.775E-02
1.925E-01†
6.812E-02
3.157E-01†
7.491E-02
2.064E-01†
5.613E-02
8.864E-01†
5.756E-01
1.027E+00†
4.700E-01
14-0-0
C-MOEA/D
1.256E-01†
7.028E-02
1.397E-01†
5.436E-02
2.805E-01†
4.214E-02
2.590E-01†
3.513E-02
1.103E+00†
2.994E-01
1.307E+00†
2.080E-01
1.564E+00†
4.235E-01
1.579E+00†
3.706E-01
4.810E-01†
5.243E-02
2.130E-01†
4.634E-02
3.806E-01†
8.949E-02
2.6.0E-01†
9.628E-02
1.180E+00†
3.775E-01
1.250E+00†
5.298E-02
14-0-0
WilcoxonâĂŹs rank sum test at a 0.05 significance level is performed between PPS-MOEA/D and each of the other five CMOEAs. † and ‡ denote that the
performance of the corresponding algorithm is significantly worse than or better than that of PPS-MOEA/D, respectively. Where âĂŹS-D-IâĂŹ indicates
PPS-MOEA/D is superior to, not significantly different from or inferior to the corresponding compared CMOEAs.
TABLE II
HV RESULTS OF PPS-MOEA/D AND THE OTHER FIVE CMOEA S ON LIR-CMOP1-14.
Test Instances
mean
LIR-CMOP1
std
mean
LIR-CMOP2
std
mean
LIR-CMOP3
std
mean
LIR-CMOP4
std
mean
LIR-CMOP5
std
mean
LIR-CMOP6
std
mean
LIR-CMOP7
std
mean
LIR-CMOP8
std
mean
LIR-CMOP9
std
mean
LIR-CMOP10
std
mean
LIR-CMOP11
std
mean
LIR-CMOP12
std
mean
LIR-CMOP13
std
mean
LIR-CMOP14
std
Wilcoxon-Test (S-D-I)
PPS-MOEA/D
1.016E+00
1.580E-03
1.349E+00
1.010E-03
8.703E-01
2.650E-03
1.093E+00
2.467E-03
1.462E+00
2.919E-04
1.129E+00
1.771E-04
3.015E+00
2.663E-03
3.017E+00
1.139E-03
3.570E+00
2.242E-01
3.241E+00
3.077E-04
4.390E+00
2.217E-04
5.614E+00
1.6.0E-01
5.710E+00
1.275E-02
6.193E+00
1.310E-02
–
MOEA/D-IEpsilon
1.014E+00†
2.433E-03
1.348E+00†
1.318E-03
8.682E-01†
3.919E-03
1.093E+00
2.458E-03
1.461E+00†
1.332E-03
9.259E-01†
4.235E-01
2.863E+00†
1.958E-01
2.936E+00†
1.855E-01
3.707E+00‡
1.878E-02
3.241E+00
2.482E-04
4.232E+00
1.840E-01
6.093E+00
1.578E-01
5.711E+00
1.301E-02
6.182E+00†
1.093E-02
8-5-1
MOEA/D-Epsilon
9.6.0E-01†
3.280E-02
1.283E+00†
2.880E-02
7.976E-01†
3.929E-02
1.021E+00†
4.191E-02
4.297E-02†
2.353E-01
5.398E-02†
2.214E-01
3.031E-01†
9.070E-01
1.060E-01†
5.488E-01
2.737E+00†
1.483E-01
2.885E+00†
1.023E-01
3.341E+00†
2.573E-01
4.884E+00†
3.168E-01
4.6.0E-01†
1.301E+00
1.334E+00†
2.452E+00
14-0-0
MOEA/D-CDP
7.538E-01†
8.947E-02
1.061E+00††
1.078E-01
4.856E-01†
4.305E-02
7.349E-01†
5.436E-02
1.630E-01†
4.426E-01
1.880E-01†
3.873E-01
3.743E-01†
9.577E-01
5.173E-01†
1.054E+00
2.770E+00†
1.843E-01
2.879E+00†
1.364E-01
3.353E+00†
2.568E-01
4.827E+00†
3.284E-01
4.628E-01†
1.423E+00
8.813E-01†
1.969E+00
14-0-0
MOEA/D-SR
9.956E-01†
2.915E-02
1.337E+00†
1.466E-02
5.914E-01†
1.072E-01
8.147E-01†
8.699E-02
1.817E-01†
4.389E-01
3.021E-01†
4.616E-01
9.878E-01†
1.271E+00
1.099E+00†
1.198E+00
2.752E+00†
1.640E-01
2.927E+00†
1.347E-01
3.383E+00†
2.899E-01
5.032E+00 †
1.755E-01
1.892E+00†
2.572E+00
1.270E+00†
2.288E+00
14-0-0
C-MOEA/D
7.414E-01†
1.222E-01
1.070E+00†
9.099E-02
4.707E-01†
4.089E-02
7.308E-01†
5.162E-02
9.721E-02†
3.699E-01
2.328E-02†
1.275E-01
2.035E-01†
7.522E-01
1.658E-01†
6.114E-01
2.770E+00†
1.841E-01
2.888E+00†
9.765E-02
3.243E+00†
2.551E-01
4.890E+00 †
3.453E-01
6.293E-01†
1.712E+00
1.795E-01†
2.598E-01
14-0-0
WilcoxonâĂŹs rank sum test at a 0.05 significance level is performed between PPS-MOEA/D and each of the other five CMOEAs. † and ‡ denote that the
performance of the corresponding algorithm is significantly worse than or better than that of PPS-MOEA/D, respectively. Where âĂŹS-D-IâĂŹ indicates
PPS-MOEA/D is superior to, not significantly different from or inferior to the corresponding compared CMOEAs.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
3
10 0
PF without constraints
Feasible regions
Constraint boundary
2.5
10
3.7
3.6
1.5
1
10 -1
Mean of HV
Mean of IGD
f2
2
10 -2
3.5
3.4
0.5
10 -3
0
0
0.5
1
1.5
f1
2
2.5
3
3.3
1
2
3
4
5
6
7
8
9
10
11
1
2
3
4
Indexes of CMOEAs
(a) The landscape of LIR-CMOP9
5
6
7
8
9
10
11
Indexes of CMOEAs
(b) IGD
(c) HV
Fig. 5. The feasible and infeasible regions of LIR-CMOP9 are plotted in (a). In (b)-(c), the mean values of IGD and HV of MOEA/D-IEpsilon and
PPS-MOEA/D with different values of l are plotted. In the horizontal axises of (b)-(c), the index ’1’ represents MOEA/D-IEpsilon, indexes ’2-11’ denote
PPS-MOEA/D with l = 10, 20, 30, 40, 50, 60, 70, 80, 90 and 100.
0.6
2
1.5
0.5
1.5
0.4
IGD
IGD
IGD
1
1
0.3
0.2
0.5
0.5
0.1
0
0
1
2
3
4
5
6
0
1
2
(a) LIR-CMOP6
3
4
5
6
1
2
(b) LIR-CMOP7
3
4
5
6
(c) LIR-CMOP11
Fig. 6. The box plots of IGD values for the six tested CMOEAs. The horizontal axis shows the index of each tested CMOEAs, where 1, 2, 3, 4, 5 and 6
denote PPS-MOEA/D, MOEA/D-IEpsilon, MOEA/D-Epsilon, MOEA/D-CDP, MOEA/D-SR and C-MOEA/D respectively.
5
6
PF with constraints
Infeasible regions
4
3
PF without constraints
Infeasible regions
5
4
2
3
f2
f2
3
f2
PF without constraints
Feasible regions
Constraint boundary
2.5
1.5
2
1
0
0
1
2
3
4
5
2
1
1
0.5
0
0
1
2
f1
(a) LIR-CMOP6
Fig. 7.
3
f1
4
5
6
0
0
(b) LIR-CMOP7
0.5
1
1.5
f1
2
2.5
3
(c) LIR-CMOP11
Illustrations of the feasible and infeasible regions of LIR-CMOP6, LIR-CMOP7 and LIR-CMOP11.
are pulled to the feasible and non-dominated area by adopting
an improved epsilon constraint-handling technique. The value
of epsilon level can be set properly according to the maximum
overall constraint violation obtained at the end of the push
search stage. The comprehensive experiments indicate that the
proposed PPS achieves competitive or statistically significant
better results than the other five CMOEAs on most of the
benchmark problems.
It is also worthwhile to point out that there has been
very little work regarding using information of landscape
affected by constraints to solve CMOPs. In this context, the
proposed PPS provides a viable framework. Obviously, a lot
of work need to be done to improve the performance of PPS,
such as, the augmented constraint-handling mechanisms in
the pull stage, the enhanced strategies to switch the search
behavior and the data mining methods and machine learning
approaches integrated in the PPS framework. For another
future work, the proposed PPS will be implemented in the nondominated framework, such as NSGA-II, to further verify the
effect of PPS. More other CMOPs and real-world constrained
engineering optimization problems will also be used to test
the performance of the PPS embedded in different MOEA
frameworks.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
5
25
True PF
PPS-MOEA/D
100
True PF
MOEA/D-IEpsilon
20
f2(x)
3
2
1
15
10
5
0
0.5
1
1.5
0
2
60
40
20
0
5
10
f1(x)
0
15
0
2
4
f1(x)
(a) PPS-MOEA/D
60
True PF
MOEA/D-SR
True PF
C-MOEA/D
f2(x)
40
20
f2(x)
40
30
8
(c) MOEA/D-Epsilon
60
True PF
MOEA/D-CDP
40
6
f1(x)
(b) MOEA/D-IEpsilon
50
f2(x)
True PF
MOEA/D-Epsilon
80
f2(x)
4
f2(x)
11
20
20
10
0
0
1
2
3
4
0
5
0
2
4
f1(x)
3
4
5
(f) C-MOEA/D
120
True PF
PPS-MOEA/D
40
True PF
MOEA/D-IEpsilon
100
30
10
f2(x)
15
60
20
40
10
5
20
0
0.5
1
1.5
2
2.5
0
3
0
10
20
f1(x)
30
0
40
2
4
6
8
10
f1(x)
(b) MOEA/D-IEpsilon
50
(c) MOEA/D-Epsilon
60
80
True PF
MOEA/D-CDP
40
0
f1(x)
(a) PPS-MOEA/D
True PF
MOEA/D-SR
True PF
C-MOEA/D
60
f2(x)
20
f2(x)
40
30
40
20
20
10
0
True PF
MOEA/D-Epsilon
80
f2(x)
f2(x)
2
f1(x)
(e) MOEA/D-SR
20
f2(x)
1
The non-dominated solutions achieved by each algorithm on LIR-CMOP6 during the 30 independent runs are plotted in (a)-(f).
25
0
5
10
f1(x)
(d) MOEA/D-CDP
Fig. 9.
0
f1(x)
(d) MOEA/D-CDP
Fig. 8.
0
6
15
0
0
10
20
30
f1(x)
(e) MOEA/D-SR
40
50
0
0
5
10
f1(x)
(f) C-MOEA/D
The non-dominated solutions achieved by each algorithm on LIR-CMOP7 during the 30 independent runs are plotted in (a)-(f).
15
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
5
40
True PF
PPS-MOEA/D
20
10
3
2
1
0
2
4
6
0
8
True PF
MOEA/D-Epsilon
20
f2(x)
f2(x)
f2(x)
25
True PF
MOEA/D-IEpsilon
4
30
0
12
15
10
5
0
2
f1(x)
4
6
0
8
0
1
2
f1(x)
(a) PPS-MOEA/D
(b) MOEA/D-IEpsilon
40
3
(c) MOEA/D-Epsilon
8
30
True PF
MOEA/D-CDP
True PF
MOEA/D-SR
30
4
f1(x)
True PF
C-MOEA/D
6
20
f2(x)
f2(x)
f2(x)
20
4
10
10
0
2
0
5
10
15
0
0
2
f1(x)
(d) MOEA/D-CDP
Fig. 10.
4
6
f1(x)
(e) MOEA/D-SR
8
0
0
10
20
30
40
50
f1(x)
(f) C-MOEA/D
The non-dominated solutions achieved by each algorithm on LIR-CMOP11 during the 30 independent runs are plotted in (a)-(f).
ACKNOWLEDGMENT
This research work was supported by Guangdong Key
Laboratory of Digital Signal and Image Processing, the National Natural Science Foundation of China under Grant
(61175073, 61300159, 61332002, 51375287), Jiangsu Natural Science Foundation (BK20130808) and Science and
Technology Planning Project of Guangdong Province, China
(2013B011304002).
R EFERENCES
[1] K. Deb, Multi-objective optimization using evolutionary algorithms.
John Wiley & Sons, 2001, vol. 16.
[2] C. A. C. Coello, “Theoretical and numerical constraint-handling techniques used with evolutionary algorithms: a survey of the state of the
art,” Computer Methods in Applied Mechanics and Engineering, vol.
191, no. 11–12, pp. 1245–1287, 2002.
[3] T. BDack, F. Hoffmeister, and H. Schwefel, “A survey of evolution
strategies,” in Proceedings of the 4th international conference on genetic
algorithms, 1991, pp. 2–9.
[4] A. Homaifar, C. X. Qi, and S. H. Lai, “Constrained optimization via
genetic algorithms,” Simulation, vol. 62, no. 4, pp. 242–253, 1994.
[5] J. A. Joines and C. R. Houck, “On the use of non-stationary penalty
functions to solve nonlinear constrained optimization problems with
ga’s,” in Evolutionary Computation, 1994. IEEE World Congress on
Computational Intelligence., Proceedings of the First IEEE Conference
on. IEEE, 1994, pp. 579–584.
[6] J. C. Bean and A. ben Hadj-Alouane, A dual genetic algorithm for
bounded integer programs, 1993.
[7] D. W. Coit, A. E. Smith, and D. M. Tate, “Adaptive penalty methods for
genetic optimization of constrained combinatorial problems,” INFORMS
Journal on Computing, vol. 8, no. 2, pp. 173–182, 1996.
[8] A. Ben Hadj-Alouane and J. C. Bean, “A genetic algorithm for the
multiple-choice integer program,” Operations research, vol. 45, no. 1,
pp. 92–101, 1997.
[9] Y. G. Woldesenbet, G. G. Yen, and B. G. Tessema, “Constraint handling
in multiobjective evolutionary optimization,” IEEE Transactions on
Evolutionary Computation, vol. 13, no. 3, pp. 514–525, June 2009.
[10] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist
multiobjective genetic algorithm: NSGA-II,” IEEE Transactions on
Evolutionary Computation, vol. 6, no. 2, pp. 182–197, Apr 2002.
[11] T. Takahama and S. Sakai, “Constrained optimization by ε constrained
particle swarm optimizer with ε-level control,” in Soft Computing as
Transdisciplinary Science and Technology. Springer, 2005, pp. 1019–
1029.
[12] T. P. Runarsson and X. Yao, “Stochastic ranking for constrained evolutionary optimization,” IEEE Transactions on evolutionary computation,
vol. 4, no. 3, pp. 284–294, 2000.
[13] Z. Fan, W. Li, X. Cai, K. Hu, H. Lin, and H. Li, “Angle-based constrained dominance principle in moea/d for constrained multi-objective
optimization problems,” in IEEE Congress on Evolutionary Computation, 2016, pp. 460–467.
[14] B. Li, J. Li, K. Tang, and X. Yao, “Many-objective evolutionary
algorithms: A survey,” ACM Computing Surveys (CSUR), vol. 48, no. 1,
p. 13, 2015.
[15] E. Mezura-Montes and C. A. Coello Coello, “Constraint-handling in
nature-inspired numerical optimization: Past, present and future,” Swarm
and Evolutionary Computation, vol. 1, no. 4, pp. 173–194, Dec. 2011.
[16] Z. Cai and Y. Wang, “A multiobjective optimization-based evolutionary
algorithm for constrained optimization,” Evolutionary Computation,
IEEE Transactions on, vol. 10, no. 6, pp. 658–675, 2006.
[17] T. Ray, H. K. Singh, A. Isaacs, and W. Smith, “Infeasibility driven
evolutionary algorithm for constrained optimization,” in Constrainthandling in evolutionary optimization. Springer, 2009, pp. 145–165.
[18] Y. Wang, Z. Cai, Y. Zhou, and W. Zeng, “An adaptive tradeoff model
for constrained evolutionary optimization,” IEEE Transactions on Evolutionary Computation, vol. 12, no. 1, pp. 80–92, 2008.
[19] B. Y. Qu and P. N. Suganthan, “Constrained multi-objective optimization
algorithm with an ensemble of constraint handling methods,” Engineering Optimization, vol. 43, no. 4, pp. 403–416, 2011.
[20] K. Deb, “An efficient constraint handling method for genetic algorithms,”
Computer methods in applied mechanics and engineering, vol. 186,
no. 2, pp. 311–338, 2000.
[21] K. Miettinen, Nonlinear Multiobjective Optimization. Springer Science
& Business Media, 1999, vol. 12.
[22] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithm
based on decomposition,” IEEE Transactions on evolutionary computation, 2007.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
[23] Z. Fan, W. Li, X. Cai, H. Li, H. Huang, and Z. Cai, “An improved epsilon
constraint handling method embedded in MOEA/D for constrained
multi-objective optimization problems,” in 2016 IEEE Symposium Series
on Computational Intelligence, Dec 2016.
[24] Z. Fan, W. Li, X. Cai, H. Huang, Y. Fang, Y. You, J. Mo, C. Wei,
and E. D. Goodman, “An improved epsilon constraint-handling method
in MOEA/D for cmops with large infeasible regions,” arXiv preprint
13
arXiv:1707.08767, 2017.
[25] P. A. Bosman and D. Thierens, “The balance between proximity
and diversity in multiobjective evolutionary algorithms,” Evolutionary
Computation, IEEE Transactions on, vol. 7, no. 2, pp. 174–188, 2003.
[26] E. Zitzler and L. Thiele, “Multiobjective evolutionary algorithms: a comparative case study and the strength pareto approach,” IEEE Transactions
on Evolutionary Computation, vol. 3, no. 4, pp. 257–271, 1999.
| 2cs.AI
|
Towards Parameterized Regular Type Inference
Using Set Constraints
arXiv:1002.1836v1 [cs.LO] 9 Feb 2010
F.B UENO ,1 J.NAVAS ,2
2
AND
M.H ERMENEGILDO1,3
1 Technical U. of Madrid (Spain)
National University of Singapore (Republic of Singapore)
3 IMDEA Software (Spain)
Abstract. We propose a method for inferring parameterized regular types for
logic programs as solutions for systems of constraints over sets of finite ground
Herbrand terms (set constraint systems). Such parameterized regular types generalize parametric regular types by extending the scope of the parameters in the
type definitions so that such parameters can relate the types of different predicates. We propose a number of enhancements to the procedure for solving the
constraint systems that improve the precision of the type descriptions inferred.
The resulting algorithm, together with a procedure to establish a set constraint
system from a logic program, yields a program analysis that infers tighter safe
approximations of the success types of the program than previous comparable
work, offering a new and useful efficiency vs. precision trade-off. This is supported by experimental results, which show the feasibility of our analysis.
1 Introduction
Type inference of logic programs is the problem of computing, at compile time, a representation of the terms that the predicate arguments will be bound to during execution of
the program. This kind of type inference involves not only assigning types to procedure
arguments out of a predefined set of type definitions, as in traditional type inference,
but also the more complex problem of inferring the type definitions themselves, similarly to what is done in shape analysis. Although most logic programming languages
are either untyped or allow mixing typed and untyped code, inferring type information
for the entire program is important since it allows the compiler to generate more efficient code and it has well-known advantages in detecting programming errors early.
For instance, simple uses of such information include better indexing, specialized unification, and more efficient code generation. More advanced uses include compile-time
garbage collection and non-termination detection. There are also other areas in which
type information can be useful. For example, during verification and debugging it can
provide information to the programmer that is not straightforward to obtain by manual
inspection of the program.
In this paper we use the set constraint-based approach [14,15]. We propose an algorithm for solving a set constraint system that relates the set of possible values of the
program variables, by transforming it into a system whose solutions provide the type
definitions for the variables involved in parameterized regular form. We focus on types
(a) Program:
(b) Parametric regular types:
:- type A1(X)
-> [] | [X|A1(X)].
:- type A2(W)
-> W.
:- type A3(Y,Z) -> Z | [Y|A3(Y,Z)].
append([],L,L).
(c) Parameterized regular types:
:- type A1 -> [] | [X|A1].
append([X|Xs],Ys,[X|Zs])::- type A2 -> .
append(Xs,Ys,Zs).
:- type A3 -> A2 | [X|A3].
(d) Local parameters:
(e) Global parameters:
:- typing append(A1(E),A2(T),A3(E,T)).
:- typing nrev(R1,R2).
:- typing nrev(N1(E),N2(E)),
:- type R1 -> [] | [X|R1].
append(A1(E),A2([]),A3(E,[])). :- type R2 -> [] | [X|R2].
:- typing append(A1,A2,A3).
Fig. 1. Parametric vs parameterized regular types
which are conservative approximations of the meaning of predicates, and hence, overapproximations of the success set (in contrast to the approach of inferring well-typings,
as in, e.g., [3,21], which may differ from the actual success set of the program). Type
inference via set constraint solving was already proposed in [12,15]. However, most existing algorithms [13,14,9] are either too complicated or lacking in precision for some
classes of programs. We try to alleviate these problems by, first, generating simple equations; second, using a comparatively straightforward procedure for solving them; and,
third, using a non-standard operation during solving that improves precision by “guessing” values. At the same time, we attack a more ambitious objective since our resulting
types, that we call parameterized, are more expressive than in previous proposals.
Consider for instance the append/3 program in Fig. 1.a, and assume the type descriptors A1, A2, and A3 for each predicate argument. Most state-of-the-art analyses will
simply infer that the first argument (type A1) is a list, leaving open the types of the other
two arguments. In Fig. 1.b we show classical parametric types for append/3. We are
unaware of an existing proposal that is able to infer them as an approximation to the
success set of the predicate (see Sec. 5). Even if we had an analysis that inferred them,
they are still less expressive (or need more elaboration for the same expressiveness)
than our proposal, as we will discuss.
The parametric types in Fig. 1.b denote the expected list type for A1 which is parametric on some X. Note that A2 is unbound since it may be instantiated to any term.
On the other hand, A3 is an open-ended list of elements of some type Y whose tail is of
some type Z. A key observation is that while there is a clear relation between the type of
the elements of A1 and A3, and between the type of the tails in A2 and A3, these relations
are not captured. In Fig. 1.c, we show a desirable, more accurate type for append/3. It
denotes that the type of A3 is that of open-ended lists of elements of the same type as
A1 with tail of the same type as A2.
The relations between the types of arguments can be captured with parametric types
only if type parameters are instantiated. For example, the first typing in Fig. 1.d captures
the desired relations for append/3. Instead, by using global type parameters, as we
propose, the parametric type definitions may exhibit the required exact relations right
from the inference of the type definitions alone. Although the absence of typings to
express the same relations is a small advantage, a bigger one might be expected with
regard to the analysis. By using parametric types, an analysis should not only infer type
definitions but also typings showing the exact values for the type parameters. It is not
clear at all how this might be done (note that typings are not the types of calls). Our
proposal infers type definitions alone, as usual, and yet are at least as expressive as
parametric types with typings.
Consider a program construct where two different predicates share a variable, so
that the corresponding arguments have the same type. This happens, for example, with
the arguments of nrev/2 and append/3 in the classical naive reverse program (see
Ex. 1). This property cannot be captured with (standard) typings for the predicates if
parametric type definitions are used. One would need something like the second typing
in Fig. 1.d. However, this is not usually a valid typing (it types several predicate atoms
at the same time), and it is also not intuitive how it could be inferred. In contrast, the
parameterized type definition of Fig. 1.e, together with those of (c), easily capture the
property, by sharing the type variable global parameter X. Therefore, more precise types
(called parameterized regular types [19]) can be produced if the scope of each type
variable in a type definition is broader than the definition itself, so that the types of
different arguments can be related.
As discussed in detail in Sec. 5, we believe that no previous proposal exists for
inferring types for logic programs with the expressive power of parameterized regular
types. In addition, our proposal fits naturally in the set constraint-based approach. In
this context we also define a number of enhancements to the solving procedure. The
result is a simple but powerful type inference analysis. Our preliminary experiments
also show that our analysis runs in a reasonable amount of time.
2 Preliminaries
Let V be a non-terminal which ranges over (set) variables V , and let f range over functors (constructors) F of given arity n ≥ 0. Set expressions, set expressions in regular
form, and in parameterized form are given by E in the following grammars:
Expressions:
E ::= 0/ | V | f (E1 , . . . , En ) | E1 ∪ E2 | E1 ∩ E2
Regular form:
E ::= 0/ | N N ::= V | f (V1 , . . . ,Vn ) | N1 ∪ N2
Parameterized form: E ::= 0/ | N N ::= N1 ∪ N2 | R R ::= V | f (V1 , . . . ,Vn ) | V ∩ R
The meaning of a set expression is a set of (ground, finite) terms, and is given by the
following semantic function µ under an assignment σ from variables to sets of terms:
/ σ) = 0/
µ(0,
µ(V, σ) = σ(V )
µ(E1 ∪ E2 , σ) = µ(E1 , σ) ∪ µ(E2 , σ)
µ(E1 ∩ E2 , σ) = µ(E1 , σ) ∩ µ(E2 , σ)
µ( f (E1 , . . . , En ), σ) = { f (t1 , . . . ,tn ) | ti ∈ µ(Ei , σ)}
Let E1 and E2 be two set expressions, then a set equation (or equation, for short)
is of the form E1 = E2 . A set equation system is a set of set equations. A solution σ
of a system of equations S is an assignment that maps variables to sets of terms which
satisfies: µ(e1 , σ) = µ(e2 , σ) for all (e1 = e2 ) ∈ S. We will write S ⊢ e1 = e2 iff every
solution of S is a solution of e1 = e2 . A set equation system is in top-level form if in all
expressions of the form f (x1 , . . . , xn ), all the xi are variables. The top-level variables of
a set expression are the variables which occur outside the scope of a constructor.
A standard set equation system is one in which all equations are of the form V = E
(i.e., lhs are variables and rhs set expressions) and there are no two equations with the
same lhs. Equations where the rhs is also a variable will be called aliases. In a standard
set equation system variables which are not in the lhs of any equation are called free
variables (since they are not constrained to any particular value). Variables which do
appear in lhs are called non-free.
A regular set equation system is one which is standard, all rhs are in regular form,
has no top-level variables except for aliases, and also no free variables. A regular set
equation system is in direct syntactic correspondence with a set of regular type definitions. Regular types are equivalent to regular term grammars where the type definitions
are the grammar rules. In a regular set equation system the set variables act as the type
symbols, and each equation of the form x = e1 ∪ . . . ∪ en acts as n grammar rules of the
form x ::= e j , 1 ≤ j ≤ n. By generalizing regular set equation systems to allow free
variables what we obtain is the possibility of having parameters within the regular type
definitions. However, when free variables are allowed intersection has to be allowed
too: given that free variables are not constrained to any particular value, intersections
cannot be “computed out.”
A leaf-linear set equation system [19] is one which is standard, all rhs are in parameterized form, and all top-level variables are free. Note that leaf-linear set equation
systems are the minimal extension of regular equation systems in the above mentioned
direction, in the sense that intersections are reduced to a minimal expression: several
free variables and only one (if any) constructor expression. More importantly, a leaflinear set equation system is more expressive than parametric regular types, since parameters have the scope of the whole system, instead of the particular type definition in
which they occur, as in parametric type definitions.
3 Type inference
In this section, we present the different components of our analysis method for inferring
parameterized regular types. First, a set equation system is derived from the syntax of
the program, then, the system is solved, and finally it is projected onto the program variables providing the type definitions for such variables. The resulting equation system
is in solved form, i.e., it is leaf-linear. Such a system will be considered a fair representation of the solution to the original set equation system, since it denotes a set of
parameterized regular types. When the system is reduced to solved form, the parameters
of such a solution are the free variables.
Generating a set equation system for a program. Let P be a program, ΠP the set of
predicate symbols in P, ranked by their arity, and P| p the set of rules defining predicate
p in program P. Our analysis assumes that all rules in P have been renamed apart so
that they do not have variables in common. For each p ∈ ΠP of arity n we associate a
signature of p defined as Σ(p) = p(x1 , . . . , xn ) where {x1 , . . . , xn } is an ordered set of n
new variables, one for each argument of p. For atom A, let [A] j denote its j-th argument.
For a predicate p ∈ ΠP with arity n we define C p and the initial set equation system, E,
for P as follows. In order to avoid overloading symbol ∪, to clarify the presentation, we
will use ∪· for the usual set union, while ∪ will stand for the symbol occurring in set
expressions.
S
E = · {C p | p ∈ ΠP }
C p = CHead ∪· CBody
(1)
S
CHead = { x j = {[H] j | (H:-B) ∈ P| p } | x j ∈ vars(Σ(p)) }
T
CBody = { y = {[Σ(A)]i | [A]i = y, A ∈ B} | (H:-B) ∈ P| p , y ∈ vars(B) }
∪· { y = [Σ(A)]i ∩ t | [A]i = t, A ∈ B, (H:-B) ∈ P| p , t 6∈ vars(B), y fresh var }
Example 1. Take signatures append(A1,A2,A3) and nrev(N1,N2) in the following
program for naive reverse. The equation system for nrev/2 is Cnrev/2 = CH ∪· CB .
nrev([],[]).
nrev([X|Xs],Ys):- nrev(Xs,Zs), append(Zs,[X],Ys).
CH = { N1 = [] ∪ [X|Xs], N2 = [] ∪ Ys }
CB = { Xs = N1, Ys = A3, Zs = N2 ∩ A1, W = [X] ∩ A2 }
Note that a system E which results from Eq. 1 is in standard form. Moreover, to
put it also in top-level form we only need to repeatedly rewrite every subexpression of
every equation of E of the form f (e1 , . . . , e j , . . . , en ) into f (e1 , . . . , y j , . . . , en ), whenever
e j is not a variable, adding to E equation y j = e j , with y j a new fresh variable, until no
further rewriting is possible. The new equations added to E are in turn also rewritten in
the same process. We call the resulting system Eq(P). Obviously, Eq(P) is equivalent
to E in Eq. 1.
Analysis of the program. In order to analyze a program and infer its types, we follow
the call graph of the program bottom-up, as explained in the following. First, the call
graph of the program is built and its strongly connected components analyzed. Nodes
in the same component are replaced by a single node, which corresponds to the set
of predicates in the original nodes. The (incoming or outgoing) edges in the original
nodes are now edges of the new node. The new graph is partitioned into levels. The
first level consists of the nodes which do not have outgoing edges. Each successive
level consists of nodes which have outgoing edges only to nodes of lower levels. The
analysis procedure processes each level in turn, starting from the first level. Predicates
in the level being processed can be analyzed one at a time or all at once.
Each graph level is a subprogram of the original program. To analyze a level, equations are set up for its predicates as by Eq. 1, and copies of the solutions already obtained
for the predicates in lower levels are added. To do this, the signatures and solutions of
the predicates in lower levels are renamed apart. The new signatures replace the old
ones when used in building the equations for the subprogram. The new copies of the
solutions are added to the set equation system. For a given predicate, there is a different
copy for each atom of that predicate which occurs in the subprogram being analyzed.
Example 2. Consider the following (over-complicated) contrived predicate to declare
two lists identical, which calls predicate nrev/2 of Ex. 1 twice.
same(L1,L2):- nrev(L1,L), nrev(L,L2).
Predicate nrev/2, which is in a lower scc, would have been analyzed first. Taking signatures same(S1,S2) and nrev(N1,N2), the solution for nrev/2 would be:
{ N1 = [] ∪ [X|N1], N2 = [] ∪ [X|N2] }
Since there are two calls to this predicate in the above definition for same/2, two copies
of the above equations (renaming the above equations for hN1, N2i into hN11, N12i and
hN21, N22i) would be added. The initial equation system for same/2 is thus:
{ S1 = L1, S2 = L2, L1 = N11, L2 = N22, L = N12 ∩ N21,
N11 = [] ∪ [X1|N11], N21 = [] ∪ [X2|N21],
N12 = [] ∪ [X1|N12], N22 = [] ∪ [X2|N22]
}
Note that the two arguments of same/2 have, in principle, list elements of different type
(X1 and X2). That they are in fact of the same type will be recovered during the solving
of the equations, in particular, the equation for L, the rhs of which is an intersection
(unification). This will be done by an operation we propose (BIND) which will make
an appropriate “guess”, as explained later.
Note that the copies are required because the type variables involved might get new
constraints during analysis of the subprogram. Such constraints are valid only for the
particular atom related to the given copy. Also, as the example above shows, using a
single copy would impose constraints, because of the sharing of the same parameter
between equations referring to the same and single copy (as it would have been the case
with X in the example, had we not copied the solution for nrev/2). Such constraints
might be, in principle, not true (though in the example it is finally true that both lists
have elements of the same type). Thus, copies represent the types of the program atoms
occurring at the different program points (i.e., different call patterns).
Note also that equations with intersection in the expression in the rhs will be used
to capture unification, as it is the case also in the example. This occurs sometimes
by simplification of intersection (procedure SIMP below), but most of the times from
operation BIND, that mimics unification. Such equations are particularly useful. This
is also true even if the equation is such that its lhs is a variable which does not occur
anywhere else in the program (as it is the case of W in Ex. 1). Additionally, in these cases,
if the rhs finally becomes the empty set, denoting a failure, this needs to be propagated
explicitly, since the lhs is a variable not related to the rest of the equations.
The details of the solving procedure are explained below. The procedure is based
on the usual method of treating one equation at a time, the lhs of which is a variable,
and replacing every (top-level) occurrence of the variable by the expression in the rhs
everywhere else. Once the system is set up as explained, it is first normalized and then
solved as described in the following.
Normal form. The normal form used is Disjunctive Normal Form (DNF), plus some
/ which transform expressions
simplifications based on equality axioms for ∩, ∪, and 0,
into parameterized form. Once the system is in top-level form, two auxiliary algorithms
are used to rewrite equations to achieve normal form. These algorithms are based on
semantic equivalences, and therefore preserve solutions.
The first algorithm, DNF, puts set expressions in a set equation in disjunctive normal
form. Note that if all expressions of an equation q are in top-level form then those of
equation DNF(q) are in parameterized form, except for nested occurrences of 0/ and
possibly several occurrences of constructor expressions in conjuncts. This is taken care
of in the second algorithm, SIMP.
SIMP simplifies set expressions in an equation system E by repeatedly rewriting
every subexpression based on several equivalences until no further rewriting is possible,
as follows:
e ∩ 0/ ❀ 0/
e∩e ❀ e
e ∪ 0/ ❀ e
e∪e ❀ e
e1 ∪ (e1 ∩ e2 ) ❀ e1
f (e1 , . . . , em ) ∩ g(d1 , . . . , dn ) ❀ 0/ if f 6≡ g or n 6≡ m
f (e1 , . . . , en ) ∩ f (d1 , . . . , dn ) ❀ f (y1 , . . . , yn )
if ∀ j.1 ≤ j ≤ n.y j = e j ∩ d j
8. f (y1 , . . . , yn ) ❀ 0/ if ∃i.1 ≤ i ≤ n.E ⊢ yi = 0/
1.
2.
3.
4.
5.
6.
7.
INTER . SIMPLIFICATION
INTER . ABSORPTION
UNION SIMPLIFICATION
UNION ABSORPTION
SUBSUMPTION
CLASH
INTER . DISTRIBUTION
EMPTINESS
/ For this test a straightforward
Note that line 8 makes use of the check E ⊢ yi = 0.
adaptation of the type emptiness test of [7] to parameterized definitions is used. Note
also that the rewriting in line 7 deserves some explanation. If the set equation system
E does not contain an equation y j = e j ∩ d j for some j then the equation is added to
E, with y j a new fresh variable. Otherwise, y j from the existing equation is used. If
equations were not added, it would prevent full normalization of the set expressions. If
variables were not reused, but instead new variables added each time, it would prevent
the global algorithm from terminating.
Solving recurrence equations. Since the set equation system is kept in standard form
at all times, the number of different equations that need to be considered is very small.
The procedure CASE reduces a given equation into a simpler one. It basically takes
care only of recurrences. We call recurrence an equation x = e where x occurs top-level
in e. There are four cases of recurrences that are dealt with:
1.
2.
3.
4.
x = x ❀ x = 0/
x = x ∩ e ↔ x ⊆ e ❀ x = 0/
x = x∪e ↔ e ⊆ x ❀ x = e
x = (x ∩ e1 ) ∪ e2 ↔ e2 ⊆ x ⊆ e1 ❀ x = e2
Note that in all cases we have chosen for variable x the least solution of all possible
ones allowed by the corresponding recurrence. We are thus taking a minimal sufficient
solution, in the sense that the resulting set of terms would be the smallest possible one
that still approximates the program types. This is possible because we keep equations
in standard form, so that their lhs is always a variable and there is only one equation per
variable. If the equation turns into a recurrence, then the program contains a recursion
which does not produce solutions (e.g., an infinite failure).
Example 3. Consider the following program with signatures p(P), q(Q), and r(R):
Algorithm 1 SOLVE
Input: a set equation system E in standard form.
Output: a set equation system S in solved form.
1: initialize S
2: repeat
3:
initialize C
4:
repeat
5:
subtract q from E
6:
hq′ ,Ci ← SIMP(DNF(q),C,S)
7:
q′′ ← CASE(q′ ,S)
8:
let q′′ be of the form x = e
9:
replace every top-level occurrence of x in E and in S by e
10:
add q′′ to S
11:
until E is empty
12:
repeat
13:
for all q ∈ S do
14:
hq,Ci ← SIMP(DNF(q),C,S)
/ and S ⊢ x = 0/ then
15:
if q is of the form x = e, e 6= 0,
16:
replace q by x = 0/ in S
17:
end if
18:
end for
19:
until no equation q is replaced
/ ∈ S do
20:
for all (x = 0)
21:
if x occurs in Eq(P) for some clause c then
22:
for all variable y in the head of c do
23:
replace the equation in S with lhs y by y = 0/
24:
end for
25:
end if
26:
end for
27:
for all (x = e) ∈ S do
28:
replace every top-level occurrence of x in C by e
29:
end for
30:
add C to E
31:
if E is empty then
32:
E ← BIND(S)
33:
end if
34: until E is empty
p(X):- p(X).
q(a).
q(Y):- q(Y).
r(b).
r(Z):- q(Z), r(Z).
It is easy to see that its initial set equation system will progress towards the following
form: { P = P, Q = a ∪ Q, R = b ∪ (Q ∩ R) }.
The global algorithm. The global algorithm for solving a set equation system is given
as Algorithm SOLVE. In order to facilitate the presentation, the set equation system is
partitioned into three: E, C, and S. E is the initial equation system (generated from the
program as in Eq. 1), S is the solved system (i.e., the output of the analysis), and C is
an auxiliary system with the same form as E. This forces the use of SIMP in lines 6
and 14 with two input equation systems instead of one. SIMP may add new equations,
/ is carried along
and these are added to C. Also, in SIMP the test in line 8 (E ⊢ y j = 0)
in S (i.e., it should be read as S ⊢ y j = 0/ when called from SOLVE).
Once more, we use the test S ⊢ x = 0/ (line 15 of Algorithm SOLVE). This is related
to the other use of this test in SIMP. Note that SIMP is invoked again at line 14 of
Algorithm SOLVE. Together, the loop in lines 12-19 of Algorithm SOLVE and line 8
of SIMP perform a complete type emptiness propagation. Because of this, the effect of
the rest of cases of SIMP, which propagate symbol 0/ in subexpressions, is to achieve
an implicit check for non-emptiness of intersections. The rationale behind the loop in
lines 20-26 is different. It solves a lack of propagation of failure for some programs,
due to the form of the initial equation system.
Example 4. In the program below, failure of p/1 would not be detected unless step 23
is performed, even if failure of q/2 is detected because of type emptiness.
p(X):- q(b,X).
q(a,a).
Finally, the BIND procedure invoked at line 32 of Algorithm SOLVE is not needed
for obtaining the solved equation system, since S is already in solved form when BIND
is called, but it may improve the precision of the analysis.
Example 5 (SOLVE with binding of free variables). We show the analysis of a contrived
example which exposes the strength of our approach in propagating types through the
type variables that act as parameters. We consider signature append(A1,A2,A3) already solved with solution: A1 = [] ∪ [X|A1], A3 = A2 ∪ [X|A3]; and concentrate
on the analysis of:
appself(A,B):- append(A,[],B).
We show the contents of E at the beginning of Algorithm SOLVE and of S at line 32 (C
and E are empty). We only show the equations for the program variables of appself/2.
The equations missing are those for the solution of append/3 –note that A2 is free– and
aliases of variables A and B for the arguments of the signature of appself/2:
E={ A=A1, B=A3, W=A2 ∩ [] }
S={ A=[] ∪ [X|A1], B=A2 ∪ [X|A3], W=A2 ∩ [] }
Now, BIND can give a more precise value to the only free variable that appears in an
intersection, A2 (A2 = []). We now include the equation for A3, which is relevant:
E={A2=[]}
S={A=[] ∪ [X|A1], B=A2 ∪ [X|A3], W=A2 ∩ [], A3=A2 ∪ [X|A3]}
After a second iteration of the main loop (lines 2–34) we have E empty again, and:
S={ A=[] ∪ [X|A1], B=[] ∪ [X|A3], W=[], A3=[] ∪ [X|A3] }
Finally, after projection on the variables of interest (which are in fact those of the signature of appself/2, but we use its program variables A,B for clarity instead):
S={ A=[] ∪ [X|A], B=[] ∪ [X|B] }
Binding free variables to values. Since the equation system is already solved, i.e.,
it is in solved form, when BIND is applied, free variables can at this moment take any
value. BIND will then mimic unification by binding free variables to the minimal values
required so that the expressions involving them have a solution. In order to do this, all
subexpressions with the form of an intersection (i.e., unification) with a free variable
are considered. These are called conjuncts.
Since the set expressions are in parameterized form, conjuncts are all of the form
e1 ∩e2 , where e1 is a, possibly singleton, conjunction of free variables, and e2 a, possibly
non-existing, constructor expression. Let e1 be of the form x1 ∩ . . . ∩ xn , n ≥ 1, and e
be e2 if it exists or a new variable, otherwise. A set of candidates is proposed of the
form {xi = e | 1 ≤ i ≤ n}. For an equation q, let Candidates(q) denote the set of sets
of candidates for each conjunct in q. For defining procedure BIND we need to consider
the following relation between equations. Let q 7→ q′ when equation q is the equation
taken by SIMP at a given time, and q′ the new equation added at step 7 of Algorithm
SIMP from q. Let 7→∗ denote the transitive closure of 7→. BIND constructs the formula
below, and synthesizes from it a suitable set of set equations.
^
_
_
q∈Eq(P) q7→∗ q′
^
e
(2)
c∈Candidates(q′ ) e∈c
Example 6. Let the following contrived program for alternate/2, which resembles
the way the classical program for the towers of Hanoi problem alternates the arguments
representing the pegs that hold the disks across recursions:
alternate(A1,B1).
alternate(A2,B2):- alternate(B2,A2).
It is easy to see that, for a signature alternate(P1,P2), the solution will be the system
of the following two equations:
P1 = A1 ∪ B1,
P2 = A1 ∪ B1
The analysis of a program containing an atom of the form alternate(a,b) in a
clause body, will be faced with the equations:
W1 = (A1 ∪ B1) ∩ a,
W2 = (A1 ∪ B1) ∩ b,
which at line 32 of Algorithm SOLVE would be solved as:
W1 = (A1 ∩ a) ∪ (B1 ∩ a),
W2 = (A1 ∩ b) ∪ (B1 ∩ b),
so that BIND will find two original equations (in Eq(P); not related by 7→∗ ) with two
candidates each ({{A1=a}, {B1=a}} for the equation for W1 and {{A1=b}, {B1=b}} for
the equation for W2). Thus, BIND obtains the formula:
( A1 = a ∨ B1 = a ) ∧ ( A1 = b ∨ B1 = b ),
which is equivalent to:
( A1 = a ∧ B1 = b ) ∨ ( A1 = b ∧ B1 = a ),
and results in the following two
set equations, which “cover” all solutions of the above formula:
A1 = a ∪ b,
B1 = a ∪ b,
NFTA
PTA
Benchmark # Desc. Prec. Time (s) Error Prec. Time (s) Error
append
blanchet
dnf
fib
grammar
hanoi
mmatrix
mv
pvgabriel
pvqueen
revapp
serialize
zebra
Total
3
1
2
2
1
5
3
3
2
2
2
2
7
35
1
1
2
0
0
1
3
1
0
1
2
2
1
13
0.000
0.052
1.156
0.000
0.016
0.020
0.008
0.020
0.184
0.020
0.004
0.044
0.080
1.604
n
y
n
n
y
n
n
n
n
n
n
n
n
2
3
1
2
2
1
5
3
3
2
2
2
2
7
35
0.000
0.032
1.892
0.000
0.124
0.028
0.008
0.036
0.124
0.028
0.004
0.164
0.372
2.812
y
y
n
y
y
n
y
y
n
y
y
y
y
Bad call type
append(A,a,A)
attacker(s)
dnf(X,a(z1, o(z2,z3)))
fib(a,X)
parse([boxes,fly],S)
hanoi(5,a,b,c,[mv(e,f)])
mmult([1,2],[[1,2],[3,4]],X)
mv([1,3,1],[b,c,a],X)
pv init([1,2],X)
queens(4,[a,b,c,d])
rev([1,2],[a,b])
serialize(‘‘hello’’,[a,b,c])
zebra(E,S,J,U,second,Z,W)
10
Table 1. Experimental results for NFTA and parameterized type analysis (PTA)
and (imply expressions for P1 and P2 which) correctly approximate the success types
of alternate/2 when called as in alternate(a,b).
In the synthesis of set equations from the formula in Eq. 2, ∨ turns into ∪ and ∧
turns into ∩. The details of how to do this in a complete synthesis procedure are the
subject of current work. Our implementation currently deals with the simpler cases,
including those similar to the example above and those in which all candidates involve
only one free variable (as in Ex. 5).
4 Preliminary experimental evaluation
In order to study the practicality of our method we have implemented a prototype analyzer in Ciao (http://www.ciaohome.org [4]) and processed a representative set of
benchmarks taken mostly from the PLAI (the Ciao program analyzer) and GAIA [22]
sets. We chose the Non-deterministic Finite Tree Automaton (NFTA)-based analysis [9]
for comparison. We believe that this is a fair comparison since it also over-approximates
the success set of a program in a bottom-up fashion (inferring regular types). Its implementation is publicly available (http://saft.ruc.dk/Tattoo/index.php) and it is
also written in Ciao. We decided not to compare herein with top-down, widening-based
type analyses (e.g. [16,22,23]), since it is not clear how they relate to our method. We
leave this comparison as interesting future work.
The results are shown in Table 1. The fourth and seventh columns show analysis
times in seconds for both analyses on an Intel Core Duo 1.33GHz CPU, 1GB RAM,
and Ubuntu 8.10 Linux OS. The time for reading the program and generating the constraints is omitted because it is always negligible compared to the analysis time. Column # Desc. shows the number of type descriptors (i.e., predicate argument positions)
which will be considered for the accuracy test (all type descriptors are considered for
the timing results). For simplicity, we report only on the argument positions belonging to the main predicate. Note that execution of these predicates often reaches all the
predicates in the program and thus the accuracy of the types for those positions is often
a good summary of precision for the whole program. Columns labeled Prec. show the
number of type descriptions inferred by NFTA and our approach that are different from
type any. To test precision further, we have added to each benchmark a query that fails,
which is shown in the last column (Bad call type). Columns labeled Error show whether
this is captured by the analysis or not.
Regarding accuracy, our experiments show that parameterized types allow inferring
type descriptors with significant better precision. Our approach inferred type descriptions different from any for every one of the 35 argument positions considered, while
NFTA inferred only 13. Moreover, those type descriptors were accurate enough to capture type emptiness (i.e., failure) in 10 over 13 cases, whereas NFTA could only catch
two errors. The cases of dnf, hanoi, and pvgabriel deserve special attention, since our
method could not capture emptiness for them. For dnf the types inferred are not precise enough to capture that the second argument is not in disjunctive normal form. This
is due to the lack of inter-variable dependencies in our set-based approach. However,
emptiness is easily captured for other calls such as dnf(X,b). A similar reasoning
holds for hanoi since the analysis infers that the types of the pegs are the union of a, b,
c, e, and f. Finally, the success of pvgabriel depends on a run-time condition, and thus
no static analysis can catch the possible error.
Regarding efficiency, we expected our analysis to be slower than other less expressive methods (like NFTA), since our method is more expressive than previous proposals.
Table 1 shows that this is indeed the case, but the differences are reasonable for the selected set of benchmarks, specially considering the improvement in accuracy. Further
research is of course needed using larger programs (see Sect. 6), but we find the results
clearly encouraging.
5 Related work
Type inference, i.e., inferring type definitions from a program, has received a lot of
attention in logic programming. Mishra and Reddy [18,19] propose the inference of
ground regular trees that represent types, and compute an upper approximation of the
success set of a program. The types inferred are monomorphic. Zobel [24] also proposes
a type inference method for a program, where the type of a logic program is defined
as a recursive (regular) superset of its logical consequences. However, the inference
procedure does not derive truly polymorphic types: the type variables are just names
for types that are defined by particular type rules.
In [17] an inference method called type reconstruction is defined, which derives
types for the predicates and variables of a program. The types are polymorphic but they
are fixed in advance. On the contrary, our inference method constructs type definitions
during the analysis. In [8] the idea is that the least set-based model of a logic program
can be seen as the exact Herbrand model of an approximate logic program. This approximate program is a regular unary program, which has a specific syntactic form which
limits its computational power. For example, type parameters cannot be expressed. The
work of [11,20] also builds regular unary programs, and their type inference derives
such programs. However, the types remain monomorphic.
Heintze and Jaffar [12] defined an elegant method for semantic approximation,
which was the origin of set-based analysis of logic programs. The semantics computes, using set substitutions, a finite representation of a model of the program that
is an approximation of its least model. Unfortunately, the algorithm of [12] was rather
complicated and practical aspects were not addressed. Such practical aspects were addressed instead in subsequent work [14,13,15], simplifying the process. However, our
equations are still simpler: for example, we do not make use of projections, i.e., expressions of the form f i (x), where f is a constructor, the meaning of, e.g., y = f 1 (x) being
that x ⊇ f (y, . . .). Also, and more importantly, these approaches do not take advantage
of parameters, and the types obtained are again always monomorphic.
Type inference has also been approached using the technique of abstract interpretation. A fixpoint is computed, in most cases with the aid of a widening operation to limit
the growth of the type domain. This technique is also used in [11]. However, none of
the analyses proposed to date (e.g., [16,22,23]) is parametric. It is not clear how the approach based on abstract domains with widening compares to the set constraints-based
approach. An interesting avenue for future work, however, is to define our analysis as
a fixpoint. This is possible in general due to a result by Cousot and Cousot [6] which
shows that rewriting systems like ours can be defined alternatively in terms of a fixpoint
computation.
Directional types [1,5] are based on viewing a predicate as a procedure that maps
a call type to a success type. This captures some dependencies between arguments, but
they are restricted to monomorphic types. It would be interesting to build this idea into
our approach. We believe that some imprecision found in analyzing recursive calls with
input arguments could be alleviated by using equations which captured both calls and
successes. The relation of these with directional types might be worth investigating.
Bruynooghe, Gallagher, Humbeeck, and Schrijvers [3,21] develop analyses for type
inference which are also based on the set constraints approach. However, their analyses
infer well-typings, so that the result is not an approximation of the program success set,
as in our case. In consequence, their algorithm is simpler, mainly because they do not
need to deal with intersection. In [21] the monomorphic analysis of [3] is extended to the
polymorphic case, using types with an expressiveness comparable to our parameterized
types.
The set constraints approach to type analysis is also taken in NFTA [9], which is
probably the closest to ours. However, the types inferred in this analysis, which are approximations of the program success set, are not parametric and as a result the analysis
is less precise than ours, as shown in our experiments. The work of [2] aims at inferring
parametric types that approximate the success set. Although independently developed,
it includes some of the ideas we propose. However, the authors do not use the set-based
approach, and their inference algorithm is rather complex. The latter made the authors
abandon that line of work and the approach was not developed, switching instead to
well-typings.
6 Concluding remarks and future work
To conclude, using the set constraints approach we have proposed a simple type inference based on the rewriting of equations which, despite its simplicity, allows enhanced
expressiveness. This is achieved by using type variables with a global scope as true parameters of the equations. The improved expressiveness allows for better precision, at
an additional cost in efficiency, which is, nevertheless, not high.
During our tests, we have identified potential bottlenecks in our analysis which may
appear in large programs and which are worth investigating in future research. Some
practical issues were already addressed by Heintze in [14], that we have not considered,
concentrating instead in this paper on the precision and soundness of our approach. In
particular, it is well-known that a naive DNF expansion may make the size of an expression grow exponentially. However, we use simple rules for minimizing the number of
expressions (such as computing only intersections that will survive after simplification)
which work quite well in practice. Even for larger programs we expect these rules to
be effective, based on the experience of [14]. An efficient method for storing the new
intersections generated in line 7 in SIMP is also needed for the scalability of the system.
We use the same simple technique as [14]. Note that the size of the table is in the worst
case 2N where N is the number of original variables in the program. However, the size
of the table actually grows almost linearly in our experiments (we omitted these results
due to space limitations). Even if the size of the table were to grow faster, we believe
that we can mitigate this effect with similar techniques (e.g., use of BDDs) to [9,10],
because of the high level of redundancy. We have also observed that the replacement
performed in lines 9 and 28 of SOLVE may be expensive for large programs. We think
that there are many opportunities for reducing this limitation (e.g., dependency directed
updating [14]).
Acknowledgments: The authors would like to thank John Gallagher and the anonymous referees
for useful comments on previous versions of this paper. This work was funded in part by EU
projects FET IST-15905 MOBIUS, IST-215483 SCUBE, FET IST-231620 HATS, and 06042ESPASS, Ministry of Science projects 1TIN-2008-05624 DOVES, Ministry of Industry project
FIT-340005-2007-14, CAM project S-0505/TIC/0407 PROMESAS, and Singapore Ministry of
Education Academic Research Fund No. RP-252-000-234-112.
References
1. Alexander Aiken and T. K. Lakshman. Directional type checking of logic programs. In
In Proceedings of the 1st International Static Analysis Symposium, pages 43–60. SpringerVerlag, 1994.
2. M. Bruynooghe and J. Gallagher. Inferring Polymorphic Types from Logic Programs. In
International Symposium on Logic-based Program Synthesis and Transformation (LOPST
R’04). Preproceedings, July 2004.
3. M. Bruynooghe, J. Gallagher, and W. Humbeeck. Inference of Well-typings for Logic Programs with Application to Termination Analysis. In 12th International Static Analysis Symposium (SAS’05), volume 3672 of LNCS, pages 35–51. Springer-Verlag, 2005.
4. F. Bueno, D. Cabeza, M. Carro, M. Hermenegildo, P. López-Garcı́a, and G. Puebla (Eds.).
The Ciao System. Ref. Manual (v1.13). Technical report, School of CS (UPM), 2006. Available at http://www.ciaohome.org.
5. W. Charatonik. Directional Type Checking for Logic Programs: Beyond Discriminative
Types. In Proc. of ESOP 2000, pages 72–87. LNCS 1782, 2000.
6. P. Cousot and R. Cousot. Formal language, grammar and set-constraint-based program analysis by abstract interpretation. In Proc. ACM Conf. on Functional Programming Langguages
and Computer Architecture, pages 170–181. ACM Press, New York, NY, 1995.
7. P.W. Dart and J. Zobel. A Regular Type Language for Logic Programs. In Types in Logic
Programming, pages 157–187. MIT Press, 1992.
8. T. Frühwirth, E. Shapiro, M.Y. Vardi, and E. Yardeni. Logic programs as types for logic
programs. In Proc. LICS’91, pages 300–309, 1991.
9. J. Gallagher and G. Puebla. Abstract Interpretation over Non-Deterministic Finite Tree Automata for Set-Based Analysis of Logic Programs. In Int’l. Symp. on Practical Aspects of
Decl. Languages, number 2257 in LNCS, pages 243–261. Springer-Verlag, January 2002.
10. J. P. Gallagher, K. S. Henriksen, and G. Banda. Techniques for scaling up analyses based on
pre-interpretations. In ICLP, pages 280–296, 2005.
11. J.P. Gallagher and D.A. de Waal. Fast and precise regular approximations of logic programs.
In Pascal Van Hentenryck, editor, Proc. of the 11th International Conference on Logic Programming, pages 599–613. MIT Press, 1994.
12. N. Heintze and J. Jaffar. A finite presentation theorem for approximating logic programs. In
Proc. 17th POPL, pages 197–209, 1990.
13. N. Heintze and J. Jaffar. An engine for logic program analysis. In IEEE-LICS, pages 318–
328, 1992.
14. Nevin Heintze. Practical aspects of set based analyses. In International Joint Conference
and Symposium on Logic Programming, 1992.
15. Nevin Charles Heintze. Set based program analysis. PhD thesis, Pittsburgh, PA, USA, 1992.
16. G. Janssens and M. Bruynooghe. Deriving Descriptions of Possible Values of Program
Variables by means of Abstract Interpretation. Journal of Logic Programming, 13(2 and
3):205–258, July 1992.
17. T.K. Lakshman and U.S. Reddy. Typed prolog: A semantic reconstruction of the MycroftO’Keefe type system. In International Logic Programming Symposium. MIT Press, 1991.
18. P. Mishra. Towards a theory of types in Prolog. In Int’l. Symp. on Logic Programming, pages
289–298, Silver Spring, MD, February 1984. IEEE Computer Society.
19. P. Mishra and U. Reddy. Declaration-Free Type Checking. In 12th ACM Symposium on
Principles of Programming Languages, pages 7–21, January 1985.
20. H. Saglam and J. Gallagher. Approximating logic programs using types and regular descriptions. Technical Report CSTR-94-19, Department of Computer Science, University of
Bristol, Bristol BS8 1TR, 1994.
21. T. Schrijvers, M. Bruynooghe, and J. Gallagher. From Monomorphic to Polymorphic Welltypings and Beyond. In LOPSTR’08. Springer-Verlag, 2008.
22. P. Van Hentenryck, A. Cortesi, and B. Le Charlier. Type analysis of prolog using type graphs.
In ACM SIGPLAN’94 Conference on Programming Language Design and Implementation,
pages 337–348. ACM SIGPLAN Notices Vol. 29 No. 6, 1994.
23. C. Vaucheret and F. Bueno. More Precise yet Efficient Type Inference for Logic Programs.
In International Static Analysis Symposium, volume 2477 of Lecture Notes in Computer
Science, pages 102–116. Springer-Verlag, September 2002.
24. J. Zobel. Derivation of Polymorphic Types for Prolog Programs. In Int’l. Conf. on Logic
Programming, pages 817–838. MIT Press, May 1987.
| 6cs.PL
|
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.5, September 2017
NEURO FUZZY MODELLING FOR
PREDICTION OF CONSUMER PRICE INDEX
Godwin Ambukege, Godfrey Justo and Joseph Mushi
Department of
Computer Science and Engineering, University of Dar es Salaam, Tanzania.
ABSTRACT
Economic indicators such as Consumer Price Index (CPI) have frequently used in predicting future
economic wealth for financial policy makers of respective country. Most central banks, on guidelines of
research studies, have recently adopted an inflation targeting monetary policy regime, which accounts for
high requirement for effective prediction model of consumer price index. However, prediction accuracy by
numerous studies is still low, which raises a need for improvement. This manuscript presents findings of
study that use neuro fuzzy technique to design a machine-learning model that train and test data to predict
a univariate time series CPI. The study establishes a matrix of monthly CPI data from secondary data
source of Tanzania National Bureau of Statistics from January 2000 to December 2015 as case study and
thereafter conducted simulation experiments on MATLAB whereby ninety five percent (95%) of data used
to train the model and five percent (5%) for testing. Furthermore, the study use root mean square error
(RMSE) and mean absolute percentage error (MAPE) as error metrics for model evaluation. The results
show that the neuro fuzzy model have an architecture of 5:74:1 with Gaussian membership functions (2, 2,
2, 2, 2), provides RMSE of 0.44886 and MAPE 0.23384, which is far better compared to existing research
studies.
KEYWORDS
Fuzzy inference system, Neural Networks, Neural Fuzzy, Consumer Price Index, Prediction,
1. INTRODUCTION
Economists and policy makers around the globe put efforts on understanding economic factors,
such as conditions in foreign trade, marketable surplus of agriculture, and noneconomic, to
improve and maintain economic status of their respective countries [1]. Accurate prediction of
macro-economic indicators have proven to support good understanding of economy. To date,
various economic indicators of interest can predicted for specific sectors of economy include
Consumer Price Index (CPI), inflation rates, Gross Domestic Product (GDP), birth rates,
unemployment rates, and stock markets [2]. Among these economic indicators, CPI is the key
economic indicator that measures the change over time in the purchasing cost of a fixed basket of
goods and services that are consumed by a representative sample of households in a given
country [3].
CPI is the measure of price level of goods and services in market basket that usually used to
determine inflation rate in respective community. Various researches have proven that the use of
CPI is helpful in predicting future developments in business cycle [4]. Thus, accurate prediction
DOI : 10.5121/ijaia.2017.8503
33
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.5, September 2017
of CPI plays a significant role for better planning of economic strategies in almost every
institution such as government, financial institutions, academia, investors in industries and
agriculture, savings and credit cooperative unions, and individual consumer as well.
Various CPI prediction models have been proposed, however, their prediction accuracy is still
low. This can be justified from the study by [5], [6], and [7]. The low prediction accuracy for
these models can be caused by factors such as small volume of data used, improper selection of
optimization techniques, the use of single prediction technique independently, selection for
number of input variables or due to the expertise of the model designer.
Therefore, in this research study, the Artificial Intelligence (AI) approach was applied by
combining neural networks and fuzzy logic to develop a neuro fuzzy model so as to improve
prediction accuracy of CPI. The sample data was taken from Tanzania National Bureau of
Statistics (TNBS).
1.1 CPI in Tanzania
Tanzania started to compile National Consumer Price Index (NCPI) on yearly basis since 1965,
then on quarterly basis from 1974 to 1994 and from then to date the NCPI is calculated on
monthly basis and released to the public on 8th day of the subsequent month [8]. NCPI is also
known as total CPI or all item index, for simplicity, in this paper, the CPI will mean NCPI. CPI is
used to compute an average measure of price inflation for the household sector as a whole, to
adjust wages as well as social security and other benefits to compensate for the changes, normally
rising of consumer prices, it acts as one of macro-economic indicators in adjustment of
government fees and charges, adjustment of payments in commercial contracts, and international
comparison [9].
The CPI is the weighted average of the price of goods and services computed by taking the cost
of market basket in a given year dividing the cost of market basket in a base year or reference
year and the results multiplying by 100. For example if we have two goods A and B in the market
basket and that CY denotes the expenditure of A and B in current year and BY denotes the
expenditure of A and B in base year then, the CPI for current year is given by equation 1
CPI = (CY / BY) * 100
(1)
The CPI for each group is calculated using formula in equation 1, and the total CPI is then
computed as the sum of the product of group CPI with its respective weight. If CPIc represents
current CPI for one group and Wc denotes its associated weight, then, the total CPI (NCPI) for
groups of goods and services n to k is calculated using equation 2.
(2)
1.2 Adaptive Neural Fuzzy Inference Systems
A neural fuzzy model is a hybrid model which is found by combining Artificial Neural Networks
and Fuzzy Logic techniques [10]. The neural fuzzy systems, also known as network based fuzzy
systems, can be trained to develop IF-THEN fuzzy rules and determine membership functions for
input and output variables of the system. Basically the neuro fuzzy system is a fuzzy system
34
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.5, September 2017
which contains membership functions that are tuned using training ability of neural networks.
Research studies have shown that hybrid models perform better when solving a particular
problem compared to an individual model alone [11], [12]. The integration of two techniques
overcomes the constraints of individual model by hybridization of various methods. The
motivation on developing a network based fuzzy system comes from the fact that, artificial neural
networks are good at recognizing patterns, they have a clear established architecture with learning
and generalisation ability. However, they are not good at explaining how they reach their
decision. Neuro fuzzy models can be applied to solve problems in fields of medicine, economic
forecasting, defence techniques by military and in engineering design.
1.3 Architecture of Neuro Fuzzy Systems
Network based fuzzy systems combine the merits of connectionist neural networks and fuzzy
approaches as a soft computing component, and rule generation from ANN has become popular
due to its capability of providing some insight to the user about the knowledge embedded within
the network [13]. Hybrid models are important when considering wide range of application
domains. The architecture of neuro fuzzy system is similar to multi-layered neural network with
additional of two adaptive layers. For simplicity we assume neuro fuzzy system with two inputs
(x and y) and one output (f), the architecture consists of five layers as shown in Figure 1. Two IF
THEN rules can be considered to represent this architecture.
Rule (1) IF x is A1 and y is B1 THEN f1 = p1x + q1y + r1.
Rule (2) IF x is A2 and y is B2 THEN f2 = p2x + q2y + r2.
Where: x and y are inputs, Ai and Bi are the fuzzy sets, fi are outputs, pi, qi, and ri are consequent
parameters that are determined during training process.
The first layer accepts crisp inputs and converts to the corresponding membership functions
(MF). This is an adaptive layer which fuzzifies inputs to membership grades of linguistic
variables. The output O1 of this layer is the fuzzy set of the inputs and is given by equation 3.
O1,i = μAi(x) for i = 1, 2
O1,i = μBi(y) for i = 1, 2
(3)
Where μA(x) and μB(y) are membership functions which can be of any type such as Gaussian
membership function or Bell shaped. For example, for Gaussian MF, it is given by equation 1.2.
(4)
Where ci and σi are centre and width of the ith fuzzy set Ai respectively.
The second layer (represented by π) is the fuzzy rule layer and each neuron corresponds to a
single fuzzy rule. A fuzzy rule neuron receives inputs from the fuzzification neurons that
represent fuzzy sets in the rule antecedents. The layer calculates the weight of each rule. The
output of this layer is calculated as the product of the incoming signal.
O2,i = wi = μA,i(x). μBi(y) , i = 1, 2
(5)
35
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.5, September 2017
The third layer (represented by N) is called normalized layer. This layer calculates the firing
strengths of each rule from previous layers. The output of this layer is computed using equation 6
O3,i = wi = wi/(w1+w2) , i = 1, 2
(6)
The fourth layer is the output membership layer (y). Neurons in this layer represent fuzzy sets
used in the consequent of fuzzy rules. The output of this layer is the product of the normalised
firing strength of the rule obtained in previous layer and first order polynomial.
Figure 1: The Architecture of Neuro Fuzzy System
O4,i = wifi = wi(piv + qid + ri) , i = 1, 2
(7)
Where fi is the node function, p, q, r are consequent parameters which are determined
during training process.
The last layer is the defuzzification layer which represents a crisp output of the neuro fuzzy
system. This layer sums up all incoming signal and is computed using equation 8.
O5,i = ∑wifi= ∑wifi/(w1+w2), i =1, 2
(8)
2. METHODOLOGY
2.1 Data Preparation
The first step was to prepare the data. Sample data were CPI data points for sixteen years from
January 2000 to December 2015. The data collected were transformed into matrix vector having
one row and many columns. The data were pre-processed so as to be acceptable by the Matrix
Laboratory (MATLAB) computing platform. The data were separated into training samples
(95%) and testing samples (5%). The data from January 2000 to March 2015 were taken as
training data sets and from April 2015 to December 2015 were for model validation. One month
36
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.5, September 2017
ahead forecast technique was selected for both in sample and out of sample data. For instance,
CPI data from January 2000 to May 2000 were used to forecast CPI for June 2000, from February
2000 to June 2000 was used to forecast CPI for July 2000 and so on. This means that, the
forecasted data point at one month were taken as independent variable in predicting CPI for the
next month. In this research work, MATLAB software was used for simulating the experiments in
the process of designing and testing neuro fuzzy model. Root Mean Square Error (RMSE) and
Mean Absolute Percentage Error (MAPE) were used as error metric as proposed by [14], [15],
[16].
2.2 Training and Testing the Model
The process of model development can be explained using Figure 2. After defining the inputs and
output sets for both training and testing, the parameters for membership functions (MF) were
initialized, the appropriate number of input variables and output variable were selected. Both
commands (genfis1 and genfis2) were used for initializing membership function parameters to
generate Fuzzy Inference System (FIS) and the performance for each method was recorded.
Figure 2: Model development process
37
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.5, September 2017
Both Back propagation (BP) and hybrid method which combines BP and least square were used
as training optimization methods. The training process was done using anfis command and
different number of training epochs were tested while observing the values of Root Mean Square
Error (RMSE) and Mean Absolute Percentage Error (MAPE). The error goal was set to zero,
initial step size to 0.01, step decrease rate 0.9 and step size increase rate 1.1. These training
parameters are important in controlling the learning process of the model.
After finishing the training process, the final membership functions and training error from the
training data sets were produced. The command evalfis was applied to study the performance by
applying inputs data to the fuzzy system without output data. The output produced here is the
system output and was compared against the desired output for training and test data sets. The
training errors were calculated using the formula in equation 9 and 10.
(9)
Where xi and yi are the values of the ith observations in x (predicted value) and y (target value),
respectively and N is the total number of observations.
(10)
Where Abs denotes absolute value, At is the actual value (target) and Ft is the forecasted value, n
is the number of observations.
If the training error is not satisfactory the training process repeats. A number of simulation
experiments were conducted while changing the training parameters such as number of inputs,
type of membership functions, membership grades, and number of training epochs, in each
experiment so as to obtain satisfactory training results. Table 1 shows sample training parameters.
Table 1: Neuro Fuzzy Training Parameters
Sn
4
5
Parameter
Fuzzy system
generation method
Type of Membership
Function (MF)
Number of inputsoutputs
Number of MF per
input
Training algorithm
6
Number of training
epoch
1
2
3
Test cases
genfis1
genfis2
Triangular
Trapezoidal
Generaliz
ed Bell
Gaussian
2-1 => 6-1
2 => 4
Backpropagation
Hybrid
(Backpropagat
ion + Least
Square
30 => 5,000
38
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.5, September 2017
3. RESULTS AND DISCUSSION
3.1Training and Test Results
Table 2 shows the results of sample simulation experiments conducted to train and test different
neuro fuzzy models using the MATLAB software. This table does not include all simulation
experiments that were conducted; the table indicates few of them.
Table 2: Sample results for simulation experiments
Sn
MF
MFs per
input
Epochs
Architect
ure
RMSE
Training
RMSE
Testing
MAPE
Testing
1
Gaussian
3,3,3,2
500
4:119:1
0.36528
15.4779
8.312542
2
Triangular
3,3,3,3
450
4:174:1
0.44014
112.042
64.6368
3
Triangular
3,2,2,2,2
1000
5:107:1
0.40803
54.0926
31.12558
4
Generalize
d bell
3,3,2,3,2
750
5:229:1
0.18846
60.3491
34.97454
5
Generalize
d bell
3,3,2,2,2
1500
5:156:1
0.19256
65.0363
34.08347
6
Triangular
2,2,2,2
1000
4:40:1
0.59783
168.6596
97.1052
7
Generalize
d bell
2,3,3,2,2
500
5:156:1
0.24493
45.7328
26.28622
8
Gaussian
2,2,2,2,2
1000
5:74:1
0.40345
0.68746
0.332627
9
Gaussian
2,2,2,2,2
650
5.74:1
0.64241
0.44886
0.233839
10
Generalize
d bell
2,2,2,2,2
650
5:74:1
0.41263
1.183
0.649738
11
Generalize
d bell
2,2,3,3
650
4:82:1
0.45368
20.5294
12.25982
On average, genfis1 produced better results than genfis2. The triangular membership functions
produced a higher forecasting error compared to other membership functions, Gaussian
membership functions (2, 2, 2, 2, 2) in record number 9, was selected as the best model since it
performed well compared to others. The model has RMSE of 0.44886 and MAPE of 0.233839.
The bell shaped having 5 inputs and membership functions (3,3,2,3,2) shown in record number 4
produced good training results at 750 epochs but was not selected due to huge forecasting error
(RMSE of 60.3491). Over fitting was noticed when the number of training epochs were increased
39
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.5, September 2017
to 900 and 1000 for Gaussian membership functions (2,2,2,2,2) of which the results was not good
during testing (RMSE of 0.68746 and MAPE of 0.332627).
The obtained results imply that, some membership functions such as triangular do not suit well in
modeling some applications such as those related to predictions, for this case CPI prediction. Also
when modeling neural fuzzy, the model with low error during training does indicate it is the
best model, for example, in this case; the training error was 0.64241 in record 9 of Table 4.2 and
0.18846 in record 4. The model in record 4 was not selected as the best model due to high error
during testing (60.3491).
In this work, a new Sugeno type neuro fuzzy model was developed to predict the consumer price
index. The architecture of the proposed model is shown in Figure 2. The selected Sugeno FIS has
five layers, having 5 neurons in input layer, 10 neurons in input membership functions, 32 IF
THEN rules, 32 neurons in output membership functions and 1 neuron in output layer. This
architecture is represented as 5:74:1 in three layers model which means 5 input neurons, 74
neurons in hidden layers and 1 output neuron. The architecture has 2 Gaussian membership
functions per input, which correspond to the set of (low and high) for CPI. Figure 3 shows
portions of the rule viewer (rules) to represent the relationship between inputs and outputs. This
indicates that, for any given five consecutive months’ CPI, with the proposed model, it is possible
to predict the next month CPI.
Figure 2: Architecture of proposed model
40
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.5, September 2017
Figure 3: Rule viewer to interpret the rules
3.2 Prediction Results by Proposed Model
Table 3: Forecasting Results for Selected Neuro Fuzzy Model
Month
Target
Predicted
Error
Apr-15
157.21
156.5460181
0.663981936
May-15
157.86
157.8974576
-0.037457551
Jun-15
158.12
158.2561941
-0.136194111
Jul-15
158.78
158.5041795
0.275820524
Aug-15
158.81
159.3802297
-0.570229734
Sep-15
159.04
159.2140768
-0.174076814
Oct-15
159.17
159.5692438
-0.399243834
Nov-15
160.49
159.6562615
0.833738462
Dec-15
161.24
161.4976339
-0.257633946
41
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.5, September 2017
Table 4 Performance of Chosen Neuro Fuzzy Model
Model
Membership functions
RMSE
MAPE
Neuro fuzzy (5:74:1)
Gaussian (2, 2, 2, 2, 2)
0.44886
0.23383928
As shown in Table 3, the chosen model was then used to predict CPI for 9 months from April
2017 to December 2015. As shown, the prediction errors are small. Table 4 shows the average
prediction errors for 9 months in RMSE (0.44886) and MAPE (0.233839)
3.3 Results Discussion
In this research work, the neuro fuzzy models developed by [7], [17] and [18], have used as the
benchmark. The results of the developed neuro fuzzy model was compared with these research
works on CPI prediction and found that the proposed model performs better.
[7] researched on neuro fuzzy modelling technique using CPI data for a period of 15 years from
January 2000 to January 2014. They used seven input variables and one output to forecast the US
future CPI using subtractive clustering (genfis2) method for generating FIS. They trained the
network for 10,000 epochs and obtained the values for RMSE = 0.837. The use of large number
of inputs (7) per training pair, and genfis2 FIS initialization method during modelling process
could be the cause of low prediction results.
In this research work has found that, in modelling neuro fuzzy for CPI prediction, Gaussian
membership functions provide better prediction results when we use grid partition (genfis1) as
FIS generation method, low number of inputs per training vector (5 in this case) and small
number of membership functions per input (2 in this research). Table 5 shows prediction accuracy
comparison for the proposed work with existing one.
Table 5 Comparison of Proposed Model with Other Models
Prediction technique
RMSE
MAPE
Neuro fuzzy
3.1865
2.9771
Neuro fuzzy
1.4679
(Rosadi,
Subanar,
Suhartono, 2013)
Neuro fuzzy
0.829
(Enke and Mehdiyev, 2014)
Neuro fuzzy
0.44886
0.23384
Reference
(Ucenic and Atsalakis, 2009)
and
Proposed
42
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.5, September 2017
4. CONCLUSION
This research study applied the concepts of artificial intelligence, specifically, the neuro fuzzy
technique. The study has presented step by step procedures on how to develop the forecasting
model for consumer price index using this technique. The data for this research work were taken
from the Tanzania National Bureau of Statistics for 16 years from January 2000 to December
2015. The Root Mean Square Error and Mean Absolute Percentage Error were used as accuracy
metrics to evaluate the performance of the developed models. The findings from this study were
compared with existing research works on modelling neuro fuzzy for CPI prediction and showed
that the proposed model provides better prediction accuracy.
This research work has shown an improvement on prediction accuracy of the consumer price
index. The findings of this research should be used as the basis for building an application that
can be used for predictions of consumer price index in real time environment. The appropriate
authorities such as central banks and statistical bureaus which compile macro-economic
indicators, specifically the consumer price index, should use the findings of this study as a guide
in the prediction of consumer price index.
REFERENCES
[1] Ngowi, H. P. (2009). Economic Development And Change In Tanzania Since Independence: The
Political Leadership Factor. African Journal Of Political Science And International Relations, 3(4),
259–267.
[2] Ababio, K. A. (2012). Comparative Study Of Stock Price Forecasting Using Arima And Arimax
Models. Kwame Nkrumah University Of Science And Technology.
[3] Tanzania National Bureau Of Statistics. (2015). National Consumer Price Index For July 2015.
[4] Olsen, A. R. (2009). Business Cycle Forecasting Through Economic Indicators: A Dynamic Approach.
Copenhagen Business School.
[5] Edward, N. (2011). Modelling And Forecasting Using Time Series Garch Models : An Application Of
Tanzania Inflation Rate Data. University Of Dar Es Salaam.
[6] Haider, A., & Hanif, M. N. (2009). Inflation Forecasting In Pakistan Using Artificial Neural Networks.
Pakistan Economic And Social Review, 47(1), 123–138.
[7] Enke, D., & Mehdiyev, N. (2014). A Hybrid Neuro-Fuzzy Model To Forecast Inflation. In Procedia
Computer Science (Vol. 36, Pp. 254–260). Philadelphia: Elsevier Masson Sas.
[8] Tanzania National Bureau Of Statistics. (2016). Rebasing Of National Consumer Price Index. Dar Es
Salaam.
[9] Msokwa, Z. E. (2012). Fixed Basket Laspeyres ’ Method Compared To Modified Laspeyres ’ Method
In Computing Consumer Price Indices. Dar Es Salaam.
[10] John, M. T. (2008). Artificial Intelligence: A System Approach (1st Ed., Vol. 180). New Delhi:
Infinity Science Press Llc.
43
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.5, September 2017
[11] Delnavaz, B. (2014). Forecasting Of The Stock Price Index By Using Fuzzy-Neural Network And
Genetic Algorithms. Journal Of Applied Science And Agriculture, 9(July), 109–117.
[12] Fuller, R. (2000). Introduction In Neural Fuzzy Systems (2nd Editio). Berlin: Springer Science &
Business Media.
[13] Walia, N., Kumar, S., & Singh, H. (2015). A Survey On Applications Of Adaptive Neuro Fuzzy
Inference System. International Journal Of Hybrid Information Technology, 8(11), 343–350.
[14] Chai, T., & Draxler, R. R. (2014). Root Mean Square Error (Rmse) Or Mean Absolute Error (Mae)? –
Arguments Against Avoiding Rmse In The Literature. Geoscientific Model Development, 7(3), 1247–
1250.
[15] Gunawardana, A., & Shani, G. (2009). A Survey Of Accuracy Evaluation Metrics Of
Recommendation Tasks. The Journal Of Machine Learning Research, 10, 2935–2962.
[16] Small, G. R., & Wong, R. (2002). The Validity Of Forecasting. In International Conference
Christchurch (Pp. 1–14). Sydney University Of Technology.
[17] Rosadi, D., Subanar, T., & Suhartono. (2013). A Nalysis Of Financial Time Series Data Using
Adaptive Neuro Fuzzy Inference System ( Anfis ). International Journal Of Computer Science Issues,
10(2), 491–496.
[18] Ucenic, C. I., & Atsalakis, G. (2009). Forecasting Cpi Using A Neural Network With Fuzzy Inference
System. In L. Sakalauskas, C. Skiadas, & E. K. Zavadskas (Eds.), Applied Stochastic Models And
Data Analysis (Pp. 309–315). Vilnius.
AUTHORS
Godwin Ambukege received MSC (Computer Science) at
University of Dar es
Salaam (UDSM) in Tanzania, PGD (IT) at Amity University (India). He currently works
as Senior Systems administrator at Teofilo Kisanji University in Mbeya Tanzania. He is
also a member of Embedded and Intelligent systems research group at UDSM. His
research interests include; Artificial intelligence, machine learning, expert systems and
data analysis.
Dr. Godfrey Justo received his PHD (London), MSC (Math) Zimbabwe. He is
Engagement and Management Advisor at Tanzania Data Lab (dLab). He is also a senior
lecturer and expert in Software Engineering at UDSM, College of ICT. He has been
working with various organisations such as World Bank, Japan International
Cooperation Agency (JICA) and German Academic Exchange Service (DAAD) as
consultant.
Joseph Mushi received his Msc. Degree in Communications and Information Systems
Engineering at Wuhan University of Technology, Wuhan, P.R. China and Bsc.
Computer Science at UDSM, Tanzania. He is the Facility Manager of Tanzania Data
Lab (dLab) with background in Data and Statistical Analysis. He has been involved in
numerous open data related activities as technical advisor, including capacity building
programmes funded by World Bank and HDIF. He also holds an assistant lecturer
position at the University of Dar es Salaam (UDSM), Department of Computer Science
and Engineering.
44
| 2cs.AI
|
Generalizing the Kawaguchi-Kyan bound to
stochastic parallel machine scheduling
Sven Jäger∗ and Martin Skutella†
arXiv:1801.01105v2 [cs.DM] 4 Jan 2018
Institut für Mathematik, Technische Universität Berlin
January 8, 2018
Abstract
Minimizing the sum of weighted completion times on m identical parallel machines is one of the most important and classical scheduling problems. For the stochastic variant where processing times of jobs are random variables, Möhring, Schulz, and Uetz (1999) presented the first and
still best known approximation result achieving, for arbitrarily many machines, performance ratio 1 + 21 (1 + ∆), where ∆ is an upper bound on
the squared coefficient of√variation of the processing times. We prove
performance ratio 1 + 12 ( 2 − 1)(1 + ∆) for the same underlying algorithm – the Weighted Shortest Expected Processing Time (WSEPT) rule.
For the special case of deterministic scheduling
(i.e., ∆ = 0), our bound
√
matches the tight performance ratio 12 (1 + 2) of this algorithm (WSPT
rule), derived by Kawaguchi and Kyan in a 1986 landmark paper. We
present several further improvements for WSEPT’s performance ratio,
one of them relying on a carefully refined analysis of WSPT yielding, for
every fixed number
of machines m, WSPT’s exact performance ratio of
√
order 21 (1 + 2) − O(1/m2 ).
1
Introduction
In an archetypal machine scheduling problem, n independent jobs have to be
scheduled on m identical parallel machines or processors. Each job j is specified
by its processing time pj > 0 and by its weight wj > 0. In a feasible schedule,
every job j is processed for pj time units on one of the m machines in an
uninterrupted fashion, and every machine can process at most one job at a
time. The completion time of job j in some schedule S is denoted by CjS . The
goal isPto compute a schedule S that minimizes the total weighted completion
n
S
time
j=1 wj Cj . In the standard classification scheme of Graham, Lawler,
Lenstra,
P and Rinnooy Kan [8], this NP-hard scheduling problem is denoted by
P || wj Cj .
∗ [email protected]
† [email protected]
1
Weighted Shortest Processing Time Rule. By a well-known result of
Smith [24], sequencing the jobs in order of non-increasing ratios wj /pj gives an
optimal single-machine schedule. List scheduling in this order is known as the
Weighted Shortest Processing Time (WSPT) rule
√ and can also be applied to
identical parallel machines, where it is a 12 (1 + 2)-approximation algorithm;
see Kawaguchi and Kyan [14]. A particularly remarkable aspect of Kawaguchi
and Kyan’s work is that, in contrast to the vast majority of approximation
results, their analysis does not rely on some kind of lower bound. Instead, they
succeed in explicitly identifying
√ a class of worst-case instances. In particular,
the performance ratio 12 (1 + 2) is tight: For every ε > 0 there is a problem
√
instance for which WSPT has approximation ratio at least 12 (1 + 2) − ε. The
instances achieving these approximation ratios, however, have large numbers
of machines when ε becomes small. Schwiegelshohn [20] gives a considerably
simpler version of Kawaguchi and Kyan’s analysis.
Stochastic Scheduling. Many real-world machine scheduling problems exhibit a certain degree of uncertainty about the jobs’ processing times. This
characteristic is captured by the theory of stochastic machine scheduling, where
the processing time of job j is no longer a given number pj but a random variable pj . As all previous work in the area, we always assume that these random
variables are stochastically independent. At the beginning, only the distributions of these random variables are known. The actual processing time of a job
becomes only known upon its completion. As a consequence, the solution to
a stochastic scheduling problem is no longer a simple schedule, but a so-called
non-anticipative scheduling policy. Precise definitions on stochastic scheduling
policies are given by Möhring, Radermacher, and Weiss [16]. Intuitively, whenever a machine is idle at time t, a non-anticipative scheduling policy may decide
to start a job of its choice based on the observed past up to time t as well as the
a priori knowledge of the jobs processing time distributions and weights. It is,
however, not allowed to anticipate information about the future, i.e., the actual
realizations of the processing times of jobs that have not yet finished by time t.
It follows from simple examples that, in general, a non-anticipative scheduling policy cannot yield an optimal schedule for each possible realization of the
processing times. We are therefore looking for a policy which minimizes the
objective in expectation. For the stochastic scheduling problem considered in
this paper, the goal is to find a non-anticipative scheduling policy that minimizes the expected
P total weighted completion time. This problem is denoted by
P |pj ∼ stoch|E[ wj Cj ].
Weighted Shortest Expected Processing Time Rule. The stochastic
analogue of the WSPT rule is greedily scheduling the jobs in order of nonincreasing ratios wj /E[pj ]. Whenever a machine is idle, the Weighted Shortest
Expected Processing Time (WSEPT) rule immediately starts the next job in
this order. For a single machine this is again optimal; see Rothkopf [18]. For
identical parallel machines, Cheung, Fischer, Matuschke, and Megow [3] and
Im, Moseley, and Pruhs [11] independently show that WSEPT does not even
achieve constant performance ratio. More precisely, for every R > 0 there is
a problem instance for which WSEPT’s expected total weighted completion
time is at least R times the expected objective function value of an optimal
2
non-anticipative scheduling policy. Even in the special case of exponentially
distributed processing times, Jagtenberg, Schwiegelshohn, and Uetz [12] show a
lower bound of 1.243 on WSEPT’s performance. On the positive side, WSEPT
is an optimal policy for the special case of unitP
weight jobs with stochastically
ordered processing times, P |pj ∼ stoch(st )|E[ Cj ]; see Weber, Varaiya, and
Walrand [25]. Moreover, Weiss [26, 27] proves asymptotic optimality of WSEPT.
Möhring, Schulz, and Uetz [17] show that the WSEPT rule achieves performance
1
), where ∆ is an upper bound on the squared coefficient
ratio 1 + 12 (1 + ∆)(1 − m
of variation of the processing times.
Further Approximation Results from
P the Literature. While there is a
PTAS for the deterministic problem P || wj Cj [23], no constant-factor
Papproximation algorithm is known for the stochastic problem P |pj ∼ stoch|E[ wj Cj ].
WSEPT’s performance ratio 1+ 12 (1+∆) (for arbitrarily many machines) proven
by Möhring et al. [17] is the best hitherto known performance ratio. The only
known approximation ratio not depending on the jobs’ squared coefficient of
variation ∆ is due
P to Im et al. [11] who, for the special case of unit job weights
P |pj ∼ stoch|E[ Cj ], present an O(log2 n+m log n)-approximation algorithm.
1
The performance ratio 1 +P
2 (1 + ∆) has been carried over to different generalizations of P |pj ∼ stoch|E[ wj Cj ]. Megow, Uetz, and Vredeveld [15] show
that it also applies if jobs arrive online in a list and must immediately and irrevocably be assigned to machines, on which they can be sequenced optimally.
An approximation algorithm with this performance ratio for the problem on
unrelated parallel machines is designed by Skutella, Sviridenko, and Uetz [22].
If these two features are combined, i.e., in the online list-model with unrelated
machines, Gupta, Moseley, Uetz, and Xie [9] develop a (8 + 4∆)-approximation
algorithm.
The performance ratios are usually larger if jobs are released over time: In
the offline setting, the best known approximation algorithm has performance
ratio 2 + ∆; see Schulz [19]. This performance ratio is also achieved for unrelated machines [22] and by a randomized online algorithm [19]. There exist
furthermore a deterministic (max{2.618, 2.309+1.309∆})-approximation for the
online setting [19] and a deterministic (144 + 72∆)-approximation for the online
setting on unrelated parallel machines [9].
Our Contribution and Outline. We present the first progress on the
approximability of the basic stochastic scheduling problem on identical
parallel machinesPwith expected total weighted completion time objective
P |pj ∼ stoch|E[ wj Cj ] since the seminal work of Möhring et al. [17]; see
Figure 1. We prove that WSEPT achieves performance ratio
(p
)
(2m
−
k
)k
−
k
1
m
m
m
√
,
(1 + ∆), (1)
1 + 21 min
m
1 + min{2, 2 + 2∆}
√
√
where km := 1 − 21 2 m is the nearest integer to 1 − 12 2 m. Notice that,
for every number of machines m, the performance ratio given
√ by the first term
of the minimum in (1) is bounded from above by 1 + 12 ( 2 − 1)(1 + ∆), and
√
√
for m → ∞ it converges to this bound. As (1 + min{2, 2 + 2∆})−1 ≤ 2 − 1
for all ∆ > 0, when considering an arbitrary number of machines, the second
3
[17]
performance ratio
5
3
1
1
+
(1
+
)
∆
2
)
3
2
1+
1+
4
3
1
2 (1
+
√
2)
1
6
(1
√ 2 − 1)
1(
+∆
[Cor. 1]
[Cor. 2]
[Cor. 3]
2
max{2, 1 + ∆}
[14] 1 +
1√
}
2+2∆
1
n{2,
i
m
2 1+
(1 +
∆)
1
∆
0
1
1
2
3
2
Figure 1: Bounds on WSEPT’s performance ratio
term in the minimum dominates the first term. In the following, we list several
points that emphasize the significance of the new performance ratio (1).
• For the special case of deterministic scheduling (i.e., ∆ = 0), the machineindependent √
performance ratio in (1) matches the Kawaguchi-Kyan
bound 12 (1 + 2), which is known to be tight [14]. In particular, we dissolve the somewhat annoying discontinuity of the best previously known
bounds [14, 17] at ∆ = 0; see Figure 1.
• Againp for deterministic jobs,
our machine-dependent bound
1 + 21 ( (2m − km )km − km )/m is tight and slightly improves the 30 years
old Kawaguchi-Kyan bound for every fixed number of machines m; see
Figure 2.
• For exponentially distributed processing times (∆ = 1), our results imply
that WSEPT achieves performance ratio 4/3. This solves an open problem
by Jagtenberg et al. [12], who give a lower bound of 1.243 on WSEPT’s
performance and ask for an improvement of the previously best known
upper bound of 2 − 1/m due to Möhring et al. [17].
• WSEPT’s performance bound due to Möhring et al. [17] also holds for
the MinIncrease policy, introduced by Megow et al. [15], which is a fixedassignment policy, i.e., before the execution of the jobs starts, it determines
for each job on which machine it is going to be processed. Our stronger
bound, together with a lower bound in [22], shows that WSEPT actually
beats every fixed-assignment policy.
4
performance ratio
1.21
1.2
1.19
m
1.18
0
2
4
6
8
10
12
14
16
18
20
22
24
p
Figure 2: Graph of the function m 7→ 1 + 12 ( (2m − km )km − kmP
)/m, which for
m ∈ N gives the worst-case approximation ratio of WSPT for P || wj Cj with m
machines (dots), compared to the machine-independent Kawaguchi-Kyan bound
The improved performance ratio in (1) is derived as follows. In Section 2 we
present one of the key results of this paper (see Theorem 1 below): If WSPT
has performance ratio 1 + β for some β, then WSEPT achieves performance ratio 1 + β(1 + ∆) for the√stochastic scheduling problem. For the Kawaguchi-Kyan
√
bound 1 + β = 12 (1 + 2), this yields performance ratio 1 + 12 ( 2 − 1)(1 + ∆).
It is also interesting to notice that the performance ratio of Möhring et al. [17]
follows from this theorem by plugging in 1+β = 3/2−1/(2m), which is WSPT’s
performance ratio obtained from the bound of Eastman, Even, and Isaacs [4].
We generalize Theorem 1 to performance ratios w.r.t. the weighted sum of
α-points as objective function, where the α-point of a job j is the point in
time when it has been processed for exactly αpj time units.
The theorems derived in Section 2 provide tools to carry over bounds for
the WSPT rule to the WSEPT rule. The concrete performance ratio for the
WSEPT rule obtained this way thus depends on good bounds for the WSPT
rule. In Section 3 we derive performance ratios for WSPT w.r.t. the weighted
sum of α-points objective. For α = 12 this performance ratio follows easily
from a result by Avidor, Azar, and Sgall [1] for the 2-norm of the machine load
vector. As a consequence we obtain performance ratio 1 + 16 max{2, 1 + ∆}
for WSEPT. By optimizing
the choice of α, we finally obtain the performance
√
ratio 1 + 21 (1 + min{2, 2 + 2∆})−1 (1 + ∆). The various bounds on WSEPT’s
performance derived in Sections 2 and 3 are illustrated in Figure 1. Finally, in
Section 4 the analysis of Schwiegelshohn [20] for the WSPT rule is refined for
every fixed number of machines m, entailing the machine-dependent bound for
the WSEPT rule in (1).
2
Performance ratio of the WSEPT rule
Let ∆ ≥ Var[pj ]/E[pj ]2 for all j ∈ {1, . . . , n}. In Theorems 1 and 2 we demonstrate how performance ratios for the WSPT rule for deterministic scheduling
can be carried over to stochastic scheduling. Theorem 1 starts out from a per-
5
formance ratio for WSPT with respect to the usual objective function: the
weighted sum of completion times. In Theorem 2 this is generalized insofar as
a performance ratio of WSPT with respect to the weighted sum of α-points is
taken as a basis.
Theorem 1. If theP
WSPT rule on m machines has performance ratio 1 + βm
for the problem P || wj Cj , then the
P WSEPT rule achieves performance ratio
1 + βm (1 + ∆) for P |pj ∼ stoch|E[ wj Cj ].
The reason why the bound for WSPT does not directly carry over to WSEPT
is that under a specific realization of the processing times the schedule obtained
by the WSEPT policy may differ from the WSPT schedule for this realization.
Still, under every realization the WSEPT schedule is a list schedule. Hence,
usually a bound that is valid for every list schedule is used: The objective
value of a list schedule on m machines is at most 1/m times the objective value
of the list schedule on a single machine plus (m − 1)/m times the weighted
sum of processing times. This bound, holding because a list schedule assigns
each job to the currently least loaded machine, is applied realizationwise to
obtain a corresponding bound on the expected values in stochastic scheduling
(cf. [17, Lemma 4.1]), which can be compared to an LP-based lower bound on
the expected total completion time under an optimal policy.
In order to benefit from the precise bounds known for the WSPT rule nevertheless, we regard the following modified stochastic scheduling problem: For
each job, instead of its weight wj , we are given a weight factor ρj . The actual weight of a job is ρj times its actual processing time, i.e., if a job takes
longer, it also becomes more important. The goal is again to minimize the total
weighted completion time. For the thus defined stochastic scheduling problem
list scheduling in order of ρ-values has the nice property that it creates a WSPT
schedule in every realization. So, for this scheduling problem any performance
ratio of WSPT directly carries over to this list scheduling policy. In the following proof of Theorem 1 we first compare the expected total weighted completion
time of a WSEPT schedule for the original problem to the expected objective
value list scheduling in order of ρj for the modified problem, then apply the
performance ratio of the WSPT rule, and finally compare the expected total
weighted completion time of the optimal schedule for the modified problem to
the expected objective value of the optimal policy for the original problem. The
transitions between the two problems lead to the additional factor of 1 + ∆ in
the performance ratio.
P
Proof. Consider an instance of P |pj ∼ stoch|E[ wj Cj ] consisting of n jobs
and m machines, and let β := βm and ρj := wj /E[pj ] for j ∈ {1, . . . , n}.
For every realization (p1 , . .P
. , pn ) of the processing times we consider the instance I(p1 , . . . , pn ) of P || wj Cj which consists of n jobs with processing
times p1 , . . . , pn and weights ρ1 p1 , . . . , ρn pn , so that the jobs in the instance
I(p1 , . . . , pn ) have Smith ratios ρ1 , . . . , ρn for all possible realizations. Therefore, for every realization (p1 , . . . , pn ) the schedule obtained by the WSEPT
policy is a WSPT schedule for I(p1 , . . . , pn ). Let CjWSEPT (p1 , . . . , pn ) denote
the completion time of job j in the schedule obtained by the WSEPT policy
under the realization (p1 , . . . , pn ), let Cj∗ (I(p1 , . . . , pn )) denote its completion
∗
time in an optimal schedule for I(p1 , . . . , pn ), and let CjΠ (p1 , . . . , pn ) denote j’s
6
completion time in the schedule constructed by an optimal stochastic scheduling policy under the realization (p1 , . . . , pn ). For every realization (p1 , . . . , pn )
of the processing times, since the WSEPT schedule follows the WSPT rule for
I(p1 , . . . , pn ), its objective value is bounded by
n
X
(ρj pj )CjWSEPT (p1 , . . . , pn ) ≤ (1 + β) ·
j=1
n
X
(ρj pj )Cj∗ (I(p1 , . . . , pn )).
j=1
On the other hand, as the schedule obtained by an optimal stochastic scheduling
policy is feasible for the instance I(p1 , . . . , pn ), it holds that
n
n
X
X
∗
(ρj pj )Cj∗ (I(p1 , . . . , pn )) ≤
(ρj pj )CjΠ (p1 , . . . , pn ).
j=1
j=1
Putting these two inequalities together and taking expectations, we get the
inequality
n
n
X
X
∗
E
ρj pj CjWSEPT ≤ (1 + β) · E
ρj pj CjΠ ,
j=1
j=1
∗
∗
where CjWSEPT = CjWSEPT (p1 , . . . , pn ) and CjΠ = CjΠ (p1 , . . . , pn ). Using the
latter inequality, we can bound the expected total weighted completion time of
the WSEPT rule:
n
n
X
X
E
wj CjWSEPT =
ρj E[pj ]E[CjWSEPT ]
j=1
(∗)
=
n
X
j=1
ρj E[pj CjWSEPT ] −
n
X
ρj Var[pj ]
j=1
j=1
n
n
X
X
= E
ρj pj CjWSEPT −
ρj Var[pj ]
j=1
j=1
n
n
X
X
∗
≤ (1 + β) · E
ρj pj CjΠ −
ρj Var[pj ]
j=1
= (1 + β) ·
n
X
j=1
∗
ρj E[pj CjΠ ] −
j=1
(∗)
= (1 + β) ·
n
X
n
X
∗
ρj E[pj ]E[CjΠ ] +
j=1
= (1 + β) ·
≤ (1 + β) ·
n
X
j=1
n
X
ρj Var[pj ]
j=1
n
X
ρj Var[pj ] −
j=1
∗
wj E[CjΠ ] + β
∗
n
X
ρj Var[pj ]
j=1
n
X
ρj Var[pj ]
j=1
n
X
wj E[CjΠ ] + ∆β
j=1
≤ (1 + β(1 + ∆)) ·
wj E[pj ]
j=1
n
X
∗
wj E[CjΠ
n
X
∗
] = (1 + β(1 + ∆)) · E
wj CjΠ .
j=1
j=1
7
The equalities marked with (∗) hold because for any stochastic scheduling policy Π and all j
E[pj CjΠ ] = E[pj (SjΠ + pj )] = E[pj SjΠ ] + E[p2j ]
= E[pj ]E[SjΠ ] + E[pj ]2 + Var[pj ] = E[pj ]E[CjΠ ] + Var[pj ],
where SjΠ denotes the starting time of job j under policy Π. The independence
of pj and SjΠ follows from the non-anticipativity of policy Π, and the last
∗
inequality uses the fact that E[pj ] ≤ E[CjΠ ] for every job j.
By plugging in the Kawaguchi-Kyan bound, we immediately get the following
performance ratio (see Figure 1).
√
Corollary 1. The WSEPT rulePhas performance ratio 1 + 12 ( 2 − 1)(1 + ∆)
for the problem P |pj ∼ stoch|E[ wj Cj ].
For α ∈ [0, 1] the α-point CjS (α) of a job j is the point in time at which it has
been processed for αpj time units. Introduced by Hall, Shmoys, and Wein [10] in
order to convert a preemptive schedule into a non-preemptive one, the concept
of α-points is often used in the design of algorithms (see e.g. [5, 2, 6, 21]). In
contrast, we use them to define an alternative objective function in order to
improve the analysis of the WSEPT rule.
Pn
We consider as objective function the weighted sum of α-points j=1 wj CjS (α)
for α ∈ [0, 1], generalizing the weighted sum of completion times. For
Pnevery α
the weighted sum of α-points differs only by the constant (1 − α) j=1 wj pj
from the weighted sum of completion times. So as for optimal solutions the
objective functions are equivalent. The same applies
the stochastic variant:
Pto
n
Here the two objectives differ by the constant (1−α) j=1 wj E[pj ], whence they
have the same optimal policies. We now generalize Theorem 1 to the (expected)
weighted sum of α-points.
Theorem
2. If the WSPT rule has performance ratio 1 + β for the problem
P
P || wj Cj (α), then the WSEPT
rule has performance ratio 1 + β(1 + ∆) for
P
the problem P |pj ∼ stoch|E[
w
C
j j (α)] and 1 + β · max{1, α(1 + ∆)} for the
P
problem P |pj ∼ stoch|E[ wj Cj ].
The
P proof uses the same idea as the proof of Theorem 1: The bound for
P || wj Cj (α) is again applied realizationwise to the modified stochastic problem described above.
Proof. We use the notation of the proof of Theorem 1 for α-points Cj (α) instead
of completion times Cj . Then we have by assumption that
n
n
X
X
(ρj pj )CjWSEPT (α)(p1 , . . . , pn ) ≤ (1 + β)
(ρj pj )Cj∗ (α)(I(p1 , . . . , pn ))
j=1
j=1
n
X
∗
≤ (1 + β)
(ρj pj )CjΠ (α)(p1 , . . . , pn ),
j=1
where the last inequality holds because the schedule obtained by the optimal
stochastic scheduling policy is feasible for I(p1 , . . . , pn ). This carries over to
8
expected values:
n
n
X
X
∗
E
ρj pj CjWSEPT (α) ≤ (1 + β) · E
ρj pj CjΠ (α) .
j=1
(2)
j=1
Furthermore, for any non-anticipative stochastic scheduling policy
E[pj CjΠ (α)] = E[pj SjΠ ] + αE[p2j ] = E[pj ]E[SjΠ ] + α(E[pj ]2 + Var[pj ])
= E[pj ]E[CjΠ (α)] + αVar[pj ]
=
E[pj ]E[CjΠ ]
(3)
2
+ αVar[pj ] − (1 − α)E[pj ] .
(4)
Therefore,
n
n
X
X
E
wj CjWSEPT (α) =
ρj E[pj ]E[CjWSEPT (α)]
j=1
(3)
=
n
X
j=1
ρj E[pj CjWSEPT (α)] − α
j=1
n
X
ρj Var[pj ])
j=1
n
n
X
X
ρj Var[pj ]
ρj pj CjWSEPT (α) − α
= E
j=1
j=1
n
n
X
X
(2)
∗
ρj pj CjΠ (α) − α
ρj Var[pj ]
≤ (1 + β) · E
j=1
= (1 + β) ·
n
X
j=1
∗
ρj E[pj CjΠ (α)] − α
j=1
(3)
= (1 + β) ·
n
X
ρj Var[pj ]
j=1
n
X
∗
ρj E[pj ]E[CjΠ (α)] + α
j=1
= (1 + β) ·
≤ (1 + β) ·
n
X
j=1
n
X
n
X
ρj Var[pj ] − α
j=1
∗
wj E[CjΠ (α)] + αβ
∗
ρj Var[pj ]
j=1
n
X
wj
E[pj ]
| {z }
j=1
≤
≤ (1 + β(1 + ∆)) ·
n
X
∗
wj E[CjΠ
ρj Var[pj ]
j=1
n
X
wj E[CjΠ (α)] + αβ∆
j=1
n
X
∗
E[C Π (α)]
j
α
n
X
∗
(α)] = (1 + β(1 + ∆)) · E
wj CjΠ (α) ,
j=1
j=1
9
and
n
n
X
X
E
wj CjWSEPT =
ρj E[pj ]E[CjWSEPT ]
j=1
(4)
=
n
X
j=1
ρj E[pj CjWSEPT (α)] −
j=1
n
X
ρj (αVar[pj ] − (1 − α)E[pj ]2 )
j=1
n
n
X
X
= E
ρj pj CjWSEPT (α) −
ρj (αVar[pj ] − (1 − α)E[pj ]2 )
j=1
j=1
n
n
X
X
(2)
∗
≤ (1 + β) · E
ρj pj CjΠ (α) −
ρj (αVar[pj ] − (1 − α)E[pj ]2 )
j=1
= (1 + β) ·
n
X
j=1
∗
ρj E[pj CjΠ (α)] −
j=1
(4)
= (1 + β) ·
∗
ρj E[pj ]E[CjΠ ] +
j=1
= (1 + β) ·
≤ (1 + β) ·
n
X
j=1
n
X
∗
n
X
j=1
j=1
n
X
∗
n
X
n
X
ρj (αVar[pj ] − (1 − α)E[pj ]2 )
wj E[CjΠ ] + β
ρj (αVar[pj ] − (1 − α)E[pj ]2 )
ρj (α∆ − (1 − α))E[pj ]2
j=1
∗
wj E[CjΠ ] + β(α(1 + ∆) − 1)
j=1
≤ (1 + β) ·
ρj (αVar[pj ] − (1 − α)E[pj ]2 )
j=1
= (1 + β) ·
n
X
j=1
wj E[CjΠ ] + β
n
X
ρj (αVar[pj ] − (1 − α)E[pj ]2 )
j=1
n
X
−
n
X
n
X
j=1
∗
wj
∗
∈[0,E[CjΠ ]]
wj E[CjΠ ] + β · max{0, α(1 + ∆) − 1}
j=1
E[pj ]
| {z }
n
X
∗
wj E[CjΠ ]
j=1
n
X
∗
≤ (1 + β · max{1, α(1 + ∆)}) · E
wj CjΠ .
j=1
Theorem 1 follows from Theorem 2 by plugging in α = 1.
3
Performance ratios for WSPT with weighted
sum of α-points objective
P
In this section we derive performance
ratios for P || wj Cj (α). The two clasP
sical performance ratios for P || wj Cj of Eastman, Even, and Isaacs [4] and
of Kawaguchi, and Kyan [14] can both be generalized to this problem. The
Eastman-Even-Isaacs bound can be generalized for every α ∈ (0, 1], whereas
10
the Kawaguchi-Kyan bound carries over only for α ∈ [ 12 , 1]. In return, the
generalized Kawaguchi-Kyan bound is better for these α.
For a problem instance I denote by N (I) the job set, by CjWSPT (α)(I) the αpoint of job j in the WSPT schedule for I, and by Cj∗ (α)(I) the α-point of job j
in some fixed (‘the’) optimal schedule for I. Hence CjWSPT (1)(I) = CjWSPT (I) is
the completion time of j and CjWSPT (0)(I) = SjWSPT (I) is the starting time of j
in the WSPT schedule, and analogously for the optimal schedule. Furthermore,
WSPT
let MiWSPT (I) and Mi∗ (I) denote the load of the i-th machine and Mmin
(I)
∗
and Mmin (I) denote the load of the least loaded machine, in the WSPT schedule
and the optimal schedule for I, respectively. Moreover, let WSPTα (I) and
OPTα (I) denote the weighted sum of α-points of the schedule obtained by
the WSPT rule and of the optimal schedule, respectively. Finally, denote by
λα (I) := WSPTα (I)/OPTα (I) the approximation ratio of the WSPT rule for
the instance I. The WSPT rule does not specify which job to take first if multiple
jobs have the same ratio wj /pj . Since we are interested in the performance of
this rule in the worst case, we have to prove the performance ratio for all possible
orders of these jobs. Hence, we may assume that this is done according to an
arbitrary order of the jobs given as part of the input.
It is a well-known fact (see e.g. [20]) that for the weighted sum of completion
times objective the worst case of WSPT occurs if all jobs have the same Smith
ratio wj /pj . This generalizes to the weighted sum of α-points objective.
P
Lemma 1. For every P
α ∈ [0, 1] and every instance I of P || wj Cj (α) there
is an instance I 0 of P || pj Cj (α) with the same number of machines such that
λα (I 0 ) ≥ λα (I).
Proof. The proof proceeds in the same way as Schwiegelshohn’s proof [20]. Assume that w1 /p1 ≥ w2 /pj ≥ · · · ≥ wn /pn and that the jobs are scheduled in this
order in the WSPT schedule for I. Then define the instances Ik0 , k = 1, . . . , n
consisting of k jobs with wj0 = p0j := pj . Then for every k ∈ {1, . . . , n} it holds
that CjWSPT (α)(Ik0 ) = CjWSPT (α)(I) for every job j ∈ {1, . . . , k}. Therefore, if
we set ρn+1 := 0, we get
WSPTα (I) =
n
X
wj CjWSPT (α)(I)
=
n
X
=
n
X
(ρk − ρk+1 ) pj CjWSPT (α)(I)
j=1
=
n
X
ρj pj CjWSPT (α)(I)
j=1
j=1
=
n
X
k=j
(ρk − ρk+1 )
k
X
k=1
j=1
n
X
k
X
(ρk − ρk+1 )
k=1
pj CjWSPT (α)(I)
pj CjWSPT (α)(Ik0 ) =
j=1
n
X
(ρk − ρk+1 )WSPTα (Ik0 ).
k=1
On the other hand, for every k ∈ {1, . . . , n}, scheduling every job j ∈ {1, . . . , k}
as in the optimal schedule for I is feasible for Ik0 . Hence, we can bound the
11
optimal objective value for Ik by OPTα (Ik0 ) ≤
OPTα (I) =
n
X
wj Cj∗ (α)(I) =
n
X
j=1
=
pj Cj∗ (α)(I). Therefore,
ρj pj Cj∗ (α)(I)
j=1
n
X
(ρk − ρk+1 ) pj Cj∗ (α)(I)
j=1
=
j=1
n
X
n
X
Pk
k=j
(ρk − ρk+1 )
k=1
k
X
pj Cj∗ (α)(I) ≥
j=1
n
X
(ρk − ρk+1 )OPTα (Ik0 )
k=1
because ρk − ρk+1 ≥ 0 for all k ∈ {1, . . . , n}. Hence,
Pn
(ρk − ρk+1 )WSPTα (Ik0 )
OPTα (I)
≤ Pk=1
λα (I) =
n
0
WSPTα (I)
k=1 (ρk − ρk+1 )OPTα (Ik )
0
WSPTα (Ik )
≤ max
= max λα (Ik0 ).
0
k∈{1,...,n} OPTα (Ik )
k∈{1,...,m}
For unit Smith ratio instances the WSPT rule is nothing but list scheduling
according to an arbitrary given order. Restricting to them has the benefit that
the objective value of a schedule S can be computed easily from its machine
loads, namely
n
m
X
1X S 2
pj CjS ( 12 ) =
(M ) .
(5)
2 i=1 i
j=1
This classical observation can for example be found in the paper of Eastman et
al. [4].
For the sum of the squares of the machine loads as objective function Avidor,
Azar, and Sgall [1] showed that WSPT has performance ratio 4/3. So this also
also holds for the weighted sum of 21 -points. By plugging it in into Theorem 2,
we get the following corollary.
Corollary 2. The WSEPT P
rule has performance ratio 1 + 16 max{2, 1 + ∆} for
the problem P |pj ∼ stoch|E[ wj Cj ].
Now we generalize the bound of Eastman, Even, and Isaacs [4].
Theorem 3 (Generalized Eastman-Even-Isaacs bound). For every α ∈ (0, 1]
the WSPT rule has performance ratio
m−1
1
1+
≤1+
2αm
2α
P
for the problem P || wj Cj (α).
Proof. By Lemma 1 we can restrict to the case that wj = pj for all jobs j. Let
I1 be the instance consisting of the same jobs but only one machine. On this
instance the WSPT schedule is optimal. Moreover, the following two inequalities
hold.
n
n
n
X
1 X
1 X
pj SjWSPT (I) ≤
pj SjWSPT (I1 ) =
pj Sj∗ (I1 ),
(6)
m
m
j=1
j=1
j=1
n
X
j−1
n
pj Cj∗ ( 21 )(I) ≥
1 X
pj Cj∗ ( 12 )(I1 ).
m j=1
12
(7)
The first holds because a WSPT schedule is a list schedule and the starting time
of any job j is at most the average machine load caused by all jobs preceding j
in the list. The second inequality, due to Eastman, Even, and Isaacs [4], follows
from Equation (5) and the convexity of the square function. Putting these two
inequalities together yields
n
X
pj CjWSPT (α)(I) =
n
X
j=1
pj SjWSPT (I) + α
j=1
n
X
n
X
j=1
(6)
p2j ≤
n
n
X
1 X
pj Sj∗ (I1 ) + α
p2j
m j=1
j=1
n
n
n
1 X 2 (7) X
1 X 2
1
pj Cj∗ ( 21 )(I1 ) + α −
pj ≤
p
=
pj Cj∗ ( 12 )(I) + α −
m j=1
2m j=1
2m j=1 j
j=1
=
n
X
n
pj Cj∗ (α)(I)
j=1
=
1+
m−1
2αm
n
n
m−1X 2 X
m−1X
+
pj ≤
pj Cj∗ (α)(I)
pj Cj∗ (α)(I) +
2m j=1
2αm j=1
j=1
X
n
pj Cj∗ (α)(I).
j=1
Remark. The generalized Eastman-Even-Isaacs bound does P
not lead to better
performance ratios for the WSEPT rule for P |pj ∼ stoch|E[ wj Cj ] than the
bound of Möhring et al. [17], as plugging in β = m−1
2αm into Theorem 2 leads to
a performance ratio of
m−1
1
1
1+
· max{1, α(1 + ∆)} ≥ 1 + (1 + ∆) 1 −
.
2αm
2
m
So far, by choosing α = 1 and α = 12 we have derived the two performance ratios for the WSEPT rule labeled by [Cor. 1] and [Cor. 2] in Figure 1.
These are better than those following from Theorem 3. Besides, the proofs of
Schwiegelshohn [20] and of Avidor et al. [1] of the underlying bounds for WSPT
are quite similar. Both consist of a sequence of steps that reduce the set of
instances to be examined. In every such reduction step it is shown that for any
instance I of the currently considered set there is an instance I 0 in a smaller set
for which the approximation ratio of WSPT is not better. This can be generalized to arbitrary α ∈ [ 12 , 1]. The resulting performance ratios for WSPT lead by
means of Theorem 2 to a family of different performance ratios for the WSEPT
rule. Note that the performance ratio of WSEPT following from the result of
Avidor et al. for α = 21 has better behavior for large ∆, while the performance
ratio following from Kawaguchi and Kyan’s result for α = 1 is better for small
∆. This behavior generalizes to α ∈ [ 21 , 1]: the smaller the underlying α, the
better the ratio for large ∆ but the worse the ratio for small ∆. Finally, we take
for every ∆ > 0 the minimum of all the derived bounds.
Theorem 4 (Generalized Kawaguchi-Kyan bound). For every α ∈ [ 12 , 1] the
WSPT rule has performance ratio
1+
for P ||
P
1
√
2α + 8α
wj Cj (α), and this bound is tight.
Combining this bound with Theorem 2 yields for every α ∈ [ 12 , 1] the perp
√
formance ratio 1 + 21 max{1/(α + 2α), (1 + ∆)/(1 + 2/α)} of WSEPT for
13
P
P |pj ∼ stoch|E[ wj Cj ]. This is minimized at α := 1/ min{2, 1 + ∆}, yielding
the following performance ratio (see Figure 1).
P
Corollary 3. For P |pj ∼ stoch|E[ wj Cj ] the WSEPT rule has performance
ratio
1
1
p
· (1 + ∆).
1+ ·
2 1 + min{2, 2(1 + ∆)}
Proof of Theorem 4. The proof of Theorem 4 is analogous to the proof of
Schwiegelshohn [20], consisting of a sequence of lemmas that reduce the set of
instances to consider until a worst-case instance is determined. From now on,
let α ∈ [ 12 , 1]. Assuming that p1 ≥ · · · ≥ pn , let
n
X
1
` := max j ∈ {1, . . . , m} pj ≥
pj 0 .
m−j+1 0
j =j
Then we call the ` jobs with largest processing times long jobs and denote the
set of long jobs by L.
P
Lemma 2. For every instance
I of P || pj Cj (α) and every ε > 0 there is an
P
instance I 0 = I 0 (ε) of P || pj Cj (α) with the same number of machines such
that λα (I 0 ) ≥ λα (I) and
WSPT 0
(i) Mmin
(I ) = 1,
WSPT 0
WSPT 0
(ii) every job j with SjWSPT (I 0 ) < Mmin
(I ) fulfills CjWSPT (I 0 ) ≤ Mmin
(I )
0
and pj < ε,
(iii) in the optimal schedule for I 0 every machine either is used only by a single
∗
long job or has load Mmin
(I 0 ).
Like in Schwiegelshohn’s paper, the lemma is proven by scaling the instance
WSPT
and splitting all jobs with SjWSPT < Mmin
until they satisfy the conditions.
Proof. As in the proof by Schwiegelshohn [20], the reduction relies on the obWSPT
servation that a job j with SjWSPT < Mmin
can be replaced by two jobs (j, 1)
0
0
WSPT
and (j, 2) with p(j,1) + p(j,2) = pj and SjWSPT + p0(j,1) ≤ Mmin
in such a way
that the WSPT rule schedules the new jobs one after the other on the same machine and during the same time slot as the old job and that the approximation
ratio does not decrease. (Thereto the position of the second job in the input
list must be chosen appropriately.) The reason that the approximation ratio
can only increase is that the transformation reduces the objective value of the
WSPT schedule by exactly
αp2j − (αp2(j,1) + p0(j,2) (p0(j,1) + αp(j,2) )) = (2α − 1)p0(j,1) p0(j,2) =: δ ≥ 0,
and the objective value of the optimal schedule is reduced by at least this
amount. (Replacing j in the optimal schedule by the new jobs gives a feasible schedule). So
δ
WSPTα − δ
= λα +
(λα − 1) ≥ λα
OPTα − δ
OPTα − δ
14
t=1
Figure 3: Optimal schedule and WSPT schedule for instance satisfying the
conditions of Lemma 2
because δ < OPTα and λα ≥ 1.
WSPT
until the conSuch job splitting is applied to jobs with SjWSPT < Mmin
WSPT
ditions of the lemma are satisfied: First, every job jutting out over Mmin
in
WSPT
the WSPT schedule is split so that the first part ends at Mmin , then the jobs
are split in such a way that they can be evenly distributed onto the machines
without a long job in the optimal schedule, and finally, they are split until they
WSPT
are smaller than Mmin
(I) · ε. After the splitting, the whole instance is scaled
WSPT
by Mmin (I) in order to fulfill the first condition.
Note that the restriction to α ≥ 12 is needed for this lemma because for
smaller α splitting jobs increases the objective value and can thence reduce the
performance ratio.
From now on, we focus on instances I that fulfill the requirements of Lemma 2
WSPT
for someP0 < ε < Mmin
(I). For a subset J ⊆ N of jobs we write
p(J ) := j∈J pj . We call the jobs in the set
WSPT
S := {j ∈ {1, . . . , n} | SjWSPT (I) < Mmin
}
short jobs. This set is disjoint fromP
L because all jobs j ∈ S have processing
n
1
WSPT
0
> ε for all jobs j ∈ L.
time pj < ε, while pj ≥ p` ≥ m−`+1
j 0 =` pj ≥ Mmin
Finally, we call the jobs in
M := N (I) \ (S ∪ L)
medium jobs. In the optimal schedule for an instance I of the type of Lemma 2,
every machine i that does not process only a single long job has load
∗
Mi∗ (I) = Mmin
(I) = p(M ∪ S)/(m − |L|). While some of them may process a
medium job together with some short jobs, the rest only process short jobs (see
Figure 3).
P
Lemma 3. For every instance I of P || pj Cj (α) satisfying the conditions of
Lemma 2 there is an instance I 0 with λα (I 0 ) ≥ λα (I) that still satisfies the
conditions of Lemma 2 and has the additional property that all non-short jobs
require equal processing time.
Proof. The proof is an adapted version of the proof of Corollary 5 in the paper
of Schwiegelshohn [20]. We assume that processing times are rational. Let
15
P
A := j∈L∪M p2j and t := p(L ∪ M)2 /A ∈ Q. Let q ∈ N be the denominator
of t (as reduced fraction). The instance I 0 is defined as follows: The number
m0 of machines is set to qm. Furthermore, there are qt ∈ N non-short jobs of
size p0 := A/p(L ∪ M). Finally, for every short job of I there are q short jobs
in I 0 . We assume that these can be distributed evenly onto all qm machines
and additionally, together with the medium jobs, they can be distributed in a
balanced manner onto the machines that do not process a long job. (If this
is not the case, we split the short jobs appropriately beforehand.) Then the
conditions of Lemma 2 remain valid and all non-short jobs have the same size.
Let S 0 be the set of short jobs in I 0 and let N S 0 be the set of non-short jobs in
I 0 . First notice that
X
X
p0j =
qt · p0
= q
pj and
0
j∈N S
X
(p0j )2
j∈L∪M
= qt · (p0 )2
=
X
q
j∈N S 0
p2j .
j∈L∪M
Therefore, the objective value of the WSPT schedule is only scaled by q in the
course of this transformation:
X
X
WSPTα (I 0 ) =
p0j CjWSPT (α)(I 0 ) +
p0j CjWSPT (α)(I 0 )
j∈N S 0
j∈S 0
=
X
p0j CjWSPT (α)(I 0 ) +
X
p0j + α
j∈N S 0
j∈S 0
X
(p0j )2
j∈N S 0
= q
X
pj CjWSPT (α)(I) +
X
j∈S
j∈L∪M
X
X
= q
pj + α
X
p2j
j∈L∪M
pj CjWSPT (α)(I) +
j∈S
pj CjWSPT (α)(I)
j∈L∪M
= q · WSPTα (I).
It remains to be shown that the objective value of an optimal schedule for
I 0 is at most q times the optimal objective value for I. This follows from the
following calculation, where the inequality marked with (∗) is proven below.
X
X
1 X
(p0j )2
OPTα (I 0 ) =
p0j Cj∗ (α)(I 0 ) =
p0j Cj∗ ( 21 )(I 0 ) + α −
2
0
0
0
j∈N (I )
j∈N (I )
j∈N (I )
0
m
X
1 X 0 2
(5) 1 X
=
Mi∗ (I 0 )2 + α −
(pj ) +
(p0j )2
2 i=1
2
0
0
j∈S
j∈N S
m
X
X
(∗) 1 X
1
q
M ∗ (I)2 + α −
p2j + q
p2j
≤ q
2 i=1 i
2
j∈S
j∈L∪M
X
X
1
(5)
= q
pj Cj∗ ( 21 )(I) + α −
p2j
2
j∈N (I)
j∈N (I)
X
∗
=q
pj Cj (α)(I) = q · OPTα (I).
j∈N (I)
16
In order to prove the inequality (∗), we have to show that
0
m
X
Mi∗ (I 0 )2
≤q
i=1
m
X
Mi∗ (I)2 .
i=1
If p0 ≤ 1 + qtp0 /m0 = 1 + tp0 /m, i.e., if the new jobs are medium jobs, then
all machines have load 1 + tp0 /m in the optimal schedule for I 0 . Therefore,
0
m
X
Mi∗ (I 0 )2
i=1
2
p(S) + p(L ∪ M)
= m (1 + tp /m) = qm
m
Pm
2
m
∗
X
i=1 Mi (I)
= qm
≤q
Mi∗ (I)2 ,
m
i=1
0
0
2
where the last inequality follows from the convexity of the square function.
In the other case, when the non-short jobs are long, there are qt machines
with load p0 and q(m − t) machines with load m/(m − t) in the optimal schedule
∗
for I 0 . We define k := p(M)/Mmin
(I) and
à :=
X
∗
p2j + k · Mmin
(I)2 ,
t̃ :=
j∈L
p(L ∪ M)2
,
Ã
p̃ :=
Ã
.
p(L ∪ M)
The convexity of the square function implies that
(|L| + k) ·
j∈L
2
∗
pj + k · Mmin
(I) /(|L| + k)
j∈L
∗ (I)2
p2j + k · Mmin
P
t̃ = (|L| + k) ·
P
≤ |L| + k (8)
and
∗
k · Mmin
(I) = k ·
p(M)
k
2
=
p(M)2
p(M)2
≥
= |M|
k
|M|
p(M)
|M|
2
≥
X
p2j ,
j∈M
whence à ≥ A, t̃ ≤ t, and p̃ ≥ p0 . Therefore,
m
m
≤
< p0 ≤ p̃,
m−t
m − t̃
where the inequality in the middle holds because by assumption p0 > 1 + tp0 /m.
m
m
= p(N (I)) = t̃ · p̃ + (m − t̃) · m−
Since additionally, t · p0 + (m − t) · m−t
, the
t̃
convexity of the square function implies that
0 2
t · (p ) + (m − t) ·
m
m−t
2
2
≤ t̃ · p̃ + (m − t̃) ·
m
m − t̃
2
.
(9)
For an illustration of this formula, see Schwiegelshohn [20]. The following cal-
17
culation concludes the proof.
0
m
X
i=1
Mi∗ (I 0 )2
2
2 !
m
m
0 2
= q t(p ) + (m − t)
= qt(p ) + q(m − t)
m−t
m−t
!
2
(9)
m
≤ q t̃p̃2 + (m − t̃)
m − t̃
(8)
2
m
m2
2
2
= q t̃p̃ +
≤ q t̃p̃ +
m − |L| − k
m − t̃
2 !
m
= q à + (m − |L| − k)
m − |L| − k
X
∗
∗
= q
p2j + k · Mmin
(I)2 + (m − |L| − k) · Mmin
(I)2
0 2
j∈L
=q·
m
X
Mi∗ (I)2 .
i=1
Since by Lemma 2 reducing ε can only increase the approximation ratio,
the worst-case approximation ratio is approached in the limit ε → 0, which
we will subsequently further investigate. In the limit the sum of the squared
processing times of the short jobs is negligible, wherefore the limits for ε → 0
of the objective values of the WSPT schedule and the optimal schedule for an
instance I(ε) of the type of Lemma 3 only depend on two variables: the ratio s
between the numbers of non-short jobs and machines and the duration x of the
non-short jobs. The limit of the objective value of the WSPT schedule is given
by
m
lim WSPTα (I(ε)) =
+ smx(1 + αx).
ε→0
2
For the optimal schedule the formula depends on whether the non-short jobs
are medium or long. In the first case it is given by
m
1
lim OPTα (I(ε)) = (sx + 1)2 + α −
smx2 .
ε→0
2
2
and in the second case by
lim OPTα (I(ε)) = αsmx2 +
ε→0
m
.
2(1 − s)
So we have to determine the maximum of the function
λM (s, x) :=
m
2
m
2 (sx
+ smx(1 + αx)
2sx(αx + 1) + 1
= 2 2
s x + sx((2α − 1)x + 2) + 1
+ 1)2 + (α − 21 )smx2
on {(s, x) | 0 ≤ s < 1, 0 ≤ x ≤ 1/(1 − s)} and the maximum of
λL (s, x) :=
m
2
+ smx(1 + αx)
(1 − s)(2sx(αx + 1) + 1)
=
m
αsmx2 + 2(1−s)
2αs(1 − s)x2 + 1
on the region {(s, x) | 0 ≤ s < 1, 1/(1 − s) ≤ x}.
18
∂
λM is positive on the feasible region, so for every
The partial derivative ∂x
1
fixed s the maximum of λM (s, ·) is attained at x = 1−s
, corresponding to the case
that the non-short jobs are long. This case is also captured by the function λL .
For x → ∞ the function λL converges to one. Hence, for every s the maximum of λL (s, ·) must be attained at a finite point x. The partial derivative
∂
∂x λL has only one positive root, namely
p
xs := (αs + (2(1 − s) + αs)αs)/(2αs(1 − s)) > 1/(1 − s).
By plugging this in, we obtain
p
λL (s, xs ) = 1 + 21 ( (2(1 − s) + αs)αs/α − s).
The only root of the derivative of the function s 7→ λL (s, xs ) that is less than 1
is
√
s := 1/(2 + 2α).
Plugging this in yields the worst-case performance ratio
1+
1
√ .
2α + 8α
Like the proofs of Kawaguchi and Kyan [14] and of Avidor et al. [1], this proof
shows how the worst-case
of short jobs of total
p
√ instances look like: They consist
length m and 1/(2 + 2α)m long jobs of length 1 + 2/α. For α ∈ {1/2, 1} we
recover the worst case instances of Avidor et al. and of Kawaguchi and Kyan.
4
Performance ratio of the WSPT rule for a fixed
number of machines
P
In this section we analyze the WSPT rule for the problem P || wj Cj with a
fixed number m of machines. The problem instances
of Kawaguchi and Kyan [14]
√
whose approximation ratios converge to (1 + 2)/2 consist of a set of infinitesimally
√ short jobs with total
√ processing time
√ m, and a set of k jobs of length
1+ 2, where k/m → 1− 2/2. Since 1− 2/2 is irrational, the worst case ratio
can only be approached if the number of machines goes to infinity. Rounding
√
these instances for a fixed m by choosing k as the nearest integer to 1 − 22 m
√
(in the following denoted by 1 − 22 mP) yields at least a lower bound on
the worst-case approximation ratio for P || wj Cj . It is, however, possible that
for a particular m there is an instance for which the WSPT rule has a larger
approximation ratio than for the rounded instance of Kawaguchi and Kyan. As
we will see, the worst-case instances for any fixed m actually look almost as the
rounded Kawaguchi-Kyan instances with the only difference that the length of
the long jobs depends as a function on m.
P
Theorem 5. For P || wj Cj the WSPT rule has performance ratio
$
p
√ '
2
1 (2m − km )km − km
,
where km :=
1−
m .
1+
2
m
2
Moreover, this bound is tight for every fixed m ∈ N.
19
In the remainder we prove this theorem. Lemmas 1 and 2 hold in particular
for the weighted sum of completion times. Since the described transformations
do not change the number of machines, also for a fixed number m of machines
the worst case occurs in an instance of the form described in Lemma 2. However,
we cannot apply Lemma 3 when m is fixed because the transformation in this
lemma possibly changes the number of machines. As this is not allowed in our
setting, we have to find different reductions. We first reduce to instances with
at most one medium job and then reduce further to instances where all long
jobs have equal length. Similar reductions are also carried out by Kalaitzis,
Svensson, and Tarnawski [13].
P
Lemma 4. For every instance I of P || pj Cj satisfying the conditions of
Lemma 2 there is an instance I 0 with the same number of machines and
λ(I 0 ) ≥ λ(I) that still satisfies the conditions of Lemma 2 and has the additional property that there is at most one medium job.
∗
Proof. We replace the set M of medium jobs in I by bp(M)/Mmin
(I)c jobs of
∗
0 :=
∗
∗
size Mmin (I) and one job of size p
p(M) − Mmin (I)bp(M)/Mmin
(I)c (see
Figure 4). In the thus defined instance I 0 the job of size p0 is the only potentially
medium job. Moreover, the properties of Lemma 2 remain true. Let M̃ be the
set of new jobs in the instance I 0 . Then we have
X
X
X
X
p0j −
pj =
p0j −
pj = 0 and
j∈N (I 0 )
X
(p0j )2
j∈N (I 0 )
j∈N (I)
−
X
p2j
j∈M
j∈M̃
X
=
j∈N (I)
(p0j )2
−
X
p2j =: δ > 0
j∈M
j∈M̃
by the convexity of the square function. Since in the WSPT schedules for
both problem instances all jobs in M resp. M̃ have the same starting time
WSPT :=
WSPT
WSPT 0
Mmin
Mmin
(I) = Mmin
(I ), and all other jobs remain unchanged, we
have
X
X
X
p0j SjWSPT (I 0 ) =
p0j SjWSPT (I 0 ) +
p0j SjWSPT (I 0 )
j∈N (I 0 )
j∈N (I 0 )\M̃
X
=
j∈M̃
WSPT
p0j SjWSPT (I 0 ) + Mmin
j∈N (I 0 )\M̃
X
=
X
+
WSPT
Mmin
=
X
pj
j∈M
pj SjWSPT (I) +
X
pj SjWSPT (I)
j∈M
j∈N (I)\M
X
p0j
j∈M̃
pj SjWSPT (I)
j∈N (I)\M
=
X
pj SjWSPT (I).
j∈N (I)
Therefore, we get
WSPT(I 0 ) =
X
j∈N (I 0 )
=
X
j∈N (I)
X
p0j CjWSPT (I 0 ) =
pj SjWSPT (I)+
p0j SjWSPT (I 0 ) +
j∈N (I 0 )
X
j∈N (I)
p2j +δ =
X
j∈N (I)
20
X
(p0j )2
j∈N (I 0 )
pj CjWSPT (I)+δ = WSPT(I)+δ.
t=1
Figure 4: Optimal schedule and WSPT schedule for instance after the transformation of Lemma 4
After the transformation the schedule that puts every long job on a machine of
its own and balances the loads of the remaining machines is still optimal. In
this schedule all machines have the same load as in the optimal schedule for I.
Therefore, we have
m
X
X
(5) 1
(p0j )2
OPT(I 0 ) =
p0j Cj∗ (I 0 ) = (Mi∗ (I 0 ))2 +
2
i=1
j∈N (I 0 )
j∈N (I 0 )
m
X
1 X
1
(5) X
= (Mi∗ (I))2 +
p2j + δ =
pj Cj∗ (I)+ δ = OPT(I)+δ/2.
2 i=1
2
X
j∈N (I)
j∈N (I)
Together this yields
λ(I 0 ) =
WSPT(I 0 )
WSPT(I) + δ
δ
λ(I)
=
=
λ(I)
+
1
−
.
OPT(I 0 )
OPT(I) + δ/2
OPT(I) + δ/2
2
As λ(I) < 2, the second summand is non-negative, implying that λ(I 0 ) ≥ λ(I).
For the instance shown in Figure 3 the optimal and the WSPT schedule of
the reduced instance are shown in Figure 4.
P
Lemma 5. For every instance I of P || pj Cj satisfying the conditions of
Lemma 4 there is an instance I 0 with the same number of machines and
λ(I 0 ) ≥ λ(I) that still satisfies the conditions of Lemma 4 where additionally
all long jobs have equal processing time.
Proof. We replace the set L of long jobs by |L| jobs with processing time p0j :=
p(L)/|L|. Then these jobs are still long, and no other jobs become long during
this transformation, so the set L0 of long jobs in I 0 consists exactly of the newly
defined jobs. Besides, the sets of medium and short jobs are not affected by this
transformation. The WSPT schedule for I 0 schedules the jobs in the same way
as for I. Similarly, it is still optimal to schedule each long job on a machine of
its own, and totally balance the loads of the remaining machines. Therefore, I 0
still satisfies the conditions of Lemma 4. The optimal and the WSPT schedule
after the transformation are depicted in Figure 5.
21
Note that
X
j∈N (I 0 )
X
X
p0j −
pj =
X
j∈N (I 0 )
p0j −
j∈L0
j∈N (I)
(p0j )2 −
X
X
p2j =
pj = 0 and
j∈L
(p0j )2 −
j∈L0
j∈N (I)
X
X
p2j =: −δ < 0,
j∈L
where the second inequality follows again from the convexity of the square function.
Because in in the WSPT schedules for I and I 0 all long jobs start at time
WSPT :=
WSPT
WSPT 0
Mmin
Mmin
(I) = Mmin
(I ) and all other jobs are unmodified, the
same calculation as in the proof of Lemma 4 shows that
X
X
p0j SjWSPT (I 0 ) =
pj SjWSPT (I),
j∈N (I 0 )
j∈N (I)
and thus,
X
WSPT(I 0 ) =
j∈N (I 0 )
=
X
X
p0j CjWSPT (I 0 ) =
j∈N (I 0 )
pj SjWSPT (I)+
j∈N (I)
X
p2j −δ
X
=
j∈N (I)
X
p0j SjWSPT (I 0 ) +
(p0j )2
j∈N (I 0 )
pj CjWSPT (I)−δ
= WSPT(I)−δ.
j∈N (I)
In the optimal schedule for I and I 0 all long jobs start at time 0, so we also have
that
X
X
p0j Sj∗ (I 0 ) =
pj Sj∗ (I),
j∈N (I 0 )
j∈N (I)
and hence,
X
OPT(I 0 ) =
j∈N (I 0 )
=
X
X
p0j Cj∗ (I 0 ) =
pj Sj∗ (I) +
j∈N (I)
p0j Sj∗ (I 0 ) +
j∈N (I 0 )
X
p2j − δ =
j∈N (I)
X
(p0j )2
j∈N (I 0 )
X
pj Cj∗ (I) − δ = OPT(I) − δ.
j∈N (I)
Together, this results in the inequality
λ(I 0 ) =
WSPT(I) − δ
δ
WSPT(I 0 )
=
= λ(I) +
(λ(I) − 1) ≥ λ(I)
0
OPT(I )
OPT(I) − δ
OPT(I) − δ
because δ < OPT(I).
The reduction used in the proof is illustrated in Figure 5.
As in Section 2 we will analyze the limit for ε → 0. The limits of the objective
values of the WSPT schedule and the optimal schedule for an instance I(ε) of
the type of Lemma 5 depend only on three variables: two real variables, viz.
the length x of the long jobs and the length y of the medium job (y = 0 if no
medium job exists), and one integer variable, namely the number k of long jobs.
They are given by
m
+ kx(1 + x) + y(1 + y),
lim WSPT(I(ε)) =
ε→0
2
(m + y)2
y2
lim OPT(I(ε)) = k · x2 +
+ .
ε→0
2(m − k)
2
22
t=1
Figure 5: Optimal schedule and WSPT schedule for instance after the transformation of Lemma 5
In Figure 6 these formulas are illustrated via two-dimensional Gantt charts
(see [4, 7]) for the three different types of single-machine schedules used by the
WSPT schedule and the optimal schedule, respectively. In order to describe a
valid scheduling instance of the prescribed type, the values x, y, and k must lie
in the domains
i
( h
m
if k < m − 1,
∈ 0, m−k−1
y+m
x∈
,∞ .
k ∈ {0, . . . , m − 1},
y
m−k
=0
if k = m − 1,
Lemma 6. The maximum of the ratio
λm (x, y, k) :=
m
2
+ kx(1 + x) + y(1 + y)
k · x2 +
(m+y)2
2(m−k)
+
y2
2
=
(m − k)(2kx2 + 2kx + 2y 2 + 2y + m)
(m − k)(2kx2 + y 2 ) + (y + m)2
p
on the feasible domains is 1 + 21 ( (2m − km )km − km )/m, and it is attained at
1√
m
km := 1 −
2 m ,
ym := 0,
.
xm := p
2
(2m − km )km − km
Proof. This set of possible combinations is closed. Since for x → ∞ this ratio
converges to 1, the maximum is attained at some point (x∗ , y ∗ , k ∗ ).
Assume now that y ∗ 6= 0. Then k ∗ < m−1, and the function y 7→ λ(x∗ , y, k ∗ )
is thus quasi-convex (since the derivative has at most one root in the feasible
region and is non-positive at y = 0). Therefore, the maximum is attained
at the boundary, i.e., y ∗ = m/(m − k − 1). But then the job is in fact a
long job, and by Lemma 5 it has the same length as the other long jobs, i.e.,
x = y = m/(m − k − 1). By the definitions of k, x, and y, however, this instance
is properly described by the parameter values k 0 = k + 1, y 0 = 0, and x0 = x.
So we have shown that y ∗ = 0, and only the values x and k maximizing
λ0m (x, k) :=
(m − k)(2kx2 + 2kx + m)
(m − k)2kx2 + m2
remain to be determined. For every fixed k, the maximum is attained at
xk,m := p
m
(2m − k)k − k
23
,
1
2
x2
k×
k×
x
x2
y2
1×
1
1
2
m+y 2
m−k
− y2
1×
1
2
y2
y
1
(m − k − 1) ×
1
2
m+y
m−k
2
(m − k − 1) ×
1
2
1
Figure 6: Limit two-dimensional Gantt charts (see [4, 7]) for ε → 0 of the optimal
schedule and the WSPT schedule for instance I(ε) satisfying the conditions of
Lemma 5, illustrating the different parts of the objective values
24
√
λ00m (k)
(2m−k)k−k
Figure 7: Graph of the function (m, k) 7→
= 1+
2m
the worst-case performance ratio with m machines and k long jobs
showing
as can be seen by calculating the roots of the derivative. Plugging this in, we
obtain the univariate function
p
(2m − k)k − k
00
,
λm (k) := 1 +
2m
whose maximum over {0, . . . , m−1} is to be determined. Notice that λ00m (0) = 1
is not maximal. As a function on the interval [1, m − 1], the function λ00m is
concave (since the second derivative is negative) and continuous. Furthermore,
√
2m2 − 1
um := m −
2
1
1
00
00
satisfies
the
equation
λ
(u
−
)
=
λ
u
+
. Therefore,
for every
m
m
m
m
2
2
1
1
1
1
ξ ∈ um − 2 , um + 2 and every ζ ∈ [1, m − 1] \ um − 2 , um + 2 it holds that
00
00
In particular, the unique integer k in the interval
λ
m (ξ) 1≥ λm (ζ).
um − 2 , um + 21 satisfies that λ00m (k) ≥ λ00m (k 0 ) for all k 0 ∈ {1, . . . , m − 1}.
We prove that
$
√ '
2
km :=
m
1−
2
√
always lies in this interval. Firstly, km ≤
1 − m22 m + 12 < um + 12 . Secondly,
l
√
√
√
notice that 2m2 is irrational such that
2m2 ≤ 2m2 − 1+1 as the interval
√
i
√
2m2 − 1, 2m2 does not contain an integer; therefore,
km
m
l√
$√
%
√
2m2
2−1+1
1
2m
2m2 − 1 + 1
= m−
≥ m−
≥ m−
= um − .
2
2
2
2
Figure 7 shows the function (m, k) 7→ λ00m (k) with marked (thick) lines for
integral values of k. One can also see the concave dependence on k (thin lines).
This concludes the proof of Theorem 5. In Figure 2 the graph of the function m 7→ λm (xm , 0, km ), whose values at integral m are exactly the worst-case
25
performance ratio
[17]
3
2
1
1 (1
+ 4
)
+∆
1+
5
4
1+
1
21
1
√ − 1)(
1( 3
+∆
)
[Cor. 4]
[Cor. 3]
4
1√
}
2+2∆
{2 ,
n
i
m
+
(1 + ∆
)
[14]
1
∆
0
1
1
2
3
2
Figure 8: Bound on WSEPT’s performance ratio for two machines
approximation ratios for instances with m machines, is depicted. The jumps
and kinks occur when the number km of long jobs in the worst-case instance
changes. By taking the √
limit for m → ∞, we obtain alternative proof of the
performance ratio 21 (1 + 2) by Kawaguchi and Kyan [14], avoiding the somewhat complicated transformation and case distinction in the proof of Lemma 3
and Schwiegelshohn’s proof [20]. √
For increasing m the tight performance ratio
converges quite quickly to 12 (1 + 2): the difference lies in O(1/m2 ). By plugging in the machine-dependent performance ratio into Theorem 1, we obtain
the following performance ratio for the WSEPT rule.
P
Corollary 4. For instances of the problem P |pj ∼ stoch|E[ wj Cj ] with m
machines the WSEPT rule has performance ratio
p
(2m − km )km − km
1
1+ ·
(1 + ∆).
2
m
This bound is better than the bound of Corollary 3 only if m and ∆ both
are small. Even for two machines, it is outdone for large ∆ (see Figure 8).
5
Open problem
For every fixed value of ∆,
√ one obtains the machine-independent performance
ratio of 1 + 21 (1 + min{2, 2 + 2∆})−1 (1 + ∆) for instances with squared coefficient of variation bounded by ∆. On the other hand, for every fixed m our
performance bound tends to infinity when ∆ goes to infinity, so that it does
not imply a constant performance ratio (independent of ∆) for instances with a
constant number of machines. As far as the authors know, the question if such
a ∆-independent constant performance ratio of WSEPT for a fixed number of
26
machines exists is still open. The examples of Cheung et al. [3] and of Im et
al. [11] only show that no constant performance ratio can be given when ∆ and
m are allowed to go simultaneously to infinity.
References
[1] A. Avidor, Y. Azar, and J. Sgall. Ancient and new algorithms for load
balancing in the `p norm. Algorithmica, 29(3):422–441, Mar. 2001.
[2] C. Chekuri, R. Motwani, B. Natarajan, and C. Stein. Approximation techniques for average completion time scheduling. SIAM Journal on Computing, 31(1):146–166, Jan. 2001.
[3] W. C. Cheung, F. Fischer, J. Matuschke, and N. Megow. A Ω(∆1/2 ) gap
example for the WSEPT policy. Cited as personal communication in Uetz:
MDS Autumn School Approximation Algorithms for Stochastic Optimization, 2014.
[4] W. L. Eastman, S. Even, and I. M. Isaacs. Bounds for the optimal scheduling of n jobs on m processors. Management Science, 11(2):268–279, Nov.
1964.
[5] M. X. Goemans. Improved approximation algorithms for scheduling with
release dates. In M. Saks, editor, Proceedings of the Eighth Annual ACMSIAM Symposium on Discrete Algorithms, pages 591–598. Society for Industrial and Applied Mathematics, 1997.
[6] M. X. Goemans, M. Queyranne, A. S. Schulz, M. Skutella, and Y. Wang.
Single machine scheduling with release dates. SIAM Journal on Discrete
Mathematics, 15(2):165–192, 2002.
[7] M. X. Goemans and D. P. Williamson. Two-dimensional Gantt charts and
a scheduling algorithm of Lawler. SIAM Journal on Discrete Mathematics,
13(3):281–294, May 2000.
[8] R. L. Graham, E. L. Lawler, J. K. Lenstra, and A. H. G. Rinnooy Kan. Optimization and approximation in deterministic sequencing and scheduling:
a survey. In P. L. Hammer, E. L. Johnson, and B. H. Korte, editors, Discrete Optimization II, volume 5 of Annals of Discrete Mathematics, pages
287–326. Elsevier, 1979.
[9] V. Gupta, B. Moseley, M. Uetz, and Q. Xie. Stochastic online scheduling on
unrelated machines. In F. Eisenbrand and J. Koenemann, editors, Integer
Programming and Combinatorial Optimization: 19th International Conference, Proceedings, pages 228–240, Cham, 2017. Springer International
Publishing, arXiv:1703.01634.
[10] L. A. Hall, D. B. Shmoys, and J. Wein. Scheduling to minimize average
completion time: Off-line and on-line algorithms. In E. Tardos, editor, Proceedings of the Seventh Annual ACM-SIAM Symposium on Discrete Algorithms, pages 142–151, Philadelphia, PA, USA, 1996. Society for Industrial
and Applied Mathematics.
27
[11] S. Im, B. Moseley, and K. Pruhs. Stochastic scheduling of heavy-tailed
jobs. In E. W. Mayr and N. Ollinger, editors, 32nd International Symposium on Theoretical Aspects of Computer Science, volume 30 of Leibniz
International Proceedings in Informatics (LIPIcs), pages 474–486. Schloss
Dagstuhl–Leibniz-Zentrum fuer Informatik, 2015.
[12] C. Jagtenberg, U. Schwiegelshohn, and M. Uetz. Analysis of Smith’s rule in
stochastic machine scheduling. Operations Research Letters, 41(6):570–575,
Nov. 2013.
[13] C. Kalaitzis, O. Svensson, and J. Tarnawski. Unrelated machine scheduling
of jobs with uniform Smith ratios. In P. N. Klein, editor, Proceedings of
the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms,
pages 2654–2669, Philadelphia, PA, USA, 2017. Society for Industrial and
Applied Mathematics, arXiv:1607.07631.
[14] T. Kawaguchi and S. Kyan. Worst case bound of an LRF schedule for
the mean weighted flow-time problem. SIAM Journal on Computing,
15(4):1119–1129, Nov. 1986.
[15] N. Megow, M. Uetz, and T. Vredeveld. Models and algorithms for stochastic online scheduling. Mathematics of Operations Research, 31(3):513–525,
Aug. 2006.
[16] R. H. Möhring, F. J. Radermacher, and G. Weiss. Stochastic scheduling problems I — General strategies. Zeitschrift für Operations Research,
28(7):193–260, Nov. 1984.
[17] R. H. Möhring, A. S. Schulz, and M. Uetz. Approximation in stochastic
scheduling: The power of LP-based priority policies. Journal of the ACM,
46(6):924–942, Nov. 1999.
[18] M. H. Rothkopf. Scheduling with random service times. Management
Science, 12(9):707–713, May 1966.
[19] A. S. Schulz. Stochastic online scheduling revisited. In B. Yang, D.-Z. Du,
and C. A. Wang, editors, Combinatorial Optimization and Applications:
Second International Conference. Proceedings, pages 448–457, Berlin, Heidelberg, 2008. Springer.
[20] U. Schwiegelshohn. An alternative proof of the Kawaguchi–Kyan bound
for the largest-ratio-first rule. Operations Research Letters, 39(4):255–259,
July 2011.
[21] M. Skutella. A 2.542-approximation for precedence constrained single
machine scheduling with release dates and total weighted completion
time objective. Operations Research Letters, 44(5):676–679, Sept. 2016,
arXiv:1603.04690.
[22] M. Skutella, M. Sviridenko, and M. Uetz. Unrelated machine scheduling
with stochastic processing times. Mathematics of Operations Research,
41(3):851–864, Aug. 2016.
28
[23] M. Skutella and G. J. Woeginger. A PTAS for minimizing the total weighted
completion time on identical parallel machines. Mathematics of Operations
Research, 25(1):63–75, Feb. 2000.
[24] W. E. Smith. Various optimizers for single-stage production. Naval Research Logistics Quarterly, 3(1-2):59–66, Mar. 1956.
[25] R. R. Weber, P. Varaiya, and J. Walrand. Scheduling jobs with stochastically ordered processing times on parallel machines to minimize expected
flowtime. Journal of Applied Probability, 23(3):841–847, Sept. 1986.
[26] G. Weiss. Approximation results in parallel machnies stochastic scheduling.
Annals of Operations Research, 26(1):195–242, Dec. 1990.
[27] G. Weiss. Turnpike optimality of Smith’s rule in parallel machines stochastic scheduling. Mathematics of Operations Research, 17(2):255–270, May
1992.
29
| 8cs.DS
|
Generic Description of Well-Scoped, Well-Typed Syntaxes
Gergő Érdi
[email protected]
We adapt the technique of type-generic programming via descriptions pointing into a universe to
the domain of typed languages with binders and variables, implementing a notion of syntax-generic
programming in a dependently typed programming language.
We present an Agda library implementation of type-preserving renaming and substitution (including proofs about their behaviour) “once and for all” over all applicable languages using our technique.
1 Introduction
Type theory-based tools like Agda[22] or Coq[6] are frequently used to study formal languages, which
are often typed, and involve structures that bind and that reference names. All these aspects: syntax,
scoping rules, and typing rules, need to be modeled, so that we can prove properties and implement
transformations in the metalanguage of our choice.
The usual approach to this modeling work is to create bespoke datatypes in the metalanguage for
each object language, and write, from scratch, all proofs and transformations, as needed, in terms of
those datatypes. While there are standard techniques for creating the right representation (see, among
others, [2, 19, 20, 17, 7]), these techniques are still ad-hoc in that they are not formalized in the metatool: they happen “outside”, and so all proofs and programs need to be written again and again as the
language changes.
This paper describes a generic approach to modeling languages, by representing the definition of any
given language as scrutable data instead of an ad-hoc datatype. This allows us to derive the datatype of
terms (in varying levels of static guarantees), the implementation of renaming and substitution, and the
proofs of these transformations behaving as expected, once and for all from this single description.
The full Agda implementation of the library described here is available at http://gergo.erdi.
hu/projects/generic-syntax/.
2 Motivation
Suppose we want to use a type theory-based tool to study a typed language; in our running example,
this language is going to be the simply typed lambda calculus[13]. The only types in this language are
function types and a single base type:
𝖽𝖺𝗍𝖺 𝖳𝗒 ∶ 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
∙
∶ 𝖳𝗒
_▹_ ∶ 𝖳𝗒 → 𝖳𝗒 → 𝖳𝗒
The first goal is to give a representation of terms of our object language; in other words, a datatype
whose values correspond to terms.
There is a spectrum of static semanics we could capture in this representation:
DRAFT March 30, 2018
This work is licensed under the Creative Commons
Attribution-Share Alike License.
2
Generic description of well-scoped, well-typed syntaxes
• The datatype could describe only the syntactical form of our language (using some type to represent
variable names), with each constructor corresponding to an abstract grammar rule:
𝖽𝖺𝗍𝖺 𝖥𝗈𝗋𝗆 (𝑛𝑎𝑚𝑒 ∶ 𝖲𝖾𝗍) ∶ 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗏𝖺𝗋 ∶ 𝑛𝑎𝑚𝑒
→ 𝖥𝗈𝗋𝗆 𝑛𝑎𝑚𝑒
𝗅𝖺𝗆 ∶ 𝑛𝑎𝑚𝑒 → 𝖥𝗈𝗋𝗆 𝑛𝑎𝑚𝑒
→ 𝖥𝗈𝗋𝗆 𝑛𝑎𝑚𝑒
_·_ ∶ 𝖥𝗈𝗋𝗆 𝑛𝑎𝑚𝑒 → 𝖥𝗈𝗋𝗆 𝑛𝑎𝑚𝑒 → 𝖥𝗈𝗋𝗆 𝑛𝑎𝑚𝑒
• One step up, to represent well-scoped expressions, is to use de Bruijn indices for variables[8], and
track the number of variables in scope at the type level[4]:
𝖽𝖺𝗍𝖺 𝖤𝗑𝗉𝗋 (𝑉 ∶ ℕ) ∶ 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗏𝖺𝗋 ∶ 𝖥𝗂𝗇 𝑉
→ 𝖤𝗑𝗉𝗋 𝑉
𝗅𝖺𝗆 ∶ 𝖤𝗑𝗉𝗋 (𝟣 + 𝑉 )
→ 𝖤𝗑𝗉𝗋 𝑉
_·_ ∶ 𝖤𝗑𝗉𝗋 𝑉 → 𝖤𝗑𝗉𝗋 𝑉 → 𝖤𝗑𝗉𝗋 𝑉
This allows us, for example, to talk about closed expressions:
𝖢𝗅𝗈𝗌𝖾𝖽𝖤𝗑𝗉𝗋 ∶ 𝖲𝖾𝗍
𝖢𝗅𝗈𝗌𝖾𝖽𝖤𝗑𝗉𝗋 = 𝖤𝗑𝗉𝗋 𝟢
𝖼𝗅𝗈𝗌𝖾 ∶ ∀ {𝑉 } → 𝖤𝗑𝗉𝗋 𝑉 → 𝖢𝗅𝗈𝗌𝖾𝖽𝖤𝗑𝗉𝗋
𝖼𝗅𝗈𝗌𝖾 {𝗓𝖾𝗋𝗈} 𝑒 = 𝑒
𝖼𝗅𝗈𝗌𝖾 {𝗌𝗎𝖼 𝑛} 𝑒 = 𝖼𝗅𝗈𝗌𝖾 (𝗅𝖺𝗆 𝑒)
• Finally, we can create a representation for well-typed terms, either by using the standard technique
of typed pointers (into a context of types of variables in scope) to represent variables, and a datatype
for terms indexed by the object-level type of said term[2] (an intrinsically typed representation):
𝖽𝖺𝗍𝖺 𝖢𝗍𝗑 ∶ 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
∅ ∶ 𝖢𝗍𝗑
_, _ ∶ 𝖢𝗍𝗑 → 𝑇 𝑦 → 𝖢𝗍𝗑
𝖽𝖺𝗍𝖺 𝖵𝖺𝗋 ∶ 𝖢𝗍𝗑 → 𝑇 𝑦 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗏𝗓 ∶ ∀ {𝑡} {𝛤 }
→ 𝖵𝖺𝗋 (𝛤 , 𝑡) 𝑡
𝗏𝗌 ∶ ∀ {𝑢 𝑡} {𝛤 } (𝑣 ∶ 𝖵𝖺𝗋 𝛤 𝑡) → 𝖵𝖺𝗋 (𝛤 , 𝑢) 𝑡
𝗂𝗇𝖿 𝗂𝗑𝗅 𝟤𝟥 _·_
𝖽𝖺𝗍𝖺 𝖳𝗆 (𝛤 ∶ 𝖢𝗍𝗑) ∶ 𝖳𝗒 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗏𝖺𝗋 ∶ ∀ {𝑡} → 𝖵𝖺𝗋 𝛤 𝑡
→ 𝖳𝗆 𝛤 𝑡
𝗅𝖺𝗆 ∶ ∀ {𝑢 𝑡} → 𝖳𝗆 (𝛤 , 𝑡) 𝑢
→ 𝖳𝗆 𝛤 (𝑡 ▹ 𝑢)
_·_ ∶ ∀ {𝑢 𝑡} → 𝖳𝗆 𝛤 (𝑢 ▹ 𝑡) → 𝖳𝗆 𝛤 𝑢 → 𝖳𝗆 𝛤 𝑡
Or, alternatively, using typed Parametric Higher-Order Syntax[12] representation:
𝖽𝖺𝗍𝖺 𝖧𝖮𝖳𝗆 (𝑉 ∶ 𝖳𝗒 → 𝖲𝖾𝗍) ∶ 𝖳𝗒 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗏𝖺𝗋 ∶ ∀ {𝑡} → 𝑉 𝑡
→ 𝖧𝖮𝖳𝗆 𝑉 𝑡
𝗅𝖺𝗆 ∶ ∀ {𝑢 𝑡} → (𝑉 𝑡 → 𝖧𝖮𝖳𝗆 𝑉 𝑢)
→ 𝖧𝖮𝖳𝗆 𝑉 (𝑡 ▹ 𝑢)
_·_ ∶ ∀ {𝑢 𝑡} → 𝖧𝖮𝖳𝗆 𝑉 (𝑢 ▹ 𝑡) → 𝖧𝖮𝖳𝗆 𝑉 𝑢 → 𝖧𝖮𝖳𝗆 𝑉 𝑡
G. Érdi
3
Of course, there are various correspondences between these four representations – for example, we
can erase the types from a 𝖳𝗆 into an 𝖤𝗑𝗉𝗋, or resolve the names in a well-scoped 𝖥𝗈𝗋𝗆 to get an 𝖤𝗑𝗉𝗋 using
de Bruijn indices. We can write these, and more, correspondences as functions implementing mappings
between these types; but there is nothing inherently tying together 𝖥𝗈𝗋𝗆, 𝖤𝗑𝗉𝗋, 𝖳𝗆 and 𝖧𝖮𝖳𝗆 since they
are four independent datatypes as far as our metalanguage Agda is concerned.
Moreover, even once we write these functions, we are usually not done yet by this point. Languages
with variables have functorial (renaming) and monadic (substitution) structure that is needed for the syntactic accounting of most reduction rules. One approach is to represent renamings and simultaneous
substitutions as functions from variables:
.
_→_ ∶ ∀ {𝑎 𝑏 𝑐} {𝐴 ∶ 𝖲𝖾𝗍 𝑎} → (𝐴 → 𝖲𝖾𝗍 𝑏) → (𝐴 → 𝖲𝖾𝗍 𝑐) → 𝖲𝖾𝗍 _
.
𝐹 → 𝐺 = ∀ {𝑡} → 𝐹 𝑡 → 𝐺 𝑡
𝖱𝖾𝗇 ∶ 𝖢𝗍𝗑 → 𝖢𝗍𝗑 → 𝖲𝖾𝗍
.
𝖱𝖾𝗇 𝛤 𝛥 = 𝖵𝖺𝗋 𝛥 → 𝖵𝖺𝗋 𝛤
.
𝗆𝖺𝗉 ∶ ∀ {𝛤 𝛥} → 𝖱𝖾𝗇 𝛤 𝛥 → 𝖳𝗆 𝛥 → 𝖳𝗆 𝛤
𝖲𝗎𝖻 ∶ 𝖢𝗍𝗑 → 𝖢𝗍𝗑 → 𝖲𝖾𝗍
.
𝖲𝗎𝖻 𝛤 𝛥 = 𝖵𝖺𝗋 𝛥 → 𝖳𝗆 𝛤
.
𝖻𝗂𝗇𝖽 ∶ ∀ {𝛤 𝛥} → 𝖲𝗎𝖻 𝛤 𝛥 → 𝖳𝗆 𝛥 → 𝖳𝗆 𝛤
Another approach is to use some data structures to describe renamings and simultaneous substitutions. This has the benefit that equality of the contents of the renaming/substitution automatically lifts to
extensional equality of their action. Here, and in the rest of this paper, we are going to use, and reserve
the more evocative names 𝗋𝖾𝗇 and 𝗌𝗎𝖻 for, this second approach, following [10]: representing renamings
as order-preserving embeddings of contexts, and substitutions as a snoc-list of terms:
𝗂𝗇𝖿 𝗂𝗑 𝟥 _⊇_
𝖽𝖺𝗍𝖺 _⊇_ ∶ 𝖢𝗍𝗑 → 𝖢𝗍𝗑 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝖽𝗈𝗇𝖾 ∶ ∅ ⊇ ∅
𝖽𝗋𝗈𝗉 ∶ ∀ {𝑡 𝛤 𝛥} → 𝛤 ⊇ 𝛥 → 𝛤 , 𝑡 ⊇ 𝛥
𝗄𝖾𝖾𝗉 ∶ ∀ {𝑡 𝛤 𝛥} → 𝛤 ⊇ 𝛥 → 𝛤 , 𝑡 ⊇ 𝛥 , 𝑡
.
𝗋𝖾𝗇−𝗏𝖺𝗋 ∶ ∀ {𝛤 𝛥} → 𝛥 ⊇ 𝛤 → 𝖵𝖺𝗋 𝛤 → 𝖵𝖺𝗋 𝛥
𝖽𝖺𝗍𝖺 𝖤𝗇𝗏 (𝐴 ∶ 𝑇 𝑦 → 𝖲𝖾𝗍) ∶ 𝖢𝗍𝗑 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
∅ ∶ 𝖤𝗇𝗏 𝐴 ∅
_, _ ∶ ∀ {𝑡 𝛥} → (𝜎 ∶ 𝖤𝗇𝗏 𝐴 𝛥) → (𝑒 ∶ 𝐴 𝑡) → 𝖤𝗇𝗏 𝐴 (𝛥 , 𝑡)
.
𝗅𝗈𝗈𝗄𝗎𝗉 ∶ ∀ {𝐴 𝛥} → 𝖤𝗇𝗏 𝐴 𝛥 → 𝖵𝖺𝗋 𝛥 → 𝐴
_⊢∗ _ ∶ 𝖢𝗍𝗑 → 𝖢𝗍𝗑 → 𝖲𝖾𝗍
𝛤 ⊢∗ 𝛥 = 𝖤𝗇𝗏 (𝑇 𝑚 𝛤 ) 𝛥
4
Generic description of well-scoped, well-typed syntaxes
.
𝗌𝗎𝖻−𝗏𝖺𝗋 ∶ ∀ {𝛤 𝛥} → 𝛤 ⊢∗ 𝛥 → 𝖵𝖺𝗋 𝛥 → 𝑇 𝑚 𝛤
𝗌𝗎𝖻−𝗏𝖺𝗋 = 𝗅𝗈𝗈𝗄𝗎𝗉
.
𝗋𝖾𝗇 ∶ ∀ {𝛤 𝛥} → 𝛤 ⊇ 𝛥 → 𝖳𝗆 𝛥 → 𝖳𝗆 𝛤
𝗋𝖾𝗇 𝜌 (𝗏𝖺𝗋 𝑣) = 𝗏𝖺𝗋 (𝗋𝖾𝗇−𝗏𝖺𝗋 𝜌 𝑣)
𝗋𝖾𝗇 𝜌 (𝗅𝖺𝗆 𝑒) = 𝗅𝖺𝗆 (𝗋𝖾𝗇 (𝗄𝖾𝖾𝗉 𝜌) 𝑒)
𝗋𝖾𝗇 𝜌 (𝑓 · 𝑒) = 𝗋𝖾𝗇 𝜌 𝑓 · 𝗋𝖾𝗇 𝜌 𝑒
.
𝗌𝗎𝖻 ∶ ∀ {𝛤 𝛥} → 𝛤 ⊢∗ 𝛥 → 𝖳𝗆 𝛥 → 𝖳𝗆 𝛤
𝗌𝗎𝖻 𝜎 (𝗏𝖺𝗋 𝑣) = 𝗌𝗎𝖻−𝗏𝖺𝗋 𝜎 𝑣
𝗌𝗎𝖻 𝜎 (𝗅𝖺𝗆 𝑒) = 𝗅𝖺𝗆 (𝗌𝗎𝖻 (𝗌𝗁𝗂𝖿𝗍 𝜎) 𝑒)
𝗌𝗎𝖻 𝜎 (𝑓 · 𝑒) = 𝗌𝗎𝖻 𝜎 𝑓 · 𝗌𝗎𝖻 𝜎 𝑒
We omit the definition of some helper functions here, because there is nothing interesting about them.
In fact, it would be fair to say there is nothing interesting in the implementation of 𝗆𝖺𝗉/𝗋𝖾𝗇 and 𝖻𝗂𝗇𝖽/𝗌𝗎𝖻
either; however, they are included here so we can note that they are all defined in terms of pattern matching
on the constructors of 𝖳𝗆. This means not only that the implementation of these two functions need to
be changed every time we wish to change our object language; but also, all proofs about their behaviour
as well; proofs such as the following:
𝗋𝖾𝖿 𝗅𝗋 ∶ ∀ {𝛤 } → 𝛤 ⊇ 𝛤
𝗋𝖾𝖿 𝗅𝗋 {∅} = 𝖽𝗈𝗇𝖾
𝗋𝖾𝖿 𝗅𝗋 {𝛤 , _} = 𝗄𝖾𝖾𝗉 𝗋𝖾𝖿 𝗅𝗋
𝗋𝖾𝗇−𝗋𝖾𝖿 𝗅 ∶ ∀ {𝛤 𝑡} (𝑒 ∶ 𝖳𝗆 𝛤 𝑡) → 𝗋𝖾𝗇 𝗋𝖾𝖿 𝗅𝗋 𝑒 ≡ 𝑒
𝗋𝖾𝗇−𝗋𝖾𝖿 𝗅 (𝗏𝖺𝗋 𝑣) 𝗋𝖾𝗐𝗋𝗂𝗍𝖾 𝗋𝖾𝗇−𝗏𝖺𝗋−𝗋𝖾𝖿𝗅 𝑣 = 𝗋𝖾𝖿𝗅
𝗋𝖾𝗇−𝗋𝖾𝖿 𝗅 (𝗅𝖺𝗆 𝑒) 𝗋𝖾𝗐𝗋𝗂𝗍𝖾 𝗋𝖾𝗇−𝗋𝖾𝖿𝗅 𝑒 = 𝗋𝖾𝖿 𝗅
𝗋𝖾𝗇−𝗋𝖾𝖿 𝗅 (𝑓 · 𝑒) 𝗋𝖾𝗐𝗋𝗂𝗍𝖾 𝗋𝖾𝗇−𝗋𝖾𝖿𝗅 𝑓 | 𝗋𝖾𝗇−𝗋𝖾𝖿𝗅 𝑒 = 𝗋𝖾𝖿𝗅
Just like the definition of 𝗋𝖾𝗇, arguably there is nothing interesting about the proof of 𝗋𝖾𝗇−𝗋𝖾𝖿𝗅 and
similar theorems. However, for a lot of applications, they still need to be written, by hand, and updated
whenever the definition of 𝖳𝗆 changes. In fact, the key motivation for this paper was encountering an
Agda library[16] implementing the usual renaming and substitution lemmas that starts with the following
comment:
-- Copy/paste everything from below,
-- then add/remove cases as necessary,
-- depending on the definition of the syntax.
How can we turn these functions and proofs into a reusable library?
G. Érdi
5
3 Generic description of syntaxes
The standard approach[3, 18, 21, 5] for writing type-generic programs in a dependently-typed setting is
to create a datatype that can be used to point into the universe of types. Generic functions and proofs
can then be written by recursing, simultaneously, on a code of a type and values of the type that is the
interpretation of the given code. If needed, the coded representation can also be chosen to only encode a
circumscribed sub-universe of all types.
Similarly, the road to writing functions like type erasure and simultaneous substitution, and proofs like
renaming by reflexivity, once and for all, is paved with writing them not in terms of a bespoke datatype for
one given object language. Instead, we create in a library a way of writing a description of a language, and
then, given a particular description by the user, synthesize typed and untyped representations, functions
between these representations, and proofs about these functions in the form of syntax-generic programs.
Assuming we have 𝑇 𝑦 : 𝖲𝖾𝗍 provided by the user, the central datatype for describing languages is
𝖣𝖾𝗌𝖼 defined below:
𝖽𝖺𝗍𝖺 𝖡𝗂𝗇𝖽𝖾𝗋 ∶ 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝖻𝗈𝗎𝗇𝖽 𝗎𝗇𝖻𝗈𝗎𝗇𝖽 ∶ 𝖡𝗂𝗇𝖽𝖾𝗋
𝖲𝗁𝖺𝗉𝖾 ∶ ℕ → ℕ → 𝖲𝖾𝗍
𝖲𝗁𝖺𝗉𝖾 𝑛 𝑘 = 𝖵𝖾𝖼 (𝖵𝖾𝖼 𝖡𝗂𝗇𝖽𝖾𝗋 𝑛) 𝑘
𝖽𝖺𝗍𝖺 𝖣𝖾𝗌𝖼 ∶ 𝖲𝖾𝗍𝟣 𝗐𝗁𝖾𝗋𝖾
𝗌𝗀 ∶ (𝐴 ∶ 𝖲𝖾𝗍) (𝑘 ∶ 𝐴 → 𝖣𝖾𝗌𝖼) → 𝖣𝖾𝗌𝖼
𝗇𝗈𝖽𝖾 ∶ (𝑛 ∶ ℕ) {𝑘 ∶ ℕ} (𝑠ℎ𝑎𝑝𝑒 ∶ 𝖲𝗁𝖺𝗉𝖾 𝑛 𝑘) (𝑤𝑡 ∶ 𝖵𝖾𝖼 𝑇 𝑦 𝑛 → 𝖵𝖾𝖼 𝑇 𝑦 𝑘 → 𝑇 𝑦 → 𝖲𝖾𝗍) → 𝖣𝖾𝗌𝖼
The 𝗌𝗀 constructor is the way to put arbitrary user-specified data into the syntax nodes. Because
the rest of the description depends on the stored piece of data, this is also the mechanism by which the
alternatives of the object grammar can be encoded. This matches the corresponding behaviour of [1].
The idea behind the 𝗇𝗈𝖽𝖾 constructor will become clear when we get to deriving a typed representation: the goal is to enable a global view of all newly-bound variables and the types of all sub-terms. 𝑛 is
the number of newly bound variables, and 𝑘 is the number of subterms; 𝑠ℎ𝑎𝑝𝑒 describes the subterm structure: which newly bound variable is in scope in which subterm. 𝑤𝑡 is the user-supplied well-typedness
constraint: a proposition over the types of the 𝑛 newly-bound variables, the types of the 𝑘 sub-terms, and
the type of the term itself.
To give a feel for 𝖣𝖾𝗌𝖼riptions, let’s look at some examples, in increasing order of complexity. For
function 𝖠𝖯𝖯lication, we introduce no new variables, so the shape just prescribes two subterms. The welltypedness constraint ensures the type of the function matches the type of the argument and the result.
𝖠𝖯𝖯 ∶ 𝖣𝖾𝗌𝖼
𝖠𝖯𝖯 = 𝗇𝗈𝖽𝖾 𝟢 ([] ∷ [] ∷ [])
𝜆 { [] (𝑡1 ∷ 𝑡2 ∷ []) 𝑡0 → 𝑡1 ≡ 𝑡2 ▹ 𝑡0 }
𝖫𝖠𝖬bda abstractions can be added either in Curry or Church style. Both bind one new variable and
has one subterm (in which the newly bound variable is in scope). The well-typedness constraint connects
the type of the newly bound variable 𝑡′ and the type of the subterm 𝑢 to the type of the whole lambda
abstraction 𝑡0 . In Church style, the parameter type is stored in a 𝗌𝗀 node and also used in the typing
constraint:
6
Generic description of well-scoped, well-typed syntaxes
𝖽𝖺𝗍𝖺 𝖲𝗍𝗒𝗅𝖾 ∶ 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝖢𝗎𝗋𝗋𝗒 𝖢𝗁𝗎𝗋𝖼𝗁 ∶ 𝖲𝗍𝗒𝗅𝖾
𝖫𝖠𝖬_𝗌𝗍𝗒𝗅𝖾 ∶ 𝖲𝗍𝗒𝗅𝖾 → 𝖣𝖾𝗌𝖼
𝖫𝖠𝖬 𝖢𝗎𝗋𝗋𝗒 𝗌𝗍𝗒𝗅𝖾 = 𝗇𝗈𝖽𝖾 𝟣 ((𝖻𝗈𝗎𝗇𝖽 ∷ []) ∷ [])
𝜆 { (𝑡′ ∷ []) (𝑢 ∷ []) 𝑡0 → 𝑡0 ≡ 𝑡′ ▹ 𝑢 }
𝖫𝖠𝖬 𝖢𝗁𝗎𝗋𝖼𝗁 𝗌𝗍𝗒𝗅𝖾 = 𝗌𝗀 𝖳𝗒 𝜆 𝑡 → 𝗇𝗈𝖽𝖾 𝟣 ((𝖻𝗈𝗎𝗇𝖽 ∷ []) ∷ [])
𝜆 { (𝑡′ ∷ []) (𝑢 ∷ []) 𝑡0 → 𝑡 ≡ 𝑡′ × 𝑡0 ≡ 𝑡 ▹ 𝑢 }
For (non-recursive) 𝖫𝖤𝖳, there are two subterms for the one newly bound variable: the first one is the
definition, and the second one is the rest of the term where the new variable is in scope. Of course, the
type of the definition 𝑢′ must match the type of the newly introduced variable 𝑢; additionally, the type 𝑡0
of the whole term is the same as the type 𝑡 of term where the new variable is in scope:
𝖫𝖤𝖳 ∶ 𝖣𝖾𝗌𝖼
𝖫𝖤𝖳 = 𝗇𝗈𝖽𝖾 𝟣 ((𝗎𝗇𝖻𝗈𝗎𝗇𝖽 ∷ []) ∷ (𝖻𝗈𝗎𝗇𝖽 ∷ []) ∷ [])
𝜆 { (𝑢 ∷ []) (𝑢′ ∷ 𝑡′ ∷ []) 𝑡0 → 𝑢 ≡ 𝑢′ × 𝑡0 ≡ 𝑡′ }
The only tweak we need to make for the recursive version (i.e. 𝖫𝖤𝖳𝖱𝖤𝖢) is to make the newly bound
variable visible in the definition as well:
𝖫𝖤𝖳𝖱𝖤𝖢 ∶ 𝖣𝖾𝗌𝖼
𝖫𝖤𝖳𝖱𝖤𝖢 = 𝗇𝗈𝖽𝖾 𝟣 ((𝖻𝗈𝗎𝗇𝖽 ∷ []) ∷ (𝖻𝗈𝗎𝗇𝖽 ∷ []) ∷ [])
𝜆 { (𝑢 ∷ []) (𝑢′ ∷ 𝑡′ ∷ []) 𝑡0 → 𝑢 ≡ 𝑢′ × 𝑡0 ≡ 𝑡′ }
4 Representations for terms
Now that we have a way of describing our language, it is time to start reaping benefits from it. First, to
do anything meaningful with terms, we need a way to represent them, i.e. some Agda type (for a given
𝖣𝖾𝗌𝖼ription) whose (Agda) values correspond to terms of the object language.
4.1
𝖥𝗈𝗋𝗆: Syntax only
For the representation of purely syntactic 𝖥𝗈𝗋𝗆s with explicit names (with no scoping rules enforced),
we simply take the arity of new names and the arity of subterms from the 𝑛 and 𝑘 arguments of 𝗇𝗈𝖽𝖾.
Crucially, and in contrast to [1], because there is no 𝖣𝖾𝗌𝖼 continuation in 𝗇𝗈𝖽𝖾 that would depend on any
typing information, we can just ignore 𝑤𝑡 and 𝑠ℎ𝑎𝑝𝑒:
𝗆𝗈𝖽𝗎𝗅𝖾 𝖴𝗇𝗌𝖼𝗈𝗉𝖾𝖽 (𝑑𝑒𝑠𝑐 ∶ 𝖣𝖾𝗌𝖼) (𝑛𝑎𝑚𝑒 ∶ 𝖲𝖾𝗍) 𝗐𝗁𝖾𝗋𝖾
𝗆𝗎𝗍𝗎𝖺𝗅
𝖽𝖺𝗍𝖺 𝖥𝗈𝗋𝗆 ∶ 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗏𝖺𝗋 ∶ 𝑛𝑎𝑚𝑒
→ 𝖥𝗈𝗋𝗆
𝖼𝗈𝗇 ∶ 𝖢𝗈𝗇 𝑑𝑒𝑠𝑐 → 𝖥𝗈𝗋𝗆
𝖽𝖺𝗍𝖺 𝖢𝗈𝗇 ∶ 𝖣𝖾𝗌𝖼 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗌𝗀 ∶ ∀ {𝐴 𝑘} 𝑥 → 𝖢𝗈𝗇 (𝑘 𝑥) → 𝖢𝗈𝗇 (𝗌𝗀 𝐴 𝑘)
𝗇𝗈𝖽𝖾 ∶ ∀ {𝑛 𝑘 𝑠ℎ 𝑤𝑡} (𝑛𝑠 ∶ 𝖵𝖾𝖼 𝑛𝑎𝑚𝑒 𝑛) (𝑒𝑠 ∶ 𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 𝑘)
G. Érdi
7
→ 𝖢𝗈𝗇 (𝗇𝗈𝖽𝖾 𝑛 {𝑘} 𝑠ℎ 𝑤𝑡)
𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 ∶ ℕ → 𝖲𝖾𝗍
𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 = 𝖵𝖾𝖼 𝖥𝗈𝗋𝗆
4.2
𝖤𝗑𝗉𝗋: Untyped, well-scoped
The untyped representation is still indexed by the number of variables in scope. At any inductive site, we
increase that by the number of newly bound variables that are in scope, as prescribed by the 𝑠ℎ𝑎𝑝𝑒 in the
𝗇𝗈𝖽𝖾 constructor of 𝖣𝖾𝗌𝖼. Because this representation is untyped, the well-typedness constraint 𝑤𝑡 is not
used.
𝖼𝗈𝗎𝗇𝗍𝖵 ∶ ∀ {𝑛} → 𝖵𝖾𝖼 𝖡𝗂𝗇𝖽𝖾𝗋 𝑛 → ℕ
𝖼𝗈𝗎𝗇𝗍𝖵 []
=𝟢
𝖼𝗈𝗎𝗇𝗍𝖵 (𝖻𝗈𝗎𝗇𝖽 ∷ 𝑏𝑠) = 𝗌𝗎𝖼 (𝖼𝗈𝗎𝗇𝗍𝖵 𝑏𝑠)
𝖼𝗈𝗎𝗇𝗍𝖵 (𝗎𝗇𝖻𝗈𝗎𝗇𝖽 ∷ 𝑏𝑠) = 𝖼𝗈𝗎𝗇𝗍𝖵 𝑏𝑠
𝗆𝗈𝖽𝗎𝗅𝖾 𝖴𝗇𝗍𝗒𝗉𝖾𝖽 (𝑑𝑒𝑠𝑐 ∶ 𝖣𝖾𝗌𝖼) 𝗐𝗁𝖾𝗋𝖾
𝗆𝗎𝗍𝗎𝖺𝗅
𝖽𝖺𝗍𝖺 𝖤𝗑𝗉𝗋 (𝑉 ∶ ℕ) ∶ 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗏𝖺𝗋 ∶ 𝖥𝗂𝗇 𝑉
→ 𝖤𝗑𝗉𝗋 𝑉
𝖼𝗈𝗇 ∶ 𝖢𝗈𝗇 𝑉 𝑑𝑒𝑠𝑐 → 𝖤𝗑𝗉𝗋 𝑉
𝖽𝖺𝗍𝖺 𝖢𝗈𝗇 (𝑉 ∶ ℕ) ∶ 𝖣𝖾𝗌𝖼 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗌𝗀 ∶ ∀ {𝐴 𝑘} 𝑥 → 𝖢𝗈𝗇 𝑉 (𝑘 𝑥) → 𝖢𝗈𝗇 𝑉 (𝗌𝗀 𝐴 𝑘)
𝗇𝗈𝖽𝖾 ∶ ∀ {𝑛 𝑘 𝑠ℎ 𝑤𝑡} (𝑒𝑠 ∶ 𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 𝑉 𝑠ℎ)
→ 𝖢𝗈𝗇 𝑉 (𝗇𝗈𝖽𝖾 𝑛 {𝑘} 𝑠ℎ 𝑤𝑡)
𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 ∶ ∀ {𝑛 𝑘} → ℕ → 𝖲𝗁𝖺𝗉𝖾 𝑛 𝑘 → 𝖲𝖾𝗍
𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 𝑉 = 𝖠𝗅𝗅 (𝜆 𝑏𝑠 → 𝖤𝗑𝗉𝗋 (𝖼𝗈𝗎𝗇𝗍𝖵 𝑏𝑠 + 𝑉 ))
4.3
𝖳𝗆: Intrinsically well-typed
First, to compute the extended context of subterms in the generic case when there are multiple newly
bound variables, we make an object type-ornamented version of _+_(à la [20]) and 𝖼𝗈𝗎𝗇𝗍𝖵:
_<><_ ∶ 𝖢𝗍𝗑 → 𝖫𝗂𝗌𝗍 𝑇 𝑦 → 𝖢𝗍𝗑
𝛤 <>< [] = 𝛤
𝛤 <>< (𝑡 ∷ 𝑡𝑠) = (𝛤 , 𝑡) <>< 𝑡𝑠
For the intrinsically typed representation, we generalize the standard technique of representing GADTs[23]:
the indices corresponding to the types of the newly bound variables and the subterms are existentially
bound, and the index of the term itself is turned into a parameter; then, an explicit witness of welltypedness restores guardedness. This is a generalization in the sense that well-typedness can be an arbitrary proposition, not necessarily just a conjunction of equalities.
8
Generic description of well-scoped, well-typed syntaxes
𝗏𝗂𝗌𝗂𝖻𝗅𝖾 ∶ ∀ {𝑛} → 𝖵𝖾𝖼 𝖡𝗂𝗇𝖽𝖾𝗋 𝑛 → 𝖵𝖾𝖼 𝑇 𝑦 𝑛 → 𝖫𝗂𝗌𝗍 𝑇 𝑦
𝗏𝗂𝗌𝗂𝖻𝗅𝖾 []
[]
= []
𝗏𝗂𝗌𝗂𝖻𝗅𝖾 (𝖻𝗈𝗎𝗇𝖽 ∷ 𝑏𝑠) (𝑡 ∷ 𝑡𝑠) = 𝑡 ∷ 𝗏𝗂𝗌𝗂𝖻𝗅𝖾 𝑏𝑠 𝑡𝑠
𝗏𝗂𝗌𝗂𝖻𝗅𝖾 (𝗎𝗇𝖻𝗈𝗎𝗇𝖽 ∷ 𝑏𝑠) (_ ∷ 𝑡𝑠) = 𝗏𝗂𝗌𝗂𝖻𝗅𝖾 𝑏𝑠 𝑡𝑠
𝗆𝗈𝖽𝗎𝗅𝖾 𝖳𝗒𝗉𝖾𝖽 (𝑑𝑒𝑠𝑐 ∶ 𝖣𝖾𝗌𝖼) 𝗐𝗁𝖾𝗋𝖾
𝗆𝗎𝗍𝗎𝖺𝗅
𝖽𝖺𝗍𝖺 𝖳𝗆 (𝛤 ∶ 𝖢𝗍𝗑) ∶ 𝑇 𝑦 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗏𝖺𝗋 ∶ ∀ {𝑡} → 𝖵𝖺𝗋 𝛤 𝑡
→ 𝖳𝗆 𝛤 𝑡
𝖼𝗈𝗇 ∶ ∀ {𝑡} → 𝖢𝗈𝗇 𝛤 𝑡 𝑑𝑒𝑠𝑐 → 𝖳𝗆 𝛤 𝑡
𝖽𝖺𝗍𝖺 𝖢𝗈𝗇 (𝛤 ∶ 𝖢𝗍𝗑) (𝑡 ∶ 𝑇 𝑦) ∶ 𝖣𝖾𝗌𝖼 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗌𝗀 ∶ ∀ {𝐴 𝑘} 𝑥 → 𝖢𝗈𝗇 𝛤 𝑡 (𝑘 𝑥) → 𝖢𝗈𝗇 𝛤 𝑡 (𝗌𝗀 𝐴 𝑘)
𝗇𝗈𝖽𝖾 ∶ ∀ {𝑛 𝑘 𝑠ℎ 𝑤𝑡} (𝑡𝑠0 ∶ 𝖵𝖾𝖼 𝑇 𝑦 𝑛) {𝑡𝑠 ∶ 𝖵𝖾𝖼 𝑇 𝑦 𝑘}
(𝑒𝑠 ∶ 𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 𝛤 𝑡𝑠0 𝑠ℎ 𝑡𝑠)
{{_ ∶ 𝑤𝑡 𝑡𝑠0 𝑡𝑠 𝑡}}
→ 𝖢𝗈𝗇 𝛤 𝑡 (𝗇𝗈𝖽𝖾 𝑛 𝑠ℎ 𝑤𝑡)
𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 ∶ ∀ {𝑛 𝑘} → 𝖢𝗍𝗑 → 𝖵𝖾𝖼 𝑇 𝑦 𝑛 → 𝖲𝗁𝖺𝗉𝖾 𝑛 𝑘 → 𝖵𝖾𝖼 𝑇 𝑦 𝑘 → 𝖲𝖾𝗍
𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 𝛤 𝑡𝑠0 = 𝖠𝗅𝗅𝟤 (𝜆 𝑏𝑠 → 𝖳𝗆 (𝛤 <>< 𝗏𝗂𝗌𝗂𝖻𝗅𝖾 𝑏𝑠 𝑡𝑠0 ))
4.4
𝖧𝖮𝖳𝗆: Typed PHOAS
Although the PHOAS representation is not used in the rest of this paper, we included it here for the sake
of completeness.1 It differs from 𝖳𝗆 only in that 𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 subterms are represented as functions from
𝑉 ariables to (higher-order) terms:
𝗆𝗈𝖽𝗎𝗅𝖾 𝖯𝖧𝖮𝖠𝖲 (𝑑𝑒𝑠𝑐 ∶ 𝖣𝖾𝗌𝖼) 𝗐𝗁𝖾𝗋𝖾
𝗆𝗎𝗍𝗎𝖺𝗅
𝖽𝖺𝗍𝖺 𝖧𝖮𝖳𝗆 (𝑉 ∶ 𝑇 𝑦 → 𝖲𝖾𝗍) ∶ 𝑇 𝑦 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗏𝖺𝗋 ∶ ∀ {𝑡} → 𝑉 𝑡
→ 𝖧𝖮𝖳𝗆 𝑉 𝑡
𝖼𝗈𝗇 ∶ ∀ {𝑡} → 𝖢𝗈𝗇 𝑉 𝑡 𝑑𝑒𝑠𝑐 → 𝖧𝖮𝖳𝗆 𝑉 𝑡
𝖽𝖺𝗍𝖺 𝖢𝗈𝗇 (𝑉 ∶ 𝑇 𝑦 → 𝖲𝖾𝗍) (𝑡 ∶ 𝑇 𝑦) ∶ 𝖣𝖾𝗌𝖼 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗌𝗀 ∶ ∀ {𝐴 𝑘} 𝑥 → 𝖢𝗈𝗇 𝑉 𝑡 (𝑘 𝑥) → 𝖢𝗈𝗇 𝑉 𝑡 (𝗌𝗀 𝐴 𝑘)
𝗇𝗈𝖽𝖾 ∶ ∀ {𝑛 𝑘 𝑠ℎ 𝑤𝑡} (𝑡𝑠0 ∶ 𝖵𝖾𝖼 𝑇 𝑦 𝑛) {𝑡𝑠 ∶ 𝖵𝖾𝖼 𝑇 𝑦 𝑘}
(𝑒𝑠 ∶ 𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 𝑉 𝑡𝑠0 𝑠ℎ 𝑡𝑠)
{{_ ∶ 𝑤𝑡 𝑡𝑠0 𝑡𝑠 𝑡}}
→ 𝖢𝗈𝗇 𝑉 𝑡 (𝗇𝗈𝖽𝖾 𝑛 𝑠ℎ 𝑤𝑡)
𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 ∶ ∀ {𝑛 𝑘} → (𝑇 𝑦 → 𝖲𝖾𝗍) → 𝖵𝖾𝖼 𝑇 𝑦 𝑛 → 𝖲𝗁𝖺𝗉𝖾 𝑛 𝑘 → 𝖵𝖾𝖼 𝑇 𝑦 𝑘 → 𝖲𝖾𝗍
𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 𝑉 𝑡𝑠0 = 𝖠𝗅𝗅𝟤 (𝜆 𝑏𝑠 𝑡 → (𝖠𝗅𝗅 𝑉 (𝗏𝗂𝗌𝗂𝖻𝗅𝖾 𝑏𝑠 𝑡𝑠0 ) → 𝖧𝖮𝖳𝗆 𝑉 𝑡))
1
Our full library implementation also includes a conversion function from 𝖳𝗆 to 𝖧𝖮𝖳𝗆
G. Érdi
9
4.5 Example: Type erasure
The building blocks introduced so far already allow us to write some interesting functions generically for
all syntaxes. As an example, here is a conversion function between the typed and the untyped representation of terms. Because the two representations are computed from a single shared 𝖣𝖾𝗌𝖼ription, the type of
𝗎𝗇𝗍𝗒𝗉𝖾 is more valuable as a specification than it would be if it converted between two ad-hoc datatypes.
𝗎𝗇𝗍𝗒𝗉𝖾−𝗏𝖺𝗋 ∶ ∀ {𝛤 𝑡} → 𝖵𝖺𝗋 𝛤 𝑡 → 𝖥𝗂𝗇 (𝗌𝗂𝗓𝖾 𝛤 )
𝗎𝗇𝗍𝗒𝗉𝖾−𝗏𝖺𝗋 𝗏𝗓 = 𝗓𝖾𝗋𝗈
𝗎𝗇𝗍𝗒𝗉𝖾−𝗏𝖺𝗋 (𝗏𝗌 𝑣) = 𝗌𝗎𝖼 (𝗎𝗇𝗍𝗒𝗉𝖾−𝗏𝖺𝗋 𝑣)
𝗎𝗇𝗍𝗒𝗉𝖾 ∶ ∀ {𝛤 𝑡} → 𝖳𝗆 𝛤 𝑡 → 𝖤𝗑𝗉𝗋 (𝗌𝗂𝗓𝖾 𝛤 )
𝗎𝗇𝗍𝗒𝗉𝖾 (𝗏𝖺𝗋 𝑣) = 𝗏𝖺𝗋 (𝗎𝗇𝗍𝗒𝗉𝖾−𝗏𝖺𝗋 𝑣)
𝗎𝗇𝗍𝗒𝗉𝖾 (𝖼𝗈𝗇 𝑒) = 𝖼𝗈𝗇 (𝗎𝗇𝗍𝗒𝗉𝖾−𝖼𝗈𝗇 𝑒)
𝗐𝗁𝖾𝗋𝖾
𝗎𝗇𝗍𝗒𝗉𝖾∗ ∶ ∀ {𝛤 𝑛 𝑘 𝑠ℎ 𝑡𝑠0 𝑡𝑠} → 𝖳.𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 {𝑛} {𝑘} 𝛤 𝑡𝑠0 𝑠ℎ 𝑡𝑠 → 𝖴.𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 (𝗌𝗂𝗓𝖾 𝛤 ) 𝑠ℎ
𝗎𝗇𝗍𝗒𝗉𝖾∗
[]
= []
𝗎𝗇𝗍𝗒𝗉𝖾∗ {𝑠ℎ = 𝑏𝑠 ∷ _} (𝑒 ∷ 𝑒𝑠) = 𝗌𝗎𝖻𝗌𝗍 𝖤𝗑𝗉𝗋 (𝗌𝗂𝗓𝖾−<>< _ 𝑏𝑠 _) (𝗎𝗇𝗍𝗒𝗉𝖾 𝑒) ∷ 𝗎𝗇𝗍𝗒𝗉𝖾∗ 𝑒𝑠
𝗎𝗇𝗍𝗒𝗉𝖾−𝖼𝗈𝗇 ∶ ∀ {𝛤 𝑡 𝑐} → 𝖳.𝖢𝗈𝗇 𝛤 𝑡 𝑐 → 𝖴.𝖢𝗈𝗇 (𝗌𝗂𝗓𝖾 𝛤 ) 𝑐
𝗎𝗇𝗍𝗒𝗉𝖾−𝖼𝗈𝗇 (𝗌𝗀 𝑥 𝑒)
= 𝗌𝗀 𝑥 (𝗎𝗇𝗍𝗒𝗉𝖾−𝖼𝗈𝗇 𝑒)
𝗎𝗇𝗍𝗒𝗉𝖾−𝖼𝗈𝗇 (𝗇𝗈𝖽𝖾 𝑡𝑠0 𝑒𝑠) = 𝗇𝗈𝖽𝖾 (𝗎𝗇𝗍𝗒𝗉𝖾∗ 𝑒𝑠)
5 Type-preserving renaming and substitution
Now that we have a typed representation, implementing type-preserving renaming and substitution becomes a straightforward matter of recursing simultaneously on the description (implicitly) and a given
term of the language corresponding to that description (explicitly):
𝗆𝗎𝗍𝗎𝖺𝗅
.
𝗋𝖾𝗇 ∶ ∀ {𝛤 𝛥} → 𝛤 ⊇ 𝛥 → 𝖳𝗆 𝛥 → 𝖳𝗆 𝛤
𝗋𝖾𝗇 𝜌 (𝗏𝖺𝗋 𝑣) = 𝗏𝖺𝗋 (𝗋𝖾𝗇−𝗏𝖺𝗋 𝜌 𝑣)
𝗋𝖾𝗇 𝜌 (𝖼𝗈𝗇 𝑐) = 𝖼𝗈𝗇 (𝗋𝖾𝗇−𝖼𝗈𝗇 𝜌 𝑐)
𝗋𝖾𝗇−𝖼𝗈𝗇 ∶ ∀ {𝑑 𝛤 𝛥 𝑡} → 𝛤 ⊇ 𝛥 → 𝖢𝗈𝗇 𝛥 𝑡 𝑑 → 𝖢𝗈𝗇 𝛤 𝑡 𝑑
𝗋𝖾𝗇−𝖼𝗈𝗇 𝜌 (𝗌𝗀 𝑥 𝑘)
= 𝗌𝗀 𝑥 (𝗋𝖾𝗇−𝖼𝗈𝗇 𝜌 𝑘)
𝗋𝖾𝗇−𝖼𝗈𝗇 𝜌 (𝗇𝗈𝖽𝖾 𝑡𝑠0 𝑒𝑠) = 𝗇𝗈𝖽𝖾 𝑡𝑠0 (𝗋𝖾𝗇∗ 𝜌 𝑒𝑠)
𝗋𝖾𝗇∗ ∶ ∀ {𝛤 𝛥} {𝑛 𝑘 𝑡𝑠0 𝑠ℎ 𝑡𝑠} → 𝛤 ⊇ 𝛥 → 𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 {𝑛} {𝑘} 𝛥 𝑡𝑠0 𝑠ℎ 𝑡𝑠 → 𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 𝛤 𝑡𝑠0 𝑠ℎ 𝑡𝑠
𝗋𝖾𝗇∗
𝜌 []
= []
𝗋𝖾𝗇∗ {𝑠ℎ = 𝑏𝑠 ∷ _} 𝜌 (𝑒 ∷ 𝑒𝑠) = 𝗋𝖾𝗇 (𝗄𝖾𝖾𝗉∗ (𝗏𝗂𝗌𝗂𝖻𝗅𝖾 𝑏𝑠 _) 𝜌) 𝑒 ∷ 𝗋𝖾𝗇∗ 𝜌 𝑒𝑠
𝗆𝗎𝗍𝗎𝖺𝗅
.
𝗌𝗎𝖻 ∶ ∀ {𝛤 𝛥} → 𝛤 ⊢∗ 𝛥 → 𝖳𝗆 𝛥 → 𝖳𝗆 𝛤
𝗌𝗎𝖻 𝜎 (𝗏𝖺𝗋 𝑣) = 𝗌𝗎𝖻−𝗏𝖺𝗋 𝜎 𝑣
𝗌𝗎𝖻 𝜎 (𝖼𝗈𝗇 𝑐) = 𝖼𝗈𝗇 (𝗌𝗎𝖻−𝖼𝗈𝗇 𝜎 𝑐)
10
Generic description of well-scoped, well-typed syntaxes
𝗌𝗎𝖻−𝖼𝗈𝗇 ∶ ∀ {𝛤 𝛥 𝑡 𝑐} → 𝛤 ⊢∗ 𝛥 → 𝖢𝗈𝗇 𝛥 𝑡 𝑐 → 𝖢𝗈𝗇 𝛤 𝑡 𝑐
𝗌𝗎𝖻−𝖼𝗈𝗇 𝜎 (𝗌𝗀 𝑥 𝑐)
= 𝗌𝗀 𝑥 (𝗌𝗎𝖻−𝖼𝗈𝗇 𝜎 𝑐)
𝗌𝗎𝖻−𝖼𝗈𝗇 𝜎 (𝗇𝗈𝖽𝖾 𝑡𝑠0 𝑒𝑠) = 𝗇𝗈𝖽𝖾 𝑡𝑠0 (𝗌𝗎𝖻∗ 𝜎 𝑒𝑠)
𝗌𝗎𝖻∗ ∶ ∀ {𝛤 𝛥 𝑛 𝑘 𝑠ℎ 𝑡𝑠0 𝑡𝑠} → 𝛤 ⊢∗ 𝛥 → 𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 {𝑛} {𝑘} 𝛥 𝑡𝑠0 𝑠ℎ 𝑡𝑠 → 𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 𝛤 𝑡𝑠0 𝑠ℎ 𝑡𝑠
𝗌𝗎𝖻∗
𝜎 []
= []
∗
𝗌𝗎𝖻 {𝑠ℎ = 𝑏𝑠 ∷ _} 𝜎 (𝑒 ∷ 𝑒𝑠) = 𝗌𝗎𝖻 (𝗌𝗁𝗂𝖿𝗍∗ (𝗏𝗂𝗌𝗂𝖻𝗅𝖾 𝑏𝑠 _) 𝜎) 𝑒 ∷ 𝗌𝗎𝖻∗ 𝜎 𝑒𝑠
Finally, we can fulfill our original goal of proving properties of renamings and substitutions once and
for all. Our library implementation contains all proofs from [16] adapted to use our generic representation,
thus avoiding the need to change proofs as the object language is changed. As an illustration, here is
𝗋𝖾𝗇−𝗋𝖾𝖿𝗅 from earlier, generically for all syntaxes:
𝗆𝗎𝗍𝗎𝖺𝗅
𝗋𝖾𝗇−𝗋𝖾𝖿𝗅 ∶ ∀ {𝛤 𝑡} → (𝑒 ∶ 𝖳𝗆 𝛤 𝑡) → 𝗋𝖾𝗇 𝗋𝖾𝖿𝗅𝗋 𝑒 ≡ 𝑒
𝗋𝖾𝗇−𝗋𝖾𝖿𝗅 (𝗏𝖺𝗋 𝑣) 𝗋𝖾𝗐𝗋𝗂𝗍𝖾 𝗋𝖾𝗇−𝗏𝖺𝗋−𝗋𝖾𝖿𝗅 𝑣 = 𝗋𝖾𝖿𝗅
𝗋𝖾𝗇−𝗋𝖾𝖿𝗅 (𝖼𝗈𝗇 𝑒) 𝗋𝖾𝗐𝗋𝗂𝗍𝖾 𝗋𝖾𝗇−𝖼𝗈𝗇−𝗋𝖾𝖿𝗅 𝑒 = 𝗋𝖾𝖿 𝗅
𝗋𝖾𝗇−𝖼𝗈𝗇−𝗋𝖾𝖿𝗅 ∶ ∀ {𝑐 𝛤 𝑡} (𝑒 ∶ 𝖢𝗈𝗇 𝛤 𝑡 𝑐) → 𝗋𝖾𝗇−𝖼𝗈𝗇 𝗋𝖾𝖿𝗅𝗋 𝑒 ≡ 𝑒
𝗋𝖾𝗇−𝖼𝗈𝗇−𝗋𝖾𝖿𝗅 (𝗌𝗀 𝑥 𝑒)
𝗋𝖾𝗐𝗋𝗂𝗍𝖾 𝗋𝖾𝗇−𝖼𝗈𝗇−𝗋𝖾𝖿𝗅 𝑒 = 𝗋𝖾𝖿𝗅
𝗋𝖾𝗇−𝖼𝗈𝗇−𝗋𝖾𝖿𝗅 (𝗇𝗈𝖽𝖾 𝑡𝑠 𝑒𝑠) 𝗋𝖾𝗐𝗋𝗂𝗍𝖾 𝗋𝖾𝗇∗ −𝗋𝖾𝖿𝗅 𝑒𝑠 = 𝗋𝖾𝖿𝗅
𝗋𝖾𝗇∗ −𝗋𝖾𝖿𝗅 ∶ ∀ {𝛤 𝑛 𝑘 𝑠ℎ 𝑡𝑠0 𝑡𝑠} (𝑒𝑠 ∶ 𝖢𝗁𝗂𝗅𝖽𝗋𝖾𝗇 {𝑛} {𝑘} 𝛤 𝑡𝑠0 𝑠ℎ 𝑡𝑠) → 𝗋𝖾𝗇∗ 𝗋𝖾𝖿𝗅𝗋 𝑒𝑠 ≡ 𝑒𝑠
𝗋𝖾𝗇∗ −𝗋𝖾𝖿𝗅
[]
= 𝗋𝖾𝖿𝗅
∗
𝗋𝖾𝗇 −𝗋𝖾𝖿𝗅 {𝛤 } {𝑠ℎ = 𝑏𝑠 ∷ _} {𝑡𝑠0 } (𝑒 ∷ 𝑒𝑠) 𝗋𝖾𝗐𝗋𝗂𝗍𝖾 𝗄𝖾𝖾𝗉∗ −𝗋𝖾𝖿𝗅 {𝛤 } (𝗏𝗂𝗌𝗂𝖻𝗅𝖾 𝑏𝑠 𝑡𝑠0 ) =
𝖼𝗈𝗇𝗀𝟤 _∷_ (𝗋𝖾𝗇−𝗋𝖾𝖿𝗅 𝑒) (𝗋𝖾𝗇∗ −𝗋𝖾𝖿𝗅 𝑒𝑠)
6 Simply Typed Lambda Calculus, generically
In this section, we show details of modeling the simply typed lambda calculus by giving its 𝖣𝖾𝗌𝖼ription
and using the derived representations for terms.
Using the definitions from section 3, we put it all together using a simple datatype for tagging nodes
by their grammar production rules:
𝖽𝖺𝗍𝖺 ‵ 𝖲𝖳𝖫𝖢 ∶ 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
‵
𝖺𝗉𝗉 ‵ 𝗅𝖺𝗆 ∶ ‵ 𝖲𝖳𝖫𝖢
𝖲𝖳𝖫𝖢 ∶ 𝖣𝖾𝗌𝖼
𝖲𝖳𝖫𝖢 = 𝗌𝗀 ‵ 𝖲𝖳𝖫𝖢 𝜆
{ ‵ 𝖺𝗉𝗉 → 𝖠𝖯𝖯
; ‵ 𝗅𝖺𝗆 → 𝖫𝖠𝖬 𝖢𝗎𝗋𝗋𝗒 𝗌𝗍𝗒𝗅𝖾
}
Note that there is no explicit mention of variable occurrences (the 𝗏𝖺𝗋 constructor of our original 𝖳𝗆
datatype in section 2) in this description. That is because in the framework we are presenting, variables
are understood to be (unavoidably) part of the syntax. In fact, the concept of variables is the crucial
G. Érdi
11
differentiator of the syntaxes we consider here, versus arbitrary inductive definitions. This is because the
functorial and monadic structure we expect of terms is defined in terms of variables.
For the various datatypes representing terms, we can recover _·_ and 𝗅𝖺𝗆 (the constructors of the
datatypes from section 2) with some pattern synonyms[27].2 For the 𝖳𝗒𝗉𝖾𝖽 representation, the pattern
synonyms include matching on the explicit well-typedness witnesses to refine the scrutinee 𝖳𝗆’s type index whenever they are used; for the 𝖴𝗇𝗍𝗒𝗉𝖾𝖽 and 𝖴𝗇𝗌𝖼𝗈𝗉𝖾𝖽 representations, the pattern synonyms contain
less detail while matching on the same shape:
𝗈𝗉𝖾𝗇 𝖴𝗇𝗍𝗒𝗉𝖾𝖽 𝖲𝖳𝖫𝖢
-- _·_ : ∀ {V} → Expr V → Expr V → Expr V
𝗉𝖺𝗍𝗍𝖾𝗋𝗇 _·_ 𝑓 𝑒 = 𝖼𝗈𝗇 (𝗌𝗀 ‵ 𝖺𝗉𝗉 (𝗇𝗈𝖽𝖾 (𝑓 ∷ 𝑒 ∷ [])))
-- lam : ∀ {V} → Ty → Expr (1 + V) → Expr V
𝗉𝖺𝗍𝗍𝖾𝗋𝗇 𝗅𝖺𝗆 𝑡 𝑒 = 𝖼𝗈𝗇 (𝗌𝗀 ‵ 𝗅𝖺𝗆 (𝗌𝗀 𝑡 (𝗇𝗈𝖽𝖾 (𝑒 ∷ []))))
𝗈𝗉𝖾𝗇 𝖴𝗇𝗌𝖼𝗈𝗉𝖾𝖽 𝖲𝖳𝖫𝖢 𝑁𝑎𝑚𝑒
-- _·_ : Form → Form → Form
𝗉𝖺𝗍𝗍𝖾𝗋𝗇 _·_ 𝑓 𝑒 = 𝖼𝗈𝗇 (𝗌𝗀 ‵ 𝖺𝗉𝗉 (𝗇𝗈𝖽𝖾 [] (𝑓 ∷ 𝑒 ∷ [])))
-- lam : Name → Ty → Form → Form
𝗉𝖺𝗍𝗍𝖾𝗋𝗇 𝗅𝖺𝗆 𝑛 𝑡 𝑒 = 𝖼𝗈𝗇 (𝗌𝗀 ‵ 𝗅𝖺𝗆 (𝗌𝗀 𝑡 (𝗇𝗈𝖽𝖾 (𝑛 ∷ []) (𝑒 ∷ []))))
𝗈𝗉𝖾𝗇 𝖳𝗒𝗉𝖾𝖽 𝖲𝖳𝖫𝖢
-- _·_ : ∀ {Γ t u} → Tm Γ (t ▹ u) → Tm Γ t → Tm Γ u
𝗉𝖺𝗍𝗍𝖾𝗋𝗇 _·_ 𝑓 𝑒 = 𝖼𝗈𝗇 (𝗌𝗀 ‵ 𝖺𝗉𝗉 (𝗇𝗈𝖽𝖾 [] (𝑓 ∷ 𝑒 ∷ []) {{𝗋𝖾𝖿𝗅}}))
-- lam : ∀ {Γ t u} → Tm (Γ , t) u → Tm Γ (t ▹ u)
𝗉𝖺𝗍𝗍𝖾𝗋𝗇 𝗅𝖺𝗆 𝑒 = 𝖼𝗈𝗇 (𝗌𝗀 ‵ 𝗅𝖺𝗆 (𝗇𝗈𝖽𝖾 (_ ∷ []) (𝑒 ∷ []) {{𝗋𝖾𝖿𝗅}}))
6.1 Example: Desugaring let bindings
The following example demonstrates that once we define these handy pattern synonyms, writing nongeneric intensional transformations is no more unwieldy than with a direct, datatype-based approach.
We define, in one fell swoop, languages in two 𝖥𝗅𝖺𝗏𝗈𝗎𝗋s in Church and Curry style, with ‵ 𝗅𝖾𝗍 bindings
only permitted in the 𝗌𝗎𝗀𝖺𝗋𝖾𝖽 flavour. (𝗅𝖾𝗍 is unfortunately an Agda keyword, hence the name 𝗅𝖾𝗍𝗏𝖺𝗋 for
the pattern synonym)
𝖽𝖺𝗍𝖺 𝖥𝗅𝖺𝗏𝗈𝗎𝗋 ∶ 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
𝗌𝗎𝗀𝖺𝗋𝖾𝖽 𝖽𝖾𝗌𝗎𝗀𝖺𝗋𝖾𝖽 ∶ 𝖥𝗅𝖺𝗏𝗈𝗎𝗋
𝖽𝖺𝗍𝖺 ‵ 𝖲𝖳𝖫𝖢 ∶ 𝖥𝗅𝖺𝗏𝗈𝗎𝗋 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
2
Since Agda’s pattern synonyms are not typed, there is no way to attach a type signature to them[26]. In this paper, we will
give signatures in comments just for clarity.
12
Generic description of well-scoped, well-typed syntaxes
‵
‵
𝖺𝗉𝗉 ‵ 𝗅𝖺𝗆 ∶ ∀ {𝜑} → ‵ 𝖲𝖳𝖫𝖢 𝜑
𝗅𝖾𝗍 ∶ ‵ 𝖲𝖳𝖫𝖢 𝗌𝗎𝗀𝖺𝗋𝖾𝖽
𝖲𝖳𝖫𝖢 ∶ 𝖥𝗅𝖺𝗏𝗈𝗎𝗋 → 𝖲𝗍𝗒𝗅𝖾 → 𝖣𝖾𝗌𝖼
𝖲𝖳𝖫𝖢 𝜑 𝑠 = 𝗌𝗀 (‵ 𝖲𝖳𝖫𝖢 𝜑) 𝜆
{ ‵ 𝖺𝗉𝗉 → 𝖠𝖯𝖯
; ‵ 𝗅𝖺𝗆 → 𝖫𝖠𝖬 𝑠 𝗌𝗍𝗒𝗅𝖾
; ‵ 𝗅𝖾𝗍 → 𝖫𝖤𝖳
}
𝖳𝗆 ∶ 𝖥𝗅𝖺𝗏𝗈𝗎𝗋 → 𝖲𝗍𝗒𝗅𝖾 → 𝖢𝗍𝗑 → 𝖳𝗒 → 𝖲𝖾𝗍
𝖳𝗆 𝜑 𝑠 = 𝖳𝗒𝗉𝖾𝖽.𝖳𝗆 (𝖲𝖳𝖫𝖢 𝜑 𝑠)
𝗉𝖺𝗍𝗍𝖾𝗋𝗇 _·_ 𝑓 𝑒 = 𝖼𝗈𝗇 (𝗌𝗀 ‵ 𝖺𝗉𝗉 (𝗇𝗈𝖽𝖾 [] (𝑓 ∷ 𝑒 ∷ []) {{𝗋𝖾𝖿𝗅}}))
𝗉𝖺𝗍𝗍𝖾𝗋𝗇 𝗅𝖺𝗆 𝑒
= 𝖼𝗈𝗇 (𝗌𝗀 ‵ 𝗅𝖺𝗆 (𝗇𝗈𝖽𝖾 (_ ∷ []) (𝑒 ∷ []) {{𝗋𝖾𝖿𝗅}}))
′
𝗉𝖺𝗍𝗍𝖾𝗋𝗇 𝗅𝖺𝗆 𝑡 𝑒 = 𝖼𝗈𝗇 (𝗌𝗀 ‵ 𝗅𝖺𝗆 (𝗌𝗀 𝑡 ((𝗇𝗈𝖽𝖾 (_ ∷ []) (𝑒 ∷ []) {{𝗋𝖾𝖿𝗅 , 𝗋𝖾𝖿𝗅}}))))
𝗉𝖺𝗍𝗍𝖾𝗋𝗇 𝗅𝖾𝗍𝗏𝖺𝗋 𝑒0 𝑒 = 𝖼𝗈𝗇 (𝗌𝗀 ‵ 𝗅𝖾𝗍 (𝗇𝗈𝖽𝖾 (_ ∷ []) (𝑒0 ∷ 𝑒 ∷ []) {{𝗋𝖾𝖿𝗅 , 𝗋𝖾𝖿𝗅}}))
Then we can define a desugaring transformation in exactly the same way as we would do in direct
style; the type ensures that the function is object type-preserving, and the result contains no let bindings.
The transformation of both Curry and Church style languages is implemented in a single function.
.
𝖽𝖾𝗌𝗎𝗀𝖺𝗋 ∶ ∀ {𝑠 𝜑 𝛤 } → 𝖳𝗆 𝜑 𝑠 𝛤 → 𝖳𝗆 𝖽𝖾𝗌𝗎𝗀𝖺𝗋𝖾𝖽 𝑠 𝛤
𝖽𝖾𝗌𝗎𝗀𝖺𝗋
(𝗏𝖺𝗋 𝑣)
= 𝗏𝖺𝗋 𝑣
𝖽𝖾𝗌𝗎𝗀𝖺𝗋
(𝑓 · 𝑒)
= 𝖽𝖾𝗌𝗎𝗀𝖺𝗋 𝑓 · 𝖽𝖾𝗌𝗎𝗀𝖺𝗋 𝑒
𝖽𝖾𝗌𝗎𝗀𝖺𝗋 {𝖢𝗎𝗋𝗋𝗒} (𝗅𝖺𝗆 𝑒)
= 𝗅𝖺𝗆 (𝖽𝖾𝗌𝗎𝗀𝖺𝗋 𝑒)
′
𝖽𝖾𝗌𝗎𝗀𝖺𝗋 {𝖢𝗁𝗎𝗋𝖼𝗁} (𝗅𝖺𝗆 𝑡 𝑒) = 𝗅𝖺𝗆′ 𝑡 (𝖽𝖾𝗌𝗎𝗀𝖺𝗋 𝑒)
𝖽𝖾𝗌𝗎𝗀𝖺𝗋 {𝖢𝗎𝗋𝗋𝗒} (𝗅𝖾𝗍𝗏𝖺𝗋 𝑒0 𝑒) = 𝗅𝖺𝗆 (𝖽𝖾𝗌𝗎𝗀𝖺𝗋 𝑒) · 𝖽𝖾𝗌𝗎𝗀𝖺𝗋 𝑒0
𝖽𝖾𝗌𝗎𝗀𝖺𝗋 {𝖢𝗁𝗎𝗋𝖼𝗁} (𝗅𝖾𝗍𝗏𝖺𝗋 𝑒0 𝑒) = 𝗅𝖺𝗆′ _ (𝖽𝖾𝗌𝗎𝗀𝖺𝗋 𝑒) · 𝖽𝖾𝗌𝗎𝗀𝖺𝗋 𝑒0
6.2 Example: Normalization
We have formalized a normalization proof of 𝖲𝖳𝖫𝖢 based on the approach presented in [25]. Since our
terms are intrinsically typed, and the simultaneous substitution function is type-preserving, we can avoid
a lot of otherwise lengthy type preservation proofs. We show our formalization in broad strokes here to
highlight the use of the generic syntax library.
Small-step semantics of STLC is defined with a type-preserving reduction relation between typed
terms. 𝗌𝗎𝖻 is crucially used in defining the reduction rule for applying a lambda abstraction.
𝖽𝖺𝗍𝖺 𝖵𝖺𝗅𝗎𝖾 ∶ ∀ {𝛤 𝑡} → 𝖳𝗆 𝛤 𝑡 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
↓𝗅𝖺𝗆 ∶ ∀ {𝛤 𝑡 𝑢} → (𝑒 ∶ 𝖳𝗆 (𝛤 , 𝑢) 𝑡) → 𝖵𝖺𝗅𝗎𝖾 {𝛤 } {𝑢 ▹ 𝑡} (𝗅𝖺𝗆 𝑒)
𝗂𝗇𝖿 𝗂𝗑 𝟣𝟫 _==>_ _==>∗ _
𝖽𝖺𝗍𝖺 _==>_ {𝛤 } ∶ ∀ {𝑡} → 𝖱𝖾𝗅 (𝖳𝗆 𝛤 𝑡) 𝗅𝗓𝖾𝗋𝗈 𝗐𝗁𝖾𝗋𝖾
𝖺𝗉𝗉−𝗅𝖺𝗆 ∶ ∀ {𝑡 𝑢} {𝑒 ∶ 𝖳𝗆 _ 𝑢} → (𝑓 ∶ 𝖳𝗆 _ 𝑡) → 𝖵𝖺𝗅𝗎𝖾 𝑒
𝖺𝗉𝗉−𝖿𝗎𝗇 ∶ ∀ {𝑡 𝑢} {𝑓 𝑓 ′ ∶ 𝖳𝗆 _ (𝑢 ▹ 𝑡)} → 𝑓 ==> 𝑓 ′ → ∀ 𝑒
→ 𝗅𝖺𝗆 𝑓 · 𝑒 ==> 𝗌𝗎𝖻 (𝗋𝖾𝖿𝗅 , 𝑒) 𝑓
→ 𝑓 · 𝑒 ==> 𝑓 ′ · 𝑒
G. Érdi
13
𝖺𝗉𝗉−𝖺𝗋𝗀 ∶ ∀ {𝑡 𝑢} {𝑓 ∶ 𝖳𝗆 _ (𝑢 ▹ 𝑡)} {𝑒 𝑒′ } → 𝖵𝖺𝗅𝗎𝖾 𝑓 → 𝑒 ==> 𝑒′ → 𝑓 · 𝑒 ==> 𝑓 · 𝑒′
_==>∗ _ ∶ ∀ {𝛤 𝑡} → 𝖱𝖾𝗅 (𝖳𝗆 𝛤 𝑡) _
_==>∗ _ = 𝖲𝗍𝖺𝗋 _==>_
The bulk of the proof proceeds exactly as in [25] towards the final goal of
𝗇𝗈𝗋𝗆𝖺𝗅𝗂𝗓𝖺𝗍𝗂𝗈𝗇 ∶ ∀ {𝑡} → (𝑒 ∶ 𝖳𝗆 ∅ 𝑡) → 𝖧𝖺𝗅𝗍𝗌 𝑒
At one point in the course of proving 𝗇𝗈𝗋𝗆𝖺𝗅𝗂𝗓𝖺𝗍𝗂𝗈𝗇, we arrive at a proof obligation for
𝖺𝗉𝗉−𝗌𝗎𝖻 ∶ 𝗅𝖺𝗆 (𝗌𝗎𝖻 (𝗌𝗁𝗂𝖿𝗍 𝜎) 𝖿) · 𝖾 ==>∗ 𝗌𝗎𝖻 (𝜎 , 𝖾) 𝖿
which can be filled in by delegating most of the heavy lifting to our library of proofs; namely the following
three:
𝗌𝗎𝖻−⊢∗ ⊇ ∶ (𝜎 ∶ 𝛤 ⊢∗ 𝛩) (𝜌 ∶ 𝛩 ⊇ 𝛥) (𝑒 ∶ 𝖳𝗆 𝛥 𝗍) → 𝗌𝗎𝖻 (𝜎 ⊢∗ ⊇ 𝜌) 𝑒 ≡ 𝗌𝗎𝖻 𝜎 (𝗋𝖾𝗇 𝜌 𝑒)
𝗌𝗎𝖻−𝗋𝖾𝖿 𝗅 ∶
(𝑒 ∶ 𝖳𝗆 𝛤 𝗍) → 𝗌𝗎𝖻 𝗋𝖾𝖿𝗅 𝑒 ≡ 𝑒
𝗌𝗎𝖻−⊢⊢∗ ∶ (𝜎2 ∶ 𝛤 ⊢∗ 𝛩) (𝜎1 ∶ 𝛩 ⊢∗ 𝛥) (𝑒 ∶ 𝖳𝗆 𝛥 𝗍) → 𝗌𝗎𝖻 (𝜎2 ⊢⊢∗ 𝜎1 ) 𝑒 ≡ 𝗌𝗎𝖻 𝜎2 (𝗌𝗎𝖻 𝜎1 𝑒)
Extending our language with a new boolean type and the corresponding eliminator for conditional
expressions involved changing the 𝖣𝖾𝗌𝖼ription to
𝖽𝖺𝗍𝖺 ‵ 𝖲𝖳𝖫𝖢 ∶ 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
‵
𝖺𝗉𝗉 ‵ 𝗅𝖺𝗆 ‵ 𝗍𝗋𝗎𝖾 ‵ 𝖿 𝖺𝗅𝗌𝖾 ‵ 𝗂𝖿 ∶ ‵ 𝖲𝖳𝖫𝖢
𝖲𝖳𝖫𝖢 ∶ 𝖣𝖾𝗌𝖼
𝖲𝖳𝖫𝖢 = 𝗌𝗀 ‵ 𝖲𝖳𝖫𝖢 𝜆
{ ‵ 𝖺𝗉𝗉 → 𝗇𝗈𝖽𝖾 𝟢 ([] ∷ [] ∷ [])
𝜆 { []
(𝑡1 ∷ 𝑡2 ∷ [])
𝑡0 → 𝑡1 ≡ 𝑡2 ▹ 𝑡0 }
‵
; 𝗅𝖺𝗆 → 𝗇𝗈𝖽𝖾 𝟣 ((𝖻𝗈𝗎𝗇𝖽 ∷ []) ∷ []) 𝜆 { (𝑡 ∷ []) (𝑢 ∷ [])
𝑡0 → 𝑡0 ≡ 𝑡 ▹ 𝑢 }
; ‵ 𝗍𝗋𝗎𝖾 → 𝗇𝗈𝖽𝖾 𝟢 []
𝜆 { []
[]
𝑡0 → 𝑡0 ≡ 𝖻𝗈𝗈𝗅 }
; ‵ 𝖿𝖺𝗅𝗌𝖾 → 𝗇𝗈𝖽𝖾 𝟢 []
𝜆 { []
[]
𝑡0 → 𝑡0 ≡ 𝖻𝗈𝗈𝗅 }
‵
; 𝗂𝖿
→ 𝗇𝗈𝖽𝖾 𝟢 ([] ∷ [] ∷ [] ∷ []) 𝜆 { []
(𝑏 ∷ 𝑡1 ∷ 𝑡2 ∷ []) 𝑡0 → 𝑏 ≡ 𝖻𝗈𝗈𝗅 × 𝑡1 ≡ 𝑡0 × 𝑡2 ≡ 𝑡0 }
}
and adding new axioms to the 𝖵𝖺𝗅𝗎𝖾 and ==> propositions:
𝖽𝖺𝗍𝖺 𝖵𝖺𝗅𝗎𝖾 ∶ ∀ {𝛤 𝑡} → 𝖳𝗆 𝛤 𝑡 → 𝖲𝖾𝗍 𝗐𝗁𝖾𝗋𝖾
↓𝗅𝖺𝗆 ∶ ∀ {𝛤 𝑡 𝑢} → (𝑒 ∶ 𝖳𝗆 (𝛤 , 𝑢) 𝑡) → 𝖵𝖺𝗅𝗎𝖾 {𝛤 } {𝑢 ▹ 𝑡} (𝗅𝖺𝗆 𝑒)
↓𝗍𝗋𝗎𝖾 ∶ ∀ {𝛤 } → 𝖵𝖺𝗅𝗎𝖾 {𝛤 } 𝗍𝗋𝗎𝖾
↓𝖿𝖺𝗅𝗌𝖾 ∶ ∀ {𝛤 } → 𝖵𝖺𝗅𝗎𝖾 {𝛤 } 𝖿 𝖺𝗅𝗌𝖾
𝖽𝖺𝗍𝖺 _==>_ {𝛤 } ∶ ∀ {𝑡} → 𝖱𝖾𝗅 (𝖳𝗆 𝛤 𝑡) 𝗅𝗓𝖾𝗋𝗈 𝗐𝗁𝖾𝗋𝖾
𝖺𝗉𝗉−𝗅𝖺𝗆 ∶ ∀ {𝑡 𝑢} {𝑒 ∶ 𝖳𝗆 _ 𝑢} → (𝑓 ∶ 𝖳𝗆 _ 𝑡) → 𝖵𝖺𝗅𝗎𝖾 𝑒
→ 𝗅𝖺𝗆 𝑓 · 𝑒 ==> 𝗌𝗎𝖻 (𝗋𝖾𝖿𝗅 , 𝑒) 𝑓
𝖺𝗉𝗉−𝖿𝗎𝗇 ∶ ∀ {𝑡 𝑢} {𝑓 𝑓 ′ ∶ 𝖳𝗆 _ (𝑢 ▹ 𝑡)} → 𝑓 ==> 𝑓 ′ → ∀ 𝑒
→ 𝑓 · 𝑒 ==> 𝑓 ′ · 𝑒
𝖺𝗉𝗉−𝖺𝗋𝗀 ∶ ∀ {𝑡 𝑢} {𝑓 ∶ 𝖳𝗆 _ (𝑢 ▹ 𝑡)} {𝑒 𝑒′ } → 𝖵𝖺𝗅𝗎𝖾 𝑓 → 𝑒 ==> 𝑒′ → 𝑓 · 𝑒 ==> 𝑓 · 𝑒′
𝗂𝖿−𝖼𝗈𝗇𝖽 ∶ ∀ {𝑡} {𝑏 𝑏′ } → 𝑏 ==> 𝑏′ → (𝑡ℎ𝑛 𝑒𝑙𝑠 ∶ 𝖳𝗆 _ 𝑡)
→ 𝗂𝖿 𝑏 𝑡ℎ𝑛 𝑒𝑙𝑠 ==> 𝗂𝖿 𝑏′ 𝑡ℎ𝑛 𝑒𝑙𝑠
𝗂𝖿−𝗍𝗋𝗎𝖾 ∶ ∀ {𝑡} → (𝑡ℎ𝑛 𝑒𝑙𝑠 ∶ 𝖳𝗆 _ 𝑡)
→ 𝗂𝖿 𝗍𝗋𝗎𝖾 𝑡ℎ𝑛 𝑒𝑙𝑠 ==> 𝑡ℎ𝑛
𝗂𝖿−𝖿 𝖺𝗅𝗌𝖾 ∶ ∀ {𝑡} → (𝑡ℎ𝑛 𝑒𝑙𝑠 ∶ 𝖳𝗆 _ 𝑡)
→ 𝗂𝖿 𝖿𝖺𝗅𝗌𝖾 𝑡ℎ𝑛 𝑒𝑙𝑠 ==> 𝑒𝑙𝑠
but, upholding the promise of a generic library for manipulating terms, none of the proofs involving
syntactic transformations, like 𝖺𝗉𝗉−𝗌𝗎𝖻 or its building blocks, needed any updating.
14
Generic description of well-scoped, well-typed syntaxes
7 Conclusions and future work
In this paper, we have described the datatype 𝖣𝖾𝗌𝖼 that can be used to give a machine-readable definition of a statically typed language with variable-binding structures. We have implemented a library in
the dependently typed programming language Agda that provides untyped, well-scoped, and well-typed
representations of terms; type-preserving renaming and substitution; and proofs of their well-behaving.
We have demonstrated the use of the library by formalizing a normalization proof of the simply-typed
lambda calculus; and then observed the ease with which the object language can be changed or extended
without labouriously rewriting the syntactic transformations and proofs.
We have also recovered object language-specific functions like let-desugaring and (in the full version
of the library) typechecking that are implemented by pattern matching on pattern synonyms that exhibit
the correspondence between the description-derived representation and the traditional standard approach.
One possible direction for future work is support for languages with multiple binding namespaces
by multiplexing the context. These contexts can be orthogonal, like the truth and validity hypotheses in
modal logic à la [24], where variable occurrences would be indexed by which namespace they point into;
this index can then be used by the well-typedness constraint to enforce restrictions on occurrence. Or the
contexts could correspond to multiple sorts and be nested, for polymorphic languages like System F[14]:
terms have types have kinds, and the types of term variables available in a given context can refer to type
variables. This requires substitution to be able to shift on multiple axes at the same time, as discussed in
[15].
Since well-typedness is represented as an arbitrary proposition over subterm types, bound variable
types, and the type of the term itself, it can be any relation between them, not just a conjunction of
equalities. We have not explored this avenue yet, but it seems applicable to supporting languages with
subtyping[9], or those that use unification in their type system[11] (and not just in implementing a typechecker). However, treating renamings and substitutions as exact type-preserving may be inadequate in
these settings.
References
[1] Guillaume Allais, Robert Atkey, James Chapman, Conor McBride & James McKinna (2018): A Scope-andType Safe Universe of Syntaxes with Binding, Their Semantics and Proofs. Available at https://github.
com/gallais/generic-syntax.
[2] Guillaume Allais, James Chapman, Conor McBride & James McKinna (2017): Type-and-scope safe programs
and their proofs. In: Proceedings of the 6th ACM SIGPLAN Conference on Certified Programs and Proofs,
ACM, pp. 195–207.
[3] Thorsten Altenkirch & Conor McBride (2003): Generic programming within dependently typed programming. In: Generic Programming, Springer, pp. 1–20.
[4] Thorsten Altenkirch & Bernhard Reus (1999): Monadic presentations of lambda terms using generalized
inductive types. In: International Workshop on Computer Science Logic, Springer, pp. 453–468.
[5] Stevan Andjelkovic (2011): A family of universes for generic programming.
[6] Bruno Barras, Samuel Boutin, Cristina Cornes, Judicaël Courant, Jean-Christophe Filliatre, Eduardo
Gimenez, Hugo Herbelin, Gerard Huet, Cesar Munoz, Chetan Murthy et al. (1997): The Coq proof assistant reference manual. Ph.D. thesis, Inria.
[7] Nick Benton, Chung-Kil Hur, Andrew J Kennedy & Conor McBride (2012): Strongly typed term representations in Coq. Journal of automated reasoning 49(2), pp. 141–159.
G. Érdi
15
[8] Richard S Bird & Ross Paterson (1999): De Bruijn notation as a nested datatype. Journal of functional
programming 9(1), pp. 77–91.
[9] Luca Cardelli, Simone Martini, John C Mitchell & Andre Scedrov (1994): An extension of system F with
subtyping. Information and Computation 109(1-2), pp. 4–56.
[10] James Maitland Chapman (2009): Type checking and normalisation. Ph.D. thesis, University of Nottingham.
[11] Olaf Chitil (2001): Compositional explanation of types and algorithmic debugging of type errors. In: ACM
SIGPLAN Notices, 36, ACM, pp. 193–204.
[12] Adam Chlipala (2008): Parametric higher-order abstract syntax for mechanized semantics. In: ACM Sigplan
Notices, 43, ACM, pp. 143–156.
[13] Alonzo Church (1940): A formulation of the simple theory of types. The journal of symbolic logic 5(2), pp.
56–68.
[14] Jean-Yves Girard (1971): Une Extension De L’Interpretation De Gödel a L’Analyse, Et Son Application a
L’Elimination Des Coupures Dans L’Analyse Et La Theorie Des Types. In: Studies in Logic and the Foundations of Mathematics, 63, Elsevier, pp. 63–92.
[15] Jonas Kaiser, Steven Schäfer & Kathrin Stark (2017): Autosubst 2: Towards Reasoning with Multi-Sorted
de Bruijn Terms and Vector Substitutions. In: Proceedings of the Workshop on Logical Frameworks and
Meta-Languages: Theory and Practice, ACM, pp. 10–14.
[16] András Kovács (2017): Example of complete STLC substitution calculus. https://gist.github.com/
AndrasKovacs/bd6a6333e4eecd7acb0eb9d98f7e0cb8.
[17] Daniel R Licata & Robert Harper (2009): A universe of binding and computation. In: ACM Sigplan Notices,
44, ACM, pp. 123–134.
[18] Per Martin-Löf & Giovanni Sambin (1984): Intuitionistic type theory. 9, Bibliopolis Napoli.
[19] Conor McBride (2005): Type-preserving renaming and substitution.
[20] Conor McBride (2013): Dependently Typed Metaprogramming (in Agda).
[21] Bengt Nordström, Kent Petersson & Jan M Smith (1990): Programming in Martin-Löf’s type theory. 200,
Oxford University Press Oxford.
[22] Ulf Norell (2008): Dependently typed programming in Agda. In: International School on Advanced Functional Programming, Springer, pp. 230–266.
[23] Simon Peyton Jones, Dimitrios Vytiniotis, Stephanie Weirich & Geoffrey Washburn (2006): Simple
unification-based type inference for GADTs. In: ACM SIGPLAN Notices, 41, ACM, pp. 50–61.
[24] Frank Pfenning & Rowan Davies (2001): A judgmental reconstruction of modal logic. Mathematical structures in computer science 11(4), pp. 511–540.
[25] Benjamin C. Pierce, Arthur Azevedo de Amorim, Chris Casinghino, Marco Gaboardi, Michael Greenberg,
Cǎtǎlin Hriţcu, Vilhelm Sjöberg & Brent Yorgey (2017): Software Foundations. Electronic textbook. Version
5.0. http://www.cis.upenn.edu/~bcpierce/sf.
[26] The Agda Team (2016): Release notes for Agda 2 version 2.3.2. http://wiki.portal.chalmers.
se/agda/pmwiki.php?n=Main.Version-2-3-2. Accessed: 2017-12-30.
[27] Mark Tullsen (2000): First Class Patterns? In: International Symposium on Practical Aspects of Declarative
Languages, Springer, pp. 1–15.
| 6cs.PL
|
arXiv:1704.00513v1 [cs.DC] 3 Apr 2017
Optimizing Communication by Compression for
Multi-GPU
Scalable Breadth-First Searches
This Master thesis has been carried out by Julian Romera
at the
Ruprecht-Karls-Universität Heidelberg
under the supervision of
JProf Dr Holger Fröning
Ruprecht-Karls-Universität Heidelberg
Institut für Technische Informatik
Master thesis
submitted by
Julian Romera
born in Madrid, Spain
2016
Optimizing Communication by Compression for Multi-GPU
Scalable Breadth-First Search:
Die Breitensuche, auch Breadth First Search (BFS) genannt, ist ein fundamentaler Bestandteil vieler Graph Operationen wie spanning trees, shortest path oder
auch betweenness centrality. Ihre Bedeutung steigt stets an, da sie immer mehr
zu einem zentralen Aspekt vieler populärer Datenstrukturen wird, welche intern
durch Graph Strukturen repräsentiert werden. Die Parallelisierung des BFS Algorithmus erfordert die Verteilung der Daten auf mehrere Prozessoren und wie die
Forschung zeigt, ist die Leistungsfähigkeit durch das Netzwerk begrenzt [31]. Aus
diesem Grund ist es wichtig Optimierungen auf die Kommunikation zu konzentrieren, was schluessendlich die Leistung dieses wichtigen Algorithmus erhöht. In
dieser Arbeit wird ein alternativer Kompressionsalgorithmus vorgestellt. Er unterscheidet sich von existierenden Methoden dadurch, dass er die Charakteristika
der Daten berücksichtigt, was die Kompression verbessert. Ein weiterer Test zeigt
zudem, wie dieser BFS Algorithmus von traditionellen instruktionsbasierten Optimierungen in einer verteilten Umgebung profitiert. Als letzten Schritt werden
aktuelle Techniken aus dem Hochleistungsrechnen und andere Arbeiten in diesem
Bereich betrachtet.
Optimizing Communication by Compression for Multi-GPU
Scalable Breadth-First Search:
The Breadth First Search (BFS) algorithm is the foundation and building block
of many higher graph-based operations such as spanning trees, shortest paths and
betweenness centrality. The importance of this algorithm increases each day due
to it is a key requirement for many data structures which are becoming popular
nowadays. These data structures turn out to be internally graph structures.
When the BFS algorithm is parallelized and the data is distributed into several
processors, some research shows a performance limitation introduced by the
interconnection network [31]. Hence, improvements on the area of communications
may benefit the global performance in this key algorithm.
In this work it is presented an alternative compression mechanism. It differs with
current existing methods in that it is aware of characteristics of the data which
may benefit the compression.
Apart from this, we will perform a other test to see how this algorithm (in a distributed scenario) benefits from traditional instruction-based optimizations. Last,
we will review the current supercomputing techniques and the related work being
done in the area.
Keywords: Breath-First Search, Graph500, Compression.
Dedicated to Catren
4
Contents
1 Introduction
1.1 Introduction . .
1.2 Motivation . . .
1.3 Contributions .
1.4 Structure of the
.
.
.
.
8
8
8
9
10
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
11
11
14
17
18
18
19
21
22
23
24
24
25
26
26
27
27
30
30
31
3 Related work
3.1 Optimizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Other implementations . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
32
32
33
34
. . . .
. . . .
. . . .
thesis
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2 Background
2.1 Milestones in supercomputing . . . . . . . . . . . . . . .
2.2 Global architectural concepts . . . . . . . . . . . . . . .
2.3 Clusters and High Performance Computing . . . . . . . .
2.3.1 Hardware architectures . . . . . . . . . . . . . .
2.3.2 Interconnection networks in HPC . . . . . . . . .
2.3.3 The classification of supercomputers used in HPC
2.3.4 The popularization of the multicore architectures
2.3.5 Energy efficiency . . . . . . . . . . . . . . . . . .
2.4 General Purpose Graphic Proccessor Units . . . . . . . .
2.4.1 GPGPU architectures . . . . . . . . . . . . . . .
2.4.2 Regular and non-regular problems . . . . . . . . .
2.5 Message Passing Interface . . . . . . . . . . . . . . . . .
2.6 Graph computations . . . . . . . . . . . . . . . . . . . .
2.6.1 Graphs . . . . . . . . . . . . . . . . . . . . . . . .
2.6.2 Graphs partitioning . . . . . . . . . . . . . . . . .
2.6.3 Graph 500 challenge . . . . . . . . . . . . . . . .
2.7 Input data . . . . . . . . . . . . . . . . . . . . . . . . . .
2.7.1 Synthetic data . . . . . . . . . . . . . . . . . . . .
2.7.2 Real world graphs (datasets) . . . . . . . . . . . .
5
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Contents
6
4 Problem Analysis
4.1 Our initial implementation . . . . . . . . . . . . . . . . . .
4.1.1 Optimizations table in “Baseline” implementation
4.1.2 Our “Baseline” algorithm . . . . . . . . . . . . . .
4.1.3 General communication algorithm . . . . . . . . . .
4.1.4 Data and communications . . . . . . . . . . . . . .
4.2 Instrumentation . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Instrumented zones . . . . . . . . . . . . . . . . . .
4.3 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 Communication overhead . . . . . . . . . . . . . . .
4.3.2 Instruction overhead . . . . . . . . . . . . . . . . .
5 Optimizing data movement by compression
5.1 Concepts . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Compression algorithms (codecs or schemes) . . . . .
5.3 Compression libraries . . . . . . . . . . . . . . . . . .
5.3.1 Lemire et al. . . . . . . . . . . . . . . . . . . .
5.3.2 Turbo-PFOR . . . . . . . . . . . . . . . . . .
5.3.3 Alenka CUDA . . . . . . . . . . . . . . . . . .
5.4 Integration of the compression . . . . . . . . . . . . .
5.4.1 Experimental performance comparison . . . .
5.4.2 Observations about experimental performance
5.4.3 Usage of compression thresholds . . . . . . . .
5.4.4 Communication before and after compression
.
.
.
.
.
.
.
.
.
.
.
6 Optimizing instruction overhead
6.1 Scalar optimizations . . . . . . . . . . . . . . . . . . .
6.1.1 Strength reduction . . . . . . . . . . . . . . . .
6.2 Vectorization . . . . . . . . . . . . . . . . . . . . . . .
6.2.1 Compiler optimizations . . . . . . . . . . . . . .
6.3 Thread parallelism . . . . . . . . . . . . . . . . . . . .
6.3.1 Usage of fork-join model when suitable . . . . .
6.3.2 Optimal thread scheduling . . . . . . . . . . . .
6.3.3 Thread contention prevention . . . . . . . . . .
6.3.4 NUMA control / NUMA aware . . . . . . . . .
6.4 Memory access . . . . . . . . . . . . . . . . . . . . . .
6.4.1 Locality of data access . . . . . . . . . . . . . .
6.4.2 Merge subsequent loops . . . . . . . . . . . . .
6.4.3 Variable caching in the outer loop . . . . . . . .
6.5 Communication patterns . . . . . . . . . . . . . . . . .
6.5.1 Data transfer grouping . . . . . . . . . . . . . .
6.5.2 Reduce communication overhead by overlapping
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
36
36
36
37
37
38
39
39
40
40
41
.
.
.
.
.
.
.
.
.
.
.
42
42
49
53
53
55
56
56
57
61
62
63
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
65
65
65
69
73
75
76
76
77
77
79
79
79
79
80
80
80
Contents
7
6.6
81
Summary of implemented optimizations . . . . . . . . . . . . . . . . .
7 Final Results
7.1 Experiment platforms . . . . . . . . . . . .
7.2 Results . . . . . . . . . . . . . . . . . . . .
7.2.1 Scalability analysis . . . . . . . . .
7.2.2 Overhead of the added compression
7.2.3 Instruction overhead analysis . . .
7.2.4 Compression analysis . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
84
84
84
88
91
93
93
8 Conclusions
95
9 Future work
96
References
99
Acknowledgments
105
Deposition
106
Chapter 1
Introduction
1.1
Introduction
The study of large Graphs datasets has been a very popular subject for some
years. Fields such as Electronic engineering, Bioscience, the World Wide Web, Social networks, Data mining, and the recent expansion of “Big Data” result in larger
amounts of data each time, which need to be explored by the current and state-ofthe-art graph algorithms. One result of this recent expansion of large-sized graph
is the establishment of benchmarks that engage organisations in the development of
new techniques and algorithms that solve this new challenge.
Graph 500 1 is a new benchmark created in 2010 which serves this purpose. It uses
the Breadth-first Search (BFS) algorithm as a foundation. This benchmark contrasts
with other well-known and more mature tests like LINPACK (used in the Top 500 2
Challenge), in that the latter executes linear equations in a computation-intensive
fashion. In contrast, the former makes use of graph based data-intensive operations
to traverse the graph.
In this work we focus on possible optimizations of the Breadth-first Search (BFS)
algorithm.
1.2
Motivation
As stated before, as the growth of data volumes increases at a very high rate and
the structures containing this data use an internal graph representation, the research
in graph algorithms also increases.
1
http://www.graph500.org
2
http://www.top500.org/
8
Chapter 1: Introduction
9
Some of the current research done on the area of the BFS algorithms is framed
under the Graph 500 challenge.
The work done on improving our previous implementation of the garph500 benchmarch (“Baseline”), as well as the research of other authors in this area, shows
that there is a severe impact in the overall performance of the distributed version
of the algorithm due to the latency introduced by the the data movement between
processors.
1.3
Contributions
We believe that the conclusions presented in this work may benefit a broad variety of Breadth-first Search implementations. Compressing the data movement may
alleviate the main bottleneck of the distributed version of the algorithm: the communication. In this work we make use of a Compressed Sparse Row (CSR) matrix,
2D data partitioning and Sparse vector multiplications (SpMVM). By the analysis of
the transmitted data (a SpMV vector represented as an integer sequence) we have
been able to successfully achieve over 90% in terms of data transfer reduction and
over an 80% of communication time reduction. All this is achieved thanks to specific
numerical proprieties the the data of the analyzed graph (Section 2.6).
Furthermore, we will also show how different techniques to reduce the instruction
overhead may improve the performance, in terms of Traversed Edges per Second
TEPS (Section 2.6.3) and time, of a graph500 implementation. Finally, we will
discuss further optimizations. Here follows the main contributions of this work:
1. As mayor goal, we have used a Bit Packing compression scheme (Section 5.2
with delta compression on top, optimized for the SIMD instruction set of Intel™ x86 architecture. The reasoning and criteria of the used algorithm are
discussed in further detail on Section 5.4. Also, a scalability study before and
after integrating the compression has been performed in sections 5 and 7.
2. In addition to the added compression library (which contains several compression codecs), two more libraries have been integrated (partially). With this we
enable our graph500 application to test more codecs for other different graphs
and problem sizes. This is described in further detail in section 5.3. The newly
added implementations are not fully operative yet. These new packages offer
benefits such as new compression codecs, different input vector sizes (which
would allow bigger problem sizes and the avoidance of the performance penalty
imposed by the conversion of the full vector, required by the current library),
and a GPU-based implementation. This converts our graph500 implementation
on a modular compression test-bench.
Chapter 1: Introduction
10
3. As part of this work, the instruction overhead has been reduced. With this, we
have tested the effects of some common instruction transformation techniques
at different scales in our BFS implementation. The used techniques have been
described more in detail on Section 6 (Optimizing instruction overhead)
4. An external instrumentation (Score-P) has been added to help with the detection of the bottlenecks, optimization of our code and validation of the final
results. The reasoning for the profiler selection criteria is discussed in Section
4.2. As a summary, the selected profiler provides a fine-grain way to gather a
high level of details with low overhead 3 .
5. Some other tasks have been performed to our Graph 500 implementation.
Briefly, these have been:
– The application’s built process has been improved to reduce the building
time. Also, a pre-processing of the compiled code allows a more fine-grain
tuning. For this purpose we have chosen Maketools 4 . As result, we detect
the maximum capabilities of the target system at pre-compile time.
– We have created scripts to automate the visualization and interpretation
of the results. For this we combine c/c++ with R.
1.4
Structure of the thesis
The thesis is formed by 4 main parts (i) a background section about the involved
elements of the thesis. (ii) An analysis of our problem with two possible solutions:
compression and instruction overhead reduction. (iii) Finally in the last part, we
discuss the result and see how much do they mach with the purposed solutions for
our problem.
3
http://www.vi-hps.org/projects/score-p/
4
https://www.gnu.org/software/make/
Chapter 2
Background
As the background of optimizations made to a Graph 500 application, a short
timeline of the history of Supercomputing will be listed to place the cited concepts on
the time. Some of the core optimizations technologies, used in both the compression
and the main Graph 500 implementation will be reviewed.
Regarding Supercomputing, it will be provided a background on some of its core
concepts: High Performance Computing (HPC), General Processing Graphic Processing Units (GPGPU) and Message Passing Interface (MPI).
Last, concepts about graphs (in a Linear Algebra context), and the datasets used
often on graphs literature will also be described.
2.1
Milestones in supercomputing
As an introduction to Supercomputing, some of the personal names and milestones
of its history1 are listed below. Many of these will be referenced along in this work.
• 1930’s - Fist theoretical basis of computing.
– 1936: Alan Turing develops the notion of a “Universal machine”
through his paper “On Computable Numbers” [50]. It is capable of computing anything that is computable.
• 1940’s - The foundation of computing.
– 1945: John von Neumann introduces the concept of a stored program in
the draft report on the EDVAC design. This way he creates the design of
a sequential computer [53] (known as the von Neumann architecture).
– 1947: Bell Laboratories invents the Transistor.
1
http://www.computerhistory.org
11
Chapter 2: Background
12
– 1948: Claude Shannon writes “A Mathematical Theory of Communication” (Sec. 5.1 ) [45]
• 1950’s - The first commercial general-purpose computers are born.
– 1951: UNIVAC-1 is launched. This is the very first commercial generalpurpose computer. Even though it is general-purpose, it is focus on home
use.
– 1953: IBM announces Model 650. Stores the data on rotating tape. This
is the first mass-produced computer.
– 1953: IBM announces Model 701. The first Mainframe 2
– 1956: UNIVAC announces an UNIVAC built with transistors.
• 1960’s - During this period the IBM Corporation dominated the general purpose early computing industry. This was the era of the Mainframes. There are
new additions to the foundation of computing.
– 1965: Gordon Moore coins the Moore’s Law. This predicts the number
of in-chip transistors throughout time.
– 1966: Michael Flynn proposes a classification of computer architectures.
This is known as Flynn’s Taxonomy. It is still used nowadays.
– 1968: CDC 7600 is launched by Seymour Cray. It is considered by
many the first true Supercomputer.
• 1970’s - First Supercomputers enter in scene. They increase their performance
through Vector Processors Units among other methods.
– 1971: Intel Corporation launches its first chip: Intel 4004
– 1972: Cray Research Inc. is created by Seymour Cray. He is consider as
the father of Supercomputing.
– 1975: Cray Research Inc completes the development of Cray 1 - The
first Vector-processor Supercomputer.
– 1978: Intel Corporation introduces the first 16-bit processor, the 8086.
• 1980’s - Many technological improvements. First IBM/Intel personal computers (PC). First Connection Machine Supercomputers.
2
These computers, also referred as “Big Iron”, receive their name for the shape of their cabinet
structure. They are often used in statistical computation, banking transactions, etc.
Chapter 2: Background
13
– 1982: Cray Research Inc introduces the Cray X-MP Supercomputer.
This version uses Shared Memory and a vector-processor. It is a ’cleanedup’ version of the CRAY-1. Successive versions of this Supercomputer
increase the number of CPUs, raise the clock frequency and expand the
instruction size from 24 bits to 32.
– 1983: Thinking Machines introduces CM-1 , the first Connection Machine (CM). It is based upon the SIMD classification.
– 1984: IBM introduces the IBM PC/AT based on the chip Intel 80286.
The chip works at 16 bits.
– 1985: Intel Corporation introduces the 80386 chip with 32-bit processing
and on-chip memory management.
• 1990’s - Next step in the Supercomputer era. First Massively Parallel architectures. The home computer industry is based on ’clones’ of the original IBM
PC.
– 1991: Thinking Machines introduces CM-5, a new version of their Connection Machine running on an RISC SPARC, and replacing the connection
network of the previous CM-2. The CM-5 changes, this way, to a MIMD
design. This is the first NUMA architecture.
– 1992: Touchstone Delta Is an experiment carried out in Caltech with 64x
Intel 8086 microprocessors. It opens a door to a new era of parallelism.
This is later referred as “The attack of the killer micros”. Here, a large
number of Commercial off-the-shelf (COTS) microprocessors, invaded a
world dominated by “Big Iron” Vector computers.
– 1993: Intel Corporation releases “Pentium” chip. Personal computers
continue to grow.
– 1993: top500.org ranking is created. It uses a linear algebra LINPACK
benchmark.
– 1997: The Intel ASCI Red Supercomputer was developed based on the
Touchstone Delta experiment. This Massively parallel Supercomputer was
the fastest in the world until the early 2010’s.
– 1999: IBM PowerPC 440 microprocessor is launched. This 32-bit RISC
high performance core will be the main processor of many future HPP
architectures like the Blue Gene/L or Cray XT3
• 2000’s - The fastest Supercomputers are Massively Parallel architectures.
– 2001: General Purpose GPU processing (GPGPU) begins its development by the advent of the programmable shaders and floating point
units on graphics processors.
Chapter 2: Background
14
– 2004: Intel Corporation announces the cancelation of two of their processors [36]. This is known as the End of Frequency Scaling. The new
speedups are based upon parallel techniques developed in the previous Supercomputing era. With the Multicore the personal computer industry
also go parallel.
• 2010’s - The only Supercomputer architectures in Top500 3 are Massively Parallel: MPPs and Clusters.
– 2010: graph500.org challenge is created. It uses graph-based data insensitive computation over a Breadth-first Search (BFS) algorithm.
– 2012: green.graph500.org challenge is created. The current energy
awareness leads to this energy focused variation of the Graph 500 Challenge. This new benchmark is also based upon the BFS algorithm but
reflexes the power efficiency in the results.
2.2
Global architectural concepts
Continuing with the background about High Performance Computing (HPC) and
related technologies like GPGPUs, we will introduce in this section some concepts
which will be referenced in the document. This concepts encompass the internal
methods used in the compression algorithm (SIMD / Vectorization [28]), other optimizations made to the selected compression algorithm (Super-Scalar optimizations,
improvements on the Pipelinig [68]). These concept will also be referenced in this
Background section for a better understanding of HPC, GPGPUs and Supercomputing in general.
The Von Neumann design
It was in 1945 when the physicist and mathematician John von Neumann and
others, designed the first programmable architecture. It was called Von Neumann
architecture and had an intrinsically serial design. In this, there is only one processor
executing a series of instructions and the execution flow occur in the same order as
it appears in the original program [53].
Moore’s forecast
Gordon Moore, in 1965, created a forecast for the number of in-chip transistors
on a silicon microprocessor [35]. It was able to predicted the number of transistors on
3
http://www.top500.org/statistics/overtime/
Chapter 2: Background
15
a chip troughtout the time: “In-chip transistror number would double every 2 years”
(Figure 2.1).
This prediction was true for a long time and the clock frequency of microprocessors
was increasing due to deeper Pipelines and more transistors. In the early 2000’s the
micro processors reached a temperature limit [35]. Because of this, clock frequency
has not been since 2005 (Sec. 2.3.5).
This, has ultimately led to Multi-core technology (a chip with multiple cores)
allowing a similar speed. As a result, nowadays software applications are required to
have parallelism to benefit from this [34].
Moore's Law −
"The number of transistors incorporated
in a chip will approximately double every 24 months"
Intel
Itanium
2000Mtr
8.0e+08
1.2e+09
1.6e+09
AMD 6100
1800Mtr
NVIDIA
G80GPU
681Mtr
0.0e+00
4.0e+08
Number of transistors (tr)
2.0e+09
Number of transistors (tr)
Fairchil
9tr
Intel
4004
2300tr
Motorola Intel
68020 80486
250.000tr 1.2Mtr
Intel
Pentium
45Mtr
Motorola
68060
2Mtr
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Year
Figure 2.1: Moore’s Law [62]
Pipeline instruction parallelism
The Instruction Pipelining technique uses instruction-level parallelism inside a
single processor. The basic instruction cycle is chopped into a series which is called
Chapter 2: Background
16
a Pipeline. This allows a quicker execution throughput. Put simply, instead of processing each instruction sequentially (completing the instruction before the next one),
each instruction is split into sub steps. These sub steps can be executed in parallel.
Superscalar architectures
A Superscalar processor is an architecture in which the processor is able to issue
multiple instructions in a single clock. This is achieved using redundant facilities to
execute an instruction.
Each of the superscalar replicated execution units is a resource within a single
CPU, such units may be an Arithmetic Logic Unit, a Vector Processing Unit, bitwise
operators, or sets of CPU registers (Intel™ Multithreading). This excludes the case
of the replicated unit being a separate processing unit (for example other core in a
Multi-core processor).
Note that, while a superscalar design is typically also pipelined, Pipelining and
Superscalar architectures are different concepts.
Flynn’s taxonomy
In 1996, Michael Flynn creates the earliest classification systems for parallel and
sequential computers [20]. The categories in the taxonomy may be seen in Table 2.1
Single Data Stream
Multiple Data Streams
Single Instruction Stream
Multiple Instruction Streams
SISD
SIMD
MISD
MIMD
Table 2.1: Flynn’s Taxonomy
He classified systems no matter if they were functioning using a single set or
multiple sets of instructions. Also to whether (or not) the instructions were using a
single set or multiple sets of data. Flynn’s Taxonomy is still in use today.
According to the authors Patterson and Hennessy, “Some machines are hybrids
of these categories, but this classic model has survived because it is simple, easy to
understand, and gives a good approximation” [26].
Two terms in Flynn’s Taxonomy relevant to this work: SIMD and MIMD are
described below.
1. Simple Instruction Multiple Data (SIMD)
In a SIMD architecture, the parallelism is in the data. There is only one program
counter, and it moves through a set of instructions. One of these instructions
may operate on multiple data elements in parallel, at a time.
Chapter 2: Background
17
Examples in this category are a modern GPGPU, or Intel SSE and AltiVec
Instruction-sets.
2. Multiple Instruction Multiple Data (MIMD)
MIMD are autonomous processors simultaneously executing different instructions on different data.
Distributed systems are generally categorized as MIMD architectures. In these
systems each processor has an individual memory. Also, each processor has no
knowledge about the memory in other processors. In order to share data, this
must be passed as a message from one processor to another.
Examples of this category are a Multicore-superscalar processor, a modern MPP,
or a Cluster.
As of 2015, all the Top500 Supercomputers are within this category.
Types of parallelism according to their level of synchronization
Other way to classify the parallelism is by the concept of Granularity. In this
context we relate the amount of computation with the amount of communication due
to synchronization. This classification would be as follows:
• fine-grained parallelism In this class of granularity the tasks have small
computation and high amount of synchronization. As a downside of this level
of synchronization the overhead due to communication is bigger.
• coarse-grained parallelism In this type of granularity, the tasks are bigger,
with less communication between them. As a result the overhead due to communication is smaller.
In order to attain a good parallel performance, the best balance between load
and communication overhead needs to be found. If the granularity is too fine, the
application will suffer from the overhead due to communication. On the other side,
if the granularity is too coarse, the performance can suffer from load imbalance.
2.3
Clusters and High Performance Computing
High Performance Computing (HPC), or supercomputing, is a kind of computation focused in high consuming computational operations where specific parallel
paradigms and techniques are required. Some examples of these are predictive models, scientific computation, or simulations. The types of Supercomputers (and their
features), and how they evolved in history until today, are defined in this section.
Chapter 2: Background
18
2.3.1
Hardware architectures
Based upon the taxonomy purposed by Flynn we show a general classification of
the HPC architectures. This can be see in Figure 2.2.
Parallel architectures
MIMD
SIMD
UMA
Bus
Multicomputers
Multiprocessors
Vector
Array
COMA
Switched
ccNUMA
NUMA
ncNUMA
Shared memory
MPP
Cluster
Grid
Cube
Message passing
Figure 2.2: Architectural classification in HPC [38].
2.3.2
Interconnection networks in HPC
As this topic is very broad we will only review a concept related to networking,
GPGPU, hardware architectures and in an extend, to this work.
The Remote Direct Memory Access (RDMA) is a technique created to solve a
problem in networking applications. First, the problem is the performance penalty
imposed when a system (sometimes, a processor unit within the main system) accesses
the memory of another system (sometimes the original one) involving the Operating
system (e.g. CPU, caches, memory, context switches) of the first one. The solution
for this problem consists of the access via the network device to the peer system.
Some implementations of this technique use hardware specific vendors as Infiniband™ , and others give support to already existing technologies as Ethernet, such
as RoCE and iWarp.
The specific BFS implementation based on GPGPU, in which multiple GPUS
might be allocated within the same system, would suffer a high penalty without this
technique. In our tests, the RDMA support is enabled and managed by the Message
Passing Interface MPI implementation (Section 2.5)
Chapter 2: Background
2.3.3
19
The classification of supercomputers used in HPC
The first computers in the market were mainly built for governmental, scientific
or military purposes. It could be stated that computing started with Supercomputing. Here we list a classification of supercomputers according to similar design
characteristics and sorted chronologically.
Vector supercomputers
These were the first to appear [62]. In these architectures, the parallelism was
achieved with the help of Vectorizing Compilers. The applications needed to be
adapted for their use with these compilers.
In this scenario some loops could sometimes be vectorized through annotations in
the language. Other loops needed to be arranged in order to remove the dependencies.
The main benefits of these Supercomputers were for programs that would involve
a lot of array processing.
They enter in the Flynn’s Taxonomy (Table 2.1 page 16) under the SIMDs category and unlike other SIMD architectures like the (described in the next section)
Connection Machines, the modern GPGPUs or the Intel SSE / AltiVec instruction
sets these used larger and variable vector sizes.
Data-parallel architectures
This was a brief moment in time. In this category enter the initial Connection
Machines. These architectures also enter in the Flynn’s SIMD category.
Here, the data-parallel language will distribute the work. The algorithms were
very clever and involved data-parallel operations. Examples of those operations were
segmented scans, Sparse Matrix operations or reductions.
Some parallel languages were C*, CMF (Connection Machine Fortran) or *Lisp
[62].
Shared Memory
In these systems, several processors share a common memory space. The programming model in these systems is based on Threads and Locks.
The algorithms were some variation of Parallel Random Access Machine (PRAM)
algorithm, or hardware-based algorithms like cache coherence implementations. This
latter were supported by hundreds of technical papers. In this classification enter the
sub-taxonomy listed in Figure 2.2. Two of this types will be described below.
1. in a UMA model all the processors share the physical memory uniformly and
access time to a memory location is independent of which processor makes the
request.
Chapter 2: Background
20
2. in a NUMA model the memory access time depends on the memory location
relative to the processor: the local memory of a processor will be accessed faster
than its remote memory (memory local to another processor or memory shared
between other processors). This type of architectures are common in cluster
programming and have relevance in this work due to the new advances made
over the distributed BFS algorithm for big cluster supercomputers. A diagram
of the local and remote banks of memory may be seen in figure 2.3. This
architecture will be referenced in section 3.
processor core & L2 cache
RAM
L3
L3
RAM
RAM
L3
L3
RAM
shared L3 cache
8 core xeon E5 4640
Figure 2.3: Detail of a NUMA architecture. Intel™ SandyBridge-EP system [61].
Multicomputers
The Multicomputers replaced the Vector Supercomputers. The Shared memory
machines were still popular but in contrast with them, these new architectures provided thousands of processes rather than just tens. These two models are not incompatible and are often combined. The programming model was made using Message
Passing. In its early days each organization used its own message-passing implementation. Later, the scientific community created an standard, Message Passing
Interface or simply MPI [62]. From an algorithm point of view, there had to be made
complete changes to the code. This was called domain decomposition and required a
complete re-structuring of the program, which needed to be split.
These architecture designs enter in Flynn’s MIMD category. In this group there
are two main classes: Massively Parallel Processor (MPP) and Commodity Clusters
(or simple Clusters). Both are described below.
Chapter 2: Background
21
1. MPP Massively Parallel Processor (MPP) are complex and expensive systems.
An MPP is a single computer system which sometimes is formed by many nodes.
The nodes contain many connected processors which are tightly coupled by
specialized interconnect networks (e.g: Hypercube, Mesh or Torus [1]) [26, 17].
Examples of MPPs are the IBM Blue Gene series or the Cray Supercomputers
XT30 and XC40 among others. As it will be difined in the section Graph
computations (Section 2.6), these systems are referred as Lightweight by the
Graph 500 BoF 4 .
2. Clusters systems are usually cited as: “a home-made version of an MPP for
a fraction of the price” [47]. However this is not always true, as it can be the
example of the Tianhe-2, in China. These, are based upon independent systems
(Usually Shared Memory Systems) connected through high speed networks (e.g.
Infiniband, Myrinet, Gigabit Ethernet, etc).
As of November 2015, the Top10 architectures in the Top500
and 8 MPPs (being the fastest, the Tianhe-2 Cluster in China)
5
list are 2 Clusters
In the Top10 of the Graph500 6 list, as of November 2015, there is 1 Cluster
and 9 MPPs (being the fastest, the K-computer MPP in Japan)
2.3.4
The popularization of the multicore architectures
Due to the change in the manufacturing process of the silicon-chips, the rest of
the world followed suit.
The required power to enhance the performance (by increasing the area of the
circuit block) varies at a higher rate than the performance does. (Figure 2.4) [62].
One direct effect of the latter, is that the generated heat increases (proportional
to the consumed power). As a result, alleviating the effects of this heat, and keeping
the silicon chips within an operative range is difficult. The physical explanation of
this is:
The generated heat in a chip is proportional to its Power consumption P [41],
which is given by (1). On this, when the frequency increases also the power consumption and the heat rise.
C −→ The Capacitance being switched per clock cycle (proportional to the number of transistors)
V −→ Voltage
4
http://www.graph500.org/bof
5
http://top500.org/lists/2015/11/
6
http://www.graph500.org/results_nov_2015
Chapter 2: Background
22
Parallelism needed
Power consumption increases at a higher rate than performance
Core
Independent core
Core
Core
Core
Shared L3 Cache
Integrated cores
Mobile phone Processor
0.1 Watts
4 GFLOP/s
Intel Nehalem ® Processor
x1000 power consumption
x10 performance
100 Watts
50 GFLOP/s
Figure 2.4: Power consumption increases at a higher rate than performance [62].
F −→ The processor frequency
P =C ∗V2∗F
(1)
This is what ultimately led to Intel’s cancellation of several microprocessors in
2004 [36]. Also, this moment in time is generally cited as the “End of Frequency
Scaling”. As this stopped the increasement of the frequency, many cores started
to be added into one single chip. There are also more reasons (economical) which
affected the evolutions in this chip technology. For example, the relationship between
the profits and the technology costs (smaller technologies are more expensive). This
ultimately led to an slow down to new smaller technologies [62] (Fig. 2.5).
From the point of view of the software, this evolution made threads a popular
programming model in the last years (as it had been learned a lot about of them with
Shared Memory architectures, used in Supercomputing). Also, other programming
models like the less restricted “fork-join” threads used in CILK, started to be more
popular and practical.
2.3.5
Energy efficiency
After going through the limitations that heat imposes to chips, and how the energy
consumption rises with the in-chip number of transistors, it can be better understood
why energy efficiency is becoming today one of the main design constrains in High
Performance Computing.
Chapter 2: Background
23
190
Economical factors and technology −
Chips that are getting smaller
150
130
110
90
70
50
30
Size of transistors bought per dollar (nm)
170
Number of transistors bought per dollar (millions)
Size of transistors bought per dollar (nm)
20m
20m
19m
2012
2014
2015
16m
0 10
11.2m
2.6m
4.4m
2002
2004
7.3m
2006
2008
2010
Year
Figure 2.5: Manufacturing costs for smaller technologies become uneconomical [62].
To illustrate the importance of this, a real example (using modern Supercomputers) will be is used.
As of 2015, in the United States of America, the cost per megawatt is around 1
million dollars [62] and the consumed energy usually follows a chip scaling. With this
in mind, in the scenario that we are going to describe, in 2008 a 1 petaflop system
used 3 megawatts to run. Following this scaling, in 2018 a 1 exaflop Supercomputer
will require 20 megawatts.
To contrast the importance of the energy saving plans, a 20 megawatts supercomputer would require an approximate budget of 20 Million dollars which was a quarter
of the yearly budget destined to nursing in that country in 2015. [62]).
The current energy cost in super computers is serious and not sustainable. In
the last years researching on energy efficiency has moved also to algorithms and
software design. As it will be illustrated in Section 2.6, new benchmarks (like the
green.graph500 7 ) are being created to measure this Energy impact of computation.
2.4
General Purpose Graphic Proccessor Units
Overview
7
http://green.graph500.org/
Chapter 2: Background
24
Energy Consumption −
Evolution in time
19
Supercomputer
15
13
11
0 1 2 3 4 5 6 7 8 9
Energy (Megawatts per machine)
17
Tihanhe−2
1992
1995
1998
2001
2004
2007
2010
2013
2016
Year
Figure 2.6: Power consumption increases at a higher rate than performance [62].
2.4.1
GPGPU architectures
GPUS are widely used as commodity components in modern-day machines. In the
case of NVIDIA™ architectures, a GPU consists of many individual execution units
(SMs/ SMXs), each of which executes in parallel with the others (Figure 2.7). During
runtime, threads on each execution unit are organized into thread blocks, and each
block consists of multiple 32-thread groups, called Warp. If threads within a Warp
are set to execute different instructions, they are called “diverged”. Computations
in diverged threads are only partially parallel, thus reducing the overall performance
significantly.
The GPU includes a large amount of device memory with high bandwidth and high
access latency, called global memory. In addition, there is a small amount of shared
memory on each execution unit, which is essentially a low latency, high bandwidth
memory running at register speeds. Due to such massive amounts of parallelism,
GPUs have been adopted to accelerate data and graph processing. However, this
maximum potential parallelism is many times difficult to achieve.
2.4.2
Regular and non-regular problems
• In regular code, control flow and memory references are not data dependent,
for example SpMVM is a good example. Knowing only the source code, the
input size, and the starting addresses of the matrix and vectors we can predict
Chapter 2: Background
25
GPU
SMX 1
…
SMX n
SMX 2
INSTRUCTION CACHE
WARP SCHEDULERS
REGISTER FILE
CORE
CORE
CORE
CORE
…
CORE
CORE
CORE
CORE
CORE
…
CORE
CORE
CORE
CORE
CORE
…
CORE
CORE
CORE
CORE
CORE
…
CORE
INTERCONNECT NETWORK
SHARED MEMORY (L1 CACHE)
L2 CACHE
GLOBAL MEMORY
Figure 2.7: Simplified architecture of an NVIDIA™ Kepler GPU
the program behavior on an in-order processor.
• In irregular code, both control flow and memory addresses may be data dependent. The input values determine the program’s runtime behavior, which
therefore cannot be statically predicted. For example, in a binary-search-tree
implementation, the values and the order in which they are processed affect the
control flow and memory references. Processing the values in sorted order will
generate a tree with only right children whereas the reverse order will generate
a tree with only left children.
Graph-based applications in particular tend to be irregular. Their memory-access
patterns are generally data dependent because the connectivity of the graph and
the values on nodes and edges determine which graph elements are computed next
accessed, but values are unknown before the input graph is available and may change
dynamically.
2.5
Message Passing Interface
The Message Parsing Interface (MPI) resulted as a standardisation of the multiple
implementations of communication protocols in early supercomputing.
MPI only defines an interface of calls, datatypes or restriction that are separately
implemented by other organizations. Two popular implementations of MPI are OpenMPI and MPICH.
On MPI exist two types of communication calls:
Chapter 2: Background
26
• Point-to-Point calls. Allow the direct data send and receive from / to other
node. Some examples are MPI_send, MPI_isend (non-blocking) or MPI_Recv.
• Collective communications allow us to send data from / to multiple nodes
at a time. Some examples are MPI_Reduce, MPI_Allgatherv, MPI_scatter.
2.6
Graph computations
As examined in the Introduction section (Section 1.1) Graphs and Graphs computations are becoming more important each day.
In a first part of this section it will be reviewed the Graphs structures and their
characteristics.
On a second part of the section it will be discussed the Graph 500 challenge. About
the latter it will be described what is it main motivation, what parts conform a Graph
500 application, what is the basic algorithm in a non parallelized application. Lastly
it will be discussed why the parallelization of this algorithm (BFS) is theoretically and
computationally difficult.
Also other recent benchmark (Green Graph 500), which is very related with this
one, will be briefly described.
2.6.1
Graphs
The graph algorithms encompass two classes of algorithms: traversal algorithms
and analytical iterative algorithms. A description of both follows:
• Traversal algorithms involve iterating through vertices of the graph in a
graph dependent ordering. Vertices can be traversed one time or multiple times.
This class includes search algorithms (such as breadth-first search or depthfirst search), single source shortest paths, minimum spanning tree algorithm,
connectivity algorithms and so on and so forth.
• Analytically iterative algorithms involve iterating over the entire graph
multiple times until a convergence condition is reached. This class of algorithms can be efficiently implemented using the Bulk Synchronous Parallel
(BSP) model. Algorithms in this class include page rank, community detection, triangle counting and so on.
We will focus in a parallel implementation of the traversal algorithm Breadth-first
Search (BFS). This algorithms requires some form of data partition, concurrency
control, and thread-level optimizations for optimal performance.
• Degree of a graph is the number of edges connected to a vertex.
Chapter 2: Background
27
• Distance between to vertices on a graph is the value of the shortest path
between to vertices.
2.6.2
Graphs partitioning
Parallel SpMV
1D Partitioning-based BFS (Vertex)
In 1D data partitions each processor gets a row of vertices of the adjacency matrix. This kind of partitioning is the default one and has network communication
complexity of θ(n × m), where n is the number of rows, and m is the number of
columns.
A=
A1
A2
..
.
Ak
2D Partitioning-based BFS (Edge)
The 2D-partition based BFS is known to perform better than the 1D in graph with
low degree. Unfortunately, for high degree graphs the situation becomes the opposite.
First of all, as a pre-condition for doing this kind of partition that adjacency matrix
must be symmetric in relation to the number of processors, i.e with 36 processors we
need to be able to divide the matrix into 6 × 6 blocks.
(1)
(1)
(C)
A
· · · A1,C
.1,1
..
..
.
.
.
.
(1)
(1)
AR,1 · · · AR,C
..
..
..
.
.
.
A= .
..
.
..
..
.
(C)
A1,1 · · · A(C)
1,C
.
..
..
..
.
.
2.6.3
(C)
AR,1 · · · AR,C
Graph 500 challenge
Some concept attaining the Graph 500 challenge are described below.
Chapter 2: Background
28
Performance the Traverse Edges per Second (TEPS) is used as the main performance metric of the Graph 500 applications. TEPS is proposed by the graph500
benchmark to report the throughput of supercomputers on graph processing.
Energy efficiency is also a metric which measures the efficiency in terms of energy
expressed as traversed edges per joule. Since platforms have heterogeneous hardware
resources and different power consumptions, their performance (TEPS) is divided by
their power to calculate the energy
Provided reference implementations The Graph 500 provides five reference implementations with different characteristics. This implementations have the following
characteristics
1. Sequential This reference implementation uses a sequential BFS algorithm
and introduces no parallelism but the data vectorizations introduced by the
compiler. It could fit for example a UMA architecture, using one only processor
on its CPU, such as a mobile phone or a tablet.
2. OpenMP The benefit of this implementation comes from the hand of the introduced thread parallelism which may be achieved with OpenMP and a multicore
architecture. An example architecture which would fit in this category would
be a laptop or a Personal computer.
3. Cray XMT This implementation is specific for Cray™ MPPs. Takes advantage
on the specific interconnectors of this HPC supercomputers.
4. MPI On this implementation the parallelism is based on Flynn’s MIMD category (Table 2.1 and Figure 2.2). This is the problem is partitioned and divided
into several processors through Message Passing Interface (Section 2.5)
5. 2D data partitioned MPI.This implementation has been added with posteriority and is based on the 2D problem partitioning purposed in [12]. As the
previous one distribute the load among processors using MPI. This reference
implementation is the more structurally similar to our one. .
All these implementations are based on CPU processors.
Problem sizes
According to the Graph 500 consortium problem sizes enter into different categories depending on their size. The size is expressed in logarithmic scale and represents the number of total vertices inside the graph. An overview of the categories
may be seen in table Table 2.2.
Chapter 2: Background
29
As it will be described in section 2.7 and it is cited briefly here, The graph 500
challenge makes use of a Kronecker generator [44] to create the graph. The generated
graph (as required in this challenge) contains 16 edges per vertex (Edge factor) . This
latter is the reason why the Graph 500 applications works over sparse graphs.
In the table 2.2 Real world graphs whit a higher number of edges per vertex
(degree of the graph) have an scale greater than 30. As it was previously described
in this same section a node with logarithmic scale of 30 would have 230 vertices and
234 Edges.
One note which will also be referenced next in the document is that the meaning
of using an Sparse vector with elements of (SpMV) of 64-bit means that a Row
(or column for symmetric 2d partitions) of the CSR matrix representing the whole
graph, also has 64 elements. This would mean that the maximum represented graph
size could have 264 Vertices. If we think about this in terms of compression. Is our
compressor algorithm works over 32-bit integers (integer is other way of calling the
bitmap SpMV) would mean that the maximum archivable graph size (in Table 2.2)
would be small size.
Problem class
Toy (level 10)
Mini (level 11)
Small (level 12)
Medium (level 13)
Large (level 14)
Huge (level 15)
Scale
26
29
32
36
39
42
Edge factor Approx. storage size in TB
16
0.0172
16
0.1374
16
1.0995
16
17.5922
16
140.7375
16
1125.8999
Table 2.2: Problem sizes for a Graph 500 BFS Benchmark
Structure and Kernels
Algorithm 1 A Graph 500 application in pseudocode
Graph generation (not timed) “Graph Generation as a list of Edges”
Kernel 1 (timed) “Graph Construction and conversion to any space-efficient
format such as CSR or CSC”
3: for i in 1 to 64 do
4:
Kernel 2 (timed) “Breadth First Search (timed)”
5:
Validation (timed) “Validation for BFS tree using 5 rules”
1:
2:
A graph 500 application is formed by 4 different steps, some of which are timed
and some of which not. Also, has two differentiated core parts called kernels, in
the first one improvements over the structure containing the graph are allowed and
Chapter 2: Background
30
timed. The second kernel contains the code of the BFS algorithm. This latter step is
repeated 64 times and on each of them the results are verified.
As a result of a benchmark are required some statistical parameters, such as the
mean execution time or the Traverse Edges per Second metric (TEPS) calculated as
an harmonic mean of all the 64 required iterations. The algorithm of the benchmark
may be seen as pseudocode in table 1.
The Green Graph 500 Challenge
The Green Graph 500 8 challenge is an Energy-Aware Benchmark launched in the
year 2012. It is based on the similar technical principles than the Graph 500. Some
aspects like the metrics (TEPS/Watt), the kernels structure, or the submission rules
[27] differ from the previous benchmark.
As it was discussed in the previous section Looking for energy efficient technologies
(Section 2.3.5), energy is becoming a concern between the HPC community due to
the increasingly maintenance costs.
Due to this fact new improved algorithms and implementations are starting to be
focused also on a energy-efficient implementation.
2.7
2.7.1
Input data
Synthetic data
In the graph 500 challenge’s algorithm of table 1, the step on line 1 generates a
list of edges, that in next steps will be converted into other structures.
The graph 500 consortium sets some specifications for the generated graph and for
the used generator. The generator [44] builds open scale graphs which have similar
statistic distributions to the graphs observed in the real world. the specifications on
the Graph 500 challenge require, apart from the usage of a Kronecker generator, a
number of 16 edges per vertex.
This graphs generated synthetically this way have been target of many statistical
research. In the figure 2.8 it may be seen a typical graph 500 graph, following a Power
law distribution. One reason for showing how a graph looks like statistically, is to
see the comparison with that data that it is distributed by the processors and will be
compressed later.
Kroneker generated graphs meet the following relations:
vertices = 2scale ∧ edges = 2scale × edgef actor
8
http://green.graph500.org/
Chapter 2: Background
31
0
50 100
250
150
0 50
200
0
1000 2000 3000 4000
Empirical and theoretical CDFs
P−P plot
50 100
Data
200
0.4
0.0
0.8
0.4
0
0.8
Theoretical quantiles
Empirical probabilities
Data
0.0
CDF
Q−Q plot
Empirical quantiles
0.04
0.02
0.00
Density
Empirical and theoretical dens.
0.2
0.4
0.6
0.8
Theoretical probabilities
Figure 2.8: Mass distribution of a synthetically generated graph. The
distribution has been generated using [18]
2.7.2
Real world graphs (datasets)
A real world graph usually is greater than a scale 30 kronecker generated graph.
Other differences with kronecker are the number of edges per vertex. real world
graphs have a much larger edge-factor, and thus a higher diameter (maximum distance
between two vertex in the graph).
Chapter 3
Related work
3.1
Optimizations
1. 2D decomposition This optimization is described in the section Graph computation.
2. Vertex sorting Vertex sorting is a technique that helps to alleviate one of the
major problems of the graph computations. Big size graphs are represented
with structures with a very low spatial locality. This means that for example
to traverse a graph with a distributed algorithm, the required data to compute
the computation may be allocated in other processor. Unfortunately there is no
known way to prevent this. The Vertex sorting technique is combined with the
previous one (2D decomposition) in such a way that with the decomposition
each processor would have a subset of the Adjacency matrix and by adding an
extra phase just at the begging to relabel the vertex, we add extra locality to
the edges allocated in a same processor.
3. Direction optimization . Beamer et al. [7, 6] proposed two new techniques
to deal with the unnecessary vertex computations when performing the distributed BFS algorithm. As an introduction, the usual way to transverse a
graph is called Top-down: the adjacency matrix is used in such direction that
provides us with the neighbour nodes to the ones in our frontier queue. One
characteristic of the Top-bottom algorithm is that performs bad with high degree graphs (graphs with many neighbours in one vertex). to avoid unnecessary
computations. the Adjacency matrix is duplicated to be able to use the opposite
direction to traverse. This way when the number of neighbours is greater than
a δ, the number of nodes in vertex in the opposite direction is computed and
the minimum among them is chosen ase direction to follow.
4. Data compression Since the latency in the communication between the processors limits the maximum computational speed in the nodes by stalling them
32
Chapter 3: Related work
33
without new data, this method intend to alleviate this effect. This is the main
goal of this work.
5. Sparse vector with pop counting. The pop instruction (__popc in CUDA
C) are instructions to deal with bitmaps with hardware support. Some new
implementations of the graph500 use this optimization.
6. Adaptive data representation
7. Overlapped communication The technique of overlapping communication
(also called software pipelining) allows to reduce the stall time due to communication. (Described in further detail in section 6.5).
8. Shared memory
9. GPGPU General purpose GPUs are reviewed in section 5.3.2.
10. NUMA aware With NUMA aware specific libraries it is possible to “pin” a
full data structure to an specific (core, processor, bank of memory) in systems
with the possible penalty of accesses from one processor to a bank of memory
belonging to other processor.
3.2
Other implementations
1. Hybrid (CPU and GPGPU)
• “B40C” (Merrill et al.) [33]
• “LoneStar-Merrill”
1
[9]
• “BFS-Dynamic Parallelism” (Zhang et al.) [65]
• “MapGraph” [21]
• “GunRock” (Wang et al.) [55]
• “Enterprise” (Liu et al.) [30]
• SC11 (Ueno et al.)
• ISC12 (Ueno et al.) [52]
• SC12 (Ueno et al.) [51]
2. CPU-based
• ISC14 (Fujisawa et al.) [60]
1
http://iss.ices.utexas.edu/?p=projects/galois/lonestargpu
Chapter 3: Related work
34
Optimizations
2D decomposition
Vertex sorting
Direction optimization
Data compression
Sparse vector with pop counting
Adaptive data representation
Overlapped communication
Shared memory
GPGPU
NUMA-aware
SC11
X
X
ISC12 SC12 ISC14
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
Table 3.1: Other implementations
3.3
Compression
In this subsection we present a list of other work on compression applied to
Breadth-first Search implementations. We discuss their algorithm and compare this
to ours.
1. “Compression and Sieve: Reducing Communication in Parallel Breadth
First Search on Distributed Memory Systems” (Huiwei et al.) [31]
In this work a 1D partition-Based BFS using SpMV multiplications is analyzed
and with and without compression on its communication. The approach followed to compress the sparse vectors is to state that they have low cardinality
as they are bitmaps. The compression codecs they analyze enter into the category of Bitmap Indexes. The selected codec is WAH (Word Aligned Hybrid)
a patented codec with an medium to low compression ratio but on the other
hand an acceptable speed. The result of the study shows a reduction in the
BFS time of over the 50% thanks to the compression.
2. “Highly Scalable Graph Search for the Graph500 Benchmark” (Ueno
et al.) [51].
The implementation of Ueno et al. had a good position in the graph 500 challenge during June 2012. It was it that year when they introduced compression.
In this case the compression was implemented over a 2D Partition-based BFS
executing on a GPFPU device. In contrast with Huiwei et al. Uneno treated
the data as a sequence of integers instead of a sequence of bits. Also they
chose a Varint-based encoder (VLQ) with neither a good compression ratio nor
a good compression / decompression speed. The main goal of this algorithm
was its implementation simplicity. Once selected they implemented it on both
Chapter 3: Related work
35
the GPU device and the CPU. Also they reimplemented two more versions, a
32-bit based and a 64-bit.
A not about GPGPU devices is their current lack of native arithmetic support
for 64-bit integers. The choose of a full 64-bit compression on the device would
have make the algorithm to incur on a performance penalty. For this reason and
also to avoid performing compression for low amount of data they implement
(as we have also done) a threshold mechanism.
Chapter 4
Problem Analysis
As aforementioned in the Introduction the optimizations presented in this work
have been implemented on a previously developed code [25].
In this section, we will analyze the previous implementation and decide (based
on the results of the analysis of the code and the instrumentation experiments) what
changes will be made to increase the overall performance.
Also, we will review the optimizations already applied to our code based on other
state of-the-art research (discussed in Section 3 - Relared work). Next, we will explain how we have measured the data using internal (coded timers) and external instrumentation (profilers). In the case of the external instrumentation we will compare
the available candidates and review the chosen one according to our needs.
In addition, we will also include the measured data of our instrumentations. Regarding the communicated data, it will be highlighted what is its type, as it will be
useful to understand the effectiveness of the chosen compression scheme.
Finally, and based on the data of the instrumentation results, we will analyse and
decide what approaches are considered to be feasible to alleviate the detected issues
and optimize the overall performance.
4.1
Our initial implementation
A list of the optimizations implemented on the Baseline version of this work is
listed in the table 4.1.
4.1.1
Optimizations table in “Baseline” implementation
The list of optimizations may be seen on table 4.1.
36
Chapter 4: Problem Analysis
Optimizations
2D decomposition
Vertex sorting
Direction optimization
Data compression
Sparse vector with pop counting
Adaptive data representation
Overlapped communication
Shared memory
GPGPU
NUMA-aware
37
Baseline implementation
X
X1
X
Table 4.1: Other implementations
1
Feature to be implemented in this work
4.1.2
Our “Baseline” algorithm
In our initial algorithm, where the data is partitioned in 2D fashion, each of the
processors operates over is own assigned block of the global Matrix. The
As is was discussed in the section General Purpose Graphic Processor Units (Section 2.4), some requirements are needed to build an optimal algorithm over a GPGPU
device.
In this section it will be listed the pseudo-code of the used parallel BFS algorithm
implemented over CUDA. This used algorithm is based on the implementation of Merrill et Al. [33, 25] and includes modifications to adapt the code to a 2D-partitioning.
4.1.3
General communication algorithm
The code presented in this work makes use of a 2-Dimensional partitioning of the
CSR initial Graph Matrix.
The 2D partitioning were proposed as an alternative to the 1D to reduce the
amount of data transmitted through the network [63].
Being G(V, E) a Graph with |V | Vertexes and |E| Edges, on an original 1D strategy the distribution unit across the network would be the Vertexes performing a
complexity of O(P ) where P is the total number of Processors.
On the other hand, on a 2D strategy
√ the transferred unit are the Edges. This
leads to a complexity reduction of O( P ).
Described briefly, on a 2D partitioning (the used one) there are two phases where
each processo operates and transmit part of their Matrixes.
On the first sub-phase the BFS operates over rows and as result performs an
“AllReduce” Operation that transmits the Column. This may be viewed on 2 and 3
Chapter 4: Problem Analysis
38
Algorithm 2 Our BFS algorithm based on replicated-CSR and SpMV
Input : s: source vertex id
Output : Π: Predecesor list
f: Frontier queue / Next Queue
t: Current queue
A: Adjacency matrix
Π: Visited node /predecessor list
⊗: SpMV multiplication
: element-wise multiplication
1: f (s) ← s
2: for each processor Pi,j in parallel do
3:
while f 6= ∅ do
4:
T ransposeV ector(fi,j )
5:
fi ← ALLGATHERV(fi , P∗,j )
6:
ti ← Ai,j ⊗ fi
7:
ti,j ← ALLTOALLV(ti , Pi,∗ )
8:
ti,j ← ti,j Πi
9:
Πi,j ← Πi,j + ti,j
10:
fi,j ← ti,j
of Algorithm 3.
The next sub-phase Multicasts rows [25] Line 4 of the algorithm.
Algorithm 3 Simplified communication algorithm for 2D-partitioning
1: while “there are still Vertexes to discover” do
2:
updateDataStructures ← BFS_iteration
3:
column_communication
4:
row_communication
5: combine_predecessorList
4.1.4
Data and communications
This section details the types of data that we will use in our communications.
These are based on two structures, the first one consists of a sequence of long integers
used to distribute vertexes between the nodes 1 [25].
1
Update: In our implementation, for the used graph generator and the maximum reachable Scale
Factor , these integers are in the range [1..224 ] and are sorted in ascending order. Being the smallest,
the first one in the sequence.
Chapter 4: Problem Analysis
39
A second structure is used to prevent nodes of re-processing already visited vertexes.
The Frontier Queue The frontier Queue is the main data to be compressed. Each
of the elements on its sequence is a sparse vector of bits (SpMV). The analysis that
we have performed to multiple big buffers of this data shows an uniform distribution,
slightly skewed (Section 5).
The Predecessor list bitmap The bitmap is used in the last call of the BFS
iteration is a 32-bit sequence of integers that is used to convert the frontier queue in
the predecessor list.
4.2
Instrumentation
In this section It will be covered the selection of the Performance analysis tool.
It will be discussed the used criteria, the needed requirements in our application and
some of the available the options.
As mayor requirements it will be set a low execution overhead, and the ability to
wrap specific code zones for measurement.
Minor requirements will be the availability of data display through console. This
feature will enable the data to be tabbed more easily into diagrams.
• TAU 2 Originally implemented in the University of Oregon. Offers console
interface. Narrow code wrapping is not implemented.
• Vampir
3
Offers graphical interface. We have not tested this option.
• Score-P 4 Offers both graphical (using Cube 5 ) and console interface. Allow
direct code zone wrapping.
4.2.1
Instrumented zones
Listed below are the instrumented zones corresponding to the second Kernel
(timed), which involves the communication, the BFS iteration.
2
https://www.cs.uoregon.edu/research/tau/home.php
3
https://www.vampir.eu/
4
http://www.vi-hps.org/projects/score-p/
5
http://www.scalasca.org/
Chapter 4: Problem Analysis
40
• BFSRUN_region_vertexBroadcast initial distribution of the Start vertex.
Performed only once.
• BFSRUN_region_localExpansion Execution of the BFS algorithm on the
GPU device.
• BFSRUN_region_testSomethingHasBeenDone Checks if the queue is
empty (completion). To do so performs an MPI_Allreduce(isFinished, LogicAND) between all nodes.
• BFSRUN_region_columnCommunication. Performs a communication
of the Frontier Queue vectors between ranks in the same column. This steps
contains the next one.
• BFSRUN_region_rowCommunication. Performs a communication of the
Frontier Queue vectors between ranks in the same row.
• BFSRUN_region_allReduceBitmap. Merges the predecessors in all nodes
and check performs the validation step of the results. Run only once at completion of the BFS iteration.
4.3
Analysis
In this section it will be analyzed the previous instrumentation and instruction
complexity data.
As a result, a solution will be given for each of the issues found. The implementation, results and final conclusions, follow in the next sections of this document.
4.3.1
Communication overhead
As is has been mentioned at the beginning of this chapter, our Baseline implementation is mainly based upon a GPGPU BFS kernel, based in the implementation
of Merrill et al. [25], with a 2D data partition, which enables the implementation to
perform within a multi-GPU platform and in a multiple node environment through
Message Passing Interface (MPI).
In the results of the previous work, it was noticed a loss of performance in comparison to the original Merrill et al. [33] implementation. This was despite the application
was running on multiple distributed multiGPU nodes within a fast cluster.
As a result of analyzing the reasons of the performance loss, it was decided to
mitigate latency generated in the communications through compressing the transfers.
That new improvement is the work presented in this document.
Chapter 4: Problem Analysis
4.3.2
41
Instruction overhead
As we were going to add a compression system which would be enabled at compile
time, we still had space to add and test new improvements. The analysis of the code
showed for example that the execution using thread parallelism (OpenMP) was slower
than the one where the application was running on a single thread. Even though the
experiment platforms has capabilities for multiple thread execution.
In a more on detail analysis we saw that some of the code entering in the timed
parts of the application performed matrix intensive operations and allowed parallelism
by the compilers. We also saw that many scalar improvements could be applied
without much complexity.
In the last part we ran the compiler in inspection mode to retrieve the successfully
applied optimizations in our Baseline implementation. Also, and to set a reference
point with other distributed BFS implementation, we also performed this test on
those. The other implementations are: (i) a Graph500 2D partition-based with MPI
reference implementation. (ii) and the latest state-of-the-art Graph 500 implementation using GPGPU [51]. The comparison may be seen in table .
Chapter 5
Optimizing data movement by
compression
By compressing the data movements in the graph500 execution, we intend to solve
the issues detected in the previous section (Section 4.3.1).
In this section we will first start describing some compression-related concepts for a
better understanding of the optimizations. We will describe the available compression
techniques. For each of them, we will expose how they match the Graph 500 data types.
We will also discuss the related and most state-of-the-art compression algorithms (also
known as codecs or schemes). After describing the algorithms we will focus on the
available open-sourced library choices. Last, we will focus on the integration and
discuss the decisions taken here.
5.1
Concepts
We list below related compression terms that are used in this work and its related
literature [68, 28, 11, 46, 29, 58, 3].
Information Theory
The Information Theory was the result of applying statistical analysis to measure
the maximum compression achievable in a communication channel. It was created
by Claude Shannon in 1948, in his paper A Mathematical Theory of Communication
[45].
In the successive years, the Information theory has applied statistics and optimal
codes to assign the shorter codes to the most probable symbols on sequence to increase
that way the compression ratio (defined in this section). This ratio depends on some
parameters of the data, like its Entropy (also defined in this section).
42
Chapter 5: Optimizing data movement by compression
43
Information Retrieval Systems (IRSs)
An Information Retrieval Systems (IRSs) is a system which searches and retrieves
data in environments of large data-volumes (for example Database or the World Wide
Web).
To achieve this, modern techniques like Indexing are used. These techniques make
use of specific compression schemes which have required to be adapted to numerical
data (indexes). This benefits our Graph 500 application in two ways:
1. Since our application also uses numerical data - on each communication it transmits a “Frontier queue” (which is a sequence of integers), we are able to use an
integer-specific encoding for our purpose. Also, because of the current importance of this IRS systems, this algorithms improves quickly.
2. These techniques, due to the requirements of this IRS systems, have evolved to
offer high speeds with low-latencies. This factor is important since one requirement in our application is the performance.
Indexer
An Indexer is a type of the IRS. As defined before they manage the access to the
stored indexes. A part of this Indexer systems is usually compressed due to the high
amount of data that they need to store.
Common web search engines like Google1 are examples of Indexers. In the case
of Google, the indexing system is based in the Variable Byte algorithm (an specific
variant called varint-GB) [13] which will be discussed in further detail below.
Inverted Indexes
The Inverted indexes are the most commonly used data-structures (among other
options) to implement indexer systems [67]. They consist of two parts:
1. A lexicon: it is a dictionary-like structure that contains the terms that appear
in the documents plus a “document frequency”. The latter is somehow similar
to a “ranking”, and indicates in how many documents those terms appear.
2. Inverted list (also called Posting list), is an structure which contains (for each
term) the document identifiers, its frequency, and a list of placements within
the documents.
This latter part usually becomes very large when the volume of the data is high,
so it is often compressed to reduce the issue. The used schemes are required to
meet this criteria:
1
https://www.google.com/
Chapter 5: Optimizing data movement by compression
44
• they need to considerably reduce the size of the Inverted list by achieving
a high compression ratio.
• the need to allow a very fast decompression to not affect the search speed.
Shannon’s Entropy
As aforementioned Shannon’s Information theory settled the foundation of the
data compression. One of the main concepts of this theory: the “Entropy” enables us
to know the maximum theoretical limit of an optimum compression. The definition
of the concept is given below.
Put simply, Shannon’s Entropy can be defined as the amount of unpredictability
of an event in terms of information. In Shannon’s Entropy events are binary.
According to Shannon, in a distribution where a binary event happens with probability pi , and the event is sampled N times, the total amount of information we
receive is given by (2) 2 .
H(X) = −
N
−1
X
pi log2 pi
(2)
i=0
To illustrate the concept we will use the example of tossing a coin with equal
result probability. In this case, because we have no information about the possible
result (Each side of the coin has p(x) = 0.5), the Entropy (Unpredictability) will be
maximum according to Equation (2) (H(X) = 1). Visually, this can be seen3 in the
Figure 5.1.
Limits of compressibility
The minimum number of bits per symbol will be given by the Entropy of the
symbols in the sequence. Expression (3).
h
i
bits = H(X)
(3)
Compression Ratio
According to the Information Theory [45], the maximum compression resulting
from a data sequence, depends on its probability density function. Accordingly, the
bigger the predictability in the sequence, the greater the achieved compression.
2
3
This expression applies to discrete distributions.
To display Entropy as a continuous function it has been approximated to H(p) = −plog(p) −
(1 − p) log (1 − p) [10].
Chapter 5: Optimizing data movement by compression
45
0.0
0.2
0.4
H(X)
0.6
0.8
1.0
Shannon's Entropy Function − H(X)
0.0
0.2
0.4
0.6
0.8
1.0
Pr(X=1)
Figure 5.1: Shannon’s Binary Entropy Function
The resulting compression is proportional to the compression ratio (greater is
better). This ratio is shown in the expression (4).
This concept will be useful to measure the impact of the optimization of our
communications in the final results.
ratio =
original_size
compressed_size
(4)
“Lossless” and “Lossy” compressions
These terms are related with compression and thus defined below. First, the
compression of our data needs to meet these requirements:
1. the size of the encoded data needs to be lower than the size of the original size.
2. the encoding algorithm must allow recovering the original sequence from the
compressed sequence, without any data loss.
When both previous criteria are met we are talking about “Lossless codecs”. These
are the ones used in this work.
In other literature, only the first criteria above is required, and some tolerance
is allowed in the second one. These are called “lossy codecs”. They are often used
to reduce the size of media data (images, sound and video). These are not being
discussed in this work.
Chapter 5: Optimizing data movement by compression
46
Compression algorithms, Schemes, Codecs and Encodings
In this work the compression algorithms will also be named schemes, codecs or
encodings. These four terms are equivalent.
Dependency
To perform the compression of a full integer set some of the codecs operate over
each integer independently. Others act over a group of integers and return a sequence
in which each compressed block have dependencies to the others. These dependencies are called oblivious (independent) and list-adaptive (dependant) in the literature
[57, 11]. They also appear as Integer encoders and Integer sequence encoders in the
literature [23].
Usually because independent (oblivious) codecs compress each value by its own
and discard the information about the rest may miss global characteristics which
could benefit the global compression. Codecs belonging to this method are reviewed
on this section even though the main focus will be on the dependent (list-adaptive)
schemes.
Compression techniques
The compression techniques discussed below are the foundation blocks of the
algorithms reviewed in this document.
Next to these techniques are listed others (“Novel techniques and optimizations”)
that are usually combined with the former.
As a result, very optimized algorithms, operating over a very specific data domain
(integers and integer sequences) are built. These encodings are, as of the moment of
writing this document, the start-of-the-art integer compression algorithms.
1. Dictionary Compression This is the first of the main compression technique
exposed in this document and encompass multiple algorithms families.
The dictionary technique is the most general compression method, and can be
applied to any data domain. Usually this technique is made by using Huffman or Arithmetical optimal codes. The symbols to be compressed (dictionary
elements) are assigned a dense code according to their relative probability of appearance in the document. The higher the probability of the symbol the shorter
the assigned code. This technique has the downside of requiring a full recoding
if new symbols are added. A way to check the efficacy of this method for our
data, is by generating the empirical Entropy (defined next, in this section) of
our distribution of numbers.
Chapter 5: Optimizing data movement by compression
47
2. Prefix Suppression PS (or length encoding) This is the second global
compression mechanism described in the document. This technique is applied
to data know a-priory to be a integer. Prefix suppression (PS) suppresses the 0’s
at the beginning of each element in the sequence. As a direct result, we reduce
the sequence’s domain. This technique generates a bit-aligned variable-sized
sequence.
3. Frame of Reference (FOR) The next global compression mechanism described this document is Frame-of-reference. This method was first described
in [24]. This compression technique is applied to a sequence of integers. The
way it works is by splitting the sequence of elements in n blocks, so that for
each block B
∃ min ∈ B · ∀ i ∈ B · (i‘ = i − min)
(5)
This process is repeated for all blocks in the sequence. The size of each of the
n blocks B is:
h
i
size = log2 (max + 1 − min)
(6)
This technique contrasts with PS in that the generated sequence has a fixed
size. This resulted sequence has a byte-level alignment.
This family of algorithms will be the main focus of this work (discussed below)
as they give a good compression ratio and low compression / decompression
times.
New techniques focused on integer/ integer sequences compression
These optimizations are applied to the previous techniques to improve the algorithm. The technique Binary Packing described below is a full technique per se, and
has been used as the core compression algorithm in this work (S4-BP128 codec). Any
of these techniques below may be combined.
1. Delta compression, differential coding or delta-gaps technique The
concept is firstly described in [37]. The delta compression, may also appear
with the names Differential coding, delta-gaps, d-gaps or δ-gaps in the literature.
[68, 28, 11, 46].
The technique works by storing the distances between elements instead of the
elements themselves (which have lower size than the main integers).
Regarding the research tendency over this technique, it is changing in the last
few years:
Chapter 5: Optimizing data movement by compression
48
• the initial work was focused on the gap probability distributions to achieve
the optimal compression ratio. For example, depending on the frequency
of each term we could adjust the selected Huffman or Golomb’s codings
[56].
• the most recent work emphasizes in a better compression ratio and lower
decompression latency [49].
2. Binary packing This technique implements both the concepts of Frame-ofReference (FOR) and differential coding. If in FOR each integer in a range is
coded and then represented in reference to that range. For example, for a block
of size 128, and integers in the range [2000, 2127] the can be stored using 7 bits
as offsets of the minimum value of the range (i.e. 2000). Using the expression
(6) the resulting size can be seen in (7)
h
i
log2 (2127 + 1 − 2000) = 7
(7)
With the technique of differential coding, now the minimum value is coded and
the rest of elements are represented as increments δ to that value.
With this technique there are two main factors which give theresulting size of
the compressed sequence. These are:
• the number of bits b used to encode the blocks (in the worst case block)
• the block length B
Initial schemes of this technique, refered in the literature as PackedBinary [4]
or AFOR-1 [14], use a fixed block size of 128 or 32 integers. Future variations
of the algorithm make use of variable size blocks (AFOR-2, AFOR-3) [16].
3. Patched coding and Exceptions Binary coding may lead at some cases to
very bad efficiency, resulting sometimes in compressions needing more bits than
the original representation. This is for the example the case of compressing
sequences like the following:
5, 120, 300, 530, 990, 25, 980, 799273656454
(8)
In the sequence given by (8) in case of only needing to compress the 7th first
numbers the required bits would be
h
i
log2 (990 + 1 − 5) = 9, 9
(9)
Unfortunately, if the 8th number would need to be included in the block, the
compression would even result in a size bigger than 32-bit per integer.
Chapter 5: Optimizing data movement by compression
49
To alleviate this problem Zukowski-et-al [68] proposed the concept of exceptions,
where the the 8th number in this sample sequence would be stored in a separate
space. This is the concept of exceptions, and will encompass all values greater
than 2b , being b the value defined in the Binary packing concept.
5.2
Compression algorithms (codecs or schemes)
Names/Families
Unary
Gamma
Huffman-int
Golomb
Rice (Ricek )
Variable Byte
Simple family
FOR family
PDICT
Lempel-Ziv family
Alignment Dependency Technique Target data
bit
Independent
PS
Integers
bit
Independent
PS
Integers
bit
Independent
PS
Integers
bit
Group
PS
Integers
bit
Group
PS
Integers
byte
Independent
PS
Ints. Seq.
word
Group
PS
Ints. Seq.
byte
Group
FOR
Ints. Seq.
Dict.
Ints. Seq.
Dict.
low H(x) Seq.
Table 5.1: Compression algorithms (also called codecs or schemes)
To describe the compression algorithms used in this document, we will group them
first, according to their resulting alignment in the compressed sequence. These groups
and their features are shown in Table 5.1 . In addition to this, some terms used here
are described in the introduction of the current section.
The main compression groups to be described are Bit aligned, Byte aligned, Word
aligned and Dictionary based.
A. Bit aligned algorithms
(a) Unary This oblivious (group independent) codec represents an integer as
N one bits followed by a zero bit which acts as a terminator.
(b) Gamma This oblivious codec consists of two parts. In the first it represents an integer as log2 N bits in Unary scheme. In the second part the
f loor(log2 N ) bits of the integer are also represented with Unary. This
scheme is usually inefficient as integers become larger. To avoid this sometimes Delta compression is used in the second half of the encoding.
(c) Huffman-int This scheme is also an independent integer encoder. This
is a variation of Gamma in which the first half (the first log2 N bits) are
encoded whith Huffman optimal codes instead of Unary.
Chapter 5: Optimizing data movement by compression
50
(d) Rice This group dependent (List-adaptative) integer sequence encoder conN
sists of two parts. In the first part the f loor( 2∗k
) bits are encoded using in
Unary. The secon part uses (N ∗ mod2 ∗ k) bits for a given k. This schema
is also known as Ricek in literature.
(e) Golomb This group dependent integer sequence encoder is similar to Ricek .
In this scheme k is a power of 2.
B. Byte aligned algorithms
(a) Frame Of Reference Family This codecs are group dependent sequence
integer encoders. As described above the FOR technique has multiple
scheme implementations. Some of the differences of these schemes are
explained here.
i. PFOR [68] Original Patched coding Frame-of-Reference algorithm.
In this scheme the exceptions where stored without being compressed
including only its relative order in the linked list.
ii. PFOR2008 [64] as a compression optimization it was proposed the
compression of the exception storage, using 8, 16 or 32 bits.
iii. NewPFOR [59] The exceptions are stored in a dedicated part of the
output. This area is divided in two subareas. Simple-16 is used as
exception encoder in one of the areas.
iv. OptPFOR [59] This scheme is very similar to the previous NewPFOR but the size of the exceptions subareas is selected accordingly
to achieve better compression ration an lower decompression latency.
Simple-16 is also used as encoder scheme for the exceptions.
v. ParaPFOR [5] This is a variation of NewPFOR optimized for the
execution on GPGPUs. As an improvement, the compression of the
exceptions using Single-16 is moved from the exception part to the
original segment. This modifications worsens the compression ratio
but leads to a much faster decompression.
vi. Simple-PFOR Fast-PFOR SIMD-FastPFOR [29] In this codec,
exceptions are placed in 32 different reserved areas of the output.
These are encoded using Vectorized (with SIMD instructions) Binary
packing. The main differences in Simple and FastPFOR schemes is
in how they compress high bits of the exceptions values. In all other
aspects Simple and FastPFOR are identical. Regarding the SIMDPFOR codec, it is identical to FastPFOR except that its packs relies
on vectorised bit packing for some of the exception values.
vii. S4-BP1284 This codec was also introduced by Lemire et al. [28].
It uses the Bit packing technique described above with blocks (B ) of
4
This is the codec that has been used in this work.
Chapter 5: Optimizing data movement by compression
51
128 integers. For each block, the bit-width (b ) ranges fron 0 to 32
bits. The “S4” stands for 4-integer SIMD. This algorithm uses the
concept of Metablock to refer to 4 blocks of 128 integers (Partitioned
that way to perform better the SIMD instructions). Before Bit packing
is performed, the metablocks use some bits to contain two b values as
a header. VByte compression is added to the last blocks that can not
be grouped in a metapackage. This is the fastest Lemire et al. family
of codecs [28] and is the one used in this work.
viii. TurboPFOR A fixed size Vectorized Binary packing is also used to
encode the exceptions. As novelty, a flag bitmap is included to map
the placement of the exceptions 5 .
(b) Variable Byte Code
This popular codec is group independent (Oblivious), as it encodes each
value on its own. As a difference with the aforementioned Oblivious codecs
(Unary, Gamma and Huffman-int), this one uses byte alignment instead of
bit. A benefit of this schemes over the previous ones is that even though
the compression ratio is usually worse, the decoding speed is faster [43].
The main algorithm is known under many names: v-byte, Variable Byte,
variable-byte, var-byte, VByte, varint, VInt, or VB. An algorithm that
enter this category is the Group Varint Encoding (varint-GB) [13], used
by Google. Some improved codecs over the initial Variable Byte algorithm
will be listed below.
i. varint-G8IU [46] This scheme is based on Intel ™ SSE3 SIMD instructions. By the usage of vectorisation Stephanoph-et-al outperformed by a 300% the original VByte. They also showed that by using
vectorisation in the CPU it is possible to improve the decompression
performance more than 50%. The varint-G8IU is an SIMD optimized
variation of the varint-GB used by Google [29]. It was patented by its
authors.
ii. maskedvbyte [40] This version also uses a SIMD instruction set
based on Intel ™ SSE2 to achieve the performance improvements over
the original VByte algorithm. The codec adds a Mask (Bitmap) to be
able to perform the decoding faster with the vectorisation.
iii. maskedvbyte Variable-Length Quantity (VLQ). Only provide in this
work to reference the compression added to their Graph 500 implementation by [51].
C. Word aligned algorithms
5
https://github.com/powturbo/TurboPFor/issues/7
Chapter 5: Optimizing data movement by compression
52
(a) Simple family This family of algorithms was firstly described in [2]. It
is made by storing as many possible integers within a single n-bit sized
word. Within the word, some bits are used to describe the organization
of the data. Usually this schemes give a good compression ratio but are
generally slow.
i. VSimple 6 This codec is a variant of Simple-8b. It can use word sizes
of 32, 40 and 64 bits. As other improvements it uses a “Run Length
Encoding” 7 .
ii. Simple-8b [3] Whereas the Variable-byte family of codecs use a fixed
length input to produce a variable sized output, these schemes may
be seen as the opposite process, with multiple variable integers they
produce a fixed size output. In the case of this particular scheme a
64-bit word is produced. It uses a 4-bit fixed size header to organize
the data, and the remaining 60 bits of the word for the data itself. It
uses a 14-way encoding approach to compress positive integers. The
encoded positive integers may be in a range of [0, 228 ). It is considered
to be the most competitive codec in the Simple family for 64-bit CPUs
[3].
iii. Simple-9, Simple-16 [2, 58] These two formats (also named S9 and
S16 respectively in the literature) use a word size of 32 bits. The
former codec uses a 9-way encoding approach to compress positive
integers, but in some cases bits are wasted depending on the sequence.
To alleviate the issue the later codec uses a 16-way encoding approach.
Both of them use a 4-bit fixed header for describing the organization
in the data and 28 bits for the data itself. Due to the limitation that
this implies: they are restricted to integers in the range [0, 228 ) they
will not be reviewed in further detail as do not meet the Graph 500
requirements for a vertex size (48-bit integers).
D. Dictionary based algorithms
(a) Lempel Ziv Family and PDICT Dictionary algorithms are based in
the Dictionary technique described above. Some algorithms that use this
technique are the Lempel Ziv algorithms family (commonly named using
a LZxx prefix) and PDICT [68]. Their performance compressing / decompressing integer sequences is outperformed by the integer domain-specific
algorithms aforementioned. For this reason they will stay of the scope of
this work.
6
https://github.com/powturbo/TurboPFor
7
Note from the author inside the source code files.
Chapter 5: Optimizing data movement by compression
5.3
53
Compression libraries
For the usage of each of described algorithms in our Graph500 implementation we
could have either implemented from scratch each of the above algorithms or either use
a state-of-the-art library containing some of those algorithms. As time is a limitation
in this work and the compression algorithms are beyond of the scope of our work,
we have used several open sourced libraries based on the work of other authors. The
selection criteria of each of those libraries is described below.
Regarding these libraries, and because we noted that all of them presented different
benefits, we opted to implement a modular library integration, so we were able to
easily include (or remove) any of them so we could and test and demonstrate the
positive impact of each. The library integration has been made with the following
design parameters in mind:
• The design is simple and modular. This allows new maintainers to add new
desired compression libraries with the minimum effort. This has been done by
the usage of a software design pattern called “Factory” [22]. In this way we
have managed to improve of three of the Graph 500 ’s evaluation criteria:
1. Minimal development time: addition of new libraries by new maintainers
in an small amount of time.
2. Maximal maintainability: modification and debugging of the current code.
3. Maximal extensibility: addition of new libraries by new maintainers with
a minimal impact on the already developed code.
• Minimal possible impact in the performance of the BFS kernel. We have managed to achieve this by creating the Factory call prior to the initialization of the
second Kernel (BFS cycles and validations) and passing a memory reference to
the compression / decompression routine.
5.3.1
Lemire et al.
This library8 , has been created as proof-of-concept of various academical works [68,
28, 24]. One of the main contributions of this work is the achievement of a very high
performance and compresion-ratio. This has very high value in the compression of
Inverted indexes (Section 5.1), which are very used nowadays (as it has been exposed
before).
Lemire et al. achieved this high performance by implementing a version of the
algorithm using SIMD instruction sets in the CPUs.
8
https://github.com/lemire/FastPFor/
Chapter 5: Optimizing data movement by compression
54
Features
Languages
Input types/ sizes
Compressed type
License
Available codecs
Required hardware
Technology
Pros
Cons
Integrated in code
Lemire et al.
C++,Java
Unsigned Integer
32-bit
Unsigned Integer
32-bit (vector)
Apache 2.0
S4-BP128,
FastPFOR,
SimplePFOR,
SIMDFastPFor,
v-byte, maskedvint, varintG8IU
SSE4.1+
SIMD
Clean code
Integer size adds
extra latency because conversion
is required .
Yes
TurboPFOR
C,C++,Java
Unsigned Integer
32/ 64-bit
Unsigned Integer
32/ 64-bit (vector)
GPL 2.0
“Turbo-PFOR”,
Lemire et al.’s ,
Simple8b/9/16,
LZ4,
v-byte,
maskedvbyte,
varintG8IU
SSE2+
SIMD
Very good compression ratio
Source code is
difficult to understand.
Partially
Alenka Project
CUDA C
Unsigned Integer
32/ 64-bit
8-bit
aligned
Structure.
Apache 2.0
Para-PFOR
Nvidia GPGPU
GPGPU
Clean code. Very
good latency
Compression less
optimized than
other schemes
Partially
Table 5.2: Main feature comparison in three used compression libraries
The base of this work is the Frame-of-Reference algorithm first purposed by
Goldstein-et-al [24], and later improved by Heman-et-al who among other additions
to the algorithm, also enhanced the performance by using Superscalar techniques [68].
Some characteristics about the Lemire’s implementation are listed below.
Compression input types and sizes This library allows only unsigned 32-bit integer input type.
Compression output The output type is the same as the input: unsigned 32-bit
integer.
Included codecs FastPFOR, SIMD-FastPFOR, S4-BP128(Used), Vbyte, maskedvbyte, Varint, Varint-G8IU.
Pros and Cons Good compression ratio for the best case scenario. A 32-bit integer
is reduced to near 4-bit (value near to the Entropy) (ratio = 32
= 8). Also,
4
Chapter 5: Optimizing data movement by compression
55
the library has been implemented in a clean way and the source code is easy to
understand.
As disadvantages to fit this library in a Graph 500 application, we have found
the next:
The implementation restricts the size of the compressed 32-bit integers. This
limitations implies:
• for each compression-decompression call, a size and type pre-conversion
must be carried out carried out to adjust the type to the required by the
Graph 500.
• to avoid integer overflows in the previous conversions, it is required to
perform a pre-check over all the integers in the Graph just after the graph
generation.
5.3.2
Turbo-PFOR
The Turbo-PFOR9 library is an open sourced project containing a full set of
state-of-the-art integer compression algorithms from various authors. The library
main contribution is an improved version of the Lemire et al. [28] implementation,
mixed with the ideas and code from the other authors. The Lemire’s code has been
fully optimized10 and the library itself makes a massive use of the CPU’s SIMD
instructions.
The package also includes an statistical Zipf Synthetic Generator which allows
to pass the skewness of the distribution as a parameter. The statistical generator
allows to include into the compared terms the value of the empirical Entropy of the
generated integer sequence.
The package also compares the new optimized implementation with other stateof-the-art from other libraries. As the result of the optimizations, the author claims
a better compression ratio and a higher decompression speed for integer sequences
more similar to phenomenological Pareto distributions.
Compression input types and sizes The library allows sizes of 32 and 64 bits
integers without sign.
Compression output The implementation relays on the C Standard Template Library (STL) and the output type is the same than the input.
9
10
https://github.com/powturbo/TurboPFor
https://github.com/powturbo/TurboPFor/issues/7
Chapter 5: Optimizing data movement by compression
56
Included codecs “Turbo-PFOR”, Opt-PFOR, Lemire et al.’s set of codecs, Simple8b, Simple-9, Simple-16, LZ4, Vbyte, maskedvbyte, varint-G8IU.
Pros and Cons Similar latency than the analogous codec in Lemire et al. library,
with an 2X compression ratio improvement.
Also, the integer input size matches the “64-bit unsigned integer” required by
Graph 500. This prevents our code from the pre-conversion step in lines 2, 4,
6, 8 of algorithm 5.
As a possible downside, the code is difficult to understand since it heavily relays
on C macros.
5.3.3
Alenka CUDA
The Alenka11 project is aimed to be a fast database implementation with an inbuilt
compressed inverted indexed based on CUDA and GPGPU technology. The query
implementation is also based in this latter technology.
The main compression algorithm is an state-of-the-art Patched Frame-Of-Reference
Para-PFOR (Section 5.2). Its main features are:
Compression input types and sizes One of the main benefits of this implementation is that implements a compression for unsigned integers. These size and sign
suit the Graph 500 requirements which are “unsigned 64-bit integers”. Therefore it is not needed the pre conversion step in lines 2, 4, 6, 8 of algorithm
5 . The possible integer input types are 32 and 64 bits without sign.
Compression output the algorithm generates a data structure containing information to re-ensemble the data in the other peer.
Included codecs Para-PFOR implementation over CUDA (Section 5.2).
Pros and Cons As an advantage, the code is clean (easy to read) and short. As a
possible downside may be pointed that the transformation of the output structure to an equivalent MPI structure to perform the communication may require
some extra programming effort. However, this coding is mostly complete.
5.4
Integration of the compression
As described in further detail in other parts of this document, the integration of
these compression libraries (which combine a broad spectrum of HPC technologies)
has been designed flexible (“pluglable”) so that it is simple to add a new “library”
from other authors. This library addition would involve some steps like:
11
https://github.com/antonmks/Alenka
Chapter 5: Optimizing data movement by compression
57
1. The 3rd-party compression library need to be copied to the Graph 500 ’s compression/
folder.
2. The “Factory object” must become aware of that a new library. This is done by
selecting a name for the new class within that object and adjusting the input
and output types. For example, it would be correct to copy an already existing
one, choosing a new name, and setting the types.
3. A compression-specific Object for that new library must exist. This Class
contains a wrapper of the new library’s compression & decompression calls so
they match our application’s calls. An specific file, that can be copied and
modified exists for this purpose. In this new created object it is required to
adjust the Class’s name and the new types.
4. Finally, the file containing a mapping between that compression type and its
MPI equivalent needs to be updated.
Other aspects regarding the integration are the low impact of the integration of
the integration on our code. The code performs only one call to create the compression object, which remains in memory. This is performed outside the Second kernel
(BFS + Communication) to minimize its effect on the the overall performance. The
compression object is passed as a memory position to the BFS calls.
Other benefit of this implementation is that the compression library may be updated regardless from the Graph 500 implementation.
5.4.1
Experimental performance comparison
In order to select the compression scheme that best suits our purposes, a comparison table (Table 5.4 and Table 5.5) containing the results of our compression
experiments is presented below. In these tables, the parameters used in the headers
are described below.
• C. ratio % shows the achieved compression-ratio represented as percentile.
The lower the better. The formulat can be seen in expression 10
ratio% =
1
∗ 100
compression_ratio
(10)
• Bits/Integer Resulting bits required to represent the 32-bit integer. This value
has as lower limit the value of entropy (Section 5.1) of the sequence distribution.
The lower the better.
• C Speed MI/s represents the compression speed measured in Millions (106 )
of 32-bit integers processed per second. The higher the better.
58
Chapter 5: Optimizing data movement by compression
• D Speed MI/s analogous to the previous parameter. For decompression speed.
The higher the better.
The aim of generating the data is to be able to gather compression statistics from
each codec and see the one that best suits our needs: we require good decompression
speed (like an IRSs), but also a good compression speed, to not penalize the compression calls. A good compression-ratio is as important as the two previous aspects.
For building the table we will be using a synthetical Zipf integer sequence generator.
The generator is specific for generating sequences of indexes (integers) similar to the
ones appearing in web pages, search engines indexes and databases indexes (Further
discussion about those topics in the next sections of this chapter).
Regarding graphs datasets, they have been study to follow Power law distributions
(like Zipf, pareto or lognormal) [54]. So, that could be a clue of that using a Zipf
generator would give a real insight about the data we intend to compress. Unfortunately, regarding our data, we do not compress the vertices. Instead, we do compress
sparse vectors (SpMV) represented as integers. These integer sequence have some
own numerical characteristicswhich are going to help us to choose the correct codec
for of Frontier Queue compression.
If we could do a summary of what family of codecs to choose depending of the
properties of the data to compress, it would be as follows:
• For compressing an unique integer instead of a sequence, our option would be a
Prefix Suppression (PS) type codec. Whiting the available codecs in this broad
family we would choose the one which better adjust to or latency , speed and
compression-ratio requirements.
• If we had a sequence of integers where the distance from one number to the
other is a-priori known to not be very high, the Frame Of Reference (FOR)
type with a possible delta compression on top would be a good option.
• If the data is an integer sequence, where the numbers are known to be low
(applying this family over high numbers could result counterproductive), then
the Varint family could be a good option,
• If the cardinality of the data to compress is low, for example bitmaps with ’0’s
and ’1’s, a Bitmap Index family would be a good option. Downsides of bitmap
indexes are their low speed and their heavy memory footprint.
• for any other case where we have no knowledge about properties of the data to
be compressed a general Purpose dictionary compression (for example Lempel
Ziv family) would be a good choice.
To gather information about the sequence of integers to be compressed (the SpMVs
from our Frontier queue) we have manually extracted several big buffers (bigger than
Chapter 5: Optimizing data movement by compression
59
20M 12 integers) into files and analysed them afterward with an statistical package
[15]. The result of the tests showed an slightly left-skewed Uniform distribution
(Figure5.2). In terms of compressibility the term Uniform distribution is similar
to randomness. Because random sequences have a high entropy (Section 5.1), the
compression would most likely be bad. Even though the entropy of those sequences
was high: 15-bit out of 32-bit of an integer, we have been able to perform a very high
compression (over 90% reduction) very close to the entropy value. This has been
thanks to the fact of (1) the data was sorted due to our BFS implementation (2)
the gaps in the sorted sequence where small. (3) this kind of sequences have raised
interest due to the Big data + Indexer phenomena and have the support of efficient
algorithms nowadays.
50000
20000
0
0
20000 40000 60000
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0
20000
40000
60000
P−P plot
0
20000 40000 60000
Data
0.4
0.0
0.4
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.8
Empirical and theoretical CDFs
Empirical probabilities
Theoretical quantiles
0.8
Data
0.0
CDF
Q−Q plot
Empirical quantiles
1.0e−05
0.0e+00
Density
Empirical and theoretical dens.
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.2
0.4
0.6
0.8
1.0
Theoretical probabilities
Figure 5.2: Mass distribution of a long SpMVM vector in our Frontier
Queue. The distribution detection has been performed using [15]
Also, as aforementioned we are including in this work a compression based on
GPU (which we believe can provide interesting performance benefits in comparison
to the two ones presented in this work - based on CPU’s SIMD instruction-sets).
12
M stands for Millions
60
Chapter 5: Optimizing data movement by compression
Experiment platform
Prob. Distribution
Empirical Entropy
Skewness
Integer Range
Total integers in sample
Experiment platform
Prob. Distribution
Total integers in sample
Extracted Frontier Queue from our BFS
“Creek”
Uniform (slightly skewed)
14.58845 bit
-0.003576766
[1-65532]
29899
TREC GOV2 Corpus
“Creek”
Power Law
23918M
Table 5.3: Configuration used in both the real sample of a Frontier Queue
transfer and the TREC GOV2 Corpus. Both tests have been performed
in the “Creek” experiment platform (Table 7.1).
Regarding compression implemented on GPUs, it has to be noted than currently,
due to the limitation of a maximum integer size of 32-bit in the GPUs cores’s ALUs,
the use of GPU compression on a graph 500 application would limit the maximum
achievable scale of the BFS to 32 (232 V ertices). The option of using a 64-bit integers
with non native hardware support would penalize the compression. [51].
In order to test compression our chosen Frame of Reference codec in GPGPU we
are configuring our benchmarking tool to use a data corpus 13 used in papers from
other authors, where the algorithm for our CUDA compression (Para-PFOR) is tested
with the same dataset and on a similar platform. Our used system is a node of our
“Creek” cluster (Table 7.1).
The result of this comparison can be seen in Table 5.5. As a first result it can be
noted that the decompression speeds of the Para-PFOR algorithm on a Fermi architecture are faster than any other SIMD techniques implemented on CPU. Regarding
the compression ratio we can observe that it is lower than other more optimized algorithms. The reasoning behind this is that the Para-PFOR scheme sacrifices its
exception management to benefit the overall decompression speed [5].
Our results of this theoretical simulation14 of the CPU and CUDA compressions
are the following.
• the performance of the Para-PFOR using GPGPUs performs better than other
SIMD based algorithms, but lacks of a good compression ratio. The latter
13
TREC GOV2 corpus without sensible data, and valid licence for this work. http://ir.dcs.
gla.ac.uk/test_collections/access_to_data.html
14
Based on a comparison with other existent research [5] on a comparable environment.
Chapter 5: Optimizing data movement by compression
61
disadvantage may be alleviated by using the exception management of other
PFOR codecs.
• the usage of the compression directly in the GPU device may result interesting
because it may minimize the memory footprint of the used data structures of
our BFS (in case of used encoded buffers directly in the card). Unfortunately,
even though we would be able to use more data of the graph in the limited
GPU devices’ memory, we would still be limited to an scale of 32, if we want to
keep the performance of the compression algorithm low (no native 32-bit integer
support on GPU devices [51]).
5.4.2
Observations about experimental performance
As it can be seen in the figures of the previous results, the described above algorithms are focused for their usage in IRSs and therefore designed to achieve a high
compression ratio with high decompression speed (Table 5.4).
The higher importance of the decompression speed over the compression one is
founded in that these systems need to access (decompress) inverted indexes at near
real-time with optimally almost zero latency.
In our integration of these integer sequence compression schemes for a graph 500
application we should focus on the schemes offering a good compression ratio and
balanced good compression ←→ decompression speeds.
In this work three compression libraries have been integrated (two of them partially). Each of them features a main optimised codec. Based on Table 5.4 and Table
5.5 (resulted from comparing results from similar test environments), we will give an
insight on how each codec fits in the Graph 500 application.
1. Turbo-PFOR “Turbo-PFOR” This library has only been partially integrated. The included codec Turbo-PFOR offers a near 50% better compression
ratio than any other in this work. The decompression speed performs as well as
SIMD-FastPFOR (Lemire et al. library). As downside the compression speed
is lower. We believe this scheme may offer promising results.
2. Lemire et al. “S4-BP128” This is the library used and integrated in this
work. Its most competitive codec S4-BP128-D4 (tested in this work) offers both
a good compression ratio and good (and balanced) compression and decompression speeds.
3. Alenka CUDA “Para-PFOR” The partially integrated Alenka library, includes as main algorithm Para-PFOR implemented over CUDA. This codec has
been described, and also theoretically compared here. The comparison has been
Chapter 5: Optimizing data movement by compression
62
made by adapting the data from other works (a dataset15 has been executed
in a system with similar capabilities than the one in the paper). As a result
this codec seems promising because with a close compression ratio to the one
used in this work (S4-BP128 ) it outperforms all codecs in terms of decompression speed 16 . The compression speeds are not listed in the reviewed technical
document and therefore unknown.
C. ratio (%)
45,56
47.15
47.15
47.15
47.15
47.21
56.14
57.86
68.69
68.69
68.69
100.00
-
Bits/Integer
14.58
15.09
15.09
15.09
15.09
15.11
17.96
18.52
21.98
21.98
21.98
32.00
-
C. speed MI/s
3094.58
428.43
1762.95
1762.75
3207.85
182.82
320.88
427.30
587.20
629.42
3399.26
-
D. speed MI/s
5068.97
4694.21
2850.89
2503.91
4701.06
1763.01
1259.19
579.62
381.70
282.95
4535.60
-
Codec
H(x)empirical
VB-TurboBP128
TurboPFOR
TurboBP128
TurboFOR
S4-BP128 (Used)
varint-G8IU
VSEncoding
TurboVbyte
Variable Byte
MaskedVByte
Copy (No C / D)
CUDA ParaPFOR
Table 5.4: Comparison of the compression codecs on the “Creek” experiment platform (Table 7.1). The test uses 64-bit unsigned integers as data
(even though the benchmark implementation also allows 32-bit tests and
uses the latter as reference size). The empathized figures on the table
represent the more optimum values on each column. The use integer size
meets criteria of Graph 500 for the vertexes.
5.4.3
Usage of compression thresholds
By compression thresholds we refer to the minimal size of the integer sequence,
required to trigger the compression call. The motivation behind this concept is that
the compression routines add an extra overhead to the Processor Unit (CPU or GPU,
depending of the used library). This overhead is sometimes not justified in terms of
benefit from the compression.
15
TREC GOV2 corpus without sensible data, and valid licence for this work. http://ir.dcs.
gla.ac.uk/test_collections/access_to_data.html
16
This result has been calculated using theoretical data.
Chapter 5: Optimizing data movement by compression
63
C. ratio (%) Bits/Integer C Speed MI/s D Speed MI/s
Function
14.04
4.49
224.83
742.90
TurboPFOR
14.64
4.68
190.15
548.45
VSimple
16.56
5.30
242.81
396.39
LZ4
19.58
6.27
511.25
641.76
TurboVbyte
21.75
6.96
996.56
1244.29
S4-BP128 (Used)
26.00
8.32
1197.39
1240.31
TurboFOR
28.01
8.96
407.32
358.39
Vbyte FPF
31.87
10.20
135.34
879.72
varint-G8IU
100.00
32.00
1005.10
1005.62
Copy (No C/D)
27.54
N/A
N/A
1300.00 CUDA ParaPFOR
Table 5.5: S4-BP128 and Turbo-PFOR compared to a similar CUDA ParaPFOR setup in an external paper. The data used is the TREC GOV2
Corpus.
This is the case of very short sequence of integers where the time of performing Compression −→ Transmission −→ Decompression is greater than the one
achieved by just Transmission.
Regarding this optimization, we have not noticed significant benefit in its use.
However, it will be discussed in section Future work (Section 9) this technique may
help other improvements.
5.4.4
Communication before and after compression
In this section it will be described the algorithms of the communication with the
integrated compression call. As it was described in the section Graphs 500 Challenge
(Section 2.6.3), the specifications of the Graph 500 benchmark specify Vertex labels of
48-bit unsigned integers. To be compressed, our integer sequences (defined as 64-bit
integer) we have proposed three codecs (i) S4-BP128 (used), (ii) Turbo-PFOR and
(iii) Para-PFOR (over CUDA). The first one operates over 32-bit integers. The last
two codecs operate over 64-bit integers.
To be able to fit the 64-bit integer sequence, on the used compressor (S4-BP128,
32-bit), we have favourably used the fact that our Graph generator did not overpassed
the 32-bit range [0, 232 ), due to the Scale used in the experiments. As a result of this
the 64-bit integers could be pre-converted to 32-bit and post-converted again to 64-bit.
The synthesis of the compression-integration is detailed in the Algorithm
64
Chapter 5: Optimizing data movement by compression
Algorithm 4 Our BFS algorithm with SpVM and compression
Input : s: source vertex id
Output : Π: Predecesor list
f: Frontier queue / Next Queue
t: Current queue
A: Adjacency matrix
Π: Visited node /predecessor list
⊗: SpMV multiplication
: element-wise multiplication
1: f (s) ← s
2: for each processor Pi,j in parallel do
3:
while f 6= ∅ do
4:
T ransposeV ector(fi,j )
0
5:
fi ← compress(fi )
00
0
6:
fi ← ALLGATHERV(fi , P∗,j )
00
7:
fi ← decompress(fi )
8:
ti ← Ai,j ⊗ fi
0
9:
ti ← compress(ti )
00
0
10:
ti,j ← ALLTOALLV(ti , Pi,∗ )
00
11:
ti,j ← decompress(ti,j )
12:
ti,j ← ti,j Πi
13:
Πi,j ← Πi,j + ti,j
14:
fi,j ← ti,j
Algorithm 5 Integration of compression. Detailed
1:
2:
3:
4:
procedure Compression(input, output_compressed64)
input32 ← transform32(input)
output32 ← compress_library_call(input32)
output_compressed64 ← transform64(output32)
5:
6:
7:
8:
procedure Decompression(input_compressed64, output)
input32 ← transform32(input_compressed64)
output32 ← decompress_library_call(input32)
output ← transform64(output32)
Chapter 6
Optimizing instruction overhead
As a second group of optimizations made to our BFS base implementation we will
be testing some optimizations based upon the efficient programming of our algorithm.
In these optimizations we will be covering five areas: (i) scalar optimizations (ii)
vectorization (iii) thread parallelism (iv) memory access, and (v) communication
patterns. For each of these subsections, small code snippets illustrating the optimizations will be given. At the end of the section, it will be included a checkbox table
showing which optimizations have been added. The results are listed in the Final
Ressults section (Section 7).
6.1
Scalar optimizations
Scalar operations / data are those which involve common data types or functions
where the data consist of a single value which can not be operated in parallel. Examples of these in C language are an integer, a char or a data struct, and also any
operation involving the use of any on them.
In this section, some methods to improve the performance with scalar operations
will be discussed. On each technique it will be annotated whether (or not) that
technique has been used in the Graph 500 implementation, and where.
6.1.1
Strength reduction
This technique has been intensively used in our code. By Strength reduction we
cover transformations in the data and functions that avoid either unnecessary value
conversions (or floating point operations), and expensive arithmetical functions by
the usage of its reciprocal. Some sample transformations are listed below using short
code snippets.
1. Usage of the correct value for its specific data type. In C language this would
65
Chapter 6: Optimizing instruction overhead
66
// Bad: 2 is "int"
long b=a*2;
// Good: 2L is "long"
long b=a*2L;
// Bad: overflow
long n=100000*100000;
// Good: correct
long n=100000L*100000L;
// Bad: excessive
float p=6.283185307179586;
// Good: accurate
float p=6.283185f;
// Bad: 2 is "int"
float q=2*p;
// Good: 2.0f is "float"
float q=2.0f*p;
// Bad: 1e9 is "double"
float r=1e9*p;
// Good: 1e9f is "float"
float r=1e9f*p;
// Bad: 1 is "int"
double t=s+1;
// Good: 1.0 is "double"
double t=s+1.0;
Figure 6.1: Strength reduction in variables and constants.
encompass the usage of the suffixes “L”,“U” or “LU” for Longs, Unsigned values
or both the previous.
This will also include the case of the suffix “f” for a float value. By doing
this and keeping a coherence in the arithmetic or floating point operations,
unnecessary conversions are avoided.
2. The replacement of arithmetical operations (with big overhead) by a reciprocal
(with less overhead) will also improve the results in the case of operations in
critical loops. This is the case of replacements of Divisions (slower) by multiplications (faser).
Replacement of expressions with much overhead with cheaper ones
By expensive expressions we mean long operations involving some of the previous division/multiplication or transcendental functions which can be partially precomputed and transformed manually outside a critical loop, and substituted with an
equivalent expression.
Chapter 6: Optimizing instruction overhead
67
// Bad: 3.14 is a double
float x = 3.14;
// Good: 3.14f is a float
float X = 3.14f;
// Bad: sin() is a
// double precision function
float s = sin(x)
// Good: sin() is a
// single precision function
float s = sinf(x)
// Bad: round() takes double
// and returns double
long v = round(X);
// Good: lroundf() takes float
// and returns long
long V = lroundf(x);
// Bad: abs() is not from IML // Good: fabsf() is from IML
// it takes int and returns int// It takes and returns a float
int V = abs(x);
float v = fabsf(X);
Figure 6.2: Strength reduction with funtions
// Common Subexpression Elimination
// change division
for (i = O; i < n; i++)
A[i] /= B;
// Algebraic transformations
// old:
for (i = O; i < n; i++)
P[i] = (Q[i]/R[i])/S[i];
// to multiplication
const float Br = 1.0f/B;
for (i = 0; i < n; i++)
A[i] *= Br;
// new:
for (i = 0; i < n; i++)
P[i] = Q[i]/(R[i]*S[i]);
Figure 6.3: The operations on the right have less overhead, than their left
counterparts.
Chapter 6: Optimizing instruction overhead
68
Algebraic transformations. Usage of equivalent functions with less overhead
The usage of transcendental functions with hardware support instead of their analogous. replacement of exp(), log() or pow(x, -0.5) with exp2(), pow2() or 1.0/sqrt().
This kind of strength reduction can also accelerate the global result when the operation is performed inside a critical loop.
// No hardware support
// Equivalent functions with hardware support
double r = pow(r2, -0.5);
double r = 1.0/sqrt(r2);
double v = exp(x);
double v = exp2(x*1.44269504089);
double y = y0*exp(log(x/x0)*
double y = y0*exp2(1og2(x/xO)*
1og(y1/y0)/1og(x1/x0));
log2(y1/y0)/1og2(x1/xO));
Figure 6.4: Usage of transcendental functions with hardware support
Complexity reduction in loops by the use of caching
This technique, also called loop hoisting or loop invariant code-motion in the
literature can be placed also in the next subsection memory access optimizations.
This technique involves loops consisting of several nested loops. This kind of
transformations are usually detected and performed by the compiler. In some cases
where the invariant variable is a vector, the compiler may skip it.
In this optimization, a memory position (e.g. a vector) indexed with the outer
loop index does not need to be re-accessed in the inner loop. This variable can be
copied to a scalar variable (“cached”) outside the inner loop and therefore prevent
the positioning within the vector.
Efficient bitwise operations
The optimal bitwise transformations have been used intensively in this implementation. Bit-level operations are faster than their higher level analogous. Also, the
usage of optimal methods which prevent branching whilst the bitwise operation is
performed makes the result more optimal.
Control of floating point semantics
These optimizations (also called Precision control) are usually activated in the
compiler and therefore would also fit in the subsection about compiler optimizations.
However, because they enable a balance between precision and speed of floating point
scalars are placed in this section.
Chapter 6: Optimizing instruction overhead
69
The control is done at compile time and allows the selection of multiple levels of
precision. The lower the level of precision in the Arithmetic Logic Unit, the faster the
resulting code will be.
Because the Graph 500 test is not a floating point intensive benchmark, this
implementation, as well as others like Titech’s [51] use the lowest possible precision
mode to gain performance.
6.2
Vectorization
As previously defined in the section Scalar optimizations a variable based upon
multiple data values allows the possibility of being vectorized. This is for example
the case of an array in C.
By saying that the variable can be vectorized we mean that in case some requirements are met, the variable will be processed by a Vector Processing Unit (VPU)
within the CPU.
This will be done with the help of specific Instruction-sets (SSE, AVX, etc in x86
architectures or AltiVec in PowerPC ones). These instruction-sets are an example of
SIMD designs described in Section 2.2.
When a vector is processed in SIMD fashion (by the VPU ), as opposed to when it
is processed as an scalar by the Arithmetic Processing Unit (ALU), the calculations
are accelerated by a factor of the number of values fitting the register. As it can be
seen in the table 6.1, an Intel™ AVX instruction (256-bit width register) operating
over 32-bit integers will enable an assembly instruction to operate over 8 integers at
once.
A way to look at this is, if the code is not using VPUs where it could use them,
it is performing only 18 th of the possible achievable arithmetic performance.
To manage the usage of the SIMD instruction-set there are two possible approaches.
1. The usage of in-line assembly or special functions in higher level languages called
intrinsics.
2. Regular code in a way that allows the compiler to use vector instructions for
those loops. This approach is more portable between architectures (requires a
re-compilation). Also, it is more portable for future hardware with different
instruction-sets.
The vectorization made in this work focuses on the second approach: automaticvectorization by the compiler.
Chapter 6: Optimizing instruction overhead
70
As the concept of aligned buffer is used multiple times in this sections we define
this below.
T *p is n-T aligned if p % n = 0
Where T is the type of the pointer and p contains its memory address.
void main()
{
const int n=8;
int i;
// intel compiler buffer allocation using 256-bit alignment (AVX)
int A[n] __attribute__((a1igned(32)));
int B[n] __attribute__((aligned(32)));
// Initialization. There is scalar, thus not vectorized
for (i=0; i<n; i++)
A[i]=B[i]=i;
// This loop will be auto-vectorized
for (i=0; i<n; i++)
A[i]+=B[i];
}
Figure 6.5: Automatic vectorization of a for-loop on the main() section.
Code for Intel compilers.
Optimization of critical loops
This is one of the main focus on instruction overhead reduction in this work. An
introduction to this technique has done in the Vectorisation section above.
As it has been described, the usage of this technique over code with many loop
iterations and array operations, may get a higher benefit from this. However, some
requirements need to be met for the automatic-vectorisation being performed. These
will be explained later in this section.
The main benefit of the usage of this technique on our application is in the second
Kernel (BFS + Validation), and more specifically on the validation code.
Our validation code is an adaptation of the reference code provided by the Graph
500 1 community, with the addition of a 2D partitioning [51]. The matrixes operations
are performed by the CPU and iterated with non optimized loops which miss the
chance of any possible vectorization.
1
http://www.graph500.org/referencecode
Chapter 6: Optimizing instruction overhead
71
// compiler is hinted to remove data aliasing with the ‘‘restrict’’ keyword.
void vectorizable_function(double * restrict a, double * restrict b)
{
size_t i;
// hint to the compiler indicating that buffers
// has been created with 128-bit alignment (SSE+ in Intel)
double *x = __builtin_assume_aligned(a, 16);
double *y = __builtin_assume_aligned(b, 16);
// for-loop size, is known
for (i = 0; i < SIZE; i++)
{
// vector-only operations
x[i] = y[i + 1];
}
}
Figure 6.6: Automatic vectorization of a buffer containing data dependencies (aliasing). The buffers are used at a different block level (function
call) than they have been declared.
These optimizations based on array vectorisations have been proved to be successful on other state-of-the-art implementations like Ueno et al. [51]. Also, in other
works not directly related with the Graph 500, which also makes intensive usage of
arrays (the compression library [28]) this optimizations have resulted successful.
In order to successfully manage automatic-vectorisation the requirements listed
below must be met.
• The loop must be a for-loop. Other kinds like while will not be vectorised.
• The number of iterations of the loop must be known either in compile or run
time.
• The loop cannot contain complex memory accesses, i.e. memory accesses must
have a regular pattern (ideally by the use a unit stride. e.g. V[n+STRIDE]).
Note that some compilers may manage these issues, however this is not the
general rule.
• The loop must not contain branching (for example if-then-else statements).
• The data to be vectorised must not depend on other index pointing to itself (for
example, a compiler will not vectorise data in the form V[i]= V[i-1]. Sometimes
72
Chapter 6: Optimizing instruction overhead
Instruction Set Year and Intel Processor Vector registers
SSE
1999, Pentium III
128-bit
SSE2
2001, Pentium 4
128-bit
SSE3-SSE4.2
2004 - 2009
128-bit
AVX
2011, Sandy Bridge
256-bit
AVX2
2013, Haswell
256-bit
IMCI
2012, Knights Corner
512-bit
AVX-512
(future) Knights Landing
512-bit
Packed Data Types
32-bit single precision FP
8- to 64-bit integers; SP & DP FP
(additional instructions)
single and double precision FP
integers, additional instructions
32- and 64-bit ints. Intel Xeon Phi
32- and 64-bit ints. SP & DP FP
Table 6.1: SIMD instruction sets in Intel™ x86 architectures. The column
Vector registers shows the width of the Vector processing unit in bits.
This value is the optimum alignment for buffers/ arrays / vectors which
are prone to be vectorized.
even though a dependency may exist in the code, the compiler can be made
aware to ignore it with the use of language keywords, like for example restrict
in C)
In addition to these restrictions which applies to the loop, the compiler must
be aware of this action by the use of a -O flag. Also, the internal structure of
the data vector (this will be explained in further detail later in this chapter) must
be declared “aligned” in memory. Only when the code is programmed accordingly,
the compiler will detect and automatically vectorize the loops. Even though these
improvements may have impact in overall results , due to time constrains, they have
been implemented partially. This is reflected in the section Future work (Section 9).
In other implementations of Graph 500, [51], the vectorization of array structures
executing on the CPU has been performed on detail. For that matter a fixed alignment
of 128-bit (would allow Intel ™ SSE* instructions on x86 architectures) is used for
the general memory data buffers. Also, other data structures which require temporal
locality use a different alignment constants(“CACHE_LINE_ALIGNMENT ”, set to
32-bit for 32-bit integers). These tasks are performed on compile-time by the use of
wrappers over the functions that allocate the memory.
Regularization of vectorization patterns
In auto-vectorisation, even though the compiler detects loops with incorrect alignment, it will try to vectorize them by adding scalar iterations at the beginning (called
“peel”), and also at the end (called “tail”). The rest of the vector will be vectorized
normally. To perform a regularization of the vector a padding is introduced through
a second container. Also, the compiler is made aware of this not required operation
by the usage of a “pragma”.
This optimization could be done in our Graph 500 application to increase the
Chapter 6: Optimizing instruction overhead
73
amount of vectorised loops. These transformations have not been implemented and
are listed here as reference.
Usage of data structures with stride memory access
When a data structure must be designed to solve a problem, sometimes two approaches are possible. Both are illustrated with code snippets below.
// Example of an Array of Structures (AoS) - To avoid
struct AoS { // Elegant, but ineffective data layout
float X, y, Z, q;
};
AoS irregularMemoryAccess[m];
// Example of Struct of Array (SoA). Preferred data structure
struct SoA {
// Data layout permits effective vectorization due to the
// uniform (n-strided) memory access
const
float
float
float
float
};
int m;
* X; //
* y; //
* Z; //
* q; //
Array of m-coordinates of charges
...y-coordinates...
...etc.
These arrays are allocated in the constructor
Figure 6.7: Usage of data structures with stride memory access. Array of
Structures (AoS) over Structure of Arrays (SoA)
6.2.1
Compiler optimizations
To perform the optimizations, one possible strategy is the use of a preprocessor
tool that prepares the build based on selected parameters, or detected features in the
host machine’s hardware.
The selection of build tool was difficult as currently exist multiple, and very similar
options (e.g. cmake, Automake, etc). In our case we selected our decision on the ease
of porting our current Built scripts to other tool. For this matter we chose Automake,
a mature project with active support developed by GNU.
74
Chapter 6: Optimizing instruction overhead
Apart from being able to select options in our build or automatically detect optimal hardware features, it is possible to select good build parameters based on good
practices. Those are exposed below.
Usage of an unique and optimal instruction-set
As it can be extracted from white papers of commercial or open source compilers[42],
a good practise to follow when creating the binary code for an application is to specify the target architecture on which that code will run. This will enable the compile
to select that known applicable software support for the specified hardware. Also
the avoidance of including several instruction-sets which will increase the size of the
generated binary code may impact on the overall performance.
As it has been specified in the introduction of this subsection, by the usage of a
preprocessor tool, several tests are done before the build to detect the host platform.
The assigned architecture will depend on this result.
The Table 6.2 compares the used compiler parameters for different Graph 500
implementations. The explicit usage of a pre checked target architecture may be seen
under the column march in that table.
Inclusion of performance oriented features
Some research has been done on the benefits on applying Inter-procedural optimizations (IPO) to array-intensive applications to improve the general cache hit rate
[66]. By contrast with other implementations like [51] this implementation applies
this technique.
Other performance oriented parameters added to the compiler are the -O and
-pipe flags. In Table 6.2 can be seen a comparison of the different configurations of
common Graph 500 implementations.
Removal of non performance-oriented features
In contrast with the previous compiler parameters, other parameters are known to
reduce the overall performance [39]. The removal of the -fPie flag from our previous
implementation may have influenced in the performance increasement.
Ensure auto-vectorisation is enabled in the compiler
All the requirements listed in the section Vector optimisations are enabled by
making the compiler aware of this. Usually vector optimisations are performed when
the -O level is greater then a value. The value depends on the compiler. For the
specific case of GNU’s gcc compiler this value equals “2” (-O2).
Chapter 6: Optimizing instruction overhead
Implementation Maketool
Graph500 ’s ref.
No
Ours (previous)
No
Ours (new)
automake
Titechs
No
Lonestar-merril
cmake
march CUDA caps. -O
No
No
4
No
No
3
Yes
compile-time 3
No
No
3
No
run-time
3
75
fmath pipe ipo other
Yes
No No
Yes
Yes No -fPie
Yes
Yes Yes
Yes
No Yes
-g
No
No No
-g
Table 6.2: Comparisom of several build parameter for several open
graph500 implementations. Maketool =Build framework (e.g. automake,
cmake, none). march =whether or not the target architecture is detected
at pre-process time. CUDA caps =whether or not CUDA capabilities are detected (and when). -0=default optimization level. fmath =Control of floating point semantics by default. pipe =pipe optimization enabled by default. ipo =Linker-level optimization enabled by default (Inter-procedural
Optimizations). other =other compile flags.
Ensure auto-vectorisation has been successfully enabled
The Compiler reporter is a reporting mechanism that outputs all the successfully
performed optimizations automatically. It is usually activated by a flag.
By checking that the code transformations that have been performed we can
confirm the changes in the generated code.
Due to time restrictions this testing was not fully performed in our Graph 500
application, and the part of Vectorisation is only partially verified.
host% icpc -c code.cc -qopt-report -qopt-report-phase:vec
host% cat vdep.optrpt
LOOP BEGIN at code.cc(4,1)
<Multiversioned v1>
remark #25228: LOOP WAS VECTORIZED
LOOP END
Figure 6.8: Ensure auto-vectorisation has been successfully enabled. Example for an Intel™ compiler
6.3
Thread parallelism
The previusly described optimizations act over the achievable parallelism in the
data layer. In this new layer it is possible to generate serial codes executed concur-
76
Chapter 6: Optimizing instruction overhead
rently by the creation of “threads”. In HPC, this threads make use of the parallelism
that introduces the multiple cores within the processor.
These threads must be used with a parallel Framework. Some example implementations of parallel threads Frameworks are Pthreads, OpenMP, Cilk+, Boost threads.
We will focus this work on OpenMP as it is the most popular and standardised
framework, with the most active support. It implements one of the most common
parallel patterns in HPC, the parallel-for or Map-reduce model. On it, when a for
needs to iterate n items and uses this model, the framework distributes the items by
X.
Concurrent paradigms may increase the overall performance. However, they have
the downside of an added programming complexity. The writes and reads in shared
memory by multiple codes may create “race conditions”.
Put simple a race condition is a non-deterministic state where the multiple processes have own copies of a variable and operate over it. When at some point the
variable is copied to a global position, different threads will change the value and it
will be the value of the slowest process, the one remaining in memory.
In our Graph 500 implementation we were required to constantly test the application when the OpenMP threads was re-adjusted.
For a better understanding of the section Thread contention below, the concept of
synchronisation of threads, and solutions for the race condition issue will be defined
next. The problem of race conditions is usually solved “locking” one or more threads.
This is called synchronization.
6.3.1
Usage of fork-join model when suitable
In contrast with the parallel-for (or map-reduce model), the OpenMP framework
allows also the usage of a very mature model based on Parent and Child processes.
At some cases when it is not possible the usage of a Map-reduce model, it is
recommended to study the possibility of including this one.
6.3.2
Optimal thread scheduling
The threads framework OpenMP, in the Map-reduce model (parallel-for) allows
the implicit specification on how the load of tasks per thread will be balanced. This
is done by specifying a scheduling policy. The three main policies available are listed
below. Also, it is specified the one that we consider the most suitable.
static(n) this is the default policy. Each thread has a fixed number n of tasks to
perform. Even though other threads have already finished their tasks, if one
task in one thread takes longer to run, the other threads must wait.
Chapter 6: Optimizing instruction overhead
77
dynamic(n) this policy assigns dynamically the tasks according to the load in other
threads. One downside of this policy is that threads may not have enough
temporal locality resulting on bad cache hit rates.
guided(n, min) this is the policy chose in our implementation. It mixes both previous scheduling policies. It is basically a dynamic policy with a minimum
number of fixed tasks in the thread. The temporal and spatial locality of the
accessed data may result in a better cache hit rate.
A note about the scheduling policies is that the correct selection is better done
when the execution has been analysed with a profiler.
6.3.3
Thread contention prevention
The threads contention may have two main causes. The first one is an excess of
synchronisation (defined above). The second one is called false sharing and is defined
next.
As it was explained before an address is n-aligned when it is multiple of n. In
the case of false sharing the retrieved data of some threads (even though is aligned)
is partially divided between two cache lines. This can translate into overhead due
to complex memory accesses. In the case of a big-sized critical loop, this alignment
issue may affect the maximum performance benefit resulting from the use of threads.
The issue is solved by adding a padding which corrects the alignment of the vector.
This optimization would be recommended in the critical loops of the validation
code.
// Padding to avoid sharing a cache line between threads
const int paddingBytes = 32; // Intel AVX+
const int paddingElements = paddingBytes / sizeof(int);
const int mPadded = m + (paddingElements-m%paddingElements);
int hist_containers[nThreads][mPadded]; // New container
Figure 6.9: Prevent thread contention by preventing false sharing: Use
of a new padded container.
6.3.4
NUMA control / NUMA aware
In Shared memory systems, platforms with multiple processors have access to
different banks of physical memory. The banks which physically belong to other
Chapter 6: Optimizing instruction overhead
78
processor are called “remote memory” for a given processor. The remote memory
access have a greater latency penalty compare to local memory access (This can be
seen in the Figure 2.3). This same effect happens in multicore systems and each core’s
first level cache.
During the initial assignation of threads in this NUMA model, a thread running
in one processor may be using the memory of a different one. All of this results in
the threads adding overhead to the application. This issue affecting Non Uniform
Memory Access (NUMA) architectures can be alleviated with several approaches.
• Because memory allocation occurs not during initial buffer declaration, but
upon the first write to the buffer (“first touch”), for better performance in
NUMA systems, data can be initialized with the same parallel pattern as during data usage. This approach can be seen on Figure 6.10. By using this technique an initial assignation is done using the threads framework - this prevents
memory-access issues in the threads.
• Other approach is the one used by the current Top 1 implementation in the
Graph 500. They have implemented their own library 2 which allows to “pin”
any data structure from memory to any specific CPU, memory bank o core.
In our work, after the analysis of the results we have included in the section Future
work (Section 9) a revision of our thread scheduling policy, as well as the introduction
of this technique to alleviate the bad performance when the thread execution is active.
// Original initialization loop
for i < M
for j < N
A[i * M + j] = 0.0f
// Loop
#pragma
for i <
for j
A[i
initialization with ‘‘First Touch’’
omp parallel for
M
< N
* M + j] = 0.0f
Figure 6.10: Pseudo-code of allocation on First Touch.
2
ULIBC - https://bitbucket.org/yuichiro_yasui/ulibc
Chapter 6: Optimizing instruction overhead
6.4
79
Memory access
The memory access optimizations performed by alignments and more sequential
structures, (as it was stated previously in the document), will help in the vectorisation
of many loops. In this section a different approach is defined. This technique helps
also with critical loops and may be an advantage in some zones of our second kernel
(BFS + verification).
6.4.1
Locality of data access
The “Rule of Thumb” for memory optimization is the locality of data access in
space and in time.
• By spatial locality we mean a correlative access to the data in memory. This
is managed by a proper election of the used data structures. The use of Struct
of Arrays (SoA) over Arrays of Structs (AoS) improves this situation. The
proper packing of the data to avoid having to iterate in 1-stride fashion at the
beginning and end of the structure (padding the structure) also improves the
locality.
• By temporal locality we mean that the required data at one point is as close
as possible to the one in the next point. We manage this by changing the order
of the operations (e.g., loop tiling, loop fusion or loop hoisting).
6.4.2
Merge subsequent loops
In this operation in different but subsequent loops (with same iterator distance)
may be groped if the flow of the program allows this. This produces that the retrieved memory block to the CPU cache can be reused more close in time. This
transformation can be seen in Figure 6.12.
6.4.3
Variable caching in the outer loop
This optimization (loop hoisting) has been described in the previous section Scalar
Optimizations. It is listed here as it is also a memory-access based optimization. As
previously stated, this optimization has been included in many parts of the code
(both kernels of the Graph 500 application). This transformation can be seen in
Figure 6.14.
Chapter 6: Optimizing instruction overhead
80
// Original code
for (i = 0, i < m, i++)
for (j = 0; ; j < n; j++)
compute(a[i], b[j]); // Memory access is unit-stride in j
// Step 1: strip-mine
for (i = O; i < m; i++)
for (jj = 0; jj < n; j += TILE)
for (j = jj; j < jj + TILE; j++)
compute(a[i], b[j]); // Same order of operation
// as original
// Step 2: permute
for (jj = 0; jj < n; j += TILE)
for (i = O; i < m; i++)
for (j = jj; j < jj + TILE; j++)
compute(a[i], b[j]); // Re-use to j=jj sooner
Figure 6.11: Improve temporal locality by using loop-tiling (Cache blocking)
6.5
Communication patterns
Previously in this section, the optimizations over the multiple data in one instruction was defined as the first layer of parallelism. Next, in the case of the threads which
are generally used in HPC to control the cores inside a processor, it was defined the
second layer of parallelism.
In this section it will be defined a third layer. On it, the “parallel processors” are
each one of the nodes communicating by message passing (MPI).
6.5.1
Data transfer grouping
Small and subsequence data transfers can be prevented by grouping them in a
MPI custom datatype. This improves the communication in two ways: (i) a longer
usage of the network, and hence a better throughput is achieved (ii) the memory
access improves when relaying on the internal MPI packing system (Zero copy).
We have applied these optimizations several times to alleviate the overhead of
successive MPI_broadcasts and MPI_allgatherv for the pre distribution of communication metadata (i.e. buffer sizes in the next transfer)
6.5.2
Reduce communication overhead by overlapping
The usage of this technique implies the continuous execution of tasks during the
delays produced by the communication delay. To perform this optimization an asyn-
Chapter 6: Optimizing instruction overhead
81
// Original code
for (i = 0, i < m, i++)
compute(a[i], b[j]);
for (i = 0, i < m, i++)
secondcompute(a[i], b[j]);
// Loop fusion
for (i = 0, i < m, i++)
compute(a[i], b[j]);
secondcompute(a[i], b[j]);
Figure 6.12: loop fusion
chronous communication pattern, and a double (Send and Receive) buffers must be
used.
This optimization has been proposed as other possible improvement in the communications and is listed un the the section Future work (Section 9).
6.6
Summary of implemented optimizations
In this section we have covered many optimizations, from which only a subset have
been added into our work. Mainly, the added optimizations have been as follows:
1. Scalar optimizations. From this group we have added all of the listed optimizations and covered most of out code.
2. Vectorization. The optimizations covered in this section been the main aim of
our instruction overhead reduction. These optimizations have been added with
more or less success (this will be discussed in the section Final results, Section
7)
3. Thread parallelism. As we have mentioned above, to deal with threads we
have used the OpenMP framework. When first analyzing our Baseline implementation we noticed a paradoxical performance loss when using threads
compared to the 1 single thread case. This problem has been constant during
all the development of our new 2 versions. In the last version we have achieved
to slightly increase the performance of the OpenMP version over the 1-threaded
one.
4. Memory access. The memory access optimizations usually require the refactoring of large portions of code. For this reason sometimes it becomes more
82
Chapter 6: Optimizing instruction overhead
// Original code
for (i = 0, i < m, i++)
for (j = 0, j < n, j++)
for (k = 0, k < p, k++)
// unnecessary and multiple access
// to a memory block with complexity
// O(m*n*k)
compute1(a[i][i]);
compute2(a[i][i]);
// New code
for (i = 0, i < m, i++)
const int z = a[i][i];
for (j = 0, j < n, j++)
for (k = 0, k < p, k++)
compute1(z);
compute2(z);
Figure 6.13: variable caching to provide better temporal and spatial locality (loop hoisting)
difficult the optimization of an already created code the its new reimplementation. Hence, we have only applied a few of this optimizations to the new
structures that have been required to implement.
5. Communication patterns. Among these optimizations we have only covered the regrouping of individual transfers performed serially into a custom
MPI_Datatype which would achieve in one transmission a bigger amount of
data and a bigger use of the network resources. This modifications have been
made majorly for the broadcasted metadata prior to other communications.
Regarding the communication overlapping it has successfully proved to be a
successful addition to a Graph 500 application [60, 51] (Section 3) Despite of
this, we have not implemented it in our code.
Chapter 6: Optimizing instruction overhead
// Non optimized communication
if (r == 0) {
// Overlapping comp/comm
for(int i=0; i<size; ++i) {
arr[i] = compute(arr, size);
}
MPI_Send(arr, size, MPI_DOUBLE, 1, 99, comm); } else {
MPI_Recv(arr, size, MPI_DOUBLE, 0, 99, comm, &stat); }
Figure 6.14: No communication overlapped with computation
// Optimized communication. Software pipeline
if (r == 0) {
MPI_Request req=MPI_REQUEST_NULL:
for(int b=0; b < nblocks; ++b) {
if(b) {
// pipelining
if(req != MPI_REQUEST_NULL) MPI_Wait(&req, &stat);
MPI_Isend(&arr[(b-1)*bs], bs, MPI_DOUBLE, 1, 99, comm, &req); }
for(int i=b*bs; i<(b+1)*bs; ++i) {
arr[i] = compute(arr, size);
}
}
MPI_Send(&arr[(nblocks-1)*bs], bs, MPI_DOUBLE, 1, 99, comm);
}
else {
for(int b=0; b<nblocks; ++b) {
MPI_Recv(&arr[b*bs], bs, MPI_DOUBLE, 0, 99, comm, &stat);
}
}
Figure 6.15: No communication overlapped with computation
83
Chapter 7
Final Results
In this section we present the results from our experiments and try to solve the
problem detected in the previous section Analysis (Section 4.3). In this same section
we will discuss the results and set the initial thought for the next sections, Conclusions
and Future work.
The chapter starts with data related with the systems used for testing. One is
allocated in the Georgia Institute of Technology; other, is in the Rupretcht-Karls Universität Heidelberg.
The results will be grouped depending on the type of performed optimization: (i)
communication compression or (ii) instruction overhead optimizations.
7.1
Experiment platforms
The Tables 7.1 and 7.3 show the main details of the used architectures. The
big cluster (Keeneland) has been used to perform main tests. The development and
analysis of the compression has been performed in the smaller one.
A third table 7.2 shows the specifications of the platform (also allocated in University of Heidelberg) used to test the overhead added by the compression. The
experiments performed on this third platform have not been included as they have
only been concluded partially. Even though, some conclusions of the executions on
this machine are are added in this chapter. The table is therefore added, as a reference
if needed.
7.2
Results
For this work we are going to focus in three main points of our development. This
will help us to identify our two main goals: compression and instruction overhead.
They are described below.
84
Chapter 7: Final Results
Operating System
CUDA Version
MPI Version
System
Number of Nodes
Number of CPUs / node
Number of GPUs / node
Processor
Number of cores
Number of threads
L1 cache size
L2 cache size
L3 cache size
Memory
Max. vector register width
PCIe Speed (CPU ←→ GPU)
QPI Speed (CPU ←→ CPU)
GPU Device
Cores
Total system Cores
Compute capability
Memory
Memory Bandwidth
Interconnect
Rate
85
64-bit Ubuntu 12 with kernel 3.8.0
7.0
1.8.3 (OpenMPI)
SMP Cluster
8
1
2
Intel E5-1620 @ 3.6 GHz
4
2
4 x 64 KB
4 x 256 KB
1 x 10 MB
4 GB (DDR3-1333), ECC disabled
256-bit, Intel AVX
5 GT/s
none
NVIDIA GTX 480 (Fermi architecture)
15 (SMs) x 32 (Cores/SM) = 480 (Cores)
7680 (Cores)
2.0
1.5 GB, ECC disabled
177.4 GB/s
GigabitEthernet (Slow network fabric)
1 Gbit/sec
Table 7.1: Experiment platform “Creek”.
1. Baseline was our first BFS implementation. It its studied in the section Analysis. The Baseline implementation does not compress data and is a good scenario
to study traditional overhead reduction on a graph500 application
2. No-compression is our latest version without compression. The later has
been disabled at compile time and introduces zero overhead. Measuring the
difference in performance between this and the previous one (none of them have
compression capabilities) is a good chance to measure the instruction-oriented
improvements.
3. Compression same code the prior one. It has its compression capabilities
enabled.
Here follows a description of the diagrams listed below. The diagrams have been
generated from the trace files resulting from the executions. The three main types of
graphs shown here are (i) strong scaling diagrams (ii) weak scaling (iii) time breakdowns. To know what conclusions we are getting from each type of graph they will
be described.
1. Strong scaling keeps some value fixed and shows evolution over other parameter. To explain this lets say we use in our system two variables. We can increase
Chapter 7: Final Results
86
Operating System
CUDA Version
MPI Version(s)
System
Number of Nodes
Number of CPUs / node
Number of GPUs / node
Processor
Number of cores
Number of threads
L1 cache size
L2 cache size
L3 cache size
Memory
Max. vector register width
PCIe Speed (CPU ←→ GPU)
QPI Speed (CPU ←→ CPU)
GPU Device
Cores
Total system Cores
Compute capability
Memory
Memory Bandwidth
Interconnect
Rate
64-bit Ubuntu 14 with kernel 3.19.0
7.0
1.8.3 (OpenMPI)
SMP System
1
2
16
Intel E5-2667 @ 3.2 GHz
8
1
8 x 64 KB
8 x 256 KB
1 x 20 MB
251 GB (DDR3-1333), ECC disabled
256-bit, Intel AVX
8 GT/s
8 GT/s
NVIDIA K80 (Kepler architecture)
13 (SMX s) x 192 (Cores/SMX ) = 2496 (Cores)
39936 (Cores)
3.7
12 GB, ECC disabled
480 GB/s
Internal PCIe 3.0 bus
PCIe’s rate
Table 7.2: Experiments platform “Victoria”.
the (1) processors or the (2) size of our problem. In Strong scaling we set one
with a fixed value. The we see the evolution of the other one. This kind of
diagrams let us now how the system behaves with a big peak of load. Lets set
this analogy: I have a computing device and a program running that is taking
more load each time. The strong scaling shows me a “vertical scalability”, i.e.
whether my system is going to function without getting blocked. In this kind
of diagrams we are looking forward to see a ascending line, which tell us that
the system can manage the load.
2. Weak scaling Analogous to the previous one, we may vary both of the parameters. In such a way that the x-axis stays constant to that increasement: the
ratio of the two increased variables stays constant. We can see this as an “horizontal scaling” where we can see how our system behaves against many parallel
tasks. We can see this with an easy example. This would be the scalability of
a “modern cloud” when thousand of millions of clients access at the same time.
The previous scaling would be the behaviour of an individual computer system
agains a very high load. When we look at this type of diagrams we expect to
see a horizontal line that shows that our system is capable to manage the load
and stay even.
3. In the time breakdown we will chunk into pieces the our application and
visualize through bar-plots the times of each region. This will give insights of
Chapter 7: Final Results
Operating System
CUDA Version
MPI Version(s)
System
Number of Nodes
Number of CPUs / node
Number of GPUs / node
Processor
Number of cores
Number of threads
L1 cache size
L2 cache size
L3 cache size
Memory
Max. vector register width
PCIe Speed (CPU ←→ GPU)
QPI Speed (CPU ←→ CPU)
GPU Device
Cores
Total system Cores
Compute capability
Memory
Memory Bandwidth
Interconnect
Rate
87
64-bit Ubuntu 14
7.0
1.6.4 (OpenMPI)
SMP Cluster
120 (used 64)
2
3 (used 1)
Intel E5-2667 @ 2.8 GHz
6
2
6 x 64 KB
6 x 256 KB
1 x 12 MB
251 GB (DDR3), ECC disabled
128-bit, Intel SSE4
5 GT/s
6.4 GT/s
NVIDIA M2090 (Fermi architecture)
16 (SMs) x 32 (Cores/SM) = 512 (Cores)
32768 (Cores)
2.0
6 GB, ECC disabled
178 GB/s
Infiniband (Fast network fabric)
40 Gb/sec (4X QDR)
Table 7.3: Experiments platform “Keeneland”
1
1
http://keeneland.gatech.edu/
the bottlenecks of the application.
The next part of the chapter will be an individual exam of each of the figures.
The analysis will be done afterwards, so at this point just some clarifications will be
given. The figures listed below are (A) 3 Strong scaling analysis (B) two weak scaling
figures (C) two time breakdown diagrams.
A. Strong scaling diagrams
(a) Figure 7.1 (a) shows the behaviour of the application for each one of
the three scenarios on setting a high scale factor fixed and observing the
evolution at each number of processors. In this diagram we prefer a high
stepped ascending line.
(b) Figure 7.1 (b) is analogous to the previous case. The Baseline is not
shown as we currently don’t have data of times at these time (we do have
of TEPS). The measured parameter here is the evolution of time at different processor-number scenarios. In this case we would prefer a stepped
line again. However, because we are measuring time we want this to be
descending.
Chapter 7: Final Results
88
(c) Figure 7.1 (c) shows an different picture of what is happening. We set
the number of processor fixed (to one of the highest our application can
reach). Then we observe the results for different problem sizes. Again, we
look forward a stepped ascending line.
B. Weak scaling diagrams
(a) Figure 7.2 (a) and 7.2 (b) show both, the variation of the number of
processors and the variation of the problem size in the same axis (X). The
scaling is used in logarithmic scale. We do not use a constant number
of vertices or edges per increased number of processors. This is due to
limitations in our experiments and the fact that since we are using a 2D
partitioning of the graph, the required number of processors grows exponentially. Even-though we do not keep the ratio constant we approximate
this thanks to the logarithmic increasement of the scale. Regarding the
diagrams we measure both Time and TEPS. Again, we would prefer a
plain horizontal line as result. As it has been previously noted we are not
keeping our ratio constant, so son ascending skewness will be normal.
C. Time breakdown diagrams
(a) Figure 7.3 (a) and 7.3 (b) will let us see one of the main goals of this
work: the effect compression directly over each zone of the application
On the next section we will analyse these results and prepare the Conclusion in
the next chapter.
7.2.1
Scalability analysis
To get conclusions about scalability we are first choosing the Compression scenario
and the No compression one. We will be leaving the Baseline for the next subsection. Regarding the diagrams, we will be looking at the Strong and Weak scalability
diagrams (the time breakdown will be used deduct conclusions about compression).
As we explained at the beginning of this chapter in the strong types of diagrams
we will prefer stepped ascending lines (for TEPS) and descending (for Times. The
more degree in the line, the best. Looking at Figures 7.1 (a) and 7.1 (b) we see the
stepped lines. The scenarios shows a better capability to perform high in the case of
compression when compared to no compression. This, regardless of being the main
BFS algorithm and the state-of-the-art improvements included on this, the one that
will let grow more stepped in the diagram.
We are going to start noticing an important behaviour of our application, from
now own so it is going to be now when it is going to be described. We will notice that
some odd scales (and also number of processors. The explanation follows) produce a
Chapter 7: Final Results
89
Scale 24 − Strong scaling
9
e+
0
42
1.
08
e+
49
9.
4.
7
5e
+0
8
Traversed Edges per Second (TEPS)
1.
90
e+
09
Compression
No−compression
Baseline
gpus:16
gpus:25
gpus:36
gpus:49
Number of Processors
(a) The three scenarios in strong scaling.
Varying problem size. Fixed processors to 49
Time − fixed scale of 24
74
2.
08
1.
85
e−
8.
01
57
e+
08
1.
24
e+
09
Total time (s)
e−
01
1.
63
e
+0
9
3.
63
e−
01
Compression
No−compression
02
e−
scale:17
scale:19
scale:22
scale:24
62
scale:15
scale:25
gpus:16
9.
8.
59
e+
07
4.
72
e+
Traversed Edges per Second (TEPS)
2.
01
e+
09
Compression
No−compression
Problem size (Scale factor)
gpus:25
gpus:36
gpus:64
Number of Processors
(b) diagram showing Time in strong scaling.
(c) Number of processors stays fixed.
Figure 7.1
Chapter 7: Final Results
90
Time − Weak scaling
TEPS − Weak scaling
e+
09
2.
0
09
1.
5e
+
08
4e
+
9.
Traversed Edges per Second (TEPS) per Processor
Compression
No−compression
Baseline
+0
8
scale:23
gpus:25
scale:24
gpus:36
0e
scale:22
gpus:16
scale:25
gpus:49
4.
5.
9e
−0
2
1.
6e
−
01
2.
6e
−
01
Total time (s) per Processor
3.
7
e−
01
Compression
No−compression
scale:22
gpus:16
scale:23
gpus:25
Scale factor per GPU
scale:24
gpus:36
scale:25
gpus:49
Scale factor per GPU
(a) Weak scaling. Times
(b) Weak scaling. TEPS
Figure 7.2
0.40
0.45
Expansion
Row com
Col com
Pred list red
Total time
0.30
0.25
Total time (s)
0.20
0.15
0.15
0.20
0.25
0.35
0.30
0.35
Expansion
Row com
Col com
Pred list red
Total time
0.00
0.00
0.05
0.05
0.10
0.10
Total time (s)
Time breakdown − Compression − Scale 24
0.50
0.40
Time breakdown − No compression − Scale 24
gpus:16
gpus:25
gpus:36
gpus:64
gpus:16
Number of processors
gpus:25
gpus:36
gpus:64
Number of processors
(a) Time breakdown. Compression disabled
(b) Time breakdown. Compression enabled
Figure 7.3
Chapter 7: Final Results
91
very stepped low peak (in the case of TEPS) or a ascending peak in the case of Times.
The reason for that lays on the management of odd units (processor ranks or matrixes
with odd number of elements). We are partitioning our symmetric NxN matrix on a
2D fashion. For doing this some ranks of the initial MPI_COMM_WORLD will be designed
to columns and some others to rows. Then, and from now on the communication
happens using these rank partitions. The problem arises when we have to divide
by 2 the rank and the number is odd. Some nodes the number of ranks on each
node will be different. In our application we make use of a custom MPI_Allreduce
algorithm [48] which makes use of Point to Point MPI calls for communication. When
that implementation is taken to our application and we have to deal with the odd
ranks, we have to create the concept of “residuum” to deal with this. Within our
application we make use of residuums in two points: (i) in the column communication
(ii) in the predecessor list reduction step which occurs at the end of each BFS iteration
implementations. The residuum approach has two mayor downsides. In a similar way
matrixes where Nx N is an odd number adds complexity when dividing the Matrix
in Row and columns (the matrix are sparse-binary and operated through bitwise
operations)
1. Complexity. The residuum is performed through the addition of many extra
loops, branches, conditions, MPI difference cases, special cases when dealing
with the bitmaps, more code in the Matrix structure class to deal with this
problem, and so on and so forth. All this sums up and makes our implementation
very complex and prone to crashes. In fact, we know that some scales (which
turn out to be odd) will not run without crashes. (scales 17 and 25).
2. Low performance The two places where these residuums are located are the
column and predecessor reduction steps. In the case of the latter, due to the
mechanism to deal with the residuum we can not add compression. As a result
the transmitted data in this phase, which takes time only once during the BFS
iteration, sums more transmitted data (11 times in for 216 Edges) than the
whole transmitted (compressed) during the rest of the iteration.
We see a good scalability Strong and Weak fashions for the tested data, using
compression.
7.2.2
Overhead of the added compression
To measure the added overhead by integrating compression 1 , we performed tests
on a third system.
1
our BFS implementation is a proof-of-the-concept implementation to test compression using 3rd
party compression libraries and codecs.
92
Chapter 7: Final Results
Briefly this is a multiGPU platform with a NUMA architecture: two processors.
Each of the processors has attached 8x NVIDIA K80 GPUs. For this experiment
we expect that the GPU-BFS kernel uses direct card-to-card communication using
the network device (RDMA), avoiding that way the cpu-memory path. As our compression is allocated in the CPU execution layer (we even use the CPU vectorization
to introduce data-parallelism), we are going to be able to separate the execution of
the BFS (GPU+GRAM+PCIe3+Network device+RDMA) with the rest of the computation, buffer management, BFS queues operated with the CPU, and lastly the
compression calls. This latter operations will make use of the 2 CPUS, 2xRAM (one
allocated to each processor), and a QPI high speed bus as interconnection.
Because the speed of this setup is very high and the latencies of its “network”
(will be mostly bus communication: PCIe 3 and QPI) leave not much space to gain
speed by compressing data, we are dealing with the perfect scenario to test how much
latency are we adding. First of all to understand the introduced overhead, they are
listed and explained in further detail below.
1. the obvious computational cost of the compression (which is pretty low in the
chosen compression scheme, see section 5)
2. the data conversion before and after compression to adapt it to the requirements of the 3rd party compression libraries. This data conversions have a high
memory footprint cost which may be greater than the computational cost of
the compression.
3. NUMA architectures incur in an extra cost as their processors may access banks
of memory allocated in other processor. The cost of these accesses is expensive
[60]. This add up to the expensive memory footprints from the previous buffer
copies.
4. Due to the fact of this being a compression codecs test-bench, the compression
libraries have been integrated modularily. Unfortunately, many times “easy to
use” equals low performance. The point being that the “modularity” has been
implemented in C++ using a kind of hierarchy called (virtual keyword) which
allows to choose the inherited class of other class at runtime. This prevents
the compiler from knowing what class will be used, and to inline it. Each
compression call using the Lemire et al. library incurs in two virtual call (the
virtual keyword is also used in Lemire et al.).
As expected the overhead has been detected in the previous implementation. Also,
we have noticed performance loss in other situation: In communication between MPI
ranks using shared memory. The performance which compressions in these cases is
lower to the one between ranks allocated in different machines (rank communicating
through RDMA do not enter here)
Chapter 7: Final Results
93
The problem could be solved easily without many changes with the use of the
implemented compression threshold (Section 5). At the beginning of the application
it could be implemented a quick MPI_COMM_WORLD ranks test to detect low
compression performance between ranks. We give more details aboit this idea in the
Future work section (Section 9)
7.2.3
Instruction overhead analysis
In this section, we will compare our Baseline implementation with the transformed application (leaving off the compression capabilities).
We have focused the optimizations on Vectorization. We intend to make use of
the Vector Processor Unit (VPU) included in the CPU of modern systems. As an
example, a x86 system with SSE™ a 128-bit width VPU, would be able to operate
with 4x 32-bit integers at a time. The performance boost is 4X.
Other aspects we have focused are the efficient memory access, and thread parallelism.
To see the successfulness of our optimizations we will be comparing also the scaling
diagrams. This time we will be focusing on the Baseline and the No-compression
versions.
From what it can be seen in the results the No-compress versions outperforms
the Baseline in all scenarios (low and high scales , few and many processors). Even
thought these results show success in these changes, it is difficult to assert what
changes produced the better (or even some) benefit.
7.2.4
Compression analysis
Compression is the main aim of this work and for better gathering conclusions from
the results we will be using other sources of data, apart from the diagrams generated
in this experiments. The diagrams responsible from giving us information about the
successfulness (or not) of the compressed data movement are the time breakdowns
(Figures 7.3 (a), 7.3 (b)). The zones where the compression could be applied are
three: (i) row communication, (ii) column communication and (iii) predecessor list
reduction.
Zones (i), and (ii) have been made capable of compressing their data. For the case
of (iii) several reasons have made it difficult to have it compressed in this work. For
that matter, we present an algorithm for performing that step with compressed data
transfers.
The efficiency of the compression has also been measured using a more accurate
procedure in the cluster Creek, in University of Heidelberg. The downside, perhaps
of using this cluster is that its only 16 GPUs do not allow profiling with big scale
transfers.
Chapter 7: Final Results
94
Next follow the tables generated from the instrumentation results using the profiler.
Initial Data (Bytes) Compressed data (Bytes) Reduction (%)
Vertex Broadcast
8192
8192
0,0%
Predecessor reduction
7.457.627.888
7.457.627.888
0,0%
Column communication 7.160.177.440
610.980.576
91,46%
Row communication
4.904.056.832
403.202.512
91,77%
Table 7.4: Reduction with compression in terms of data volume. Measurement with Scalasca and ScoreP. Experiment on Creck platform, scale
22, 16 GPUs
Vertex Broadcast
Predecessor reduction
Compressed column communication
Compressed row communication
Initial time (s) Compression times (s) Reduction (%)
0.162604
0.159458
0,0%
164.169409
154.119694
0,0%
156.615386
31.035062
80,18%
79.772907
13.497788
83,07%
Table 7.5: Reduction with compression in terms of data volume. Measurement with Scalasca and ScoreP. Experiment on Creck platform, scale
22, 16 GPUs
Chapter 8
Conclusions
The conclusions of this work are based on the subsection Analysis in the previous
section Results (Sections 4.3.1 and 4.3.2).
In this work we have reviewed (i) how does a compression based on “Binary
Packing” (Section 5) improves a distributed BFS algorithm (ii) The effects of applying
the traditional instruction optimizations to the same distributed algorithm.
Regarding whether or not have we successfully met our goal set in chapter 4 of
improving the overall performance of the Baseline implementation it could be stated
that yes. For the analyzed scenario, and using a relatively high scale and a relatively
high number of processors we have significantly increased
1. the “horizontal” scalability: our distributed BFS distributes better the load
among the nodes. This is an indirect effect of improving the communication
latency.
2. the “vertical” scalability: our algorithm can deal with bigger amounts of
load on an unique processor.
(a) the direct reason for this are the instruction overhead optimizations
(vectorization, memory access and thread parallelism). For the particular
case seen on Figure 7.1 (a), the direct improvement is constant and stays
over the 100% (relation between non compressed and baseline).
(b) the indirect reason for this to happen is the previous optimization in the
compression: since the bottleneck in the algorithm is the communication,
each of the processors stay still a certain amount of time awaiting to receive
the Frontier Queues from its neighbours. The Figure 7.1 (a) in previous
chapter shows a 200% improvement of the compression implementation
in relation to the baseline.
95
Chapter 9
Future work
We believe that some good steps to follow in order to develop a good Graph 500
implementation would be:
A In case we want to run our BFS on GPGPU
(a) Use the Graph 500 implementation SC12 (Ueno et al.) [51] for the BFS
kernel. This implementations was already top 1 in the list in the year 2013.
(Section 3).
(b) The previous implementation already compress its compression. their used
scheme named Variable-Length Quantity VLQ belongs to the Varint family (Section 5). Two characteristics of varint codecs are (i) that they do not
perform well with big cardinalities, (ii) are slower and have smaller compression ration than a FOR + delta compression scheme. This latter about
the FOR compression is also only valid if the data meets some criteria (as
the varints codecs). The criteria for FOR codecs is ordered sequences of integers (unordered is also possible but requires an initial sorting) with low
distances between them. Since the Frontier Queue in the sparse graphs
that we are studying meet both criteria, this codec is ideal.
(c) Our last step would be to add the NUMA-control1 library that has given
the position #1 to Fujisawa et al. to deal with the memory movements in
our application (including the ones in the compression integration)
B In case we want to run our BFS on a NUMA cluster
(a) As, to our knowledge, there is no open state-of-the-art Graph500 implementation, we would need to develop it based on the paper of the Top 1
implementation [60].
1
ULIBC - https://bitbucket.org/yuichiro_yasui/ulibc
96
Chapter 9: Future work
97
(b) Since the current Top 1 does not perform compressed movement of data,
and since we also know that the algorithm includes SpMV, CSR and 2D
graph partitioning, it would be a good fit for the compression scheme
presented here.
(c) The last step would also be to add the NUMA control library that they
feature to the memory structures in our Compression integration
As we partially described in section 7 when dealing with the problem of the
compression-added overhead at some cases, a solution would be to learn about the
ranks in the topology and know whether or not apply compression. A more optimum
approach for the second case described in that section (nodes with shared memory,
there the compression will not have any benefit, but only overhead) would be instead
of (1) add an initial discovery cycle (2) broadcast results (3) manage structures. we
could do the inverse way: since the 2D partitioning allows manually to create the
partitioning for which nodes would go in the row column and the same for the row,
we could keep the initial steps: (1) initial discovery of the topology (2) broadcast of
the data, and now (3) would be: setup the application rank based on the map: more
amount of communication −→ different MPI rank group (rows or columns)
Other proposed work related with compression, but in a different scope, would
encompass the use a Lempel Ziv fast compression algorithm to compress the visited neighborhood bitmap. This has been proved to be successful in other previous
work (Huiwei-et-al [31]) and results on a great communications transfer reduction.
However, as our bitmap transfer is small (as proven in instrumentation results, Section 4.2) the implementation-cost vs performance-benefit balance must be previously
evaluated.
The usage of an Hybrid (MPI + OpenMP) model for parallelizing the collective
operations using spare cores, shows an experimental speedup of 5X in other previous
works in real world applications [32]. The revision of the dynamic task scheduling of
our OpenMP thread model may also add an extra performance.
One feature wich differes with many others is our implementation is a custom
MPI_Allreduce function that we is used to reduce the column communication. The
implementation is based on the Rabensefner-et-al algothithm [48] which is dated in
2005. As it was described in the section 7, this algorithm uses some point-to-point
communication calls which make neccessary the creation of an special case called
residuum for some scenarios where |Rows x Columns| is odd (Section 7). The usage
of a more new algorithm, with the possibility of being implemented over GPGPU
may be a good improvement, as done in the work of Faraji-et-al [19].
Another possible addition to our implementation is the new acceleration framework OpenACC2 . It has been successfully proved to work for other high computing
2
http://www.openacc.org/
Chapter 9: Future work
98
problems [8], and the results may be interesting on Graph5003 .
3
http://www.graph500.org/
References
[1] N. R. Adiga, M. A. Blumrich, D. Chen, P. Coteus, A. Gara, M. E. Giampapa, P. Heidelberger, S. Singh, B. D. Steinmacher-Burow, T. Takken, M. Tsao, and P. Vranas.
Blue gene/l torus interconnection network. IBM J. Res. Dev., 49(2):265–276, March
2005.
[2] Vo Ngoc Anh and Alistair Moffat. Inverted index compression using word-aligned
binary codes. Inf. Retr., 8(1):151–166, January 2005.
[3] Vo Ngoc Anh and Alistair Moffat. Index compression using 64-bit words. Softw. Pract.
Exper., 40(2):131–147, February 2010.
[4] Vo Ngoc Anh and Alistair Moffat. Index compression using 64-bit words. Softw. Pract.
Exper., 40(2):131–147, February 2010.
[5] Naiyong Ao, Fan Zhang, Di Wu, Douglas S. Stones, Gang Wang, Xiaoguang Liu,
Jing Liu, and Sheng Lin. Efficient parallel lists intersection and index compression
algorithms using graphics processing units. Proc. VLDB Endow., 4(8):470–481, May
2011.
[6] Scott Beamer, Krste Asanovic, David A. Patterson, Scott Beamer, and David Patterson. Searching for a parent instead of fighting over children: A fast breadth-first search
implementation for graph500. Technical report.
[7] Scott Beamer, Aydın Buluc, Krste Asanovic, and David A. Patterson. Distributed
memory breadth-first search revisited: Enabling bottom-up search. Technical Report
UCB/EECS-2013-2, EECS Department, University of California, Berkeley, Jan 2013.
[8] Stu Blair, Carl Albing, Alexander Grund, and Andreas Jocksch. Accelerating an mpi
lattice boltzmann code using openacc. In Proceedings of the Second Workshop on
Accelerator Programming Using Directives, WACCPD ’15, pages 3:1–3:9, New York,
NY, USA, 2015. ACM.
[9] M. Burtscher, R. Nasre, and K. Pingali. A quantitative study of irregular programs
on gpus. In Workload Characterization (IISWC), 2012 IEEE International Symposium
on, pages 141–151, Nov 2012.
[10] Chris Calabro. The Exponential Complexity of Satisfiability Problems. PhD thesis, La
Jolla, CA, USA, 2009. AAI3369625.
99
100
References
[11] Matteo Catena, Craig Macdonald, and Iadh Ounis. On inverted index compression for
search engine efficiency. Lecture Notes in Computer Science, 8416:359–371, 2014.
[12] Fabio Checconi, Fabrizio Petrini, Jeremiah Willcock, Andrew Lumsdaine, Anamitra Roy Choudhury, and Yogish Sabharwal. Breaking the speed and scalability barriers
for graph exploration on distributed-memory machines. In Proceedings of the International Conference on High Performance Computing, Networking, Storage and Analysis,
SC ’12, pages 13:1–13:12, Los Alamitos, CA, USA, 2012. IEEE Computer Society Press.
[13] Jeffrey Dean. Challenges in building large-scale information retrieval systems: Invited
talk. In Proceedings of the Second ACM International Conference on Web Search and
Data Mining, WSDM ’09, pages 1–1, New York, NY, USA, 2009. ACM.
[14] Renaud Delbru, Stephane Campinas, and Giovanni Tummarello. Searching web data:
an entity retrieval and high-performance indexing model. Web Semantics: Science,
Services and Agents on the World Wide Web, 10(0), 2012.
[15] Marie Laure Delignette-Muller and Christophe Dutang. fitdistrplus: An R package for
fitting distributions. Journal of Statistical Software, 64(4):1–34, 2015.
[16] Jean-Paul Deveaux, Andrew Rau-Chaplin, and Norbert Zeh. Adaptive tuple differential
coding. In Database and Expert Systems Applications, 18th International Conference,
DEXA 2007, Regensburg, Germany, September 3-7, 2007, Proceedings, pages 109–119,
2007.
[17] J. Dongarra, T. Sterling, H. Simon, and E. Strohmaier. High-performance computing:
clusters, constellations, mpps, and future directions. Computing in Science Engineering, 7(2):51–59, March 2005.
[18] Stefan Evert and Marco Baroni. zipfR: Word frequency distributions in R. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics,
Posters and Demonstrations Sessions, pages 29–32, Prague, Czech Republic, 2007. (R
package version 0.6-6 of 2012-04-03).
[19] Iman Faraji and Ahmad Afsahi. Gpu-aware intranode mpi_allreduce. In Proceedings of
the 21st European MPI Users’ Group Meeting, EuroMPI/ASIA ’14, pages 45:45–45:50,
New York, NY, USA, 2014. ACM.
[20] Michael J. Flynn and Kevin W. Rudd. Parallel architectures. ACM Comput. Surv.,
28(1):67–70, March 1996.
[21] Zhisong Fu, Michael Personick, and Bryan Thompson. Mapgraph: A high level api
for fast development of high performance graph analytics on gpus. In Proceedings of
Workshop on GRAph Data Management Experiences and Systems, GRADES’14, pages
2:1–2:6, New York, NY, USA, 2014. ACM.
References
101
[22] Erich Gamma, Richard Helm, Ralph Johnson, and John Vlissides. Design Patterns:
Elements of Reusable Object-oriented Software. Addison-Wesley Longman Publishing
Co., Inc., Boston, MA, USA, 1995.
[23] Veluchamy Glory and Sandanam Domnic. Compressing inverted index using optimal
fastpfor. JIP, 23(2):185–191, 2015.
[24] J. Goldstein, R. Ramakrishnan, and U. Shaft. Compressing relations and indexes. In
Data Engineering, 1998. Proceedings., 14th International Conference on, pages 370–
379, Feb 1998.
[25] Matthias Hauck. Scalable breadth-first search using distributed gpus. Master’s thesis,
University of Heidelberg, 2014.
[26] John L. Hennessy and David A. Patterson. Computer Architecture, Fifth Edition: A
Quantitative Approach. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA,
5th edition, 2011.
[27] Hoefler, T. GreenGraph500 Submission Rules.
greengraph500rules.pdf.
http://green.graph500.org/
[28] D. Lemire, L. Boytsov, and N. Kurz. SIMD Compression and the Intersection of Sorted
Integers. ArXiv e-prints, January 2014.
[29] Daniel Lemire and Leonid Boytsov. Decoding billions of integers per second through
vectorization. CoRR, abs/1209.2137, 2012.
[30] Hang Liu and H. Howie Huang. Enterprise: Breadth-first graph traversal on gpus. In
Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, SC ’15, pages 68:1–68:12, New York, NY, USA, 2015.
ACM.
[31] H. Lv, G. Tan, M. Chen, and N. Sun. Compression and Sieve: Reducing Communication in Parallel Breadth First Search on Distributed Memory Systems. ArXiv e-prints,
August 2012.
[32] Aurèle Mahéo, Patrick Carribault, Marc Pérache, and William Jalby. Optimizing
collective operations in hybrid applications. In Proceedings of the 21st European MPI
Users’ Group Meeting, EuroMPI/ASIA ’14, pages 121:121–121:122, New York, NY,
USA, 2014. ACM.
[33] Duane Merrill, Michael Garland, and Andrew Grimshaw. Scalable gpu graph traversal.
SIGPLAN Not., 47(8):117–128, February 2012.
[34] Pierre Michaud, Andrea Mondelli, and André Seznec. Revisiting clustered microarchitecture for future superscalar cores: A case for wide issue clusters. ACM Trans. Archit.
Code Optim., 12(3):28:1–28:22, August 2015.
102
References
[35] Gordon E. Moore. Readings in computer architecture. chapter Cramming More Components Onto Integrated Circuits, pages 56–59. Morgan Kaufmann Publishers Inc.,
San Francisco, CA, USA, 2000.
[36] N. Y. Times. Intel Halts Development Of 2 New Microprocessors.
[37] Wee Keong Ng and Chinya V. Ravishankar. Block-oriented compression techniques for
large statistical databases. IEEE Trans. Knowl. Data Eng., 9(2):314–328, 1997.
[38] David A. Patterson and John L. Hennessy. Computer Organization and Design: The
Hardware/Software Interface. Morgan Kaufmann Publishers Inc., San Francisco, CA,
USA, 3rd edition, 2007.
[39] Mathias Payer. Too much pie is bad for performance. 2012.
[40] Jeff Plaisance, Nathan Kurz, and Daniel Lemire. Vectorized vbyte decoding. CoRR,
abs/1503.07387, 2015.
[41] Jan M. Rabaey. Digital Integrated Circuits: A Design Perspective. Prentice-Hall, Inc.,
Upper Saddle River, NJ, USA, 1996.
[42] Om P Sachan. White paper: Intel compilers for linux: Application porting guide.
Technical report, Intel Corporation, 2014.
[43] Falk Scholer, Hugh E. Williams, John Yiannis, and Justin Zobel. Compression of
inverted indexes for fast query evaluation. In Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval,
SIGIR ’02, pages 222–229, New York, NY, USA, 2002. ACM.
[44] C. Seshadhri, Ali Pinar, and Tamara G. Kolda. An in-depth analysis of stochastic
kronecker graphs. J. ACM, 60(2):13:1–13:32, May 2013.
[45] C.E. Shannon. A mathematical theory of communication. Bell System Technical
Journal, The, 27(3):379–423, July 1948.
[46] Alexander A. Stepanov, Anil R. Gangolli, Daniel E. Rose, Ryan J. Ernst, and
Paramjit S. Oberoi. Simd-based decoding of posting lists. In Proceedings of the 20th
ACM International Conference on Information and Knowledge Management, CIKM
’11, pages 317–326, New York, NY, USA, 2011. ACM.
[47] Andrew S. Tanenbaum and James R. Goodman. Structured Computer Organization.
Prentice Hall PTR, Upper Saddle River, NJ, USA, 4th edition, 1998.
[48] Rajeev Thakur, Rolf Rabenseifner, and William Gropp. Optimization of collective
communication operations in mpich. International Journal of High Performance Computing Applications, 19(1):49–66, 2005.
[49] Andrew Trotman. Compressing inverted files, 2003.
References
103
[50] Alan M. Turing. On computable numbers, with an application to the Entscheidungsproblem. Proceedings of the London Mathematical Society, 2(42):230–265, 1936.
[51] K. Ueno and T. Suzumura. Parallel distributed breadth first search on gpu. In High
Performance Computing (HiPC), 2013 20th International Conference on, pages 314–
323, Dec 2013.
[52] Koji Ueno and Toyotaro Suzumura. Highly scalable graph search for the graph500
benchmark. In Proceedings of the 21st International Symposium on High-Performance
Parallel and Distributed Computing, HPDC ’12, pages 149–160, New York, NY, USA,
2012. ACM.
[53] John von Neumann. First draft of a report on the edvac. IEEE Ann. Hist. Comput.,
15(4):27–75, October 1993.
[54] Jianyong Wang, Wojciech Cellary, Dingding Wang, Hua Wang, Shu-Ching Chen, Tao
Li, and Yanchun Zhang, editors. Web Information Systems Engineering - WISE 2015
- 16th International Conference, Miami, FL, USA, November 1-3, 2015, Proceedings,
Part I, volume 9418 of Lecture Notes in Computer Science. Springer, 2015.
[55] Yangzihao Wang, Andrew Davidson, Yuechao Pan, Yuduo Wu, Andy Riffel, and
John D. Owens. Gunrock: A high-performance graph processing library on the gpu.
SIGPLAN Not., 50(8):265–266, January 2015.
[56] Ian H. Witten, Alistair Moffat, and Timothy C. Bell. Managing Gigabytes (2Nd Ed.):
Compressing and Indexing Documents and Images. Morgan Kaufmann Publishers Inc.,
San Francisco, CA, USA, 1999.
[57] Hao Yan, Shuai Ding, and Torsten Suel. Compressing term positions in web indexes.
In Proceedings of the 32Nd International ACM SIGIR Conference on Research and
Development in Information Retrieval, SIGIR ’09, pages 147–154, New York, NY,
USA, 2009. ACM.
[58] Hao Yan, Shuai Ding, and Torsten Suel. Inverted index compression and query processing with optimized document ordering. In Proceedings of the 18th International
Conference on World Wide Web, WWW ’09, pages 401–410, New York, NY, USA,
2009. ACM.
[59] Hao Yan, Shuai Ding, and Torsten Suel. Inverted index compression and query processing with optimized document ordering. In Proceedings of the 18th International
Conference on World Wide Web, WWW ’09, pages 401–410, New York, NY, USA,
2009. ACM.
[60] Y. Yasui and K. Fujisawa. Fast and scalable numa-based thread parallel breadthfirst search. In High Performance Computing Simulation (HPCS), 2015 International
Conference on, pages 377–385, July 2015.
104
References
[61] Y. Yasui, K. Fujisawa, and K. Goto. Numa-optimized parallel breadth-first search on
multicore single-node system. In Big Data, 2013 IEEE International Conference on,
pages 394–402, Oct 2013.
[62] Kathy Yelick. The endgame for moore’s law: Architecture, algorithm, and application
challenges. In Federated Computing Research Conference, FCRC ’15, pages 6–, New
York, NY, USA, 2015. ACM.
[63] Andy Yoo, Edmond Chow, Keith Henderson, William McLendon, Bruce Hendrickson,
and Umit Catalyurek. A scalable distributed parallel breadth-first search algorithm
on bluegene/l. In Proceedings of the 2005 ACM/IEEE Conference on Supercomputing,
SC ’05, pages 25–, Washington, DC, USA, 2005. IEEE Computer Society.
[64] Jiangong Zhang, Xiaohui Long, and Torsten Suel. Performance of compressed inverted
list caching in search engines. In Proceedings of the 17th International Conference on
World Wide Web, WWW ’08, pages 387–396, New York, NY, USA, 2008. ACM.
[65] Peter Zhang, Eric Holk, John Matty, Samantha Misurda, Marcin Zalewski, Jonathan
Chu, Scott McMillan, and Andrew Lumsdaine. Dynamic parallelism for simple and
efficient gpu graph algorithms. In Proceedings of the 5th Workshop on Irregular Applications: Architectures and Algorithms, IA3 ’15, pages 11:1–11:4, New York, NY, USA,
2015. ACM.
[66] W. Zhang, G. Chen, M. Kandemir, and M. Karakoy. Interprocedural optimizations
for improving data cache performance of array-intensive embedded applications. In
Proceedings of the 40th Annual Design Automation Conference, DAC ’03, pages 887–
892, New York, NY, USA, 2003. ACM.
[67] Justin Zobel, Alistair Moffat, and Kotagiri Ramamohanarao. Inverted files versus
signature files for text indexing. ACM Trans. Database Syst., 23(4):453–490, December
1998.
[68] M. Zukowski, S. Heman, N. Nes, and P. Boncz. Super-scalar ram-cpu cache compression. In Data Engineering, 2006. ICDE ’06. Proceedings of the 22nd International
Conference on, pages 59–59, April 2006.
Acknowledgments
I would like thank the CEG group in the Ruprecht-Karls Universität Heidelberg
for having given me this opportunity to challenge myself with this project. I would
also like to thank all the members of the department for their helpfulness at every
moment.
Specially, I would like to thank, my advisor in this thesis; to the student previously
in charge of this project and from whom I have continued the work; to the colleague
and PhD student who has helped me with the german translations in this work; and
to the second Professor responsible of the project. The latter, in charge of performing
the tests (many times during weekends and Bank holidays).
105
Erklärung:
Ich versichere, dass ich diese Arbeit selbstständig verfasst habe und keine anderen als
die angegebenen Quellen und Hilfsmittel benutzt habe.
Heidelberg, den (Datum)
..........................................
| 8cs.DS
|
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
1
Universal channel coding
for general output alphabet
Masahito Hayashi Fellow, IEEE
arXiv:1502.02218v3 [cs.IT] 30 Jun 2017
Abstract
We propose two types of universal codes that are suited to two asymptotic regimes when the output alphabet
is possibly continuous. The first class has the property that the error probability decays exponentially fast and we
identify an explicit lower bound on the error exponent. The other class attains the epsilon-capacity the channel
and we also identify the second-order term in the asymptotic expansion. The proposed encoder is essentially based
on the packing lemma of the method of types. For the decoder, we first derive a Rényi-relative-entropy version
of Clarke and Barron’s formula the distance between the true distribution and the Bayesian mixture, which is of
independent interest. The universal decoder is stated in terms of this formula and quantities used in the information
spectrum method. The methods contained herein allow us to analyze universal codes for channels with continuous
and discrete output alphabets in a unified manner, and to analyze their performances in terms of the exponential
decay of the error probability and the second-order coding rate.
Index Terms
Universal coding; information spectrum; Bayesian; method of types
I. I NTRODUCTION
In wireless communication, the channel is described with continuous output alphabet, e.g., additive
white Gaussian noise (AWGN) channel and Gaussian fading channel. In these cases, it is not so easy to
identify the channel even though the channel is stationary and memoryless. Then, it is needed to make
a code that achieves good performances for any channel in a set of multiple stationary and memoryless
channels, e.g. a set of MIMO (multiple-input and multiple-output) Gaussian channels (e.g. [15], [13],
[16], [17], [18]). More precisely, it is desired to construct a code that works well for any stationary
and memoryless channel in a given parametric family of possible single-antenna/multi-antenna AWGN
channels for a real wireless communication alphabet.
In the discrete case, to resolve this problem, Csiszár and Körner[4] proposed universal channel coding
by employing the method of types. Since their code construction depends only on the input distribution
and the coding rate, it does not depend on the form of the channel, which is a remarkable advantage. They
also provide an explicit form of a lower bound of the exponential decreasing rate of the decoding error
probability. However, their method works only when input and output alphabets have finite cardinality.
Hence, their method cannot be applied to any continuous output alphabet while several practical systems
have a continuous output alphabet. Indeed, even in the continuous output case, universal channel codes
have been discussed for MIMO Gaussian channels [17], [18], [19], [20], in which, this problem was often
discussed in the framework of compound channel. Although the studies [24], [18], [19] did not cover the
general discrete memoryless case, the paper [20] covered the MIMO Gaussian channels as well as the
general discrete memoryless case1 . However, they did not provide any explicit form of a lower bound
This research was partially supported by the MEXT Grant-in-Aid for Scientific Research (A) No. 23246071, and was also partially
supported by the National Institute of Information and Communication Technology (NICT), Japan. The Centre for Quantum Technologies
is funded by the Singapore Ministry of Education and the National Research Foundation as part of the Research Centres of Excellence
programme. This paper was presented in part at 2014 the International Symposium on Information Theory and Its Applications, Melbourne,
Australia from 26 to 29 October 2014 [47].
Masahito Hayashi is with the Graduate School of Mathematics, Nagoya University, Furocho, Chikusaku, Nagoya, 464-860, Japan, and
Centre for Quantum Technologies, National University of Singapore, 3 Science Drive 2, Singapore 117542. (e-mail: [email protected])
1
While the review paper [17] discussed a universal code in the both cases, their treatments were separated.
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
2
of the exponential decreasing rate of the decoding error probability. Therefore, it is desired to invent
a universal channel code satisfying the following two conditions. (1) The universal channel code can
be applied to the discrete memoryless case and the continuous case in a unified way. (2) The universal
channel code has an explicit form of a lower bound of the exponential decreasing rate of the decoding
error probability.
Even in the discrete case, Csiszár and Körner[4]’s analysis is restricted to the case when the transmission
rate is strictly smaller than the capacity. When the transmission rate equals the capacity, the asymptotic
minimum error depends on the second-order coding rate [10], [11], [12]. For the fixed-length source
coding and the uniform random number, this kind of analysis was done in [14]. Recently, with the second
order rate, Polyanskiy [21] addressed a universal code in the framework of compound channel. Also,
Moulin [48] discussed a universal code in the discrete memoryless case after the conference version of
this paper [47]. The papers [8], [9], [22], [23] addressed the optimal second order rate for the mixed
channel. However, no study discussed the universal channel coding with the second-order coding rate
for the continuous case although the case of fixed continuous channels was discussed with the higher
order analysis including the second-order analysis by Moulin [40]. Hence, it is also desired to propose a
universal channel code working with the second-order coding rate even with the continuous case. Further,
the universal coding with the second order analysis has another problem as follows. The conventional
second order analysis [10], [11], [12] has meaning only when the mutual information is the first order
coding rate. However, the set of such channels has measure zero. So, such a analysis might be not so
useful when we do not know the transition matrix of the channel.
In this paper, we deal with the universal coding with a general output alphabet (including the continuous
case) and derive the exponential decreasing rate of the average error probability and the second order
analysis. Further, to resolve the above problem for the second order analysis, we introduce the perturbation
with the order O( 11 ) of the channel. Notice that such a perturbation for distribution is often employed
n2
[30, 8.11 Theorem]. Under this perturbation, we derive the second order analysis, which is the maximum
order to achieve the asymptotic constant average error probability. Indeed, even if the input alphabet is
continuous, we usually use only a finite subset of the input alphabet for encoding. Hence, it is sufficient
to realize a universal channel code for the case when the input alphabet is finite and the output alphabet is
continuous. In a continuous alphabet, we need an infinite number of parameters to identify a distribution
when we have no assumption for the true distribution. In statistics, based on our prior knowledge, we often
assume that the distribution belongs to a certain parametric family of distributions as in [20]. In particular,
an exponential family is often employed as a typical example. This paper adopts this typical assumption
as one of our assumptions. This paper addresses three assumptions. One is that the output distribution
PY |X=x belongs to an exponential family on a general set Y for each element x of a given finite set X . As
explained in Example 4, this assumption covers the usual setting with finite-discrete case because the set
of all distributions on a given finite-discrete set forms an exponential family. This assumption also covers
the Gaussian fading channel and the multi-antenna Gaussian channel as addressed in Examples 5 and 6.
However, in real wireless communication, the additive noise is not subject to Gaussian distribution [41].
In this case, the channel is not given as an exponential family of channels when the fading coefficient is
unknown or the distribution of the noise contain unknown parameter, To cover this case, we consider two
general conditions in this paper. Under these assumptions, we provide a universal code with an explicit
form of a lower bound of the exponential decreasing rate of the decoding error probability.
To construct our universal encoder, we employ the method given for the quantum universal channel
coding [5], in which, the packing lemma [4] is employed independently of the output alphabet. That
is, the paper [5] showed that the encoder given by the packing lemma can simulate the average of the
decoding error probability under the random coding. Since the method of [5] does not depend on the
output alphabet, it works well with a continuous alphabet.
To construct our universal decoder, we focus on the method given for the quantum universal channel
coding [5]. The paper [5] considered a universally approximated output distribution by employing the
method of types in the sense of the maximum relative entropy. Then, it employed the decoder constructed
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
3
by the information spectrum method based on the approximating distributions. However, in a continuous
alphabet, we cannot employ the method of types so that it is not so easy to give a universally approximated
output distribution in the sense of maximum relative entropy. That is, we cannot directly employ the method
in the paper [5].
To resolve this problem, we focus on Clarke and Barron formula [3] that shows that the Baysian average
distribution well approximates any independent and identical distribution in the sense of relative entropy,
i.e., Kullback-Leibler divergence. Its quantum extension was shown in [6]. Their original motivation is
rooted in universal data compression. However, they did not discuss the α-Rényi relative entropy. In this
paper, we evaluate the quality of this kind of approximation in terms of the α-Rényi relative entropy. Then,
we apply the universal decoder constructed by the information spectrum method based on the Baysian
average distributions. Modifying the method given in [5], we derive our lower bound of the error exponent
of our universal code. We also derive the asymptotic error of our universal code with the second-order
coding rate.
The remaining part of this paper is the following. Section II explains notations and our assumptions
In this section, we provide three examples of channels whose output distributions form an exponential
family. Section III explains our formulation and obtained results. In Section IV, we give notations for the
method of types and our universal encoder. This part is similar to the previous paper [5]. In Section V,
based on the result by Clarke and Barron [3], we derive an α-Rényi-relative-entropy version of Clarke
and Barron formula as another new result. Section VI gives our universal decoder and its properties to be
applied in the latter sections. In Section VII, we prove our lower bound of our error exponent. In Section
VIII, we prove the universal achievability for the second order sense. Appendixes are devoted in several
lemmas used in this paper.
II. P RELIMINARY
A. Information quantities
We focus on an input alphabet X := {1, . . . , d} with finite cardinality and an output alphabet Y that
may have infinite cardinality and is a general measurable set. In this paper, the output alphabet Y is treated
as a general probability space with a measure µ(dy) because this description covers the probability space
of finite elements and the set of real values. Hence, when the alphabet Y is a discrete set including a
finite set, the measure µ(dy) is chosen to be the counting measure. When the alphabet Y is a vector space
over the real numbers R, the measure µ(dy) is chosen to be the Lebesgue measure. When we treat a
probability distribution P on the alphabet Y, it is restricted to a distribution absolutely continuous with
respect µ(dy). In the following, we use the lower case p(y) to express the Radon-Nikodym derivative of P
with respect to the measure µ(dy), i.e., the probability density function of P so that P (dy) = p(y)µ(dy).
In general, a channel from X to Y is described as a collection W of conditional probability measures
Wx on Y for all inputs x ∈ X . Then, we impose the above assumption to Wx for any x ∈ X . So, we
have Wx (dy) = wx (y)µ(dy). We denote the conditional probability density function by w = (wx )x∈X .
When a distribution on X is given by a probability distribution P , and a conditional distribution on a
set Y with the condition on X is given by V , we define the joint distribution W ×P
P on X × Y by
W × P (B, x) := W (B|x)P (x), and the distribution W · P on Y by W · P (B) := x W (B|x)P (x)
for a measurable set B ⊂ Y. Also, we define the notations w × P andP
w · P as w × P (y, x)µ(dy) :=
W × P (dy, x) = wx (y)P (x)µ(dy) and w · P (y)µ(dy) := W · P (dy) = x∈X wx (y)P (x)µ(dy). We also
employ the notations WP := W · P and wP := w · P .
We denote the expectation under the distribution P by EP [ ]. Throughout this paper, the base of
the logarithm is chosen to be e. For two distributions P and Q on Y, we define the relative entropy
)
) s
D(P kQ) := EP [log p(Y
], and the value sD1+s (P kQ) := log EP [( p(Y
) ] for s > −1 when these
q(Y )
q(Y )
p(Y ) s
p(Y )
expectations EP [log q(Y ) ] and EP [( q(Y ) ) ] exist. The function s 7→ sD1−s (P kQ) is a concave function
)
for s ∈ [0, 1] because −sD1−s (P kQ) can be regarded as a cumulant generating function of − log p(Y
.
q(Y )
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
4
For s ∈ [−1, 0) ∪ (0, ∞), the Rényi relative entropy D1+s (P kQ) is defined as D1+s (P kQ) := sD1+ss(P kQ) .
Here, the function sD1+s (P kQ) is defined for s = 0, but the Rényi relative entropy D1+s (P kQ) is not
necessarily defined for s = 0. In addition, we can define the max relative entropy
p(y)
Dmax (P kQ) := inf a a ≥ log
almost every where with respect to y under the distribution P .
q(y)
(1)
So, we have Dmax (P kQ) = limα→∞ Dα (P kQ).
Given a channel W from X to Y and a distribution P on X , we define the value sI1−s (P, W ) for
s ∈ [0, 1] as
X
sI1−s (P, W ) := − sup log
P (x)e−sD1−s (Wx kQ) = inf sD1−s (W × P kQ × P ).
(2)
Q
Q
x
Since the minimum of concave functions is a concave function, the function s 7→ sI1−s (P, W ) is also a
concave function. In fact, Hölder inequality guarantees that
Z X
1
(3)
sI1−s (P, W ) = −(1 − s) log (
P (x)wx (y)1−s ) 1−s µ(dy)
x
when the RHS exists [35][34, (34)]. For s ∈ (0, 1], when sI1−s (P, W ) is finite, I1−s (P, W ) is defined
)
. The quantity I1−s (P, W ) with (3) is the same as the Gallager function [7] with different
as sI1−s (P,W
s
parametrization for s. We also define the mutual information
Z
X
X
wx (y)
I(P, W ) :=
P (x)D(Wx kW · P ) =
P (x) wx (y) log
µ(dy)
(4)
w · P (y)
Y
x
x
and its variance
i2
wx (Y )
− I(P, W )
V (P, W ) :=
P (x)EWx log
w · P (Y )
x
Z
2
X
wx (y)
=
P (x) wx (y) log
− I(P, W ) µ(dy).
w · P (y)
Y
x
h
X
(5)
When the channel satisfies some suitable conditions, we have
lim I1−s (P, W ) = I(P, W ).
s→0
(6)
For example, when the alphabet Y has a finite cardinality, the relation (6) holds by choosing the measure
µ(dy) to be the counting measure.
B. Exponential family
To state the conditions for the main results, in this subsection, we consider a parametric family of
distributions {Pθ }θ∈Θ on the alphabet Y with a finite-dimensional parameter set Θ ⊂ Rk . Here, we
assume that the Radon-Nikodym derivative pθ (y) is differentiable with respect to θ. To cover a class
of proper channels, we introduce an exponential family of distributions, which covers so many useful
distributions in statistics, and has been widely recognized as a key concept of statistics [1], [36]. A
parametric family of distributions {Pθ }θ∈Θ on the measurable set Y is called an exponential family with
a parametric space Θ ⊂ Rk when a distribution P0 is absolutely continuous with respect to a measure
µ(dy) on a measurable set Y and the parametric family of distributions {Pθ }θ∈Θ is written as [36]
Pk
pθ (y) = p0 (y)e
j=1
θ j gj (y)−φ(θ)
with generators gj (y) and satisfies the following conditions.
(7)
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
A1
The potential function φ(θ) equals the cumulant generating function of gj , i.e.,
Z
Pk
j
φ(θ)
e
=
p0 (y)e j=1 θ gj (y) µ(dy) < ∞
5
(8)
Y
for θ ∈ Θ.
A2
φ is a C 2 function on Θ, i.e., the Hessian matrix Jθ = (Jθ|i,j )i,j of φ is continuous on Θ.
The function ψ is called the potential. In the above assumption, the Hessian matrix Jθ equals the covariance
matrix of the random variables (gj (Y ))i under the distribution Pθ . In statistics, the Hessian matrix Jθ
is called Fisher information matrix, and expresses the bound of the precision of the estimation of the
parameter θ.
Example 1 (Multinomial distributions): The set of multinomial distributions in a finite output set Y =
{0, 1, . . . , m} forms an exponential family. We choose the generator gj (y) := δj,y for j = 1, . . . , m and
1
define p0 (y) := m+1
. Remember that µ(dy) is given as the counting measure. For parameters θ = (θj )
with j = 1, . . . , m, the distributions pθ (y) is given as
y
Pemθ j when y ≥ 1
1+ j=1 eθ
(9)
pθ (y) :=
Pm1 θj when y = 0.
1+
e
j=1
Pm
j
The potential φ(θ) is log(1 + j=1 eθ ) − log(m + 1).
Example 2 (Poisson distributions): The set of Poisson distributions is a one-parameter exponential
−1
family on the set of natural numbers N := {0, 1, . . .}. We set p0 (y) := en! and choose the generator
nθ−eθ
g(y) := y. Then, we have pθ|P oi (y) := e n! with the potential φ(θ) = eθ − 1.
Example 3 (Gaussian distributions): The set of Gaussian distributions forms a one-parameter exponential
family on R. We set p0 (y) :=
(y)2
2
e−
√
(θ)2
.
2
2π
(y−θ)2
and choose the generator g(y) := y. Then, we have pθ (y) :=
e− √ 2
2π
with the potential φ(θ) =
Now, we consider a parametric family of channels from X to Y with a finite-dimensional parameter
set The channel is parameterized as {Wθ }θ∈Θ with a finite-dimensional parameter set Θ ⊂ Rk .
Now, using the conditions for an exponential family, we make a condition for a family for the channels
{Wθ }θ∈Θ . A parametric family of channels {Wθ }θ∈Θ is called an exponential family of channels with
a parametric space Θ ⊂ Rk when a channel W0 = (W0,x ) from X to Y is composed of absolutely
continuous distributions W0,x with respect to a measure µ(dy) and the parametric family of channels
{Wθ }θ∈Θ is written as
Pk
wθ,x (y) = w0,x (y)e
j=1
θ j gj,x (y)−φx (θ)
for any x ∈ X with generators gj,x(y) and satisfies the following conditions.
B1
The potential function φx (θ) equals the cumulant generating function of gj,x , i.e.,
Z
Pk
j
φx (θ)
e
=
w0,x (y)e j=1 θ gj,x (y) µ(dy) < ∞
(10)
(11)
Y
for θ ∈ Θ.
B2
φx is a C 2 function on Θ, i.e., the Hessian matrix Jθ,x = (Jθ,x|i,j )i,j of φ is continuous on Θ.
When Y equals X , this definition of exponential family has been discussed in many papers in the context
of Markovian processes [25], [26], [27], [28], [29], and is different from the definition in the papers [42],
[43], [44], [45]. Here, for convenience, we say that a family of channels satisfies Condition A when all
of the above conditions hold, i.e., it is an exponential family of channels.
Here, we list typical examples as follows. All of the below examples satisfy Condition B.
Example 4 (Finite set): Consider a finite input set X = {1, . . . , d} and a finite output set Y =
{0, 1, . . . , m}. Then, the measure µ(dy) is chosen to be the counting measure. We choose the generators
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
6
g(i,j),x (y) := δi,x δj,y with i = 1, . . . , d, j = 1, . . . , m. For parameters θ = (θi,j ), the exponential family of
channels is given as
x,y
Pemθ x,j when y ≥ 1
1+ j=1 eθ
wθ,x (y) :=
(12)
Pm1 θx,j when y = 0.
1+
e
j=1
Then, the set of output distributions {Wθ }θ forms an exponential family for x ∈ X .
Example 5 (Gaussian fading channel): Assume that the output set Y is the set of real numbers R and
the input set L is a finite set {1, . . . , d}. We choose d elements x1 , . . . , xd ∈ R as input signals. The
additive noise Z is assumed to be subject to the Gaussian distribution with the expectation b and the
variance v. Then, we assume that the received signal Y is given by the scale parameter a as
Y = axi + Z.
(13)
The class of these channels is known as Gaussian fading channels, and its compound channel is discussed
2
in the paper [19]. We choose the generators g1,ℓ (y) := − y2 , g2,ℓ (y) := yxℓ , and g3,ℓ (y) := y. By using the
three parameters θ1 := v1 , θ2 := av , and θ3 := vb , the channels form an exponential family of channels as
(y−axℓ −b)2
1
wθ,ℓ (y) := √
e− 2v
2πv
1
2
(θ 2 xℓ +θ 3 )2
−y
1
2
3
θ
= √ e− 2 θ +y(θ xℓ +θ )− 2θ1 .
(14)
2π
Example 6 (multi-antenna Gaussian channel): We consider the constant multi-antenna (MIMO) Gaussian channel when the sender has t antennas and the receiver has r antennas. Then, the output set Y is
given as the set of r-dimensional real numbers Rr and the input set L is given as a finite set {1, . . . , d}.
In this case, the sender chooses d elements ~x1 , . . . , ~xd ∈ Rt as input signals. The r-dimensional additive
~ is assumed to be subject to the r-dimensional Gaussian distribution with the expectation ~b ∈ Rd
noise Z
and the covariance matrix vi,i′ . When the input is ~xℓ , we assume that the received signal Y~ is written by
an r × t matrix (ai,j ) as
t
X
~
Y~ =
ai,j xjℓ + Z.
(15)
j=1
The class of these channels is known as′ constant multi-antenna Gaussian channels [13]. We choose three
i i
kinds of generators g(1,i,i′ ),ℓ′ (y) := − y 2y , g(2,i,j),ℓ′ (y) := y i xjℓ′ , and g(3,i),ℓ′ (y) := y i . Then, we define three
′
kinds of parameters: The symmetric matrix (θ1,i,i )i,i′ isPdefined as the inverse matrix ofPthe covariance
′
′
′ ′
matrix vi,i′ . The matrix (θ2,i,i )i,i′ is defined as θ2,i,j := i′ θ1,i,i ai′ ,j . The vector θ3,i := i′ θ1,i,i bi . So,
the channels form an exponential family of channels as
det(θ1 ) − 1 Pi,i′ θ1,i,i′ (yi −Pj ai,j xj −bi )(yi′ −Pj′ ai′ ,j′ xj′ −bi′ )
ℓ
ℓ
wθ,ℓ (~y ) := √
e 2
2π
det(θ1 ) − 1 Pi,i′ θ1,i,i′ yi yi′ +Pi yi (θ3,i +Pj θ2,i,j xj )− 1 Pi,i′ ((θ1 )−1 )i,i′ (θ3,i +Pj θ2,i,j xj )(θ3,i′ +Pj θ2,i′ ,j xj )
ℓ .
ℓ
ℓ
2
e 2
= √
(16)
2π
Example 7 (Poisson channel): When the signal is too weak, we have Poisson channel as follows.
Assume that the output set Y is N and the input set X is a finite set {1, . . . , d}. Then, the measure µ(dy)
is chosen to be the counting measure. We choose three kinds of generators gi,x (y) := δi,x y. By using the
parameter θ = (θi )di=1 , the Poisson channel is given as an exponential family of channel with
wθ,x (y) := Pθx |P oi (y).
For the definition of Pθx |P oi , see Example 2.
Here, we summarize the number in Table I.
(17)
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
7
TABLE I
N UMBERS USED IN THIS PAPER
k
d
n
Dimension of parameter
Cardinality of input alphabet
No. of use of channels
C. General assumptions
In real wireless communication, the additive noise is not subject to Gaussian distribution [41]. Now, we
consider the case when the additive noise Z in (13) is subject to a general parametric family of distributions
{Pθ }θ on Y = R, which is not a Gaussian distribution. Then, the probability density function of the output
distributions are given as wθ,ℓ (y) = pθ (y −axℓ ), and do not form an exponential family in general. Further,
even when the distribution Pθ of the additive noise Z is known, when the fading coefficient a is unknown,
the set of our channels does not form an exponential family in general. Hence, it is needed to relax the
condition of exponential family for practical purpose. As a preparation, we consider the following condition
for a family of distributions {Pθ }θ∈Θ on the measurable set Y with Θ ⊂ Rd .
C1
The parametric space Θ is compact.
C2
The likelihood ratio derivative lθ|i (y) := ∂θ∂ i log pθ (y) exists for i = 1, . . . , k even on the boundary
of Θ.
R
C3
The Fisher information matrix Jθ|i,j := Y pθ (y)lθ|i (y)lθ|j (y)µ(dy) can be defined. Also, the map
θ 7→ Jθ|i,j is continuous for i, j = 1, . . . , d.
C4
For any s > 0, the map (θ, θ ′ ) 7→ D1+s (Pθ kPθ′ ) is continuous. Also, when s > 0 is fixed, the
P
D
(P kP
)
convergence 2 limǫ→0 1+s ǫθ2 θ+ξǫ = i,j ξj ξi Jθ|i,j holds and is uniform for θ and ξ ∈ {ξ ∈
Rd |kξk = 1}.
We say that a family of distributions satisfies Condition C when all of the above conditions hold.
For example, when an exponential family has a compact parameter space, the likelihood ratio derivative
lθ|i (y) is given as gj (y)−EPθ [gj (Y )] so that it satisfies Condition C. Under this condition, to prove several
required properties, we can effectively employ the compactness of the parameter space Θ. For example,
when a probability density function P is differentiable on Y = R, the support P is Y = R, and the
integral satisfies the condition
Z ∞
dp(y)
/p(y))2p(y)dy < ∞
(18)
(
dy
−∞
the family of distribution pθ (y) := p(y − θ) satisfies Condition C [2].
Next, we extend Condition C to a family of channels Wθ = (Wθ,x ) from a discrete alphabet X to a
general alphabet Y with the parameter set Θ.
D1
The parametric space Θ is compact.
D2
The likelihood ratio derivative lθ,x|i (y) := ∂θ∂Ri log wθ,x (y) exists for i = 1, . . . , k and x ∈ X .
D3
The Fisher information matrix Jθ,x|i,j := Y wθ,x (y)lθ,x|i (y)lθ,x|j (y)µ(dy) can be defined for
x ∈ X . Also, the map θ 7→ Jθ,x|i,j is continuous for i, j = 1, . . . , d and x ∈ X .
D4
For any s > 0 and x ∈ X , the map (θ, θ ′ ) 7→ D1+s (Wθ,x kWθ,x′ ) is continuous. Also, when
P
D
(W kW
)
s > 0 and x ∈ X are fixed, the convergence 2 limǫ→0 1+s θ,xǫ2 θ,x+ξǫ = i,j ξj ξi Jθ,x|i,j holds
and is uniform for θ and ξ ∈ {ξ ∈ Rk |kξk = 1}.
Hence, we say that a family of channels satisfies Condition D when all of the above conditions hold. For
example, when an exponential family of channels has a compact parameter space, it satisfies Condition D.
For a given distribution P on Y = R for the additive noise, we consider the channel Wθ (y|i) = P (y − θi )
with a compact space Θ. Even when the additive noise in a wireless communication is not subject to
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
8
Gaussian distribution, when the distribution P satisfies the condition (18), the family of channels {Wθ }θ
satisfies Condition D.
However, it is better to remove the assumption of the compactness. For this purpose, we introduce the
conditions E1 and E4 instead of C1 and C4 for a family of distributions {Pθ }θ∈Θ on the measurable set
Y with Θ ⊂ Rk .
E1
There exists a sequence of compact subsets {Θi }∞
i=1 satisfying the following conditions. (1)
Θi ⊂ Θi+1 . (2) For any θ ∈ Θ, there exists Θi such that θ ∈ Θi .
E4
The uniformity of the convergence in C4 holds on any compact subspace Θi .
We say that a family of channels satisfies Condition E when the conditions E1 and E4 hold as well as
C2 and C3.
Next, we extend Condition E to a family of channels Wθ = (Wθ,x ) from a discrete alphabet X to a
general alphabet Y with the parameter set Θ. Instead of D1 and D4, we introduce the conditions F1 and
F4.
F1
For each x ∈ X , there exists a sequence of compact subsets {Θi }∞
i=1 satisfying the following
i
i+1
i
conditions. (1) Θ ⊂ Θ . (2) For any θ ∈ Θ, there exists Θ such that θ ∈ Θi .
F4
The uniformity of the convergence in D4 holds on any compact subspace Θi .
We say that a family of channels satisfies Condition F when the conditions F1 and F4 hold as well as
D2 and D3. Since Condition F does not require compactness, any exponential family of channels satisfies
Condition F.
Example 8: Assume that the output set Y is the set of real numbers R and the input set L is a finite
set {1, . . . , d}. We choose d elements x1 , . . . , xd ∈ R as input signals. We assume that the additive noise
Z in (13) is subject to a general parametric family of distributions {Pθ }θ on Y = R satisfying Condition
E. Our family of channels {Wθ }θ is given as wθ,ℓ (y) = pθ (y − axℓ ). Since {Pθ }θ on Y = R satisfies
Condition E, this family of channels satisfies Condition F.
Example 9: Assume that the output set Y is the set of real numbers R and the input set L is a finite set
{1, . . . , d}. We choose d elements x1 , . . . , xd ∈ R as input signals. We assume that the additive noise Z
in (13) is subject to a distribution P satisfying (18). However, we do not know the fading coefficient θ.
Our family of channels {Wθ }θ∈R is given as wθ,ℓ (y) = p(y − θxℓ ). Due to the condition (18), this family
of channels satisfies Condition F.
III. M AIN
RESULTS
A. Exponential evaluation
In this paper, we address the n-fold stationary memoryless channel of W
Qθn, i.e., we focus onn the
n
n
channel Wθn whose probability density function is defined as wθ,x
n (y ) :=
i=1 wθ,xi (yi ) with x :=
n
n
n
(x1 , . . . , xn ) ∈ X and y := (y1 , . . . , yn ) ∈ Y . When the set of messages is Mn := {1, . . . , Mn }, the
encoder is given as a map En from Mn to X n , and the decoder is given as a map Dn from Y n to Mn .
The triple Φn := (Mn , En , Dn ) is P
calledP
a code, and the size Mn is often written as |Φn |. The decoding
n
n
−1
error probability is eθ (Φn ) := M1n M
i=1
j6=i Wθ,En (i) (Dn (j)).
Theorem 1: Given real numbers R and R1 with R1 > R, a distribution P on X , and a family of
channels {Wθ }θ , we assume that the family of channels {Wθ }θ satisfies Condition B, D, or F, and that
sI1−s (P, Wθ ) can be defined for any s ∈ [0, 1]. Then, there exists a sequence of codes Φn with the size
|Φn | = enR satisfying that
lim
n→∞
−1
log eθ (Φn ) ≥ min( max (sI1−s (P, Wθ ) − sR1 ), R1 − R)
s∈[0,1]
n
(19)
for every θ ∈ Θ.
The proof of Theorem 1 is given as the combination of Sections IV, VI, and VII. Section IV gives the
encoder and Section VI gives the decoder and a part of the evaluation of the decoding error probability.
Section VII evaluates the decoding error probability by using several formulas given in Sections IV and
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
9
VI. While R expresses the coding rate, our code needs another rate R1 > R, which decides the form of
the universal decoder. This is because our decoder employs the likelihood ratio test of a certain simple
hypothesis testing, and the rate R1 describes the likelihood ratio. Hence, it is crucial to decide the rate
R1 for the construction of our universal code.
For simplicity, we consider the case when we can identify the parameter θ, we can choose Θ0 to be
{θ}. In this case, the best choice of R1 is the real number R1 satisfying maxs∈[0,1] (sI1−s (P, Wθ )−sR1 ) =
1
R1 −R, which is given as R1 = R+maxs∈[0,1] 1+s
(sI1−s (P, Wθ )−sR), which can be seen as a special case
1
(sI1−s (P, Wθ ) − sR).
of Lemma 1 later. Hence, our lower bound of the exponent is maxP maxs∈[0,1] 1+s
1
This value is strictly smaller than Gallager’s exponent maxs∈[0,1] 1−s (sI1−s (P, Wθ ) − sR)[7].
Now, we consider the general case when we assume that the true channel parameter θ belongs to a
subset Θ0 ⊂ Θ. Then, we propose the following method to choose the rate R1 and the distribution P to
identify our code for a given transmission rate R < supP inf θ∈Θ0 I(P, Wθ ).
(M1) When argmaxP inf θ∈Θ0 I(P, Wθ ) is a non-empty set, we fix an element P1 in this set. Otherwise,
we choose a distribution P1 such that inf θ∈Θ0 I(P1 , Wθ ) is sufficiently close to supP inf θ∈Θ0 I(P, Wθ )).
Next, we choose R1 ∈ (R, inf θ∈Θ0 I(P1 , Wθ )). Then, the exponential decreasing rate is greater
than min(maxs∈[0,1] (sI1−s (P, Wθ ) − sR1 ), R1 − R) > 0 when the true parameter is θ ∈ Θ0 .
Now, we consider another method to improve the bound min(maxs∈[0,1] (sI1−s (P, Wθ ) − sR1 ), R1 − R).
For this purpose, we prepare the following lemma whose proof of Lemma 1 is given in Appendix A.
Lemma 1: When the function s 7→ sI1−s (P, Wθ ) is a C 1 function for any θ ∈ Θ, we have
max inf min( max (sI1−s (P, Wθ ) − sR1 ), R1 − R)
R1
θ∈Θ0
s∈[0,1]
= inf max min( max (sI1−s (P, Wθ ) − sR1 ), R1 − R)
θ∈Θ0
R1
s∈[0,1]
1
(sI1−s (P, Wθ ) − sR).
θ∈Θ0 s∈[0,1] 1 + s
= inf max
(20)
The maximum value maxR1 min(maxs∈[0,1] (sI1−s (P, Wθ ) − sR1 ), R1 − R) is attained when R1 = R +
1
maxs∈[0,1] 1+s
(sI1−s (P, Wθ ) − sR). In particular, the maximum value
maxR1 inf θ∈Θ0 min(maxs∈[0,1] (sI1−s (P, Wθ ) − sR1 ), R1 − R) is attained when R1 = R+
1
(sI1−s (P, Wθ ) − sR).
inf θ∈Θ0 maxs∈[0,1] 1+s
Based on Lemma 1,we propose the following method to choose the rate R1 and the distribution P ,
which improve the above method.
1
(M2) When argmaxP inf θ∈Θ0 maxs∈[0,1] 1+s
(sI1−s (P, Wθ )−sR) is a non-empty set, we fix an element
1
P2 in this set. Otherwise, we choose a distribution P2 such that inf θ∈Θ0 maxs∈[0,1] 1+s
(sI1−s (P2 , Wθ )−
1
(sI1−s (P, Wθ ) − sR). Then, we choose
sR) is arbitrarily close to supP inf θ∈Θ0 maxs∈[0,1] 1+s
R1∗ := argmaxR1 inf θ∈Θ0 min(maxs∈[0,1] (sI1−s (P2 , Wθ ) − sR1 ), R1 − R).
1
For simplicity, we consider the case when maxP inf θ∈Θ0 maxs∈[0,1] 1+s
(sI1−s (P, Wθ ) − sR) exists. The
worst case of the lower bound of the exponential decreasing rate of (M2) is inf θ∈Θ0 min(maxs∈[0,1] (sI1−s (P2 , Wθ )−
sR1∗ ), R1∗ − R). Due to (20) in Lemma 1, the lower bound is calculated as
inf min( max (sI1−s (P2 , Wθ ) − sR1∗ ), R1∗ − R)
θ∈Θ0
s∈[0,1]
= max inf min( max (sI1−s (P2 , Wθ ) − sR1 ), R1 − R)
R1
θ∈Θ0
s∈[0,1]
1
(sI1−s (P2 , Wθ ) − sR)
θ∈Θ0 s∈[0,1] 1 + s
1
1
(sI1−s (P, Wθ ) − sR) ≥ inf max
(sI1−s (P2 , Wθ ) − sR).
= max inf max
P θ∈Θ0 s∈[0,1] 1 + s
θ∈Θ0 s∈[0,1] 1 + s
= inf max
(21)
1
(sI1−s (P2 , Wθ ) − sR) expresses the worst case of the lower bound of the
Since inf θ∈Θ0 maxs∈[0,1] 1+s
exponential decreasing rate of (M1), (M2) provides a better lower bound than (M1).
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
10
B. Second order evaluation
When we need to realize the transmission rate close to the capacity, we need to employ the second
order analysis.
Theorem 2: Given real numbers R1∗ and R2∗ , a distribution P on X , and a family of channels {Wθ }θ ,
we assume that the family of channels {Wθ }θ satisfies Condition A, B, or C, and that 0 < I(P, Wθ ) < ∞
and 0 < V (P, Wθ ) < ∞ for any parameter θ ∈ Θ. Then, there exists a sequence of codes Φn with the
1
∗
∗ √
size |Φn | = enR1 + nR2 −n 4 satisfying the following conditions. (1) When I(P, Wθ ) > R1∗ , we have
lim eθ (Φn ) = 0.
n→∞
(22)
(2) When I(P, Wθ ) = R1∗ , we have
lim eθ (Φn ) ≤
n→∞
Z
√
R∗
2
V (P,Wθ )
−∞
1
x2
√ exp(− )dx.
2
2π
(23)
The proof of Theorem
2 is given in Section VIII. Therefore, in order that the upper bound is smaller
R∗
2
√
R
2
V (P,Wθ )
√1 exp(− x )dx < 1 , R∗ needs to be a negative value. For example, when
than 21 , i.e., −∞
2
2
2
2π
Θ = {θ}, we choose P as an element of {P ′|I(P, Wθ ) = maxP I(P, Wθ ), V (P, Wθ ) = V− }∗ with V− :=
R
R √V2− 1
2
minP {V (P, Wθ )|I(P, Wθ ) = maxP I(P, Wθ )}. Then, we can realize the minimum error −∞ √2π exp(− x2 )dx
√ ∗
with the coding length n max I(P, Wθ ) + nR2 when R2∗ ≤ 0 [10], [11], [12].
However, most of channels Wθ do not satisfy the condition I(P, Wθ ) = R1∗ for a specific first order
rate R1∗ . That is, Theorem 2 gives the asymptotic average error probability 0 or 1 as the coding result for
most of channels Wθ . This argument does not reflect the real situation properly because many channels
have their average error probability between 0 and 1, i.e., 0 nor 1 is not an actual value of average error
probability. To overcome this problem, we introduce the second order parameterization θ1 + √1n θ2 ∈ Θ.
This parametrization is applicable when the unknown parameter
belongs to the neighborhood of θ1 . Given
√
a parameter θ, we choose the parameter θ2 := (θ − θ1 ) n. This parametrization is very conventional in
statistics. For example, in statistical hypothesis testing, when we know that the true parameter θ belongs
to the neighborhood of θ1 , using the new parameter θ2 , we approximate the distribution family by the
Gaussian distribution family [39]. In this sense, the χ2 -test gives the asymptotic optimal performance. As
another example, this kind of parametrization is employed to discuss local minimax theorem [30, 8.11
Theorem].
The range of the neighborhood of this method depends on n and θ1 . In the realistic case, we have
some error for our guess of channel and the number n is finite. When the range of the error of our
prior estimation of the channel is included in the neighborhood, this method effectively works. When
our estimate of channel has enough precision as our prior knowledge, we can expect such a situation. In
this scenario, we may consider the following situation. We fix the parameter θ1 as a basic (or standard)
property of the channel. We choose the next parameter θ2 as a fluctuation depending on the daily changes
or the individual specificity of the channel. Here, n is chosen depending on the our calculation ability
of encoding and decoding. Hence, it is fixed priorly to the choice of θ2 . Then, Theorem 2 is refined as
follows.
Theorem 3: Given real numbers R1∗ and R2∗ , a distribution P on X , and a family of channels {Wθ1 + √1 θ2 }
n
with the second order parameterization, we assume the following conditions in addition to the assumptions
of Theorem 3. The function θ 7→ I(P, Wθ ) is a C 1 function on Θ, and the function θ 7→ V (P, Wθ ) is
a continuous function on Θ. Under these conditions, there exists a sequence of codes Φn with the size
1
∗
∗ √
|Φn | = enR1 + nR2 −n 4 satisfying the following conditions. (1) When I(P, Wθ1 ) > R1∗ , we have
lim eθ1 + √1
n→∞
θ
n 2
(Φn ) = 0.
(24)
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
11
(2) When I(P, Wθ1 ) = R1∗ , we have
lim e
n→∞
θ1 + √1n θ2
(Φn ) ≤
Z
R∗ −f (θ2 )
√2
V (P,Wθ )
−∞
1
x2
√ exp(− )dx,
2
2π
(25)
P
θ)
where f (θ2 ) := i ∂I(P,W
|θ=θ1 θ2i . Further, when the convergence (72) is uniform with respect to θ2
∂θ i
in any compact subset, there exists an upper bound ēθ1 + √1 θ2 (Φn ) of eθ1 + √1 θ2 (Φn ) such that the bound
n
n
ēθ1 + √1 θ2 (Φn ) converges to the RHS of (25) uniformly for θ2 in any compact subset.
n
The proof of Theorem 3 is given in Section VIII.
Now, we consider how to realize the transmission rate close to the capacity, In this case, we need
to know that the true parameter θ0 belongs to the neighborhood of θ1 . More precisely, given a real
number C > 0 and the parameter θ1 , weP
need to assume that the true parameter θ0 belongs to the set
1
{θ1 + √n θ2 |kθ2 k ≤ C}, where kθk := ki=1 (θi )2 . To achieve the aim, we choose the distribution P
such that I(P, Wθ1 ) equals the capacity of the channel Wθ1 . Then, we choose R1∗ to be the capacity,
i.e., I(P, Wθ1 ). To keep the average error probability
approximately less than ǫ, we choose R2∗ as the
R∗
−f
(θ
)
R √V2 (P,W2θ ) 1
x2
√
exp(−
)dx ≤ ǫ. Then, using Theorem 3, we
maximum number satisfying maxθ2 :kθ2 k≤C −∞
2
2π
can realize a rate close to the capacity and the average error probability approximately less than ǫ. When
we have larger order perturbation than √1n , our upper bound of limn→∞ eθ (Φn ) becomes 0 or 1 because
such a case corresponds to the case when kθ2 k goes to infinity. Hence, we find that √1n is the maximum
order of the perturbation to discuss the framework of the second order.
One might doubt the validity of the above approximation, however, it can be justified as follows. When
the true parameter belongs to the neighborhood of θ1 satisfying I(P, Wθ1 ) = R1∗ , the above theorem
gives the approximation of average error probability. Theorem 3 reflects such a real situation. Indeed, the
convergence of the upper bound ēθ (Φn ) is not uniform for θ in the neighborhood of such a parameter θ1
because the limit is discontinuous at θ1 . However, the convergence of eθ1 + √1 θ2 (Φn ) is uniform for θ2 in
n
any compact subset. The reason for the importance of the second order asymptotics is that the limiting
error probability can be used for the approximation of the true average error probability because the
convergence is uniform for the second order rate in any compact subset [31]. Due to the same reason, the
RHS of (25) can be used for the approximation of our upper bound of the true average error probability.
IV. M ETHOD OF TYPES AND UNIVERSAL ENCODER
In this section, we define our universal encoder for a given distribution P on X and a real positive
number R < H(P ) by the same way as [5]. Although the contents of this section is the same as a part
of [5], since the paper [5] is written with quantum terminology, we repeat the same contents as a part of
[5] for readers’ convenience.
The key point of this section is to provide a code to satisfy the following property by using the method
of types. In information theory, we usually employ the random coding method. However, to construct
universal coding, we cannot employ this method because a code whose decoding error probability is less
than the average might depend on the true channel. Hence, we need to choose a deterministic code whose
decoding error probability is less than the average. To resolve this problem, we employ a deterministic
code whose decoding error probability can be upper bounded by a polynomial times of the average of
the decoding error probability under the random coding. To realize this idea, we employ the method of
types.
First, we prepare notations for the method of types. Given an element xn ∈ X n , we define the integer
nx := |{i|xi = x}| for x ∈ X and the empirical distribution TY (xn ) := ( nn1 , . . . , nnd ), which is called a
type. The set of types is denoted by Tn (X ). For P ∈ Tn (X ), a subset of X n is defined by:
TP := {xn ∈ X n |The empirical distribution of xn is equal to P }.
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
12
Hence, we define the distribution
n
PTP (x ) :=
1
|TP |
xn ∈ TP
xn ∈
/ TP .
0
Then, we define the constant cn,P by
1
cn,P
n
n
Pd
PTP (x′ ) = P n (x′ ) = e
i=1
ni log P (i)
(26)
= e−nH(P ) .
(27)
for x′ n ∈ TP . So, the constant cn,P is bounded as
cn,P ≤ |Tn (X )|.
(28)
Further, the sequence of types V = (v1 , . . . , vd ) ∈ Tn1 (X ) × · · · × Tnd (X ) is called a conditional type
for xn when the type of xn is ( nn1 , . . . , nnd ) [4]. We denote the set of conditional types for xn by V (xn , X ).
For any conditional type V for xn , we define the subset of X n :
n
TV (xn ) := x′ ∈ X n |TY ((x1 , x′1 ), . . . , (xn , x′n )) = V · P ,
where P is the empirical distribution of xn .
According to Csiszár and Körner[4], the proposed code is constructed as follows. The main point of
this section is to establish that Csiszár-Körner’s Packing lemma [4, Lemma 10.1] provides a code whose
performance is essentially equivalent to the average performance of random coding in the sense of (30).
Lemma 2: For a positive number R > 0, there exists a sufficiently large integer N satisfying the
following. For any integer n ≥ N and any type P ∈ Tn (X ) satisfying R < H(P ), there exist Mn :=
1
en(R−n 4 ) distinct elements
M̂n := {En,P,R (1), . . . , En,P,R (Mn )} ⊂ TP
such that the inequality
|TV (xn ) ∩ (M̂n \ {xn })| ≤ |TV (xn )|e−n(H(P )−R)
(29)
holds for every xn ∈ M̂n ⊂ TP and every conditional type V ∈ V (xn , X ).
This lemma is shown in Appendix C from Csiszár and Körner[4, Lemma 10.1]. Now we define our
universal encoder En,P,R by using Lemma 2. Note that this encoder M̂n does not depend on the output
alphabet because the employed Packing lemma treats the conditional types from the input alphabet to the
input alphabet. Due to this property, as shown below, the decoding error probability is upper bounded by
a polynomial times of the average of the decoding error probability under the random coding. Now, we
transform the property (29) to a form applicable to our evaluation.
Using the encoder M̂n , we define the distribution PM̂n as
(
1
xn ∈ M̂n
Mn
PM̂n (xn ) :=
0 xn ∈
/ M̂n .
Now, we focus on the permutation group Sn on {1, . . . , n}. For any xn ∈ X n , we define an invariant
subgroup Sxn ⊂ Sn , where Sn is the permutation group with degree n:
Sxn := {g ∈ Sn |g(xn ) = xn }.
P
For V ∈ V (xn , X ), the probability g∈Sxn |Sx1n | PM̂n ◦ g(x′ n ) does not depend on the element x′ n ∈
P
P
P
TV (xn ) ⊂ TP . Since x′ n ∈TV (xn ) g∈Sxn |Sx1n | PM̂n ◦ g(x′n ) = x′ n ∈TV (xn ) PM̂n (x′ n ) = |TV (xn ) ∩ (M̂n \
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
13
P
{xn })| · M1n , any element x′ n ∈ TV (xn ) ⊂ TP satisfies g∈Sxn |Sx1n | PM̂n ◦ g(x′ n ) =
any element x′ n ∈ TV (xn ) ⊂ TP satisfies
X 1
|TV (xn ) ∩ M̂n | 1
n
PM̂n ◦ g(x′ ) =
·
|Sxn |
|TV (xn )|
Mn
g∈S n
|TV (xn )∩M̂n |
|TV (xn )|
· M1n . Thus,
x
=
(a)
|TV (xn ) ∩ (M̂n \ {xn })|
|TV (xn )|Mn
1
(b)
n
1
1
≤e−nH(P ) en 4 = P n (x′ )en 4 =
cn,P
n
1
PTP (x′ )en 4
(30)
when the conditional type V is not identical, i.e., V (x|x′ ) 6= δx,x′ , where (a) and (b) follow from (29)
and (27), respectively. Notice that cn,P is defined in (27). Relation (30) holds for any x′ n (6= xn ) ∈ TP
because there exists a conditional type V such that x′ n ∈ TV (xn ) and V is not identical.
V. α-R ÉNYI - RELATIVE - ENTROPY VERSION OF C LARKE -BARRON FORMULA
In this section, we discuss an α-Rényi-relative-entropy version of Clarke-Barron formula [3] for a family
{Pθ }θ∈Θ of distributions on a probability space Y. In the quantum paper [5], the key idea to evaluate
the decoding error probability is to upper bound the max relative entropy Dmax (Pθn kQ(n) ) between each
independent and identical distribution Pθn and a certain distribution Q(n) . That is, as shown in [5], when
the output alphabet Y is a finite set, we can find a distribution Q(n) such that the max relative entropy
Dmax (Pθn kQ(n) ) behaves as O(log n), i.e.,
Dmax (Pθn kQ(n) ) = O(log n) for ∀θ ∈ Θ.
(31)
For the detail, see Remark 2. However, when the output alphabet Y is not a finite set, it is not easy to
directly evaluate the max relative entropy. Hence, in this paper, we focus on an α-Rényi-relative-entropy
version of Clarke-Barron formula as follows. For a given distribution ν on the parameter
R space Θ, we
define the mixture distribution Qnν [{Pθ }] with the density function qνn [{Pθ }](y n ) := Θ qθn (y n )ν(θ)dθ.
When we need to clarify the family of distributions {Pθ }, we simplify it by Qnν .
Lemma 3: Assume that ν(θ) is continuous for θ and the support of ν is Θ. Then, an exponential family
{Pθ } satisfies that
k
(32)
D1+s (Pθn kQnν ), D1+n (Pθn kQnν ) ≤ log n + O(1)
2
n
1
1
k
k
+ log det Jθ + log
−
log(1 + s) + o(1)
(33)
D1+s (Pθn kQnν ) ≤ log
2
2π 2
ν(θ) 2s
for s > 0, as n goes to infinity. When the continuity for ν(θ) is uniform for θ in any compact set, the
constant O(1) on the RHS of (32) can be chosen uniformly in any compact set for θ.
The proof of Lemma 3 is given in Appendix B. Clarke-Barron’s paper [3] showed the relation
k
D(Pθn kQnν ) = log n + O(1)
(34)
2
as n goes to infinity. Since D(Pθn kQnν ) ≤ D1+s (Pθn kQnν ) for s > 0, combining (33), we have
k
D1+s (Pθn kQnν ) = log n + O(1)
(35)
2
as n goes to infinity. That is, to show (35), it is sufficient to show (33).
Now, we explain how to refine the condition (31) with a non-finite alphabet Y by using Lemma 3.
Since
Pθn ({y n |pnθ (y n ) > qνn (y n )enδ })
′
′
n
n
≤e−ns δ+s D1+s′ (Pθ kQν ) ,
(36)
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
14
(33) implies that
1
log Pθn ({y n |pnθ (y n ) > qνn (y n )enδ })
n
s′
′
≥s δ − D1+s′ (Pθn kQnν )
n
′
k
s
′
log n + O(1)
≥s δ −
n 2
→s′ δ
−
(37)
for s′ > 0. Since s′ is arbitrary, we have
1
lim − log Pθn ({y n |pnθ (y n ) > qνn (y n )enδ }) = ∞.
(38)
n→∞
n
Since δ > 0 can be chosen arbitrarily small, (38) can be regarded as the refinement of (31). (See the
definition of the max relative entropy in (1).)
Next, we consider a family of distribution {Pθ }θ∈Θ to satisfy Condition C. In this case, we choose the
discrete mixture distribution QnC as follows. Since Θ is compact, the subset Θ[n] := √1n Zk ∩ Θ has finite
k
cardinality, which increases with the order O(n 2 ).PThen, we define the distribution QnC [{Pθ }θ∈Θ ] by the
probability density function qCn [{Pθ }θ∈Θ ](y n ) := θ∈Θ[n] |Θ1[n] | pnθ (y n ) and have the following lemma by
simplifying it by QnC .
Lemma 4: When a family {Pθ } satisfies Condition E, we have
k
D1+s (Pθn kQnC ) ≤ log n + O(1)
(39)
2
for s > 0, as n goes to infinity. More precisely, there is an upper bound Aθ,n of D1+s (Pθn kQnC ) satisfying
the following. Aθ,n does not depend on s > 0, and the quantity Aθ,n − k2 log n converges uniformly for
θ.
Proof: We fix ǫ > 0. Due to the assumption and the definition of Θ[n] , we can choose an integer N
k
. Due
to Condition C, for
satisfying the following conditions. Notice that maxθ∈Θ,θ′ ∈Θ[n] kθ − θ ′ k2 = 4n
P
((max
J
ξi ξj )+ǫ)k
ξ:kξk=1
i,j θ|i,j
n ≥ N and θ ∈ Θ, we can choose θ ′ ∈ Θ[n] such that D1+s (Pθ kPθ′ )P≤
8n
((maxξ:kξk=1 i,j Jθ|i,j ξi ξj )+ǫ)k
1
pn (y n ) ≤ qCn (y n ), using the constant BC := maxθ∈Θ
, we have
|Θ[n] | θ′
8
n
n
n
n
esD1+s (Pθ kQC ) ≤ |Θ[n] |s esD1+s (Pθ kPθ′ ) ≤ |Θ[n] |s esn
BC
n
= |Θ[n] |s esBC .
. Since
(40)
That is,
D1+s (Pθn kQnC ) ≤ log |Θ[n] | + BC .
(41)
Since the compactness of Θ implies log |Θ[n] | = k2 log n + O(1), we obtain the desired argument.
Now, we consider Condition E, which is weaker than Condition C. In this case, we choose the discrete
mixture distribution QnE as follows. Since Θi is compact, the subset Θ[n,i] := √1n Zk ∩ Θi has finite
k
cardinality, which increases with the order O(n 2 ). Then, P
we defineP
the distribution QnE [{Pθ }, {Θi }] by
∞
6
1
n
i
n
n n
the probability density function qE [{Pθ }, {Θ }](y ) :=
i=1 π 2 i2
θ∈Θ[n,i] |Θ[n,i] | pθ (y ) and have the
following lemma by simplifying it by QnE .
Lemma 5: When a family {Pθ } satisfies Condition E, we have the same inequality as (39) with QnE .
More precisely, there is an upper bound Bθ,n of D1+s (Pθn kQnE ) satisfying the similar condition as Aθ,n .
The difference is that the quantity Bθ,n − k2 log n converges uniformly for θ in any compact subset. because
Θ is not necessarily compact. Hence, the obtained inequality contains the statement like (33).
Proof: For any compact subset Θ′ ⊂ Θ and ǫ > 0, we choose i such that Θ′ ⊂ Θi . We fix ǫ > 0. Due to
the assumption and the definition of Θ[n,i] , we can choose an integer N satisfying the following conditions.
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
15
k
. For n ≥ N and θ ∈ Θ′ , we can choose θ ′ ∈ Θ[n,i] such
4n
((maxξ:kξk=1 i,j Jθ|i,j ξi ξj )+ǫ)k
6
. Since π2 i2 |Θ
pn′ (y n ) ≤ qEn (y n ), using the constant
that D1+s (Pθ kPθ′ ) ≤
8n
[n,i] | θ
P
((maxξ:kξk=1 i,j Jθ|i,j ξi ξj )+ǫ)k
, we have
BE := maxθ∈Θ′
8
Notice that maxθ∈Θ,θ′ ∈Θ[n,i] kθ −Pθ ′ k2 =
n
n
esD1+s (Pθ kQE ) ≤
π 2s i2s
n
n
|Θ[n,i]|s esD1+s (Pθ kPθ′ ) .
s
6
(42)
That is,
D1+s (Pθn kQnE ) ≤ log
π 2 i2
+ log |Θ[n,i] | + BE .
6
(43)
So, we obtain the desired argument.
Remark 1: The paper [46] seems discuss a topic related to α-Rényi-relative-entropy version of ClarkeBarron formula. The main issue of the paper [46] is to evaluate D(Pθn kQnν ) not but Dα (Pθn kQnν ). For this
purpose, the paper [46] discusses the α-Rényi relative entropy Dα (Pθ kPθ′ ) between two close points θ
and θ ′ .
Remark 2: In the relation to (38), we consider the finite output alphabet case under the same assumption
as Lemma 3. We additionally assume that {Pθ } equals the set of distributions on Y. So, for any type
P ∈ Tn (Y), there exists a parameter θP ∈ Θ such that P = PθP . For a constant c > 0, a type P ∈ Tn (Y),
and an positive integer n, there exists a constant cn,P,c such that when the distribution Pθ is close to the
type P ∈ Tn (Y), i.e., D(Pθ kP ) ≥ nc , we have cn,P,cPθn (y n ) ≥ PTP (y n ). Similar to (28), the constant cn,P,c
increases only polynomially
P with respect to n.
j
Since D(Pθ kPθP ) ∼
= 21 i,j Jθ|i,j (θi − θPi )(θj − θP ), the probability ν(BP,n,c) of the set BP,n,c := {θ ∈
Θ|D(Pθ kP ) ≥Pnc } behaves as the inverse of a polynomial of n. The mixture distribution Qν is evaluated
n n
as Qnν (y n ) ≥ P ∈Tn (X ) ν(BP,n,c )c−1
n,P,c P (y ). This fact shows that for any δ > 0, there exists N such
that we have
X
enδ Qnν (y n ) ≥
P (y n) ≥ Pθn (y n )
(44)
P ∈Tn (Y)
for any integer n ≥ N and any θ ∈ Θ. Hence, Dmax (Pθn kQnν ) ≤ nδ, which is a stronger statement than
(38).
More strongly, the quantum paper [5] showed that there exists a distribution Q(n) such that Dmax (Pθn kQ(n) )
behaves as O(log n). That is, the paper [5] essentially chosen
such a distribution Q(n) in (31) to be the
P
uniform mixture of the uniform distributions PTP , i.e., P ∈Tn (Y) |Tn1(Y)| PTP [33]. This distribution Q(n)
satisfies Dmax (Pθn kQ(n) ) ≤ log |Tn (Y)|. The paper [5] employed this property to construct the universal
channel coding even in the quantum setting. However, in the infinite output alphabet case, we cannot
show (44). Hence, we need this long discussion here. The inequality (44) in finite systems also played
an essential role in the papers [33], [49], [6].
VI. U NIVERSAL DECODER
The aim of this section is to make a universal decoder, which does not depend on the parameter θ and
depends only on the distribution P , the family of channel {Wθ }, prior distributions on Θ, the coding
rate R, and another rate R1 . The other rate R1 is decided by the distribution P , the family {Wθ }, and R
as discussed in the method (M1) or (M2) (large deviation) and Section VIII (second order). The required
property holds for any choice of the prior distributions on Θ.
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
16
A. Exponential family of channels
In this subsection, we construct a decoder only for exponential family of channels, i.e., under Condition
B. In Subsection VI-B, under Condition D or F, we can construct it in the same way by replacing Qnw by
QnC or QnE due to Lemmas 4 and 5. In the decoding process, we need to extract the message that has large
correlation with detected output signal like the maximum mutual information decoder [4]. However, since
the cardinality of the output alphabet Y is not necessarily finite, we cannot directly apply the maximum
mutual information decoder because it can be defined only with finite input and output alphabets. Here,
to overcome this problem, we need to extract the message that has large correlation with detected output
signal. To achieve this aim, we firstly construct output distributions that universally approximate the true
output distribution and the mixture of the output distribution. Then, we apply the decoder for information
spectrum method [32] to these distributions. Using this idea, we construct our universal decoder. That is,
for a given number R > 0 and a distribution P on X , we define our universal decoder for our universal
encoder En,P,R given in Section IV by using the type P = ( nn1 , . . . , nnd ).
(n)
We choose prior distributions ν with support Θ. For a given xn ∈ TP , we define the distribution Qxn
on Y n as follows. First, for simplicity, we consider the case when
n
x′ = (1, . . . , 1, 2, . . . , 2, . . . , d, . . . , d) ∈ TP .
| {z }
| {z } | {z }
n1
(n)
n2
(45)
nd
In this case, Qx′ n is defined as Qnν 1 [{Wθ,1 }] × Qnν 2 [{Wθ,2 }] × · · · × Qnν d [{Wθ,d }]. For a general element
(n)
(n)
xn ∈ TP , the distribution Qxn on Y n is defined by the permuted distribution of Qx′ n by the application
n
of the permutation converting x′ to xn .
Now, to make our decoder, we apply the decoder for information spectrum method [32] to the case
(n)
when thePoutput distribution is given by Qxn and the mixture of the output distribution is given by
(n)
(n)
QP := xn ∈TP PTP (xn )Qxn . Then, we define the subset
(n)
(n)
D̂xn := {y n |qxn (y n ) ≥ enR1 qP (y n )}.
(46)
Given a number R1 > R and an element i ∈ Mn , we define the subset Di := D̂En,P,R (i) \ ∪i−1
j=1 D̂En,P,R (i) ,
−1
inductively. Finally, we define our universal decoder Dn,P,R,R1 as Dn,P,R,R
(i)
:=
D
and
our
universal
i
1
1
code Φn,P,R,R1 := (enR−n 4 , En,P,R , Dn,P,R,R1 ).
Now, we discuss several important properties related to our universal decoder by using Lemma 3.
For xn ∈ TP and s > 0, (33) of Lemma 3 guarantees that
(n)
n
D1+s (Wθ,x
n kQxn )
X k
≤
log(nP (x)) + O(1).
2
x∈X
(47)
Hence, similar to (37), (47) implies that
1
n
n
n
n
nδ (n) n
qxn (y )}) ≥ s′ δ
lim inf − log Wθ,x
n ({y |wθ,xn (y ) > e
n→∞
n
′
′
for s > 0. Since s is arbitrary, we have
1
n
nδ (n) n
n
n
n
qxn (y )}) = ∞.
lim − log Wθ,x
n ({y |wθ,xn (y ) > e
n→∞
n
Modifying the derivation in (37), we have
√
1
n
n
n
n
nδ (n) n
qxn (y )}) ≥ s′ δ + o(1)
− √ log Wθ,x
n ({y |wθ,xn (y ) > e
n
(48)
(49)
(50)
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
17
for s′ > 0. Here, the term o(1) in (50) can be chosen as a arbitrary small constant uniformly with respect
to n and θ in any compact
3.
Pset due to Lemma
(n)
(n)
n
Further, since QP = xn ∈TP PTP (x )Qxn , the information processing inequality for Rényi relative
(n)
(n)
(n)
P
n
n
n
entropy yields that esD1+s (Wθ ·PTP kQP ) ≤ xn ∈TP PTP (xn )esD1+s (Wθ,xn kQxn ) = esD1+s (Wθ,xn kQxn ) , which
implies
(n)
D1+s (Wθn · PTP kQP )
X k
≤
log(nP (x)) + O(1).
2
x∈X
(51)
−1
(n)
log Wθn · PTP {y n |wθn · PTP (y n ) > enδ qP (y n )} = ∞.
n→∞ n
(52)
Here, the constant O(1) can be chosen uniformly with respect to θ in any compact set due to Lemma 3.
Similar to (49), (51) implies that
lim
B. General assumption
Next, under more general assumption, i.e., Conditions D and F, we construct the universal decoder
by replacing Qnw by QnC or QnE due to Lemmas 4 and 5. Under Condition D, first, for simplicity, we
(n)
consider the case of (45). In this case, Qx′ n (y n ) is defined as QnC1 [{Wθ,1 }θ∈Θ ] × QnC2 [{Wθ,2 }θ∈Θ ] ×
(n)
· · · × QnCd [{Wθ,d }θ∈Θ ]. For a general element xn ∈ TP , the distribution Qxn on Y n is defined by the
(n)
(n)
permuted distribution of Qx′ n by the application of the permutation converting x′ n to xn . Using QP :=
P
(n)
n
xn ∈TP PTP (x )Qxn , we define the decoder as (46). So, in the same way, we can show (49), (50), (51),
and (52) by using (39) of Lemma 4. In particular, the constants o(1) and O(1) in (50) and (51) can be
chosen uniformly with respect to θ due to the same reason, respectively.
(n)
Under Condition F, in the case of (45), the distribution Qx′ n is defined as QnE1 [{Wθ,1 }, {Θi }] ×
(n)
QnE2 [{Wθ,2 }, {Θi }] × · · · × QnEd [{Wθ,d }, {Θi }]. For a general element xn ∈ TP , the distribution Qxn
(n)
on Y n is defined by the permuted distribution of Qx′ n by the application of the permutation converting
P
(n)
(n)
x′ n to xn . Using the distribution QP := xn ∈TP PTP (xn )Qxn , we define the decoder as (46). So, in the
same way, we can show (49), (50), (51), and (52) by using Lemma 4. Again, the constants o(1) and O(1)
in (50) and (51) can be chosen uniformly with respect to θ due to the same reason, respectively.
VII. E RROR EXPONENT
In this section, using the property (28) of type, the encoder property (30), and the decoder properties
(49) and (52), we will prove Theorem 1, i,e., show that the code Φn,P,R,R1 satisfies that
−1
log eθ (Φn,P,R,R1 ) ≥ min( max (sI1−s (P, Wθ ) − sR1 ), R1 − R).
n→∞ n
s∈[0,1]
lim
(53)
Since the decoder properties (49) and (52) holds for all of Conditions B, D, and F, the following proof
is valid under Conditions B, D, and F.
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
18
First, we have
Mn
X
1
n
Wθ,E
(Dic )
n,P,R (i)
M
n
i=1
Mn
X
1 n
≤
(Wθ,En,P,R (i) (D̂Ec n,P,R (i) )
M
n
i=1
X
n
+
Wθ,E
(
D̂
))
E
(j)
n,P,R
n,P,R (i)
j6=i
=
Mn
X
1 n
Wθ,En,P,R (i) (D̂Ec n,P,R (i) )
M
n
i=1
+
Mn
X
1 X n
Wθ,En,P,R (i) (D̂En,P,R (j) ) .
Mn i6=j
j=1
(54)
In the following, we evaluate the first term of (54). Any element xn ∈ TP satisfies that
c
n
Wθ,x
n (D̂xn )
(n)
(n)
n
n
n
nR1
=Wθ,x
qP (y n )})
n ({y |qxn (y ) < e
(n)
n
n(R1 +δ)
n
n
n
qP (y n )})
≤Wθ,x
n ({y |wθ,xn (y ) < e
(n)
n
n
n
n
nδ
+ Wθ,x
qxn (y n )})
n ({y |wθ,xn (y ) > e
(n)
(n)
(55)
(n)
n
nδ
n
n
n
qxn (y n ), the condition qxn (y n ) < enR1 qP (y n ) implies wθ,x
because when wθ,x
n (y ) ≤ e
n (y ) <
(n)
en(R1 +δ) qP (y n ). To get the latter exponent, we choose any elements s ∈ [0, 1] and xn ∈ TP . Using
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
the measure µn (dy n ) := µ(dy1) · · · µ(dyn ), we have
19
(n)
n
n(R1 +δ)
n
n
n
qP (y n )})
Wθ,x
n ({y |wθ,xn (y ) < e
Z
n 1−s ns(R1 +δ) (n) n s n
n
e
qP (y ) µ (dy n )
≤
wθ,x
n (y )
Yn
Z
1
(a) X
n 1−s ns(R1 +δ) (n) n s n
n
=
e
qP (y ) µ (dy n )
wθ,x
n (y )
|TP | Y n
xn ∈T
P
(b)
≤|Tn (X )|
Z
X
n n
P (x )
·
Yn
xn ∈X n
(c)
(n)
n
n 1−s ns(R1 +δ)
wθ,x
e
qP (y n )s µn (dy n )
n (y )
=|Tn (X )|ens(R1 +δ)
Z X
(n)
n 1−s
n n
n
qP (y n )s µn (dy n )
P (x )wθ,xn (y )
·
Yn
(d)
xn ∈X n
≤|Tn (X )|ens(R1 +δ)
1
Z X
1−s
n 1−s 1−s n
n
P n (xn )wθ,x
·
µ (dy n)
n (y )
·
Z
Yn
xn ∈X n
(n)
Yn
s
qP (y n ) s µn (dy n )
1s
=|Tn (X )|ens(R1 +δ)
Z X
1
1−s
n
n 1−s 1−s n
·
µ (dy n)
P n (xn )wθ,x
n (y )
Yn
xn ∈X n
ns(R1 +δ)
=|Tn (X )|e
1
n(1−s)
Z X
1−s
µ(dy)
·
P (x)wθ,x (y)1−s
Y
x∈X
n(s(R1 +δ)−sI1−s (P,Wθ ))
=|Tn (X )|e
,
(56)
where (b) and (d) follow from (28) and the Hölder inequality, respectively. Step (c) follows from the
exchange of the orders of finite sum
R andn the integral.
n 1−s ns(R1 +δ) (n) n s n
Since the value of the integral Y n wθ,x
e
qP (y ) µ (dy n ) does not change when the
n (y )
order of y1 , . . . , yn , we have the relation
Z
Z
n
n 1−s ns(R1 +δ) (n) n s n
n
n
n 1−s ns(R1 +δ) (n) n s n
wθ,xn (y ) e
qP (y ) µ (dy ) =
wθ,x
e
qP (y ) µ (dy n )
n′ (y )
Yn
Yn
with xn′ 6= xn ∈ TP , which implies Step (a).
n
n
n
n
c
n
Due to (49) and (55), the exponential decreasing rate of Wθ,x
n (D̂xn ) equals that of wθ,xn ({y |wθ,xn (y ) <
(n)
en(R1 +δ) qP (y n )}). Thus, using (56), we have
1
c
n
lim − log Wθ,x
n (D̂xn ) ≥ −s(R1 + δ) + sI1−s (P, Wθ ).
n→∞
n
Since δ > 0 is arbitrary, we have
1
c
n
lim − log Wθ,x
n (D̂xn ) ≥ −sR1 + sI1−s (P, Wθ ).
n→∞
n
(57)
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
20
Next, we proceed to the second term of (54). In the following, we simplify Wθ · P to be Wθ,P .
Mn X
1 X
n
Wθ,E
(D̂En,P,R (j) )
n,P,R (i)
Mn j=1
i6=j
(a)
=
Mn X
X
1
X
j=1 i6=j g∈SEn,P,R (j)
=
Mn
X
X
j=1 g∈SEn,P,R (j)
1
(b) en 4
≤
cn,P
1
n4
Mn X
X
j=1 xn ∈X n
1
n
Wθ,g(E
(D̂En,P,R (j) )
n,P,R (i))
|SEn,P,R (j) | Mn
1
|SEn,P,R (j) |
X
x′ n (6=En,P,R (j))∈TP
n
n
PM̂n (g(x′ ))Wθ,x
′ n (D̂En,P,R (j) )
n
PTP (xn )Wθ,x
n (D̂En,P,R (j) )
Mn
e X
=
W n · PTP (D̂En,P,R (j) )
cn,P j=1 θ
1
=en 4
Mn
X
j=1
1
n4
(n)
(n)
Wθn · PTP {y n |qEn,P,R (j) (y n ) ≥ enR1 qP (y n )}
Mn
e X
(n)
≤
Wθn · PTP {y n |qEn,P,R (j) (y n ) ≥ en(R1 −δ) wθn · PTP (y n )}
cn,P j=1
1
Mn
en 4 X
(n)
Wθn · PTP {y n |wθn · PTP (y n ) > enδ qP (y n )}
+
cn,P j=1
1
Mn
en 4 X
(n)
(n)
≤
e−n(R1 −δ) QEn,P,R (j) {y n |qEn,P,R (j) (y n ) ≥ en(R1 −δ) wθn · PTP (y n )}
cn,P j=1
1
Mn
en 4 X
(n)
Wθn · PTP {y n |wθn · PTP (y n ) > enδ qP (y n )}
+
cn,P j=1
1
1
Mn
en 4 X
en 4
(n)
≤
e−n(R1 −δ) +
Mn Wθn · PTP {y n |wθn · PTP (y n ) > enδ qP (y n )}
cn,P j=1
cn,P
1
√
en 4
(n)
Mn e−n(R1 −δ) + e n Mn Wθn · PTP {y n |wθn · PTP (y n ) > enδ qP (y n )}
=
cn,P
(58)
n
n
where (a) follows from the relation Wθ,g(E
(D̂En,P,R1 (j) ) = Wθ,E
(D̂En,P,R1 (j) ) for g ∈ SEn,P,R (j)
n,P,R (i))
n,P,R (i)
and (b) follows from (30).
(n)
Since Wθn · PTP {y n |wθn · PTP (y n ) > enδ qP (y n )} satisfies the condition (52) and
1
n4
limn→∞ n1 log cen,P Mn = limn→∞ n1 log Mn < ∞,
1
en 4
−1
(n)
log
Mn Wθn · PTP {y n |wθn · PTP (y n ) > enδ qP (y n )} = ∞.
lim
n→∞ n
cn,P
(59)
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
21
1
Since
limn→∞ −1
n
log
en 4
cn,P
Mn e−n(R1 −δ) = R1 − δ − R, (58) and (59) imply that
Mn X
−1
1 X
Wn
(D̂En,P,R (j) ) ≥ R1 − δ − R.
lim
log
n→∞ n
Mn j=1 i6=j θ,En,P,R (i)
(60)
Since δ > 0 is arbitrary, we have
Mn X
−1
1 X
lim
Wn
(D̂En,P,R (j) ) ≥ R1 − R.
log
n→∞ n
Mn j=1 i6=j θ,En,P,R (i)
(61)
Combining (54), (57), and (61), we obtain (53).
VIII. S ECOND ORDER
In this section, we show Theorems 2 and 3. First, we show Theorem 2. Next, we show Theorem 3
by modifying the proof of Theorem 2. These proofs are based on the decoder properties (50) and (51)
as well as the discussion in Section VII. To satisfy the condition of Theorem 2, we modify the encoder
given in Section
IV and the decoder given in Section VI as follows.
For our encoder, we choose R to
R∗
R∗
1
be R1∗ + √n2 , and for our decoder, we choose R1 to be R1∗ + √n2 + n2/3
. The modified code is denoted by
Φn,P,R∗1 ,R∗2 .
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
22
A. Evaluation of second term in (54)
We evaluate the first and the second terms in (54), separately. Here, we show that the second term in
(54) goes to zero. Similar to (58), we have
Mn X
1 X
Wn
(D̂En,P,R (j) )
Mn j=1 i6=j θ,En,P,R (i)
1
Mn
en 4 X
W n · PTP (D̂En,P,R (j) )
=
cn,P j=1 θ
1
Mn Z
en 4 X
(n)
(n)
=
wθn · PTP (y n )qP (y n )s−1QP (y n )1−s µn (dy n )
cn,P j=1 D̂E
n,P,R (j)
1
Z
Mn
s Z
1−s
(a) en 4 X
(n) n 1−s
n
n
n
n
n
n 1s (n) n s−1
1−s
s
qP (y ) µ (dy )
≤
wθ · PTP (y ) qP (y ) µ (dy )
cn,P j=1
D̂En,P,R (j)
D̂En,P,R (j)
1
Mn Z
s
1
s−1
en 4 X
(n)
(n)
wθn · PTP (y n ) s QP (y n ) s µn (dy n ) QP (D̂En,P,R (j) )1−s
≤
cn,P j=1
n
Y
1
Mn
(n)
en 4 X
s( 1 −1)D 1 (Wθn ·PTP kQP ) (n)
s
e s
QP (D̂En,P,R (j) )1−s
=
cn,P j=1
1
Mn
(n)
en 4 X
(1−s)D 1 (Wθn ·PTP kQP ) (n)
s
=
e
QP (D̂En,P,R (j) )1−s
cn,P j=1
1
(n)
en 4
(1−s)D 1 (Wθn ·PTP kQP ) (n)
s
Mn e
QP (D̂En,P,R (j) )1−s
=
cn,P
1
1
en 4 nR∗1 +√nR∗2 −n 41 (1−s)D 1 (Wθn ·PTP kQ(n)
∗
∗ √
(n)
P ) −nR1 − nR2 −n 3
s
e
QEn,P,R (j) (D̂En,P,R (j) )1−s
≤
e
e
cn,P
1 −n 13 (1−s)D 1 (Wθn ·PTP kQ(n)
P )
s
=
e
,
e
cn,P
(62)
where (a) follows from Hölder inequality.
1
Here, we choose s to be n+1
. Since (51) implies
(n)
(1 − s)D 1 (Wθn · PTP kQP ) = (1 −
s
≤(1 −
1
(n)
)D1+n (Wθn · PTP kQP )
n+1
1 k
) log n + O(1),
n+1 2
(63)
(62) yields
Mn X
1 X
n
lim
Wθ,E
(D̂En,P,R (j) ) = 0.
n,P,R (i)
n→∞ Mn
j=1 i6=j
(64)
That is, the second term of (54) goes to zero.
B. Evaluation of first term in (54)
Now, we evaluate the first term in (54), which is upper bounded by two terms in√(55). So, we evaluate
the first and the second terms in (55). For an arbitrary δ2 > 0, we substitute δ2 / n into δ in (55). So,
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
23
(50) implies that the second term in (55) goes to zero. Further, for an arbitrary δ3 > 0, the first term in
(55) is upper bounded as
√
n(R∗2 +δ2 ) (n) n
qP (y )})
√
∗
∗
n
(y n )})
enR1 + n(R2 +δ2 +δ3 ) wθ,P
∗
n
nR1 +
n
n
n
Wθ,x
n ({y |wθ,xn (y ) < e
n
n
n
n
≤Wθ,x
n ({y |wθ,xn (y ) <
n
n
n
−
n
+ Wθ,x
n ({y |wθ,P (y ) < e
√
nδ3 (n) n
qP (y )}).
(65)
We also have
√
(n)
n
n
n
− nδ3
n
qP (y n )})
Wθ,x
n ({y |wθ,P (y ) < e
X 1
√
n
n
n
− nδ3 (n) n
n
qP (y )})
=
Wθ,x
n′ ({y |wθ,P (y ) < e
|TP |
n′
x ∈TP
X
√
n
n
n
n
− nδ3 (n) n
≤|Tn (X )|
P n (x)Wθ,x
qP (y )})
n′ ({y |wθ,P (y ) < e
xn′ ∈X n
n
n
=|Tn (X )|Wθ,P
({y n |wθ,P
(y n ) < e−
√
− nδ3
≤|Tn (X )|e
√
− nδ3
≤|Tn (X )|e
(n)
√
n
QP ({y n |wθ,P
(y n )
nδ3 (n) n
qP (y )})
√
(n)
< e− nδ3 qP (y n )})
→ 0.
(66)
√
Therefore, due to (65) and (66), the limit of the
first term in (55) with δ = δ2 / n not larger than the
√
n
nR∗1 + n(R∗2 +δ2 +δ3 ) n
n
n
n
wθ,P (y n )}). So, (55) and (50) guarantee that
limit of limn→∞ Wθ,xn ({y |wθ,xn (y ) < e
∗
n
nR1 +
n
n
n
c
n
lim Wθ,x
n (D̂xn ) ≤ lim Wθ,xn ({y |wθ,xn (y ) < e
√
n(R∗2 +δ2 +δ3 )
n→∞
n→∞
n
wθ,P
(y n )})
(67)
for arbitrary δ2 > 0 and δ3 > 0.
When I(P, Wθ ) > R1∗ ,
∗
n
nR1 +
n
n
n
Wθ,x
n ({y |wθ,xn (y ) < e
√
n(R∗2 +δ2 +δ3 )
n
wθ,P
(y n )}) → 0.
Now, we show that any element xn ∈ TP satisfies that
nR∗1 +
n
n
n
n
Wθ,x
n ({y |wθ,xn (y ) < e
√
n(R∗2 +δ2 +δ3 )
n
wθ,P
(y n )}) →
Z
R∗ +δ2 +δ3
√2
V (P,Wθ )
−∞
1
x2
√ exp(− )dx
2
2π
(68)
(69)
Pn
w n n (Y n )
wθ,xi (Yi )
1
∗
when I(P, Wθ ) = R1∗ . We have n1 log wθ,x
n (Y n ) − R1 = n
i=1 (log wθ,P (Yi ) − D(Wθ,xi kWθ,P )). When n
θ,P
goes to infinity, each element x ∈ X appears in xn ∈ TP infinitely times. So, the central limit theorem
implies (69) with the assumptions of Theorem 2, So, when I(P, Wθ ) = R1∗ , combining (67), (66), and
(69), we have
Z √R∗2 +δ2 +δ3
V (P,Wθ )
x2
1
c
n
√ exp(− )dx.
lim Wθ,x
(70)
n (D̂xn ) ≤
n→∞
2
2π
−∞
In summary, since δ2 and δ3 are arbitrary in (70), we have
n
c
limn→∞ Wθ,x
=0
n (D̂xn )
R∗
2
√
R
V
(P,W
θ)
c
n
≤ −∞
limn→∞ Wθ,x
n (D̂xn )
when I(P, Wθ ) > R1∗
√1
2π
2
exp(− x2 )dx when I(P, Wθ ) = R1∗ .
(71)
Hence, the relations (54), (71), and (64) yield (23). Therefore, we complete our proof of Theorem 2.
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
24
C. Proof of Theorem 3
Now, we proceed to the proof of Theorem 3. Similarly, we evaluate the average error probability by
discussing the first and second terms of (54), separately. Fortunately, the upper bound in the RHS of (62)
goes to zero uniformly with respect to θ in any compact set because the constant O(1) in (51) can be
chosen uniformly with respect to θ in any compact set. So, the second term of (54) goes to zero uniformly
with respect to θ in any compact set.
Now, we evaluate the first term of (54), which is upper bounded by two terms in (55). Due to (50), the
second term of (55) goes to zero uniformly with respect to θ in any compact set because the term o(1)
in (50) can be chosen as a arbitrary small constant uniformly with respect to θ in any compact set. Then,
the remaining term is the first term of (55), which is upper bounded by two terms in (65). The second
n
n
n
n
term of
(65) goes to zero due to (66). Therefore, it is enough to evaluate limn→∞ Wθ,x
n ({y |wθ,xn (y ) <
√
∗
∗
n
enR1 + n(R2 +δ2 +δ3 ) wθ,P
(y n )}), which is the first term of (65).
Since the function θ 7→ I(P, Wθ ) is a C 1 function on Θ and the function θ 7→ V (P, Wθ ) is a continuous
w n n (Y n )
√1 f (θ2 ) + o( √1 ) and
function on Θ, the likelihood ratio n1 log wθ,x
n (Y n ) has the expectation I(P, Wθ1 ) +
n
n
θ,P
the variance n1 V (P, Wθ ) + o( n1 ) with θ = θ1 +
√1 θ2
n
for xn ∈ TP . This property yields the relation
∗
√
∗
n
n
n
n
nR1 + nR2 n
lim wθ,x
wθ,P (y n )})
n ({y |wθ,xn (y ) < e
n→∞
when I(P, Wθ1 ) > R1∗
0 R∗ −f (θ )
Z √2 2
=
1
x2
V (P,Wθ )
1
√
)dx when I(P, Wθ1 ) = R1∗
exp(−
2
2π
−∞
(72)
for xn ∈ TP . The convergence is uniform with respect to θ2 in any compact set. Since δ2 and δ3 are
arbitrary, we obtain the desired argument. Since the uniformity in any compact subset has been discussed
in the above discussion. the proof of Theorem 3 is completed, now.
IX. C ONCLUSION
We have proposed a universal channel coding for general output alphabet including continuous output
alphabets. Although our encoder is the same as the encoder given in the previous paper [5], we cannot
directly apply the decoder given in [5] because it is not easy to make a distribution that universally
approximates any independent and identical distribution in the sense of maximum relative entropy in the
continuous alphabet. To overcome the difficulty, we have invented an α-Rényi-relative-entropy version of
Clarke and Barron’s formula for Bayesian average distribution. That is, we have shown that the Bayesian
average distribution well approximates any independent and identical distribution in the sense of αRényi relative entropy. Then, we have made our universal decoder by applying the information spectrum
method to the Bayesian average distribution. We have lower bounded the error exponent of our universal
code, which implies that our code attains the mutual information rate. Since our approach covers the
discrete and continuous cases and the exponential and the second-order type evaluations for the decoding
error probability, our method provides a unified viewpoint for the universal channel coding, which is an
advantage over the existing studies [4], [13], [17], [18], [19], [20], [21], [22], [24].
Further, we have introduced the parametrization θ1 + √θ2n for our channel, which is commonly used in
statistics for discussing the asymptotic local approximation by normal distribution family [39], [30]. This
parametrization matches the second order parameterization of the coding rate. So, we can expect that this
parametrization is applicable to the case when we have an unknown small disturbance in the channel.
Here, we compare our analysis on universal coding with the compound channel [37], [38], [21]. In
the compound channel, we focus on the worst case of the average error probability for the unknown
channel parameter. Hence, we do not evaluate the average error probability when the channel parameter
is not the worst case. However, in the universal coding [4], we evaluate the error probability for all
possible channels. Hence, we can evaluate how better the average error probability of each case is than
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
25
the worst case. In particular, our improved second order analysis in Theorem 3 clarifies its dependence
of the unknown parameter θ2 .
ACKNOWLEDGMENT
The author is grateful for Dr. Hideki Yagi to explaining MIMO Gaussian channels and informing
the references [13], [15], [16], [17], [18], [19], [20], [21], [22]. He is grateful to the referee for Dr.
Vincent Tan for helpfu comments. He is also grateful to the referee of the first version of this paper for
helpful comments. The author is partially supported by JSPS Grants-in-Aid for Scientific Research (A)
No. 23246071 and (A) No.17H01280. The author is also partially supported by the National Institute of
Information and Communication Technology (NICT), Japan.
A PPENDIX A
P ROOF OF L EMMA 1
When inf θ∈Θ0 I(P, Wθ ) ≤ R, all terms in (20) are zero. So, we can assume that inf θ∈Θ0 I(P, Wθ ) > R
without loss of generality.
Firstly, we show that
max min( max (sI1−s (P, Wθ ) − sR1 ), R1 − R)
R1
s∈[0,1]
1
(sI1−s (P, Wθ ) − sR).
s∈[0,1] 1 + s
= max
(73)
Since the function R1 7→ maxs∈[0,1] (sI1−s (P, Wθ ) − sR1 ) is monotone decreasing and continuous and the
function R1 7→ R1 − R is monotone increasing and continuous, there exists a real number R1∗ > R such
that maxs∈[0,1] (sI1−s (P, Wθ ) − sR1∗ ) = R1∗ − R. We also choose s∗ := argmaxs∈[0,1] (sI1−s (P, Wθ ) − sR1∗ ).
Here, we assume that s∗ ∈ (0, 1). s∗ satisfies dsI1−sds(P,Wθ ) |s=s∗ = R1∗ . Since (s∗ I1−s∗ (P, Wθ ) − s∗ R1∗ ) =
(s∗ I1−s∗ (P,Wθ )−s∗ R)
R−s∗ I1−s∗ (P,Wθ )
. So, maxs∈[0,1] (sI1−s (P, Wθ ) − sR1∗ ) =
and
R1∗ − R, we have R1∗ =
∗
1+s
1+s∗
R−s∗ I1−s∗ (P,Wθ )
dsI1−s (P,Wθ )
|s=s∗ =
.
ds
1+s∗
(1+s)
dsI1−s (P,Wθ )
−sI
(P,W )−R
1−s
θ
θ )−sR)
ds
with respect to s is
and (1 +
Since the first derivative of (sI1−s (P,W
1+s
(1+s)2
dsI1−s (P,Wθ )
(sI1−s (P,Wθ )−sR)
s)
−sI1−s (P, Wθ )−R is monotone decreasing for s ∈ [0, 1], s∗ := argmaxs∈[0,1]
ds
1+s
R−s∗ I1−s∗ (P,Wθ )
(sI1−s (P,Wθ )−sR)
dsI1−s (P,Wθ )
. So, we find that maxs∈[0,1]
|s=s∗ =
=
satisfies the same condition
ds
1+s∗
1+s
(s∗ I1−s∗ (P,Wθ )−s∗ R)
∗
∗
. Thus, we obtain (73) when s ∈ (0, 1). When s = 0, we can show s∗ = 0, which
1+s∗
implies (73). Similarly, we can show (73) when s∗ = 1.
Since
max inf min( max (sI1−s (P, Wθ ) − sR1 ), R1 − R) ≤ inf max min( max (sI1−s (P, Wθ ) − sR1 ), R1 − R),
R1
θ∈Θ0
θ∈Θ0
s∈[0,1]
R1
s∈[0,1]
(74)
it is sufficient to show there exists R1 such that
1
(sI1−s (P, Wθ ) − sR).
θ∈Θ0 s∈[0,1] 1 + s
inf min( max (sI1−s (P, Wθ ) − sR1 ), R1 − R) ≥ inf max
θ∈Θ0
s∈[0,1]
(75)
1
(sI1−s (P, Wθ ) − sR). Given a parameter θ ∈ Θ0 , using
We choose R1 to be R + inf θ∈Θ0 maxs∈[0,1] 1+s
1
the function f (s, θ) := 1+s (sI1−s (P, Wθ ) − sR) and sθ := argmaxs∈[0,1] f (s, θ), we have
f (sθ , θ) ≥ ′inf max
f (s′ , θ ′ ),
′
θ ∈Θ0 s ∈[0,1]
(76)
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
26
which implies that
max (sI1−s (P, Wθ ) − sR1 ) ≥ (sθ I1−sθ (P, Wθ ) − sθ R1 )
s∈[0,1]
=(sθ I1−sθ (P, Wθ ) − sθ R) + sθ ′inf max
f (s′ , θ ′ )
′
θ ∈Θ0 s ∈[0,1]
′
=f (sθ , θ) + sθ (f (sθ , θ) − ′inf max
f (s , θ ′ )) ≥ f (sθ , θ)
′
θ ∈Θ0 s ∈[0,1]
1
(sI1−s (P, Wθ ) − sR) = R1 − R.
≥ inf max
θ∈Θ0 s∈[0,1] 1 + s
(77)
Thus,
1
(sI1−s (P, Wθ ) − sR),
θ∈Θ0 s∈[0,1] 1 + s
min( max (sI1−s (P, Wθ ) − sR1 ), R1 − R) = inf max
s∈[0,1]
(78)
which implies (75).
A PPENDIX B
P ROOF OF α-R ÉNYI RELATIVE ENTROPY VERSION OF C LARKE -BARRON FORMULA
A. Preparation
To show Lemma 3, we prepare several formulas used in proofs of (32) and (33). Under the assumption
of exponential family, we consider the logarithmic derivatives lθ (y) := (lθ,1 (y), . . . , lθ,k (y)), where
∂
∂
lθ,j (y) := j log pθ (y) = gj (y) − j φ(θ).
(79)
∂θ
∂θ
The Fisher information matrix Jθ,i,j is given as
Z
∂
∂2
l (y)pθ (y)µ(dy).
(80)
Jθ,i,j := i j φ(θ) =
i θ,j
∂θ ∂θ
Y ∂θ
Due to the above assumption, Jθ,i,j is continuous for θ ∈ Θ. Hence, we have
∂2
log pθ (y) = −Jθ,i,j
(81)
∂θi ∂θj
which is independent of y.
Define
n
1 X ∂
lθ,j;n (y n ) := √
log pθ (yi ).
(82)
n i=1 ∂θj
Since pθ is an exponential family, we define
n
Jθ,j,j ′;n
1 X ∂2
log pθ (yi ),
:= −
n i=1 ∂θj ∂θj ′
(83)
which is also independent of for y n ∈ Y n .
In the following discussion, we employ the Laplace approximation (Laplace method of approximation).
To use the Taylor expansion at θ0 , we choose θ1 (θ) between θ and θ0 . Then, we have
pnθ (y n )
ν(θ)
pnθ0 (y n )
Pn
=
e
i=1
log pθ (yi )
pnθ0 (y n )
Pn
=
e
ν(θ)
i=1 log pθ0 (yi )+
Pk
j
j
j=1 (θ −θ0 )
Pn
∂
i=1 ∂θ j
log pθ (yi )|θ=θ0 +
Pk
pnθ0 (y n )
T
=e−n(θ−θ0 )
Jθ (θ)
√
1
(θ−θ0 )+ n(θ−θ0 )T lθ0 ;n (y n )
2
ν(θ),
j,j ′ =1
′
j
j′
(θ j −θ0 )(θ j −θ0 )
2
Pn
∂2
i=1 ∂θ j ∂θ j ′
log pθ (yi )|θ=θ1
ν(θ)
(84)
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
27
where the final equation follows from the properties of lθ;n (y n ) := (lθ,j;n (y n )) and the matrix Jθ .
Next, for an arbitrary ǫ > 0, we choose a neighborhood Uθ0 ,δ := {θ|kθ − θ0 k ≤ δ} such that Jθ ≤
Jθ0 (1 + ǫ) and ν(θ) ≥ ν(θ0 )(1 − ǫ) for θ ∈ Uθ0 ,δ . For θ ∈ Uθ0 ,δ , we have
T
e−n(θ−θ0 )
Jθ (θ)
√
1
(θ−θ0 )+ n(θ−θ0 )T lθ0 ;n (y n )
2
ν(θ)
Jθ (1+ǫ)
√
−n(θ−θ0 )T 0 2
(θ−θ0 )+ n(θ−θ0 )T lθ0 ;n (y n )
≥e
=e
1
l
(y n )T Jθ−1 (1+ǫ)−1 lθ0 ;n (y n )
2 θ0 ;n
0
ν(θ0 )(1 − ǫ)
√
−( n(θ−θ0 )−(Jθ0 (1+ǫ))−1 lθ0 ;n (y n ))T
e
(85)
Jθ (1+ǫ) √
0
( n(θ−θ0 )−(Jθ0 (1+ǫ))−1 lθ0 ;n (y n ))
2
ν(θ0 )(1 − ǫ).
(86)
B. Proof of (32)
Now, we prove (32). We focus on the set Byn := {z ∈ Rk |z T lθ0 ;n (y n ) ≥ 0, kzk ≤ 1} for y n . For
n ≥ δ12 , we have
Z
qνn (y n )
pnθ (y n )
=
ν(θ)dθ
n
n
pnθ0 (y n )
Θ pθ0 (y )
Z
pnθ (y n )
≥
ν(θ)dθ
n
n
Uθ0 ,δ pθ0 (y )
Z
J (1+ǫ)
(a)
√
T θ0
(θ−θ0 )+ n(θ−θ0 )T lθ0 ;n (y n )
2
ν(θ0 )(1 − ǫ)dθ
≥
e−n(θ−θ0 )
Uθ0 ,δ
Z
Jθ (1+ǫ)
(b) 1
−z T 0 2
z+z T lθ0 ;n (y n )
e
= k
ν(θ0 )(1 − ǫ)dz
√
n 2 kzk≤ nδ
Z
J (1+ǫ)
(c) 1
T θ0
z+z T lθ0 ;n (y n )
2
ν(θ0 )(1 − ǫ)dz
e−z
≥ k
n 2 Byn
Z
kJθ k(1+ǫ)
(d) 1
0
−
2
≥ ke
ν(θ0 )(1 − ǫ)dz
n2
By n
(e)
≥
1
k
n2
−
e
kJθ k(1+ǫ)
0
2
k
π2
(1 − ǫ),
2Γ( k2 + 1)
(87)
√
where (a), (b), (c), and (d) follow from (84) and (85), the relation z = n(θ − θ0 ), the relation n ≥ δ12 ,
and the relation kJθ0 k ≥ z T Jθ0 z for kzk ≤ 1, respectively. The inequality (e) is shown because the
k
volume of Byn is 2Γ(πk2+1) . That is, we have
2
kJθ k(1+ǫ) 2Γ( k + 1)
pnθ0 (y n )
0
2
2
≤e
k ·
k .
n
n
n 2 qν (y )
(1 − ǫ)π 2
1
Therefore, for n ≥
1
,
δ2
(88)
using (88) we have
k
n
n
n−s 2 esD1+s (Pθ0 kQν )
n
pθ0 (Y n ) s
−s k2
=n EPθn
0
qνn (Y n )
kJ k(1+ǫ)
θ0
2Γ( k2 + 1) s
s
2
≤EPθn e
k
0
(1 − ǫ)π 2
kJθ k(1+ǫ)
2Γ( k2 + 1) s
0
s
2
≤e
.
k
(1 − ǫ)π 2
(89)
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
28
Substituting n into s in (89), we have
kJ k(1+ǫ)
θ0
2Γ( k2 + 1)
k
1
−n k2 nD1+n (Pθn0 kQn
ν)
2
− log n = log n
≤ log e
e
.
k
2
n
(1 − ǫ)π 2
kJθ k(1+ǫ)
0
2Γ( k2 +1)
2
is a constant, we obtain (32).
Since log e
k
D1+n (Pθn kQnν )
(1−ǫ)π 2
C. Proof of (33)
When the continuity in the assumption is uniform for θ in any compact subset, we can choose a
common constant δ > 0 in a compact subset in Θ. Since the constant δ decides the range for n in the
above discussion, the constant on the RHS of (32) can be chosen uniformly for θ in any compact subset.
Next, for a deeper analysis for proving (33), we fix an arbitrary small real number ǫ > 0. Then, we
choose a sufficiently large real number R and a large integer N1 such that the complement Cnc of the set
Cn := {y n |klθ0 ;n (y n )k < R} satisfies
n
Pθ0 (Cnc ) ≤ ǫ
n
(90)
for n ≥ N1 . Then, we evaluate pqnν (y(yn)) under the assumption klθ0 ;n (y n )k < R as follows. Then, we can
θ0
choose sufficiently large N2 such that
Z
Jθ (1+ǫ)
−(z−(Jθ0 (1+ǫ))−1 lθ0 ;n (y n ))T 0 2
(z−(Jθ0 (1+ǫ))−1 lθ0 ;n (y n ))
e
dz
√
kzk≤ nδ
≥
k
(2π) 2
k
1
(det(Jθ0 )) 2 (1 + ǫ) 2
(1 − ǫ)
(91)
for n ≥ N2 and y n satisfying that klθ0 ;n (y n )k < R because the limit of LHS of (91) is
Thus, when n ≥ N2 , we have
Z
pnν (y n )
pnθ (y n )
=
w(θ)dθ
n
n
pnθ0 (y n )
Θ pθ0 (y )
Z
pnθ (y n )
≥
ν(θ)dθ
n
n
Uθ0 ,δ pθ0 (y )
(a)
1
n T
−1
−1
k
(2π) 2
1
k
(det(Jθ0 )) 2 (1+ǫ) 2
.
n
≥e 2 lθ0 ;n (y ) Jθ0 (1+ǫ) lθ0 ;n (y )
Z
J (1+ǫ) √
√
−1
n T θ0
( n(θ−θ0 )−(Jθ0 (1+ǫ))−1 lθ0 ;n (y n ))
2
·
ν(θ0 )(1 − ǫ)dθ
e−( n(θ−θ0 )−(Jθ0 (1+ǫ)) lθ0 ;n (y ))
Uθ0 ,δ
(b) − k
2
=n
Z
·
(c)
√
kzk≤ nδ
− k2
≥n
1
e 2 lθ0 ;n (y
e
n )T J −1 (1+ǫ)−1 l
n
θ0 ;n (y )
θ0
−1 l
n T
θ0 ;n (y ))
e−(z−(Jθ0 (1+ǫ))
1
l
(y n )T Jθ−1 (1+ǫ)−1 lθ0 ;n (y n )
2 θ0 ;n
0
Jθ (1+ǫ)
0
(z−(Jθ0 (1+ǫ))−1 lθ0 ;n (y n ))
2
ν(θ0 )(1 − ǫ)dz
k
ν(θ0 )(1 − ǫ)
2
(2π) 2
k ,
1
(det(Jθ0 )) 2 (1 + ǫ) 2
√
where (a), (b), and (c) follow from (84) and (86), the relation z = n(θ − θ0 ), and (91).
(92)
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
29
Now, we introduce a notation. RFor a distribution P , a subset S of the probability space, and a random
variable X, we denote the value S X(ω)P (dω) by EP |S [X]. Therefore, using this notation, we have
k
n
n
n−s 2 esD1+s (Pθ0 kQν )
h pn (Y n ) s i
k
θ0
=n−s 2 EPθn
0
qνn (Y n )
h pn (Y n ) s i
h pn (Y n ) s i
k
θ0
θ0
−s k2
n
c
=n−s 2 EPθn |Cn
+
n
E
Pθ |Cn
0
0
qνn (Y n )
qνn (Y n )
sk i
s
h s
(a)
2
2
− 2 lθ0 ;n (y n )T Jθ−1 (1+ǫ)−1 lθ0 ;n (y n )
−s
−2s (det(Jθ0 )) (1 + ǫ)
0
n
≤EPθ |Cn e
ν(θ0 ) (1 − ǫ)
sk
0
(2π) 2
h kJθ0 k(1+ǫ) 2Γ( k + 1) s i
2
2
+ EPθn |Cnc es
k
0
(1 − ǫ)π 2
sk
s
(b)
(det(Jθ0 )) 2 (1 + ǫ) 2
−(1+ǫ)−1 s2 lθ0 ;n (y n )T Jθ−1 lθ0 ;n (y n )
−s
−2s
0
≤EPθn e
ν(θ0 ) (1 − ǫ)
sk
0
(2π) 2
kJθ k(1+ǫ) 2Γ( k + 1) s
0
2
2
+ es
ǫ,
k
(1 − ǫ)π 2
(93)
where (a) and (b) follow from (88), (92) and (90), respectively.
Due to the central limit theorem, the random variable lθ0 ;n (Y n )T Jθ−1
l (Y n ) asymptotically obeys the
0 θ;n
−1
−(1+ǫ) 2s lθ0 ;n (Y n )T Jθ−1 lθ0 ;n (Y n )
0
] converges
χ2 distribution with the degree k. Hence, the expectation EPθn [e
0
1
to (1+s(1+ǫ)−1 )k/2 . Thus,
k
n
n
lim n−s 2 esD1+s (Pθ0 kQν )
n→∞
s
sk
(det(Jθ0 )) 2 (1 + ǫ) 2
1
ν(θ0 )−s (1 − ǫ)−2s
≤
sk
−1
k/2
(1 + s(1 + ǫ) )
(2π) 2
kJθ k(1+ǫ) 2Γ( k + 1) s
0
2
2
ǫ.
+ es
k
(1 − ǫ)π 2
(94)
Since ǫ > 0 is arbitrary, we have
k
n
n
lim n−s 2 esD1+s (Pθ0 kQν )
n→∞
s
1
(det(Jθ0 )) 2
≤
ν(θ0 )−s
.
sk
k/2
(1 + s)
(2π) 2
(95)
Taking the logarithm, we have
k
s
sk
k
log(2π),
lim sD1+s (Pθn0 kQnν ) − s log n ≤ − log(1 + s) − s log ν(θ0 ) + log det(Jθ0 ) −
n→∞
2
2
2
2
which implies (33).
(96)
A PPENDIX C
P ROOF OF L EMMA 2
To prove Lemma 2, we recall packing lemma [4, Lemma 10.1]. We assume that X and Y are finite
sets. For a type P ∈ Tn (X ), let V(P, Y) be the set of conditional types of P from X to Y. Then, we
have the following proposition.
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
30
Proposition 1 ([4, Lemma 10.1]): For every R > δ > 0, there exists a sufficiently large integer N
satisfying the following. For every n ≥ N and every type P ∈ Tn (X ) satisfying H(P ) > R, there exist at
least exp[n(R−δ)] distinct sequences xni ∈ Tp such that every pair of stochastic matrices V , V̂ ∈ V(P, Y)
and every i satisfy the inequality
[
n
n
(97)
TV (xi ) ∩
TV̂ (xj ) ≤ |TV (xni ) exp[−n|I(P, V̂ ) − R|+ ].
j6=i
δn
Proof of Lemma 2: Indeed, the proof of Proposition 1 in [4] employs only the property log
goes to
n
3
∞ as n → ∞ even when δ depends on n. Hence, we can apply Proposition 1 to the case with δ = n− 4 .
3
Now, we apply Proposition 1 to the case when Y = X , V̂ (x|x′ ) = δx,x′ and δ = n− 4 . Since xni ∈ TP ,
we have V(P, Y) = V (xni , X ). We also have TV̂ (xnj ) = {xnj } for any j. The relation I(P, V̂ ) = H(P )
implies that −n|I(P, V̂ ) − R|+ = −(nH(P ) − R). So, (97) is the same as (29). Hence, we obtain Lemma
2.
R EFERENCES
[1] E. L. Lehmann and George Casella. Theory of point estimation. Springer Texts in Statistics. Springer-Verlag, New York, second edition,
1998.
[2] M. Akahira, “Loss of information of a statistic for a family of non-regular distributions,” Annals of the Institute of Statistical Mathematics,
48(2), 349 – 364 (1996).
[3] B. S. Clarke and A. R. Barron, “Information-theoretic asymptotics of Bayes methods,” IEEE Trans. Inform. Theory, 36, 453–471
(1990).
[4] I. Csiszár and J. Körner, Information Theory: Coding Theorems for Discrete Memoryless Systems, Second edition (Cambridge University
Press, 2011)
[5] M. Hayashi, “Universal coding for classical-quantum channel,” Com. Math. Phys., Vol.289, No.3, 1087-1098 (2009).
[6] M. Hayashi, “Universal approximation of multi-copy states and universal quantum lossless data compression,” Comm. Math. Phys.,
Vol.293, No. 1, 171-183, (2010).
[7] R. G. Gallager, Information Theory and Reliable Communication, John Wiley & Sons, 1968.
[8] H. Yagi and R. Nomura, “Single-letter characterization of epsilon-capacity for mixed memoryless channels,” in Proc. 2014 IEEE ISIT,
Honolulu, Hawaii, USA, June 29 -July 4, pp. 2874 - 2878.
[9] H. Yagi and R. Nomura, “Channel dispersion for well-ordered mixed channels decomposed into memoryless channels,” in Proc. 2014
IEEE ISITA, Melbourne, Victoria, Australia, 26-29 October, pp. 35 - 39.
[10] V. Strassen, “Asymptotische Abschatzungen in Shannons Informationstheorie,” In Trans. Third Prague Conf. Inf. Theory, pages 689-723,
Prague, 1962.
[11] M. Hayashi, “Information Spectrum Approach to Second-Order Coding Rate in Channel Coding,” IEEE Trans. Inform. Theory, vol.
55, no.11, 4947 – 4966, 2009.
[12] Y. Polyanskiy, H.V. Poor, and S. Verdú, “Channel coding rate in the finite blocklength regime,” IEEE Trans. Inform. Theory, vol. 56,
no. 5,2307 – 2359, 2010.
[13] E. Telatar, “Capacity of multi-antenna Gaussian channels,” European Trans. Telecomm., vol. 10, no. 6, pp. 585-595, Dec. 1999.
[14] M. Hayashi, “Second-Order Asymptotics in Fixed-Length Source Coding and Intrinsic Randomness,” IEEE Trans. Inform. Theory,
vol. 54, 4619 - 4637, 2008.
[15] G.J. Foschini and M.J. Gans, “On limits of wireless communications in a fading environment when using multiple antennas,” Wireless
Personal Communications, vol. 6, pp. 311-335, 1998.
[16] T.L. Marzetta and B.M. Hochwald, “Capacity of a mobile multiple-antenna communication link in Rayleigh flat fading,” IEEE Trans.
on Inf. Theory, vol. 59, no. 12, pp. 139-157, Jan. 1999.
[17] A. Lapidoth and P. Narayan, “Reliable communication under channel uncertainty,” IEEE Trans. on Inf. Theory, vol. 44, no. 6, pp.
2148-2177, Oct. 1998.
[18] Y. Sun, C.E. Koksal, and N.B. Shroff, “Capacity of compound MIMO Gaussian channels with additive uncertainty,” IEEE Trans. on
Inf. Theory, vol. 59, no. 12, pp. 8267-8274, Dec. 2013.
[19] N. Merhav, “Universal decoding for memoryless Gaussian channels with a deterministic interference,” IEEE Trans. on Inf. Theory, vol.
39, no. 4, pp. 1261-1269, July. 1993.
[20] M. Feder and A. Lapidoth, “Universal decoding for channels with memory,” IEEE Trans. on Inf. Theory, vol. 44, no. 5, pp. 1726-1745,
Sep. 1998.
[21] Y. Polyanskiy, “On dispersion of compound DMCs,” Proc. 51th Annual Allerton Conf., Allerton, IL, USA, 2013.
[22] W. Yang, G. Durisi, T. Koch, and Y. Polyanskiy, “Quasi-static SIMO fading channels at finite blocklength,” Proc. IEEE Int. Symp. on
Inf. Theory, Istanbul, Turkey, July 2013.
[23] M. Tomamichel, and V. Y. F. Tan, “Second-order coding rates for channels with state,” preprint available at arXiv:1305.6789v2, May
2014.
M. HAYASHI: UNIVERSAL CHANNEL CODING FOR GENERAL OUTPUT ALPHABET
31
[24] R. L. Dobrushin, “Optimum information transmission through a channel with unknown parameters,” Radio Eng. Electron., vol. 4, no.
12, pp. 1-8, 1959.
[25] B. R. Bhat, “On exponential and curved exponential families in stochastic processes,” Math. Scientist, Vol. 13, 121-134 (1988).
[26] B. R. Bhat, Stochastic models : analysis and applications, New Delhi: New Age International (2000).
[27] I. L. Hudson, “Large Sample Inference for Markovian Exponential Families with Application to Branching Processes with Immigration,”
Austr. J. Statist. Vol. 24, 98-112 (1982).
[28] U. Kuchler and M. Sorensen, “Exponential Families of Stochastic Processes: A Unifying Semimartingale Approach,” International
Statistical Review, Vol. 57, No. 2, 123-144 (1989).
[29] P. D. Feigin, “Conditional Exponential Families and a Representation Theorem for Asympotic Inference,” The Annals of Statistics, Vol.
9, No. 3, 597-603 (1981).
[30] A. W. van der Vaart, Asymptotic Statistics, Cambridge Series in Statistical and Probabilistic Mathematics), Cambridge University Press
(June 19, 2000).
[31] M. Hayashi, “Second Order Analysis Based on Information Spectrum” IEEE Information Theory Society Newsletter, Vol. 62, No. 1,
March, (2012).
[32] S. Verdú, T. S. Han, “A general formula for channel capacity,” IEEE Trans. Inf. Theory, 40, 1147–1157 1994.
[33] M. Tomamichel, and M. Hayashi, “Operational Interpretation of Renyi Information Measures via Composite Hypothesis Testing Against
Product and Markov Distributions,” arXiv:1511.04874 (2015); IEEE International Symposium on Information Theory (ISIT2016),
Barcelona, 10 -15 July 2016. pp. 585-589.
[34] M. Hayashi, “Quantum wiretap channel with non-uniform random number and its exponent and equivocation rate of leaked information,”
IEEE Trans. on Inf. Theory, vol. 61, no. 10, pp. 5595-5622 (2015).
[35] R. Sibson. “Information radius,” Zeitschrift fur Wahrscheinlichkeitstheorie und Verwandte Gebiete 14(2), pp 149–160 (1969).
[36] S. Amari, H. Nagaoka, Methods of Information Geometry, (AMS & Oxford University Press, Oxford, 2000).
[37] D. Blackwell, L. Breiman, and A. Thomasian, “The capacity of a class of channels,” Ann. Math. Stat., pp. 1229-1241, 1959.
[38] J. Wolfowitz, Coding Theorems of Information Theory. Englewood Cliffs, NJ: Prentice-Hall, 1962.
[39] E. L. Lehmann and J. P. Romano, Testing Statistical Hypotheses, Springer Texts in Statistics (2005).
[40] P. Moulin, “The Log-Volume of Optimal Codes for Memoryless Channels, Asymptotically Within A Few Nats,” arXiv 1311.0181
(2013).
[41] S. Niranjayan and N. C. Beaulieu, “Analysis of Wireless Communication Systems in the Presence of Non-Gaussian Impulsive Noise
and Gaussian Noise,” 2010 IEEE Wireless Communication and Networking Conference, 2010.
[42] K. Nakagawa and F. Kanaya, “On the converse theorem in statistical hypothesis testing for markov chains,” IEEE Trans. Inform. Theory,
Vol. 39, No. 2, 629-633 (1993).
[43] H. Nagaoka, “The exponential family of Markov chains and its information geometry” Proceedings of The 28th Symposium on
Information Theory and Its Applications (SITA2005), Okinawa, Japan, Nov. 20-23, (2005).
[44] M. Hayashi and S. Watanabe, “Information Geometry Approach to Parameter Estimation in Markov Chains,” Annals of Statistics.
Volume 44, Number 4, 1495-1535 (2016)
[45] S. Watanabe and M. Hayashi, “Finite-length Analysis on Tail probability for Markov Chain and Application to Simple Hypothesis
Testing,” Accepted for publication in Annals of Applied Probability.
[46] D. Haussler and M. Opper, “Mutual information, metric entropy and cumulative relative entropy risk,” Annals of Statistics vol. 25, no.
6, pp. 2451 – 2492, 1997
[47] M. Hayashi, “Universal Channel Coding with Continuous Output System,” 2014 International Symposium on Information Theory and
Its Applications, Melbourne, Australia, 26 - 29, October 2014. pp. 21-25
[48] P. Moulin, “Coding Redundancy of Finite-Blocklength Universal Channel Coding,” IEEE International Symposium on Information
Theory (ISIT2015), Hong Kong, June 14 - June 19, 2015.
[49] M. Hayashi and M. Tomamichel, “Correlation Detection and an Operational Interpretation of the Rényi Mutual Information,” Journal
of Mathematical Physics, vol. 57, 102201 (2016)
| 7cs.IT
|
Published as a conference paper at ICLR 2018
I NTERACTIVE G ROUNDED L ANGUAGE ACQUISITION
AND G ENERALIZATION IN A 2D W ORLD
arXiv:1802.01433v2 [cs.CL] 21 Feb 2018
Haonan Yu, Haichao Zhang & Wei Xu
Baidu Research
Sunnyvale, CA 94089, USA
{haonanyu,zhanghaichao,wei.xu}@baidu.com
A BSTRACT
We build a virtual agent for learning language in a 2D maze-like world. The
agent sees images of the surrounding environment, listens to a virtual teacher, and
takes actions to receive rewards. It interactively learns the teacher’s language from
scratch based on two language use cases: sentence-directed navigation and question answering. It learns simultaneously the visual representations of the world,
the language, and the action control. By disentangling language grounding from
other computational routines and sharing a concept detection function between
language grounding and prediction, the agent reliably interpolates and extrapolates to interpret sentences that contain new word combinations or new words
missing from training sentences. The new words are transferred from the answers
of language prediction. Such a language ability is trained and evaluated on a population of over 1.6 million distinct sentences consisting of 119 object words, 8 color
words, 9 spatial-relation words, and 50 grammatical words. The proposed model
significantly outperforms five comparison methods for interpreting zero-shot sentences. In addition, we demonstrate human-interpretable intermediate outputs of
the model in the appendix.
1
I NTRODUCTION
Some empiricists argue that language may be learned based on its usage (Tomasello, 2003). Skinner
(1957) suggests that the successful use of a word reinforces the understanding of its meaning as well
as the probability of it being used again in the future. Bruner (1985) emphasizes the role of social
interaction in helping a child develop the language, and posits the importance of the feedback and
reinforcement from the parents during the learning process. This paper takes a positive view of the
above behaviorism and tries to explore some of the ideas by instantiating them in a 2D virtual world
where interactive language acquisition happens. This interactive setting contrasts with a common
learning setting in that language is learned from dynamic interactions with environments instead of
from static labeled data.
Language acquisition can go beyond mapping language as input patterns to output labels for merely
obtaining high rewards or accomplishing tasks. We take a step further to require the language to be
grounded (Harnad, 1990). Specifically, we consult the paradigm of procedural semantics (Woods,
2007) which posits that words, as abstract procedures, should be able to pick out referents. We
will attempt to explicitly link words to environment concepts instead of treating the whole model
as a black box. Such a capability also implies that, depending on the interactions with the world,
words would have particular meanings in a particular context and some content words in the usual
sense might not even have meanings in our case. As a result, the goal of this paper is to acquire
“in-context” word meanings regardless of their suitability in all scenarios.
On the other hand, it has been argued that a child’s exposure to adult language provides inadequate
evidence for language learning (Chomsky, 1991), but some induction mechanism should exist to
bridge this gap (Landauer & Dumais, 1997). This property is critical for any AI system to learn infinite sentences from finite training data. This type of generalization problem is specially addressed in
our problem setting. After training, we want the agent to generalize to interpret zero-shot sentences
of two types:
1
Published as a conference paper at ICLR 2018
Training
Testing
NAV “Move to north of avocado.” “Go to east of rabbit.”
QA
(Answer: “Watermelon”)
(a)
“Go to east of avocado.”
“Can you reach watermelon?”
“What is in northwest?” “What is east of avocado?” “What is the color of watermelon?”
(Answer: “Nothing”)
(Answer: “Red”)
(b)
(c)
(d)
Figure 1: An illustration of XWORLD and the two language use cases. (a) and (b): A mixed training
of NAV and QA. (c): Testing ZS1 sentences contain a new combination of words (“east” and “avocado”) that never appear together in any training sentence. (d): Testing ZS2 sentences contain a
new word (“watermelon”) that never appears in any training sentence but is learned from a training
answer. This figure is only a conceptual illustration of language generalization; in practice it might
take many training sessions before the agent can generalize. (Due to space limitations, the maps are
only partially shown.)
1) interpolation, new combinations of previously seen words for the same use case, or
2) extrapolation, new words transferred from other use cases and models.
In the following, we will call the first type ZS1 sentences and the second type ZS2 sentences. Note
that so far the zero-shot problems, addressed by most recent work (Hermann et al., 2017; Chaplot
et al., 2018) of interactive language learning, belong to the category of ZS1. In contrast, a reliable interpretation of ZS2 sentences, which is essentially a transfer learning (Pan & Yang, 2010) problem,
will be a major contribution of this work.
We created a 2D maze-like world called XWORLD (Figure 1), as a testbed for interactive grounded
language acquisition and generalization.1 In this world, a virtual agent has two language use cases:
navigation (NAV) and question answering (QA). For NAV, the agent needs to navigate to correct
places indicated by language commands from a virtual teacher. For QA, the agent must correctly
generate single-word answers to the teacher’s questions. NAV tests language comprehension while
QA additionally tests language prediction. They happen simultaneously: When the agent is navigating, the teacher might ask questions regarding its current interaction with the environment. Once
the agent reaches the target or the time is up, the current session ends and a new one is randomly
generated according to our configuration (Appendix B). The ZS2 sentences defined in our setting
require word meanings to be transferred from single-word answers to sentences, or more precisely,
from language prediction to grounding. This is achieved by establishing an explicit link between
grounding and prediction via a common concept detection function, which constitutes the major
novelty of our model. With this transferring ability, the agent is able to comprehend a question containing a new object learned from an answer, without retraining the QA pipeline. It is also able to
navigate to a freshly taught object without retraining the NAV pipeline.
It is worthwhile emphasizing that this seemingly “simple” world in fact poses great challenges for
language acquisition and generalization, because:
˝ The state space is huge. Even for a 7 ˆ 7 map with 15 wall blocks and 5 objects selected from
119 distinct classes, there are already octillions (1027 ) of possible different configurations, not
to mention the intra-class variance of object instances (see Figure 16 in the appendix). For two
configurations that only differ in one block, their successful navigation paths could be completely
different. This requires an accurate perception of the environment. Moreover, the configuration
constantly changes from session to session, and from training to testing. In particular, the target
changes across sessions in both location and appearance.
1
https://github.com/PaddlePaddle/XWorld
2
Published as a conference paper at ICLR 2018
˝ The goal space implied by the language for navigation is huge. For a vocabulary containing only
185 words, the total number of distinct commands that can be said by the teacher conforming to
our defined grammar is already over half a million. Two commands that differ by only one word
could imply completely different goals. This requires an accurate grounding of language.
˝ The environment demands a strong language generalization ability from the agent. The agent has
to learn to interpret zero-shot sentences that might be as long as 13 words. It has to “plug” the
meaning of a new word or word combination into a familiar sentential context while trying to
still make sense of the unfamiliar whole. The recent work (Hermann et al., 2017; Chaplot et al.,
2018) addresses ZS1 (for short sentences with several words) but not ZS2 sentences, which is a
key difference between our learning problem and theirs.
We describe an end-to-end model for the agent to interactively acquire language from scratch and
generalize to unfamiliar sentences. Here “scratch” means that the model does not hold any assumption of the language semantics or syntax. Each sentence is simply a sequence of tokens with each
token being equally meaningless in the beginning of learning. This is unlike some early pioneering systems (e.g., SHRDLU (Winograd, 1972) and A BIGAIL (Siskind, 1994)) that hard-coded the
syntax or semantics to link language to a simulated world–an approach that presents scalability issues. There are two aspects of the interaction: one is with the teacher (i.e., language and rewards)
and the other is with the environment (e.g., stepping on objects or hitting walls). The model takes
as input RGB images, sentences, and rewards. It learns simultaneously the visual representations
of the world, the language, and the action control. We evaluate our model on randomly generated
XWORLD maps with random agent positions, on a population of over 1.6 million distinct sentences
consisting of 119 object words, 8 color words, 9 spatial-relation words, and 50 grammatical words.
Detailed analysis (Appendix A) of the trained model shows that the language is grounded in such
a way that the words are capable to pick out referents in the environment. We specially test the
generalization ability of the agent for handling zero-shot sentences. The average NAV success rates
are 84.3% for ZS1 and 85.2% for ZS2 when the zero-shot portion is half, comparable to the rate
of 90.5% in a normal language setting. The average QA accuracies are 97.8% for ZS1 and 97.7%
for ZS2 when the zero-shot portion is half, almost as good as the accuracy of 99.7% in a normal
language setting.
2
M ODEL
Our model incorporates two objectives. The first is to maximize the cumulative reward of NAV
and the second is to minimize the classification cost of QA. For the former, we follow the standard
reinforcement learning (RL) paradigm: the agent learns the action at every step from reward signals.
It employs the actor-critic (AC) algorithm (Sutton & Barto, 1998) to learn the control policy (Appendix E). For the latter, we adopt the standard supervised setting of Visual QA (Antol et al., 2015):
the groundtruth answers are provided by the teacher during training. The training cost is formulated
as the multiclass cross entropy.
2.1
M OTIVATION
The model takes two streams of inputs: images and sentences. The key is how to model the language
grounding problem. That is, the agent must link (either implicitly or explicitly) language concepts
to environment entities to correctly take an action by understanding the instruction in the current
visual context. A straightforward idea would be to encode the sentence s with an RNN and encode
the perceived image e with a CNN, after which the two encoded representations are mixed together.
Specifically, let the multimodal module be M, the action module be A, and the prediction module
be P, this idea can be formulated as:
`
˘
NAV: A` MpRNNpsq, CNNpeqq˘
(1)
QA: P MpRNNpsq, CNNpeqq .
Hermann et al. (2017); Misra et al. (2017); Chaplot et al. (2018) all employ the above paradigm. In
their implementations, M is either vector concatenation or element-wise product. For any particular
word in the sentence, fusion with the image could happen anywhere starting from M all the way
to the end, right before a label is output. This is due to the fact that the RNN folds the string of
words into a compact embedding which then goes through the subsequent blackbox computations.
3
P
Prediction
𝐌0
𝐌/
ℎ
A
Action
F
Vision
ℎ
𝑒
𝜙
𝑥*+,Multimodal
𝑠
NAV: “Move to the red apple.”
QA: “Say the color of avocado.”
L
Grounding
Published as a conference paper at ICLR 2018
𝑐
𝑎
𝑣
Figure 2: An overview of the model. We process e by always placing the agent at the center via
zero padding. This helps the agent learn navigation actions by reducing the variety of target representations. c, a, and v are the predicted answer, the navigation action, and the critic value for policy
gradient, respectively. φ denotes the concept detection function shared by language grounding and
prediction. MA generates a compact representation from xloc and h for navigation (Appendix C).
Therefore, language grounding and other computational routines are entangled. Because of this, we
say that this paradigm has an implicit language grounding strategy. Such a strategy poses a great
challenge for processing a ZS2 sentence because it is almost impossible to predict how a new word
learned from language prediction would perform in the complex entanglement involved. Thus a
careful inspection of the grounding process is needed.
2.2
A PPROACH
The main idea behind our approach is to disentangle language grounding from other computations
in the model. This disentanglement makes it possible for us to explicitly define language grounding
around a core function that is also used by language prediction. Specifically, both grounding and
prediction are cast as concept detection problems, where each word (embedding) is treated as a
detector. This opens up the possibility of transferring word meanings from the latter to the former.
The overall architecture of our model is shown in Figure 2.
2.2.1
E XPLICIT GROUNDING
We begin with our definition of “grounding.” We define a sentence as generally a string of words
of any length. A single word is a special case of a sentence. Given a sentence s and an image
representation h “ CNNpeq, we say that s is grounded in h as x if
I) h consists of M entities where an entity is a subset of visual features, and
II) x P t0, 1uM with each entry xrms representing a binary selection of the mth entity of h. Thus
x is a combinatorial selection over h.
Furthermore, x is explicit if
III) it is formed by the grounding results of (some) individual words of s (i.e., compositionality).
We say that a framework has an explicit grounding strategy if
IV) all language-vision fusions in the framework are explicit groundings.
For our problem, we propose a new framework with an explicit grounding strategy:
`
˘
NAV: A` MA px, CNNpeqq˘
QA: P MP px, CNNpeqq ,
(2)
where the sole language-vision fusion x in the framework is an explicit grounding. Notice in the
above how the grounding process, as a “bottleneck,” allows only x but not other linguistic information to flow to the downstream of the network. That is, MA , MP , A, and P all rely on grounded
4
Published as a conference paper at ICLR 2018
results but not on other sentence representations. By doing so, we expect x to summarize all the
necessary linguistic information for performing the tasks.
The benefits of this framework are two-fold. First, the explicit grounding strategy provides a conceptual abstraction (Garnelo et al., 2016) that maps high-dimensional linguistic input to a lowerdimensional conceptual state space and abstracts away irrelevant input signals. This improves the
generalization for similar linguistic inputs. Given e, all that matters for NAV and QA is x. This guarantees that the agent will perform exactly in the same way on the same image e even given different
sentences as long as their grounding results x are the same. It disentangles language grounding
from subsequent computations such as obstacle detection, path planning, action making, and feature
classification, which all should be inherently language-independent routines. Second, because x is
explicit, the roles played by the individual words of s in the grounding are interpretable. This is
in contrast to Eq. 1 where the roles of individual words are unclear. The interpretability provides a
possibility of establishing a link between language grounding and prediction, which we will perform
in the remainder of this section.
2.2.2
I NSTANTIATION OF EXPLICIT GROUNDING
Let h P RN ˆD be a spatially flattened feature cube (originally in 3D, now the 2D spatial domain
collapsed into 1D for notational simplicity), where D is the number of channels and N is the number
of locations in the spatial domain. We adopt three definitions for an entity:
1) a feature vector at a particular image location,
2) a particular feature map along the channel dimension, and
3) a scalar feature at the intersection of a feature vector and a feature map.
Their grounding results are denoted as xloc ps, hq P t0, 1uN , xfeat ps, hq P t0, 1uD , and xcube ps, hq P
t0, 1uN ˆD , respectively. In the rest of the paper, we remove s and h from xloc , xfeat , and xcube
for notational simplicity while always assuming a dependency on them. We assume that xcube is a
low-rank matrix that can be decomposed into the two:
xcube “ xloc ¨ xfeat | .
To make the model fully differentiable, in the following we relax the definition of grounding so that
xloc P r0, 1sN , xfeat P r0, 1sD , and xcube P r0, 1sN ˆD . The attention map xloc is responsible for
image spatial attention. The channel mask xfeat is responsible for selecting image feature maps, and
is assumed to be independent of the specific h, namely, xfeat ps, hq “ xfeat psq. Intuitively, h can be
modulated by xfeat before being sent to downstream processings. A recent paper by de Vries et al.
(2017) proposes an even earlier modulation of the visual processing by directly conditioning some
of the parameters of a CNN on the linguistic input.
Finally, we emphasize that our explicit grounding, even though instantiated as a soft attention mechanism, is different from the existing visual attention models. Some attention models such as Xu et al.
(2015); de Vries et al. (2017) violate definitions III and IV. Some work (Andreas et al., 2016a;b; Lu
et al., 2016) violates definition IV in a way that language is fused with vision by a multilayer perceptron (MLP) after image attention. Anderson et al. (2017) proposes a pipeline similar to ours
but violates definition III in which the image spatial attention is computed from a compact question
embedding output by an RNN.
2.2.3
C ONCEPT DETECTION
With language grounding disentangled, now we relate it to language prediction. This relation is a
common concept detection function. We assume that every word in a vocabulary, as a concept, is
detectable against entities of type (1) as defined in Section 2.2.1. For a meaningful detection of
spatial-relation words that are irrelevant to image content, we incorporate parametric feature maps
into h to learn spatial features. Assume a precomputed xfeat , the concept detection operates by
sliding over the spatial domain of the feature cube h, which can be written as a function φ:
φ : h, xfeat , u ÞÑ χ,
N
where χ P R is a detection score map and u is a word embedding vector. This function scores the
embedding u against each feature vector of h, modulated by xfeat that selects which feature maps to
5
Published as a conference paper at ICLR 2018
𝑥%&'(
𝑥"#$
“What is the color of the object in the northeast?”
Figure 3: An illustration of the attention cube xcube “ xloc ¨ xfeat | , where xloc attends to image
regions and xfeat selects feature maps. In this example, xloc is computed from “northeast.” In order
for the agent to correctly answer “red” (color) instead of “watermelon” (object name), xfeat has to
be computed from the sentence pattern “What ... color ...?”
use for the scoring. Intuitively, each score on χ indicates the detection response of the feature vector
in that location. A higher score represents a higher detection response.
While there are many potential forms for φ, we implement it as
φph, xfeat , uq “ h ¨ pxfeat ˝ uq,
(3)
D
where ˝ is the element-wise product. To do so, we have word embedding u P R where D is equal
to the number of channels of h.
2.2.4
P REDICTION BY CONCEPT DETECTION
For prediction, we want to output a word given a question s and an image e. Suppose that xloc
and xfeat are the grounding results of s. Based on the detection function φ, MP outputs a score
vector m P RK over the entire lexicon, where each entry of the vector is:
mrks “ x|loc φph, xfeat , uk q “ x|loc χk ,
(4)
where uk is the kth entry of the word embedding table. The above suggests that mrks is the result
of weighting the scores on the map χk by xloc . It represents the correctness of the kth lexical entry
as the answer to the question s. To predict an answer
`
˘
Ppmq “ argmax softmaxpmq .
k
Note that the role of xfeat in the prediction is to select which feature maps are relevant to the question s. Otherwise it would be confusing for the agent about what to predict (e.g., whether to predict
a color or an object name). By using xfeat , we expect that different feature maps encode different
image attributes (see an example in the caption of Figure 3). More analysis of xfeat is performed in
Appendix A.
2.2.5
G ROUNDING BY CONCEPT DETECTION
To compute xcube , we compute xloc and xfeat separately.
We want xloc to be built on the detection function φ. One can expect to compute a series of score
maps χ of individual words and merge them into xloc . Suppose that s consists of L words twl u
with wl “ uk being some word k in the dictionary. Let τ psq be a sequence of indices tli u where
0 ď li ă L. This sequence function τ decides which words of the sentence are selected and
organized in what order. We define xloc as
`
˘
xloc “ Υ φph, 1, wl1 q, . . . , φph, 1, wli q, . . . , φph, 1, wlI q
(5)
“ Υpχl1 , . . . , χli , . . . , χlI q,
6
Published as a conference paper at ICLR 2018
Flip
Cross correlation
“northwest”
“northwest of apple”
“apple”
Figure 4: A symbolic example of the 2D convolution for transforming attention maps. A 2D convolution can be decomposed into two steps: flipping and cross correlation. The attention map of
“northwest” is treated as an offset filter to translate that of “apple.” Note that in practice, the attention is continuous and noisy, and the interpreter has to learn to find out the words (if any) to perform
this convolution.
where 1 P t0, 1uD is a vector of ones, meaning that it selects all the feature maps for detecting wli .
Υ is an aggregation function that combines the sequence of score maps χli of individual words. As
such, φ makes it possible to transfer new words from Eq. 4 to Eq. 5 during test time.
If we were provided with an oracle that is able to output a parsing tree for any sentence, we could set
τ and Υ according to the tree semantics. Neural module networks (NMNs) (Andreas et al., 2016a;b;
Hu et al., 2017) rely on such a tree for language grounding. They generate a network of modules
where each module corresponds to a tree node. However, labeled trees are needed for training.
Below we propose to learn τ and Υ based on word attention (Bahdanau et al., 2015) to bypass the
need for labeled structured data.
We start by feeding a sentence s “ twl u of length L to a bidirectional RNN (Schuster & Paliwal,
1997). It outputs a compact sentence embedding semb and a sequence of L word context vectors wl .
Each wl summarizes the sentential pattern around that word. We then employ a meta controller
called interpreter in an iterative manner. For the ith interpretation step, the interpreter computes the
word attention as:
“
‰
$
Word attention: oil 9 ÿ
exp Scos ppi´1 , wl q
’
’
’
i
’
oil wl
’
& Attended context: w “
l
ÿ
(6)
τ˚
i
’
Attended
word:
s
“
oil wl
’
’
’
’
l
%
Interpreter state: pi “ GRUppi´1 , wi q
where Scos is cosine similarity and GRU is the gated recurrent unit (Cho et al., 2014). Here we use
τ ˚ to represent an approximation of τ via soft word attention. We set p0 to the compact sentence
embedding semb . After this, the attended word si is fed to the detection function φ. The interpreter
aggregates the score map of si by:
$
`
˘
’
Detection: y 1 “ softmax φph, 1, si q
’
&
Map transform: xiloc “ y 1 ˚ y i´1
Υ
(7)
Map update gate: ρi “ σpW pi ` bq
’
’
%
Map update: y i “ ρi xiloc ` p1 ´ ρi qy i´1
where ˚ denotes a 2D convolution, σ is sigmoid, and ρi is a scalar. W and b are parameters to be
learned. Finally, the interpreter outputs xIloc as xloc , where I is the predefined maximum step.
Note that in the above we formulate the map transform as a 2D convolution. This operation enables
the agent to reason about spatial relations. Recall that each attention map xloc is egocentric. When
the agent needs to attend to a region specified by a spatial relation referring to an object, it can
translate the object attention with the attention map of the spatial-relation word which serves as a
2D convolutional offset filter (Figure 4). For this reason, we set y 0 as a one-hot map where the map
7
Published as a conference paper at ICLR 2018
center is one, to represent the identity translation. A similar mechanism of spatial reasoning via
convolution was explored by Kitaev & Klein (2017) for a voxel-grid 3D representation.
By assumption, the channel mask xfeat is meant to be determined solely from s; namely, which
features to use should only depend on the sentence itself, not on the value of the feature cube h.
Thus it is computed as
`
˘
xfeat “ MLP RNNpsq ,
(8)
where the RNN returns an average state of processing s, followed by an MLP with the sigmoid
activation.2
3
R ELATED WORK
Our XWORLD is similar to the 2D block world MaseBase (Sukhbaatar et al., 2016). However, we
emphasize the problem of grounded language acquisition and generalization, while they do not.
There have been several 3D simulated worlds such as ViZDoom (Kempka et al., 2016), DeepMind
Lab (Beattie et al., 2016), and Malmo (Johnson et al., 2016). Still, these other settings intended for
visual perception and control, with less or no language.
Our problem setting draws inspirations from the AI roadmap delineated by Mikolov et al. (2015).
Like theirs, we have a teacher in the environment that assigns tasks and rewards to the agent, potentially with a curriculum. Unlike their proposal of using only the linguistic channel, we have multiple
perceptual modalities, the fusion of which is believed to be the basis of meaning (Kiela et al., 2016).
Contemporary to our work, several end-to-end systems also address language grounding problems
in a simulated world with deep RL. Misra et al. (2017) maps instructions and visual observations to
actions of manipulating blocks on a 2D plane. Hermann et al. (2017); Chaplot et al. (2018) learn to
navigate in 3D under instructions, and both evaluate ZS1 generalization. Despite falling short of the
vision challenges found in these other worlds, we have a much larger space of zero-shot sentences
and additionally require ZS2 generalization, which is in fact a transfer learning (Pan & Yang, 2010)
problem.
Other recent work (Andreas et al., 2017; Oh et al., 2017) on zero-shot multitask learning treats
language tokens as (parsed) labels for identifying skills. In these papers, the zero-shot settings are
not intended for language understanding.
The problem of grounding language in perception can perhaps be traced back to the early work
of Siskind (1994; 1999), although no statistical learning was adopted at that time. Our language
learning problem is related to some recent work on learning to ground language in images and
videos (Yu & Siskind, 2013; Gao et al., 2016; Rohrbach et al., 2016). The navigation task is relevant
to robotics navigation under language commands (Chen & Mooney, 2011; Tellex et al., 2011; Barrett
et al., 2017). The question-answering task is relevant to image question answering (VQA) (Antol
et al., 2015; Gao et al., 2015; Ren et al., 2015; Lu et al., 2016; Yang et al., 2016; Anderson et al.,
2017; de Vries et al., 2017). The interactive setting of learning to accomplish tasks is similar to that
of learning to play video games from pixels (Mnih et al., 2015).
4
E XPERIMENTS
We design a variety of experiments to evaluate the agent’s language acquisition and generalization
ability. Our model is first compared with several methods to demonstrate the challenges in XWORLD.
We then evaluate the agent’s language generalization ability in two different zero-shot conditions.
Finally, we conclude with preliminary thoughts on how to scale our model to a 3D world.
2
Note that here we drop the explicitness requirement (definition III) for Eq. 8. This choice of implementation simplicity is purely based on our current problem that requires little language compositionality when
computing xfeat (unlike xloc ). One could imagine an alternative grounding that is explicit where a single-step
word attention extracts words from s to compute xfeat .
8
Published as a conference paper at ICLR 2018
XWORLD sentences
NAV command
(~570k)
“Please arrive at the location
between the cat and the dog.”
QA question
(~1m)
“What is the object that is west
of the orange?”
QA answer
(136)
“Dragon”
Figure 5: The three types of language data and their statistics.
4.1
G ENERAL SETUP
For all the experiments, both the sentences and the environments change from session to session, and
from training to testing. The sentences are drawn conforming to the teacher’s grammar. There are
three types of language data: NAV command, QA question, and QA answer, which are illustrated
in Figure 5. In total, there are „570k NAV commands, „1m QA questions, and 136 QA answers
(all the content words plus “nothing” and minus “between”). The environment configurations are
randomly generated from octillions of possibilities of a 7 ˆ 7 map, conforming to some high-level
specifications such as the numbers of objects and wall blocks. For NAV, our model is evaluated on
four types of navigation commands:
nav
nav
nav
nav
obj:
col obj:
nr obj:
bw obj:
Navigate to an object.
Navigate to an object with a specific color.
Navigate to a location near an object.
Navigate to a location between two objects.
For QA, our model is evaluated on twelve types of questions (rec * in Table 2). We refer the reader
to Appendix B for a detailed description of the experiment settings.
4.2
C OMPARISON METHODS
Four comparison methods and one ablation method are described below:
ContextualAttention [CA] A variant of our model that replaces the interpreter with a contextual
attention model. This attention model uses a gated RNN to convert a sentence to a filter which is
then convolved with the feature cube h to obtain the attention map xloc . The filter covers the 3 ˆ 3
neighborhood of each feature vector in the spatial domain. The rest of the model is unchanged.
StackedAttentionNet [SAN] An adaptation of a model devised by Yang et al. (2016) which was originally proposed for VQA. We replace our interpreter with their stacked attention model to compute
the attention map xloc . Instead of employing a pretrained CNN as they did, we train a CNN from
scratch to accommodate to XWORLD. The CNN is configured as the one employed by our model.
The rest of our model is unchanged.
VIS-LSTM [VL] An adaptation of a model devised by Ren et al. (2015) which was originally proposed for VQA. We flatten h and project it to the word embedding space RD . Then it is appended
to the input sentence s as the first word. The augmented sentence goes through an LSTM whose last
state is used for both NAV and QA (Figure 17, Appendix D).
ConcatEmbed [CE] An adaptation of a model proposed by Mao et al. (2015) which was originally
proposed for image captioning. It instantiates L as a vanilla LSTM which outputs a sentence embedding. Then h is projected and concatenated with the embedding. The concatenated vector is used
for both NAV and QA (Figure 18 Appendix D). This concatenation mechanism is also employed
by Hermann et al. (2017); Misra et al. (2017).
NavAlone [NAVA] An ablation of our model that does not have the QA pipeline (MP and P) and is
trained only on the NAV tasks. The rest of the model is the same.
9
Published as a conference paper at ICLR 2018
Ours
CA
SAN
CE
VL
NAVA
1
Reward
0
−1
−2
−3
−4
−5
0
2
4
6
6
Number of training examples (×10 )
(a)
1.0
0.8
0.6
0.4
0.2
0.0
Accuracy
Success rate
1.0
0.8
0.6
0.4
0.2
0.0
nav_obj
nav_col_obj nav_nr_obj nav_bw_obj
(b)
overall
(c)
Figure 6: The basic evaluation. (a) Training reward curves. The shown reward is the accumulated
discounted reward per session, averaged over every 8k training time steps. The shaded area of each
curve denotes the variance among 4 random initializations of the model parameters. (b) Navigation
success rates in the test. (c) The accuracies of the answers in the test (NAVA is excluded because it
does not train QA).
In the following experiments, we train all six approaches (four comparison methods, one ablation,
and our model) with a small learning rate of 1 ˆ 10´5 and a batch size of 16, for a maximum of
200k minibatches. Additional training details are described in Appendix C. After training, we test
each approach on 50k sessions. For NAV, we compute the average success rate of navigation where
a success is defined as reaching the correct location before the time out of a session. For QA, we
compute the average accuracy in answering the questions.
4.3
BASIC EVALUATION
In this experiment, the training and testing sentences (including NAV commands and QA questions)
are sampled from the same distribution over the entire sentence space. We call it the normal language
setting.3
The training reward curves and testing results are shown in Figure 6. VL and CE have quite low
rewards without convergences. These two approaches do not use the spatial attention xloc , and thus
always attend to whole images with no focus. The region of a target location is tiny compared
to the whole egocentric image (a ratio of 1 : p7 ˆ 2 ´ 1q2 “ 1 : 169). It is possible that this
difficulty drives the models towards overfitting QA without learning useful representations for NAV.
3
Although some test sentences might not be seen in training (i.e., zero-shot) due to sampling, all the content
words and their combinations (totaling a dozen thousands) are highly likely to be exhausted by training. Thus
we consider this experiment as a normal setting, compared to the zero-shot setting in Section 4.4.
10
Published as a conference paper at ICLR 2018
Both CA and SAN obtain rewards and success rates slightly worse than ours. This suggests that in
a normal language setting, existing attention models are able to handle previously seen sentences.
However, their language generalization abilities, especially on the ZS2 sentences, are usually very
weak, as we will demonstrate in the next section. The ablation NAVA has a very large variance in
its performance. Depending on the random seed, its reward can reach that of SAN (´0.8), or it can
be as low as that of CE (´3.0). Comparing it to our full method, we conclude that even though
the QA pipeline operates on a completely different set of sentences, it learns useful local sentential
knowledge that results in an effective training of the NAV pipeline.
Note that all the four comparison methods obtained high QA accuracies (ą70%, see Figure 6c),
despite their distinct NAV results. This suggests that QA, as a supervised learning task, is easier
than NAV as an RL task in our scenario. One reason is that the groundtruth label in QA is a much
stronger training signal than the reward in NAV. Another reason might be that NAV additionally
requires learning the control module, which is absent in QA.
4.4
L ANGUAGE GENERALIZATION
Our more important question is whether the agent has the ability of interpreting zero-shot sentences.
For comparison, we use CA and SAN from the previous section, as they achieved good performance
in the normal language setting. For a zero-shot setting, we set up two language conditions:
NewWordCombination [ZS1] Some word pairs are excluded from both the NAV commands and the
QA questions. We consider three types of unordered combinations of the content words: (object,
spatial relation), (object, color), and (object, object). We randomly hold out X% of the word pairs
during the training.
NewWord [ZS2] Some single words are excluded from both the NAV commands and the QA questions, but can appear in the QA answers. We randomly hold out X% of the object words during the
training.
We vary the value of X (12.5, 20.0, 50.0, 66.7, or 90.0) in both conditions to test how sensitive the
generalization is to the amount of the held-out data. For evaluation, we report the test results only
for the zero-shot sentences that contain the held-out word pairs or words. The results are shown in
Figure 7.
We draw three conclusions from the results. First, the ZS1 sentences are much easier to interpret than
the ZS2 sentences. Neural networks seem to inherently have this capability to some extent. This is
consistent with what has been observed in some previous work (Hermann et al., 2017; Chaplot et al.,
2018) that addresses the generalization on new word combinations. Second, the ZS2 sentences are
difficult for CA and SAN. Even with a held-out portion as small as X% “ 12.5%, their navigation
success rates and QA accuracies drop up to 80% and 35%, respectively, compared to those in the
normal language setting. In contrast, our model tends to maintain the same results until X “ 90.0.
Impressively, it achieves an average success rate of 60% and an average accuracy of 83% even when
the number of new object words is 9 times that of seen object words in the NAV commands and
QA questions, respectively! Third, in ZS2, if we compare the slopes of the success-rate curves with
those of the accuracy curves (as shown in Figures 7e and 7f), we notice that the agent generalizes
better on QA than on NAV. This further verifies our finding in the previous section that QA is in
general an easier task than NAV in XWORLD. This demonstrates the necessity of evaluating NAV in
addition to QA, as NAV requires additional language grounding to control.
Interestingly, we notice that nav bw obj is an outlier command type for which CA is much less
sensitive to the increase of X in ZS2. A deep analysis reveals that for nav bw obj, CA learns to
cheat by looking for the image region that contains the special pattern of object pairs in the image
without having to recognize the objects. This suggests that neural networks tend to exploit data in
an unexpected way to achieve tasks if no constraints are imposed (Kottur et al., 2017).
To sum up, our model exhibits a strong generalization ability on both ZS1 and ZS2 sentences, the
latter of which pose a great challenge for traditional language grounding models. Although we use a
particular 2D world for evaluation in this work, the promising results imply the potential for scaling
to an even larger vocabulary and grammar with a much larger language space.
11
nav_bw_obj
90.0
66.7
50.0
20.0
12.5
0.0
nav_col_obj
nav_nr_obj
nav_bw_obj
90.0
66.7
50.0
20.0
12.5
0.0
nav_obj
90.0
66.7
50.0
20.0
12.5
0.0
Held out X(%)
(a)
90.0
66.7
50.0
20.0
12.5
0.0
Success rate
nav_nr_obj
90.0
66.7
50.0
20.0
12.5
0.0
nav_col_obj
90.0
66.7
50.0
20.0
12.5
0.0
1.0
0.8
0.6
0.4
0.2
0.0
nav_obj
90.0
66.7
50.0
20.0
12.5
0.0
2
0
2
4
6
8
10
90.0
66.7
50.0
20.0
12.5
0.0
Reward
Published as a conference paper at ICLR 2018
Held out X(%)
Ours
CA
SAN
(b)
Ours
CA
SAN
Accuracy
1.0
0.8
0.6
0.4
0.2
0.0
Ours
CA
SAN
90.0
66.7
50.0
20.0
12.5
0.0
nav_bw_obj
90.0
66.7
50.0
20.0
12.5
0.0
nav_col_obj
nav_nr_obj
nav_bw_obj
90.0
66.7
50.0
20.0
12.5
0.0
nav_obj
90.0
66.7
50.0
20.0
12.5
0.0
Held out X(%)
(d)
90.0
66.7
50.0
20.0
12.5
0.0
Success rate
nav_nr_obj
90.0
66.7
50.0
20.0
12.5
0.0
nav_col_obj
90.0
66.7
50.0
20.0
12.5
0.0
1.0
0.8
0.6
0.4
0.2
0.0
(c)
nav_obj
90.0
66.7
50.0
20.0
12.5
0.0
2
0
2
4
6
8
10
90.0
66.7
50.0
20.0
12.5
0.0
Reward
Held out X(%)
Held out X(%)
Ours
CA
SAN
(e)
Ours
CA
SAN
Accuracy
1.0
0.8
0.6
0.4
0.2
0.0
Ours
CA
SAN
90.0
66.7
50.0
20.0
12.5
0.0
Held out X(%)
(f)
Figure 7: The test results of language generalization with a varying held-out portion of X%, where
X “ 0 represents the basic evaluation in Section 4.3. (a–c) ZS1. (d–f) ZS2. For either ZS1 or
ZS2, from top to bottom, the three rows represent the average navigation reward per session, the
average navigation success rate per session, and the average QA accuracy, respectively. (The plots
of nav obj in (a) and (b) are empty because there is no ZS1 sentence of this type by definition.)
4.5
H OW DOES IT ADAPT TO 3D?
We discuss the possibility of adapting our model to an agent with similar language abilities in a 3D
world (e.g., Beattie et al. (2016); Johnson et al. (2016)). This is our goal for the future, but here
we would like to share some preliminary thoughts. Generally speaking, a 3D world will pose a
12
Published as a conference paper at ICLR 2018
greater challenge for vision-related computations. The key element of our model is the attention
cube xcube that is intended for an explicit language grounding, including the channel mask xfeat and
the attention map xloc . The channel mask only depends on the sentence, and thus is expected to work
regardless of the world’s dimensionality. The interpreter depends on a sequence of score maps χ
which for now are computed as multiplying a word embedding with the feature cube (Eq. 3). A more
sophisticated definition of φ will be needed to detect objects in a 3D environment. Additionally, the
interpreter models the spatial transform of attention as a 2D convolution (Eq. 7). This assumption
will be not valid for reasoning 3D spatial relations on 2D images, and a better transform method
that accounts for perspective distortion is required. Lastly, the surrounding environment is only
partially observable to a 3D agent. A working memory, such as an LSTM added to the action
module A, will be important for navigation in this case. Despite these changes to be made, we
believe that our general explicit grounding strategy and the common detection function shared by
language grounding and prediction have shed some light on the adaptation.
5
C ONCLUSION
We have presented an end-to-end model of a virtual agent for acquiring language from a 2D world
in an interactive manner, through the visual and linguistic perception channels. After learning,
the agent is able to both interpolate and extrapolate to interpret zero-shot sentences that contain
new word combinations or even new words. This generalization ability is supported by an explicit grounding strategy that disentangles the language grounding from the subsequent languageindependent computations. It also depends on sharing a detection function between the language
grounding and prediction as the core computation. This function enables the word meanings to
transfer from the prediction to the grounding during the test time. Promising language acquisition
and generalization results have been obtained in the 2D XWORLD. We hope that this work can shed
some light on acquiring and generalizing language in a similar way in a 3D world.
ACKNOWLEDGMENTS
We thank the anonymous reviewers for providing valuable comments and suggestions. We thank the
other team members, Yuanpeng Li, Liang Zhao, Yi Yang, Zihang Dai, Qing Sun, Jianyu Wang, and
Xiaochen Lian, for helpful discussions. We thank Jack Feerick and Richard Mark for proofreading.
Finally, we specially thank Dhruv Batra for his feedback on an early version of this paper.
R EFERENCES
Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and
Lei Zhang. Bottom-up and top-down attention for image captioning and VQA. arXiv preprint
arXiv:1707.07998, 2017.
Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In
CVPR, 2016a.
Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural
networks for question answering. In ACL, 2016b.
Jacob Andreas, Dan Klein, and Sergey Levine. Modular multitask reinforcement learning with
policy sketches. In ICML, 2017.
Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual question answering. In ICCV, 2015.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly
learning to align and translate. In ICLR, 2015.
Daniel Paul Barrett, Scott Alan Bronikowski, Haonan Yu, and Jeffrey Mark Siskind. Driving under
the influence (of language). IEEE Transactions on Neural Networks and Learning Systems, 2017.
Charles Beattie, Joel Z. Leibo, Denis Teplyashin, Tom Ward, Marcus Wainwright, Heinrich Küttler,
Andrew Lefrancq, Simon Green, Vı́ctor Valdés, Amir Sadik, Julian Schrittwieser, Keith Anderson, Sarah York, Max Cant, Adam Cain, Adrian Bolton, Stephen Gaffney, Helen King, Demis
13
Published as a conference paper at ICLR 2018
Hassabis, Shane Legg, and Stig Petersen. DeepMind lab. arXiv preprint arXiv:1612.03801,
2016.
Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In
ICML, 2009.
Jerome Bruner. Child’s talk: Learning to use language. Child Language Teaching and Therapy, 1
(1), 1985.
Devendra Singh Chaplot, Kanthashree Mysore Sathyendra, Rama Kumar Pasumarthi, Dheeraj Rajagopal, and Ruslan Salakhutdinov. Gated-attention architectures for task-oriented language
grounding. In AAAI, 2018.
David L Chen and Raymond J Mooney. Learning to interpret natural language navigation instructions from observations. In AAAI, 2011.
Kyunghyun Cho, Bart van Merrienboer, Çaglar Gülçehre, Fethi Bougares, Holger Schwenk, and
Yoshua Bengio. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In EMNLP, 2014.
Noam Chomsky. Linguistics and cognitive science: Problems and mysteries. In The Chomskyan
Turn. 1991.
Harm de Vries, Florian Strub, Jérémie Mary, Hugo Larochelle, Olivier Pietquin, and Aaron C.
Courville. Modulating early visual processing by language. In NIPS, 2017.
John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and
stochastic optimization. JMLR, 12(Jul), 2011.
Haoyuan Gao, Junhua Mao, Jie Zhou, Zhiheng Huang, Lei Wang, and Wei Xu. Are you talking to a
machine? dataset and methods for multilingual image question. In NIPS, 2015.
Qiaozi Gao, Malcolm Doering, Shaohua Yang, and Joyce Y Chai. Physical causality of action verbs
in grounded language understanding. In ACL, 2016.
Marta Garnelo, Kai Arulkumaran, and Murray Shanahan. Towards deep symbolic reinforcement
learning. arXiv preprint arXiv:1609.05518, 2016.
Stevan Harnad. The symbol grounding problem. Physica D, 42, 1990.
Karl Moritz Hermann, Felix Hill, Simon Green, Fumin Wang, Ryan Faulkner, Hubert Soyer, David
Szepesvari, Wojciech Marian Czarnecki, Max Jaderberg, Denis Teplyashin, Marcus Wainwright,
Chris Apps, Demis Hassabis, and Phil Blunsom. Grounded language learning in a simulated 3d
world. In NIPS Workshop, 2017.
Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learning to
reason: End-to-end module networks for visual question answering. In ICCV, 2017.
M. Johnson, K. Hofmann, T. Hutton, and D. Bignell. The malmo platform for artificial intelligence
experimentation. In IJCAI, 2016.
I.T. Jolliffe. Principal Component Analysis. Springer Verlag, 1986.
Michał Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaśkowski. ViZDoom: A Doom-based AI research platform for visual reinforcement learning. In IEEE Conference on Computational Intelligence and Games, 2016.
Douwe Kiela, Luana Bulat, Anita L. Vero, and Stephen Clark. Virtual embodiment: A scalable
long-term strategy for artificial intelligence research. In NIPS Workshop, 2016.
Nikita Kitaev and Dan Klein. Where is misty? interpreting spatial descriptors by modeling regions
in space. In EMNLP, 2017.
Satwik Kottur, José M. F. Moura, Stefan Lee, and Dhruv Batra. Natural language does not emerge
’naturally’ in multi-agent dialog. In EMNLP, 2017.
14
Published as a conference paper at ICLR 2018
Thomas Landauer and Susan Dumais. A solution to plato’s problem: The latent semantic analysis
theory of acquisition, induction, and representation of knowledge. Psychological Review, 104,
1997.
Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention
for visual question answering. In NIPS, 2016.
Junhua Mao, Xu Wei, Yi Yang, Jiang Wang, Zhiheng Huang, and Alan L Yuille. Learning like a
child: Fast novel visual concept learning from sentence descriptions of images. In ICCV, 2015.
Tomas Mikolov, Armand Joulin, and Marco Baroni. A roadmap towards machine intelligence. arXiv
preprint arXiv:1511.08130, 2015.
Dipendra Misra, John Langford, and Yoav Artzi. Mapping instructions and visual observations to
actions with reinforcement learning. In EMNLP, 2017.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level
control through deep reinforcement learning. Nature, 518(7540), 2015.
Junhyuk Oh, Satinder P. Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization
with multi-task deep reinforcement learning. In ICML, 2017.
Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Trans. on Knowl. and Data
Eng., 22(10), October 2010.
Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question
answering. In NIPS, 2015.
Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of
textual phrases in images by reconstruction. In ECCV, 2016.
Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. In
ICLR, 2016.
M. Schuster and K. K. Paliwal. Bidirectional recurrent neural networks. IEEE Transactions on
Signal Processing, 45(11), 1997.
Jeffrey Mark Siskind. Grounding language in perception. Artificial Intelligence Review, 1994.
Jeffrey Mark Siskind. A computational study of cross-situational techniques for learning word-tomeaning mappings. Cognition, 1996.
Jeffrey Mark Siskind. Grounding the lexical semantics of verbs in visual perception using force
dynamics and event logic. Journal of Artificial Intelligence Research, 1999.
Burrhus Frederic Skinner. Verbal behavior. Copley Publishing Group, 1957.
Sainbayar Sukhbaatar, Arthur Szlam, Gabriel Synnaeve, Soumith Chintala, and Rob Fergus. Mazebase: A sandbox for learning from games. arXiv preprint arXiv:1511.07401, 2016.
Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction, volume 1. MIT
press Cambridge, 1998.
Stefanie Tellex, Thomas Kollar, Steven Dickerson, Matthew R Walter, Ashis Gopal Banerjee, Seth
Teller, and Nicholas Roy. Understanding natural language commands for robotic navigation and
mobile manipulation. In AAAI, 2011.
Michael Tomasello. Constructing a Language: A Usage-Based Theory of Language Acquisition.
Harvard University Press, 2003.
Dmitry Ulyanov.
Multicore-TSNE.
Multicore-TSNE, 2016.
https://github.com/DmitryUlyanov/
15
Published as a conference paper at ICLR 2018
L.J.P van der Maaten and G.E. Hinton. Visualizing high-dimensional data using t-SNE. Journal of
Machine Learning Research, 9, 2008.
Terry Winograd. Understanding natural language. Cognitive Psychology, 3(1), 1972.
William A. Woods. Meaning and links. AI Magazine, 28(4), 2007.
Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich
Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual
attention. In ICML, 2015.
Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks
for image question answering. In CVPR, 2016.
Haonan Yu and Jeffrey Mark Siskind. Grounded language learning from video described with sentences. In ACL, 2013.
16
Published as a conference paper at ICLR 2018
blue
brown
gray
green
orange
purple
red
yellow
apple
armadillo
artichoke
avocado
banana
bat
bathtub
beans
bear
bed
bee
beet
beetle
bird
blueberry
bookshelf
broccoli
bull
butterfly
cabbage
cactus
camel
carpet
carrot
cat
centipede
chair
cherry
circle
clock
coconut
corn
cow
crab
crocodile
cucumber
deer
desk
dinosaur
dog
donkey
dragon
dragonfly
duck
eggplant
elephant
fan
fig
fireplace
fish
fox
frog
garlic
giraffe
glove
goat
grape
greenonion
greenpepper
hedgehog
horse
kangaroo
knife
koala
ladybug
lemon
light
lion
lizard
microwave
mirror
monitor
monkey
monster
mushroom
octopus
onion
ostrich
owl
panda
peacock
penguin
pepper
pig
pineapple
plunger
potato
pumpkin
rabbit
racoon
rat
rhinoceros
rooster
seahorse
seashell
seaurchin
shrimp
snail
snake
sofa
spider
square
squirrel
stairs
star
strawberry
tiger
toilet
tomato
triangle
turtle
vacuum
wardrobe
washingmachine
watermelon
whale
wheat
zebra
east
north
northeast
northwest
south
southeast
southwest
west
7
6
5
4
3
2
blue
brown
gray
green
orange
purple
red
yellow
apple
armadillo
artichoke
avocado
banana
bat
bathtub
beans
bear
bed
bee
beet
beetle
bird
blueberry
bookshelf
broccoli
bull
butterfly
cabbage
cactus
camel
carpet
carrot
cat
centipede
chair
cherry
circle
clock
coconut
corn
cow
crab
crocodile
cucumber
deer
desk
dinosaur
dog
donkey
dragon
dragonfly
duck
eggplant
elephant
fan
fig
fireplace
fish
fox
frog
garlic
giraffe
glove
goat
grape
greenonion
greenpepper
hedgehog
horse
kangaroo
knife
koala
ladybug
lemon
light
lion
lizard
microwave
mirror
monitor
monkey
monster
mushroom
octopus
onion
ostrich
owl
panda
peacock
penguin
pepper
pig
pineapple
plunger
potato
pumpkin
rabbit
racoon
rat
rhinoceros
rooster
seahorse
seashell
seaurchin
shrimp
snail
snake
sofa
spider
square
squirrel
stairs
star
strawberry
tiger
toilet
tomato
triangle
turtle
vacuum
wardrobe
washingmachine
watermelon
whale
wheat
zebra
east
north
northeast
northwest
south
southeast
southwest
west
1
Figure 8: The Euclidean distance matrix of the 134 question groups where each group is represented
by a word label. Each row (column) represents the sampled questions that have the word label as
the answer. A matrix entry indicates the empirical expectation of the distance between the channel
masks of the sentences from two question groups. The labels are arranged into three topics: color,
object, and spatial relation. A small distance indicates that the two channel masks are similar. (Zoom
in on the screen for a better view.)
A PPENDICES
A
V ISUALIZATION AND ANALYSIS
In this section, we visualize and analyze some intermediate results of our trained model.
Channel mask xfeat . We inspect the channel mask xfeat which allows the model to select certain
feature maps from a feature cube h and predict an answer to the question s. We randomly sample
10k QA questions and compute xfeat for each of them using the grounding module L. We divide
the 10k questions into 134 groups, where each group corresponds to a different answer.4 Then
we compute an Euclidean distance matrix D where entry Dri, js is the average distance between
the xfeat of a question from the ith group and that from the jth group (Figure 8). It is clear that
4
The word “orange” is both a color word and an object word, which is why the number of groups is one less
than 119 (objects) ` 8 (spatial relations without “between”) ` 8 (colors) = 135.
17
Published as a conference paper at ICLR 2018
low
Where
is
green
What
is
between
lion
in
to
The
The
What
(a)
high
grid
apple
east
location
is in
of
blue
located
the
and
cherry
fish
?
hedgehog
?
?
is ?
?
(b)
Figure 9: Visualizations of the computation of word attention. (a) Word context vectors wl . (b) The
word attentions ol of several example questions. Each attention vector, represented by a color bar,
shows the attention accumulated over I interpretation steps.
there are three topics (object, color, and spatial relation) in the matrix. The distances computed
within a topic are much smaller than those computed across topics. This demonstrates that with the
channel mask, the model is able to look at different subsets of features for questions of different
topics, while using the same subset of features for questions of the same topic. It also implies that
the feature cube h is learned to organize feature maps according to image attributes.
Word context & attention. To intuitively demonstrate how the interpreter works, we visualize the
word context vectors wl in Eq. 6 for a total of 20k word locations l (10k from QA questions and
the other 10k from NAV commands). Each word context vector is projected down to a space of 50
dimensions using PCA (Jolliffe, 1986), after which we use t-SNE (van der Maaten & Hinton, 2008;
Ulyanov, 2016) to further reduce the dimensionality to 2. The t-SNE method uses a perplexity of 100
and a learning rate of 200, and runs for 1k iterations. The visualization of the 20k points is shown in
Figure 9a. Recall that in Eq. 6 the word attention is computed by comparing the interpreter state pi´1
with the word context vectors wl . In order to select the content words to generate meaningful score
maps via φ, the interpreter is expected to differentiate them from the remaining grammatical words
based on the contexts. This expectation is verified by the above visualization in which the context
vectors of the content words (in blue, green, and red) and those of the grammatical words (in black)
are mostly separated. Figure 9b shows some example questions whose word attentions are computed
from the word context vectors. It can be seen that the content words are successfully selected by the
interpreter.
Attention map xloc . Finally, we visualize the computation of the attention map xloc . In each of the
following six examples, the intermediate attention maps xiloc and word attentions oil (in Eq. 7) of
the three interpretation steps are shown from top to bottom. Each step shows the current attention
map xiloc overlaid on the environment image e. The last attention map x3loc is the final output of
the interpreter at the current time. Note that not all the results of the three steps are needed to
generate the final output. Some results might be discarded according to the value of the update gate
ρi . As a result, sometimes the interpreter may produce “bogus” intermediate attention maps which
do not contribute to xloc . Each example also visualizes the environment terrain map xterr (defined
in Appendix C) that perfectly detects all the objects and blocks, and thus provides a good guide
for the agent to navigate successfully without hitting walls or confounding targets. For a better
visualization, the egocentric images are converted back to the normal view.
18
Published as a conference paper at ICLR 2018
high
low
“Please move to east of cabbage .”
𝑜*%
%
𝑥"#$
𝑥&'((
Figure 10: The first example showing how xloc is computed.
19
Published as a conference paper at ICLR 2018
high
low
“The grid between
greenonion and monster is your goal .”
𝑜*%
%
𝑥"#$
𝑥&'((
Figure 11: The second example showing how xloc is computed.
20
Published as a conference paper at ICLR 2018
high
low
“Navigate
to west of
garlic
please
.”
𝑜*%
%
𝑥"#$
𝑥&'((
Figure 12: The third example showing how xloc is computed.
21
Published as a conference paper at ICLR 2018
high
low
“Navigate
to
red
beet
.”
𝑜*%
%
𝑥"#$
𝑥&'((
Figure 13: The fourth example showing how xloc is computed.
22
Published as a conference paper at ICLR 2018
high
low
“Tell
the name of the
object
which is west to snake .”
𝑜*%
%
𝑥"#$
𝑥&'((
Figure 14: The fifth example showing how xloc is computed.
23
Published as a conference paper at ICLR 2018
high
low
“What is
north
near
vacuum
?”
𝑜*%
%
𝑥"#$
𝑥&'((
Figure 15: The sixth example showing how xloc is computed.
24
Published as a conference paper at ICLR 2018
Figure 16: All the 119 ˆ 3 “ 357 object instances plus the agent (second-to-last) and the wall (last).
B
XWORLD SETUP
XWORLD is configured with 7 ˆ 7 grid maps. On each map, the open space for the agent has a size
ranging from 3 ˆ 3 to 7 ˆ 7. Smaller open spaces are set up for curriculum learning (Appendix F),
but not for testing. To keep the size of the environment image fixed, we pad the map with wall
blocks if the open space has a size less than 7. The agent has four navigation actions in total: left,
right, up, and down. In each session,
I The maximum number of time steps is four times the map size. That is, the agent only has
7 ˆ 4 “ 28 steps to reach a target.
II The number of objects on the map ranges from 1 to 5.
III The number of wall blocks on the map ranges from 0 to 15.
IV The positive reward when the agent reaches the correct location is 1.0.
V The negative rewards for hitting walls and for stepping on non-target objects are ´0.2 and ´1.0,
respectively.
VI The time penalty of each step is ´0.1.
25
Published as a conference paper at ICLR 2018
Object
Spatial relation
Color
Other
apple, armadillo, artichoke, avocado, banana, bat,
bathtub, beans, bear, bed, bee, beet,
beetle, bird, blueberry, bookshelf, broccoli, bull,
butterfly, cabbage, cactus, camel, carpet, carrot,
cat, centipede, chair, cherry, circle, clock,
coconut, corn, cow, crab, crocodile, cucumber,
deer, desk, dinosaur, dog, donkey, dragon,
dragonfly, duck, eggplant, elephant, fan, fig,
fireplace, fish, fox, frog, garlic, giraffe,
glove, goat, grape, greenonion, greenpepper, hedgehog,
horse, kangaroo, knife, koala, ladybug, lemon,
light, lion, lizard, microwave, mirror, monitor,
monkey, monster, mushroom, octopus, onion, orange,
ostrich, owl, panda, peacock, penguin, pepper,
pig, pineapple, plunger, potato, pumpkin, rabbit,
racoon, rat, rhinoceros, rooster, seahorse, seashell,
seaurchin, shrimp, snail, snake, sofa, spider,
square, squirrel, stairs, star, strawberry, tiger,
toilet, tomato, triangle, turtle, vacuum, wardrobe,
washingmachine, watermelon, whale, wheat, zebra
between,
east,
north,
northeast,
northwest,
south,
southeast,
southwest,
west
blue,
brown,
gray,
green,
orange,
purple,
red,
yellow
?, ., and,
block, by, can,
color, could, destination,
direction, does, find,
go, goal, grid,
have, identify, in,
is, locate, located,
location, me, move,
name, navigate, near,
nothing, object, of,
on, one, please,
property, reach, say,
side, target, tell,
the, thing, three,
to, two, what,
where, which, will,
you, your
Table 1: The teacher’s lexicon.
The teacher has a vocabulary size of 185. There are 9 spatial relations, 8 colors, 119 distinct object
classes, and 50 grammatical words. Every object class contains 3 different instances. All object
instances are shown in Figure 16. Every time the environment is reset, a number of object classes
are randomly sampled and an object instance is randomly sampled for each class. There are in
total 16 types of sentences that the teacher can speak, including 4 types of NAV commands and 12
types of QA questions. Each sentence type has several non-recursive templates, and corresponds
to a concrete type of tasks the agent must learn to accomplish. In total there are 1,639,015 distinct
sentences with 567,579 for NAV and 1,071,436 for QA. The sentence length varies between 2 and
13.
The object, spatial-relation, and color words of the teacher’s language are listed in Table 1. These
are the content words that can be grounded in XWORLD. All the others are grammatical words. Note
that the differentiation between the content and the grammatical words is automatically learned by
the agent according to the tasks. Every word is represented by an entry in the word embedding table.
The sentence types that the teacher can speak are listed in Table 2. Each type has a triggering
condition about when the teacher says that type of sentences. Besides the shown conditions, an
extra condition for NAV commands is that the target must be reachable from the current agent
location. An extra condition for color-related questions is that the object color must be one of the
eight defined colors. If at any time step there are multiple types triggered, we randomly sample one
for NAV and another for QA. After a sentence type is sampled, we generate a sentence according to
the corresponding sentence templates.
C
I MPLEMENTATION DETAILS
The environment image e is a 156 ˆ 156 egocentric RGB image. The CNN in F has four convolutional layers: p3, 3, 64q, p2, 2, 64q, p2, 2, 256q, p1, 1, 256q, where pa, b, cq represents a layer configuration of c kernels of size a applied at stride width b. All the four layers are ReLU activated. To
enable the agent to reason about spatial-relation words (e.g., “north”), we append an additional parametric cube to the convolutional output to obtain h. This parametric cube has the same number of
channels with the CNN output, and it is initialized with a zero mean and a zero standard deviation.
The word embedding table is initialized with a zero mean and a unit standard deviation. All the gated
RNNs (including the bidirectional RNN) in L have 128 units. All the layers in L, unless otherwise
stated, use tanh as the activation function.
26
Published as a conference paper at ICLR 2018
Sentence Type
Example
Triggering Condition
nav obj
“Please go to the apple.”
[C0] Beginning of a session. &
[C1] The reference object has a unique
name in the environment.
nav col obj
“Could you please move to the red apple?”
[C0] & [C2] There are multiple objects
that either have the same name but
different colors, or have different
names but the same color.
nav nr obj
“The north of the apple is your destination.”
[C0] & [C1]
nav bw obj
“Navigate to the grid between apple and
banana please.”
[C0] & [C3] Both reference objects
have unique names in the environment
and are separated by one block.
rec col2obj
“What is the red object?”
[C4] There is only one object that has
the specified color.
rec obj2col
“What is the color of the apple?”
[C1]
rec loc2obj
“Please tell the name of the object in the south.”
[C5] The agent is near the reference
object.
rec obj2loc
“What is the location of the apple?”
[C1] & [C5]
rec loc2col
“What color does the object in the east have?”
[C5]
rec col2loc
“Where is the red object located?”
[C4] & [C5]
rec loc obj2obj
“Identify the object which is in the east of the apple.”
[C1] & [C6] There is an object near the
reference object.
rec loc obj2col
“What is the color of the east to the apple?”
[C1] & [C6]
rec col obj2loc
“Where is the red apple?”
[C2] & [C5]
rec bw obj2obj
“What is the object between apple and banana?”
[C7] Both reference objects have unique
names in the environment and are
separated by a block.
rec bw obj2loc
“Where is the object between apple and banana?”
[C7] & [C8] The agent is near the block
which is between the two reference
objects.
rec bw obj2col
“What is the color of the object between apple
and banana?”
[C7]
Table 2: All the sixteen sentence types of the teacher.
For NAV, xloc is used as the target to reach on the image plane. However, knowing this alone does not
suffice. The agent must also be aware of walls and possibly confounding targets (other objects) in
the environment. Toward this end, MA further computes an environment terrain map xterr “ σphf q
where f P RD is a parameter vector to be learned and σ is sigmoid. We expect that xterr detects
any blocks informative for navigation. Note that xterr is unrelated to the specific command; it solely
depends on the current environment. After stacking xloc and xterr together, MA feeds them to another
CNN followed by an MLP. The CNN has two convolutional layers p3, 1, 64q and p3, 1, 4q, both with
paddings of 1. It is followed by a three-layer MLP where each layer has 512 units and is ReLU
activated.
The action module A contains a two-layer MLP of which the first layer has 512 ReLU activated
units and the second layer is softmax whose output dimension is equal to the number of actions.
We use adagrad (Duchi et al., 2011) with a learning rate of 10´5 for stochastic gradient descent
(SGD). The reward discount factor is set to 0.99. All the parameters have a default weight decay
?
of 10´4 ˆ 16. For each layer, its parameters have zero mean and a standard deviation of 1 { K,
where K is the number of parameters of that layer. We set the maximum interpretation step I “ 3.
The whole model is trained end to end.
D
BASELINE DETAILS
Some additional implementation details of the baselines in Section 4.3 are described below.
[CA] Its RNN has 512 units.
27
Published as a conference paper at ICLR 2018
CNN
𝑠
𝑒
MLP
RNN
Softmax
QA
Last
NAV
Figure 17: An overview of the baseline VL. The computations of NAV and QA only differ in the
last MLPs.
𝑠
RNN
Sequence
Average
MLP
Softmax
QA
CNN
Concat
NAV
𝑒
Figure 18: An overview of the baseline CE. The computations of NAV and QA only differ in the
last MLPs.
[VL] Its CNN has four convolutional layers p3, 2, 64q, p3, 2, 64q, p3, 2, 128q, and p3, 1, 128q. This
is followed by a fully-connected layer of size 512, which projects the feature cube to the word
embedding space. The RNN has 512 units. For either QA or NAV, the RNN’s last state goes
through a three-layer MLP of which each layer has 512 units (Figure 17).
[CE] It has the same layer-size configuration with VL (Figure 18).
[SAN] Its RNN has 256 units. Following the original approach (Yang et al., 2016), we use two
attention layers.
All the layers of the above baselines are ReLU activated.
E
E XPLORATION AND E XPERIENCE R EPLAY
The agent has one million exploration steps in total, and the exploration rate λ decreases linearly
from 1 to 0.1. At each time step, the agent takes an action a P tleft, right, up, downu with a
probability of
1
λ ¨ ` p1 ´ λq ¨ πθ pa|s, eq,
4
where πθ is the current policy, and s and e denote the current command and environment image,
respectively. To stabilize the learning, we also employ experience replay (ER) (Mnih et al., 2015).
The environment inputs, rewards, and the actions taken by the agent in the most recent 10k time steps
are stored in a replay buffer. During training, each time a minibatch tai , si , ei , ri uN
i“1 is sampled
from the buffer, using the rank-based sampler (Schaul et al., 2016) which has proven to increase the
training efficiency by prioritizing rare experiences. Then we compute the gradient as:
´
N
ÿ
`
˘`
˘
∇θ log πθ pai |si , ei q ` ∇θ vθ psi , ei q ri ` γvθ´ ps1i , e1i q ´ vθ psi , ei q ,
i“0
where i is the sample index in the batch, s1i and e1i are the command and image in the next time step,
v is the value function, θ are the current parameters, θ´ are the target parameters that have an update
delay, and γ is the discount factor. This gradient maximizes the expected reward while minimizing
the temporal-difference (TD) error. Note that because of ER, our AC method is off-policy. To avoid
introducing biases into the gradient, importance ratios are needed. However, we ignored them in
the above gradient for implementation simplicity. We found that the current implementation worked
well in practice for our problem.
28
Published as a conference paper at ICLR 2018
F
C URRICULUM LEARNING
We exploit curriculum learning (Bengio et al., 2009) to gradually increase the environment complexity to help the agent learn. The following quantities are increased in proportional to minp1, G1 { Gq,
where G1 is the number of sessions trained so far and G is the total number of curriculum sessions:
I The size of the open space on the environment map.
II The number of objects in the environment.
III The number of wall blocks.
IV The number of object classes that can be sampled from.
V The lengths of the NAV command and the QA question.
We found that this curriculum is important for an efficient learning. Specifically, the gradual changes
of quantities IV and V are supported by the findings of Siskind (1996) that children learn new words
in a linguistic corpus much faster after partial exposure to the corpus. In the experiments, we set
G “25k during training while do not have any curriculum during test (i.e., testing with the maximum
difficulty).
29
| 2cs.AI
|
A Faster Drop-in Implementation for Leaf-wise Exact Greedy
Induction of Decision Tree Using Pre-sorted Deque
arXiv:1712.06989v2 [cs.DS] 20 Dec 2017
Jianbo Ye
Pennsylvania State University
[email protected]
Abstract
This short article presents a new implementation for decision trees. By introducing presorted deques, the leaf-wise greedy tree growing strategy no longer needs to re-sort data
at each node, and takes O(kn) time and O(1) extra memory locating the best split and
branching. The consistent, superior performance – plus its simplicity and guarantee in
producing the same classification results as the standard decision trees – makes the new
implementation a drop-in replacement for leaf-wise tree induction with strong performance.
Introduction. Decision tree [6] is one of the oldest classification algorithms and has received
increasing attentions in recent years due to its use in building powerful ensemble models [1, 5, 3,
8]. It usually consists two steps: growing a binary tree that splits data points into different leaf
nodes (the induction step); prune some nodes from the developed tree to reduce the variation in
its inferences without substantially degrading the classification performance (the pruning step).
Due to its simplicity, interpretability, and generalization performance, decision tree and tree
ensembles are among the most widely adopted data analytics methods in practice.
In decision tree algorithms [14, 15, 2, 11, 18], the induction step is often the most expensive
part to compute. There have been a number of efforts trying to improve its speed. Some try to
parallelize the tree construction process [18, 9, 7, 12, 3], some try to subsample the data points
or features. A complementary approach is to speed up the algorithm itself, which is the focus
of this short article.
Prior efforts in this direction include many approximations in split-point searching, some of
which are backed up with theoretical guarantees. They are significant contributions but their
motivations might be challenged given the new message delivered in this short article — we will
discuss this issue later. What’s more important, the classical Hunt’s algorithm [6] and SPIQ [11],
as well as their variants [15, 14, 2, 18], still remain as the two dominant choices in practice,
exemplified by implementations such as Scikit-learn [13], XGBoost [3], R-package rpart [20]
and gbm [17]. Compared to the Hunt’s algorithm, SPIQ is a faster approach that pre-sorts
features and grows the tree level-wisely by scanning the entire dataset at each level. The Hunt’s
algorithm was often criticized as the need to sort data points at each splitting node, and was
often considered computationally inferior to breath-wise approach SPIQ historically [11, 16].
Nevertheless, Hunt’s algorithm which adopts a leaf-wise search strategy can easily develop a
very deep tree, a feature SPIQ would also pay a lot computations to cope with. Furthermore,
the leaf-wise search strategy realizes the possibility to prioritize which branches the tree growing
process would seek if the complete greedy tree is not constructed by the end and an early stop
criterion is used instead [19].
1
The message of this short article is simple and clear. We show that the leaf-wise search strategy
in constructing a decision tree is realizable with comparable or even faster performance with
respect to breath-wise tree growth. For an unbalanced tree where most data points are sitting
on only a small portion of leaf nodes, Hunt’s algorithm can in fact be much faster than SPIQ if a
proper data structure introduced in the short article is implemented. By maintaining pre-sorted
deques, there is no need to sort data points at each splitting node in Hunt’s algorithm. This
fact in return brings a lot speedups compared to the conventional implementation.
When we really need approximations? During the past two decades, there have been great
efforts in developing approximation methods in split-point searching for a given node. There
are mainly two rationales:
1. High memory consumption. SPIQ based approaches was claimed to be less favorable in
memory performance [3, 8], when the entire dataset is used to create node splits in a
level-wise manner. Meanwhile, in leaf-wise tree growing — many approximation methods
adopt [7, 16, 8], one only needs to pre-load a small portion of relevant data points to
create a local branching.
2. Inefficiency. The leaf-wise greedy tree growing was often criticized by the need to re-sort
features [11]. Partly for the reason to alleviate the O(kn log n) complexity of sorting,
several approximation techniques instead search a candidate subset of split points by
grouping features into several consecutive intervals and use the interval points as the
candidates [16, 10]. Such approximation takes at least O(kn log m) + O(km) time at each
node, where m ≪ n is the number of intervals, independent of the number of data points
n at the current node.
To be fair, the parallel improvement of SPIQ, SPRINT [18] can be implemented in a distributed
manner for large datasets with memory-resilient hash tables. One may wonder whether the
leaf-wise tree growing strategy is more popular because it is easier to implement. The current
standard of workstation actually allows us to handle quite large dataset to build a single decision
tree with devoted memory.
If memory is not the bottleneck for pre-sorting based method, the second inefficiency argument
is unlikely a strong reason for replacing exact greedy method with approximation method to
determine the best splits. Because as we will show in this short article, leaf-wise greedy tree
growing does not require re-sorting features. Once features are pre-sorted, it only takes O(kn)
time to find the best split points at each node. Approximation method, albeit powerful for
certain areas, often relies on extra assumptions about the data [16], and can perform inferior
than the standard method [3] or requires extra work for handling special cases. For example, a
quite popular practice to process missing attributes is to impute them with a numeric number
out of the feature range. The standard exact greedy method can faithfully find the right splits
to solely analyze the effect of a missing attribute, while the approximation method has no such
guarantee or require special modification to accurately handle the missing effect.
Remind that the advancement of hardwares in the last two decades changes the picture of how
we prioritize implementing models. For an alternative algorithm to get widely accepted, we
believe that it needs to meet several requirements [4]: (1) It must inherit the level of trust
that standard algorithms (e.g. ID3 [14], C4.5 [15], CART [2]) has attained through the many
decades of practical use; (2) it must produce significant speedups consistently; (3) it must be
simple to develop and deploy.
2
Our Implementation. Consider we have N data points, and k features. For each feature, we
pre-sort the entire dataset accordingly, and save the sorted results to a deque. Throughout the
algorithm, we assume each leaf node in the tree is associated with k sorted deques all of which
are referring the same subset of data of size n. At the beginning, all data is associated with
the root node, and the pre-sorted deques are corresponds. The Hunt’s decision tree algorithm
proceeds with the two kinds of extra operations related to the deques.
1. When the best split position is to search along a feature dimension, we traverse the
corresponding deque by computing gains against all possible candidate splits.
2. When a node is splitting data into two branches, the deque corresponding this node is
divided into a “left” deque and a “right” deque. This is achieved by first annotating each
relevant data point which branch it belongs. The orders of the elements in deques are
preserved after splitting.
Remark that both operations takes O(kn) time, and no sorting of features is needed. The
implementation details can be found in our demo repository1 , where an in-memory solution is
provided. It is worth noting that the choice of using deque rather than other containers is of two
folds: it is cache friendly in terms of traversing in the ascending order of features; it supports
O(1) time inserting and removing an element at both ends such that the second operation can
be achieved in linear time and constant extra memory.
max depth
# leaf
xgboost
ours
2
2
0.978
0.568
4
8
2.119
1.615
8
128
4.698
3.762
16
12,259
11.565
8.041
24
63,361
19.863
14.853
32
86,082
25.469
16.831
Table 1: Seconds for constructing a single tree at different max depth (no node pruning). Presorting time is excluded. Features are stored as 32-bit floats.
Experiments. We implemented Hunt’s algorithm (the tree induction step) using C++11 with
the aforementioned data structures. As for performance reference, we also compare its timing to
the fast implementation of SLIQ algorithm in XGBoost [3] on the HIGGS Data Set2 . Following
the practice of [3], we use a 1 million subsample to train the decision tree. For a fair comparison,
we experiment under the least conservative setting for both implementations. We show the
performance of our serial implementation is comparable or better than the one implemented
in XGBoost (version 0.6). The goal is, however, not to demonstrate the speed superiority of
Hunt’s algorithm over SLIQ, but to show that Hunt’s algorithm can be as efficient as SLIQ,
which to the best of our knowledge was not known to the literature. The statistics are reported
in Table 1 based on 20 repeated runs in single-thread mode on iMac 3.3 GHz Intel Core i5.
min leaf samples
# leaf
time (seconds)
1
86,082
16.831
5
31,237
10.762
10
15,146
8.750
50
3,764
7.526
100
2,205
6.938
Table 2: Seconds for constructing a single tree at different min leaf samples (no node pruning).
Pre-sorting time is excluded.
1
2
https://github.com/bobye/Decision_Tree_PDeque
https://archive.ics.uci.edu/ml/datasets/HIGGS
3
In the second experiment, we show the time performance of our method can be predictably
improved by setting the minimal number of samples of leaf node. We list the timing result
in Table 2. By setting it to a larger value, the time to construct an exact greedy tree can be
greatly reduced.
Conclusion. In this short article, we demonstrate that the Hunt’s algorithm for the leaf-wise
induction of a decision tree can be implemented efficiently in modern computer architectures
using presorted deque and have similar serial performance to the SLIQ [11] implementation in
highly optimized library XGBoost [3].
References
[1] Leo Breiman. Random forests. Machine learning, 45(1):5–32, 2001.
[2] Leo Breiman, Jerome Friedman, Charles J Stone, and Richard A Olshen. Classification and regression trees. CRC
press, 1984.
[3] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd
international conference on knowledge discovery and data mining, pages 785–794. ACM, 2016.
[4] Yufei Ding, Yue Zhao, Xipeng Shen, Madanlal Musuvathi, and Todd Mytkowicz. Yinyang k-means: A drop-in
replacement of the classic k-means with consistent speedup. In International Conference on Machine Learning, pages
579–587, 2015.
[5] Jerome H Friedman. Greedy function approximation: a gradient boosting machine. Annals of statistics, pages
1189–1232, 2001.
[6] Earl B Hunt, Janet Marin, and Philip J Stone. Experiments in induction. 1966.
[7] Ruoming Jin and Gagan Agrawal. Communication and memory efficient parallel decision tree construction. In
Proceedings of the 2003 SIAM International Conference on Data Mining, pages 119–129. SIAM, 2003.
[8] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm:
A highly efficient gradient boosting decision tree. In Advances in Neural Information Processing Systems, pages
3149–3157, 2017.
[9] Richard Kufrin. Decision trees on parallel processors. Machine Intelligence and Pattern Recognition, 20:279–306,
1997.
[10] Ping Li, Qiang Wu, and Christopher J Burges. Mcrank: Learning to rank using multiple classification and gradient
boosting. In Advances in neural information processing systems, pages 897–904, 2008.
[11] Manish Mehta, Rakesh Agrawal, and Jorma Rissanen. Sliq: A fast scalable classifier for data mining. Advances in
Database TechnologyEDBT’96, pages 18–32, 1996.
[12] Qi Meng, Guolin Ke, Taifeng Wang, Wei Chen, Qiwei Ye, Zhi-Ming Ma, and Tieyan Liu. A communication-efficient
parallel algorithm for decision tree. In Advances in Neural Information Processing Systems, pages 1279–1287, 2016.
[13] Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu
Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. Journal
of Machine Learning Research, 12(Oct):2825–2830, 2011.
[14] J. Ross Quinlan. Induction of decision trees. Machine learning, 1(1):81–106, 1986.
[15] J Ross Quinlan. C4. 5: programs for machine learning. Elsevier, 2014.
[16] Sanjay Ranka and V Singh. Clouds: A decision tree classifier for large datasets. In Proceedings of the 4th Knowledge
Discovery and Data Mining Conference, pages 2–8, 1998.
[17] Greg Ridgeway. Generalized boosted models: A guide to the gbm package, 2007.
[18] John Shafer, Rakesh Agrawal, and Manish Mehta. Sprint: A scalable parallel classi er for data mining. In Proc. 1996
Int. Conf. Very Large Data Bases, pages 544–555, 1996.
[19] Haijian Shi. Best-first decision tree learning. PhD thesis, The University of Waikato, 2007.
[20] Terry Therneau, Beth Atkinson, and Brian Ripley. rpart: Recursive partitioning and regression trees. r package
version 4.1–10, 2015.
4
| 8cs.DS
|
arXiv:1604.06687v1 [cs.DS] 22 Apr 2016
Faster Average Case Low Memory Semi-External
Construction of the Burrows-Wheeler Transform
German Tischler
Abstract. The Burrows Wheeler transform has applications in data compression as well as full
text indexing. Despite its important applications and various existing algorithmic approaches the
construction of the transform for large data sets is still challenging. In this paper we present a new
semi external memory algorithm for constructing the Burrows Wheeler transform. It is capable of
constructing the transform for an input text of length n over a finite alphabet in time O(n log 2 log n)
on average, if sufficient internal memory is available to hold a fixed fraction of the input text. In
the worst case the run-time is O(n log n log log n). The amount of space used by the algorithm in
external memory is O(n) bits. Based on the serial version we also present a shared memory parallel
algorithm running in time O( np max{log2 log n + log p}) on average when p processors are available.
1. Introduction
The Burrows-Wheeler transform (BWT) was introduced to facilitate the lossless compression of data
(cf. [3]). It has an intrinsic connection to some data structures used for full text indexing like the suffix
array (cf. [19]) and is at the heart of some compressed full text self indexes like the FM index (see [9]).
The FM index requires no more space than the k’th order entropy compressed input text plus some
asymptotically negligible supporting data structures. Many construction algorithms for the BWT are
based on its relation to the suffix array, which can be computed from the input text in time linear
in the length of that text (see e.g. [13, 22]). While these algorithms run in linear time and are thus
theoretically optimal they require O(n log n)1 bits of space for the uncompressed suffix array given an
input text of length n while the text itself can be stored in a space of n⌈log σ⌉ bits for an alphabet of
size σ where we often have σ ≪ n and in most applications σ is constant. Algorithms for computing
the suffix array in external memory have been proposed (see e.g. [2, 6]) but these algorithms require
large amounts of space and input/output in external memory. An asymptotically optimal internal
memory solution concerning time and space has been proposed [11]. However the space usage of
this algorithm is O(n) bits for constant alphabets, where an inspection of the algorithm suggests
that the actual practical memory usage of the algorithm is several times the size of the text in bits.
The practical space usage of the algorithm subsequently presented in [23] is lower (i.e. the involved
constants are smaller) while theoretically not linear. It however still requires multiple times as much
space as the input text. A sample implementation given by the authors only works for input sizes of up
to 232 (see [1]) and only handles a single level of the recursive algorithm. Given the implementation
complexity of the algorithm it remains unclear if it would scale well. Crochemore et al present an
Supported by the Wellcome Trust.
Full version of an extended abstract which appeared in the Proceedings of the 2nd International Conference on Algorithms for Big Data.
1 by log we mean log in this paper
2
2
German Tischler
algorithm computing the BWT in quadratic time with constant additional space (see [4]). In [1] Beller
et al propose a semi external algorithm for the construction of the BWT based on induced sorting.
An algorithm is called semi external if it uses non negligible amounts of internal as well as external
memory. According to the authors the algorithm scales to arbitrary input sizes and uses about one
byte (i.e. 8 bits) per input symbol in internal memory. An algorithm constructing the BWT of a text
by block-wise merging using a finite amount of internal memory is presented in [8]. The algorithm
partitions the text into a set of fixed size blocks. The run-time is O(n2 /b) for a block size of b and
a text length of n. It requires an amount of internal memory which is roughly sufficient to compute
the suffix array of a single of these block. In particular the amount of internal memory used can be
smaller than the space required for the text. In this paper we modify this algorithm to run in time
O(n log n log log n) in the worst case and O(n log2 log n) on average for the case where we are able to
keep a fixed fraction of the text in memory. Assuming the practically common case of a finite alphabet
the algorithm in [8] uses blocks of size O(n/ log n) when provided with O(n) bits of space in internal
memory so its total run-time for this setting is O(n log n). In consequence our algorithm is faster on
average and slower by O(log log n) for a very unlikely worst case. Compared to the algorithm presented
in [1] our algorithm can work with less internal memory. For DNA for instance the complete text can
be held in memory using about 2 bits per symbol which is significantly less than a full byte (8 bits)
per character. We also propose a parallel version of our algorithm for the shared memory model. This
algorithm runs in time O( np max{log2 n, log p}) on average and with high probability when using p
processors.
The rest of this paper is structured as follows. In Section 2 we introduce definitions which we
will be using throughout the paper. Section 3 gives a short overview over the algorithm of Ferragina
et al (cf. [8]) on which our new algorithm is based. In Section 4 we present an algorithm for computing
the minimal periods of all prefixes of a string in succinct space. We show how to compute the suffix
order for blocks of b suffixes drawn from a circular version of a string of length n ≥ b in optimal
time O(b) and space O(b log b) bits after collecting some repetition information about the string in a
preprocessing phase taking time O(n) and O(b log b) bits of space in Section 5. In Section 6 we show
how to merge the BWTs of two adjacent text blocks of length bl and br respectively given by their
BWT in external memory in time O ((bl + br ) log log(bl + br )) and sufficient space in internal memory
to hold a rank dictionary for the left block. We describe in Section 7 how to combine the block sorting
from Section 5 and block merging from Section 6 to obtain a BWT construction algorithm based on a
balanced instead of a skewed merge tree. In Section 8 we describe a parallel version of our algorithm.
Finally in section 9 we summarise our results and discuss some remaining open problems.
2. Definitions
For a string s = s0 s1 s2 . . . sm−1 of length |s| = m ∈ N we define s[i] = si for 0 ≤ i < m and for
s = s0 s1 . . . we define s[i] = si for 0 ≤ i. For a finite word u and a finite or infinite word v we write
their concatenation as uv. For any finite words u, x and finite or infinite words w, v such that w = uxv
we call u a prefix, v a suffix and x a factor of w. A prefix, suffix or factor of some string w is called
proper if it is not identical with w. The empty word consisting of no symbols is denoted by ǫ. For a
string s and indices 0 ≤ i ≤ j < |s| we denote the factor s[i]s[i + 1] . . . s[j] by s[i, j]. For any i, j such
that i > j the term s[i, j] denotes the empty word. A finite word w has period p iff w[i] = w[i + p]
for i = 0, . . . , |w| − p − 1 and an infinite word w has period p iff w[i] = w[i + p] for i = 0, 1, . . .. For a
finite word u and k ∈ N the k’th power uk of u is defined by u0 = ǫ and ui+1 = ui u for i ∈ N. A word
w is primitive if it is not a power of a word u such that |u| < |w|. A word u is a root of w if w = uk
for some k ∈ N. A word w is a square if there is a word u such that w = u2 . Throughout this paper
let Σ = {0, 1, . . . , σ − 1} denote a finite alphabet for some σ ∈ N, σ > 0 and let t = t0 t1 . . . tn−1 ∈ Σn
denote a finite string of length n > 0. We define the semi infinite string t̃ by t̃[i] = t[i − ⌊i/n⌋n] for
i ≥ 0. We define the suffix t̃i of t̃ as t̃i = t̃[i]t̃[i + 1] . . . and t̃i < t̃j for i, j ∈ N, i 6= j iff either t̃i = t̃j
and i < j or for the smallest ℓ ≥ 0 such that t̃[i + ℓ] 6= t̃[j + ℓ] we have t̃[i + ℓ] < t̃[j + ℓ]. The suffix
Faster Average Case Low Memory Semi-External Construction of the BWT
3
array A of t is defined as the permutation of the numbers 0, 1, . . . , n − 1 such that t̃A[i] < t̃A[i+1]
for i = 0, 1, . . . , n − 2 and the Burrows-Wheeler transform (BWT) B = b0 b1 . . . bn−1 of t is given by
bi = t̃[A[i] + n − 1] for i = 0, 1, . . . , n − 1.
3. BWT construction by block-wise merging
We give a short high level description of the algorithm by Ferragina et al. in [8] as we will be modifying
it. Unlike our algorithm it assumes the input string to have a unique minimal terminator symbol. Given
a block size b the input string t is partitioned into c = ⌈n/b⌉ blocks T0 , T1 , . . . , Tc−1 of roughly equal
size. The algorithm starts by suffix sorting the last block, computing its BWT Bc−1 and the bit array
gtc−1 which denotes for each suffix in Tc−1 but the first whether it is smaller or larger than the first.
The BWT of Ti . . . Tc−1 for i < c − 1 is computed by first computing the suffix array for the suffixes
starting in Ti by using the text of Ti and Ti+1 in memory and handling the comparison of suffixes
starting in Ti but equal until both have entered Ti+1 by using the bit vector gti+1 which explicitly
stores the result of this comparison. The BWTs of Ti and Ti+1 Ti+2 . . . Tc−1 are merged by computing
the ranks of the suffixes starting in Ti+1 Ti+2 . . . Tc−1 in the sorted set of suffixes of Ti and computing
a gap array Gi which denotes how many suffixes from Ti+1 Ti+2 . . . Tc−1 are to be placed before the
suffixes in Ti , between two adjacent suffixes in Ti and after all suffixes in Ti . This process follows a
backward search of Ti+1 Ti+2 . . . Tc−1 in Ti . Using the array Gi it is simple to merge the two BWTs
together. For computing the rank of a suffix from Ti+1 . . . Tc−1 it is necessary to know whether it is
smaller or larger than the one at the start of Ti+1 Ti+2 . . . Tc−1 as Bi is not a conventional BWT. For
further details about the algorithm the reader is referred to [8].
4. Computing the Minimal Period of the Prefixes of a String
A border of a string w is a string u such that w = uv = xu and |v| = |x| 6= 0. The border array B of
a string w is the integer array of length |w| such that B[i] stores the length of the longest border of
w[0, i] for i = 0, 1, . . . , |w| − 1. When a string w has the border array B, then the minimal period of
w[0, i] for 0 ≤ i < |w| can be obtained as i + 1 − B[i] (see e.g. Proposition 1.5 in [5]). The uncompressed
border array for a string of length b can be stored using b words, i.e. b⌈log b⌉ bits. Due to the properties
of the border array (the value at index i + 1 can be at most 1 larger than the value at index i and
all values are non-negative) it is however easy to use a succinct version in O(b) bits. For the rest
of the section let w denote a string and let B denote the border array of w. Then we can represent
B as the bit vector B̃ = 10B[0]−B[1]+1 10B[1]−B[2]+1 1 . . . 0B[b−1]−B[b−2]+1 1 and B[i] can be retrieved as
i − rank0 (B̃, select1 (B̃, i)) for i = 0, 1, . . . , b − 1 where rank0 (B̃, i) denotes the number of 0 bits in B̃ up
to and including index i and select1 (B̃, i) returns the index of the i + 1’th 1 bit in B̃. Indexes for the
rank and select operations can be stored in o(b) extra bits (see e.g. [21]). The most common indexes
for these operations are tailored for static bit vectors while the computation of the border array (see
for instance section 1.6 in [5]) incrementally extends the array while using the already existing prefix
to deduce the next value. The rank and select indexes as presented in section 6.1 in [21] are however
easily adapted for the case of only appending bits at the end of the bit vector. Both types of indexes
partition the indexed bit vector into blocks. In the case of rank the blocks consist of β0 = ⌈log2 |B̃|⌉
bits of B̃. An array R0 is used to explicitly store rank0 (B̃, β0 i) for i = 0, 1, . . . , ⌈ βn0 ⌉ − 1 where each
element of R0 takes ⌈log |B̃|⌉ bits. The blocks are again partitioned into smaller blocks of length
β1 = ⌈ log2|B̃| ⌉ (for the sake of simplicity in the discussion we assume β1 divides β0 . The method is
easily extended to the case where β0 is not exactly a multiple of β1 ). Another integer array R1 is used
to store rank0 (B̃, β1 i) − rank0 (B̃, ⌊ ββ10i ⌋) for i = 0, 1, . . . , ⌈ βn1 ⌉ − 1. Each element of R1 takes ⌈log β0 ⌉
bits. We can incrementally build the rank dictionary by adding one value to R0 and ⌈ ββ10 ⌉ values to R1
each time β0 bits have been appended to the underlying bit vector. This can trivially be done in time
O(β0 ) for each block of size β0 . The answers for a block yet to be completed can be stored explicitly in
4
German Tischler
an array of size β0 which takes β0 ⌈log β0 ⌉ = O(log2 B̃ log B̃) bits and thus is asymptotically negligible
compared to the space used for the bit vector itself. Using this method the rank operation can be
computed in time O(1) at any time and the time for adding each single bit to the vector is amortised
constant. The method for keeping constant time select operations is conceptually very similar. The
data structure partitions the bit vector B̃ such that each block contains ζ0 = ⌈log2 |B̃|⌉ bits of value
1. Again we can keep enough space to store the answers for a single incomplete block explicitly and
extend the index for select each time the total number of 1 bits in the intermediate bit vector reaches
an integer multiple of ζ0 .
Lemma 1. Let x denote a string over a finite alphabet Γ = {0, 1, . . . , γ − 1}. The border array of x can
be computed in time O(|x|) and using space O(|x| log γ) bits.
Lemma 2. Let x denote a string over a finite alphabet Γ = {0, 1, . . . , γ − 1}. The sequence of minimal
periods of the prefixes of x can be enumerated in time O(|x|) and space O(|x| log γ) bits.
Using the periodicity lemma (see Lemma 1 in [17]) we in addition obtain the following lemma.
Lemma 3. Let x denote a string over a finite alphabet Γ = {0, 1, . . . , γ − 1}. If x has any period
|x|
q ≤ ⌊ |x|
2 ⌋ then there exists a minimal period p of x such that p ≤ ⌊ 2 ⌋ and p divides all other periods
of x whose value does not exceed ⌊ |x|
2 ⌋. It is decidable in time O(|x|) and space O(x log γ) bits whether
|x|
x has any period p ≤ ⌊ 2 ⌋ and if it has any such periods then their minimum can be computed in the
same time and space bounds.
Proof. Let x denote a string over a finite alphabet Γ = {0, 1, . . . , γ − 1} and let P denote the set of
all periods of x whose value does not exceed ⌊ |x|
2 ⌋. If P is empty, then the lemma holds as x has no
relevant periods. Otherwise let p be the minimal element of P . Then for any element q of P we have
p+ q ≤ |x| as p ≤ q ≤ ⌊ |x|
2 ⌋. According to the periodicity lemma this implies that the greatest common
divisor of p and q is also a period of x. As P contains all the periods of x not exceeding ⌊ |x|
2 ⌋, p is
the minimal element of P and the greatest common divisor of p and q divides p it follows that this
greatest common divisor is p. As p is the minimal period of x it can be computed in time O(|x|) and
space O(|x| log γ) bits according to Lemma 2.
5. Sorting single blocks
The algorithm by Ferragina et al processes each single block relying on knowledge about the priorly
fully processed following block, in case of the last block the terminator. For our algorithm we need to
be able to sort a single block without knowing the complete sorted order of the next block. For this
purpose we need to be able to handle repetitions, one of the major challenges along the way, efficiently.
For the following argumentation we will need the occurring block sizes to be as similar as possible, so
we will deduce a final block size from a target block size chosen to satisfy memory constraints. Let b′
denote a preliminary target block size such that 0 < b′ ≤ n. From b′ we can deduce a final block size
b as
&
'
n
b = n .
(1)
b′
In Equation 1 the denominator denotes the number ν of blocks required for the preliminary block size
b′ . The final block size is obtained as the smallest number b of symbols required to keep this number
of blocks ν, in particular we have b ≤ b′ As b is the smallest such number we have µ′ = νb − n such
that 0 ≤ µ′ < ν and equally µ = ν − µ′ such that 1 ≤ µ ≤ ν. In consequence we can decompose the
set of indices [0, n) on t into µ sub intervals of length b and ν − µ sub intervals of length b − 1 and
define the index blocks Bi by
[ib, (i + 1)b)
for 0 ≤ i < µ
Bi =
(2)
[µb + (i − µ)(b − 1), µb + (i + 1 − µ)(b − 1) for µ ≤ i < ν
Faster Average Case Low Memory Semi-External Construction of the BWT
5
Informally we have µ blocks of size b and ν − µ blocks of size b − 1 such that the blocks of size b
precede the ones of size b − 1. Blocks of size b − 1 only exist if b does not divide n.
When comparing two suffixes of t̃ starting at indices i and j such that i < j we may encounter
two cases. In the first case we have t̃[i, j − 1] 6= t̃[j, j + (j − i − 1)]. Then the comparison requires
handling at most j − i symbol pairs, i.e. at most 2b − 2 symbols if the two suffixes start in the same
block of length b as defined above. In the second case we have t̃[i, j − 1] = t̃[j, j + (j − i − 1)], i.e. a
square with period j −i at index i in t̃. Thus situations which require us to have access to ω(b) symbols
for sorting the suffixes of a block of indices on t̃ are induced by repetitions of periods strictly smaller
than b. In consequence only repetitions with a period smaller than b are relevant for our block sorting.
Definition 1. Let B denote a block of b indices on t̃ starting at index i, i.e. the set of suffixes t̃i+j for
j = 0, 1, . . . , b − 1.
• B propagates a repetition of period p, 1 ≤ p ≤ b iff t̃i [0, b + 2p − 1] has period p.
• B generates a repetition of period p, 1 ≤ p ≤ b iff t̃[b − p, b − 1] = t̃[b, b + p − 1] and the block of b
suffixes starting at i + b propagates a repetition of period p.
If a block propagates repetitions of any periods, then there is a unique minimal period dividing all
other propagated periods according to Lemma 3. This unique minimal period can then be computed
in time O(b) and space O(b log σ) bits using minor modifications of standard string algorithms (see
Section 4). As there is a unique minimal period propagated by a block if any and as for repetition
generation we are only interested in periods which are propagated by the next block we can compute
the relevant generation properties of a block in the same time and space bounds. The suffix order on t̃
defined as above defines the order of two suffixes at indices i and j such that i 6= j by index comparison
if t̃i = t̃j . This case can only appear if the text t itself has a period p which is smaller than n and
divides n, i.e. when t = αi for some string α and some integer i > 1. For sorting the suffixes in our
blocks we can avoid this case. If t is an integer power of a string α such that |α| < b for some exponent
k > 1, then we will observe that each of our blocks propagates period p, thus this case is easily detected
given precomputed repetition propagation information. In this case the Burrows-Wheeler transform
of t can be obtained from the Burrows-Wheeler transform for α by repeating each symbol k times, e.g.
if abc is the Burrows-Wheeler transform for α and we have k = 2 then the transform for t is aabbcc.
Similar arguments can be employed to obtain the suffix array and inverse suffix array of t from the
reduced variants for α. Thus in the following we will without loss of generality assume that t is not an
integer power of a string of period smaller than b. In consequence all suffix comparisons required for
sorting the suffixes starting in a block are decided by character inequality and not by position. Using
information about short period repetitions in the input string, we are able to handle the sorting of a
single block of suffixes extending beyond the end of the block efficiently by reducing long repetitions
as we show in the following lemma.
Lemma 4. A block of b circular suffixes of t̃ can be sorted in lexicographical order using time O(b) and
space O(b log b) bits using precomputed repetition propagation data.
Proof. The pairwise order between all suffixes of the block is decided by character inequalities. However
these character inequalities may happen after Ω(n) symbols in each case. Note however that if we
compare two strings u and v lexicographically such that u = αβ k γ and v = αβ k δ we obtain the same
result as when comparing u′ = αγ and v ′ = αδ, i.e. as when we remove the repetition of β from both.
Now assume we are sorting the circular suffixes of t for a block starting at index i with block length b,
i.e. we are sorting b strings of infinite length where for each pair of strings the comparison ends after
less than n steps. When comparing a pair of such strings starting at offsets j and k in the block such
that without loss of generality j < k we may encounter the following cases.
1. There exists ℓ < (b − j) + (k − j) = b + k − 2j such that t̃[i + j + ℓ] 6= t̃[i + k + ℓ]. Then as
b + k − 2j < 2b the order between the strings would be correctly reflected in a suffix sorting of
t̃[i, i + 2b − 1], where 2b is O(b).
6
German Tischler
2. Otherwise we have t̃[i + j, i + k − 1] = t̃[i + k, i + k + (k − j) − 1], i.e. a square of period k − j
at position i + j in t̃ and this repetition extends to at least index i + b + (k + j) − 1 in t̃.
There is thus a complete instance of a rotation of the root of the square in t̃ after the end of the
considered block (the root of the repetition may overlap the block boundary or come to lie outside
of the block). In consequence the block may generate a repetition of period k − j depending on
whether the following block propagates a repetition of this period. If t̃[i + j, i + k − 1] is an
integer power of a shorter primitive string u, then let p = |u|, otherwise let p = k − j. Now we
may encounter two sub cases. If the following block does not propagate period p, then the order
between our two strings starting at index i + j and i + k in t̃ would be correctly reflected in
a suffix sorting of t̃[i, i + 2b + 2p − 1] due to the definition of the propagation of a repetition.
Note that 2b + 2p − 1 is O(b) as p ≤ k − j < b. Now assume the following block does propagate
period p. Let β = t̃[i + b, i + b + p − 1] and let α = t̃[i, i + (b − p) − 1]. Then the block can be
written as αβ. Further we know that t̃[i]t̃[i + 1] . . . = αβ m γ for some integer m > 1 + ⌈b/p⌉ and
γ[0, p − 1] 6= β. Using precomputed information about the single blocks’ repetition propagation
properties and adequate data structures it is simple to deduce the numbers m and γ in time
O(b). For each block propagating a repetition we store the index of the closest following block
in t̃ which no longer propagates this repetition in an array. This array can be computed in time
O(n) with a single scan over the list of repetitions propagated by the single blocks. Even for a
moderate block size in O(log2 n) the space required for this array is asymptotically negligible
compared to the text and it can be stored in external memory. Now for computing m and γ we
can go to the last following block still propagating the repetition and perform a naive scan of the
text until the repetition ends. This scan will terminate in time O(b). For determining the order
of the suffixes in the current block it is sufficient to sort αβ 1+⌈b/p⌉ γ which has length O(b).
Sorting the suffixes of t̃ starting in a given block of length b using the precomputed repetition
generation and propagation block properties can thus be performed in time O(b) using space O(b log b)
bits in internal memory. Given the explicit suffix sorting for a block, it is trivial to determine, whether
any other suffix in the block is lexicographically smaller or larger than the first suffix of the block and
store the resulting bit vector in external memory.
For forward searching using the suffix array it is useful to in addition have the longest common
min{|u|,|v|}
prefix (LCP) array. For two finite strings u, v let LCP(u, v) = argmaxl=0
u[0, ℓ − 1] = v[0, ℓ − 1].
For two infinite strings u, v let LCP(u, v) = ∞ if u = v and LCP(u, v) = i where i is the smallest non
negative integer such that u[i] 6= v[i] otherwise. For a block of indices [i, i + b − 1] on t̃ such that i ∈ N
let A denote the permutation of i, i + 1, . . . , i + b − 1 such that t̃A[j] < t̃A[j+1] for j = 0, 1, . . . , b − 2.
Then the LCP array of the block is defined by LCP[0] = 0 and LCP[i] = LCP(t̃A[i−1] , t̃A[i] ) for
i = 1, 2, . . . , b − 1. Using a repetition reduction method similar to the suffix sorting case we obtain the
following result.
Lemma 5. The LCP array for a block of b circular suffixes on t̃ can be computed in time O(b) and
space O(b log b) bits using precomputed repetition propagation data.
Proof. As the input string is without loss of generality not an integer power of a string of period
p < b all suffix comparisons on t̃ within a block are decided by character inequalities. In consequence
the values in the LCP array of a block are finite but the array can contain values in Ω(n) as the
considered suffixes extend beyond the end of the block. Given a finite string s and the suffix array of
s the LCP array for s can be computed in time O(|s|) and stored in an array of size O(s log s) bits
(see [16]). Thus if the block does not generate a repetition, then we can compute the LCP array for
the block in time O(b) and represent it as an array using O(b log b) bits. This involves extending the
text block by O(b) elements (see above), computing the LCP array using the suffix array and the text
and filtering the LCP array so that only suffixes inside the original block remain. If we find suffixes
outside the original block between two original suffixes than we need to combine the minimum of the
corresponding entries with the second original block value, e.g. if the block starts at index i and we
Faster Average Case Low Memory Semi-External Construction of the BWT
7
have A[j] < i + b, A[j + 1] ≥ i + b, A[j + 2] ≥ i + b, . . . , A[j + k] < i + b then the LCP value between the
suffixes at the block indices A[j] and A[j + k] is min{LCP[j + 1], LCP[j + 2], . . . , LCP[j + k]} and we
drop the LCP values relating to suffixes outside the block. Now assume the block generates a repetition
of period p. Then as described in case 2 of the proof of Lemma 4 it is sufficient to suffix sort the string
αβ 1+⌈b/p⌉ γ as defined above to obtain the order of the suffixes in the current block. This block has
length O(b), so its LCP array can be computed in time O(b) and stored in space O(b log b) bits. We
first remove the values corresponding to suffixes outside the block as we did for the non propagating
case. Then the LCP value at index j > 1 is correct if LCP[j] < min{i + b − A[j − 1], i + b − A[j]},
i.e. if the LCP between the two corresponding suffixes is too short to reach the next block for at
least one of the two. Otherwise both suffixes are equal at least until the comparison of the two has
extended to the next block. Then the stored LCP value will be too small and thus incorrect. We can
however easily correct these values using the precomputed repetition propagation information. We
can compute in time O(b) at which offset from the start of the next block the propagated repetition
first breaks. Let o denote this offset. Then we obtain the correct LCP value for index j > 0 as
LCP[j] = (i + b − max{A[j − 1], A[j]}) + o. If we store this value explicitly, then the LCP array will
take space O(b log n) bits. As we can however detect and correct the incorrect LCP values in constant
time given the suffix array and number o we can leave the LCP array as is taking space O(b log b) and
still have constant time access to the correct values by computing them as required.
Using the suffix and LCP array the time for a forward search of a pattern of length m in a block
of size b reduces from O(m log b) to O(m + log b) (see [19]).
6. Merging Pairs of Adjacent Blocks
In our modified algorithm we replace the completely skewed binary merge tree used in [8] by a balanced
binary merge tree. Consequently we will need to be able to merge blocks with a block size in Ω(n).
For merging two adjacent blocks we need the following components:
1. The BWT of the left and right block. These can be compressed and in external memory as they
will be scanned sequentially.
2. An internal memory index of the left block suitable for backward search in O(1) time per step.
An FM type index using space bl Hk + o(n log σ) bits can be used where bl is the length of the
left block and Hk denotes the k’th order entropy of the left block (see [21]).
3. The gt bit vectors for the left and right block. Scanned sequentially and thus can be read
streaming from external memory.
4. The number of circular suffixes in the left block smaller than the rightmost suffix of the right
block. Used as the start point for the backward search.
5. The gap array G.
The first three are equivalent to those used in [8]. The rank of the rightmost suffix in the right block
relative to the suffixes of the left block can be obtained by employing forward search on one or more
text blocks. If the left block is a single block which was produced by explicit suffix sorting using the
method of Section 5, then the rank can be obtained using classical forward search in the suffix array
while using the adjoined LCP array. This takes time O(n+ log b) in the worst case (on average this can
be expected to be O(log n + log b), see [25]). If the left block was already obtained by merging c blocks
together, then the desired rank can be obtained as the sum of the ranks of the suffix relative to all
single blocks composing the left block in time O(c(n+log b)). Assuming the blocks are merged together
in a balanced binary merge tree the total time used for forward searches is O( nb log nb n) in the worst
case and O( nb log nb log n) on average. If we choose b ∈ O( logn n ) then this becomes O(n log n log log n).
The memory required for the index of the left block in internal memory will be bl ⌈log σ⌉ + o(bl log σ)
bits for a left block size of bl assuming that the entropy compression is ineffective. This leaves us with
the space required for the gap array. In the original algorithm this is a conventional array in internal
memory taking space O(b log n) bits for a left block size of b. As we want to be able to merge blocks
8
German Tischler
with size in Ω(n) this space requirement is too high. Using Elias γ code (cf. [7]) we can store the gap
array for merging a left and right block of length bl and br respectively in O(bl + br ) bits of space as
we show in the following lemma.
Pl−1
Lemma 6. Let G denote an array of length ℓ such that G[i] ∈ N for 0 ≤ i < ℓ and i=0
G[i] = s for
some s ∈ N. Then the γ code for G takes O(ℓ + s) bits.
Proof. Elias γ code stores the number z > 0 as the bit sequence 0⌊log z⌋ bin(z) of length 1 + 2⌊log z⌋
where bin(z) denotes the binary representation of z (most significant to least significant bit left to
right). The γ code for G thus requires
ℓ+2
ℓ−1
X
i=0
⌊log(G[i] + 1)⌋
bits. As γ code is unable to represent the number zero but zero is a valid entry in G we add 1 to each
element. We have
bY
bX
bX
l −1
l −1
l −1
(G[i] + 1) .
log(G[i] + 1) = log
⌊log(G[i] + 1)⌋ ≤
i=0
i=0
i=0
and the last product is maximal for G[0] = G[1] = . . . = G[ℓ − 1] ≈ sℓ (conceptually due to the fact
that x(1 − x) has its maximum at x = (1 − x) = 12 , i.e. a product of two positive real numbers of
constant sum is maximal if both are chosen as half the sum). Thus the maximum space used for the
gap array in γ code is bounded by
ℓ−1
s
X
⌊log G[i] + 1⌋ ≤ ℓ 1 + 2 log
ℓ+2
+1
ℓ
i=0
For s ≤ ℓ the argument of the logarithm is between 1 and 2 thus the space usage is bounded by 3ℓ.
For s > ℓ let s = 2r l for some r ∈ R, r > 0. Then we have
ℓ(1 + 2 log( sl + 1)) ≤
≤
ℓ(1 + 2 + 2 log sl )
s
3+2r
2r (3 + 2r) = s 2r ≤ s(3 +
1
loge 2 )
which is in O(s). Thus in both cases the space is in O(ℓ + s) bits.
≤ 5s
Elias γ code however is not suitable for efficient updating as we would need it for computing
the gap array. During the computation of a gap array each step of the backward search leads to an
increment of exactly one position in the gap array. In particular the sum over the elements of the
array after s steps is s. The γ code representation of the gap array does not allow efficient updating of
single elements. We overcome this problem by producing partial sparse gap arrays, which we write to
external memory. We accumulate an amount c of indices for incrementing in internal memory before
we write a sparse variant to external memory, which implies that the sum over the sparse array written
will be c. Whenever we have produced two sparse gap arrays in external memory featuring the same
element sum, then we merge the two together into a sparse array of twice that sum. At the end of the
process we need to merge the remaining set of arrays into a single final gap array. Let G denote a gap
Pℓ−1
array of length ℓ such that i=0 G[i] = s for some s > 0 with k > 0 non-zero values. Let i0 , i1 , . . . , ik−1
denote the sequence of indices of non-zero values in G in increasing order. We store two γ coded bit
streams for representing G. In the first we store the sequence i0 + 1, i1 − i0 , i2 − i1 , . . . , ik−1 − ik−2
and in the second we store the subsequence of non-zero values in G. Note that the elements of both
sequences are strictly positive integers, so we do not need to add 1 for the γ code. The maximum
space usage for the first sequence is bounded by k(1 + 2 log kℓ ) and for the second by k(1 + 2 log ks )
which sums up to 2k(1 + log ksℓ2 )) (see the proof for Lemma 6 for details on obtaining these bounds).
√
√
with value e4log2sℓ2 where e denotes the base
For fixed s, ℓ this function has a maximum for k = 2sℓ
e
e
of the natural logarithm. As we by definition have k ≤ min{ℓ, s} this maximum can only be observed
for s ≥ e2ℓ2 ≈ .2706ℓ (and equally ℓ ≥ e2s2 ). For s < e2ℓ2 we obtain the maximum space usage for k = s
Faster Average Case Low Memory Semi-External Construction of the BWT
9
and this space maximum is then bounded by 2s(1 + log sℓ ). While computing a complete gap array we
may at an intermediate stage have no more than two arrays with a sum of c2i for i = 0, 1, . . . , ⌊log sc ⌋
when the arrays of the smallest sum we produce have sum c for a constant c ≥ 1. We will never have
two arrays for more then two of the sizes while merging and only a single one for all other sizes, but
for the sake of simplicity in obtaining a bound on the size of the sum of the arrays we will assume
that two for each size are possible. For those arrays with a sum of at most 4ℓ we can bound the space
used by
2
P⌊log 4cl ⌋
i=0
2(2i c)(1 + log 2ℓi c ) ≤
≤
≤
≤
≤
≤
≤
≤
4c
4c
P⌊log 4cl ⌋
i=0
2i (1 + log 2ℓi c )
i=0
2i (1 + log 2ℓi )
P⌊log 4cℓ ⌋
P⌊log 4cℓ ⌋
2i (log 2ℓi )
i=0
P⌊log 4cℓ ⌋ i
2 (log ℓ − i)
8c i=0
P⌈log ℓ⌉ i
8c i=0
2 (⌈log ℓ⌉ − i)
P⌈log ℓ⌉
2⌈log ℓ⌉+3 c i=0 2ii
2⌈log ℓ⌉+4 c
32cℓ
8c
Thus the space for these arrays is O(ℓ) bits as c is a constant. For partial arrays with a sum of more
than 4ℓ we can resort to dense arrays. We may have a constant number of arrays with a sum between 4ℓ
and ℓ. These take space O(ℓ) (see the proof for Lemma 6). Finally we can have a logarithmic number
of arrays with a sum larger then ℓ. Analogously to the arrays with a sum of at most 4ℓ we can deduce
that the space for these arrays is O(s) bits. In total all partial gap arrays together take space O(s + ℓ)
bits. Two sparse partial gap arrays can be merged in time linear in the number of their respective
non-zero elements which is bounded by the sum over the elements. Two of the dense partial gap arrays
we use can also be merged in time linear in the sum over their elements, as these arrays are only used
if the sum over their elements reaches a quarter of there length. Merging two partial gap arrays is in
both cases a streaming operation and requires virtually no internal memory. If the merging process
starts with partial arrays for sum 1, the final array has length ℓ and the sum over its elements is s
P⌈log s⌉
then the total merging time can be bounded by i=0 min{s, ℓ}. For merging two blocks of size bl
and br respectively this is in O(min{bl , br } log max{bl , br }) which becomes O(bl log bl ) if bl ∈ Θ(br ) as
it holds for balanced binary merge trees. If we accumulate logb2r br indices for incrementing in an array
in internal memory we can reduce the time used for merging without asymptotically increasing the
total memory footprint of the algorithm. The space used for the array in internal memory is O( logbrbr )
bits and thus asymptotically negligible compared with the (uncompressed) text. When the internal
memory array is full, then we first need to sort it so we can write the partial gap array representation.
This can be performed in time √linear in the length of the array by using a two phase
√ radix sort
with
a
number
of
buckets
in
O(
b
),
which
require
additional
internal
memory
of
O(
br ) words or
r
√
O( br log br ) bits. For both run-time and space usage the radix sort does not asymptotically increase
the resources required for the merging. Starting with partial gap arrays of sum logb2r br implies the total
number of such arrays is O(log2 br ). Consequently the binary merge tree for obtaining the final gap
array has a depth in O(log log2 br ) = O(log log br ) and it can be processed in time O(br log log br ).
The gt array for the merged block can be composed by concatenating the gt array for the left
block and an array storing the respective information for the right block computed while performing
the backward search for filling the gap array. For this purpose we need to know the rank of the leftmost
suffix in the left block. This can either be computed using forward search on the suffix arrays of the
basic blocks or extracted from a sampled inverse suffix array which can be computed along the way.
The sampled inverse suffix arrays of two blocks can just like the BWTs of the two blocks be merged
using the gap array. This is also an operation based on stream accesses, so it can be done in external
memory in time O(bl + br ) when merging two blocks of size bl and br .
10
German Tischler
7. BWT Computation by Balanced Tree Block Merging
Using the building blocks described above we can now describe the complete algorithm for computing
the BWT of t by merging basic blocks according to a balanced binary tree.
1. Choose a target block size b′ ∈ O( logn n ) and deduce a block size b = ⌈ ⌈ nn ⌉ ⌉ such that the number
b′
of blocks c satisfies c = ⌈ nb ⌉ = ⌈ bn′ ⌉ and n can be split into blocks of size b and b − 1 only. Split t
such that the blocks of size b appear before those of size b′ . This step takes constant space and
time.
2. Compute which blocks in t propagate repetitions of period at most b and for each block which is
followed by a block propagating a repetition whether it is generating this repetition. This takes
log σ
time O(n) in total and space O(b log σ) = O( nlog
n ) ⊆ O(n) bits. The result data can be stored
in external memory.
3. Compute a balanced merge tree for the blocks. Start with a root representing all blocks. If a
node containing a single block is considered produce a leaf and stop. Otherwise for an inner
node representing k > 1 blocks produce a left subtree from the ⌈ k2 ⌉ leftmost blocks and a right
subtree from ⌊ k2 ⌋ rightmost blocks in t. The tree has O(log n) nodes. Each node stores at most
two (start and end) block indices taking O(log log n) bits and two node pointers also taking space
O(log log n) bits. So the total tree takes space O(log n log log n) bits. It can be computed in time
O(log n).
4. Sort the blocks and store the resulting BWT, gt and sampled inverse suffix arrays in external memory. Using the suffix and LCP arrays of the basic blocks also compute the start
ranks necessary for the backward searches when merging the blocks together. This takes time
O(n log n log log n) in the worst case and O(n) on average and space O(b log b) = O( logn n log logn n )
= O(n) bits of internal memory.
5. Process the merge tree. Mark all leafs as finished and all inner nodes as unfinished. While there
are unfinished nodes choose any unfinished node with only finished children, merge the respective
blocks and mark the node as finished. There are O(log n) leafs and the tree is balanced, so it
has O(log log n) levels. Each single level can be processed in time O(n log log n). So the total
run time for the tree merging phase is O(n log2 log n). The maximum internal memory space
usage appears when performing the merge operation at the root of the tree. Here we need space
bl Hk + o(bl log σ) bits where bl denotes the sum of the length of the blocks in the left subtree
which is O(n) and Hk denotes the k’th order entropy of the text comprising those text blocks.
Summing over all steps the run-time of the algorithm is O(n log n log log n) in the worst case and
O(n log2 log n) on average. In practice this means we can compute the BWT of a text as long as we
are able to hold the text (more precisely the text for the left subtree of the merge tree) in internal
memory. If we can hold a fixed fraction of the text in main memory, then we can still compute the
BWT of the text in the same run-time by resorting to the original iterative merging scheme from [8].
log σ
1
We decompose the text into blocks of size b′ such that b′ ≤ nc log
n where c is the fixed fraction of the
text we can hold in internal memory and compute a partial BWT for each of these blocks where the
suffixes sorted are considered as coming from the whole text, i.e. suffix comparisons are still over t̃ and
not limited to a single of the blocks. Then we merge these blocks along a totally skewed merge tree such
that the left block always has size about b′ . The size of the set of partial sparse gap arrays required
at any time remains bounded by O(n) bits. As the number of blocks is fixed, the total asymptotical
run-time of the algorithm remains O(n log n log log n) in the worst case and O(n log2 log n) on average.
8. Parallelisation
Many of the steps of our algorithm are parallelisable. As building blocks we will need representations
of the required data structures in external memory which allow accessing parts without performing a
complete sequential scan and we will require parallel versions of
Faster Average Case Low Memory Semi-External Construction of the BWT
11
• the construction of rank indexes,
• the computation of gap arrays,
• the merging of BWT blocks,
• the merging of sampled inverse suffix arrays and
• the sorting of single blocks.
As the merging of sampled (inverse) suffix arrays is basically a simplified version of the merging of
BWT blocks (the difference is not all values are present), we will only describe the merging of BWT
blocks and leave the sub sampled merging variant as an exercise for the reader. In this paper we
consider shared memory parallelisation only, i.e. a setting where all processors involved have access to
the same internal memory. Some of the algorithms may be modified for distributed memory settings.
8.1. Data Structures in External Memory
We are using the following data structures in external memory:
• The gt bit vectors
• Burrows Wheeler transforms of blocks
• Dense γ coded gap arrays
• Sparse γ coded gap arrays
Arbitrary positions in the gt bit vectors can be accessed without further information, so no additional
information is necessary to support scanning starting from a given position.
For the Burrows Wheeler transform sequences we choose the following representation. For a
given string B = b0 . . . bm−1 over Σ and a given block size d ∈ O(log2 n) we partition B into blocks
of size d such that all but the last block have size d. Each block factor is stored using run-length,
Huffman and γ code. We first compute a run length representation of the block which gives us a
sequence of pairs comprised of symbols and number of consecutive occurrences. As an example the
factor aabbbcaaa would be transformed to (a, 2), (b, 3), (c, 1), (a, 3). This sequence of pairs is stored
using Huffman code for the symbols and Elias γ code for the run lengths. As we assume a constant
alphabet size, the dictionary for the Huffman code takes constant space per block. The space used
by the Huffman coded sequence of symbols is bounded by d⌈log σ⌉. According to our considerations
about γ code above the maximum amount of space required for storing any sequence of run lengths
for a block of length d is d(1 + 2⌈log(1 + 1)⌉) = 3d. So in the worst case the representation of the block
BWT in this code is in O(d log σ) bits. In consequence each block can be addressed using a fixed size
pointer of length in O(log(m log σ)). As we have d ∈ O(log2 n) and are only storing BWTs of length
m
m ≤ n we are storing at most O( logm2 n ) pointers. In total these require space O( log
n ) bits. When
the BWT is coded in this way we can start decoding it at any index which is a multiple of the block
size d ∈ O(log2 n). The setup procedure takes O(1) time and we can decode each following symbol in
time O(1) when performing a sequential scan (see [20, 15] for constant time encoding and decoding of
Huffman/minimum redundancy code. The code lengths of the γ code can be decoded using a lookup
table taking o(n) bits).
Dense γ coded gap arrays can be stored very similar to our BWT storage scheme for allowing
the start of a decoding run from any position in the array with very little overhead. As shown above
Pl−1
an array G of length l such that i=0 G[i] = s can be stored using O(s + l) bits, which means the
start index for the code of any position can be stored in O(log(s + l)) bits. We decompose the array
into non-overlapping blocks of length e ∈ O(log2 n) elements and store a pointer to the start of each
s+l
block. These pointers take space O( log
n ) bits in total and we can start the decoding from any index
2
on G after a setup time in O(log n). Subsequently each following element can be sequentially decoded
in constant time per element.
For the sparse γ coded gap files we have already described the code itself above. Let us assume
we are considering a gap array G of length ℓ such that the sum over the elements of G is s and G
contains k non-zero values. To facilitate the start of a decoding run at a given index on G we keep
the γ code as is, where we assume that the two γ coded sequences (distance between non-zero values
in the array and non-zero values) are stored interleaved. As the array is sparse we possibly would be
12
German Tischler
adding excessive amounts of space by adding pointers marking the next non-zero value for equidistant
indices on G. Instead we store information for every j = ⌈log n⌉2 ’th instance of a non-zero value on
G for i = 0, 1, ⌊ kj ⌋. For each such instance of a non-zero value we store the bit position of the code
pair in the γ code and the absolute index on G. The γ code itself requires O(k(1 + log ksℓ2 )) bits as
shown above, so each pointer into the code takes O(log(k log n)) bits and there are O( logk2 n ) such
log n)
pointers, so the space used for pointers is O k log(k
bits which is bounded by the space used
log2 n
for the sparse array itself. Storing the selected absolute values of non-zero indices on G takes space
O( logk n ). Using this index on the external memory representation for G we can start decoding the
array from any index on i on G by first employing a binary search on the absolute value of the largest
indexed non-zero value element in G with an index not greater than i. This element may be O(log2 n)
non-zero elements away from our desired target which need to be skipped before we reach the relevant
portion of G. Thus the setup time is O(log2 n). After the setup we can decode either the sequence
of non-zero values starting from i or the sequence of all values starting from i in constant time for
each such value. Decoding non-zero values is required for merging two sparse arrays while decoding
all values is required for merging two BWTs.
For parallel encoding of these data structures we note that all of them are easily writable in
any given number of parts to a set of files such that a simple combination of these files represents
the respective complete structure. Thus encoding the data structures does not need be performed
sequentially. If we write data to multiple files, then we need a simple additional layer of indexing meta
data designating the length of the corresponding sequence in each file. For gt bit vector, Burrows
Wheeler transform and dense γ coded gap files this additional data stores the number of bits, symbols
and gap array entries respectively in the form of a prefix sum array. This array allows us to find the
correct starting file and offset in this file given the offset in the complete sequence in logarithmic time
in the number of files. The number of files for any one such sequence is bounded by the number of
processors, so the space for the prefix sum array is not a concern as long as the number of processors
p is O( logn n ). For sparse γ coded gap files the situation is slightly more involved, as we want to be
able to start processing from a given index i on the gap file to process the entry at index i and all gap
file entries following i or given some j skip j non-zero values from the front of the file and process all
following non-zero elements. For this purpose we store two index sequences in addition. The first is a
prefix sum sequence over the number of gap array elements represented by each file. Here the number
stored in a file is obtained as
• the length of the gap array if there is only one file or
• otherwise, if there is more than one file
– the index of the last non-zero gap array value represented in the file plus one for the first
file,
– the index of the last non-zero gap array value represented in the current file minus the index
of the last non-zero gap array value represented in the previous file if the file is neither the
first nor the last and
– the length of the full gap array minus the index of the last non-zero gap array value represented in the penultimate file for the last file.
The second is the prefix sum sequence over the sequence of non-zero values stored in each file. Both
additional index sequences store O(p) values. Again the space for these can be neglected when p ∈
O( logn n ). Searching the file level indexes for the correct file and in file offset for a given global offset
in each case takes time O(log p) and is thus not critical in comparison with the in file access times
for BWT and gap files. We will thus without loss of generality below discuss a situation were the
respective streams are assumed to be given as a single file instead of a set of files for the sake of
simplicity of exposition.
Faster Average Case Low Memory Semi-External Construction of the BWT
13
8.2. Parallel Construction of Rank Indexes
There are several data structures which allow constant time rank queries on sequences over a finite
alphabet. In this paper we will consider the wavelet tree (see [10]). A prefix free code g for an alphabet
Γ is a function g : Γ 7→ {0, 1}∗ such that for all pairs of symbols (a, b) ∈ Γ2 such that a 6= b the code
g(a) is not a prefix of g(b). In the following we in addition assume that each prefix free code considered
is such that each binary string c either has a prefix which is a code of some symbol or c is a prefix
of at least one code produced by the code. This condition implies that there are no unused binary
code words. In consequence the longest code assigned to any symbol is bounded by the size of the
input alphabet. For any prefix free code g : Γ 7→ {0, 1}∗ where |Γ| > 1 let gi : Γi 7→ {0, 1}∗ denote
the sub code obtained from g by choosing Γi = {a ∈ Γ | g(a)[0] = i} and gi (a) = g(a)[1, |g(a) − 1|].
Informally gi is defined for all symbols in Γ for which the first bit of the code assigned by g is i and
this first bit is stripped off when transforming g to gi . Let fi : Γ∗ 7→ Γ∗i be the function given by the
homomorphism defined by mapping each symbol in Γi to itself and all other symbols to the empty
word, i.e. the function removing all symbols which are not in Γi . Let h : Γ 7→ {0, 1}∗ denote a prefix
free code. Then the wavelet tree concerning h for a text s ∈ Γ∗ is obtained in the following recursive
way.
• If |Γ| = 1 then produce a leaf.
• Otherwise store the bit sequence h(s0 )[0]h(s1 )[0] . . . h(s|s|−1 )[0] in an inner node. Attach a left
and right child to this inner node where the left child is the wavelet tree for f0 (s) concerning the
code h0 and the right child is the wavelet tree for f1 (s) concerning the code h1 .
When this procedure is complete, then we assign a binary code to each inner node, where we assign
ǫ to the root, c0 to each inner node which is a left child of a parent with code c and c1 to each inner
node which is a right child of a parent with code c. For the parallel construction of a wavelet tree
based on a prefix free code h from an input string observe that each interval of indices on the input
string corresponds to a unique interval of indices in the bit vector stored in any inner node of the
wavelet tree. For constructing a wavelet tree in parallel from an input string s of length m over Σ
using p processors we use the following steps.
⌋.
1. Choose a block size d = ⌊ m+p−1
p
2. Assign index block Ii = [di, min(di + d, m)) to processor i.
3. Each processor computes a symbol histogram Hi : Σ 7→ N for its interval such that Hi [a] = |{j |
j ∈ Ii and s[j] = a}|.
Pi−1
4. Compute the prefix sum arrays Ĥi defined by Ĥi [a] = j=0 Hi [a] for all a ∈ Σ. This can be
parallelised along the alphabet Σ.
5. Let C(h) = {h(a) | a ∈ Σ} and let Ĉ(h) = {c′ | c′ = c[0, i] for some c ∈ C(h) and − 1 ≤ i <
|c| − 1} (the set of all proper prefixes of codes produced by h). Let Pi : Ĉ 7→ N be defined by
P
Pi (j) = k|j prefix of h(k) Ĥ(k). Processor i computes table Pi . Note that P may be understood as
6.
7.
This
time
a function from integers to integers instead of binary strings to integers, as the elements of Ĉ may
be interpreted as the numerical value represented by the respective binary strings. Pi denotes
the bit offset of the input interval for input block i for each inner node of the constructed wavelet
tree.
Processor i produces the wavelet tree for interval Ii while instead of starting at the beginning of
the bit vector of a node starting at the bit position designated by the array Pi .
For each node of the wavelet tree the p processors compute the necessary prefix sums for the
binary rank dictionaries of each node in parallel.
approach
can be used to obtain a rank dictionary for the input string s of length m over Σ in
using p processors. This applies for block type as well as Huffman code.
O m
p
8.3. Parallel Computation of Gap Arrays
When we compute a gap array we are generally merging two adjacent text blocks. Let the left and right
block sizes be bl and br respectively and let the indices of the left block on t̃ start at index i. In the
14
German Tischler
serial version we start with the suffix at index i+bl +br −1 and perform br −1 steps of backward search
using the index for the left block. This can be parallelised by performing several backward searches
in parallel starting from several indices within the right block. If p ≤ br processors are available, then
we can choose the starting positions xj = i + bl + br − 1 − j⌈ bpr ⌉ for j = 0, 1, . . . , p − 1 and perform
ℓj = xj+1 − xj steps for j = 0, 1, p − 2 and ℓj = br − xj steps for j = p − 1. As the backward search
is based an suffix ranks in the left block we need to find the ranks of these p suffixes in the left block
using forward search on one or more suffix arrays just as we did above for the single suffix at the end of
the right block in serial mode. As we have p processors and need to find p ranks, the run-time stays the
same as in the serial version for both average and worst case. In the average case the forward searches
n
), so in practice we can expect the forward
are not a time critical factor as long as p is O( log2 n log
log n
searches to be of no concern for the execution time. When we perform p backward searches in parallel,
then we also need to be prepared to perform the corresponding increment operations on the gap array
in parallel. This is no problem if the gap array can be kept in internal memory as modern processors
have specialised operations for atomic, lock free modifications of single memory cells. If we use sparse
gap arrays in external memory, then the situation is somewhat more complicated. In the serial version
we first accumulate a number of indices in the gap array whose value is to be incremented in a list of
length k ∈ O( logb2r br ) in internal memory. If this list runs full, then it needs to be sorted and written to
external memory as a sparse γ coded gap array. Whenever there are two sparse gap arrays of the same
sum, then the two need to be merged to obtain a single sparse array of twice that sum (as the number
of non-zero elements in the files approaches the size of the left block we switch to dense arrays as
described above. Generating and merging dense γ coded gap arrays are however simplified versions of
the equivalent problems in sparse arrays, so we will not explicitly discuss this below as the algorithms
are easily deduced from the sparse case). In the parallel case we keep the concept of a single array
accumulating indices in internal memory. All p processors are appending values to this array until
it runs full. This can be implemented by using a fill pointer which is modified by atomic increments
while elements are inserted at the array index designated by the fill pointer before the increment. A
process stalls backward search when the pointer exceeds a given threshold signalling that the array
is full. When all processes have detected that the array is full, then we can perform a parallel
radix
√
sort on the array. As in the serial case this is performed in two stages, i.e. using e ∈ O( n) buckets.
Like for histogram computation in the parallel wavelet tree construction we partition the array into
p non overlapping intervals and compute a separate histogram
for each interval. This takes space
√
√
n
O(p n log n) bits which is negligible as long as p ∈ o( log n ) which is not an issue in practice. Prefix
sums are computed based on these bucket histograms to obtain the starting index for the elements for
each bucket originating from each processor input interval. The elements for bucket i appear ahead
of those for buckets with index larger than i and within each bucket the elements originating from
processor interval j appear before those for intervals with higher indices. These prefix sums can be
computed in parallel by first accumulating all element counts over ⌈ pe ⌉ buckets and then based on these
offsets filling in the final prefix sums in parallel. Using the computed prefix sums the numbers can be
sorted into their buckets in parallel while keeping the stable order necessary for a correct
sort.
radix
√
Using this radix sort method we can sort a set of the k numbers from [0, bl ] in time O max{ p bl ,k}
using p processors. After the array is sorted we again partition it into p index intervals. In each interval
in parallel we transform the now sorted sequence into run length encoding, i.e. pairs of numbers and
the number of occurrences of the number in the interval. This representation does not asymptotically
need more space than the sorted number sequence. The computation of the run length encoding in
intervals produces a correct run length encoding of the complete sequence except for the borders of the
intervals. Here the last run length in one interval may be for the same number as the first run length in
the next interval. This however can be corrected in time O(log p) using p processors, where we assign
the complete run-length to the leftmost occurrence of the number and mark the other occurrences as
invalid. After this we can produce a compact version without space between the intervals of the run
length encoded array in internal memory in time O( kp ) by counting the number of runs per interval,
Faster Average Case Low Memory Semi-External Construction of the BWT
15
computing the prefix sums over this number sequence and copying the sequence to another array in
compact form. Finally we decompose this compact run length array into at most p intervals of length
⌈k⌉
⌈ ep ⌉e where e ∈ O(log2 n) is a positive integer and write the resulting decomposition to p sparse γ
coded gap arrays such that the concatenation of the files represents the complete sparse gap array.
As each interval size but the last is a multiple of e each file starts on an index of the sequence of run
length pairs which is also a multiple of e. This file production also takes time O( kp ). Finally we can
also produce the block pointers required for accessing the file without performing a linear scan in the
same time complexity. Summarising we can transform the array of length k accumulating indices√ for
incrementing in the gap array to a sparse γ coded representation in external memory in time O( k+p br )
using p processors. Now assume we have two sparse γ coded gap arrays with ka and kb non-zero values
respectively which are to be merged. We want to perform this merging using p processors, where in
the optimal case each processor is assigned the same amount of work. Both files store e ∈ O(log2 n)
non-zero values per block. Using a binary search on these values we can determine the smallest i such
b
⌉ blocks are stored in the two files for values smaller than i. The first processor
that at least ⌈ ka +k
p
merges the values of the two input sequences from the beginning up to value i. The same scheme is
used to split the remaining data into a work package for the second processor and the rest for the
remaining processors etc. The splitting operations take time O(p log3 n). We are computing p splitting
points, each such splitting point is computed using a binary search on O(n) sorted values and accessing
each such value takes time O(log2 n). After the splitting the merging is easily done in parallel in time
3
ka +kb
b
O( ka +k
p ) so the run time of the method is O(p log n +
p ). This is a convenient procedure when
the number of processors is low compared to the size of the input. If the number of processors should
not be neglectable in comparison to the input size, i.e. if O(p log3 n) would be large, then we could
employ a different splitting method where we would decompose the merging work along an optimally
balanced binary tree and thus reduce the work for finding the splitting points to O(log p log3 n). We
would still be computing O(p) splitting points, but the splitting point search within each level of the
tree of depth O(log p) could be performed in parallel in time O(log 3 n).
8.4. Merging Burrows Wheeler Transforms
For merging two BWT sequences of length bl and br respectively let us assume we have the two
sequences in external memory and the gap array in internal or external memory. For parallel merging
of the two sequences we first compute the sum over the gap array for each interval of integer length
e ∈ O(log2 n) starting at index ie on the gap array G for i = 0, 1, . . . , bel using p processors in parallel
in time O( bpl ). These can be produced sequentially and in external memory and take space o(bl ) bits.
Then we compute the prefix sums over this sequence in parallel in time O( bpl ) and also store it in
external memory. Using these prefix sums we can assign approximately equal work loads to each
processor. Using a binary search on the prefix sums we can search for the smallest j < bl + 1 such
Pj
bl +br
that s =
i=0 (G[i] + 1) ≥ ⌈ p ⌉ and then increase j and update the sum s accordingly until
s
⌈ e ⌉e − s ≤ G[j + 1] or j = bl + 1 if no such j exists (remember that G has length bl + 1). The
binary search gives us a work package which is approximately 1p of the total length. The subsequent
increasing of j is for allowing us to produce a file with a length which is a multiple of the block size
by just considering the next gap array value G[j + 1]. The case where there is no such j can only
occur for very sparse arrays with high run length values. The first processor merges the front parts
of the two BWTs by in turn taking G[i] symbols from the second BWT stream and one symbol from
the first one for i = 0, 1, . . . , j and finally ⌈ es ⌉e − s from the second BWT stream to obtain a multiple
of the block size e. The rest of the two BWT streams is merged by the other processors. The further
splitting of the remaining data into blocks is performed just like the splitting of the data into the part
for the first processor and the rest. The only difference consists in the fact that a fraction of G[j + 1]
is already handled by the first processor and thus needs to be excluded from the remaining process.
As for merging sparse γ coded gap arrays the finding of the splitting points in this manner takes time
r
O(p log3 n) following a similar argumentation. The actual merging is then performed in time O( bl +b
p )
16
German Tischler
using p processors. In practice it may be desirable to perform the merging directly using the run length
encoding instead of symbol per symbol. In the asymptotical worst case there is however no difference
in the run time. Like in the case of merging sparse γ coded gap files we can speed up the computation
of the splitting points to O(log p log3 n) if necessary. The approach of parallelisation of BWT merging
can be extended to multi way instead of two way merging as described in [12]. The concepts remain
the same, but the searching of split points is over all of the involved prefix sum arrays of gap files
instead of just one. The merging within the single merge packets remains the same as in the serial
version.
8.5. Parallel Sorting of Single Blocks
We need to suffix sort single blocks and produce the respective block BWTs before we can merge these
structures to obtain the final BWT of the complete text. To our best knowledge there is currently
no parallel in place suffix sorting algorithm. The external memory algorithmic variants described in
[14] either distribute a total work of Ω(n log n) over p processors or work in some randomised model
of computation. As the problem of parallel suffix sorting in place is still open, we resort to using our
parallel block merging scheme to achieve some parallelism for the average case. In the highly unlikely
worst case this will again not bring any improvement over the serial version. The block size b in
the serial version is chosen to allow the suffix sorting of a single block in internal memory. For the
parallel case we deduce a reduced block size bp = ⌈ pb ⌉ which allows us to perform the suffix sorting of
p such blocks at the same time in internal memory. After these blocks have been handled we merge
the resulting BWTs until we obtain the BWTs for block size b using our parallel merging approach
described above. Note that this does not require us to store gap arrays in external memory, as we have
sufficient space in internal memory to perform the merging. In the unlikely worst case we will spend
time Ω(n log n log p) for finding the start ranks of our backward searches during gap array construction.
p
n
On average the run time is O( b log
p ) in all practically relevant cases (i.e. when p log p ∈ O( log4 n )).
8.6. Parallelisation of the Complete Algorithm
Using the building blocks described above we can now present the parallel version of our algorithm
for p processors. The steps are:
1. Choose a target block size b′′ ∈ O( logn n ) as in the serial version. Transform this block size b′′ to
′′
b′ = ⌈ bp ⌉, a target block size for parallel computing and finally deduce a block size b = ⌈ ⌈ nn ⌉ ⌉.
b′
Split the text t into blocks of size b and b − 1 only such that all the blocks of length b appear
before the ones for b − 1. This step takes constant time and space.
2. Compute which blocks in t̃ propagate repetitions of a period at most b and for each block which
is followed by a block propagating a repetition whether it is generating this repetition. We can
handle p blocks in parallel during the propagation as well as the generation checking phase as
each block can be handled independently and we have sufficient space in internal memory due to
the reduced block size compared to the serial version. This step consequently takes time O( np ) in
log σ
n log σ
total and space O(p nlog
np ) = O( log n ) ⊆ o(n) bits. The result can be stored in external memory.
3. Compute a balanced merge tree for the blocks in serial like for the serial cases. This takes time
O(log n) and space o(n) bits.
4. Sort the blocks and store the resulting BWT, gt and sampled inverse suffix arrays in external
memory while processing p blocks at a time. Using the suffix and LCP arrays of the basic blocks
also compute the start ranks necessary for the backward searches when merging blocks together.
This step takes time O(n log n(log log n+ log p)) in the worst case and O( np ) on average and space
O(b′ log b′ ) = O(n) bits of internal memory.
5. Process the merge tree. Mark all leafs as finished and all inner nodes as unfinished. While
there are unfinished nodes choose any unfinished node with only finished children, merge the
respective blocks and mark the node as finished. When merging blocks use an internal memory
gap array if the node has a distance of at most ⌈log p⌉ from any leaf in the underlying sub tree
and sparse γ coded gap arrays in external memory otherwise. Use the parallel versions of gap
Faster Average Case Low Memory Semi-External Construction of the BWT
17
array computation, merging of γ coded gap array files, merging of BWT streams and merging
sampled inverse suffix array files. The log log n levels closest to the root can be processed in
2
total time O( n logplog n ) per level or O( n log p log n ) in total. The log p other (lowest) levels of
p
the tree can be processed in time O( n log
p ). The total run-time for the merging stage is thus
O( np max{log2 log n, log p}). The space requirement in internal memory remains asymptotically
√
the same as in the serial version as long as p ∈ o lognn as described above.
The total run-time of the parallel algorithm using p processors is thus O( np max{log2 log n, log p}) on
average and with high probability and O(n log n(log log n + log p)) in the worst case. As in the serial
case we can resort to the original skewed merging approach on top of the balanced merging if there is
not sufficient internal memory to hold the index for the left half of the text in memory but sufficient
to keep an index for a fixed fraction of the text while keeping the same run-time and space bounds.
9. Conclusion
We have presented a new semi external algorithm for computing the Burrows-Wheeler transform
designed for the case where we can keep a fixed fraction of the input text in internal memory. On
average our new algorithm runs in time O(n log2 log n) and is faster then the algorithm of Ferragina
et al published in [8] while in the worst case it is only slower by a factor of O(log log n). In comparison
with the algorithm by Beller et in [1] our algorithm can be applied for the case when less than 8 bits
per symbol of internal memory are available. We have also presented a parallel version of our algorithm
which on average and with high probability runs in time O( np (log2 log n + log p)) on p processors and
should achieve this run time in all practically relevant applications. In the worst case the run time of
the parallel algorithm is O(n log n(log log n + log p)) and thus, depending on the number of processors
p, possibly even slower than the serial version. However this worst case is highly unlikely in practice.
The practically interested reader can find a partial implementation of the algorithms presented
in this paper at https://github.com/gt1/bwtb3m. The current version of the code does not implement
the repetition reduction as described in Section 5 but uses a simplified approach computing the length
of the longest common prefix of all suffixes in a block with the first suffix of the following block for
determining how far a block needs to be extended to obtain a correct sorting order. On average this
extension is by O(log n) characters. For texts featuring long repetitions of short periods (shorter than
the block size) however the method would be problematic.
One major open problem in our approach is the worst case run time. While it is not as extreme as
the worst case of some other algorithms for suffix sorting or BWT construction showing good average
case behaviour (cf. [24]), it still leaves room for improvement. This is particularly true for the case of
the parallel variant of the algorithm. The increased worst case run-time over the average case stems
solely from the forward searches on block suffix arrays to obtain the starting points for the backwards
searches during gap array construction. In the serial version a string needs to feature a very large LCP
log n
value in Ω( n log
log n ) between two of its suffixes to trigger the worst case behaviour, in the parallel
case using p processors a LCP value in Ω( n plogloglogn n ) is sufficient. So an efficient method for detecting
such high LCP values using small space would be necessary to alleviate the effects of the problem.
Note that these high LCP values are not necessarily caused by runs (cf. [18]) as the positions of the
suffixes in the string sharing a long common prefix may be too far apart to allow the common strings
to connect, i.e. build a square.
Two other open problems are the actual practical space usage of our algorithm in external memory and its I/O complexity. While the space usage in external memory is asymptotically clear with
O(n log σ) bits the involved constants are important in practice. Punctual observations in our implementation studies suggest that the involved constants are low, but a systematic analysis is desirable.
A very simple analysis in analogy to the space usage shows that the algorithm uses O(n log2 log n)
bits of I/O for merging γ coded gap arrays, O(n log σ log log n) bits of I/O for merging BWTs and
18
German Tischler
O(n log log n) bits of I/O for writing and reading gt arrays, where I/O is in all cases streaming. As for
the space usage in external memory, it would also be interesting to determine the involved constants
in future work.
References
[1] T. Beller, M. Zwerger, S. Gog, and E. Ohlebusch. Space-Efficient Construction of the Burrows-Wheeler
Transform. In O. Kurland, M. Lewenstein, and E. Porat, editors, SPIRE, volume 8214 of Lecture Notes
in Computer Science, pages 5–16. Springer, 2013.
[2] T. Bingmann, J. Fischer, and V. Osipov. Inducing Suffix and LCP Arrays in External Memory. In
P. Sanders and N. Zeh, editors, ALENEX, pages 88–102. SIAM, 2013.
[3] M. Burrows and D. Wheeler. A Block-Sorting Lossless Data Compression Algorithm. Digital Systems
Research Center. RR-124, 1994.
[4] M. Crochemore, R. Grossi, J. Kärkkäinen, and G. M. Landau. A Constant-Space Comparison-Based
Algorithm for Computing the Burrows-Wheeler Transform. In J. Fischer and P. Sanders, editors, CPM,
volume 7922 of Lecture Notes in Computer Science, pages 74–82. Springer, 2013.
[5] M. Crochemore, C. Hancart, and T. Lecroq. Algorithms on Strings. Cambridge University Press, 2007.
[6] R. Dementiev, J. Kärkkäinen, J. Mehnert, and P. Sanders. Better external memory suffix array construction. ACM Journal of Experimental Algorithmics, 12, 2008.
[7] P. Elias. Universal codeword sets and representations of the integers. Information Theory, IEEE Transactions on, 21(2):194–203, 1975.
[8] P. Ferragina, T. Gagie, and G. Manzini. Lightweight Data Indexing and Compression in External Memory.
Algorithmica, 63(3):707–730, 2012.
[9] P. Ferragina and G. Manzini. Opportunistic Data Structures with Applications. In Foundations of Computer Science, 2000. Proceedings. 41st Annual Symposium on, pages 390–398. IEEE, 2000.
[10] R. Grossi, A. Gupta, and J. S. Vitter. High-order entropy-compressed text indexes. In SODA, pages
841–850, 2003.
[11] W.-K. Hon, K. Sadakane, and W.-K. Sung. Breaking a Time-and-Space Barrier in Constructing Full-Text
Indices. In FOCS, pages 251–260. IEEE Computer Society, 2003.
[12] J. Kärkkäinen and D. Kempa. Engineering a lightweight external memory suffix array construction algorithm. In C. S. Iliopoulos and A. Langiu, editors, 2nd International Conference on Algorithms for Big
Data (ICABD2014), number 1146 in CEUR-WS Proceedings, pages 53–60, Aachen, 2014.
[13] J. Kärkkäinen and P. Sanders. Simple Linear Work Suffix Array Construction. In Automata, Languages
and Programming, pages 943–955. Springer, 2003.
[14] J. Kärkkäinen, P. Sanders, and S. Burkhardt. Linear work suffix array construction. J. ACM, 53(6):918–
936, Nov. 2006.
[15] J. Kärkkäinen and G. Tischler. Near in place linear time minimum redundancy coding. In A. Bilgin,
M. W. Marcellin, J. Serra-Sagristà, and J. A. Storer, editors, DCC, pages 411–420. IEEE, 2013.
[16] T. Kasai, G. Lee, H. Arimura, S. Arikawa, and K. Park. Linear-time longest-common-prefix computation
in suffix arrays and its applications. In A. Amir and G. M. Landau, editors, CPM, volume 2089 of Lecture
Notes in Computer Science, pages 181–192. Springer, 2001.
[17] D. E. Knuth, J. H. Morris, Jr, and V. R. Pratt. Fast Pattern Matching in Strings. SIAM Journal on
Computing, 6(2):323–350, 1977.
[18] R. Kolpakov and G. Kucherov. Finding maximal repetitions in a word in linear time. In Proceedings of
the 1999 Symposium on Foundations of Computer Science (FOCS’99), New York (USA), pages 596–604,
New-York, October 17-19 1999. IEEE Computer Society.
[19] U. Manber and G. Myers. Suffix Arrays: a New Method for On-line String Searches. SIAM Journal on
Computing, 22(5):935–948, 1993.
[20] A. Moffat and A. Turpin. On the implementation of minimum redundancy prefix codes. IEEE Transactions
on Communications, 45(10):1200–1207, 1997.
[21] G. Navarro and V. Mäkinen. Compressed Full-Text Indexes. ACM Computing Surveys (CSUR), 39(1):2,
2007.
Faster Average Case Low Memory Semi-External Construction of the BWT
19
[22] G. Nong, S. Zhang, and W. H. Chan. Two Efficient Algorithms for Linear Time Suffix Array Construction.
Computers, IEEE Transactions on, 60(10):1471–1484, 2011.
[23] D. Okanohara and K. Sadakane. A Linear-Time Burrows-Wheeler Transform Using Induced Sorting. In
J. Karlgren, J. Tarhio, and H. Hyyrö, editors, SPIRE, volume 5721 of Lecture Notes in Computer Science,
pages 90–101. Springer, 2009.
[24] S. J. Puglisi, W. F. Smyth, and A. H. Turpin. A taxonomy of suffix array construction algorithms. ACM
Comput. Surv., 39(2), July 2007.
[25] W. Szpankowski. On the Height of Digital Trees and Related Problems. Algorithmica, 6(1-6):256–277,
1991.
20
German Tischler
The Wellcome Trust Sanger Institute
Wellcome Trust Genome Campus
Hinxton, Cambridge
CB10 1SA
United Kingdom
e-mail: [email protected]
German Tischler
| 8cs.DS
|
arXiv: 1706.00679
Testing Gaussian Process with Applications to
Super-Resolution
arXiv:1706.00679v2 [math.ST] 14 Nov 2017
Jean-Marc Azaïs• , Yohann De Castro?∗ and Stéphane Mourareau◦
•
Institut de Mathématiques de Toulouse
Université Paul Sabatier, 118 route de Narbonne, 31062 Toulouse, France
?
Laboratoire de Mathématiques d’Orsay
Univ. Paris-Sud, CNRS, Université Paris-Saclay, 91405 Orsay, France
∗
Équipe MOKAPLAN
INRIA, 2 rue Simone Iff, 75012 Paris, France
◦
Laboratoire d’Analyse et de Mathématiques Appliquées
Univ. Paris-Est, Champs sur Marne, 77454 Marne La Vallée, France
Abstract: This article introduces exact testing procedures on the mean of a Gaussian
process X derived from the outcomes of `1 -minimization over the space of complex valued measures. The process X can thought as the sum of two terms: first, the convolution
between some kernel and a target atomic measure (mean of the process); second, a random perturbation by an additive (centered) Gaussian process. The first testing procedure
considered is based on a dense sequence of grids on the index set of X and we establish
that it converges (as the grid step tends to zero) to a randomized testing procedure: the
decision of the test depends on the observation X and also on an independent random
variable. The second testing procedure is based on the maxima and the Hessian of X in
a grid-less manner. We show that both testing procedures can be performed when the
variance is unknown (and the correlation function of X is known). These testing procedures can be used for the problem of deconvolution over the space of complex valued
measures, and applications in frame of the Super-Resolution theory are presented. As a
byproduct, numerical investigations may demonstrate that our grid-less method is more
powerful (it detects sparse alternatives) than tests based on very thin grids.
MSC 2010 subject classifications: Primary 62E15, 62F03, 60G15, 62H10, 62H15;
secondary 60E05, 60G10, 62J05, 94A08.
Keywords and phrases: Hypothesis Testing, Gaussian Process, Kac-Rice formula,
Super-Resolution.
Preprint version of January 8, 2018
1. Introduction
1.1. Grid-less spike detection through the “continuous” LARS
New testing procedures based on the outcomes of `1 minimization methods have attracted a
lot of attention in the statistical community. Of particular interest is the so-called “Spacing
test ”, that we referred to as S ST , based on the Least-Angle Regression Selection (LARS),
that measures the significance of the addition of a new active variable along the LARS path,
see [16, Chapter 6] for further details. Specifically, one is testing the relative distance between
consecutive “knots ” of the LARS, for instance λ1,P and λ2,P . The first knot λ1,P is the maximal
correlation between a response variable and P predictors. The second knot λ2,P is then the
correlation between some residuals and P − 1 predictors, and so on. This approach is now well
1
Azaïs, De Castro and Mourareau/1
INTRODUCTION
2
referenced among the regularized methods of high-dimensional statistics and it can be linked
to minimizing the `1 -norm over P coordinates, see for instance [16, Chapter 6].
In this paper, we focus on `1 -minimization over the space of signed measures and we ask for
testing procedures based on these solutions. Indeed, in deconvolution problems over the space
of measures [7]—e.g., Super-Resolution or line spectral estimation [8, 14, 12, 20, 11, 3]—one
may observe a noisy version of a convolution of a target discrete measure by some known
kernel and one may be willing to infer on the discrete measure. In this case, testing a particular
measure is encompassed by testing the mean of some “correlation” process, see Section 5 for
further details. In particular, remark that there is an uncountable number of predictors—while
there were P predictors previously when inferring on vectors of RP —since we are looking at
correlations between a response variable and vectors indexed by a continuum, say [0, 2π). For
instance, assume that one observes some moments (or Fourier coefficients) of some discrete
measure on the torus [0, 2π) and one is willing to infer on its support. A strategy would be
to look at correlations between the response variable and the moment curve (1, t, t2 , . . . , tn )
(or (cos(kt) ± ı sin(kt))). In this framework, the LARS algorithm does not return a sequence
of entries (among P possible coordinates) and phases as in high-dimensional statistics but
rather a sequence of locations (among the continuum [0, 2π)) and phases. In this paper, we
invoke the LARS to this framework—we referred to it as “continuous” LARS—for which an
uncountable number of active variables may enter the model. We present this extension in
Section 6 defining consecutive knots (λ1 , λ2 ). One can wonder:
• Can the Spacing test be used in the frame of Super-Resolution?
• Is there a grid-less procedure more powerful, in the sense of detecting spikes, than the
Spacing tests constructed on thin grids?
1.5
1
0.5
0
-0.5
-1
-1.5
thin grids
LARS from the process X(·)
(
1,
2 , R) from the Hessian R and
S Rice
Grid-less approach
LARS from P grid points
X(·)
(
1,P ,
testing procedures
2,P ) from
(X(tp ))P
p=1
S ST
Grid approach
Figure 1. The grid-less approach uses the Hessian and the first two “knots” (λ1 , λ2 ) of the “continuous” LARS
to build the test statistics S Rice . We compare it to the grid approach that builds a test statistics S ST using the
knots (λ1,P , λ2,P ) computed from a P points grid discretization (X(tp ))P
p=1 of the continuous process X.
Interestingly, as we will prove, the answer is no to the first question if no modifications of the
test statistic is done. Furthermore, the way that the Spacing test can be fixed to be extended to
a “grid-less” frame gives a new testing procedure S Rice that accounts for the distance between
3
1 INTRODUCTION
consecutive knots (λ1 , λ2 ) with respect to value of the Hessian at some maximal point, see
Figure 1 for a global view on our approach.
1.2. A comparative study
When the predictors are normalized, the Spacing test (ST) statistics is given by the expression
S ST (λ1,P , λ2,P ) :=
Φ(λ1,P )
Φ(λ2,P )
where Φ = 1 − Φ is the Gaussian survival function and Φ the standard normal cumulative
distribution function. In the framework of high-dimensional statistics, this statistics is exactly
distributed w.r.t. a uniform law on [0, 1] under the global null, namely S ST can be considered
as the observed significance [22, 4]. It is clear that one should not use this testing procedure
in the Super-Resolution framework since there is no theoretical guarantees in this case. Yet
the practitioner may be tempted to replace (λ1,P , λ2,P ) by (λ1 , λ2 ) given by the “continuous”
LARS. Unfortunately, this paper shows that the resulting test statistics S ST is non conservative
in this frame, i.e., it makes too many false rejections and one should avoid using it in practice,
see the green line in Figure 2.
1
1
1
0.9
0.9
0.9
0.8
0.8
0.8
0.7
0.7
0.7
0.6
0.6
0.6
0.5
0.5
0.5
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0
0
0
0.2
0.4
0.6
0.8
1
0
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
Figure 2. [Under the null] Comparison of the empirical cumulative distribution of the two statistics S Rice
(blue line, see Theorem 6) and S ST (green line) under the null hypothesis when applied to the consecutive
knots (λ1 , λ2 ) given by the “continuous” LARS in both cases. The diagonal (cdf of the uniform) is represented
in dotted black line. The model is described by the Super-Resolution framework (see Section 5) with cutoff
frequencies fc = 3, 5, 7 from left to right. The new test statistic S Rice is exactly distributed w.r.t. the uniform
law on [0, 1] under the null hypothesis.
To overcome this disappointing feature, one may be willing to consider thinner and thinner
grids and look at the limit as P tends to infinity. In this case, one can show that λ1,P tends to
the λ1 of “continuous” LARS, but λ2,P does not converge to λ2 , it converges to λ2 as shown
in (6). This results in a limit test that is a randomized version of the Spacing test that we
referred to as S Grid and presented in Theorem 3.
The second approach is to take a thin grid and to use S ST . This approach is perfectly valid,
this test statistics follows a uniform distribution under the null and it should be compared to
our new testing procedure S Rice . This numerical investigation has been performed in the frame
of Super-Resolution and it is presented in Figure 3, more details can be found in Section 5.2.
Figure 3 gives the cumulative distribution functions of the test statistics under “sparse” alternatives, i.e., when true spikes are to be detected. The larger the power, the better the test
Azaïs, De Castro and Mourareau/1
INTRODUCTION
4
detects spikes (abscissa represents the level of the test and the ordinate the probability to
detect the spike). In these sets of experiments, we can note that the testing procedure S Rice
based on some Hessian and the whole process X(·) is uniformly better than the spacing test
even if one takes very thin grids. One can see that the power (the ability to detect sparse objects, Dirac masses here) of the grid methods seems to present a limit that is always improved
by the continuous approach.
1
1
1
0.9
0.9
0.9
0.8
0.8
0.8
0.7
0.7
0.7
0.6
0.6
0.6
0.5
0.5
0.5
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0
0
0
0.2
0.4
0.6
0.8
1
0
0
0.2
0.4
0.6
0.8
1
1
1
1
0.9
0.9
0.9
0.8
0.8
0.8
0.7
0.7
0.7
0.6
0.6
0.6
0.5
0.5
0.5
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0
0
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0
0.2
0.4
0.6
0.8
1
Figure 3. [Under the alternative] Empirical cumulative distribution under the alternative of the Rice test
(blue) and the discrete grid tests with size 32 (green) , 102 (red), 322 (purple) and 502 (cyan) . The√alternative
is defined by a single atom at a random location with a weight log(N ) = log(2fc + 1) (first row) or N (second
row). In columns : fc = 3, 5, 7.
1.3. Toward a functional testing procedure
In the seminal paper [21], the authors generalize the LARS test to regularized least squares
problems with penalization P being a semi-norm over P coordinates but they do not cover
the “continuum” framework under consideration here. On the technical side, their convex dual
formulation leads to the study of a linear Gaussian process on which they invoke a Kac-Rice
formula while our convex dual formulation involves a functional framework for which Kac-Rice
formula cannot be applied.
Furthermore, exact testing procedures in High-Dimensional statistics are based on PostSelection Inference techniques. In a nutshell, note that the sub-gradient of the `1 -norm is
a vector with entries in [−1, 1] so that, in `1 -minimization, the selection event is described
by the signs (sub-gradient entry value {±1}) and the active support (indices for which the
sub-gradient is ±1) at the solution point through first order optimality conditions such as
the Karun-Kush-Tucker conditions. Post-Selection Inference is then the ability to build a
testing procedure that may depend on the aforementioned selection event. This approach
5
1 INTRODUCTION
might seem meaningless in the frame of Super-Resolution since the selection events have
measure zero—given Gaussian responses and a finite subset S ⊂ [0, 2π), the probability that
the output support of `1 minimization is exactly S is zero. Indeed, specifying the location
and the phase of the supremum of a Gaussian process is an event with null measure. Remark
also that there is an uncountable number of selection events—as many as the number of
points in the index of the process. So the tools of Post-Selection Inference cannot be applied
here and one need a new theoretical analysis of testing procedures based on the outcomes of
`1 -minimisation over the space of signed measures.
For the first time, this paper paves the way to build new testing procedures in the framework
of Super-Resolution theory (and line spectral estimation). In particular, we prove that we can
rightfully construct global null exact testing procedures on the first two knots λ1 and λ2
of the “continuous” LARS when one has a continuum of predictors, see Theorems 3 and 6
and Figure 2. These two new procedures offer the ability to test the mean of any stationary
Gaussian process with known correlation function Γ and C 2 -paths. Furthermore, one of these
tests is unbiased, see Theorem 3 and they can be both Studentized, see Theorems 4 and 8,
when variance σ 2 is unknown.
1.4. Outline
The outline of the paper is described by the following (short) table of contents:
2 Notations and problem formulation . . . . . . . . . . . . . . . . . . .
2.1 Hypothesis testing problem . . . . . . . . . . . . . . . . . . . .
2.2 The first and second knots of a Gaussian process . . . . . . . .
2.3 Estimating the variance . . . . . . . . . . . . . . . . . . . . . .
3 Passing to the limit, the grid approach . . . . . . . . . . . . . . . . .
4 The Rice method: a grid-less approach . . . . . . . . . . . . . . . . .
4.1 The known variance case . . . . . . . . . . . . . . . . . . . . . .
4.2 The unknown variance case . . . . . . . . . . . . . . . . . . . .
5 Applications to the Super-Resolution Theory . . . . . . . . . . . . .
5.1 Framework and results . . . . . . . . . . . . . . . . . . . . . . .
5.2 A numerical study . . . . . . . . . . . . . . . . . . . . . . . . .
6 LARS for complex-valued Gaussian processes . . . . . . . . . . . . .
6.1 Cameron-Martin type Assumption on the mean and Notations .
6.2 Description of the LARS . . . . . . . . . . . . . . . . . . . . . .
6.3 The second knot . . . . . . . . . . . . . . . . . . . . . . . . . .
6.4 Illustration: The two first knots of Super-Resolution . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
5
6
7
11
12
12
15
17
17
19
21
21
22
25
25
Notations and the formal problem formulation is described in Section 2. In Section 3, we
present the test statistic S Grid which is constructed taking the limit of consecutive LARS
knots (λ1,P , λ2,P ) on thinner and thinner grids (namely the number of predictors P tends to
infinity). Section 4 is the theoretical construction of our grid-less test based on consecutive
knots (λ1 , λ2 ) of the “continuous” LARS. The main result concerning the test statistic S Rice
is presented in this section. Applications to spike detection in Super-Resolution are developed
in Section 5. The general construction of the “continuous” LARS is given in Section 6. An
appendix with the proofs can be found at the end of the paper.
Azaïs, De Castro and Mourareau/2
NOTATIONS AND PROBLEM FORMULATION
6
2. Notations and problem formulation
2.1. Hypothesis testing problem
In this paper, our purpose is to test the mean value of a stationary complex-valued Gaussian
process Z with C 2 -paths indexed by [0, 2π). We assume that Z = A1 + ıA2 where A1 and A2
are two independent and identically distributed real-valued processes with C 2 -paths. Assume
that the correlation function Γ of A1 (and A2 ) satisfies
∀t ∈ (0, 2π),
|Γ(t)| < 1
(Anorm )
and let σ 2 := Var(A1 (·)) so that
Cov(A1 (s), A1 (t)) = σ 2 Γ(t − s) .
(1)
We denote by T := [0, 2π)2 the 2-dimensional torus. Assume that we observe a real-valued
process (X(z))z∈T indexed by T such that
∀z ∈ T, X(z) := A1 (t) cos θ + A2 (t) sin θ = Re e−ıθ Z(t) ,
where z = (t, θ) and Re(·) denotes the real part of a complex number. Remark that observing X
is equivalent to observe Z since we can recover Z from X and conversely. Furthermore, we
may assume that the process (X(z))z∈T satisfies
a.s. there is no point z ∈ T s.t. X 0 (z) = 0 and det(X 00 (z)) = 0,
(Adegen )
where X 0 (z) and X 00 (z) denote the gradient and the Hessian of X at point z. Note that
sufficient conditions for (Adegen ) are given by [5, Proposition 6.5] applied to (X(z))z∈T . In
particular if the distribution of X 00 (t) is non degenerated, using [5, Condition (b) of Proposition 6.5], it implies that Assumption (Adegen ) is met. Note also that Assumption (Adegen ) is
referred to as “Morse” process in [2]. Remark that (Anorm ) and (Adegen ) are mild assumptions
ensuring that Z is a non-pathological process with C 2 -paths.
This paper aims at testing the following hypotheses.
H0 : “Z is centered ”
against H1 : “Z is not centered ” .
Subtracting the mean tested by the null hypothesis, remark that this framework encompasses
any testing problem whose null hypothesis is a single hypothesis on the mean of Z.
2.2. The first and second knots of a Gaussian process
As in high-dimensional statistics, we can define the first and second knots (λ1 , λ2 ) as follows. If
we model some spatial correlation by means of the process X, the most correlated point zb ∈ T
and the maximal correlation λ1 are respectively the argument maximum and the maximum
of X defined by
zb := arg max X(z) and λ1 := X(b
z) .
(2)
z∈T
Under Assumption (Anorm ), one can check that the argument maximum is almost surely a
singleton, see Proposition 16.
7
2 NOTATIONS AND PROBLEM FORMULATION
To construct the second knot, given a fixed z ∈ T, one can equivalently consider two
regressions of X(y), as follows.
• On the one hand, the regression on X(z) that will appear in the grid method of Section 3.
Using a convenient normalisation related to the definition of the LARS knots, we set
∀y ∈ T \ {z},
X z (y) :=
where
X(y) − X(z)ρ(z − y)
X(y) − X(z)
= X(z) +
,
1 − ρ(z − y)
1 − ρ(z − y)
∀z ∈ T,
ρ(z) := Γ(t) cos θ ,
is the correlation function of the stationary Gaussian process X. One can check that X z is a
Gaussian process indexed by T \ {z} and independent of X(z).
• On the other hand, the regression on (X(z), X 0 (z)) will be needed for convergence purposes
in Section 4. With the convenient normalization, we set
∀y ∈ T \ {z},
X |z (y) :=
e −1 X 0 (z)i
X(y) − ρ(z − y)X(z) + hρ0 (z − y), Λ
.
1 − ρ(z − y)
e := −ρ00 (0) is the variance-covariance
where ρ0 is the gradient of the correlation function ρ and Λ
matrix of the derivative process of X, namely X 0 .
Since the derivative at zb is zero, note that X zb(·) = X |bz (·) and we define the second knot λ2 as
yb := arg max X zb(y) and λ2 := X zb(b
y ) = X |bz (b
y) ,
(3)
y∈T\{b
z}
where we prove that (b
y , λ2 ) are well defined and that yb is almost surely unique, see Proposition 16 and Remark 7. Furthermore, the couple (b
y , λ2 ) can be equivalently defined using the
extension of the LARS to our framework, the interested reader may consult Section 6.3.
2.3. Estimating the variance
When the variance σ 2 is unknown in (1), we precise here the assumptions and the estimator
we use to estimate the variance. In this section, except for explicit examples, we consider a real
valued Gaussian process Y not necessarily stationary defined on the 2-dimensional torus T. Let
m ≥ 2 (possibly infinite) and assume that Y admits an order m Karhunen-Loève expansion
in the sense that
Y =
m
X
i=1
ζi fi with Var(ζi ) = σ and ∀t ∈ T,
2
m
X
i=1
|fi (t)|2 = 1 ,
(KL(m))
where the equality holds in L2 (Ω) and (f1 , . . . , fm ) is a system of non-zero functions orthogonal
on L2 (T). Through our analysis, we need to consider one of the following assumptions.
• If m is finite,
∃(z1 , . . . , zm ) ∈ Tm pairwise distincts s.t.
(Y (z1 ), . . . , Y (zm )) is non degenerated.
(ND(m))
Azaïs, De Castro and Mourareau/2
NOTATIONS AND PROBLEM FORMULATION
8
• If m = ∞,
∀p ∈ N? , ∃(z1 , . . . , zp ) ∈ Tp pairwise distincts s.t.
(Y (z1 ), . . . , Y (zp )) is non degenerated.
( ND(∞))
Recall that a Gaussian vector is called non-degenerated if its variance-covariance matrix is
non-degenerated, i.e., it has full rank.
Some examples of process Y satisfying (KL(m)) and (ND(m)) with m = ∞ are given by
the normalized Brownian motion and any Gaussian stationary process with a spectrum that
admits an accumulation point, see [10, Page 203]. For instance, the process corresponding
to the Super-Resolution problem satisfies (KL(m)) and (ND(m)) with m finite, namely m is
twice the number of observed frequencies, see Section 5.
Definition 1. Let Y be a Gaussian process with constant variance σ 2 = Var(Y (·)) and satisfying Assumptions (KL(m)) and (ND(m)) with m finite. The quantity
m
2
σ
bKL
(Y
1 X 2
) :=
ζi ,
m
i=1
is called the Karhunen-Loève estimator of σ 2 .
2
Remark 1. An explicit expression of the estimator σ
bKL
is always possible from some set of
0
pairwise disjoint points z1 , . . . , zm0 with m > m. We only need to check that the variancecovariance matrix of the (Y (z1 ), . . . , Y (zm0 )) has rank m.
2
Remark 2. Sufficiency considerations imply that σ
bKL
is an optimal unbiased estimator for
the mean-squared error by Rao–Blackwell theorem.
Given the aforementioned definition, we are now able to construct variance estimators for
the process X. We assume that the complex Gaussian process Z that define X satisfies the
following hypotheses for some N ∈ N.
Z admits a complex Karhunen-Loève expansion of order N
(KLZ (N ))
and satisfies the following non-degeneracy conditions:
∀(t1 , . . . , tN ) ∈ [0, 2π)N pairwise distincts,
(Z(t1 ), Z(t2 ), . . . , Z(tN )) is non degenerated and
(NDZ (N ))
0
(Z(t1 ), Z (t1 ), Z(t3 ), . . . , Z(tN )) is non degenerated.
Our aim is to build, for each z ∈ T, two estimators of the variance σ 2 independently
from X(z) or (X(z), X 0 (z)). Indeed, in the following, we will distinguish two kind of statistics.
The first one is the limit of the finite dimensional statistic S Grid , see Section 3. The second one
is the case of the maximum over T, see Section 4. Both cases won’t use the same estimation
of σ 2 .
• In the grid situation, we define
z
Xnorm
(y) :=
X(y) − ρ(z − y)X(z)
p
,
1 − ρ2 (z − y)
9
2 NOTATIONS AND PROBLEM FORMULATION
where y belongs to T\{z}, ρ(·) denotes the correlation function of the process X and set
2
z
σ
bz2 := σ
bKL
Xnorm
(·)
which is well defined, independent of X(z) and with constant variance σ 2 . Furthermore,
it follows a σ 2 χ2 (2N −1)/(2N −1) distribution where χ2 (2N −1) denotes the chi-square
distribution with 2N − 1 degrees of freedom, see Proposition 2. Finally, we consider the
variance estimator
(4)
σ
b2 = σ
bzb2 ,
defined at point zb given by (2).
• In the continuous case, we define
e −1 X 0 (z)i
X(y) − ρ(z − y)X(z) + hρ0 (z − y), Λ
|z
Xnorm
(y) := q
,
e −1 ρ0 (z − y)i
1 − ρ2 (z − y) − hρ0 (z − y), Λ
where y belongs to T \ {z} and set
2
2
|z
σ
b|z
:= σ
bKL
Xnorm
(·)
which is well defined, independent of (X(z), X 0 (z)) and with constant variance σ 2 . Furthermore, it follows a σ 2 χ2 (2N − 3)/(2N − 3), see Proposition 2. Finally, we consider
the variance estimator
2
σ
b|2 = σ
b|b
(5)
z,
defined at point zb given by (2).
Proposition 2. Let Z satisfy (KLZ (N )) and (NDZ (N )) and set z ∈ T then the following
claims are true.
(a) σ
bz2 is well defined and follows a
(b)
2
σ
b|z
is well defined and follows a
σ 2 χ2 (2N −1)
2N −1
σ 2 χ2 (2N −3)
2N −3
distribution.
distribution.
|z
z
(·)/b
σz is independent of σ
bz , and the process Xnorm (·)/b
σ|z is indepen(c) The process Xnorm
dent of the random variable σ
b|z .
Proof. (a) Fix z = (t1 , θ1 ) ∈ T. Since Z satisfies (NDZ (N )), there exists (t2 , . . . , tN ) ∈
[0, 2π)N −1 pairwise different such that (Z(t1 ), Z(t2 ), . . . , Z(tN )) is non degenerated. Then,
considering z1 = z, zN +1 = z + (0, π/2) and
∀i ∈ {2, . . . , N }, zi = (ti , θ1 ) and zN +i = (ti , θ1 + π/2),
the vector V1 := (X(z1 ), . . . , X(z2N )) satisfies
2N = rank(X(z1 ), X(z2 ), . . . , X(z2N ))
z
z
= rank(X(z1 ), Xnorm
(z2 ), . . . , Xnorm
(z2N ))
z
z
= 1 + rank(Xnorm
(z2 ), . . . , Xnorm
(z2N ))
z
where rank denotes the rank of the covariance matrix of a random vector. Deduce that Xnorm
(·)
satisfies (ND(2N − 1)). This, in turn, implies that the 2N functions
gi (·) = fi (·) − ρ(z − .)fi (z)
Azaïs, De Castro and Mourareau/2
NOTATIONS AND PROBLEM FORMULATION
10
are in fact in a space of dimension 2N − 1 and a Gram-Schmidt orthogonalization in L2 (T)
z
z
z
gives (KL(2N − 1)) for the process Xnorm
(·). Finally, from (Xnorm
(z2 ), . . . , Xnorm
(z2N )), we
compute σ
bz2 that follows the desired distribution.
(b) In the case of the regression over (X(z), X 0 (z)), remark that
∂θ X(z) = X(t1 , θ1 + π/2) = X(z2 )
and ∂t X(z) = Re(e−ıθ1 Z 0 (t1 )) where ∂θ (resp. ∂t ) denote the partial derivative with respect
to θ (resp. t). Because of hypothesis (NDZ (N )), the two vectors V1 and
V2 := (X(z1 ), X(z2 ), Re(e−ıθ1 Z 0 (t1 )), Im(e−ıθ1 Z 0 (t1 )), X(z5 ), . . . , X(z2N ))
have rank 2N so both are invertible functions of (Re(ζ1 ), Im(ζ1 ), . . . , Re(ζN ), Im(ζN )). In particular, Im(e−ıθ1 Z 0 (t1 )) is a linear combination of V2 . Let γ1 and γ2 be the coefficients associated to X(z3 ) and X(z4 ). By triangular combination, we deduce that the distribution of
(X(z1 ), ∂θ X(z1 ), ∂t X(z1 ), γ1 X(z3 ) + γ2 X(z4 ), X(z5 ), . . . , X(z2N ))
is non-degenerated and so that (γ1 , γ2 ) 6= (0, 0). Setting ψ such that
γ1
γ2
cos(ψ) = p 2
and sin(ψ) = p 2
2
γ1 + γ2
γ1 + γ22
we get the non-degeneracy of
(X(z1 ), ∂θ X(z1 ), ∂t X(z1 ), X(z2N +1 ), X(z5 ), . . . , X(z2N ))
where z2N +1 = (t2 , θ1 + ψ). Finally, similarly to the proof of the previous point, regression,
|z
|z
scaling and independence prove that the rank of (Xnorm (z5 ), . . . , Xnorm (z2N +1 )) is 2N − 3 so
|z
2 is well defined and distributed
that Xnorm (·) satisfies KL(2N −3) and ND(2N −3) and that σ
b|z
as
σ 2 χ2 (2N −3)
.
2N −3
(c) This is a direct consequence of the independence of the angle and the norm for each
|z
z
marginal Gaussian vector build from Xnorm
or Xnorm .
Remark 3. When the complex process Z admits an infinite Karhunen-Loève decomposition,
we need the following modified hypothesis
∀p ∈ N? , ∀(t1 , . . . , tp ) ∈ [0, 2π)p pairwise distincts,
(Z(t1 ), Z(t2 ), . . . , Z(tp )) is non degenerated and
(NDZ (∞))
0
(Z(t1 ), Z (t1 ), Z(t3 ), . . . , Z(tp )) is non degenerated.
Indeed, for every enter p ≥ 1, note that from the observation of the vector (Z(t1 ), Z(t2 ), . . . , Z(tp ))
(resp. (Z(t1 ), Z 0 (t1 ), . . . , Z(tp ))) for pairwise disjoint points t1 , . . . , tp , we can construct an es2 (resp. σ
2 ), of the variance σ 2 with distribution σ 2 χ2 (2p − 1)/(2p − 1) (resp.
timator, say σ
b2p
b|2p
σ 2 χ2 (2p−3)/(2p−3)). Making p tend to infinity, classical concentration inequalities and Borel2 (resp. σ
2 ) converges almost surely to σ 2 . Thus the variance σ 2
Cantelli lemma prove that σ
b2p
b|2p
2 ) this observation,
is directly observable from the entire path of X. We still denote σ
bz2 (resp. σ
b|z
where z = z1 = (t1 , θ1 ).
11
3 PASSING TO THE LIMIT, THE GRID APPROACH
3. Passing to the limit, the grid approach
The main idea of this section is to define a sequence of grids (Gn )n≥1 on T, to construct a
sequence of test statistics (Sn )n≥1 from the values of the process X on Gn as in [4] and to pass
to the limit as n → ∞. More precisely, we consider Gn to be the grid with mesh ∆n := (2π)2−n
on T,
zbn := arg max X(z) and λ1,n := max X(z) .
z∈Gn
z∈Gn
It is the maximum of the process X when indexing by the grid. We can also define the
maximum of the regression when indexing by the grid, namely
λ2,n :=
max
y∈Gn \{b
zn }
X zbn (y) .
The Hessian at the maximum (2) on T is denoted by X 00 := X 00 (b
z ). By Assumption (Adegen ),
it is a random variable with values in the set of non degenerated negative definite matrices
of size 2 × 2. We can define a non degenerated positive quadratic form (i.e., a metric) on R2
by kvkX 00 = −v > X 00 v, for v ∈ R2 . Using this metric, we can consider the corresponding
Voronoi tessellation of Z2 . It is a regular partition of R2 by parallelograms, invariant by
translations (1, 0) and (0, 1). Denote by Vo ⊂ [−1, 1]2 the Voronoi cell of the origin in this
partition and by U := U(Vo ) the uniform distribution on this cell. We understand the law U as
a conditional law with respect to X 00 and, conditionally to X 00 , this law is taken independent
of (λ1 , λ2 ), see Lemma 12. Conditionally to X 00 , define the randomized statistics
(
)
k − 2U
k>
λ2 := λ2 ∨ λ1 + sup
X 00
,
(6)
1
e
e 21 kk
2
2
k∈Z \{0} kΛ kk
kΛ
e 12 is the square root of Λ
e = −ρ00 (0) and a ∨ b = max(a, b). A proof of the following
where Λ
result is given in Appendix A.1.
Theorem 3. Under Assumptions (Anorm ) and (Adegen ), the randomized test statistics
S Grid :=
Φ(λ1 /σ)
∼ U([0, 1]) ,
Φ(λ2 /σ)
where Φ denotes the standard Gaussian survival function. Moreover, the test with p-value S Grid
is unbiased: under the alternative H1 , it holds P{S Grid 6 α} > α for all α ∈ (0, 1).
Theorem 3 shows in particular that the statistics—referred to as the Spacing test statistics
in the introduction—given by
Φ(λ1 /σ)
(7)
S ST =
Φ(λ2 /σ)
does not follows a U([0, 1]) distribution ans leads to a non-conservative test. Indeed, observe
that almost surely λ2 ≤ λ2 so that S ST ≥ S Grid almost surely. Note that the two test statistics
differ on the event {λ2 6= λ2 } = {λ2 < λ2 }.
Now, when the variance σ 2 is unknown, we can build an estimator σ
b2 defined in (4) and
obtain a Studentized version of the previous theorem.
Azaïs, De Castro and Mourareau/4
THE RICE METHOD: A GRID-LESS APPROACH
12
Theorem 4. Assume (Anorm ), (Adegen ), (KLZ (N )) and (NDZ (N )) where 2 6 N < ∞, then
the following test statistics T Grid satisfies
T Grid :=
F m−1 (λ1 /b
σ)
∼ U([0, 1])
F m−1 λ2 /b
σ
under the null hypothesis where m = 2N , Fm−1 is the Student cumulative distribution function
b2 is defined by (4).
with m − 1 degrees of freedom, F m−1 = 1 − Fm−1 its survival function and σ
A proof can be found in Appendix A.2.
Remark 4. Only the first point of (NDZ (N )) is required for the proof. Moreover, if m = +∞,
the Student distribution is to be replaced by a standard normal distribution.
4. The Rice method: a grid-less approach
In this section, we build our test statistic directly on the entire path of the process X in a
grid-less manner. We assume that the process X satisfies Assumptions (Anorm ) and (Adegen ),
and is centered. As in the preceding section, we consider λ1 and λ2 defined by (2) and (3)
respectively.
e so that the covariance function of X
e is the correlation function ρ of X,
We denote X = σ X
e is the standardized version of X. Note that, by regression formulas and stationarity,
namely X
it holds
00
e
e 0 (z)) = −Λ
e X(z)
e
e (z) (X(z),
X
,
∀z ∈ T, E X
e by the decomposition
so that we can define the process R
e 00 (z) = −Λ
e X(z)
e
e
X
+ R(z)
e
e
e = −ρ00 (0) is the variance-covariance
where R(z)
and X(z)
are independent for any z ∈ T and Λ
0
e (t). In particular, observe that
matrix of X
e z ) + R(b
X 00 (b
z ) = −ΛX(b
z) ,
e z ). Using the Rice method of [5, Theorem 7.2] (see also [18]), it follows that
where R(b
z ) = σ R(b
the maximum λ1 has for density w.r.t the Lebesgue measure on R+ at point ` > 0
e
(cst)(−1)d E det(−ΛX(0)
+ R(0))1A` X(0) = `, X 0 (0) = 0 σ −1 φ(σ −1 `),
where φ denotes the standard Gaussian density, A` is the event {X(y) 6 `, ∀y ∈ T} and (cst),
as in the following, denotes a positive constant. The numerical values (cst) may vary from
an occurence to another and it may depend on m and σ which are assumed fixed in our
framework.
4.1. The known variance case
We begin by the known variance case. The main observation is that the method of [5, Theorem 7.2] can be extended to compute the joint distribution of (λ1 , λ2 , R(b
z )) as follows.
• Denote S the set of symmetric matrices and pick a Borel set B on D := R2 × S.
13
4 THE RICE METHOD: A GRID-LESS APPROACH
• For every z ∈ T, recall that
∀y ∈ T \ {z},
X |z (y) :=
and define
e −1 X 0 (z)i
X(y) − ρ(z − y)X(z) + hρ0 (z − y), Λ
1 − ρ(z − y)
∀z ∈ T,
λz2 :=
sup X |z (y) .
(8)
y∈T\{z}
Remark that, for fixed z ∈ T, λz2 is a.s. finite by Lemma 9, X |z (·) is independent
of (X(z), X 0 (z)) and, by way of consequence, λz2 is independent of (X(z), X 0 (z)). Furthermore, note that since T is without boundary, for z = zb, one has X 0 (z) = 0 and
λz2 = λ2 as defined by (3).
• Observe that on the event {∀y 6= z, X(y) < X(z)} one has almost surely that z = zb,
X(z) = λ1 , λz2 = λ2 and R(z) = R(b
z ). Also, a simple computation shows that
∀z ∈ T s.t. X 0 (z) = 0,
1{z=bz } = 1{∀y6=z,
X(y)<X(z)}
= 1{0<λz2 <X(z)} ,
almost surely. Hence, by unicity of zb and recalling that the set {z ; X 0 (z) = 0} is finite
under (Adegen ), we deduce that
X
1{(X(z),λz2 ,R(z))∈B}∩{0<λz2 <X(z)} = 1{(λ1 ,λ2 ,R(bz ))∈B} .
z:X 0 (z)=0
(n)
• On D define smooth lower approximations ϕB
converge when n goes to infinity i.e.
∀(`1 , `2 , r) ∈ Ω,
of the indicator function of B that
(n)
ϕB (`1 , `2 , r) −→ 1{(`1 ,`2 ,r)∈B}∩{0<`2 <`1 } .
• Apply Rice formula with weights [5, Theorem 6.4] (see also the proof of [5, Theorem 7.2])
to compute
h X
i
(n)
z
E
ϕB X(z), λ2 , R(z)
z:X 0 (z)=0
Z
= (cst)
ZT
= (cst)
T
h
i
(n)
e
E | det(−ΛX(z)
+ R(z))| ϕB (X(z), λz2 , R(z)) X 0 (z) = 0 dz
h
i
(n)
e
E | det(−ΛX(z)
+ R(z))| ϕB (X(z), λz2 , R(z)) dz
where the last equality relies on the fact that (X(z), λz2 , R(z)) is independent of X 0 (z).
• Combining the previous observations and passing to the monotone limit as n tends to ∞
in the aforementioned Rice formula with weights, we get that
P λ1 , λ2 , R(b
z) ∈ B
h X
i
=E
1{(X(z),λz2 ,R(z))∈B}∩{0<λz2 <X(z)}
z:X 0 (z)=0
Z
h
i
e
E | det(−ΛX(z)
+ R(z))|1{(X(z),λz2 ,R(z))∈B}∩{0<λz2 <X(z)} dz
T
h
i
e
= (cst) E | det(−ΛX(0)
+ R(0))|1{(X(0),λ02 ,R(0))∈B}∩{0<λ02 <X(0)} ,
h
i
e
= (cst) E det(−ΛX(0)
+ R(0))1{0<λ02 <X(0)} 1{(X(0),λ02 ,R(0))∈B} ,
= (cst)
(9)
Azaïs, De Castro and Mourareau/4
THE RICE METHOD: A GRID-LESS APPROACH
14
by stationarity and using that, on the event {0 < λ02 < X(0)}, the matrix −X 00 (0) =
e
ΛX(0)
− R(0) belongs to the set of positive definite symmetric matrices, namely S+ .
Before stating the key result on the joint density of (λ1 , λ2 , R(b
z )) we need to introduce
a dominating measure. First, recall that X(0) is independent of the pair (λ02 , R(0)). Then,
e0 , R(0))
e0 is defined as in (8) for the process X.
e
e
observe that (λ02 , R(0)) = σ × (λ
where λ
2
2
0
e
e
Denote µ1 the law of (λ2 , R(0)) and note that it does not depend on σ. Denote µσ the law
of (λ02 , R(0)) and remark that for any Borel set B of R × S, it holds µσ (σB) = µ1 (B).
Eventually, remark that
The law of (X(0), λ02 , R(0)) is dominated by Leb(R) ⊗ µσ .
(10)
where Leb(R) denotes the Lebesgue measure on R. As a consequence we can prove the following proposition.
Proposition 5. The joint law L((λ1 , λ2 , R(b
z ))) of (λ1 , λ2 , R(b
z )) satisfies for all (`1 , `2 , r) ∈
2
R × S,
dL((λ1 , λ2 , R(b
z )))
e 1 + r)1{0<` <` } σ −1 φ(σ −1 `1 ) ,
(`1 , `2 , r) = (cst) det(−Λ`
2
1
dLeb(R) ⊗ µσ
where Leb(R) ⊗ µσ is defined by (10) and S denotes the set of symmetric matrices.
Proof. Observe that the density at point `1 of X(0) with respect to the Lebesgue measure
is σ −1 φ(σ −1 `1 ) and recall (10). Now, for any Borel set B of R2 × S, note that
h
i
e
E det(−ΛX(0)
+ R(0))1{(0<λ02 <X(0)} 1{(X(0),λ02 ,R(0))∈B}
Z
e 1 + r)1{0<` <` } σ −1 φ(σ −1 `1 )d`1 µσ (d(`2 , r))
=
det(−Λ`
2
1
B
thanks to (9), which prove the result.
We can now state our result when the variance is known.
Theorem 6. Set
+
∀r ∈ S , ∀` > 0,
Z
+∞
Gr (`) :=
e + r)φ(uσ −1 )du ,
det(−Λu
`
e denotes the Hessian of the correlation function ρ of X at the origin. Under Assumpwhere Λ
tions (Anorm ) and (Adegen ), the test statistic
S Rice :=
GR(bz ) (λ1 )
GR(bz ) (λ2 )
∼ U([0, 1])
under the null H0 .
Proof. Using Proposition 5, we know that the density of λ1 at `1 and conditional to (λ2 , R(b
z )) =
(`2 , r) is equal to
e 1 + r)φ(σ −1 `1 )1{0<` <` } .
(cst) det(−Λ`
2
1
4 THE RICE METHOD: A GRID-LESS APPROACH
15
It is well known that, if a random variable Z has for cumulative density function F then F(Z)
follows an uniform distribution on [0, 1]. This implies that, conditionally to (λ2 , R(b
z )) = (`2 , r),
Gr (λ1 )
∼ U([0, 1]).
Gr (`2 )
Since the conditional distribution does not depend on (`2 , r), it is also the non conditional
distribution and it yields
GR(bz ) (λ1 )
∼ U([0, 1]) ,
GR(bz ) (λ2 )
as claimed.
4.2. The unknown variance case
e where σ > 0 is unknown. Assume that Z satisHence, suppose that we observe X = σ X
fies (KLZ (N )) and (NDZ (N )), and set m = 2N . The regression of the Hessian on (X(z), X 0 (z))
reads now
e
e
∀z ∈ T, X 00 (z) = −ΛX(z)
+ σ R(z).
because X 0 (z) is independent of (X(z), X 00 (z)) by stationarity. The variance being unknown,
we estimate it using σ
b|2 which is defined by (5). For fixed z ∈ T, by Claims (b) and (c) of
Proposition 2, we know that the following random variables or random processes
|z
X(z) , X 0 (z) ,
Xnorm (·)
and σ
b|z
σ
b|z
|z
are mutually independent. As Xnorm (·) = hz (·) X |z (·) where hz (·) is a deterministic function
and as Lemma 15 shows that R(z) can be expressed as radial limits of X |z (·) at point z, we
get that
X |z (·) R(z)
and σ
b|z are mutually independent,
,
X(z) , X 0 (z) ,
σ
b|z
σ
b|z
and by consequence
X(z) , X 0 (z) ,
λz R(z)
2
,
and σ
b|z are mutually independent.
σ
b|z σ
b|z
We turn now to the Rice formula described previously and introduce the notation
T2,z :=
λz2
and T2 := T2,bz .
σ
b|z
Denote Leb(R2 ) the Lebesgue measure on R2 and let µ1 be the joint law of the couple of
random variables (T2,0 , R(0)/b
σ|0 ). Under H0 , note that X(0) is a centered Gaussian variable
2
with variance σ and σ
b|0 /σ is distributed as a chi-distribution with m − 3 degrees of freedom,
i.e., the law of density
m−3
21− 2 m−4
s
fχm−3 (s) =
exp(−s2 /2)
Γ m−3
2
Azaïs, De Castro and Mourareau/4
THE RICE METHOD: A GRID-LESS APPROACH
16
b|0 /σ, T2,0 , R(0)/b
σ|0 ) has a denwhere Γ is the Gamma function. Then the quadruplet (X(0), σ
sity with respect to Leb(R2 ) ⊗ µ1 at point (`1 , s, t2 , r) ∈ R3 × S equal to
2
s (m − 3)
m−4
(cst) s
exp −
σ −1 φ(σ −1 `1 ).
2
Using the same method as for the proof of Proposition 5 we have the following proposition.
Proposition 7. Assume that Z satisfies (Anorm ), (Adegen ), (KLZ (N )) and (NDZ (N )), and
set m = 2N . Then, the joint distribution of λ1 , σ
b| /σ, T2 , R(b
z )/b
σ| has a density with respect
to Leb(R2 ) ⊗ µ1 at point (`1 , s, t2 , r) ∈ R3 × S+ equal to
2
s (m − 3)
m−4
e
(cst) det(−Λ`1 + σsr) s
exp −
φ(σ −1 `1 )1{0<σst2 <`1 } ,
2
where (cst) is a positive constant that may depend on m and σ.
Consequently, we derive the following result.
Theorem 8. Assume that Z satisfies (Anorm ), (Adegen ), (KLZ (N )) and (NDZ (N )), and set
m = 2N . For all r ∈ S+ , define H r (·) as
!
r
Z +∞
m
−
1
e 1 + r fm−1 t1
∀` > 0, H r (`) :=
det − Λt
dt1 ,
m−3
`
where fm−1 is the density of the Student distribution with m − 1 degrees of freedom. Under the
null H0 , the test statistic
H R(bz ) (T1 )
T Rice :=
∼ U([0, 1]),
H R(bz ) (T2 )
where T1 := λ1 /b
σ| , T2 = λ2 /b
σ| and σ
b| is defined by (5).
`1
Proof. First, using Proposition 7 and the change of variable t1 = σs
, the joint distribution of
the quadruplet (T1 , σ
b| /σ, T2 , R(b
z )/b
σ| )) at point (t1 , s, t2 , r) is given by
2
s (m − 3)
m−3
e
(cst) det(σs(−Λt1 + r))s
exp −
φ(st1 )1{0<t2 <t1 }
2
!2
r
m
−
3
m
−
1
e 1 + r)sm−1 exp − s
φ(st1 )1{0<t <t } .
= (cst) det(−Λt
2
1
m−1
2
Second, note that if X and Y are two independent random variables of density fX and fY
then the density of X/Y satisfies
Z
fX/Y (z) =
fX (zy)yfY (y)dy.
R
In our case, integrating over s and with the change of variable s ← s
p
(m − 1)/(m − 3), it
5 APPLICATIONS TO THE SUPER-RESOLUTION THEORY
17
holds
!2
m
−
3
m
−
1
ds
φ(st1 )sm−1 exp − s
m−1
2
+
R
!
r
2
Z
m−1
s (m − 1)
m−2
ss
exp −
ds
= (cst)
φ st1
m−3
2
R+
!
r
Z
m−1
χm−1 (s)ds
φ st1
= (cst)
sf √
m−3
m−1
R+
!
r
m−1
.
= fm−1 t1
m−3
Z
r
Putting together, the density of (T1 , T2 , R(b
z )/b
σ ) at point (t1 , t2 , r) is now given by
!
r
m
−
1
e 1 + r)fm−1 t1
(cst) det(−Λt
1{0<t2 <t1 } ,
m−3
and we conclude using the same trick as the one of Theorem 6.
5. Applications to the Super-Resolution Theory
5.1. Framework and results
Deconvolution over the space of complex-valued Radon measure has recently attracted a lot
of attention in the “Super-Resolution” community—and its companion formulation in “Line
spectral estimation”. A standard aim is to recover fine scale details of an image from few low
frequency measurements—ideally the observation is given by a low-pass filter. The novelty in
this body of work relies on new theoretical guarantees of the `1 -minimization over the space
of Radon measures with finite support. Some recent works on this topic can be found in the
papers [11, 7, 20, 8, 3, 13, 6, 12] and references therein.
An important example throughout this paper is given by the Super-Resolution problem
which can be stated as follows. Let ν 0 ∈ (M([0, 2π), C), k·k1 ) a complex-valued Radon measure
on the one dimensional torus identified to [0, 2π) equipped with the natural circle-wise metric.
Note that ||·||1 denotes the total variation norm on M([0, 2π)). The space (M([0, 2π), C), k·k1 )
can be defined as the topological dual space of continuous functions on [0, 2π) equipped with
the L∞ -norm.
Let N = 2fc + 1 where fc ≥ 1 is referred to as the “frequency cut-off”. Denote by DN the
Dirichlet kernel defined by
∀t ∈ [0, 2π),
DN (t) :=
sin(N t/2)
.
sin(t/2)
Consider the linear operator FN : M([0, 2π), C) → CN that maps any complex-valued Radon
measure ν onto its Fourier coefficients ck (ν) where
Z
ck (ν) :=
exp(−ıkx)ν(dx)
T
Azaïs, De Castro and Mourareau/5
APPLICATIONS TO THE SUPER-RESOLUTION THEORY
18
for integers k such that |k| 6 fc . Consider ζ = (ζk )k where ζk = ζ1,k + ıζ2,k and ζ`,k are i.i.d.
standard Gaussian random variables for |k| 6 fc and ` = 1, 2. In the Super-Resolution frame,
we observe a perturbed version of the Fourier coefficients, namely
1
y = √ FN (ν 0 ) + σζ .
N
? —the dual operator of F , remark that we observe the trigonometric polynomial
Applying FN
N
1
?
(y)
Z := √ FN
N
which reads as
∀t ∈ [0, 2π),
1
Z(t) =
N
Z
T
DN (t − x)ν 0 (dx) + σ
fc
X
k=−fc
1
√ ζk exp(ıkt).
N
(11)
Hence, one observes Z and infers on ν 0 assuming that it has finite support. To this purpose,
consider the process X defined for all (t, θ) ∈ T by
X(t, θ) := Re(e−ıθ Z(t)) = cos(θ) Re(Z(t)) + sin(θ) Im(Z(t)),
(12)
where Re and Im denote the real and imaginary part of a complex number. When ν 0 ≡ 0,
remark that the processes A1 = Re(Z) and A2 = Im(Z) are two independent and identically
distributed real-valued processes with C ∞ -paths. An elementary computation shows that X
has correlation function ρ and A1 has correlation function Γ with
ρ(z − y) = cos(θ − α)Γ(t − s) where Γ(t − s) = DN (t − s)/N
for all z = (t, θ) and y = (s, α) in T. Remark that (Anorm ) holds true for Γ. In this case, we
are testing
?
H0 : “FN
(FN (ν 0 )) ≡ 0 ”
?
against H1 : “∃t ∈ [0, 2π), FN
(FN (ν 0 ))(t) 6= 0 ” ,
or equivalently
H0 : “ν 0 ≡ 0 ”
against H1 : “∃t ∈ [0, 2π), ν 0 (t) 6= 0 ” .
Subtracting the known measure ν 0 , remark that this framework encompasses testing problem
whose null hypothesis is any single hypothesis H0 : “ν 0 ≡ ν0 ” against alternatives of the form
H1 : “∃t ∈ [0, 2π), ν 0 (t) 6= ν0 (t) ”.
Furthermore, we have the following propositions. First, we check that we can apply our
results to the Super-Resolution process.
Proposition 9. The process X defined by (12) satisfies Condition (KL(m)) and Condition (ND(m)) with m = 2N = 4fc + 2.
19
5 APPLICATIONS TO THE SUPER-RESOLUTION THEORY
Then, we derive a first result when the noise level σ si known.
Proposition 10. Under the null H0 , the test statistic
Rice
SSR
=
σ(α1 λ1 + α2 )φ(λ1 /σ) + (α1 σ 2 − α32 )Φ(λ1 /σ)
∼ U([0, 1]),
σ(α1 λ2 + α2 )φ(λ2 /σ) + (α1 σ 2 − α32 )Φ(λ2 /σ)
where Φ is the standard Gaussian cumulative distribution function , Φ = 1 − Φ its survival
function, φ its density function, (λ1 , λ2 ) is defined by ((2), (3)) and
α1 = 31 fc (fc + 1),
fc
b
α = √1 P (k 2 − α ) × Re(y eı(kbt−θ)
),
1
2
k
N
k=−fc
fc
P
b
1
k × Re(yk eı(kbt−θ) ).
α3 = √N
k=−fc
Finally, we have the following result when the noise level σ is unknown.
Proposition 11. Under the null H0 , the test statistic
Rice
TSR
=
−1 α2 F
α1 F m−3 (T1 ) + (α1 T1 + α2 )fm−3 (T1 ) − γm
3 m−1 (T1 )
∼ U([0, 1]),
−1 2
α1 F m−3 (T2 ) + (α1 T2 + α2 )fm−3 (T2 ) − γm α3 F m−1 (T2 )
where Fd is the Student cumulative distribution function with d degrees of freedom, F d = 1−Fd
its survival function, fd its density function, T1 = λ1 /b
σ| , T2 = λ2 /b
σ| , σ
b| is defined by (5) and
m−3
m
Γ
Γ
)
)
(
(
2
2
γm = m−3
m−2 Γ( m−1 )Γ( m−2 ) .
2
2
A proof of these propositions can be found in Appendix A.4.
5.2. A numerical study
5.2.1. Computation of λ2
Rice ), we need
To build our test statistic S Rice in the Super-Resolution context (namely SSR
to compute three quantities. The first one is λ1 , the maximum of X(.) over the torus T.
Its simple form allow us to use classical optimization routines, for instance fminsearch on
MATLAB or optim on R both combined with global resolution options on T. The second
one is R = R(b
z ) which appears in the test statistic through the coefficients α1 , α2 and α3 that
are simple functions of the observation y and zb. Finally, the third one is
X(y) − X(b
z)
zb
.
λ2 = λ2 = λ1 + max
y∈T
1 − ρ(b
z − y)
Contrary to λ1 , there is some indetermination problem when y is close to zb. In particular, the
approximation of zb is by definition not exact and the radial limits of X | are not numerically
achieved. A way to get around that is the integral form of the remainder in Taylor’s theorem.
In full generality, we compute
Z 1
Q
P
2
T 00
r
Iq
(1
−
h)
(y
−
z
b
)
X
(b
z
+
h(y
−
z
b
))(y
−
z
b
)
dh
q=0
zb
0
λ2 − λ1 = max Z 1
'
,
Q
y∈T
P
T 00
2
(1 − h) (y − zb) ρ (b
z + h(y − zb))(y − zb) dh
r
Jq
0
q=0
Azaïs, De Castro and Mourareau/5
APPLICATIONS TO THE SUPER-RESOLUTION THEORY
20
where Iq and Jq are given by Simpson’s rule [1, Page 886] applied to the numerator (resp.
denominator), Q is a tuning parameter that control the approximation error and r = ||y − zb||22 .
However, in the Super-Resolution case, elementary trigonometry identities give the following
simpler form of the denominator
r
2
fc
X
k=−fc
2
(k cos(α) − sin(α)) × sinc
r(k cos(α) − sin(α))
2
2
where y − zb = (r cos(α), r sin(α)) and sinc denote the cardinal sine function, i.e.
sin x
x
sinc(x) =
1
if x 6= 0,
if x = 0,
which is a numerically robust function. We conclude the optimization using the same routine
as the one of λ1 .
5.2.2. Monte-Carlo experiment
In this section we compare the cumulative distribution of several statistics of test in the case
where the variance is known, namely
• The statistics of the Rice test S Rice , given by Theorem 6, are displayed in blue.
• The statistics of the Spacing test S ST , given by (7), are displayed in green.
• The statistics of the Spacing test on grids Gn given by Φ(λ1,n )/Φ(λ2,n ) are displayed
with a color that take the respective values green, red, purple and cyan for sizes equal
to 32 , 102 , 322 , 502 .
• The grid test, based on S Grid of Theorem 3 can be viewed as the limit of the discrete
grid tests above as the size growths to infinity. As one can see in the figures, there is
some evidence that this limit is numerically reached for a size n = 502 .
We complete each graph by the diagonal to the cumulative distribution function of the uniform law on [0, 1] displayed in black. All the figures are based on 2000 simulations of the
corresponding statistics.
The first figure studies the distribution of S Rice and S ST under the Null. This figure is
displayed in the introduction (see Figure 2). The second figure deals with the grid statistic
and S Rice under various alternatives defined by a single spike and compares the power of the
Rice test with the discrete grid tests, see Figure 3. Finally, the third figure performs the same
study but with an alternative defined by two atoms, see Figure 4.
Discussion
Figure 2 suggests that the Spacing test is highly non-conservative which is a major drawback.
For instance, when fc = 7, the empirical level of the Spacing test at a nominal level of 5%
is in fact 11,3%, showing that this test is very non-conservative. For its part, the Rice test is
exact as predicted by the theory. This numerical agreement prove that the numerical algorithm
described in Section 5.2.1 is efficient.
21
6 LARS FOR COMPLEX-VALUED GAUSSIAN PROCESSES
1
1
1
0.9
0.9
0.9
0.8
0.8
0.8
0.7
0.7
0.7
0.6
0.6
0.6
0.5
0.5
0.5
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0
0
0
0.2
0.4
0.6
0.8
1
0
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
Figure 4. Same as Figure 2 except that (a) fc = 7 , (b) the alternative is defined by two spikes at random
√
locations
√ √ (with a constraint of separation) (c) the weights are now from left to right (log N, log N ); (log N, N );
( N , N ).
In Figure 3 and 4 we see that the power of the discrete grid tests may seem an increasing
function of the number of points of the grid. This power seems to converge since the curves
associated to 322 (purple) and 502 (cyan) are almost indistinguishable. This suggests that
the Rice test (blue) is always more powerful than the discrete grid test or the limit grid test.
Consequently, it seems unbiaised for any choice of alternative.
In conclusion the Rice test seems to be the best choice even if we are still not able to prove
theoretically that it is unbiased.
6. LARS for complex-valued Gaussian processes
6.1. Cameron-Martin type Assumption on the mean and Notations
The algorithm presented here can be used for a large class of complex processes Z. In this
section, we introduce new notation and a new set of assumptions. We consider a complexvalued Gaussian process Z indexed on a compact metric space K with covariance function K.
Remark 5. Note that this model encompasses our framework setting K = [0, 2π) and K =
2σ 2 Γ with Γ the correlation of A1 defined in Section 2.1. We do not assume that the process
is stationary in this section though our method can be applied to stationary processes as for
instance in the Super Resolution frame, see Section 6.4.
We assume that its covariance K is such that there exists σ > 0 such that
∀s 6= t ∈ K,
K(t, t) = 2σ 2 and K(s, t) < 2σ 2 .
(13)
The scalar 2 accounts for the contribution of the real and the imaginary part of Z and σ 2 is
the variance of the real part of Z. We assume that Z has continuous sample paths.
We present here the underlying hypothesis on the mean of the Gaussian processes under
consideration when using the LARS algorithm. This hypothesis is of Cameron-Martin type.
Indeed, the main drawback that should be avoided is when the mean cannot be represented in
the RKHS of the Gaussian process. We recall that we can define a reproducing Hilbert space
of the covariance K, see [15, Chapter 2.6] for instance. Denote (H, h·, ·iH ) this complex Hilbert
space. Also, we can invoke a Karhunen-Loève expansion of the process Z. Namely, there exist
Azaïs, De Castro and Mourareau/6
LARS FOR COMPLEX-VALUED GAUSSIAN PROCESSES
22
i.i.d. complex standard normal variables (gj )j≥1 , a real orthonormal system (ej )j≥1 on L2 (K)
and σj > 0 such that
X
X
Z − EZ =
σj gj ej and
σj2 = 2σ 2 < ∞,
j
j
where the identity holds almost surely in the Banach space of continuous functions on K
equipped with the L∞ -norm. By Mercer’s theorem, we know that
X
∀s, t ∈ K, K(s, t) =
σj2 ej (s)ej (t) ,
j
where the identity holds almost surely in the Banach space of continuous functions on K × K
equipped with the L∞ -norm. We recall also that the Hilbert space H can be defined as
H :=
nX
j
aj ej |
X |aj |2
j
σj2
<∞
o
with the inner product
X
X
X aj bj
.
h
aj ej ,
bj ej iH =
σj2
j
j
j
We observe Z and we want to estimate its mean E Z. Remark that almost surely it holds
Z − E Z ∈ H, where H is the closure of H in the space of continuous functions equipped
with the infinity norm, see e.g. [15, Corollary 2.6.11]. Remark that H is also closed in L2 (K).
Denoting by E the L2 orthogonal space of H, one has L2 (K) = H ⊕ E where the sum is
orthogonal. We denote by P (resp. P ⊥ ) the orthogonal projection onto H (resp. E). Since
almost surely Z − E Z ∈ H, remark that almost surely P ⊥ (Z) = P ⊥ (EZ) and this process
can be observed and is deterministic. Without loss of generality, we assume that P ⊥ (EZ) = 0
subtracting P ⊥ (Z) to Z. Also, we assume that
P(EZ) ∈ H .
(14)
Recall that P(EZ) = EZ and Assumption (14) gives that Z ∈ H using Z − E Z ∈ H.
6.2. Description of the “continuous” LARS
We assume that Z ∈ H and, as mentioned above, this assumption is equivalent to Assumption (14). Following standard references, e.g., [16, Chapter 5.6], the Least-Angle Regression
Selection (LARS) algorithm can be extended to Gaussian processes. To the best of our knowledge, the LARS for complex Gaussian processes has never been introduced and we present
its formulation here for the first time. Actually, the presentation given in this section can be
applied to any RKHS setting. It results in a description of the LARS in infinite dimensional
feature spaces and this framework has been dealt in [19]. However, note that the paper [19]
only concerns real signed measures and their “doubling” dimension trick [19, page 546] cannot
be used when dealing with complex measures. In particular, their result cannot be invoked
in Super-Resolution where it is of utmost importance to deal with complex measures. This
section presents the “continuous” LARS for Super-Resolution.
6 LARS FOR COMPLEX-VALUED GAUSSIAN PROCESSES
23
The LARS is a variable selection algorithm giving a sequence ((λk , µk ))k≥1 where the knots
are ordered such that λ1 ≥ λ2 ≥ . . . > 0 and µk ∈ (M(K, C), k · k1 ) is a complex-valued
measure. We recall that the space (M(K, C), k · k1 ) is defined as the dual space of the space
of continuous functions on K equipped with the L∞ -norm.
Inspired by the Super-Resolution framework—presented in Section 5, we consider Z as
some “correlation process” in the spirit of (11). In particular, the most correlated point can
be defined by (2), namely
λ1 := max |Z(t)| .
t∈K
Under Assumption (13), Proposition 16 shows that almost surely there exists a unique point t1
such that λ1 = |Z(t1 )|. Define the “active set” A as A1 = (t1 ). This active set is the support
of the second fitted solution µ2 . Set the first fitted solution to µ1 = 0 and the first residual
to Z1 = Z for initialization purposes. Observe that
(15)
Z1 (t1 ) = λ1 eıθ1 ,
|Z1 (t1 )| = λ1 ,
∀t ∈ K,
∀t 6= t1 ,
Z
Z1 (t) = Z(t) −
K(u, t)µ1 (du) ,
K
|Z1 (t)| < λ1 .
From this point we proceed iteratively. For k ≥ 2, we assume that we have found (λk−1 , µk−1 )
and Ak−1 = (t1 , . . . , tk−1 ) such that
Z
K(u, t)µk−1 (du) ,
∀t ∈ K, Zk−1 (t) := Z(t) −
K
∀t ∈ Ak−1 ,
∀t ∈
/ Ak−1 ,
|Zk−1 (t)| = λk−1 ,
|Zk−1 (t)| < λk−1 .
We want to add an other point tk to the active set and define a discrete measure µk supported
on Ak−1 while keeping the above inequalities true. From this point proceed as follows.
◦ First, we solve the least-squares fit given by
a = arg min Zk−1 (·) −
c∈Ck−1
k−1
X
i=1
ci K(ti , ·)
2
H
This program can be solved in closed form since its amounts in fitting k − 1 weights in leastsquares regression. It holds that
−1
a = Mk−1
(Zk−1 (t1 ), . . . , Zk−1 (tk−1 ))
where we denote Mk−1 = (K(ti , tj ))1≤i,j≤k−1 and we assume that Mk−1 is invertible. If Mk−1
is not invertible then we stop. The complex vector a = (a1 , . . . , ak−1 ) is referred to as the
“least-squares direction”.
P
Remark 6. Note that the function k−1
i=1 ai K(ti , ·) is the regression of Zk−1 onto the finite
dimensional space Span{K(ti , ·) ; i = 1, . . . , k − 1}.
Azaïs, De Castro and Mourareau/6
◦
LARS FOR COMPLEX-VALUED GAUSSIAN PROCESSES
24
Then, for any 0 < λ ≤ λk−1 , define
µ(λ) := µk−1 + (1 − λ/λk−1 )
k−1
X
ai δti
i=1
k−1
X
Z (λ) (·) := Zk−1 (·) + (λ/λk−1 − 1)
i=1
,
ai K(ti , ·) ,
and observe that |Z (λ) (t)| = λ for all t ∈ {t1 , . . . , tk−1 }. Indeed, it holds
Z (λ) (tj ) = Zk−1 (tj ) + (λ/λk−1 − 1)
k−1
X
ai K(ti , tj ) ,
i=1
>
= Zk−1 (tj ) + (λ/λk−1 − 1)a Mk−1 (0, . . . , 0, |{z}
1 , 0, . . . , 0) ,
jth
= λZk−1 (tj )/λk−1 ,
and recall that |Zk−1 (tj )| = λk−1 . Now, we keep track of the largest value of the “correlation”
process |Z (λ) | on the complementary set of Ak−1 while moving λ from λk−1 toward zero. We
define λk as the largest value for which there exists a point t ∈
/ Ak−1 such that |Z (λ) |(t) = λ.
Set
λk := max β > 0 ; ∃ t ∈
/ Ak−1 , s.t. |Z (β) |(t) = β ,
(16)
and tk := arg max |Z (λk ) |(s) .
s∈A
/ k−1
If tk is not unique, we add all the solutions of (16) to the active set Ak . For sake of readability,
we assume that tk is the only solution to (16). ◦ Update
Ak = (t1 , . . . , tk ) ,
µk = µ(λk ) = µk−1 + (1 − λk /λk−1 )
k−1
X
ai δti ,
i=1
Zk (·) = Z (λk ) (·) = Zk−1 (·) + (λk /λk−1 − 1)
where, for all t ∈ K,
Zk (t) = Zk−1 (t) + (λk /λk−1 − 1)
= Z(t) −
= Z(t) −
Z
K
k−1
X
k−1
X
i=1
ai K(ti , ·) ,
ai K(ti , t)
i=1
K(u, t)µk−1 (du) + (λk /λk−1 − 1)
k−1
X
ai K(ti , t) ,
i=1
Z
K(u, t)µk (du) ,
K
is the kth residual associated to the kth fitted solution µk . Remark also that
∀t ∈ {t1 , . . . , tk },
∀t 6= {t1 , . . . , tk },
and update k to k + 1 to iterate the procedure.
|Zk (t)| = λk ,
|Zk (t)| < λk ,
25
6 LARS FOR COMPLEX-VALUED GAUSSIAN PROCESSES
6.3. The second knot
First, observe that λ1 is defined as in (2) and that the two definitions agree. Indeed, recall
that X(t, θ) = Re (e−ıθ Z(t)) so that max X = max |Z| at point zb = (t1 , θ1 ) with t1 as in (15).
By optimality, it holds that λ1 = e−ıθ1 Z(t1 ).
Then, the case k = 2 is interesting since λ2 is a statistic used in the test statistics described
previously. We will see that the two definitions agree here again. For k = 2, it holds Z1 = Z
and the least squares direction is given by a = Z(t1 )/(2σ 2 ) and Z (λ) by
λ
K(t1 , t)
− 1)Z(t1 )
,
λ1
2σ 2
K(t1 , t)
= Z(t) + eıθ1 (λ − λ1 )
2σ 2
Z (λ) (t) = Z(t) + (
Multiplying by e−ıθ and taking the real part, this latter can be equivalently written as
Re (e−ıθ Z (λ) (t)) = X(z) + (λ − λ1 ) cos(θ1 − θ)
K(t1 , t)
,
2σ 2
where z = (t, θ) ∈ T. Now, recall that ρ(t, θ) := Γ(t) cos θ = cos(θ)K(0, t)/(2σ 2 ) to compute
Re (e−ıθ Z (λ) (t)) = X(z) + (λ − λ1 )ρ(z − zb) .
We deduce that
Re (e−ıθ Z (λ) (t)) ≤ λ ⇔ X(z) − λ1 ρ(z − zb) ≤ λ(1 − ρ(z − zb))
X(z) − X(b
z )ρ(z − zb)
⇔
≤λ
1 − ρ(z − zb)
⇔ X zb(z) ≤ λ
showing that the second knot λ2 is exactly the quantity defined in (3).
6.4. Illustration: The two first knots of Super-Resolution
The Super-Resolution process is defined in (11). It satisfies Condition (KLZ (N )) and Condition (NDZ (N )) with N = 2fc +1. The first point is given by the maximum of the modulus of Z,
see the red curve in Figure 5. Observe that Z1 = Z and the maximum satisfies Z1 (t1 ) = λ1 eıθ1 .
Then, we compute
λ
DN (t1 − t)
Z (λ) (t) = Z1 (t) + ( − 1)Z1 (t1 )
,
λ1
2N σ 2
where DN denotes the Dirichlet kernel. For λ > λ2 , the maximum of |Z (λ) | is achieved at a
unique point, namely t1 . For λ = λ2 , a second point achieves the maximum. This transition
defines Z2 := Z (λ2 ) , see Figure 5.
From this point, we can iterate fitting the least squares direction on the support {t1 , t2 } and
decreasing |Z2 | while a third point achieves the maximum.
Given the red curve in Figure 5, it was not obvious that the second knot would have been t2
since other local maxima seemed more significant than t2 on the red curve.
Azaïs, De Castro and Mourareau/A
PROOFS
26
Figure 5. LARS for Super-Resolution: we fit a Dirichlet kernel at the maximal correlation point t1 until the
maximal correlation in the residual is matched at a second point t2 =
6 t1 .
Appendix A: Proofs
We denote for random variables, Xn = oP (rn ) and Yn = OP (rn ) (for rn 6= 0) means that
rn−1 ||Xn || converges to 0 in probability and rn−1 ||Yn || is uniformly tight, respectively. Furthermore, we consider the following processes.
• The stationary process X(z) = X(t, θ) defined on T with covariance function given
by Cov(X(y), X(z)) = σ 2 ρ(z − y) where we recall the correlation function is given by
ρ(z − y) = cos(θ − α)Γ(t − s),
• For every z ∈ T, recall the regressions with respect to X(z)
X(y) − X(z)ρ(z − y)
X(y) − X(z)
= X(z) +
,
1 − ρ(z − y)
1 − ρ(z − y)
X(y) − ρ(z − y)X(z)
z
p
Xnorm
(y) =
.
1 − ρ2 (z − y)
∀y ∈ T \ {z}, X z (y) =
• For every z ∈ T, recall the regressions with respect to (X(z), X 0 (z))
e −1 X 0 (z)i
X(y) − ρ(z − y)X(z) + hρ0 (z − y), Λ
,
1 − ρ(z − y)
e −1 X 0 (z)i
X(y) − ρ(z − y)X(z) + hρ0 (z − y), Λ
|z
Xnorm
(y) = q
.
e −1 ρ0 (z − y)i
1 − ρ2 (z − y) − hρ0 (z − y), Λ
∀y ∈ T \ {z}, X |z (y) =
In particular, recall that zb is defined by (2) so X 0 (b
z ) = 0 and it yields that X zb = X |bz .
A.1. Proof of Theorem 3
Since the variance is known, we consider without loss of generality that σ 2 = 1. Using the
metric given by the quadratic form represented by X 00 , we can consider the closest point z n
of the grid Gn to zb by
z n = arg min kb
z − ukX 00 = arg min hb
z − u, −X 00 (b
z )(b
z − u)i .
u∈Gn
u∈Gn
27
A PROOFS
The main claim is that, while it holds λ1,n → λ1 a.s., we don’t have the same result for λ2,n ,
see Lemma 14. We begin with the following preliminary result, which is related to the result
of Azaïs-Chassan (see [9]).
Lemma 12. Conditionally to X 00 , ∆−1
z − z n ) follows a uniform distribution on V0 and this
n (b
distribution is independent from λ1 and λ2 .
Proof. Remark that zb has uniform distribution on T by stationarity and this distribution is
independent from λ1 and λ2 . Let B be a Borelian in R2 . Remark that zb − z n ∈ ∆n V0 by
definition of z n and note that z n ∈ ∆n Z2 . Conditionally to X 00 , it holds
P{∆−1
z − z n ) ∈ B} = P{∆−1
z − z n ) ∈ B ∩ V0 }
n (b
n (b
= P{(b
z − z n ) ∈ ∆n (B ∩ V0 )}
= P{b
z ∈ ∆n (B ∩ V0 + Z2 )} .
Since zb has uniform distribution on T and since V0 + Z2 is a partition of R2 , it holds that
P{b
z ∈ ∆n (B ∩ V0 + Z2 )} = P{b
z ∈ 2π(B ∩ V0 + Z2 )} =
Leb(R2 )(2π(B ∩ V0 ))
,
Leb(R2 )(2πV0 )
where Leb(R2 ) denotes the Lebesgue measure on R2 .
Lemma 13. It holds that
(a) X(b
zn ) − X(z n ) = oP (∆2n ).
(b) P{b
zn =
6 z n } → 0 as n goes to ∞.
(c) Let F be any measurable function, then F (b
zn ) − F (z n ) tends to zero in probability at
arbitrary speed.
(d) Almost surely, one has z n → zb and zbn → zb as n goes to infinity.
Proof. Let ε > 0. By definition of z n and since V0 ⊂ [−1, 1]2 , it holds that
√
||b
z − z n || 6 2∆n ,
(17)
almost surely. Since X has C 2 -paths and by Taylor expansion, one has
X(b
z ) − X(z n ) = (1/2)||b
z − z n ||2X 00 + oP (∆2n )
(18)
Since −X 00 is positive definite, there exists M > 0 sufficiently large such that
(1/M )Id2 4 −X 00 4 M Id2
where 4 denotes the Lowner ordering between symmetric matrices. Then, it holds
∀z ∈ R2 ,
(1/M )||z||2 ≤ ||z||2X 00 6 M ||z||2 .
(19)
From (17), (18) and (19), we deduce that
0 6 X(b
z ) − X(b
zn ) 6 X(b
z ) − X(z n ) = OP (∆2n ) ,
using the optimality of zb and zbn .
(20)
Azaïs, De Castro and Mourareau/A
PROOFS
28
By compactness of T, uniqueness of optimum zb ∈ R2 and C 2 -continuity of X, there exists η > 0 and a neighborhood N0 ⊂ R2 of zb ∈ R2 such that X(b
z ) − η > X(z) for any z ∈
/ N0
and
∀z ∈ N0 , (1/4)||b
z − z||2X 00 6 X(b
z ) − X(z) 6 ||b
z − z||2X 00
(21)
using again a Taylor expansion as in (18). Using (20), it holds that, on an event of probability
at least 1 − ε/4 and for n large enough, 0 6 X(b
z ) − X(b
zn ) ≤ η/2 implying that zbn ∈ N0 .
Invoke (19), (20) and (21) to deduce that zb − zbn = OP (∆n ).
Using Taylor formula again, we get that
X(b
z ) − X(b
zn ) = (1/2)||b
z − zbn ||2X 00 + oP (∆2n ) .
(22)
By optimality of zbn and z n and using (18) and (22), one gets
0 6 X(b
zn ) − X(z n ) ≤ (1/2)(||b
z − z n ||2X 00 − ||b
z − zbn ||2X 00 ) + oP (∆2n ) .
(23)
z − zbn ||2X 00 ≤ 0, we get (a).
Observing that ||b
z − z n ||2X 00 − ||b
00
Conditionally to X and in the metric defined by || · ||X 00 , there exists η 0 > 0, such that
the η 0 -neighborhood, denoted by Nη0 , of the boundary ∂V0 of V0 has relative volume (for the
Lebesgue measure) less than ε/8. More precisely, Nη0 denotes the set of points in V0 ⊂ R2
with || · ||X 00 -distance less than η 0 to the boundary of V0 . In particular,
∀k ∈ Z2 \ {0}, ∀z ∈ V0 \ Nη0 ,
||z||X 00 + η 0 6 ||z − k||X 00 ,
by Cauchy-Schwarz inequality. Using Lemma 12 and by homogeneity, we deduce that it holds
||b
z − z n ||X 00 + η 0 ∆n 6 ||b
z − g||X 00 ,
∀g ∈ ∆n Z2 \ {0},
with probability at least 1 − ε/8. It follows that
z − g||2X 00 ,
||b
z − z n ||2X 00 + (η 0 )2 ∆2n 6 ||b
∀g ∈ ∆n Z2 \ {0},
using that (a + b)2 > a2 + b2 for a, b > 0. Now, invoke (23) to get that
06−
(η 0 )2 2
∆n 1{z n 6=zbn } + oP (∆2n ) .
2
On these events, we get that, for n sufficiently large, zbn and z n must be equal except on an
event of probability at most ε/4 + ε/8 ≤ ε. Furthermore, this result holds unconditionally
in X 00 . We deduce that lim sup P{z n 6= zbn } 6 ε, proving (b). Note that (c) is a consequence of
the fact that, for n sufficiently large, zbn and z n must be equal except on an event of arbitrarily
small size. In particular, it shows that supk>n ||z k − zbk || converges towards zero in probability,
which is equivalent to almost sure converge of z n − zbn towards zero. Claim (d) follows when
remarking that (17) proves a.s. convergence of z n towards zb.
Lemma 14. As n tends to infinity, λ2,n converges in distribution to λ2 .
Proof. Let β ∈ R be such that 0 < β < 1/2, say β = 1/4. Let ε ∈ (0, 1). We can write
λ2,n = λA,n ∨ λB,n with
λA,n :=
max
u∈Gn \{b
zn } s.t.
λB,n :=
max
u∈Gn s.t.
ku−b
zn k6∆β
n
ku−b
zn k>∆β
n
X zbn (u) =: max X zbn (u),
u∈Gn,A
X zbn (u) =: max X zbn (u).
u∈Gn,B
29
A PROOFS
We first prove that λB,n → λ2 as n tends to infinity in distribution. By compactness, remark
that there exists a constant Cr > 0 such that
1 − ρ(u) > Cr kuk2 .
(24)
It also holds that
X(u) − X(b
zn )
,
1 − ρ(u − zbn )
X(u) − X(b
z)
X zb(u) = X(b
z) +
.
1 − ρ(u − zb)
X zbn (u) = X(b
zn ) +
(25)
(26)
Let us look to the rhs of (25) and (26). By Claim (d) of Lemma 13 and the continuous
mapping theorem, note that X(b
zn ) converges toward λ1 = X(b
z ) a.s. and we can omit these
terms. It remains to prove that on Gn,B the second terms are equivalent. Because of Lemma 13,
1 − ρ(u − zbn ) converges to 1 − ρ(u − z n ) at arbitrary speed. Remember that (17) gives zb − z n =
OP (∆n ) and it holds that ||u − zb|| > ∆βn , on Gn,B . It follows that there exists C > 0 such that
1 − ρ(u − zbn ) > C∆2β
n
and 1 − ρ(u − zb) > C∆2β
n ,
(27)
with probability greater than 1 − ε/2. As for the numerators, Eqs. (20) and (21) show that
for all u ∈ Gn,B
X(u) − X(b
zn )
X(b
z ) − X(b
zn )
|X(b
z ) − X(b
zn )|
−1 =
= (cst)
= OP (∆2−2β
)
n
2
X(u) − X(b
z)
X(u) − X(b
z)
||u − zb||
In this sense, we say that X(u)−X(b
zn ) is uniformly equivalent to X(u)−X(b
z ) on the grid Gn,B
in probability. Using (27) and noticing that for any u ∈ T
|ρ(u − zb) − ρ(u − zbn )| 6 ||ρ0 ||∞ ||b
z − zbn || = OP (∆n ) ,
the same result holds for the denominators, namely 1 − ρ(u − zbn ) is uniformly equivalent to
1 − ρ(u − zb) on the grid Gn,B in probability. We deduce that X zbn (u) is uniformly equivalent
to X zbn (u) on the grid Gn,B in probability and, passing tho their maximum, one can deduce
that λB,n converges to λ2 in probability.
We turn now to the study of the local part λA,n . Again, by Claim (c) of Lemma 13 we can
replace zbn by z n in the numerator of the r.h.s in (25) and we forget the first term which limit
is clearly λ1 almost surely. We perform a Taylor expansion at zb, it gives that
X(u) − X(b
z ) = (1/2)(u − zb)> X 00 (u − zb)(1 + oP (1)).
for any u ∈ Gn,A . Since z n − zb = OP (∆n ), we also get that
X(z n ) − X(b
z ) = (1/2)(z n − zb)> X 00 (z n − zb) + oP (∆2n ).
As for the denominator, invoke (17), (24) and Claim (c) of Lemma 13 to get that
1 − ρ(u − z n ) > 2Cr ∆2n ,
1 − ρ(u − zbn ) = 1 − ρ(u − z n ) + oP (∆2n ),
e − z n ) (1 + oP (1)),
and 1 − ρ(u − zbn ) = (1/2) (u − z n )> Λ(u
Azaïs, De Castro and Mourareau/A
PROOFS
30
e denotes the Hessian at point 0 of ρ. Putting all together yields
where −Λ
(u − z n )> X 00 (u + z n − 2b
z)
X(u) − X(b
zn )
=
(1 + oP (1)),
>
e
1 − ρ(u − zbn )
(u − z n ) Λ(u − z n )
for any u ∈ Gn,A . Now we know that, in distribution, zb − z n = ∆n U and we know that
u − z n = k∆n with k belonging to a certain growing subset of Z2 which limit is Z2 . Finally,
conditionally to X 00 , we obtain that
max
u∈Gn,A \{b
zn }
X zbn (u) −→ λ1 +
k>
(k − 2U)
X 00
,
1
e 2 kk
e 21 kk
k∈Z2 \{0} kΛ
kΛ
sup
in distribution.
Eventually,
consider the test statistic Sn := Φ(λ1,n )/Φ(λ
2,n ) and keep in mind that X u +
(0, π) = −X(u) and that if u belongs to Gn , u + (0, π) also belongs. So Theorem 1 of [4]
applies showing that, under the alternative, P{Sn 6 α} > α. It suffices to pass to the limit to
get the desired result.
A.2. Proof of Theorem 4
We use the same grid argument as for the proof of Theorem 3.
Let t1 , t2 , . . . , tN be pairwise distinct points of [0, 2π), θ1 ∈ [0, 2π), m = 2N and set
z1 = (t1 , θ1 ), . . . , zN = (tN , θ1 ), zN +1 = (t1 , θ1 + π/2), . . . , zm = (tN , θ1 + π/2).
Because of the first assumption of (NDZ (N )), the distribution of (X(z1 ), . . . , X(zm )) is
non degenerated. Consequently, following the proof of Proposition 2, we know that X z1 satisfies KL(m − 1) and ND(m − 1). Denote g1 , . . . , gm−1 the eigenfunctions of the KarhunenLoève (KL) representation of X (0,0) . Note that X z1 (·) has the same distribution as X (0,0) (.−z1 )
(stationarity) and that both are defined on the same space so the KL-eigenfunctions of X z1
are g1 (. − z1 ), . . . , gm−1 (. − z1 ).
Now consider Az1 = (Azi,j1 )1≤i,j≤m−1 the matrix with entries Azi,j1 = gi (zj+1 − z1 ) which is
invertible thanks to KL(m − 1) and ND(m − 1) and build so that
X z1 (z2 )
ζ1
..
.
z
= A 1 .. .
.
X z1 (zm )
ζm−1
2 (X zbn (G )), the estimator of σ 2
One possible explicit expression, among many others, of σ
bKL
n
norm
on the grid Gn , is
X (0,0) (z2 − zbn )
..
.
(0,0)
X
(zm − zbn )
2
zbn
σ
bn2 := σ
bKL
(Xnorm
(Gn )) =
1
m−1
Azbn
−1
2
,
2
which is a composition of continuous functions of zbn . In particular, as zbn converges a.s. to zb
(see Lemma 13, Claim (b)), we deduce that σ
bn2 converges a.s. to σ
bzb2 as n goes to infinity.
31
A PROOFS
Finally, since the KL estimator is unique, this estimator coincide with the estimator σ
b22
of [4] and Theorem 3 of [4] implies that
F m−1 λ1,n /b
σn
∼ U([0, 1]).
F m−1 λ2,n /b
σn
Note that λn1 converges almost surely to λ1 and λ2,n converges in distribution to λ2 (see
Lemma 14) to complete the proof.
A.3. Proof of Proposition 9
(a). We can assume that Z defined by (11) is centered and, in this case, it holds
fc
σ X
ζk exp(ıkt),
∀t ∈ [0, 2π) Z(t) = √
N k=−fc
(28)
where we recall that N = 2f c + 1 and ζk = ζk,1 + ıζk,2 for k = −fc , . . . , fc are independent
standard complex Gaussian variables. Formula (28) shows that Z satisfies (KLZ (N )).
(b). Let (t1 , . . . , tN ) ∈ [0, 2π) be pairwise differents, θ ∈ [0, 2π) and set
Z(t1 )
exp(−ıfc t1 ) . . . exp(ıfc t1 )
ζ1
..
..
..
..
=
. =: At1 ,...,tN ζ,
.
.
.
Z(tN )
exp(−ıfc tN ) . . .
exp(ıfc tN )
ζN
where At1 ,...,tN is a Vandermonde matrix, invertible as soon as ti 6= tj for all i 6= j. This
prove the first point of NDZ (N ). For the second assertion, consider h > 0 such that h <
min1≤i<j≤N −1 (ti − tj ) and the Gaussian vector
(Z(t1 ), . . . , Z(tN −1 ), Z(t1 + h))T =: At1 ,...,tN −1 ,t1 +h ζ,
where the covariance matrix At1 ,...,tN −1 ,t1 +h satisfies
Y
det(A∗t1 ,...,t1 +h At1 ,...,t1 +h ) =
1≤i<j≤N −1
N (N −1)/2
=4
| exp(ıti ) − exp(ıtj )|2
Y
2
sin
1≤i<j≤N −1
N
−1
Y
j=1
ti − tj
2
| exp(ı(t1 + h)) − exp(ıtj )|2
NY
−1
j=1
2
sin
tj − (t1 + h)
2
2
= sin (h/2) × gt1 ,...,tN −1 (h)
where gt1 ,...,tN −1 (0) 6= 0 if (ti )1≤i≤N −1 are pairwise distincts. Finally, denote by Rh the linear
transformation involving the first and the last coordinate such that
Z(t1 )
Z(t1 )
..
..
.
.
=
R
h
Z(tN −1 )
Z(tN −1 )
Z(t1 +h)−Z(t1 )
Z(t1 + h)
h
and remark that
sin2 (h/2)
1
= gt1 ,...,tN −1 (0) × 6= 0
2
h→0
h
4
lim det(A∗t1 ,...,t1 +h Rh∗ Rh At1 ,...,t1 +h ) = gt1 ,...,tN −1 (0) × lim
h→0
giving the desired non degeneracy condition.
Azaïs, De Castro and Mourareau/B
AUXILIARY RESULTS
32
A.4. Proof of Proposition 10 and Proposition 11
Easy computations give the following results for φ(·),
Z
Z +∞
Z +∞
φ(t)dt = Φ(`),
tφ(t)dt = φ(`),
`
+∞
t2 φ(t)dt = `φ(`) + Φ(`),
`
`
for fm−1 (·),
Z
+∞
fm−1
`
+∞
Z
tfm−1
`
Z
!
r
m−3
m−1
t
dt =
F m−1 `
,
m−1
m−3
!
r
√
Γ m−3
(m − 3) m − 3 Γ m
m−1
2
2
fm−3 (`),
√
t
dt =
m−3
(m − 2) m − 1 Γ m−1
Γ m−2
2
2
!
r
√
Γ m−3
m−1
(m − 3) m − 3 Γ m
2
2
× `fm−3 (`) + F m−3 (`) ,
√
t
dt =
m−1
m−2
m−3
(m − 2) m − 1 Γ 2 Γ 2
r
+∞
t2 fm−1
`
m−1
m−3
!
r
and for R,
e z ) + R(b
X 00 (b
z ) = −ΛX(b
z ),
−α2 α3
α1 0
X(b
z) +
,
=−
0 1
α3
0
where
α1 = 31 fc (fc + 1),
fc
b
α = √1 P (k 2 − α ) × Re(y eı(kbt−θ)
),
1
2
k
N
k=−fc
fc
P
b
1
k × Re(yk eı(kbt−θ) ).
α3 = √N
k=−fc
To conclude, use Proposition 9 to apply Theorem 6 and Theorem 8.
Appendix B: Auxiliary results
B.1. Regularity of X |z and new expression of R(z)
Lemma 15. X |z (y) admits radials limits as y → z. More precisely for all λ in the unit sphere
lim X |z (z + uλ) =
u→0
Proof. As u tends to zero
1 − ρ(uλ) =
λ> R(z)λ
.
e
λ> Λλ
u2 > e
(λ Λλ + o(1)).
2
Moreover, a Taylor expansion gives
X(z + uλ) = X(z) + uXλ0 (z) +
u2 00
X (z) + op (u2 ),
2 λ
33
REFERENCES
and
e + op (u2 ),
ρ0λ (uλ) = uρ00λ (0) + op (u2 ) = −uΛ
where (Xλ0 , ρ0λ ) and (Xλ00 , ρ00λ ) are directional derivative and directional Hessian. By consequence,
|z
X (z + uλ) =
=
which tends to
u2
>e
2 X(z)λ Λλ
u2
2
e −1 X 0 (z)i + uX 0 (z) +
+ hρ0 (uλ), Λ
λ
u2
>
e
2 (λ Λλ + o(1))
e + X 00 (z) + op (1)
X(z)λ> Λλ
λ
u2
>e
2 (λ Λλ
u2 00
2 Xλ (z)
+ op (u2 )
+ o(1))
e
λ> ΛX(z)
+ X 00 (z) λ
e
λ> Λλ
e
as u tends to 0 since Xλ00 (z) = λ> X 00 (z)λ. The result follows from X 00 (z) = −ΛX(z)+R(z).
B.2. Maximum of a continuous process
The following result is borrowed from [17, Theorem 3] and [23].
Proposition 16. Let {Y (t) ; t ∈ T } be a Gaussian process with continuous sample paths
defined on a compact metric space T . Suppose in addition that:
There is no two different points s, t ∈ T such that X(s) = X(t) a.s.
(29)
Then almost surely the maximum of X on T is attained at a single point.
Observe that (Anorm ) implies (29).
Remark 7. Proposition 16 can be applied to the process X |bz which is not continuous on a
compact set. We use the “pumping method” as follows. Use
(a) a parameterization of T as [0, 2π)2 ,
(b) polar coordinates for y ∈ T \ {b
z } with origin at zb,
(c) the change of parameter
y = (ρ, θ) 7→ ((ρ + 1), θ)
that transforms the non-compact set T \ {b
z } into a compact set (we have inflated the
“hole” {b
z } into a ball centered around zb with radius one) on which the process X |bz is
continuous thanks to Lemma 15.
References
[1] M. Abramowitz and I. Stegun. Handbook of Mathematical Functions. Dover, New York,
fifth edition, 1964.
[2] R. J. Adler and J. E. Taylor. Random fields and geometry. Springer Science & Business
Media, 2009.
[3] J.-M. Azaïs, Y. De Castro, and F. Gamboa. Spike detection from inaccurate samplings.
Applied and Computational Harmonic Analysis, 38(2):177–195, 2015.
REFERENCES/REFERENCES
34
[4] J.-M. Azaïs, Y. De Castro, and S. Mourareau. Power of the Spacing test for Least-Angle
Regression. Bernoulli, 2016.
[5] J.-M. Azaïs and M. Wschebor. Level sets and extrema of random processes and fields.
John Wiley & Sons Inc., 2009.
[6] T. Bendory, S. Dekel, and A. Feuer. Robust recovery of stream of pulses using convex
optimization. Journal of Mathematical Analysis and Applications, 442(2):511–536, 2016.
[7] K. Bredies and H. K. Pikkarainen. Inverse problems in spaces of measures. ESAIM:
Control, Optimisation and Calculus of Variations, 19(01):190–218, 2013.
[8] E. J. Candès and C. Fernandez-Granda. Towards a Mathematical Theory of Superresolution. Communications on Pure and Applied Mathematics, 67(6):906–956, 2014.
[9] M. Chassan and J.-M. Azais. Discretization error for the maximum of a Gaussian field.
Technical report, HAL, 2017. working paper or preprint.
[10] H. Cramer and M. Leadbetter. Stationary and Related Stochastic Processes (1967).
[11] Y. De Castro and F. Gamboa. Exact reconstruction using beurling minimal extrapolation.
Journal of Mathematical Analysis and applications, 395(1):336–354, 2012.
[12] V. Duval and G. Peyré. Exact support recovery for sparse spikes deconvolution. Foundations of Computational Mathematics, pages 1–41, 2015.
[13] C. Fernandez-Granda. Support detection in super-resolution. In 10th international conference on Sampling Theory and Applications (SampTA 2013), pages 145–148, Bremen,
Germany, July 2013.
[14] C. Fernandez-Granda. Super-resolution of point sources via convex programming. Information and Inference, page iaw005, 2016.
[15] E. Giné and R. Nickl. Mathematical foundations of infinite-dimensional statistical models,
volume 40. Cambridge University Press, 2015.
[16] T. Hastie, R. Tibshirani, and M. Wainwright. Statistical learning with sparsity: the lasso
and generalizations. CRC Press, 2015.
[17] M. A. Lifshits. On the absolute continuity of distributions of functionals of random
processes. Theory of Probability & Its Applications, 27(3):600–607, 1983.
[18] R. Lockhart, J. Taylor, R. J. Tibshirani, and R. Tibshirani. A significance test for the
lasso. Annals of statistics, 42(2):413, 2014.
[19] S. Rosset, G. Swirszcz, N. Srebro, and J. Zhu. `1 regularization in infinite dimensional
feature spaces. Lecture Notes in Computer Science, 4539:544, 2007.
[20] G. Tang, B. N. Bhaskar, P. Shah, and B. Recht. Compressed sensing off the grid. Information Theory, IEEE Transactions on, 59(11):7465–7490, 2013.
[21] J. E. Taylor, J. R. Loftus, and R. J. Tibshirani. Inference in adaptive regression via the
Kac–Rice formula. Ann. Statist., 44(2):743–770, 04 2016.
[22] R. J. Tibshirani, J. Taylor, R. Lockhart, and R. Tibshirani. Exact post-selection inference
for sequential regression procedures. Journal of the American Statistical Association,
111(514):600–620, 2016.
[23] V. Tsirel’Son. The density of the distribution of the maximum of a Gaussian process.
Theory of Probability & Its Applications, 20(4):847–856, 1976.
| 10math.ST
|
Obstacle avoidance via B-spline parameterizations of flat
trajectories
arXiv:1603.04911v1 [cs.SY] 15 Mar 2016
Florin Stoican† , Vlad-Mihai Ivănuşcㆠ, Ionela Prodan‡
Abstract— This paper considers the collision avoidance problem in a multi-agent multi-obstacle framework. The originality in solving this intensively studied problem resides in the proposed geometrical view
combined with differential flatness for trajectory generation and B-splines for the flat output parametrization. Using some important properties of these theoretical tools we show that the constraints can be validated at all times. Exact and sub-optimal constructions of the collision avoidance optimization problem
are provided. The results are validated through extensive simulations over standard autonomous aerial
vehicle dynamics.
Index Terms— Obstacle avoidance, Flat trajectory,
B-spline basis, MIP (Mixed-Integer Programming),
UAV (Unmanned Aerial Vehicle).
I. Introduction
One of the main issues in multi-obstacle, multi-agent
environments is the collision avoidance assessment. Usually the avoidance constraints have to be considered
both between an agent and fixed obstacles and between
any two agents. The problem is intensively studied, yet
actual, in the literature but is usually tilted towards
heuristic approaches or online validations. The first usually lacks stability and performances guarantees and the
later is easily boggled into numerical issues [1], [2].
An alternative approach is to solve offline the difficult
trajectory generation part of the overall problem, online
only a straightforward trajectory tracking is employed.
This reduces significantly the online computations and
allows stability and performance analysis. The caveat
is that the trajectory to be computed has to respect
the dynamics of the agent and to validate the collision
avoidance constraints at all times [3], [4].
An interesting implementation is represented by flat
trajectory design which guarantees that the corresponding system dynamics are respected (with the caveat that
state and input constraints are not easily accounted for
[5], [6]). This construction shifts the state and input
constraints into constraints over the flat output. To
handle this, usually, the flat output is project over some
basis functions which means that only the coefficients
†
Politehnica
University
of
Bucharest,
Department
of
Automation
Control
and
Systems
Engineering,
Bucharest,
Romania
[email protected],
of the projections need to be found. In this sense, Bspline functions represent an ideal choice since they have
enough flexibility [7] and nice theoretical properties (of
which we will make extensive use throughout the paper).
The present work builds on results sketched in [8]
and further advances the topic in several directions.
Foremost, we provide exact and sub-optimal formulations of the collision avoidance problems between an
agent and the obstacles and between any tho agents. In
both cases we make use of the geometrical properties of
the B-spline functions which allow to bound locally the
trajectories obtained through them. Hence, the collision
problems become separation problems between sets of
consecutive points. In the exact case, these constraints
lead to nonlinear formulation where both the control
points and the separation hyperplanes are variables. A
simplified (and hence sub-optimal approach) is to select
the separation hyperplanes from the support hyperplanes
of the obstacles therefore reducing the problem to a
mixed integer formulation.
Both methods make use of a multi-obstacole framework and are tested and compared over extensive simulations. The rest of the paper is organized as follows:
Section II tackles flat output and B-spline characterizations, Section III discusses flat trajectory generation and
Section IV presents the main results of the paper which
are then illustrated in Section V. Section VI draws the
conclusions.
Notation
The Minkowski sum of two sets, A and B is
denoted as A ⊕ B = {x : x = a + b, a ∈ A, b ∈ B}.
Convex Hull{p1 . . . pn } denotes the convex hull of set
generated by the collection of points p1 . . . pn .
II. Preliminaries
The problem of designing reference trajectories in a
multi-agent multi-obstacle environment is in general a
difficult one. A popular approach is to parametrize them
through flatness constructions [9], [10], [5]. In this section
we will describe some of the basics of flat trajectory and
their parametrization via B-spline basis functions.
A. Flat trajectories
[email protected]
‡ , Univ. Grenoble Alpes, Laboratory of Conception and Integration of Systems (LCIS EA 3747), F-26902, Valence, France,
[email protected]
A nonlinear time invariant system:
ẋ(t) = f (x(t), u(t)),
(1)
where x(t) ∈ Rn is the state vector and u(t) ∈ Rm is the
input vector is called differentially flat if there exists the
flat output z(t) ∈ Rm :
z(t) = γ(x(t), u(t), u̇(t), · · · , u(q) (t))
=
Θ(z(t), ż(t), · · · , z (q) (t)),
u(t)
=
Φ(z(t), ż(t), · · · , z (q) (t)).
(3)
Remark 1. For any system admitting a flat description,
the number of flat outputs equals the number of inputs
[10]. In the case of linear systems [11] the flat differentiability (existence and constructive forms) is implied by
the controllability property.
Within the multi-agent framework, the most important aspect of construction (2)–(3) is that it reduces the
problem of trajectory generation to finding an adequate
flat output (2). This means choosing z(t) such that, via
mappings Θ(·), Φ(·), various constraints on state and inputs (3) are verified. Since the flat output may be difficult
to compute under these restrictions, we parametrize z(t)
using a set of smooth basis functions Λi (t):
z(t) =
N
X
αi Λi (t), αi ∈ R.
z(t) =
n
X
Bi,d (t)pi = PBd (t)
(8)
i=0
(2)
such that the states and inputs can be algebraically
expressed in terms of z(t) and a finite number of its
higher-order derivatives:
x(t)
control points (7) and the B-spline functions (6a)–(6b)
T
where P = p0 . . . pn and Bd (t) = B0,d (t) . . . Bn,d (t) .
This construction yields several interesting properties
[15]:
P1) z(t) is C ∞ in any t ∈
/ T and C d−1 in any t ∈ T;
P2) at a time instant τi < t < τi+1 , z(t) depends
only on the B-splines Bi−d+1,d (t) . . . Bi,d (t); consequently, the B-spline curve z(t) lies within the
union of all convex hulls formed by all d successive
control points;
P3) the ‘r’ order derivatives of B-spline basis functions can be expressed as linear combinations of
(r)
B-splines of lower order (Bd (t) = Mr Bd−r (t)
with matrices Mr of appropriate dimensions and
content);
P4) taking the first and last d knot elements equal
(τ0 = · · · = τd−1 and τn+1 = . . . τn+d ) leads to
a clamped B-spline curve where the first and last
control points coincide with the curve’s end points.
Illustrative example
(4)
i=1
Parameter N depends on the number of constraints
imposed onto the dynamics [12].
There are multiple choices for the basis functions Λi (t).
Among these, B-spline basis functions are well-suited to
flatness parametrization due to their ease of enforcing
continuity and because their degree depends only up to
which derivative is needed to ensure continuity [7], [5].
Let us consider for exemplification the B-spline basis
functions defined by parameters n = 5 and d = 4 with the
knot vector of length m = n+d = 9 and with components
equally sampled between 0 and 1. Using these elements
we construct the B-spline curve (8) which we depict in
Fig. 1. Note that the convex-hulls determined by all d = 4
p2
B. B-splines
p1
A B-spline of order d is characterized by a knot-vector
[13], [14]
T = {τ0 , τ1 . . . τm } ,
(5)
of non-decreasing time instants (τ0 ≤ τ1 ≤ · · · ≤ τm )
which parametrizes the associated basis functions Bi,d (t):
(
1, for τi ≤ t < τi+1
,
(6a)
0 otherwise
τi+d − t
t − τi
Bi,d−1 (t) +
Bi+1,d−1 (t)
Bi,d (t) =
τi+d−1 − τi
τi+d − τi+1
(6b)
Bi,1 (t) =
for d > 1 and i = 0, 1 . . . n = m − d.
Considering a collection of control points
P = {p0 , p1 . . . pn } ,
p0
p3
p4
ConvHull{p0, p1, p2, p3}
p5
p6
Fig. 1: B-spline curve (solid red) and its control polygone
(dashed blue).
consecutive control points (Property II-B) contain the
curve:
[
z(t) ⊂
Convex Hull{pi−3 , pi−2 , pi−1 , pi }, ∀t ∈ [0, 1].
i=3...5
(7)
we define a B-spline curve as a linear combination of the
In particular, the patterned shape depicted in the figure
denotes region Convex Hull{p2 , p3 , p4 , p5 } which contains
the curve z(t) for any t ∈ [τ2 , τ3 ].
III. Flat trajectory generation
Let us consider a collection of N + 1 way-points and
the time stamps associated to them1 :
W = {ws } and TW = {ts },
(9)
for any s = 0 . . . N . The goal is to construct a flat
trajectory which passes through each way-point ws at the
time instant ts (or through a predefined neighborhood of
it [8]), i.e., to find a flat output z(t) such that
x(ts ) = Θ(z(ts ), . . . z (r) (ts )) = ws , ∀s = 0 . . . N.
Making use of the B-spline framework we provide a
vector of control points (7) and its associated knot-vector
(5) such that (10) is verified (parameter d is chosen such
that continuity constraints are respected):
(11)
where Θ̃(Bd (t), P) = Θ(PBd (t) . . . PMr Bd−r (t)) is constructed along Property II-B.
Let us assume that the knot-vector is fixed (τ0 = t0 ,
τn+d = tN and the intermediary points τ1 . . . τn+d−1 are
equally distributed along these extremes). Then, we can
write an optimization problem with control points pi
as decision variables whose goal is to minimize a cost
Ξ(x(t), u(t)) along the time interval [t0 , tN ]:
Z tN
P = arg min
||Ξ̃(Bd (t), P)||Q dt
P
(12)
t0
s.t. constraints (11) are verified
with Q a positive symmetric matrix. The cost
Ξ̃(Bd (t), P) = Ξ(Θ̃(Bd (t), P), Φ̃(Bd (t), P))
can impose any penalization we deem necessary (length
of the trajectory, input variation, input magnitude, etc).
In general, such a problem is nonlinear (due to mappings
Θ̃(·) and Φ̃(·)) and hence difficult to solve. A nonlinear
MPC iterative approach has been extensively studied [5].
With these tools at hand we can propose various
methods for collision avoidance in a multi-agent multiobstacle environment.
IV. Main idea
Let us consider a collection of polyhedral obstacles
O = {O1 . . . ONo }
(13)
and assume that the k-th agent follows a trajectory rk (t)
during the interval [t0 , tN ], generated as in (10) through a
1 In
(ii) collision avoidance between the k1 -th and k2 -th
agents (for any k1 6= k2 ):
(10)
Remark 2. Note that here we assume that the way-points
are defined over the entire state. Arguably there might be
situations where only a subspace of the state is of interest
(e.g., only the position components of the state).
Θ̃(Bd (ts ), P) = ws , ∀s = 0 . . . N,
collection of control points2 Pk = {pkj } and the associated
knot vector Tk = {τjk }.
Consequently, the collision avoidance conditions mentioned earlier can be formulated as follows:
(i) collision avoidance between the k-th agent and l-th
obstacle:
rk (t) ∈
/ Ol , ∀t ∈ [t0 , tN ],
(14)
particular we may consider just two way-points: the start and
end points of the trajectory.
rk1 (t) 6= rk2 (t), ∀t ∈ [t0 , tN ].
(15)
A. The exact case
The distinctive feature of conditions (14)–(15) is that
they require a continuous time interval ([t0 , tN ]) validation (i.e., imposing constraints at discrete time instants
tk along the interval is not deemed sufficient). Consequently, we make use of property Property II-B which
allows to bound the continuous B-spline parametrized
curve by its control points. Coupling this with the separating hyperplane theorem (a well-known construction
[16] which states that for any two disjoint convex objects
there exists a separating hyperplane) several results are
attainable.
First, we provide a slight reformulation of Proposition
1 from [17].
Proposition 1. The k-th agent is guaranteed to avoid
obstacles (13), i.e., to verify (14), if there ∃ckil ∈ Rn s.t.
>
>
max
ckil pkj ≤ min
ckil x,
(16)
j∈{i−d+1...i}
x∈Θ̃−1 (Ol )
for i = d − 1 . . . n and ∀Ol ∈ O.
Proof. Condition (16) states that there exists a hyperplane defined by its normal ckil which separates the
points {pki−d+1 . . . pki } from the obstacle Θ̃−1 (Ol ). Since,
according to Property II-B, the curve (10) is contained
in ∪i=d...n Convex Hull{pki−d+1 . . . pki } it follows that (16)
is a sufficient condition to verify (14).
Remark 3. In Proposition 1 note the use of mapping
Θ̃−1 (·). This appears because the obstacle avoidance constraint is the state-space whereas (16) is in the control
point space.
A similar reasoning is employed for the inter-agent
collision condition (15).
Proposition 2. The pair (k1 , k2 ) of agents, with k1 6=
k2 , is guaranteed to avoid collision, i.e., to validate (15),
if there ∃cki11ik22 ∈ Rn s.t.:
>
>
max
cki11ik22
pkj 1 ≤ min
cki11ik22
pkj 2 ,
j∈{i1 −d+1...i1 }
j∈{i2 −d+1...i2 }
(17)
2 For convenience we keep a common degree ‘d’ and number of
control points ‘n’. This is a reasonable assumption as long as the
agents have a common dynamic and follow similar restrictions.
for all possible pairs (i1 , i2 ) which validate
{(i1 , i2 ) : [τik11 , τik11+1 ] ∩ [τik22 , τik22+1 ] 6= ∅}.
(18)
Proof. Recall that (as per Property II-B) a region
Convex Hull{pi−d+1 . . . pi } contains the B-spline curve in
the time interval [τi , τi+1 ]. Applying this to the agents k1
and k2 means that all regions corresponding to indices
(18) should not intersect as they contain overlapping
time instants. The separation is enforced by (17) which
is a sufficient condition for (15).
Remark 4. Eq. (18) can be avoided altogether if the Bspline parametrizations share the same knot vector (i.e.,
Tk1 = Tk2 ). In such a case, variable cki11ik22 becomes cki 1 k2
and condition (17) is simplified to
>
>
cki 1 k2
pkj 2 ,
min
cki 1 k2
pkj 1 ≤
max
j∈{i−d+1...i}
j∈{i−d+1...i}
(19)
for all i ∈ {d − 1 . . . n}.
B. The sub-optimal case
Verifying (16) (or (17)) is difficult in practice due to
the presence of bi-linear terms (e.g., in Proposition 1
both cil and pkj are variables). Hereafter we propose a
simpler (and hence sub-optimal) implementation.
The main idea is that instead of letting the separating
hyperplane from Proposition 1 or 2 be itself a variable,
we choose from within a predefined pool of hyperplanes.
A natural choice is to select from the support hyperplanes of the obstacles. By definition, such a hyperplane
contains on one side the obstacle and hence, it remains
only to check whether the control points lie on the
opposite side. The selection of the active hyperplane is
done through decision variables (i.e., binary variables)
which leads to a mixed-integer pseudo-linear formulation.
To generate the collection of hyperplanes, we consider
the polyhedral sets bounding3 Θ̃−1 (Ol ) and take the
support hyperplanes which characterize them:
Hm = {x : h>
m x = km }, ∀m = 1 . . . M.
(20)
= {x :
= {x :
h>
mx
−h>
mx
≤
km },
(21)
≤ −km }.
(22)
Taking into account all possible combinations of halfspaces leads to a hyperplane arrangement which divides
the space into a collection of disjoint cells which are
completely characterized by sign tuples [18]:
!
M
[
[
[
σ(m)
H=
A(σ) =
Hm
(23)
σ∈Σ
σ∈Σ
1) the admissible domain Rn \ O:
Σ◦ = {σ : A(σ) ∩ O = ∅},
m=1
assume that Θ̃−1 (Ol ) is a polyhedral set as well. In general
this might not hold, but in that case a polyhedral approximation
can be obtained.
(24)
2) or the interdicted domain O:
Σ• = {σ : A(σ) ∩ O 6= ∅},
(25)
where Σ• ∩Σ◦ = ∅ and Σ• ∪Σ◦ = Σ. With these elements
we can provide the following corollaries.
Corollary 1. For an obstacle A(σ • ) with σ • ∈ Σ• , a
sufficient condition to guarantee4 (14) is:
k
•
k
−σ • (m)h>
m pj ≤ −σ (m)km + T αim ,
∀m = 1 . . . M, j = i − d + 1 . . . i
M
X
k
αim
≤M −1
(26a)
(26b)
m=1
for i ∈ {d − 1 . . . n}.
k
Proof. Taking in (26a) he binary variable ‘αim
= 0’
k
means that the i-th region Convex Hull{pi−d+1 . . . pki } of
the k-th agent sits on the opposite side of the obstacle
A(σ • ) with respect to the the m-th hyperplane. The
k
converse, taking ‘αjm
= 1’ means that inequality (26a)
is discarded since the right hand term is sufficiently large
to ignore the values on the left side (assuming that ‘T ’
was taken as a sufficiently large positive constant).
Condition (26b) forces that for any consecutive d + 1
points at least one of the inequalities (26a) is enforced
k
since at least one of the variables αim
has to be zero.
Furthermore, we relax Proposition 2 (in addition to the
simplification proposed in Remark 4) into the following
corollary.
Corollary 2. The pair (k1 , k2 ) of agents, with k1 6= k2 ,
is guaranteed to avoid collision, i.e., to validate (15), if:
max
j∈{i−d...i}
Each of these hyperplanes partitions the space in two
“half-spaces”:
+
Hm
−
Hm
where Σ ⊂ {−, +}M denotes the collection of all feasible
(corresponding to non-empty regions A(σ)) sign tuples.
Each of these tuples can be allocated to either
k1
h>
m pj ≤
M
X
min
j∈{i−d+1...i}
k1 k2
βim
≤M −1
k2
k1 k2
h>
m pj + T βim ,
(27)
(28)
m=1
for i ∈ {d − 1 . . . n}.
k1 k2
Proof. The binary variables βim
denote whether
1
the i-th regions Convex Hull{pki−d+1
. . . pki 1 } and
2
Convex Hull{pki−d+1
. . . pki 2 } are separated through the
k1 k2
m-th hyperplane (whenever βim
= 0 the inequality
(27) is enforced and otherwise is discarded). Eq. (28)
assures that at least one of the hyperplanes is active.
Several remarks are in order.
3 We
4 We make the simplifying assumption that to each obstacle O
l
corresponds a single sign tuple.
Remark 5. Corollary 2 considers a simultaneous computation of trajectories. An alternative is to compute them
iteratively such that from the point of view of the current
agent the obstacles to be avoided at t ∈ [τi , τi+1 ] become:
!
[
O←O∪
Convex Hull{pli−d+1 . . . pli }
(29)
l<k
where besides the obstacles (13), the i-th regions
Convex Hull{pli−d+1 . . . pli } of the previous agents (with
index l < k and whose trajectories are hence already
computed) are also considered as obstacles.
Remark 6. An agent may have a safety region around
it (i.e., because the agent cannot be reduced to a point
or due to the presence of disturbances in the dynamics).
Whatever the reason, and the modality to obtain it,
a safety region Sk can be attached to the k-th agent.
Consequently, the collision avoidance constraints (14)–
(15) become:
{rk (t)} ⊕ Sk ∈
/ Ol , ∀t ∈ [t0 , tN ],
(30)
and
{rk1 (t)} ⊕ Sk1 6= {rk2 (t)} ⊕ Sk2 , ∀t ∈ [t0 , tN ].
(31)
The previous results can be easily adapted to constraints
(30)–(31) by enlarging the obstacles (rk (t) ∈
/ Ol ⊕ {−Sk })
and by requiring a larger inter-distance between agents
respectively (rk1 (t) − rk2 (t) ∈
/ {−Sk1 } ⊕ Sk2 ).
Remark 7. Lastly, it is worth mentioning that in all
previous propositions and corollaries there is no guarantee
of feasibility for the optimization problems. The solution
is to incrementally increase the number of variables (i.e.,
the control points) until a feasible solution is reached.
permits to compute the remaining variables:
ż2 (t)
Ψ(t) = arctan
,
ż1 (t)
q
Va (t) = ż12 (t) + ż22 (t),
!
1 z̈2 (t)ż1 (t) − ż2 (t)z̈1 (t)
p
Φ(t) = arctan
.
g
ż12 (t) + ż22 (t)
(33a)
(33b)
(33c)
Note that in the heading component of the state appear
1st order and in the roll angle input appear 2nd order
derivatives of the flat outputs. Hence, if we wish to
have smooth state and input (their derivatives to be
continuous) it follows that the B-spline parametrization
has to have at least degree d = 4.
Further, we consider way-points which fix only the
position components of the state and time-stamps at
which the trajectory has to pass through them. Thus we
manage to skirt some of the thornier numerical aspects:
the dependence between the B-spline basis functions
and the position components is linear (Θ̃(Bd (t), P) =
PBd (t)), and hence the cost and constraints will be easily
written.
We take as cost to be minimized the length of the
curve since we would like to have the shortest path which
respects the constraints, i.e., Ξ̃(Bd (t), P) = ||z 0 (t)||. This
translates into the integral cost:
ZtN
0
ZtN
||z (t)||dt =
t0
||PM1 Bd−1 (t)||dt
t0
t
ZN
X
=
([PM1 ]i )T Bi,d−1 (t)Bj,d−1 (t)dt [PM1 ]j
i,j
t0
(34)
V. Illustrative example for an UAV system
We revisit the test case from [8]. A 2D 3-DOF model
(32) of an airplane in which the autopilot forces coordinated turns (zero side-slip) at a fixed altitude:
ẋ(t) = Va (t) cos Ψ(t),
ẏ(t) = Va (t) sin Ψ(t),
g tan Φ(t)
Ψ̇(t) =
Va (t)
(32)
The state variables are represented by the position
(x(t), y(t)) and the heading (yaw) angle Ψ(t) ∈ [0, 2π]
rad. The input signals are the airspeed velocity Va (t) and
the bank (roll) angle Φ(t), respectively.
We take as flat output the position
components
of
>
>
the state, z(t) = z1 (t) z2 (t) = x(t) y(t) which
where matrix M1 links Bd−1 (t) and Ḃd (t) as in property
Property II-B and [·]i extracts the i-th column from the
argument. Since the inner integrals can be computed
numerically, we have now a quadratic formulation of the
cost and we can use it for the various constructions from
Section III, see also [5] for a similar treatment of cost
computations.
For illustration purposes we consider 9 hyperplanes:
−0.5931
0.8051
4.2239
0.1814
0.1719
0.9834
−0.0044
0.9975
1.0000
−0.1323
0.2728
0.9912
H=
−0.7011 −0.7131 , h = 3.6785
0.8152 −0.5792
0.0317
0.4352
1.6598
0.9003
1.0000 −0.0075
4.5790
−0.5961 −0.8029
1.0280
which lead to a hyperplane arrangement (23) where the
n
tM I
`M I
tN L
`N L
10
0.0985
*
37.426
203.51
15
6.231
16.877
0.6412
33.469
20
17.106
16.307
1.055
202.076
25
15.034
16.202
0.6321
31.159
30
21.522
16.536
*
*
TABLE I: Evolution of trajectories characteristics for
degree d = 4 in the non-linear and mixed-integer formulations.
interdicted tuples (25)
Σ• = {(+ + + − − + + + −),
(+ − + − + + + + +), (+ − + + + − − + +)}
correspond to three obstacles. Further, we take three
way-points (initial, intermediary and final) though which
the trajectory has to pass at predefined times:
−9
0
6
W=
,
,
, TW = {0, 5, 10}.
−0.5
1.5
0
been done using Yalmip [19] and MPT Toolboxes [20] in
Matlab 2013a. The nonlinear solver used was the IPOPT
solver [21].
VI. Conclusions
This paper considers collision avoidance in a multiagent multi-obstacle framework. Using differential flatness for trajectory generation and B-splines for the flat
output parametrization we show that the restriction
can be validated at all times. Exact and sub-optimal
constructions are provided. Future work may consist in
analysis of the feasibility of the problem, relaxation of
way-point restrictions (similar with work done in [8]),
etc.
Acknowledgments
This work was supported by a grant of the Romanian
National Authority for Scientific Research and Innovation, CNCS – UEFISCDI, project number PN-II-RU-TE2014-4-2713.
References
We compute a flat trajectory (parametrized by n = 12
and d = 6) which passes through the given way-points,
minimizes the total path-length and respects one of the
following scenarios: i) without any collision avoidance
restriction; ii) with collision avoidance as in Proposition 1; iii) with collision avoidance as in Corollary 1. The
obstacles, their support hyperplanes and the resulting
trajectories are depicted in Fig. 2a.
Scenarios ii) and iii) both accomplish the task of avoiding the obstacles with comparable computation times
and path lengths. Figures 2b and 2c show details of the
collision avoidance. In both cases the 1st and 3rd convex regions are considered (Convex Hull{p0 . . . p5 } and
Convex Hull{p2 . . . p7 }) together with their separating
hyperplanes. In Fig. 2b these hyperplanes are the 1st and
respectively 3rd support hyperplane whereas in Fig. 2c
the separating hyperplanes are the
result of the optimiza>
0
tion problem and are H1 = { −0.1273 0.2913 x =
>
11}, H30 = { −0.05490.8461 x = 11}.
As mentioned earlier, the result of the optimization
problem (computation time, total length of the trajectory) depends heavily on the number of control points
n + 1 and degree d. We illustrate these evolutions in
Table I. Several remarks are in order. First, it seems
that after an initial decrease in the path length the future
reductions are negligible and at a significant computation
time for the mixed-integer method. Next, and somewhat
surprising, is that the non-linear method is extremely
sensitive to parameter variations (number of control
points, degree, positioning and number of the way-points,
etc) such that the results obtained are not trustworthy.
The one advantage of the latter over the former is that
it may provide a feasible solution for small values of n.
The collision avoidance between two agents is similar
and not depicted here. All the numerical simulations have
[1] M. Burger and K. Pettersen, “Smooth transitions between
trajectory tracking and path following for single vehicles and
formations,” in Estimation and Control of Networked Systems,
2010, pp. 115–120.
[2] A. Aguiar and J. Hespanha, “Trajectory-tracking and pathfollowing of underactuated autonomous vehicles with parametric modeling uncertainty,” IEEE Transactions on Automatic Control, vol. 52, no. 8, pp. 1362–1379, 2007.
[3] D. Mellinger, N. Michael, and V. Kumar, “Trajectory generation and control for precise aggressive maneuvers with
quadrotors,” The International Journal of Robotics Research,
pp. 027–36, 2012.
[4] I. Prodan, S. Bencatel, R.and Olaru, J. Sousa, C. Stoica, and
S.-I. Niculescu, “Receding horizon flight control for trajectory
tracking of autonomous aerial vehicles,” Control Engineering
Practice, vol. 21, no. 10, pp. 1334–1349, 2013.
[5] J. De Doná, F. Suryawan, M. Seron, and J. Lévine, “A flatnessbased iterative method for reference trajectory generation in
constrained nmpc,” in Nonlinear Model Predictive Control.
Springer, 2009, pp. 325–333.
[6] F. Suryawan, J. De Dona, and M. Seron, “Methods for trajectory generation in a magnetic-levitation system under constraints,” in 18th Mediterranean Conference on Control and
Automation, 2010, pp. 945–950.
[7] F. Suryawan, “Constrained trajectory generation and fault
tolerant control based on differential flatness and b-splines,”
Ph.D. dissertation, School of Electrical Engineering and Computer Science, The University of Newcastle, Australia, 2012.
[8] F. Stoican, I. Prodan, , and D. Popescu, “Flat trajectory
generation for way-points relaxations and obstacle avoidance,”
2015, submitted for publication at MED 2015.
[9] M. Fliess, J. Lévine, P. Martin, and P. Rouchon, “Flatness
and defect of non-linear systems: introductory theory and
examples,” International journal of control, vol. 61, no. 6, pp.
1327–1361, 1995.
[10] J. Lévine, Analysis and control of nonlinear systems: A
flatness-based approach. Springer Verlag, 2009.
[11] H. Sira-Ramírez and S. Agrawal, Differential Flatness. Marcel Dekker, New York, 2004.
[12] J. Wilkinson, The algebraic eigenvalue problem. Oxford Univ.
Press, 1965, vol. 155.
[13] W. J. Gordon and R. F. Riesenfeld, “B-spline curves and
surfaces,” Computer Aided Geometric Design, vol. 167, p. 95,
1974.
[14] N. M. Patrikalakis and T. Maekawa, Shape interrogation for
computer aided design and manufacturing. Springer Science
& Business, 2009.
3
A(+ − + + + − − + +)
2
1
0
−1
−2
−3
−10
A(+ + + − − + + + −)
−9
−8
−7
3
−6
A(+ − + − + + + + +)
−5
−4
−3
−2
−1
2.5
2
2
p1 p2
p3
p4
p7
H3
p6
0
0
p0
−1.5
−2
−10
−1
Conv Hull{p0 . . . p5 }
−8
−6
−4
−2
4
5
p6
H30
6
7
p1 p2
p3
p0
Conv Hull{p2 . . . p7 }
H10
−1.5
0
2
4
(b) detail of the nonlinear case
p7
p4
−0.5
Conv Hull{p2 . . . p7 }
H1
3
1
0.5
−1
2
p5
1.5
0.5
−0.5
1
3
2.5
1
0
(a) various collision avoidance methods
p5
1.5
way-points
no collision avoidance
nonlinear collision avoidance
mixed-integer collision avoidance
6
−2
−10
Conv Hull{p0 . . . p5 }
−8
−6
−4
−2
Fig. 2: Flat trajectories in a multi-obstacle environment.
[15] L. Piegl and W. Tiller, Curve and Surface Basics. Springer,
1995.
[16] M. Sion et al., “On general minimax theorems,” Pacific J.
Math, vol. 8, no. 1, pp. 171–176, 1958.
[17] F. Stoican, I. Prodan, and D. Popescu, “Flat trajectory generation for way-points relaxations and obstacle avoidance,”
in Control and Automation (MED), 2015 23th Mediterranean
Conference on. IEEE, 2015, pp. 695–700.
[18] J. Ferrez, K. Fukuda, and T. M. Liebling, “Cuts, zonotopes
and arrangements,” The sharpest Cut. SIAM Series on Optimization, 2001.
[19] J. Löfberg, “Yalmip : A toolbox for modeling and optimization
in MATLAB,” in Proceedings of the CACSD Conference,
Taipei, Taiwan, 2004. [Online]. Available: http://users.isy.liu.
se/johanl/yalmip
[20] M. Kvasnica, P. Grieder, and M. Baotić, “Multi-Parametric
Toolbox (MPT),” 2004. [Online]. Available: http://control.ee.
ethz.ch/~mpt/
[21] A. Wachter and L. T. Biegler, “On the implementation of a
primal-dual interior point filter line search algorithm for largescale nonlinear programming,” Mathematical Programming,
vol. 106, no. 1, pp. 25–57, 2006.
0
2
(c) detail of the mixed-integer case
4
6
| 3cs.SY
|
MSPKmerCounter:
A Fast and Memory Efficient
Doc-StartBIOINFORMATICS
Approach for K-mer Counting
Vol. 00 no. 00 2013
Pages 1–7
arXiv:1505.06550v1 [q-bio.GN] 25 May 2015
Yang Li and Xifeng Yan
Department of Computer Science, University of California at Santa Barbara
Received on XXXXX; revised on XXXXX; accepted on XXXXX
Associate Editor: XXXXXXX
ABSTRACT
Motivation: A major challenge in next-generation genome
sequencing (NGS) is to assemble massive overlapping short reads
that are randomly sampled from DNA fragments. To complete
assembling, one needs to finish a fundamental task in many leading
assembly algorithms: counting the number of occurrences of k-mers
(length-k substrings in sequences). The counting results are critical
for many components in assembly (e.g. variants detection and read
error correction). For large genomes, the k-mer counting task can
easily consume a huge amount of memory, making it impossible for
large-scale parallel assembly on commodity servers.
Results: In this paper, we develop MSPKmerCounter, a disk-based
approach, to efficiently perform k-mer counting for large genomes
using a small amount of memory. Our approach is based on a
novel technique called Minimum Substring Partitioning (MSP). MSP
breaks short reads into multiple disjoint partitions such that each
partition can be loaded into memory and processed individually.
By leveraging the overlaps among the k-mers derived from the
same short read, MSP can achieve astonishing compression ratio
so that the I/O cost can be significantly reduced. For the task of
k-mer counting, MSPKmerCounter offers a very fast and memoryefficient solution. Experiment results on large real-life short reads
data sets demonstrate that MSPKmerCounter can achieve better
overall performance than state-of-the-art k-mer counting approaches.
Availability: MSPKmerCounter is available at http://www.cs.ucsb.edu/
∼yangli/MSPKmerCounter
Contact: [email protected]
1 INTRODUCTION
High-quality genome sequencing plays an important role in genome
research. A central problem in genome sequencing is assembling
massive short reads generated by the next-generation sequencing
technologies (Mardis et al., 2008). These reads are usually
randomly extracted from samples of DNA segments. Typically
a modern technology can produce billions of short reads whose
length varies from a few tens of bases to several hundreds. For
example, massively parallel sequencing platforms, such as Illumina
(www.illumina.com), SOLiD (www.appliedbiosystems.com), and
454 Life Sciences (Roche) GS FLX (www.roche.com), can produce
reads from 25 to 500 bases in length. The short read length is
expected to further increase in the following years.
Despite the progress in sequencing techniques and assembly
methods in recent years, de novo assembly remains a computationally
c Oxford University Press 2013.
challenging task. The existing de novo assembly algorithms can
be classified into two main categories based on their internal
assembly model: (1) The overlap-layout-consensus model, used by
Celera (Myers et al., 2000), ARACHNE (Batzoglou et al., 2002),
Atlas (Havlak et al., 2004), Phusion (Mullikin et al., 2003) and
Forge (Platt et al., 2010); (2) The de Bruijn graph model, used
by Euler (Pevzner et al., 2001), Velvet (Zerbino et al., 2008),
ABySS (Simpson et al., 2009), AllPaths (Butler et al., 2008) and
SOAPdenovo (Li et al., 2010a). The overlap-layout-consensus
model builds an overlap graph between reads. Since each read can
overlap with many other reads, it is more useful for sequencing
data sets with a small number of long reads. The de Bruijn graph
approach breaks short reads to k-mers (substring of length k) and
then connects k-mers according to their overlap relations in the
reads. The de Bruijn graph approach is usually able to assemble
larger quantities (e.g., billions) of short reads with greater coverage.
Systematic comparison of these algorithms is given by Earl et al.
(2011) and Salzberg et al. (2012).
Although the de Bruijn graph approach comes up with a
good framework to reduce the computation time for assembly,
the graph size can be extremely large, for example, containing
billions of nodes (k-mers) for genomes of higher eukaryotes
like mammals. Therefore, large memory consumption is a
pressing practical problem for the de Bruijn graph based approach
(Miller et al., 2010). For the short read sequences generated from
mammalian-sized genomes, software like Euler, Velvet, AllPaths
and SOAPdenovo will not be able to finish assembling successfully
within a reasonable amount of memory. Due to this drawback, it
significantly limits the opportunity to run de novo assembly on
numerous commodity machines in parallel for large-scale sequence
analysis. This problem has also blocked other application of de
Bruijn graphs, e.g., variants discovery in Iqbal et al. (2012).
To deal with the memory issue, an error correction step is
often taken to eliminate erroneous k-mers before constructing the
de Bruijn graphs. In most NGS data sets, a large fraction of kmers arise from sequencing errors. These k-mers have very low
frequencies. In the giant panda genome sequencing experiment
(Li et al., 2010b), the error correction process could eliminate 68%
of the observed 27-mers, reducing the total number of distinct 27mers from 8.62 billion to 2.69 billion. Though error correction
is usually helpful, obtaining the k-mer frequencies itself is a
computationally demanding task for large genome data sets. One
“naive” solution is using a hash table, where keys are the k-mers
and values are the corresponding k-mer frequencies. Unfortunately,
1
Li
this approach will easily blow up main memory. For example, in the
Asian genome short read data set (Li et al., 2010a), if k = 25, there
are about 14.6 billion distinct k-mers. Assuming a load factor of 2/3
for the hash table and encoding each nucleotide with 2 bits, the kmers table would require nearly 160 GB memory. Furthermore, this
problem will become severe when the length of short reads produced
by the next-generation sequencing techniques further increases.
A recently developed program called Jellyfish (Marcais et al.,
2011) is designed to count k-mers in a memory efficient way.
It adopts a “quotienting” technique to reduce the memory
consumption of k-mers stored in a hash table. Implemented with
a multi-threaded, lock-free hash table, it is able to count k-mers up
to 31 nucleotides in length using a much smaller amount of memory
than the previous “naive” method. When there is no enough memory
to carry out the entire computation, Jellyfish will write intermediate
counting results to disk and later merge them. Since the same kmer may appear in several different intermediate results, the merge
operation is not just a simple concatenation process; it can be quite
slow. Another state-of-the-art k-mer counting algorithm, BFCounter
(Melsted et al., 2011) is based on bloom filter, a probabilistic
data structure that can also reduce memory footprint. However,
BFCounter is 3 times slower than Jellyfish when Jellyfish is able
to finish the task in memory (Melsted et al., 2011). And moreover,
it might miss some counts.
In this paper, we develop MSPKmerCounter, a disk-based
approach, to efficiently perform k-mer counting for large genomes
using a small amount of memory. Our approach is based on a
recently proposed technique called Minimum Substring Partitioning
(MSP) (Li et al., 2013). MSP breaks short reads to “super k-mers”
(substring of length greater than or equal to k) such that each “super
k-mer” contains k-mers sharing the same minimum p-substring
(p ≤ k). The effect is equivalent to compressing consecutive kmers using the original sequences. It is shown that this compression
approach does not introduce significant computing overhead, but
could lead to partitions 10-15 times smaller than the direct approach
using a hash function (Li et al., 2013), thus greatly reducing I/O
cost.
For the task of k-mer counting, MSPKmerCounter offers a very
fast and memory-efficient solution. Experiment results on large
real-life short reads data sets demonstrate that MSPKmerCounter
can achieve better overall performance than state-of-the-art k-mer
counting approaches like Jellyfish and BFcounter.
2 BACKGROUND
D EFINITION 1 (Short Read, K-Mer). A short read is a string
over alphabet Σ = {A, C, G, T } (in DNA assembly). A k-mer is
a string over Σ whose length is k. Given a short read s, s[i, j]
denotes the substring of s from the ith element to the jth element
(both inclusive). s can be broken into n − k + 1 k-mers, written as
s[1, k], s[2, k+1], . . ., s[n−k+1, n]. Two k-mers in s, s[i, k+i−1],
s[i + 1, k + i] are called adjacent in s.
We can view k-mers generated in a way that a window with width
k slides through a short read s. The adjacency relationship exists
between each pair of k-mers for which the last k-1 bases of the first
k-mer are exactly the same as the first k-1 bases of the last k-mer. .
2
D EFINITION 2 (Reverse Complement). DNA sequences can
be read in two directions: forwards and backwards with each
nucleotide changed to its Watson-Crick complement (A ↔ T
and C ↔ G). For each DNA sequence, its corresponding read
in the other direction is called reverse complement and they are
considered equivalent in bioinformatics.
In most sequencing technologies, the fragments (short reads) are
randomly extracted from the DNA sequence in either direction.
Therefore, if two k-mers, K1 and K2 , are adjacent from K1 to K2
in the short reads data set, it implies that the reverse complement kmer of K2 , say K2 ’ and the reverse complement k-mer of K1 , say
K1 ’, are adjacent from K2 ’ to K1 ’. So in an assembly processing,
each short read should be read twice, once in forward direction
and then in the reverse complement direction. However, in real
implementation, it is possible to avoid reading sequences twice
by inferring the subgraph introduced by reverse complements later
from the forward direction subgraph.
3 MINIMUM SUBSTRING PARTITIONING
Our approach to do fast and memory efficient k-mer counting is
based on a disk-based partition approach called Minimum Substring
Partitioning (MSP) (Li et al., 2013). MSP is able to partition kmers into multiple disjoint partitions, as well as retaining adjacent
k-mers in the same partition. This nice property introduces two
advantages: first, instead of being outputted as several individual
k-mers, consecutive k-mers can be compressed to “super k-mers”
(substring of length greater than or equal to k), which will greatly
reduce the I/O cost of partitioning; second, with adjacent k-mers
in the same partition, it is possible to do local assembly for each
partition in parallel and later merge them to generate the global
assembly.
D EFINITION 3 (Substring). A substring of a string s =
s1 s2 . . . sn is a string t = si+1 si+2 . . . si+m , where 0 ≤ i and
i + m ≤ n.
D EFINITION 4 (Minimum Substring). Given a string s, a lengthp substring t of s is called the minimum p-substring of s, if ∀s′ , s′ is
a length-p substring of s, s.t., t ≤ s′ (≤ defined by lexicographical
order). The minimum p-substring of s is written as minp (s).
D EFINITION 5 (Minimum Substring Partitioning). Given a string
s = s1 s2 . . . sn , p, k ∈ N , p ≤ k ≤ n, minimum substring
partitioning breaks s to substrings with maximum length {s[i, j]|i+
k − 1 ≤ j, 1 ≤ i, j ≤ n}, s.t., all k-mers in s[i, j] share the same
minimum p-substring, and it is not true for s[i, j + 1] and s[i − 1, j].
s[i, j] is also called “super k-mer”.
Minimum Substring Partitioning comes from the intuition that
two adjacent k-mers are very likely to share the same minimum
p-substring if p << k, since there is a length-(k-1) overlap
between them. Figure 1 shows a Minimum Substring Partitioning
example. In this example, the first 4 k-mers have the same
minimum 4-substring, ACAC, as highlighted in red box; and the
last 3 k-mers share the same minimum 4-substring, ACCC, as
highlighted in blue box. In this case, instead of generating all
these 7 k-mers separately, we can just compress them using the
original short read. Namely, we compress the first 4 k-mers to
MSPKmerCounter
CT GACACT T GACCCGT GGT , and output it to the partition
corresponding to the minimum 4-substring ACAC. Similarly, the
last 3 k-mers are compressed to CACT T GACCCGT GGT CAT
and outputted to the partition corresponding to the minimum 4substring ACCC. Generally speaking, given a short read s =
s1 s2 . . . sn , if the adjacent j k-mers from s[i, i + k − 1] to s[i +
j − 1, i + j + k − 2] share the same minimum p-substring t, then
we can just output substring si si+1 . . . si+j+k−2 to the partition
corresponding to the minimum p-substring t without breaking it to
j individual k-mers. If j is large, this compression strategy will
dramatically reduce the I/O cost.
(a)
CTGACACTTGACCCGTGGTCAT
(b)
CTGACACTTGACCCGT
TGACACTTGACCCGTG
GACACTTGACCCGTGG
ACACTTGACCCGTGGT
CACTTGACCCGTGGTC
ACTTGACCCGTGGTCA
CTTGACCCGTGGTCAT
CTGACACTTGACCCGTGGT
Partition
ACAC
Partition
ACAG
CACTTGACCCGTGGTCAT
…
may become empty. To address this problem, one can introduce a
hash function to wrap the number of partitions to any user-specified
partition number. Then the k-mers are likely to be evenly distributed
across partitions.
D EFINITION 7 (Minimum Substring with Reverse Complement).
Given a string s, a length-p substring t of s is called the minimum
p-substring of s, if ∀s′ , s′ is a length-p substring of s or s’ reverse
complement, s.t., t ≤ s′ (≤ defined by lexicographical order).
Definition 7 redefines minimum substring by considering the
reverse complement. With this new definition, we can make
sure each k-mer and its reverse complement k-mer are assigned
to the same partition. This property can help us save much
time and memory in the later processing (e.g. storing only the
lexicographical smaller one of a k-mer and its reverse complement
k-mer in hash table and avoiding reading each short read twice
to explicitly process reverse complement) since a k-mer and its
reverse complement are considered equivalent in bioinformatics and
the information introduced by reverse complement can be inferred
from the forward direction short reads. For simplicity reason, in the
following discussions, if not mentioned explicitly, we will ignore
the reverse complement issue. However, in our implementation and
experiments, we do consider its impact.
Partition
ACCC
4 METHODS
Fig. 1. A minimum substring partitioning example: (a)short read (b)K-mers
and MSP process
The results of the Minimum Substring Partitioning are determined
by the parameters k and p. Smaller p will increase the probability
that consecutive k-mers share the same minimum p-substring and
thus reduce the I/O cost. However, it will also introduce a problem
where the distribution of partition sizes become skewed and the
largest partition may not fit in the main memory. In the extreme
case of p = 1, the size of the largest partition is almost as
same as the size of the short reads data set and other partitions
are almost empty (assuming the four nucleotides A, C, G, T are
distributed randomly in the data set). In that case, we lose the point
of partitioning. On the other hand, larger p will make the distribution
of partition sizes evener at the cost of decreasing the probability
that consecutive k-mers share the same minimum p-substring and
thus increasing the I/O cost. In the extreme case of p → k, almost
no adjacent k-mers will share the same minimum p-substring and
thus no compression can be gained. Therefore one needs to make a
tradeoff (by varying p) between the largest partition’s size and the
I/O overhead. Fortunately, there is a quite wide range of values that
p can choose without affecting the performance of MSP (Li et al.,
2013).
D EFINITION 6 (Wrapped Partitions). Given a string set {si }, a
hash function H, the user-specified number of partitions N , for any
k-mer si,j , minimum substring partition wrapping assigns si,j to
the (H(minp (si,j )) mod N )-th partition.
Since each p-substring corresponds to one partition, the total
number of partitions in MSP is equal to 4p . When p increases, the
number of partitions will increase exponentially and many partitions
In this section, we describe the detailed method to do k-mer
counting with the adoption of the minimum substring partitioning
technique introduced in the last section.
The first step is to partition short reads. In this step, we will
cut each short read of length n into (n − k + 1) k-mers and
then dispatch these k-mers into different partitions. The Minimum
Substring Partitioning technique introduced in Section 3 is used as
our partitioning method. As mentioned before, with this partitioning
method, we can compress consecutive k-mers dispatched to the
same partition into one “super k-mer” to minimize the I/O cost.
There are several ways (e.g. straightforward, min-heap) to
implement the minimum substring partitioning. Here we adopt the
one introduced in Li et al. (2013), since it is proved to have the
best performance in practice. The details of this implementation is
described in Algorithm 1.
Algorithm 1 Minimum Substring Partitioning
Input: String s = s1 s2 . . . sn , integer k, p.
min s = the minimum p-substring of s[1, k]
min pos = the start position of min s in s
for all i from 2 to n − k + 1 do
if i > min pos then
min s = the minimum p-substring of s[i, i + k − 1]
update min pos accordingly
else
if the last p-substring of s[i, i + k − 1] < min s then
min s = the last p-substring of s[i, i + k − 1]
update min pos accordingly
end if
end if
end for
3
Li
As mentioned before, we can view k-mers generated in a way that
a window with width k slides through a short read. In Algorithm 1,
initially when the window starts at position 1, we scan the window
to find the minimum p-substring, say min s, and the start position of
min s, say min pos. Then we slide the window forward, one symbol
each time, till the right bound of the window reaches the end of
the short read. After each sliding, we test whether the min pos is
still within the range of the window. If not, we have to re-scan
the window to get new min s and min pos. Otherwise, we test
whether the last p-substring of the current window is smaller than
current min s. If yes, we set this last p-substring as new min s and
update min pos accordingly. If not, we just keep the old min s
to calculate the partition location. As described in last section,
the neighboring k-mers will likely contain the same minimum psubstring. Therefore, the re-scan of the whole window will not occur
very often. The worst case time complexity is O(nk) p-substring
comparisons. However, this algorithm is more efficient in practice
(close to O(n + lk), see detailed proof in Li et al. (2013)) when s
is broken to only a few number (l) of “super k-mers”. This is very
true in minimum substring partitioning of real short reads. Table 1
shows that the average number of breakdowns is small for several
real short reads data sets.
Table 1. Average number of breakdowns for real short reads data sets
Data Set
Budgerigar
Red tailed boa constrictor
Lake Malawi cichlid
Soybean
n
k
p
Average Breakdown (l)
150
121
101
75
59
59
59
59
10
10
10
10
5.22
3.89
2.77
1.69
Table 2. Basic facts about the four sequence data sets used in
our experiments
Format
Size (GB)
Avg Read Length
No. of Reads (million)
bird
snake
fish
soybean
fastq
106.8
150
323
fastq
181.7
121
573
fastq
137.4
101
598
fastq
40.1
75
227
frequency count disk files. The query of a k-mer’s frequency is also
very easy and efficient. Given a query k-mer, we can use MSP to
calculate its partition location and then perform binary search on
the corresponding count disk file to get the k-mer frequency.
5 EXPERIMENTAL RESULTS
In this section, we present experimental results that illustrate the
efficiency of our MSPKmerCounter on four large real-life short
reads data sets: Budgerigar (bird), Red tailed boa constrictor
(snake), Lake Malawi cichlid (fish) and soybean. (1) We first
analyze the efficiency of MSPKmerCounter by reporting the
memory and time costs, along with the temporary disk space
usage; (2) We then investigate the scalability and parallelizability
of MSPKmerCounter. We will compare MSPKmerCounter with
two state-of-the-art k-mer countering tools: Jellyfish (Marcais et al.,
2011) and BFCounter (Melsted et al., 2011). All the experiments, if
not specifically mentioned, are conducted on a server with 2.00GHz
Intel Xeon CPU and 512 GB RAM.
5.1 Sequence Datasets
Note that in Algorithm 1, every time when we capture a
minimum substring change at position j of s or we reach the
end of s, we output a “super k-mer” of s that contains the
previous minimum substring into the partition corresponding to that
minimum substring. This part of code is not presented in Algorithm
1.
After obtaining the partitions, we can use a simple hash table
whose keys are k-mers and values are k-mer counts to count the
k-mer frequencies. For each partition, break the “super k-mers” into
k-mers and insert these k-mers into a hash table. Since adjacent kmers are only different by the first and last symbol, direct bit shift
operations (A, C, G, T can be encoded using 2 bits) can be applied
here to improve the efficiency. Whenever we see a new k-mer, we
first look up the hash table to see if it is already in the hash table: if
yes, we increase the frequency count by 1; otherwise, we put this kmer along with an initial frequency value 1 into the hash table. After
processing one partition, write the entries in hash table to a disk
file1 and release the memory occupied by that hash table. Since all
the occurrences of the same k-mer will locate in the same partition,
the frequency count of a k-mer can be found in only one disk file.
This is a very good property, as we do not have to later merge these
1
Actually we will sort the k-mers in hash table before writing them back to
disk. Such sorting is used to facilitate efficient query of k-mer frequencies.
4
Four very large real-life short reads data sets are used to test
MSPKmerCounter. The first one is the sequence data of Budgerigar
(bird) obtained from bioshare.bioinformatics.ucdavis.edu/Data/hc
These short reads were sequenced from the Illumina HiSeq 2000
technology. The second one is the sequence data of Red tailed boa
constrictor (snake) downloaded from bioshare.bioinformatics.ucdavis.
These short reads were obtained with the Genome Analyzer
technology. The third one is the sequence data of Lake Malawi
cichlid (fish) downloaded from bioshare.bioinformatics.ucdavis.edu/
And the last one is the sequence data of soybean downloaded from
ftp://public.genomics.org.cn/BGI/soybean_resequencing/fas
Some basic facts about these four data sets are shown in Table 2.
5.2 K-mer Counting Efficiency
We conduct experiments to test the efficiency of our MSPKmerCounter
and compare it with two state-of-the-art k-mer counting algorithms:
Jellyfish, which is a fast, memory efficient k-mer counting tool
based on a multi-threaded, lock-free hash table optimized for
counting k-mers up to 31 nucleotides in length; and BFCounter,
which is a k-mer counting tool with greatly reduced memory
requirements based on bloom filter, a probabilistic data structure.
BFCounter is a completely in-memory kmer counting method.
Jellyfish can work both as in-memory or out-of-core. It requires
user to pre-specify the size of the hash table: if the hash table is
large enough to hold all the k-mers, it will be an in-memory method;
MSPKmerCounter
otherwise, whenever the hash table fills up, the intermediate results
will be written to disk and merged later, making it become a diskbased method. In this set of experiments, we will test Jellyfish
under these two different settings, denoted as Jellyfish(Memory)
and Jellyfish(Disk) respectively. For Jellyfish(Memory), we precalculate the number of distinct k-mers in each data set to make
sure the hash table is big enough to hold all the k-mers. For
Jellyfish(Disk), we set the hash table size to a fixed number so that it
will consistently make use of ∼11 GB memory. We set the number
of threads to 1 for all the three methods, since BFCounter only
supports single thread. For MSPKmerCounter, we set the number
of wrapped partitions to 1,000 (to reduce memory footprint) and the
minimum substring length p to 10.
Table 3 presents the memory consumption and running time for
the three methods when applied to the snake, fish and soybean data
sets2 for counting 31-mers3 .
Table 3. Comparison of memory consumption and running time for counting
31-mers on the snake, fish and soybean data sets.
Algorithm
Memory (GB)
Jellyfish(Memory)
Jellyfish(Disk)
BFCounter
MSPKmerCounter
Run Time (minutes)
snake
fish
soybean
snake
fish
soybean
110
11
38
9.6
114
11
29
9.9
43
11
13
6.3
455.5
775.2
1899.8
492.7
374.5
503.7
1299
399.2
93.6
117.7
342.2
99
Table 4 shows the temporary disk space usage for the three
methods when applied to the snake, fish and soybean data sets for
counting 31-mers.
Table 4. Comparison of temporary disk space usage
for counting 31-mers on the snake, fish and soybean
data sets.
Algorithm
Jellyfish(Memory)
Jellyfish(Disk)
BFCounter
MSPKmerCounter
Temp Disk Space Usage (GB)
snake
fish
soybean
0
332
0
217
0
197
0
168
0
44
0
43
As can be seen from Table 3, when applied to a large
sequence data set with deep coverage, our MSPKmerCounter soon
demonstrates its advantages. It uses much less memory than both
Jellyfish(Memory) and BFCounter. Its running time is close to
that of Jellyfish(Memory) and significantly shorter than that of
2
We reserve the bird data set to test scalability (See Section 5.3)
Jellyfish only supports counting k-mers whose length is smaller than 32.
BFCounter and MSPKmerCounter do not have such a constraint.
3
BFCounter. Jellyfish(Disk) was able to finish the counting task
using a small amount of memory, by writing intermediate results
to disk and later merging them. But unfortunately, its merging
process is relatively inefficient since the k-mer sets in those
intermediate results are not completely disjoint. Therefore it is much
slower than MSPKmerCounter. MSPKmerCounter requires no
additional merging steps after partial counting results are generated
from individual partitions. Also, as can be seen from Table 4,
MSPKmerCounter uses less amount of temporary disk space than
Jellyfish(Disk). Note that Jellyfish(Memory) and BFCounter do not
need to use any temporary disk space since they are completely
memory-based. Actually the memory consumption and temporary
disk space usage of our MSP-based counting method can be fully
controlled by varying the number of wrapped partitions and the
minimum substring length p. For more discussions (both theoretical
and experimental) about the sensitivity of MSP to these parameters,
please refer to Li et al. (2013).
5.3 Scalability
We then conduct experiments to test the scalability of Jellyfish,
BFCounter and our MSPKmerCounter. Specifically, we count the
k-mers in the Budgerigar data set for various levels of coverage,
using these three counting methods. In order to get different levels of
coverage, we randomly sampled the short reads data set to obtain a
desired amount of sequences. As same as the previous experiments,
here we also test Jellyfish under two different settings.
The memory consumption, running time and temporary disk
space usage for counting 31-mers in the Budgerigar(bird) data set
under various levels of coverage are shown in Figures 2(a), 2(b) and
2(c), respectively.
Figures 2(a) shows that with the increase of coverage, the
memory consumption of Jellyfish(Memory) increases significantly.
In comparison, the memory utilizations of BFCounter and
MSPKmerCounter only increase slightly. MSPKmerCounter
outperforms both Jellyfish and BFCounter in terms of memory
footprint. Note that we configure Jellyfish(Disk) to use at most 11
GB memory, so its memory consumption does not change since
coverage 5. Figure 2(b) shows that with the increase of coverage,
the running time of all counting methods increases. However, the
increasing speed of BFCounter is much higher than that of Jellyfish
and MSPKmerCounter. As the coverage increases, the running time
gap between MSPKmerCounter and Jellyfish(Disk) becomes larger
and larger, indicating MSPKmerCounter’s better scalability. Even
when compared with the purely memory-based Jellyfish(Memory),
MSPKmerCounter is only slightly slower at all coverages. Figure
2(c) shows that the temporary disk space usages of both
Jellyfish(Disk) and MSPKmerCounter increase as the coverage
increases. But the increasing speed of MSPKmerCounter is much
slower than that of Jellyfish(Disk), indicating MSPKmerCounter’s
better scalability in disk space utilization. Jellyfish(Memory) and
BFCounter need no extra disk space since they are completely
memory-based. To summary, when the coverage is low (e.g. less
than 5), the performance differences among Jellyfish, BFCounter
and MSPKmerCounter are not very big, though MSPKmerCounter
is still much faster than BFCounter and uses less memory
than both Jellyfish and BFCounter. As the coverage increases,
MSPKmerCounter quickly dominates the scene. In a high coverage
situation, the main memory is not big enough for Jellyfish to
5
Li
40
20
0
1
5
10
15
Coverage
20
(a) Memory Consumption
25
800
200
Jellyfish(Memory)
Jellyfish(Disk)
BFCounter
MSPKmerCounter
Temp Disk Space Usage (GB)
60
1000
Jellyfish(Memory)
Jellyfish(Disk)
BFCounter
MSPKmerCounter
Running time (minutes)
Memory Consumption (GB)
80
600
400
200
0
1
5
10
15
Coverage
20
150
100
50
0
1
25
Jellyfish(Memory)
Jellyfish(Disk)
BFCounter
MSPKmerCounter
(b) Running Time
5
10
15
Coverage
20
25
(c) Temp Disk Space Usage
Fig. 2. Memory consumption, running time and temporary disk space usage of Jellyfish, BFCounter and MSPKmerCounter for counting 31-mers in the
Budgerigar data set under various levels of coverage
5.4 Parallelizability
Our MSP-based k-mer counting method can easily be parallelized
to support multi-threads or be distributed to multiple machines to
enable parallel processing. There are three distinct phases in the
Minimum Substring Partitioning process. First, it reads the short
read sequences. Second, it calculates the minimum substring of
each k-mer and merges possible adjacent k-mers into “super kmers”. Last, it writes the “super k-mers” back to disk files. Phase
1 and phase 3 are I/O operations, so the speedup can be obtained by
using multi-threads to process phase 2. After partitioning, different
partitions are completely disjoint. Therefore it is helpful to use
different threads to process different partitions simultaneously.
We implemented a preliminary multi-thread version of our
MSPKmerCounter, denoted as MSPKmerCounter(MT). Since
BFCounter is not able to support multiple threads, here we
only conduct experiments to compare MSPKmerCounter(MT) with
Jellyfish, which is highly optimized to support efficient multithread processing. Figure 3 shows the running time comparison of
MSPKmerCounter(MT) and Jellyfish with the increasing number
of threads. Here k-mers are counted on the Lake Malawi cichlid
(fish) data set with k = 31. Again we test Jellyfish under two
different settings (the settings are as same as those in Section 5.2).
From Figure 3 we can see that: (1) Jellyfish(Memory) has an almost
linear speedup up to 4 threads, indicating the best parallelizability.
This is reasonable since it puts everything in memory and therefore
involves almost no I/O costs. However, as mentioned before, its
huge memory footprint will greatly limit its usage on commodity
computers. (2) Both Jellyfish(Disk) and MSPKmerCounter(MT)
exhibit good parallelizability for up to 2 threads and then levels
off. This is because these two disk based methods involve a lot of
I/O operations. At 2 threads the CPU calculation is already fast
6
enough and the I/O bandwidth has become the main bottleneck.
MSPKmerCounter(MT) is still faster than Jellyfish(Disk).
From (1) and (2) we can conclude that Jellyfish is more suitable
for powerful computers (e.g. computers with large RAM and many
cores), while MSPKmerCounter is the better choice for commodity
computers (e.g. computers with small RAM and few cores).
600
Running time (minutes)
finish all the computation in memory and therefore it has to write
intermediate results to disk and later merge them. This gives
MSPKmerCounter a chance to outperform Jellyfish in terms of
both memory and time. BFCounter has the advantage that its
memory usage does not increase a lot as the coverage increases.
However, compared with MSPKmerCounter, it still requires more
memory and much longer running time to finish the task. Moreover,
the memory consumption, running time and disk space usage
of MSPKmerCounter are fully controllable by varying several
parameters (Li et al., 2013).
500
Jellyfish(Memory)
Jellyfish(Disk)
MSPGraphCounter(MT)
400
300
200
100
0
1
2
4
Number of threads
8
Fig. 3. Running time versus #threads for Jellyfish and MSPKmerCounter
6 FUTURE WORK
There are some future avenues to pursue to further improve our
work. First, we can adopt the techniques (e.g. variable length
encoding) introduced in Jellyfish (Marcais et al., 2011) to make
space-efficient encoding of keys and reduce the memory usage of
each hash entry to further reduce the memory consumption. Second,
we can think about extending the use of MSP from counting kmers to the whole sequence assembly process. Since the k-mers in
different MSP partitions are completely disjoint and the majority
of adjacent k-mers in original short reads are retained in the same
partition, it is possible to perform local assembly (including some
error correction steps like tip removal and bubble merging) for each
partition and later “glue” these local assembly results to obtain the
global assembly results. By doing so, the whole assembly process
can be done with a very small amount of memory. And the assembly
can speed up a lot with the gains of parallel assembly of multiple
partitions.
MSPKmerCounter
7 CONCLUSION
REFERENCES
In this paper, we aimed at the computational bottlenecks in k-mer
counting, which is an important step in many genome sequence
assembly tasks. We developed a disk-based approach based on a
novel technique, Minimum Substring Partitioning (MSP), to solve
the memory overwhelming problem. MSP breaks the short reads
into multiple disjoint partitions so that each partition only requires
a very small amount of memory to process. By leveraging the
overlaps among the k-mers derived from the same read, MSP is able
to achieve astonishing compression ratio so that the I/O cost can
be greatly reduced, making the method be very efficient in terms
of time and space. Our MSP-based k-mer counting method were
evaluated on real DNA short read sequences. Experimental results
show that it can not only successfully finish the counting task on
very large data sets using a reasonable amount of memory, but also
achieve better overall performance than the existing k-mer counting
methods.
Mardis, E.R. (2008) Next-generation DNA sequencing methods, Annu. Rev. Genomics
Hum. Genet., 9, 387-402.
Myers, E.W. et al. (2000) A whole-genome assembly of Drosophila, Science, 287,
2196-2204.
Batzoglou, S. et al. (2002) ARACHNE: a whole-genome shotgun assembler, Genome
research, 12, 177-189.
Havlak, P. et al. (2004) The Atlas genome assembly system, Genome research, 14,
721-732.
Mullikin, J.C. and Ning, Z. (2003) The phusion assembler, Genome research, 13, 81-90.
Platt, D. and Evers, DJ (2010) Forge: A Parallel Genome Assembler Combining Sanger
and Next Generation Sequence Data, http://combiol.org/forge/.
Pevzner, P.A. et al. (2001) An Eulerian path approach to DNA fragment assembly,
Proceedings of the National Academy of Sciences, 9748-9753.
Zerbino, D.R. and Birney, E. (2008) Velvet: algorithms for de novo short read assembly
using de Bruijn graphs, Genome research, 18, 821-829.
Simpson, J.T. et al. (2009) ABySS: a parallel assembler for short read sequence data,
Genome research, 19, 1117-1123.
Butler, J. et al. (2008) ALLPATHS: de novo assembly of whole-genome shotgun
microreads, Genome research, 18, 810-820.
Li, R. et al. (2010) De novo assembly of human genomes with massively parallel short
read sequencing, Genome research, 20, 265-272.
Miller, J.R. et al. (2010) Assembly algorithms for next-generation sequencing data,
Genomics, 95, 315-327.
Li, R. et al. (2010) The sequence and de novo assembly of the giant panda genome,
Nature, 463(7279), 311-317.
Marcais, G. and Kingsford, C. (2011) A fast, lock-free approach for efficient parallel
counting of occurrences of k-mers, Bioinformatics, 27, 764-770.
ACKNOWLEDGEMENT
The authors would like to thank the Assemblathon website
(http://assemblathon.org) for providing the Budgerigar
(bird), the Red tailed boa constrictor (snake) and the Lake Malawi
cichlid (fish) data, and Beijing Genomics Institute for providing the
soybean data.
Funding: This research was sponsored in part by the U.S. National
Science Foundation under grant IIS-0847925 and IIS-0954125. The
views and conclusions contained herein are those of the authors
and should not be interpreted as representing the official policies,
either expressed or implied, of the U.S. Government. The U.S.
Government is authorized to reproduce and distribute reprints for
Government purposes notwithstanding any copyright notice herein.
Melsted, P. and Pritchard, J.K. (2011) Efficient counting of k-mers in DNA sequences
using a bloom filter, BMC Bioinformatics, 12, 333-339.
Salzberg, S.L. et al. (2012) GAGE: A critical evaluation of genome assemblies and
assembly algorithms, Genome research.
Earl, D. et al. (2011) Assemblathon 1: A competitive assessment of de novo short read
assembly methods, Genome research, 21, 2224-2241.
Iqbal, Z. et al. (2012) De novo assembly and genotyping of variants using colored de
Bruijn graphs, Nature genetics.
Li, Y. et al. (2013) Memory efficient minimum substring partitioning, Proceedings of
the Very Large Databases Endowment, 169-180.
7
| 5cs.CE
|
arXiv:1609.03492v1 [math.GR] 12 Sep 2016
Groups from Class 2 Algebras and the Weil Character
Christakis A. Pallikarosa and Harold N. Wardb
March 2, 2018
a
b
Department of Mathematics and Statistics, University of Cyprus, PO Box 20537, 1678 Nicosia, Cyprus
E-mail: [email protected]
Department of Mathematics, University of Virginia, Charlottesville, VA 22904, USA
E-mail: [email protected]
Abstract
We investigate the behaviour of the Weil character of the symplectic group on
restriction to subgroups arising from commutative nilpotent algebras of class 2. We
give explicit descriptions of the decomposition of the Weil character when restricted to
the unipotent radical of the stabilizer of a maximal totally isotropic subspace and to
its centralizer.
1
Introduction
Because of their significant properties, Weil representations play an important role in the
study of the representation theory of classical groups. The characters of Weil representations
have been computed by various authors; see for example [1], [5], [6], [11]. We will be making
use of some explicit results concerning the Weil characters of the symplectic group as these
are obtained in [14] via the ‘theta form’ (see also [11]). The approach in [14] is to follow the
treatment for the Weil representation given in [13].
One of the aims of the present work is to follow up in the direction of some of the
investigations in [9], [10] concerning the restriction of the Weil characters of symplectic and
unitary groups to certain subgroups, in particular to certain self-centralizing subgroups. In
[10] an explicit description of the restriction of the Weil character (respectively, the sum of
the two Weil characters) of the unitary group (respectively, symplectic group) to centralizers
of regular unipotent elements was also obtained. We remark here that the study of the
behaviour of element centralizers in Weil representations already began in [4].
At this point we introduce some notation. Fix q a power of an odd prime and consider
the Weil representation and its character ω for Sp(2n, q), regarded
as a matrix group on
0
I
. The subgroup G
V = GF (q)2n defined via the symplectic form ϕ having matrix
−I 0
1
I 0
,
of Sp(2n, q) to be considered is G = h−Ii × B where B is the set of matrices gQ =
Q I
with Q a symmetric matrix. So B is the unipotent radical of the stabilizer in Sp(2n, q) of a
maximal totally isotropic subspace of V , and G, which is the centralizer of B in Sp(2n, q), is
a maximal Abelian subgroup of Sp(2n, q). Also note that the matrices gQ − I with gQ ∈ B
belong to a nilpotent subalgebra of the full matrix algebra of 2n × 2n matrices over GF (q)
of class 2 (compare with the discussion in [9]).
The restriction ω|G is known to be multiplicity free. The main result of the paper is
Theorem 5, where we describe explicitly the irreducible characters of G (respectively, B)
appearing in the decomposition of ω|G (respectively, ω|B). As a by-product, in Section 7
we use the expression for the decomposition ω|B in order to obtain alternative derivations
for the number of solutions of the equation Q(x) = α with α ∈ GF (q), x ∈ GF (q)n and Q
a quadratic form of GF (q)n .
One can easily observe that a subgroup H of G with a multiplicity free restriction ω|H
satisfies |H| ≥ 2q n . In Section 9 we show that this bound is actually attained and we give a
construction of such a subgroup H with |H| = 2q n and ω|H being multiplicity free. Finally,
we show how an evaluation concerning the values of the Weil character on the elements of
this subgroup H provides a link with the Davenport–Hasse theorem on lifted Gauss sums.
The decompositions in our main theorem 5 corroborate results in the preprint by Gurevich
and Howe [3] (which appeared after we finished our manuscript), especially those in Section 2.
2
Preliminaries
We consider the Weil representation and its character ω for Sp(2n, q), q a power of an
odd prime, as a matrix group on V = GF (q)2n (with right action). We will be following
the notation about ω in [13], [14] and
usesome explicit results in [14] about its values.
0 I
. Writing members of V as pairs (x, y), with
The symplectic form ϕ has matrix
−I 0
x, y ∈ GF (q)n , one has
ϕ((x1 , y1 ), (x2 , y2 )) = x1 y2T − y1 xT2 ,
the superscript T standing for transpose.
The subgroup
G of Sp(2n, q) to be considered is G = h−Ii × B, where B is the set of
I 0
matrices gQ =
. Such a matrix is in Sp(2n, q) just when QT = Q; that is, Q is
Q I
symmetric (“Q” emphasizes the associated quadratic form Q(v) = vQv T on GF (q)n ). Let
S n be the space of n × n symmetric matrices over GF (q). There are q n(n+1)/2 such matrices.
The following theorem follows from results in [2], but we present a proof done in the
spirit of [9].
Theorem 1 G is Weil-free: the irreducible constituents of ω|G appear with multiplicity 1.
2
Proof. We use the orbit criterion: an Abelian subgroup H of Sp(2n, q) is Weil-free exactly
when the number of orbits of H on V is q n [9]. To count the orbits of G, recall
that (#
X
−I 0
−1
orbits) = |G|
|Gv |, Gv the stabilizer of v. Notice that the matrices −gQ =
−Q −I
v∈V
I 0
= (x + yQ, y), v = (x, y) is fixed exactly
in −B fix only 0. Since (x, y)gQ = (x, y)
Q I
when yQ = 0. We look at three cases:
v = (0, 0): then G(0,0) = |G| = 2q n(n+1)/2 .
v = (x, 0), x 6= 0: all members of B fix v and G(x,0) = q n(n+1)/2 . There are q n − 1 such v.
v = (x, y), y 6= 0: v is fixed by the elements gQ with yQ = 0. We can set up such Q by
thinking of it as a quadratic form with y in the radical. Write GF (q)n = hyi⊕W . Then Q can
be given by taking a form on W and extending it by 0 on hyi. That gives G(x,y) = q n(n−1)/2 ,
n n
the number
Xof choices for the form on W . There are q (q − 1) of these v.
Thus for
|Gv | we get
v∈V
X
|Gv | = 2q n(n+1)/2 + (q n − 1) × q n(n+1)/2 + q n (q n − 1) × q n(n−1)/2
v∈V
So |G|−1
X
= q n(n+1)/2 {2 + q n − 1 + q n − 1}
= q n × 2q n(n+1)/2 .
|Gv | = q n , as needed.
v∈V
3
Irreducible characters of G
To describe the characters of G, let ψ be the canonical additive character of GF (q), as
used in [13] (the terminology is that of [7, p. 190]): ψ(α) = e(2πi/p)tr(α) , where tr is the trace
function GF (q) → GF (p), p the prime dividing q. Each linear character of the additive
group of GF (q) is given by α → ψ(βα), β ∈ GF (q) [7, Theorem 5.7] (this is equivalent to
the nondegeneracy of the trace form (α, β) → tr(αβ)). In what follows, χ is the quadratic
character on GF (q)# (the nonzero elements) and δ = χ(−1). In [13], ρ was defined as
P
2
α∈GF (q) ψ(α ); we also have
X
ρ=
χ(β)ψ(β),
(1)
β6=0
a Gaussian sum [7, Chapter 5, Section 2]; ρ2 = δq. If Q ∈ S n , diagonalize Q and let
∆(Q) = χ(product of nonzero diagonal entries of Q).
That is, if we write GF (q)n as rad(Q) ⊕ W , ∆(Q) is χ(det(Q|W )); Q|W is the nonsingular
part of Q. If Q has even rank 2k, we call Q hyperbolic or elliptic according as Q|W is
hyperbolic or elliptic. If Q|W is hyperbolic, then W is the orthogonal sum of hyperbolic
planes, and det Q|W = (−1)k . Thus ∆(Q) = δ k . If Q|W is elliptic, then ∆(Q) = −δ k .
3
Lemma 2 The irreducible characters of B are the functions λS given by
λS (gQ ) = ψ(Tr(SQ)),
where S ∈ S n and Tr is the matrix trace. Each λS extends to two irreducible characters λ±
S
of G by the formula
λ±
S (gQ ) = ψ(Tr(SQ))
±
λS (−gQ ) = ±ψ(Tr(SQ)),
(2)
with the signs in the last equation matching on the two sides.
Proof. That this formula does give all the linear characters of B follows from the fact that
the trace form (S, Q) → Tr(SQ) on S n is nondegenerate. That, in turn, can be seen as
follows: suppose that Tr(SQ) = 0 for all Q ∈ S n . Take a basis for for GF (q)n that makes
S diagonal. Suppose that ζ is a nonzero diagonal entry of S. Choose Q to have 1 at that
position and 0 elsewhere. Then Tr(SQ) = ζ; so ζ = 0 after all. Thus S = 0. So these
characters λS are all distinct; and since there is the correct number of them, they give all
the characters of B. Then for the characters of G, we use the direct product decomposition
G = h−Ii × B to write them as claimed.
4
Values of ω on G
For the values of ω, we use results from [14]. A member −gQ of −B has −gQ − 1 invertible,
with diagonal entries −2, and [14, Section 6.5] gives
ω (−gQ ) = δ n χ(det(−gQ − 1)) = δ n χ((−2)2n ) = δ n .
(3)
g−1
As for gQ , [14, Theorem 6.7] implies that ω(g) = q n ρ− dim V χ(det Θg ). Here Θg is given in
[14, Definition 3.3]: it is the form defined on V g−1 by Θg (ug−1, v g−1 ) = ϕ(ug−1, v). We have
(x, y)g−1 = (yQ, 0). So
Θg ((x1 , y1 )g−1 , (x2 , y2 )g−1 ) = ϕ((x1 , y1)g−1 , (x2 , y2 ))
= ϕ((y1 Q, 0), (x2 , y2 ))
= y1 Qy2T .
It follows that χ(det Θg ) = ∆(Q). Thus
Proposition 3 The values of ω on G are given by
ω (gQ ) = q n ρ−rankQ ∆(Q)
ω (−gQ ) = δ n .
If Q = 0, ∆(Q) is artificially taken to be 1.
4
(4)
(5)
5
Character multiplicities in ω|G
P
The multiplicity (which is 0 or 1) of a linear character λ in ω|G is |G|−1 g∈G ω(g)λ(g).
With λ = λ±
S , we get
X
X
X
δ n ψ(−Tr(SQ))
q n ρ−rankQ ∆(Q)ψ(−Tr(SQ)) ±
ω(g)λ±
S (g) =
g∈G
Q∈S n
Q∈S n
(again, the sign on the right matches the sign in λ±
S ). The second summation is 0 if S 6= 0,
and δ n q n(n+1)/2 if S = 0.
For further computations, we need some standard group-order formulas. They are taken
from [12].
Group
Order
l
Q
GL(l, q)
q l(l−1)/2 (q i − 1)
i=1
O + (2k, q)
2q k(k−1) (q k − 1)
O − (2k, q)
2q k(k−1) (q k + 1)
O(2k + 1, q)
2q
k2
k
Q
k−1
Q
(q 2i − 1)
i=1
k−1
Q
(6)
(q 2i − 1)
i=1
(q 2i − 1).
i=1
We also need the q-binomial coefficient
n
.
of GF (q)n . By duality, nr q = n−r
q
5.1
n
r q
that gives the number of r-dimensional subspaces
S=0
When S = 0,
(ω, λ±
0 )G =
1
2q n(n−1)/2
X
ρ−rankQ ∆(Q) ±
Q∈S n
δn
,
2
since |G| = 2q n(n+1)/2 . We conclude that the first term must be 1/2:
X
ρ−rankQ ∆(Q) = q n(n−1)/2 .
Q∈S n
We shall elaborate on this in Section 8. Thus we have
Proposition 4 The multiplicity of λ±
0 in ω is
(ω, λ±
0 )G =
5
1 ± δn
.
2
(7)
5.2
S 6= 0
Here
X
ω(g)λ±
S (g) =
g∈G
λ+
S
X
q n ρ−rankQ ∆(Q)ψ(−Tr(SQ)).
Q∈S n
λ−
S
It follows that
and
appear with the same multiplicity in ω|G. Now note that
±
T
T
Tr(M SMQ) = Tr(SMQM ). So with M nonsingular, λ±
S and λM T SM also have the same
multiplicity in ω|G. Suppose that rankS = r. The number of members of S n congruent to
S (that is, of the form M T SM) is
|GL(r, q)|
n
|GL(r, q)|
n
×
=
,
×
r q
|O(S)|
|O(S)|
n−r q
where O(S) is the orthogonal group for S. We want to show that if r > 1, then this number is
more than (q n − 1)/2. That will imply that the only S 6= 0 that can appear in the characters
in ω|G are the ones of rank 1; there are q n − 1 of these [8, Theorem 13.2.47]. Again we
separate by parity.
• r = 2k: Then
|O(S)| ≤ O − (2k, q) = 2q k(k−1) (q k + 1)
k−1
Y
(q 2i − 1),
i=1
from (6)
So
|GL(2k, q)|
n
|GL(r, q)|
n
× −
≥
×
2k q |O (2k, q)|
|O(S)|
r q
=
n
Q
(q j − 1)
j=n−2k+1
2k
Q
q k(2k−1)
2q k(k−1)(q k + 1)
j=1
=
(q j − 1)
2(q k + 1)
× qk
k−1
Q
(q 2i − 1)
i=1
≥
(q j − 1)
j=n−2k+1
2(q k + 1)
k−1
Q
× qk
2
q 2i
i=1
=
n
Q
k−1
Q
(q 2i − 1)
i=1
n
Q
j=n−2k+1
n
Q
(q j − 1)
j=1
×
(q j − 1)
2k
Q
(q j − 1)
j=n−2k+1
2(q k + 1)
6
× qk .
2
We would like this to be more than (q n − 1)/2, that is,
n
Q
(q j − 1)
j=n−2k+1
2(q k + 1)
We need
n−1
Y
qn − 1
.
2
× qk >
(q j − 1) > 1 +
j=n−2k+1
1
.
qk
Since n − 1 ≥ n − 2k + 1, from k ≥ 1, and n − 2k + 1 ≥ 1, the product is nonempty
and its smallest factor is at least 2. Thus the inequality holds.
• r = 2k + 1: This time, again with the appropriate formula from (6) filled in,
|GL(r, q)|
n
×
=
r q
|O(S)|
=
n
Q
(q j − 1)
j=n−2k
2k+1
Q
j=1
n
Q
2
2q k2
− 1)
2k+1
Q
(q j − 1)
j=1
×
(q j
k
Q
(q 2j − 1)
j=1
(q j − 1)
j=n−2k
k
Q
q k(2k+1)
× qk
2 +k
× qk
2 +k
(q 2j − 1)
j=1
≥
n
Q
(q j − 1)
j=n−2k
2
k
Q
q 2j
j=1
=
n
Q
(q j − 1)
j=n−2k
.
2
We would also like this to be greater than (q n − 1)/2, and as long as k > 0, it is.
6
The decomposition of ω|G
Collecting the results of the preceding section gives
X
1 + δn + 1 − δn −
ω|G =
λ0 +
λ0 +
2
2
rankS=1
1
2q n(n−1)/2
7
X
Q∈S n
!
−
ρ−rankQ ∆(Q)ψ(−Tr(SQ)) (λ+
S +λS ).
We still need to determine which congruence class of symmetric matrices S of rank 1 actually
appears in the decomposition. (There are two such classes, corresponding to ∆(S) = 1 and
∆(S) = −1. Each class has (q n − 1)/2 members.) To do so, we examine the characters on a
small subgroup of G.
Let M be the n×n symmetric matrix with M11 =1 and all other entries 0; ∆(M) = 1. Let
I
0
H be the subgroup consisting of the matrices hα =
, α ∈ GF (q). Then ω(hα ) = q n
αM I
if α = 0, and ω(hα ) = q n ρ−1 χ(α) if α 6= 0, by (4). Moreover, λ±
βM (hα ) = ψ(αβ), for β 6= 0.
It follows that
(
)
X
n
n −1
±
−1
q +
q ρ χ(α)ψ(−αβ)
(ω, λβM )H = q
α6=0
= q n−1
= q
n−1
(
1 + δρ−1 χ(β)
X
χ(−βα)ψ(−βα)
α6=0
(1 + χ(β)δ),
)
by (1). So λ±
βM appears in ω|H just when χ(β) = δ. This implies the following:
Theorem 5
ω|G =
X
1 + δn + 1 − δn −
−
(λ+
λ0 +
λ0 +
S + λS ).
2
2
rankS=1
(8)
∆(S)=δ
In particular,
ω|B = λ0 + 2
X
λS .
(9)
rankS=1
∆(S)=δ
The Weil character is the sum of two irreducible characters, ω+ , of degree (q n + 1)/2, and
ω− , of degree (q n − 1)/2. Their values at −I are ω± (−I) = ±δ n (q n ± 1)/2 [14, Section 6], the
signs all matching. Observing the eigenvalues of −I in the corresponding representations,
we can write that
X
1 + δn + 1 − δn −
n
ω+ |G =
λδS
λ0 +
λ0 +
2
2
rankS=1
∆(S)=δ
ω− |G =
X
−δn
λS .
rankS=1
∆(S)=δ
As an immediate consequence we get that both ω+ |B and ω− |B are multiplicity free.
8
7
Confirmation of the ω|B decomposition
I 0
Recall that gQ =
. By (4), ω(gQ ) = q n ρ−rankQ ∆(Q). Let rankQ = r. Then (9) gives
Q I
ω(gQ) = q n ρ−r ∆(Q) = 1 + 2
X
ψ(Tr(SQ)).
(10)
rankS=1
∆(S)=δ
If x ∈ GF (q)n , x 6= 0, then xT x is a rank 1 symmetric matrix. Two such products xT x and
y T y are equal just when y = ±x. For x = (ξ1 , . . . , ξn ), the diagonal entries of xT x are the ξi2 .
Thus since at least one is nonzero, ∆(xT x) = 1. So all symmetric n × n rank 1 matrices can
be written as xT x (x 6= 0) or νxT x, where ν is a fixed nonsquare in GF (q). (As mentioned
above, there are (q n − 1)/2 matrices of each type.) We can rewrite (10) as follows:
X 1 + δ
1−δ
n −r
T
T
ω(gQ ) = q ρ ∆(Q) =
ψ(Tr(x xQ)) +
ψ(Tr(νx xQ)) .
(11)
2
2
x
The factors (1 ± δ)/2 pick out the S with ∆(S) = δ and adjust for the fact that each S
appears twice; the 1 on the right in (10) comes from x = 0. (Recall also that we have set
∆(0) = 1.)
Now
Tr(xT xQ) = Tr(xQxT ) = Tr(Q(x)) = Q(x),
so the preceding formula becomes
n −r
q ρ ∆(Q) =
X 1 + δ
x
1−δ
ψ(Q(x)) +
ψ(νQ(x)) .
2
2
(12)
Let σ denote the sum. To evaluate σ we need the number of times Q(x) = α for α ∈ GF (q).
These counts for nonzero α depend only on whether α is a square or not, since Q(βx) =
β 2 Q(x). Let 0 be taken on Z times (including Q(0) = 0); a given nonzero square S times;
and a given nonsquare N times. Then Z + (S + N)(q − 1)/2 = q n .
We also need the character sums
X
X
ρ−1
−ρ − 1
ψ(α) =
and
ψ(α) =
,
(13)
2
2
α nonsquare
α6=0 square
which follow from
X
α6=0
ψ(α) = −1,
X
ψ(α) −
X
α nonsquare
α6=0 square
9
ψ(α) = ρ.
Collect terms in σ according to Q(x):
1+δ 1−δ
1 + δ ρ − 1 1 − δ −ρ − 1
σ = Z
+S
+
+
2
2
2
2
2
2
1 + δ −ρ − 1 1 − δ ρ − 1
+
+N
2
2
2
2
S−N
S+N
+ δρ
.
= Z−
2
2
Then put Z = q n − (S + N)(q − 1)/2 to obtain
σ = qn −
q(S + N)
S−N
+ δρ
.
2
2
(14)
Now we need S and N. If Q0 is the nonsingular part of Q, the counts for Q are those for
Q0 multiplied by q n−r . Here are these numbers for Q, obtained from [7, Theorems 6.26 and
6.27] for Q0 :
r even:
r odd:
S
q
− δ ∆(Q)q n−r/2−1
q n−1 + δ (r−1)/2 q n−(r+1)/2 ∆(Q)
n−1
r/2
N
q
− δ ∆(Q)q n−r/2−1 = S
q n−1 − δ (r−1)/2 q n−(r+1)/2 ∆(Q).
n−1
r/2
Substituting into (14), we obtain simplifications corresponding to the parity of r.
• r even: then S = N and
q
σ = q n − (2q n−1 − 2δ r/2 ∆(Q)q n−r/2−1 )
2
r/2
= δ ∆(Q)q n−r/2 .
Since q r/2 = δ r/2 ρr , this is correctly q n ρ−r ∆(Q).
• r odd: then
q
σ = q n − 2q n−1 + δ (r−1)/2 q n−(r+1)/2 ∆(Q)δρ
2
(r+1)/2 n −(r+1)/2
= δ
q q
ρ∆(Q)
(r+1)/2 n (r+1)/2 −r−1
= δ
q δ
ρ
ρ∆(Q)
n −r
= q ρ ∆(Q),
again correct.
10
(15)
7.1
The S, N formulas
In point of fact, the formulas for S and N follow from those for ω|B. Combining (10) and
(12) gives
S−N
q(S + N)
+ δρ
.
(16)
q n ρ−r ∆(Q) = q n −
2
2
For a needed second equation, let Q′ be Q scaled by a nonsquare. Then the counts for Q′
are S ′ = N and N ′ = S, and ∆(Q′ ) = (−1)r ∆(Q). Formula (16) for Q′ reads
q n ρ−r (−1)r ∆(Q) = q n −
N −S
q(S + N)
+ δρ
.
2
2
(17)
Solving (16) and (17) for S and N produces
1 + (−1)r n−1 −r
1 − (−1)r n −r−1
q ρ ∆(Q) +
q ρ
δ∆(Q)
2
2
1 − (−1)r n −r−1
1 + (−1)r n−1 −r
q ρ ∆(Q) −
q ρ
δ∆(Q),
N = q n−1 −
2
2
S = q n−1 −
and then
q−1
2
1 + (−1)r n−1
n−1
q (q − 1)ρ−r ∆(Q).
= q
+
2
Z = q n − (S + N)
These are uniform expressions (perhaps not obvious in traditional derivations!) which give
(15) on taking r even or odd.
Incidentally, when δ = −1 or q is not a square, ρ is not rational. In that case, the formulas
follow from equating coefficients in (16) for the field Q(ρ) = Q + Qρ and again solving for S
and N.
8
A q-binomial identity
Recall equation (7):
X
ρ−rankQ ∆(Q) = q n(n−1)/2 .
Q∈S n
If rankQ is odd, then with ν a nonsquare in GF (q), ∆(νQ) = −∆(Q), and the terms for Q
and νQ in the sum cancel. Thus as ρ−2k = δ k q −k ,
⌊n/2⌋
q
n(n−1)/2
=
X
−rankQ
ρ
∆(Q) =
Q∈S n
X
k=0
11
X
Q∈S n
rankQ=2k
δ k q −k ∆(Q).
(18)
Let rankQ = 2k. As pointed out in Section 3, if Q is hyperbolic, then ∆(Q) = δ k ; and if Q
is elliptic, then ∆(Q) = −δ k .
P
δ k q −k ∆(Q). The number of terms with Q
Now we can give a specific formula for
Q∈S n
rankQ=2k
hyperbolic is
n
2k
×
|GL(2k, q)|
,
|O + (2k, q)|
×
|GL(2k, q)|
.
|O − (2k, q)|
q
and the number with Q elliptic is
n
2k
q
Thus
X
k −k
δ q
k −k
∆(Q) = δ q
Q∈S n
rankQ=2k
n
2k
× |GL(2k, q)| × δ
k
q
1
1
;
−
|O + (2k, q)| |O − (2k, q)|
the second δ k is the factor needed for ∆(Q). By (6),
1
|O + (2k, q)|
−
1
|O − (2k, q)|
1
=
q k(k−1)
k
Q
(q 2i − 1)
i=1
(note the addition of one more factor in the product). So, again by (6),
⌊n/2⌋
X
k=0
⌊n/2⌋
X
δ k q −k ∆(Q) =
Q∈S n
rankQ=2k
X
δ k q −k
k=0
=
k=0
n
2k
q k(2k−1)
× δk
q
q k(k−1)
2k
Q
(q i − 1)
i=1
k
Q
(q 2i − 1)
i=1
⌊n/2⌋
X
k−1
Y
n
2
(q 2i+1 − 1).
q k −k
2k q
i=0
Thus from (18),
q
n(n−1)/2
=
⌊n/2⌋
X
k=0
k−1
Y
n
k 2 −k
q
(q 2i+1 − 1).
2k q
i=0
Is that really true? Yes – it gives the count of the numberQof symplectic (skew-symmetric)
2
k
2i
n × n matrices over GF (q). From [12], |Sp(2k, q)| = q k
i=1 (q − 1), so the number of
12
symplectic matrices of rank 2k is
n
2k
|GL(2k, q)|
=
×
|Sp(2k, q)|
q
=
n
2k
q k(2k−1)
q
q k2
k
Q
2k
Q
(q i − 1)
i=1
(q 2i − 1)
i=1
k−1
Y
n
2
q k −k
(q 2i+1 − 1)
2k q
i=0
[8, Theorem 13.2.48], and that is to be summed from k = 0 to ⌊n/2⌋ (giving 1 at k = 0).
But the total number of symplectic n × n matrices is simply q n(n−1)/2 .
9
Minimum Weil-free subgroups of G
Suppose that H is a subgroup of G. If ω|H is also Weil-free, then |H| ≥ q n . It cannot be
that |H| = q n , because then ω|H would just be the sum of all q n linear characters of H. But
that sum is q n at I and 0 at h 6= I, whereas ω(h) 6= 0, by (4). Thus |H| ≥ 2q n . We shall show
that there are Weil-free subgroups of order 2q n . If |H| = 2q n , then H = h−Ii × (H ∩ B).
Adapting the orbit count in the proof of Theorem 1, we find that 2q n ×(number of orbits)
of such an H is
XX
H(x,y) ≥ 2q n + (q n − 1)q n + q n (q n − 1) = 2q 2n .
2q n + (q n − 1)q n +
x
y6=0
So if H is to be Weil-free, each H(x,y) with y 6= 0 must be just {I}. Again as in the proof
of Theorem 1, this means that if gQ ∈ H, with Q 6= 0, then Q must have full rank n. So
what would work is an n-dimensional subspace W of S n whose nonzero members are all
nonsingular. Then H ∩ B would be {gQ |Q ∈ W }.
To construct W , realize GF (q)n as GF (q n ) and let tr be the trace function GF (q n ) →
GF (q). Then for α ∈ GF (q n ), the function Qα given by Qα (ζ) = tr(αζ 2) is a quadratic form
on GF (q n ). The corresponding bilinear form is Bα (ξ, η) = tr(αξη). This is nondegenerate
when α 6= 0, making Qα nonsingular then. Now let W = {Qα |α ∈ GF (q n }.
9.1
An evaluation for H
Formula (12) for Qa reads
n −n
q ρ
∆(Qα ) =
X
ζ∈GF (q n )
1+δ
1−δ
ψ(Qα (ζ)) +
ψ(νQα (ζ)) .
2
2
Because ν ∈ GF (q) and ρ2 = δq, this becomes
X 1 + δ
1−δ
n n
2
2
δ ρ ∆(Qα ) =
ψ(tr(αζ )) +
ψ(tr(ναζ )) .
2
2
n
ζ∈GF (q )
13
(19)
One has to be careful with the matrix interpretation. Let ζ1 , . . . , ζn be a GF (q)-basis of
GF (q n ). Then the matrix for Bα is [tr(αζi ζj ]. This can be written as
ζ1 ζ1q
ζ ζ q
2
[tr(αζi ζj ] = . .2
.. ..
ζn ζnq
n−1
α 0
. . . ζ1q
q n−1 0 αq
. . . ζ2
.. ..
..
...
. . .
n−1
0 0
. . . ζnq
= D × diag(α, αq , . . . , αq
n−1
...
...
0
0
..
.
ζ1
ζ1q
..
.
ζ2
ζ2q
..
.
...
...
ζn
ζnq
..
.
Qn−1
αq is the
...
...
n−1
q n−1
q n−1
q n−1
q
... α
. . . ζn
ζ2
ζ1
) × DT .
Taking determinants gives
det [tr(αζi ζj ] = (det D)2
Yn−1
i=0
i
αq = (det D)2 N(α),
in which (det D)2 is the discriminant of the extension GF (q n )/GF (q) and
norm N(α) of α. Then
∆(Qa ) = χ((det D)2 )χ(N(α)).
i=0
i
Applying the automorphism ξ → ξ q to D cycles its columns; so (det D)q = (−1)n−1 det D,
the sign being that of an n-cycle. Thus χ((det D)2 ) = (−1)n−1 : (det D)2 is a square in
GF (q) only when its square-root det D is in GF (q)! For χ(N(α)) we have
χ(N(α)) = N(α)
q−1
2
qn −1 q−1
q n −1
2
= α q−1
= α 2 = X(α),
where X(α) is the quadratic character of α for the field GF (q n ) (we can determine any χ(z)
by reading it in GF (q)). All together, ∆(QαP
) = (−1)n−1 X(α).
For the right side of (19), we have that ζ ψ(tr(βζ 2)) = X(β)P , where P is the “ρ” for
GF (q n ), by the formulas in (13) for GF (q n ). Moreover, since ν is a nonsquare in GF (q),
X(ν) = (−1)n . Therefore (19) becomes
1+δ
1−δ
n n
n−1
n
δ ρ (−1) X(α) =
X(α)P +
(−1) X(α)P
2
2
1 − (−1)n
1 + (−1)n
,
+δ
= X(α)P
2
2
or
n n
δ ρ (−1)
This simplifies to
n−1
=P
1 + (−1)n
1 − (−1)n
+δ
2
2
.
P = (−1)n−1 ρn ,
on sorting by the parity of n. That is a particular instance of the Davenport-Hasse theorem
on lifted Gauss sums [7, Theorem 5.14].
14
References
[1] Gérardin, P. (1977). Weil representations associated to finite fields. J. Algebra 46:54–
101.
[2] Guralnick, R., Magaard, K., Saxl, J., Tiep, P. H. (2002). Cross characteristic representations of symplectic and unitary groups. J. Algebra 257:291–347.
[3] Gurevich, S., Howe, R. (2016). Small representations of finite classical groups.
arXiv:1609.01276v1.
[4] Heide, G., Zalesski, A. E. (2006). Passman’s problem on adjoint representations. In: W.
Chin, et al., ed. Groups, rings and algebras. Contemp. Math., 420, Amer. Math. Soc.,
Providence, RI:163–176.
[5] Howe, R. (1973). On the character of Weil’s representation. Trans. Amer. Math. Soc.
177:287–298.
[6] Isaacs, I. M. (1973). Characters of solvable and symplectic groups. Amer. J. Math.
95:594–635.
[7] Lidl, R., Niederreiter, H. (1997). Finite Fields, Cambridge University Press.
[8] Mullen, G. L., Panario, D., editors. (2013). Handbook of Finite Fields, CRC Press, Boca
Raton.
[9] Pallikaros, C. A., Ward, H. N. (2015). Commutative nilpotent closed algebras and Weil
representations. Comm. Algebra 43:4839–4859.
[10] Pallikaros, C. A., Zalesski, A. E., (2011). Weil representations and some nonreductive
dual pairs in symplectic and unitary groups. Comm. Algebra 39:2156–2178.
[11] Thomas, T. (2008). The character of the Weil representation. J. Lond. Math. Soc. (2)
77:221–239.
[12] Taylor, D. E. (1992). The Geometry of the Classical Groups, Helderman Verlag, Berlin.
[13] Ward, H. N. (1972). Representations of symplectic groups. J. Algebra 20:182–195.
[14] Ward,
H.
N.
(2016).
Matrices
for
the
http://people.virginia.edu/~hnw/MatrixWeil.pdf
15
Weil
representation.
| 4math.GR
|
A Unified Framework for Wide Area Measurement System Planning
James J.Q. Yua,∗, Albert Y.S. Lama , David J. Hilla , Victor O.K. Lia
a Department
of Electrical and Electronic Engineering, The University of Hong Kong, Pokfulam, Hong Kong
arXiv:1711.07652v1 [cs.SY] 21 Nov 2017
Abstract
Wide area measurement system (WAMS) is one of the essential components in the future power system. To make
WAMS construction plans, practical models of the power network observability, reliability, and underlying communication
infrastructures need to be considered. To address this challenging problem, in this paper we propose a unified framework
for WAMS planning to cover most realistic concerns in the construction process. The framework jointly optimizes the
system construction cost, measurement reliability, and volume of synchrophasor data traffic resulting in a multi-objective
optimization problem, which provides multiple Pareto optimal solutions to suit different requirements by the utilities. The
framework is verified on two IEEE test systems. The simulation results demonstrate the trade-off relationships among
the proposed objectives. Moreover, the proposed framework can develop optimal WAMS plans for full observability with
minimal cost. This work develops a comprehensive framework for most practical WAMS construction designs.
Keywords: Wide area measurement system, Dual-Use Line Relays, Phasor Measurement Units, construction,
optimization.
1. Introduction
Wide Area Measurement System (WAMS), as a reliable monitor of the power network, is considered one of
the most important components in the smart grid [1]. In
contrast to the current supervisory control and data acquisition (SCADA) system, measurements of the system
states are conducted at a much higher rate (5–60 samples per second versus one per 2–6 seconds). In addition,
all system phasors are developed simultaneously and continuously, rendering real-time knowledge of power system
parameters possible [2]. As a result, WAMS can significantly improve the performance of power grids by supporting more accurate state estimation, fault detection,
stability assessment, remedial control actions, etc. [3, 4].
A typical WAMS comprises synchrophasor measurement devices and Phasor Data Concentrators (PDCs) for
aggregating and relaying synchrophasor data. These two
major components form a hierarchical structure, connecting through a communication infrastructure (CI). While
Phasor Measurement Units (PMUs) are widely employed
in WAMS, the recently available Dual-Use Line Relays
(DULRs) introduce variability to the modern WAMS construction [5]. DULRs are the protection digital relays
for transmission lines and transformers which can report
synchrophasor data while providing system protection [6].
Due to its characteristics of being installed along transmission lines and at transformers, DULR is also called “branch
∗ Corresponding
author
Email addresses: [email protected] (James J.Q. Yu),
[email protected] (Albert Y.S. Lam), [email protected] (David
J. Hill), [email protected] (Victor O.K. Li)
Preprint submitted to Elsevier
PMU” in some previous research [7, 8]. Although DULR
can only monitor the voltage phasor of its adjacent bus
and the current phasor of the branch, it is still promising
due to its low construction cost [5].
Despite much work on WAMS planning, there seems
no unified WAMS planning framework that jointly optimizes multiple important objectives simultaneously for
placing both measurement devices and PDCs in the system. Moreover, despite the decreasing device costs for constructing WAMS, utilities are still a long way from achieving full-installation of PMUs and PDCs across the grid.
Meanwhile, better WAMS construction strategies are still
welcomed due to their better system reliability and costefficient properties.
In addition, much previous work suffers from unrealistic assumptions, which have been thoroughly discussed
in [5]. For instance, while some work considers installing
PMUs at buses, they should actually be placed at substations which is a collection of multiple buses. When a
PMU or DULR is being installed, the respective substation needs to be interrupted leading to a substantial cost
in WAMS construction [9]. Consequently a comprehensive model is required to account for all kinds of WAMS
installation costs.
Moreover, as PMUs are generally assumed to be installed on buses in the literature, two buses connected with
transformers are both considered observable if either one
is equipped with a PMU. However, this hypothesis relies
on a model of transformer tap positions as fixed network
parameters. The estimated bus voltages, power flows and
injections with a transformer with incorrectly modeled or
inaccurately measured tap ratio can deviate significantly
November 22, 2017
from their actual values, resulting in inaccurate system
state estimation [5, 10]. Last but not least, a majority
of previous work assumes that PMUs are equipped with
unlimited measurement channels to observe the current
phasors of all connecting branches. Other work focuses
on minimizing the number of channels required in WAMS
[11]. However, none of the existing work determines which
branches are observed by each PMU. When given optimal
branch allocations for PMUs, we can further improve the
measurement performance of a WAMS.
In this paper we propose a unified framework, aiming
to fulfill different construction requirements for WAMS.
The main contributions of this work are listed as follows:
achieve the optimal solution of this NP-hard problem, including but not limited to integer programming [12], metaheuristics [13], exhaustive search [14], weighted least square
algorithm [15], etc. Interested readers can refer to [16] and
[17] for more details of PMU placement methodologies.
Besides construction cost, system measurement reliability is also critical for WAMS construction. System
states of line outages and loss of measurements need to be
considered in order to design a robust and reliable WAMS.
An intuitive solution to address these system failures is
to install duplicate measurement devices to observe the
same bus and this is called measurement redundancy. Due
to its simplicity, this technique has been widely adopted
(see [12, 18] for examples). However this scheme may potentially lead to over-installation of PMUs in the system
[5]. As an alternative, a reliability-based PMU placement
model is proposed where the possibility of maintaining full
observability is investigated [8]. This model considers a
trade-off between the total number of PMUs installed and
the WAMS reliability, resulting in more versatile placement solutions [17].
Concurrently optimizing PMU and PDC placement is
another research direction related to WAMS planning. [4]
manipulates the placement of measurement devices and
PDCs to construct multiple data paths for the generated
synchrophasors to overcome CI failures in WAMS. [11] and
[19] try to minimize the system scale of CI to reduce the
WAMS construction cost.
There is also recent work analyzing WAMS construction from the perspective of graph theory and network
equivalency [20]. While satisfying the conventional bus
observability constraint, [21] also enables estimation of
system dynamic models by network reduction approaches.
The results can be further employed to update the offline
system model.
Besides, there is also research investigating the integration of WAMS construction with other power system applications and services [20, 22]. For instance, state estimation
is among the most important power system applications
which can greatly benefit from WAMS. In [23], the impact
of PMU placement plans on the reliability of state estimation subject to data integrity issues is investigated. The
proposed mechanism can also provide estimated system
parameters given measurement redundancy. Other analyses take power stability analysis [24, 25, 26] and oscillation
monitoring [27] into consideration. Another widely investigated integration considers power system cyber-security.
For passive cyber-attack prevention, constructing WAMS
and its protection facilities strategically is widely adopted
[28]. Much research has been conducted in this direction,
see [29, 30] for instance. References [31] and [32] provide
thorough surveys on this topic. Another interesting direction considers communication infrastructure in designing
WAMS construction and operation plans. For instance,
through proper software layer design, data communication quality-of-service can be guaranteed [33]. Cloud computing may also greatly contribute to improving the com-
• We propose a unified framework for WAMS planning, in which PMU, DULR, and PDC placements
are jointly optimized simultaneously.
• We consider a realistic cost model for WAMS construction including the power system substation interruption cost during installation.
• We consider a practical substation model with unknown transformer tap ratio, which can better facilitate the utilization of measured system synchrophasors.
• We consider the channel limits of PMUs, and jointly
determine which branches should be observed, aiming to provide full observability with the least devices.
• Pareto solutions can be determined for decision making considering various requirements of WAMS.
The rest of this paper is organized as follows. Section
II gives a brief literature review on WAMS construction
research. Section III introduces the system model which
allows us to design a unified WAMS planning framework.
Section IV formulates a multi-objective WAMS planning
problem for developing WAMS construction plans. Section V demonstrates the implementations of the proposed
framework on IEEE test systems, and compares them with
the state-of-the-art solutions. Finally, we conclude this paper in Section VI with discussions on the proposed framework.
2. Related Work
With the increasing demand of synchrophasor measurement in modern power systems, utilities need a methodology to construct WAMS optimally. Canonically most
optimal WAMS construction work focused on finding the
minimal number of PMUs to ensure full system observability, subject to pre-defined constraints. This so-called
optimal PMU placement problem and its variants have
been investigated intensively in the past two decades. A
wide range of solution methods have been developed to
2
by a PMU installed on bus i. Consequently, the current
phasors of these branches and the voltage phasor of i are
measured while the voltage phasors of Ni0 can be calculated. This makes buses i and Ni0 observable. Similarly, a
DULR placed at the i end of branch (i, j)1 measures the
voltage phasor of i and current phasor of (i, j). This makes
the voltage phasor of j calculable. Hence both buses i and
j are considered observable.
While buses with measurement devices can have their
voltage phasors directly measured, other buses need the
current phasors of their connecting branches to infer their
voltages. However, these current phasors may not always be available due to possible line outage contingencies. To specify any of these contingencies, we define a
system state s as a combination of different line outages,
where E s ⊂ E represents the set of failed power transmission lines of system state s. Moreover, we describe the
bus connectivity of s with As = (aij,s ) ∈ B|V|×|V| , where
B = {0, 1} is the Boolean domain. We have aij,s = 1 if
(i, j) ∈ E \ E s or i = j, otherwise aij,s = 0. For simplicity,
A = (aij ) is employed to stand for the bus connectivity of
system state s0 , where there is no line outage in the system. Furthermore, the bus observability of s is denoted
by Os = (oi,s ) ∈ B|V|×1 , where oi,s = 1 means bus i is
observable in s and zero otherwise.
Communication Infrastructure
GPS Clock
PMU
PDC
Central
Application
GPS Clock
DULR
Figure 1: Components of a wide area measurement system.
munication and computing network efficiencies [34]. The
above are examples of existing research on applications
and extensions of the WAMS construction problem. They
all indicate the significance of optimal WAMS construction strategies. However, due to the limitations of most
previous WAMS construction work introduced in Section
1, a generalized and realistic formulation of the WAMS
construction problem is required.
3. System Model
Fig. 1 depicts a schematic WAMS architecture, where
PMU, DULR, PDC, and the central controller form a hierarchical structure over CI, which serves as the media
for data transmission. PMU and DULR interface WAMS
with the power system and they comprise current transformers (CTs), voltage transformers (VTs), instrumentation cables, and synchronous GPS clocks. Synchrophasors
measured by these devices are transmitted to one or multiple layers of PDCs located at selected locations in the
system, where the data are aggregated, compressed, and
sorted into a time-stamped measurement stream [35]. In
general, the data stream is then fed into application software at the central controller for system state monitoring
and control decision generation with various control objectives. For simplicity, in this work we assume one layer of
PDCs is utilized in WAMS, while multiple layers can be
easily adopted into the proposed framework.
3.2. Phasor Measurement Reliability
Designing a robust WAMS for complete system observability with the presence of system failures is critical in
WAMS planning. To achieve this, one typical method is
to provide measurement redundancy for each bus in the
system. That is, we employ multiple measurement devices to provide independent observability to each bus.
However, usually more than half of the substations in the
system need to be equipped with PMUs in order to provide full observability under N − 1 contingencies, but such
over-installation is not practical [5]. Thus we do not consider full observability guarantee under contingencies in
this way.
Instead we address system observability in the probability sense. In practice, two types of failures are commonly considered in WAMS design, namely, line outages
and measurement device failures. In this work, both kinds
of failure are considered. We address the line outage failures by minimizing the number of unobservable buses for
different system states to improve the system reliability.
We also provide multiple measurement devices to observe
critical buses to minimize the impact of measurement device failures on the whole system, as will be described in
Section 4.3.3.
Each system state s with line outages has a probability
3.1. Power System and Network Observability
We model the power network with an undirected graph
G(V, E), where V and E are the sets of buses and branches,
respectively. There are K substations, each of which, denoted by Sk , comprises a subset of buses and thus we may
simply write Sk ⊂ V. For bus i ∈ V, we denote the set of
its neighboring buses by Ni = {j ∈ V|(i, j) ∈ E}.
A power system is deemed observable if the bus voltage phasor (both magnitude and angle) at every bus in the
network is known. We consider that maintaining full system observability is one of the critical objectives of WAMS,
where the voltage phasor of every bus can either be directly
measured or indirectly inferred by calculation. Considering the PMU measurement channel limit, branches associated to Ni0 ⊆ Ni , i.e., Ei = {(i, j)|j ∈ Ni0 }, are observed
1 For simplicity, we write “DULR installed on bus i” to represent
the DULR placed at the i end of branch (i, j) when no confusion will
be caused thereafter.
3
of occurrence calculated by
Y
Y
ps =
pl
(1 − pl ),
l∈E\E s
path from i to j over CI. Therefore, the maximum data
traffic originated from bus i is (|Ni |+1)LF Hi bit-hop-persecond, where Hi is the number of hops of the shortest
path from bus i to any PDCs over graph G. Similarly,
the data traffic from PDC to the central controller is defined as ηNiPDC HiPDC , where η is the data compression
ratio, NiPDC is the aggregated data rate received by PDC
installed on i, and HiPDC is the number of hops of the
shortest path from bus i to the central controller.
(1)
l∈E s
where pl is the reliability of power line l. With the measurement devices fixed in the system, the observability of
each bus can be developed through considering failed lines
in s. We define the unreliability index of each bus subject
to current measurement device allocation as
X
Ui =
ps (1 − oi,s ),
(2)
4. WAMS Planning Problem
s∈S
Based on the models discussed in Section 3, we formulate a multi-objective realistic optimal WAMS planning
problem (WPP) in this section. The objective of WPP is
to optimally place and set up PMUs, DULRs, and PDCs
in the power system such that
where S is the set of all system states. We further define
the overall system unreliability index by aggregating the
indices of all buses in the system, i.e.,
X
U WAMS =
Ui .
(3)
i∈V
• the total WAMS construction cost is minimized;
3.3. Communication Network
PMU and DULR devices measure the bus voltage and
current phasors of their attached branches. Each measurement device generates a data stream to be transmitted to a
specific PDC for data aggregation, which then passes the
compressed synchrophasor data to the central controller
for certain power applications. To reduce the possible data
transmission delay due to communication link congestion
and data queuing, we should carefully place PMUs and
DULRs on the power network in order to reduce the total
data traffic transmitted from measurement devices to both
PDCs and the central controller.
According to the IEEE Standard for Synchrophasor
Data Transfer for Power Systems [36], the maximum data
rate of traffic generated by measurement devices at bus i,
can be estimated as (|Ni | + 1)LF bit-per-second, where
L and F are the data size of each synchrophasor data
message and synchrophasor sampling rate2 , respectively.
Here current phasors of |Ni | adjacent branch and one bus
voltage phasor give |Ni | + 1 sets of measurement data.
After these synchrophasors are generated, they are sent
to one dedicated PDC for data aggregation, compression,
and relaying. Then the processed data is effectively compressed with any appropriate data compression schemes,
e.g., Exception Compression and Swing Door Trending
Compression [37], and later relayed to the central controller for data utilization. These two kinds of data transmission are achieved through CI, which is in turn supported by an existing telecommunication network. We do
not impose any limitations on the topology of CI. We assume that all substations are connected through CI, which
can be configured with any reasonable topology [38]. We
employ the distance matrix Q = (qij ) ∈ Z|V|×|V| to describe CI, where qij is the number of hops on the shortest
• the overall system unreliability index is minimized;
• the total WAMS synchrophasor data traffic from measurement devices to PDCs is minimized;
• all buses are observable with unknown transformer
tap ratio; and
• all system constraints are satisfied.
4.1. Control and Ancillary Variables
To formulate the problem, we define several control
variables, namely, M, D, P, and L. M = (mij ) ∈ B|V|×|V|
is the PMU installation and observation matrix, and it
determines whether PMUs are installed on specific buses,
and whether the current phasors of the respective connected branches are measured. Specifically, if a PMU is
installed on bus i implying that the voltage phasor of i is
measured, then the diagonal value mii will be set to one.
Moreover, if the current phasor of a branch (i, j) is measured, the corresponding non-diagonal mij value will also
be set to one.
Similarly, D = (dij ) ∈ B|V|×|V| is the DULR installation matrix, and if we have dij = 1, a DULR will be
installed on branch (i, j) on the i end. However, unlike M,
we always set all diagonal values dii ’s to zero and this will
make the subsequent formulation simpler.
We also let P = (pi ) ∈ B|V|×1 to denote the PDC installation status, where pi = 1 indicates that a PDC is
installed on bus i. Finally, L = (lij ) ∈ B|V|×|V| is introduced to allocate each PMU and its associated DULR
devices to a PDC. If we set lij = 1, PMU and all DULRs
installed on bus i will be assigned to transmit their measured synchrophasors to the PDC installed on bus j.
Besides these control variables, we also define three
ancillary variables to facilitate the formulation, including
M0 , D0 , and U. We use M0 = (m0i ) ∈ B|V|×1 as a PMU
installation indicator, where m0i = 1 if a PMU is installed
2 Here the data frame overhead is omitted as it is normally very
small when compared with the synchrophasor data message body.
4
at bus i. In addition, D0 = (d0i ) ∈ B|V|×1 is a measurement device installation indicator, where d0i = 1 if either a
PMU or a DULR is installed at bus i. Meanwhile, when
a measurement device is planned to be set up at bus i,
the associated substation Sk 3 i needs to be interrupted
for installation. Thus we use U = (ui ) ∈ Bk×1 to indicate
whether each substation will be interrupted, incurring disruption cost. Although M0 , D0 , and U can be implied
from M and D, we still adopt these ancillary variables to
simplify the problem formulation.
where c is the location of the central controller, rendering
HiPDC = qic . Multiplying the maximum data rate from
bus i with qij LF gives its data traffic to the PDC and the
term qjc ηLF specifies the compressed data traffic from
PDC to the central controller. The total data traffic in
WAMS is developed by aggregating these two components
as given in (5).
4.3. Constraints
There are a number of constraints governing the control
and ancillary variables.
4.2. Objective Functions
4.2.1. WAMS Construction Cost
The construction cost of WAMS, C WAMS , can be defined as a combination of the costs for all components involved, i.e.,
X
XX
DULR
C WAMS =
CiPMU m0i +
Cij
dij
i∈V
+
X
i∈V j∈V
CiPDC pi
i∈V
CiPMU ,
4.3.1. Control Variable Constraints
We introduce several constraints to define the characteristics of the control variables M, D, M0 , D0 , and L:
DULR
Cij
,
+
K
X
CkMS uk ,
d0i ≥
(4)
mij = 0, ∀(i, j) ∈
/ E and i 6= j
(6)
∀(i, j) ∈
/E
(7)
∀i, j ∈ V
(8)
∀i, j ∈ V
(9)
lij ≤ pj , ∀i, j ∈ E
(10)
dij = 0,
0
mi ≥ mij ,
m0i , d0i ≥ dij
k=1
CiPDC ,
where
and
are the installation costs
of the PMU, DULR, and PDC installed at bus i, respectively,
and CkMS is the disruption cost of substation k.
P
j∈V dij corresponds to the total number of DULRs installed on branches
with bus i on the i end
P
P connected
DULR
and hence
dij determines the cost of
i∈V
j∈V Cij
installing all DULRs in the system.
The above objective function presents a general cost
model for WAMS construction. However, in practice, utilities may take advantage of existing substation work that
already receives an outage to install devices. This can be
reflected in the model by setting the corresponding substation interruption cost to zero. In other words, installing
PMUs and DULRs in the substation does not incur additional interruption cost and the optimization problem will
favor the substation due to achieving lower total cost.
X
lij =
d0i ,
∀i ∈ E.
(11)
j∈V
(6) imposes that PMUs can only measure the current phasors of existing branches in the system, and (7) confines
that DULRs can only be installed on existing branches.
Meanwhile, based on the definitions of M and D given in
Section 4.1, mii = 1 if a PMU is installed on bus i, and
dij = 1 if a DULR is installed on branch (i, j) at the i
end. In such cases, we have m0i = 1 and d0i = 1 according
to (8) and (9), respectively. Finally, synchrophasors generated by measurement devices should be sent to PDCs
installed at some buses and this gives (10). Moreover, as
each measurement device should report to one PDC, and
buses without measurement devices should be excluded
from data traffic calculation, we have (11).
4.3.2. Bus Observability Constraint
According to (2), oi,s ’s are required to calculate Ui . As
each bus can be observed by either PMUs or DULRs, oi,s
is constrained by:
X
oi,s ≤
aij,s (mji + dji ) + dij , ∀i ∈ V.
(12)
4.2.2. Overall System Unreliability Index
As defined in Section 3.2, WAMS unreliability index
U WAMS gives the degree of unreliability subject to all system states. A larger U WAMS means the system has a
higher possibility of becoming incompletely observable with
line outages. Consequently, the system reliability can be
improved by minimizing U WAMS as an objective function
in the proposed multi-objective WPP.
j∈V
In (12), aij,s mji determines whether a PMU is installed
on bus j and measures the current phasor of branch (j, i),
rendering bus i observable. aij,s mji also includes the cases
with i = j, where PMU installed on i directly measures
the voltage phasor of i. Similarly, aij,s dji describes the
observability of bus i by DULR on branch (j, i) at the j
end. Minimizing U WAMS is equivalent
to maximizing all
P
P
oi,s ’s based on (12) as U WAMS = i∈V s∈S ps (1 − oi,s ),
according to (2) and (3). Thus, when bus i is observable,
the right-hand side of (12) is greater than or equal to one,
resulting in oi,s = 1.
4.2.3. Total WAMS Data Traffic
Based on our communication network model, the maximum system data traffic
P rate received by PDC on bus j
can be formulated as i∈V (|Ni + 1|)lij . We can determine
the system data traffic by aggregating the rates of all measurement devices and PDCs. Thus, the maximum system
data traffic rate DWAMS is given by:
XX
DWAMS =
(qij + qjc η)LF (|Ni | + 1)lij
(5)
i∈V j∈V
5
4.3.3. Redundant Observability Constraint
WAMS needs to ensure system observability. While
the unreliability objective can improve the WAMS reliability subject to transmission line outage failures, it is
more practical to provide N − t measurement redundancy
to the critical buses to counteract both line outages and
measurement device failures, where t is the measurement
redundancy degree. At least t + 1 devices should ensure
observability to each critical bus to support N −t measurement redundancy. Other non-critical buses should provided with full observability under s0 . We can guarantee
N − t observability for critical buses and full observability
for non-critical buses with the following:
X
[aij (mji + dji ) + dij ] ≥ t + 1, ∀i ∈ V.
(13)
4.3.6. Prohibited Substations
In practice, some substations in the power system may
not be interrupted at all due to various reasons. For example, substations serving critical loads may undermine
the power network security if disrupted. It can also be difficult to install measurement devices on rural substations
where no proper CI has been established. In such cases,
the installation of PMU and DULR will be prohibited by
introducing the following constraint:
mij = 0 and dij = 0, ∀i ∈ V prohibit , j ∈ V,
(16)
where V prohibit is the set of buses located in these prohibited substations.
4.3.7. Existing PMU Devices
In some situations, some substations may have prej∈V aij (mji + dji ) determines if bus i can be observed
installed PMUs, which can be integrated into WAMS. An
by
P the PMUs, or DULRs on branch (j, i) at the j end.
additional constraint is imposed to describe the observj∈V dij is the total number of DULRs installed on branches
ability introduced by these PMUs:
connecting to bus i at the i end, which can all provide voltage phasor measurement for i. The summation of these
mij = 1, ∀i ∈ V exist , j ∈ Viobserve ,
(17)
two terms gives the total number of measurement devices
observing i, which is greater than t to maintain N − t obwhere V exist is the set of buses with PMUs pre-installed.
servability for the critical buses. For non-critical buses, t
For each bus i in V exist , Viobserve is the set of other obis assigned with zero to guarantee basic full observability.
served buses connected to i by those branches whose curThere exists some work about identifying critical buses in
rent phasors are measurable. Moreover, the cost of these
a power system, e.g., [39], and we may utilize the results
PMUs should be removed from C WAMS , as originally (4)
thereafter to determine t for each bus in the system.
includes the cost of all PMUs in the system. This can be
achieved by modifying (8) to
4.3.4. PMU Measurement Channel Constraint
PMUs are provided with a limited number of measurem0i ≥ mij , ∀i ∈ V \ V exist , j ∈ V and m0i = 0, ∀i ∈ V exist .
ment channels for data communications. It is inevitable
(18)
that some branches cannot be observed even though the
as m0i is involved in (4) to calculate the construction cost.
adjacent buses have PMUs installed. Hence, to be more
(18) excludes the existing PMUs by setting their respective
practical, we introduce the following constraint:
m0i values to zero, thus making (4) accurate.
X
mij ≤ T PMU , ∀i ∈ V,
(14)
4.4. Multi-objective Optimization Problem
j∈V
P
j∈V\i
Utilizing previously defined objective functions and constraints, we formulate a multi-objective optimization problem as follows:
PMU
where T
is the maximum allowed number of current
phasor measurement channel for PMU.
4.3.5. Substation Interruption Constraint
When a PMU or DULR is installed, the involved substation needs to be interrupted. Here we employ U to
indicate whether each of the substations needs to be interrupted and this is illustrated with the following:
minimize
(C WAMS , U WAMS , DWAMS )
M,D,P,L,M0 ,D0 ,U
subject to
(6) − (18)
This optimization problem will develop the optimal WAMS
plan based on these conflicting objectives; minimizing WAMS
unreliability results in installing measurement devices at
all possible positions, which in turn drastically increases
the total construction cost; minimizing the total data traffic requires equipping all measurement devices with PDCs
for data compression, but this increases the cost. Therefore, a Pareto frontier will be developed for decision-making.
uk ≥ mij and uk ≥ dij , ∀i ∈ Sk , j ∈ V, k = 1, · · · , K.
(15)
As illustrated in Section 4.3.1, mii or dij is set to one if
a PMU or DULR is installed on bus i in substation k.
In such cases, the corresponding substation needs to be
interrupted, making uk = 1 by (15).
Constraints (6)–(15) describe a set of mandatory constraints on the control variables and WAMS observability
requirement. Besides, there are also some constraints corresponding to practical WAMS requirements.
6
4.5. Discussions
The multi-objective optimization problem defined above
provides solutions to a generalized WAMS construction
problem considering different constructional objectives. The
construction deduced from the solutions may take place
over a specified period of time. As such, there may be
a transitional period in which WAMS may work together
with the conventional SCADA system before being fully
deployed. The operational issues have been addressed
in previous work (see [40, 41] for examples). In addition, some previous research suggests to place PMU over
SCADA, e.g., [42, 43]. Typically this can be achieved by
including additional constraints for SCADA in the conventional PMU placement problem. Our formulated multiobjective optimization problem can also adopt these constraints with ease, and the resulting problem can handle
the WAMS construction problem considering an existing
SCADA system.
Sometimes utilities may impose a bandwidth limit for
WAMS data transmission instead of minimizing the data
traffic. This can be addressed by transforming the corresponding objective function(s) into constraints. For example, suppose that the total data traffic from bus i to
PDC at bus j is Dij . The data traffic objective (4) can
be transformed into a series of constraints, each of which
limits the bus-to-PDC data traffic as follows:
LF (|Ni | + 1)lij ≤ Dij , ∀i, j ∈ V.
(19)
The constraints for maximum construction cost (budget)
and minimum reliability index can be transformed from
(4) and (3) in a similar way, respectively.
Last but not least, besides being applicable to new
power systems without existing WAMS, the proposed formulation can also generate optimal construction plans in
the presence of existing measurement devices, which is
common in some developed grids, e.g., North America
systems [44]. In such systems, the utilities may need to
gradually introduce new devices to the existing WAMS.
The proposed formulation can still develop optimal construction plans to achieve objectives such as minimizing
construction cost and maximizing post-construction reliability.
2
7
8
9
G
G
5
G
3
Generator
6
4
Bus
Transformer
1
Load
G
Figure 2: Line diagram of the IEEE 9-bus system and its substation
setting. Dashed boxes are substations with one or multiple buses.
Table 1: Parameter Settings
CiPMU [45]
$8819.46
DULR
Cij
[45]
$5146.87
CiPDC [45]
$7750.00
CkMS [5, 9]
$40000.0
pl [8, 46]
0.99
η [37]
0.0877
LF [36]
20160 bps
T PMU [45]
2
without any transformers attached form independent substations. See Fig. 2 as an illustration of the modified IEEE
9-bus system, where the dashed boxes are substations. As
a result, the IEEE 9-bus and 57-bus systems have six and
42 substations, respectively. In addition, we assume that
the control centers are located at bus 8 and 38 for the two
test systems, respectively, which are arbitrarily selected
and their locations do not have significant impacts on the
system performance. For simplicity, we assume that CI
shares the same topology with the power network. This
assumption only influences the value of the distance matrix Q, which is a constant. Other topologies, e.g., star
and ring networks, can be easily adopted by changing the
value of Q.
The problem is a multi-objective integer linear program
(MOILP) and it can be effectively solved by addressing
its equivalent bounded weighted-sum program [31]. Each
singleton ILP is solved using Gurobi [47], a high-efficiency
numerical optimization solver. All tests are performed on a
PC with an Intel Core-i7 CPU at 3.6GHz and 32GB RAM.
Problem formulation and simulation scripts are coded with
Python.
5.1. Pareto Optimal Solutions
We perform a series of simulation on both IEEE 9-bus
and 57-bus systems. The Pareto frontiers are presented in
Fig. 3, where IEEE 9-bus and 57-bus systems have seven
and 30 Pareto optimal solutions, respectively. In both subplots, the x-axis corresponds to RWAMS , y-axis to DWAMS ,
and z-axis to C WAMS . The objective function values of all
Pareto optimal solutions are plotted as the red dots in the
3-D space. To distinguish their relative positions in the
space, we also give their projections on x-y, y-z, and x-z
2-D planes, with dashed lines as auxiliary lines.
Although the Pareto frontier is not smooth due to the
combinatorial nature of the problem, it contains all the
non-dominating solutions. The combinatorial nature of
the proposed optimization problem makes these solutions
5. Case Studies and Discussions
To assess the efficacy of the proposed framework, we
employ the IEEE 9-bus and 57-bus systems for performance evaluation. The system parameters are set according to the studies reported by various utilities. Specifically,
all costs are developed from quotations of General Electric
Grid Solutions [45] and the U.S. Department of Energy
PMU installation cost report [9], and the data frame information is derived from [36]. All parameter values are
listed in Table 1.
The test systems are modified to simulate substation
interruptions. Buses connected with transformers are considered to be located in the same substation. Other buses
7
1e6
1e5
2.5
1.0
2.0
0.8
1.5
0.6
1.0
0.4
0.5
1e 4
0
1
2
3
0.2
0.0
0.0 e5
0.5 1
1.0
1.5
2.0
2.5
3.0
3.5
4.0
1e0.03
(a) IEEE 9-bus system
0.2
0.4
0.6
0.8
1.0
4
3
2
1
0.0
0 1e6
(b) IEEE 57-bus system
Figure 3: Pareto optimal performance of (a) IEEE 9-bus system and (b) IEEE 57-bus system.
Table 2: Pareto Optimal Solutions for IEEE 9-bus System
Solution
1
2
3
4
5
6
7
Objective Values
C WAMS
U WAMS
D WAMS
1.70E+05
2.88E-04
3.44E+05
1.77E+05
9.61E-05
3.44E+05
1.85E+05
2.88E-04
3.54E+04
2.18E+05
0.00E+00
4.10E+05
2.35E+05
3.84E-04
2.48E+04
2.42E+05
0.00E+00
4.60E+04
2.84E+05
1.92E-04
2.48E+04
PMUs
DULRs
PDCs
1(1)→9, 2(2)→9, 3(3)→9
1(1)→9, 2(2)→9, 3(3)→9, 7(5,7,8)→9, 9(6,8,9)→9
1(1)→1, 2(2)→2, 3(3)→3
1(1)→2, 2(2)→2, 3(3)→2, 9(6,8,9)→2
1(1)→1, 2(2)→2, 3(3)→3, 6(4,6,9)→6, 7(5,7,8)→2
1(1)→4, 2(2)→7, 3(3)→9, 9(6,8,9)→9
1(1)→4, 2(2)→7, 3(3)→9, 6(4,6,9)→6
4(6)→9, 7(5)→9, 9(8)→9
4(6)→9
4(6)→1, 7(5)→2, 9(8)→3
4(6)→2, 5(7)→2, 7(8)→2
N/A
4(6)→4, 5(4)→5, 7(8)→7
7(5)→7, 8(9)→8
9
9
1,2,3
2
1,2,3,6
4,5,7,9
4,6,7,8,9
57-bus system. The Pareto optimal solutions are presented
in Fig. 4. From the plots it is obvious that all three objectives contradict with each other and this contradiction
exists in the hypothesis given in Section 4.4. Although
the objective functions are different, we can still find some
relationships between Figures 3b and 4. All data points
presented in Fig. 4 have a corresponding point in Fig.
3b, while points in the latter figure may become multiple points in the subplots of Fig. 4. As points in Fig. 4
are non-dominating solutions with respect to two objective
functions, they will retain their non-dominating property
with the extra objective. Meanwhile, non-dominating solutions in 3b can possibly be non-dominating on all three
pairs of objective functions, resulting in multiple points in
the 2-D plots.
Last but not least, we also test the scalability of the
proposed optimization problem. To do this, several largescale test power systems, namely, IEEE 118-bus, 145-bus,
and 300-bus systems, are studied. We are specifically interested in the tractability of such power systems and the
computational time required to produce optimal solutions.
The simulation results are presented in Table 3. We list
the computational time required to calculate the first optimal solution and all solutions. It can be observed that
the computational time drastically increases with system
size. For all test systems except the 300-bus case, the
optimization can be finished within one day. However,
we are unable to develop all Pareto optimal solutions for
not well-positioned. Changing one value in the control
variables can potentially lead to drastic changes in the
objective values, resulting in a non-smooth solution space
with numerous local optimums.
To better investigate the characteristics of the Pareto
optimal solution, the detailed WAMS construction plans
for the IEEE 9-bus system are given in Table 2 with their
respective objective function values. In the “PMUs” column, all PMUs installed are presented in the form of i(a, b,
· · · ) → j, which means that a PMU is installed on bus i
and transmits synchrophasors to the PDC located at bus
j. This PMU makes buses a, b, · · · observable by measuring the current phasors of their respective connecting
branches. Note that the PMU also measures the voltage
phasor of its associated bus. “DULRs” column also has
a similar i(a) → j pattern, where a DULR is installed on
branch (i, a) on the i end, and sends data to j. The table shows insight on the relationships between objective
function values and their corresponding WAMS plans. Solutions 1 and 2 share an identical PMU and DULR allocation scheme, and their U WAMS are the same. This is
also the case for solutions 6 and 7. Meanwhile, the total
number of PDCs installed in the system has a significant
impact on DWAMS performance, while the measurement
device allocation is relatively less influential.
We also perform another set of simulation, where we
compare two objectives each time to demonstrate their
trade-off relationships. We illustrate this test on the IEEE
8
(a) Optimize C WAMS and U WAMS
3.5 1e6
3.0
2.5
2.0
1.5
1.0
0.5
0.0
0.70 0.75 0.80 0.85 0.90 0.95 1.00
1e6
C WAMS
(b) Optimize C WAMS and DWAMS
D WAMS
D WAMS
U WAMS
1.2 1e 3
1.0
0.8
0.6
0.4
0.2
0.0
0.70 0.75 0.80 0.85 0.90 0.95 1.00 1.05
1e6
C WAMS
4.4 1e5
4.2
4.0
3.8
3.6
3.4
3.2
3.0
2.8
0.0
0.2
0.4
0.6
U WAMS
0.8
1.0
1e 3
(c) Optimize U WAMS and DWAMS
Figure 4: Pareto optimal performance when jointly optimizing (a) C WAMS and U WAMS , (b) C WAMS and DWAMS , and (c) U WAMS and
DWAMS .
the problem, leading to different optimal solutions. On the
other hand, the latter solutions maintain their construction plans but the reliability is reduced. This accords with
the nature of system reliability. The new constraint rules
out the possibility of observing Bus 6 by installing measurement devices on the bus. Therefore, fewer devices can
observe the bus and its robustness against contingencies is
undermined.
When there is a pre-installed PMU at Bus 7, all solutions in Table 4 which plan to install devices on the bus is
preserved. The remaining two, i.e., Solutions 7 and 9, are
removed since they are dominated due to the lower cost induced by other solutions. To conclude, prohibiting device
installation on buses generally decreases the system reliability and having pre-installed devices significantly reduce
the system construction cost.
Table 3: Computation Time Comparison on Different Test Systems
System
IEEE 9-bus
IEEE 57-bus
IEEE 118-bus
IEEE 145-bus
IEEE 300-bus
First optimal solution
0.04 s
11.25 s
192.85 s
473.60 s
4699.28 s
Total time
12.83 s
1348.01 s
22013.62 s
59248.27 s
-
IEEE 300-bus within reasonable time on the PC set-up.
For very large systems, we may first divide the systems
into smaller sub-systems, and then generate WAMS construction plans for them. Since a smaller problem can be
effectively solved, the solutions to all sub-systems can then
be combined to construct the sub-optimal solutions for the
whole system. In addition, as power systems are typically
sparse, the optimality gap of this approach should be insignificant. How to integrate the proposed multi-objective
optimization with this approach will be investigated in future research.
5.3. Comparisons with State-of-the-Art
To further evaluate the proposed framework, we compare our proposed model with some state-of-the-art existing work based on the IEEE 57-bus system. As no other
work consider U WAMS and DWAMS , we focus on C WAMS
for assessment. We compare the optimal solutions developed by the other work with the Pareto optimal solution
with the least WAMS construction cost, presented in Table 6. In this table, the cost performance of our proposed
WPP is presented with the detailed WAMS construction
plan, which is in the form as in Table 2 without data transmission destinations. In addition, we also study the maximum number of PMU channels needed, as it is a major
concern in real-world PMU installation process. Last but
not least, we consider the impact introduced by unknown
transformer tap ratios to the system observability, which is
listed under the “Unknown Tap” column. In this comparison, Immunity Genetic Algorithm (IGA) [48], Integer Programming (IP) approach [18], and Cellular Learning Automata [49] are employed. We employ the same substation
interruption cost and PMU installation cost as in Table 1,
and set the cost for PMUs with more CTs to $12530.00
[45]. For observability, the PMUs installed in these plans
are assumed to observe all connecting branches.
5.2. Impact of Optional Operational Constraints
In the previous test, we investigated the Pareto performance of generalized WAMS construction problem. As
introduced in Sections 4.3.6 and 4.3.7, it is also possible
that some substations are reserved from installing measurement devices or already have them installed. Here we
investigate the impact of these practical constraints on the
Pareto optimal WAMS construction plans.
For simplicity, we focus on IEEE 9-bus system to demonstrate construction results. We investigate two test cases.
In the first case, we assume that Bus 6 is prohibited from
being installed with measurement devices. In the second
one, we additionally assume that Bus 7 is equipped with a
PMU. All other configurations remain the same as stated
in Section 5.1. The Pareto optimal solutions for these test
cases are listed in Tables 4 and 5.
Comparing with Table 2, the Pareto optimal solutions
in the original test case can be grouped into two classes:
solutions with or without constructions at Bus 6. The former solutions, i.e., Solutions 5 and 7 in Table 2 are replaced
by Solutions 4, 6, 7, and 9 in Table 4. This is because introducing a new constraint changes the solution space of
9
Table 4: Pareto Optimal Solutions for IEEE 9-bus System with Bus 6 Prohibited from Construction
Objective Values
C
U WAMS
D WAMS
1.70E+05
2.88E-04
3.44E+05
1.77E+05
9.61E-05
3.44E+05
1.85E+05
2.88E-04
3.54E+04
1.92E+05
9.61E-05
3.54E+04
2.18E+05
0.00E+00
4.10E+05
2.33E+05
2.88E-04
3.01E+04
2.35E+05
1.03E-02
2.48E+04
2.42E+05
0.00E+00
4.60E+04
2.84E+05
1.01E-02
2.48E+04
Solution
PMUs
DULRs
PDCs
1(1)→7, 2(2)→7, 3(3)→7
1(1)→7, 2(2)→7, 3(3)→7, 7(5,7,8)→7, 9(6,8,9)→7
1(1)→1, 2(2)→2, 3(3)→3
1(1)→4, 2(2)→7, 3(3)→9, 7(5,7,8)→7, 9(6,8,9)→9
1(1)→2, 2(2)→2, 3(3)→2, 9(6,8,9)→2
1(1)→1, 2(2)→2, 3(3)→3
1(1)→4, 2(2)→7, 3(3)→9, 5(4,5,7)→5, 9(6,8,9)→9
1(1)→4, 2(2)→7, 3(3)→9, 9(6,8,9)→9
1(1)→4, 2(2)→7, 3(3)→9, 5(4,5,7)→5
4(6)→7, 7(5)→7, 9(8)→7
4(6)→7
4(6)→1, 7(5)→2, 9(8)→3
4(6)→4
4(6)→2, 5(4)→2, 7(8)→2
4(6)→1, 7(5)→2, 8(9)→8
N/A
4(6)→4, 5(4)→5, 7(8)→7
8(7)→8, 9(6)→9
7
7
1,2,3
4,7,9
2
1,2,3,8
4,5,7,9
4,5,7,9
4,5,7,8,9
WAMS
1
2
3
4
5
6
7
8
9
Table 5: Pareto Optimal Solutions for IEEE 9-bus System with Bus 6 Prohibited from Construction and Bus 7 Pre-installed with a PMU
Solution
1
2
3
4
5
6
7
Objective Values
C WAMS
U WAMS
D WAMS
1.25E+05
1.92E-04
3.44E+05
1.28E+05
9.61E-05
3.44E+05
1.40E+05
1.92E-04
3.54E+04
1.44E+05
9.61E-05
3.54E+04
1.73E+05
0.00E+00
4.10E+05
1.88E+05
2.88E-04
3.01E+04
1.97E+05
0.00E+00
4.60E+04
PMUs
DULRs
PDCs
1(1)→2, 2(2)→2, 3(3)→2, 7(5,7,8)→2
1(1)→2, 2(2)→2, 3(3)→2, 7(5,7,8)→2, 9(6,8,9)→2
1(1)→4, 2(2)→7, 3(3)→9, 7(5,7,8)→7
1(1)→4, 2(2)→7, 3(3)→9, 7(5,7,8)→7, 9(6,8,9)→9
1(1)→2, 2(2)→2, 3(3)→2, 7(5,7,8)→2, 9(6,8,9)→2
1(1)→1, 2(2)→2, 3(3)→3, 7(5,7,8)→2
1(1)→4, 2(2)→7, 3(3)→9, 7(5,7,8)→7, 9(6,8,9)→9
4(6)→2, 9(8)→2
4(6)→2
4(6)→4, 9(6)→9
4(6)→4
4(6)→2, 5(4)→2
4(6)→1, 8(9)→8
4(6)→4, 5(7)→5
2
2
4,7,9
4,7,9
2
1,2,3,8
4,5,7,9
From Table 6 we can see that the construction cost of
the proposed framework does not perform as good as the
others. However, none of the compared algorithms can
actually achieve full numerical observability on the test
system due to unknown transformer tap ratios. This incomplete observability may contribute to their low WAMS
construction cost. For these plans to achieve full observability in absence of tap ratios, extra measurement devices
are required, which will incur extra cost [18]. Another
possible reason of their low cost is that these solutions
introduces zero-injection buses to improve the observability. As an integer linear programming formulation of zeroinjection buses was proposed in [5], it can be employed in
our framework with minimal effort to further improve the
performance. Considering these analyses, our proposed
WPP can generate a feasible WAMS plan providing full
numerical observability with reasonable construction cost.
to develop practical optimal WAMS construction plans.
Therefore, we propose a multi-objective WPP for developing multiple Pareto optimal solutions to suit different
purposes of the utilities. We verify the proposal model on
the IEEE 9-bus and 57-bus systems. The simulation results indicate that multiple Pareto optimal solutions can
be developed. In addition, when maintaining the system
numerical observability, our proposed framework can result in optimal WAMS construction plans with minimal
cost. This work develops a comprehensive framework for
most practical WAMS construction designs.
6. Conclusion
References
Acknowledgement
This work was supported by the Theme-based Research
Scheme of the Research Grants Council of Hong Kong,
under Grant No. T23-701/14-N.
In this paper we propose a unified framework for constructing the future WAMS based on several realistic considerations. To be specific, we formulate a practical cost
model for WAMS construction, considering the fact that
the measurement devices should be installed at substations, whose interruption cost should not be ignored during installation. In addition, no real-time transformer tap
ratio information is required and this fits into the practical
situations better in the power system. Moreover, we allocate PMUs to different buses, and at the same time assign
branches to the PMUs for measurement.
In the proposed model, we consider three representative WAMS construction objectives, namely, the construction cost, system reliability, and synchrophasor data traffic. They unify the most common objectives considered
in most related work. While different WAMS may stress
on distinctive objectives, they should be jointly optimized
[1] J. D. L. Ree, V. Centeno, J. S. Thorp, A. G. Phadke, Synchronized Phasor Measurement Applications in Power Systems,
IEEE Trans. Smart Grid 1 (1) (2010) 20–27.
[2] A. G. Phadke, J. S. Thorp, K. J. Karimi, State Estimation with
Phasor Measurements, IEEE Trans. Power Syst. 1 (1) (1986)
233–238.
[3] I. Kamwa, R. Grondin, Y. Hebert, Wide-area measurement
based stabilizing control of large power systems-a decentralized/hierarchical approach, IEEE Trans. Power Syst. 16 (1)
(2001) 136–153.
[4] F. H. Fesharaki, R. A. Hooshmand, A. Khodabakhshian, Simultaneous Optimal Design of Measurement and Communication
Infrastructures in Hierarchical Structured WAMS, IEEE Trans.
Smart Grid 5 (1) (2014) 312–319.
[5] A. Pal, A. K. S. Vullikanti, S. S. Ravi, A PMU Placement
Scheme Considering Realistic Costs and Modern Trends in Relaying, IEEE Trans. Power Syst. 32 (1) (2017) 552–561.
[6] L. L. Grigsby (Ed.), Power System Stability and Control, Third
Edition, 3rd Edition, CRC Press, Boca Raton, 2012.
10
Table 6: Cost Comparison of the Proposed Framework with State of the Art For IEEE 57-bus System
Method
Cost
Proposed
7.24E+05a
IGA [48]b
5.78E+05c
5.78E+05c
9.98E+05
5.78E+05c
b,d
IP [18]
CLA [49]b
Measurement Devices
PMUs: 4(5,6), 11(9,13), 12(16,17), 15(1,14), 29(28,52), 32(31,33), 37(38,39),
47(46,48), 50(49,51), 54(53,55), 56(42,57)
DULRs:3(2), 7(8), 12(10), 18(19), 20(19), 21(22), 24(23), 25(30), 26(27), 34(35),
40(36), 43(41), 45(44)
PMUs: 1, 6, 13, 19, 25, 29, 32, 38, 51, 54, 56
PMUs: 1, 4, 13, 20, 25, 29, 32, 38, 51, 54, 56
PMUs: 1, 2, 6, 12, 14, 19, 21, 27, 29, 30, 32, 33, 41, 44, 49, 51, 53, 55, 56
PMUs: 1, 4, 13, 20, 25, 29, 32, 38, 51, 54, 56
Constraints
max {T PMU }
Unknown Tap
2
Observable
6
6
5
6
Unobservable
Unobservable
Unobservable
Unobservable
a
For a fair comparison, the PDC cost is removed to keep accordance with other methods.
Installed PMUs are assumed to be able to observe all connecting branches.
This minimal cost cannot guarantee system numerical observability without known transformer tap ratio.
d
Multiple optimal solutions are given subject to different constraints. Two most cost-efficient ones are listed for comparison.
b
c
[7] R. Emami, A. Abur, Robust Measurement Design by Placing Synchronized Phasor Measurements on Network Branches,
IEEE Trans. Power Syst. 25 (1) (2010) 38–43.
[8] O. Gomez, C. Portilla, M. A. Rı́os, Reliability Analysis of Substation Monitoring Systems Based on Branch PMUs, IEEE
Trans. Power Syst. 30 (2) (2015) 962–969.
[9] Factors Affecting PMU Installation Costs, Tech. rep., U.S. Department of Energy (Oct. 2014).
[10] G. N. Korres, P. J. Katsikas, G. C. Contaxis, Transformer tap
setting observability in state estimation, IEEE Trans. Power
Syst. 19 (2) (2004) 699–706.
[11] Z. H. Rather, Z. Chen, P. Thøgersen, P. Lund, B. Kirby, Realistic Approach for Phasor Measurement Unit Placement: Consideration of Practical Hidden Costs, IEEE Trans. Power Deliv.
30 (1) (2015) 3–15.
[12] B. Gou, Generalized Integer Linear Programming Formulation
for Optimal PMU Placement, IEEE Trans. Power Syst. 23 (3)
(2008) 1099–1104.
[13] B. Milosevic, M. Begovic, Nondominated sorting genetic algorithm for optimal phasor measurement placement, IEEE Trans.
Power Syst. 18 (1) (2003) 69–75.
[14] S. Chakrabarti, E. Kyriakides, Optimal Placement of Phasor Measurement Units for Power System Observability, IEEE
Trans. Power Syst. 23 (3) (2008) 1433–1440.
[15] N. M. Manousakis, G. N. Korres, A Weighted Least Squares
Algorithm for Optimal PMU Placement, IEEE Trans. Power
Syst. 28 (3) (2013) 3499–3500.
[16] N. M. Manousakis, G. N. Korres, P. S. Georgilakis, Taxonomy
of PMU Placement Methodologies, IEEE Trans. Power Syst.
27 (2) (2012) 1070–1077.
[17] F. Aminifar, M. Fotuhi-Firuzabad, A. Safdarian, A. Davoudi,
M. Shahidehpour, Synchrophasor Measurement Technology in
Power Systems: Panorama and State-of-the-Art, IEEE Access
2 (2014) 1607–1628.
[18] F. Aminifar, A. Khodaei, M. Fotuhi-Firuzabad, M. Shahidehpour, Contingency-Constrained PMU Placement in Power Networks, IEEE Trans. Power Syst. 25 (1) (2010) 516–523.
[19] M. B. Mohammadi, R. A. Hooshmand, F. H. Fesharaki, A New
Approach for Optimal Placement of PMUs and Their Required
Communication Infrastructure in Order to Minimize the Cost
of the WAMS, IEEE Trans. Smart Grid 7 (1) (2016) 84–93.
[20] A. Chakrabortty, P. P. Khargonekar, Introduction to wide-area
control of power systems, in: 2013 American Control Conference, 2013, pp. 6758–6770.
[21] J. E. Anderson, A. Chakrabortty, PMU placement for dynamic
equivalencing of power systems under flow observability constraints, Electric Power Systems Research 106 (2014) 51–61.
[22] M. Zima, M. Larsson, P. Korba, C. Rehtanz, G. Andersson,
Design Aspects for Wide-Area Monitoring and Control Systems,
Proc. IEEE 93 (5) (2005) 980–996.
[23] S. G. Ghiocel, J. H. Chow, G. Stefopoulos, B. Fardanesh,
D. Maragal, B. Blanchard, M. Razanousky, D. B. Bertagnolli,
Phasor-Measurement-Based State Estimation for Synchrophasor Data Quality Improvement and Power Transfer Interface
Monitoring, IEEE Trans. Power Syst. 29 (2) (2014) 881–888.
[24] A. Chakrabortty, C. F. Martin, Optimal Measurement Allocation Algorithms for Parametric Model Identification of Power
Systems, IEEE Trans. Control Syst. Technol. 22 (5) (2014)
1801–1812.
[25] J. Zhang, S. Nabavi, A. Chakrabortty, Y. Xin, ADMM Optimization Strategies for Wide-Area Oscillation Monitoring in
Power Systems Under Asynchronous Communication Delays,
IEEE Trans. Smart Grid 7 (4) (2016) 2123–2133.
[26] G. C. Zweigle, V. Venkatasubramanian, Wide-Area Optimal
Control of Electric Power Systems With Application to Transient Stability for Higher Order Contingencies, IEEE Trans.
Power Syst. 28 (3) (2013) 2313–2320.
[27] H. Li, A. Bose, V. M. Venkatasubramanian, Wide-Area Voltage
Monitoring and Optimization, IEEE Trans. Smart Grid 7 (2)
(2016) 785–793.
[28] G. Liang, J. Zhao, F. Luo, S. Weller, Z. Y. Dong, A Review of
False Data Injection Attacks Against Modern Power Systems,
IEEE Trans. Smart Grid 8 (4) (2017) 1630–1638.
[29] F. Pasqualetti, F. Dörfler, F. Bullo, Attack Detection and Identification in Cyber-Physical Systems, IEEE Trans. Autom. Control 58 (11) (2013) 2715–2729.
[30] M. S. Almas, L. Vanfretti, R. S. Singh, G. M. Jonsdottir, Vulnerability of Synchrophasor-based WAMPAC Applications to Time
Synchronization Spoofing, IEEE Trans. Smart Grid PP (99)
(2017) 1–1.
[31] M. J. Alves, J. Clı́maco, A review of interactive methods for
multiobjective integer and mixed-integer programming, European Journal of Operational Research 180 (1) (2007) 99–115.
[32] R. Deng, G. Xiao, R. Lu, H. Liang, A. V. Vasilakos, False Data
Injection on State Estimation in Power Systems - Attacks, Impacts, and Defense: A Survey, IEEE Trans. Ind. Inform. 13 (2)
(2017) 411–423.
[33] H. Gjermundrod, D. E. Bakken, C. H. Hauser, A. Bose, GridStat: A Flexible QoS-Managed Data Dissemination Framework
for the Power Grid, IEEE Trans. Power Deliv. 24 (1) (2009)
136–143.
[34] Y. Xin, A. Chakrabortty, A study on group communication in
distributed wide-area measurement system networks in large
power systems, in: 2013 IEEE Global Conference on Signal and
Information Processing, 2013, pp. 543–546.
[35] IEEE Guide for Phasor Data Concentrator Requirements for
Power System Protection, Control, and Monitoring, IEEE Std
C37.244-2013 (May 2013).
[36] IEEE Standard for Synchrophasor Data Transfer for Power Systems, IEEE Std C37.118.2-2011 (Dec. 2011).
[37] F. Zhang, L. Cheng, X. Li, Y. Sun, W. Gao, W. Zhao, Application of a Real-Time Data Compression and Adapted Protocol
Technique for WAMS, IEEE Trans. Power Syst. 30 (2) (2015)
653–662.
[38] A. Torres, N. Peña, F. Pérez, Communication network for a
wide-area special protection system, in: 2013 IEEE Grenoble
PowerTech (POWERTECH), 2013, pp. 1–5.
[39] A. Pal, G. A. Sanchez-Ayala, V. A. Centeno, J. S. Thorp, A
PMU Placement Scheme Ensuring Real-Time Monitoring of
Critical Buses of the Network, IEEE Trans. Power Deliv. 29 (2)
11
(2014) 510–517.
[40] M. Göl, A. Abur, A Hybrid State Estimator For Systems With
Limited Number of PMUs, IEEE Trans. Power Syst. 30 (3)
(2015) 1511–1517.
[41] N. Kashyap, S. Werner, Y. F. Huang, T. Riihonen, Power System State Estimation Under Incomplete PMU Observability –
A Reduced-Order Approach, IEEE J. Sel. Top. Signal Process.
8 (6) (2014) 1051–1062.
[42] X. Li, A. Scaglione, T. H. Chang, A Framework for Phasor
Measurement Placement in Hybrid State Estimation Via GaussNewton, IEEE Trans. Power Syst. 29 (2) (2014) 824–832.
[43] A. M. Glazunova, I. N. Kolosok, E. S. Korkina, PMU placement on the basis of SCADA measurements for fast load flow
calculation in electric power systems, in: 2009 IEEE Bucharest
PowerTech, 2009, pp. 1–6.
[44] PMUs and synchrophasor data flows in North America — North
American SynchroPhasor Initiative, https://www.naspi.org/
node/387.
[45] GE Grid Solutions, http://www.gegridsolutions.com/.
[46] Y. Wang, C. Wang, W. Li, J. Li, F. Lin, Reliability-Based Incremental PMU Placement, IEEE Trans. Power Syst. 29 (6) (2014)
2744–2752.
[47] Gurobi Optimization, http://www.gurobi.com/.
[48] F. Aminifar, C. Lucas, A. Khodaei, M. Fotuhi-Firuzabad, Optimal Placement of Phasor Measurement Units Using Immunity
Genetic Algorithm, IEEE Trans. Power Deliv. 24 (3) (2009)
1014–1020.
[49] S. M. Mazhari, H. Monsef, H. Lesani, A. Fereidunian, A
Multi-Objective PMU Placement Method Considering Measurement Redundancy and Observability Value Under Contingencies, IEEE Trans. Power Syst. 28 (3) (2013) 2136–2146.
12
| 5cs.CE
|
An information scaling law: ζ = 3/4
Michael C. Abbott & Benjamin B. Machta•
arXiv:1710.09351v1 [cond-mat.stat-mech] 25 Oct 2017
Institute of Physics, Jagiellonian University,
Ulica Łojasiewicza 11, 30-348 Kraków, Poland.
• Lewis-Sigler Institute and Department of Physics,
Princeton University, Princeton, NJ 08544, USA.
[email protected], [email protected]
25 October 2017
Consider the entropy of a unit Gaussian convolved over a discrete set of K points,
constrained to an interval of length L. Maximising this entropy fixes K, and we show
that this number exhibits a novel scaling law K ∼ L1/ζ as L → ∞, with exponent
ζ = 3/4. This law was observed numerically in a recent paper about optimal effective
theories; here we present an analytic derivation. We argue that this law is generic
for channel capacity maximisation, or the equivalent minimax problem. We also
briefly discuss the behaviour at the boundary of the interval, and higher dimensional
versions.
1. Introduction
In a recent paper [1] we studied the construction of optimal effective theories, defined to be
those which maximise the mutual information between their parameters and the expected
data. We wrote this in the language of Bayesian prior choice, and the optimal prior turns out
to almost always be discrete, consisting of K delta functions. As the noise 1/L in measuring
a certain parameter is taken to zero, we observed that there is a scaling law for the number
of delta functions:
K ∼ L1/ζ when L → ∞,
ζ = 3/4.
(1)
The main purpose of the present paper is to derive this law analytically.
In more detail, the setup is as follows. We consider a model which, at each point θ in
the full parameter space Θ, gives a probability distribution p( x |θ ) for x in data space X. An
effective model is a restriction of θ to lie in some subspace, and this can be expressed as a
prior p(θ ) with support only in this subspace. For a given prior, the mutual information of
interest is
I ( X; Θ) = S( X ) − S( X |Θ)
(2)
where the entropy and relative entropy are
S( X ) = −
Z
Z
dx p( x ) log p( x ),
p( x ) = dθ p( x |θ ) p(θ )
Z
Z
S( X |Θ) = dθ p(θ ) − dx p( x |θ ) log p( x |θ ) .
This quantity I ( X; Θ) is the natural measure of how much information about the data can
be encoded in the parameters, or equivalently, how much information about the parameters
can be learned from the data [2].
Mutual information (bits)
3
2
1
Gaussian
Bernoulli
ζ=1
ζ = 0.75
0
2
4
8
16
32
Number of delta functions K (log scale)
Figure 1: Scaling law for the number of delta functions as L is increased, plotting
I ( X; Θ)/ log 2 against K. The lines drawn are I ( X; Θ) = ζ log K for ζ = 1 (an
absolute bound) and ζ = 3/4 (our scaling law). The blue cross data points are for
the Bernoulli model discussed in section 3 below.
Choosing a prior by maximising this information was considered by Bernardo [3]. The
problem is mathematically identical to maximising the capacity of a communications channel,
which was known to often result in a discrete distribution [4–7]. This can be avoided by
considering instead I ( X m ; Θ) from m independent sets of data, and taking the limit m → ∞,
in which case the optimal prior (now called a reference prior) approaches Jeffreys prior [3, 8]
which is a smooth function. Up to normalization, Jeffreys prior is the volume form arising
from the Fisher information metric [9]:
Z
q
1q
det gµν (θ ),
Z = dθ det gµν (θ )
pJ ( θ ) =
Z
where
Z
∂ log p( x |θ ) ∂ log p( x |θ )
gµν (θ ) = dx p( x |θ )
.
∂θ µ
∂θ ν
This metric measures the distinguishability of the data which would be produced by different
parameter values, in units of standard deviations. The statement that we have enough data
to fix all parameters with high precision is equivalent to the statement that the manifold Θ is
long in all dimensions, measured with this metric. And the limit m → ∞ amounts to having
infinitely much data.
The main point of our paper [1], and of earlier work such as [10], is that in most situations of physical interest we have many very short directions, corresponding to irrelevant
parameters. Thus the generic point θ is very near to a boundary of Θ, rather than being
safely in the interior, and thus Jeffreys prior is a poor choice. We showed that the prior p? (θ )
which maximises mutual information I ( X; Θ) typically puts weight only on the boundaries
of Θ, and thus describes an appropriate lower-dimensional effective model. The scaling law
ensures that, along a relevant parameter in this effective model, neighbouring delta functions
became indistinguishable, giving an effectively continuous distribution.
Here however our focus is much narrower. We primarily consider a model in just one
2
dimension, measuring θ ∈ [0, L] with Gaussian noise of known variance, thus
p( x |θ ) =
e−( x−θ )
√1
σ 2π
2 /2σ2
(3)
.
It is convenient to choose units in which σ = 1, so that θ measures proper distance: gθθ = 1.
Thus L is the length of parameter space, in terms of the Fisher metric. Jeffreys prior is a
constant pJ (θ ) = 1/L. The optimal prior is
K
p? (θ ) =
∑ λ a δ ( θ − θ a ),
θ1 = 0,
θK = L.
(4)
a =1
With K delta functions there is a bound I ( X; Θ) ≤ log K on the mutual information — the
difference between certainty and complete ignorance among K = 2n outcomes is exactly n
bits of information. When finding p? (θ ) numerically in [1], we observed that as L → ∞ the
mutual information follows a line with lower slope (see figure 1):
I ( X; Θ) ∼ ζ log K,
ζ ≈ 0.75,
K 1.
(5)
Since I ( X; Θ) ∼ log L for large L, this implies the scaling law (1). This can also be stated
in terms of the number density of delta functions: since K grows slightly faster than L, the
average proper density grows without bound:
ρ0 =
K
∼ L−1+1/ζ = L1/3 ,
L
L 1.
To derive this law, Section 2 treats a field theory for the local density of delta functions ρ(θ ).
Section 3 looks at at another one-dimensional model which displays the same scaling
law, and then at the generalisation to D dimensions. And appendix A looks at paper from
almost 30 years ago which could have discovered this scaling law.
2. Derivation
For this Gaussian model (3), the relative entropy S( X |Θ) = 12 + 12 log 2π is independent of
the prior, so it remains only to calculate the entropy S( X ) in order to maximise the mutual
information (2).
On an infinite line, the entropy would be maximised by a constant p( x ), i.e. a prior
(4) with delta functions spaced infinitesimally close together. But on a very short line, we
observe that entropy is maximised by placing substantial weight at each end, with a gap
before the next delta function. The idea of our calculation is that the behaviour on a long
but finite line should interpolate between these two regimes. We work out first the cost of a
finite density of delta functions, and then the local cost of a spatially varying density, giving
us an equation of motion for the optimum ρ( x ). By solving this we learn how the density
increases as we move away from the boundary. The integral of this density then gives us K
with the desired scaling law.
Since the deviations from a constant p( x ) will be small, we write
1
p( x ) = [1 + w( x )],
L
3
Z
dx w( x ) = 0
and then expand the entropy
L
1
dx w( x )2 + O(w4 )
2L 0
1
= log L − ∑ |wk |2 + . . . .
2 k
Z
S( X ) = log L −
(6)
Here our convention for Fourier transforms is that
wk =
Z L
dx −ikx
e
w ( x ),
0
k∈
L
2π
Z.
L
Constant spacing
Consider first the effect of a prior which is a long string of delta functions at constant spacing
a, which we assume to be small compared to the standard deviation σ = 1, which in turn is
much less than the length L.1 This leads to
p( x ) =
a
L
∑
n ∈Z
2
1
√ e−(x−na) /2 .
2π
(7)
Because this is a convolution of a Dirac comb with a Gaussian kernel, its Fourier transform
is simply a product of such pieces. Let us write the transformation of the positions of the
sources as follows:
c0 ( x ) = a
∑
δ( x − na),
c0k =
⇒
n ∈Z
a
L
∑
n ∈Z
e−ikna =
∑
m ∈Z
δk−m 2π .
a
The zero-frequency part of pk is the constant term in p( x ), with the rest contributing to w( x ):
( 2
e−k /2 k ∈ 2π
2
a Z and k 6 = 0
wk = ∑ δk−m 2π e−k /2 =
a
0
else.
m 6 =0
The lowest-frequency terms at k = ±2π/a give the leading exponential correction to the
entropy:
2
2
2π
S( X ) = log L − e−q + O(e−2q ),
q=
.
(8)
a
As advertised, any nonzero spacing a > 0 (i.e. frequency q < ∞) reduces the entropy from
its maximum.
Variable spacing
Now consider perturbing the positions of the delta functions by a slowly varying function
∆( x ), and multiplying their weights by 1 + h( x ). We seek a formula for the entropy in terms
of ∆( x ), while allowing h( x ) to adjust so as to minimise the disturbance. This cannot be done
perfectly, as h( x ) is only sampled at spacing a, so only contributions at frequencies lower
than q = 2π/a will be screened. Thus we expect what survives to appear with the same
exponential factor as (8). In particular this ensures that at infinite density, no trace of ∆( x )
remains. And that is necessary in order for the limit to agree with Jeffreys prior, which is a
constant.
Figure 2 illustrates how the positions and weights of p? (θ ) compensate to leave p( x )
almost constant in the interior, in a numerical example. Below that it shows how the functions
∆( x ) and h( x ) used here mimic this effect.
1
We summarise all the scales involved in (10), see also figure 2.
4
p(θ), p(x)
0
10
20
30
b
∆(x) = ∆ cos(kx)
b ∆ sin(kx)
b
h(x) = -k
Gaussian, width = 1
Perturbed comb = c(x)
Unperturbed x = na
p(x), frequency > q=2π/a
Figure 2: Above, numerical solution p? (θ ) for L = 30, in which we observe that as
the spacing of the delta functions grows closer together, their weights compensate
to leave p( x ) almost constant, with deviations w( x ) at a wavelength comparable to
the spacing.
Below, a diagram to show the scales involved when perturbing the positions of
the delta functions in our derivation. These are arranged from longest to shortest
wavelength, see also (10).
The comb of delta functions c0 ( x ) we had above is perturbed to
c( x ) = [1 + h( x )] a ∑ δ x − na − ∆(na) = [1 + h( x )] c∆ ( x ).
n ∈Z
The effect of h( x ) is a convolution in frequency space:
∆
ck = c∆
k + ∑ hk0 ck−k0 .
k0
It will suffice to study ∆( x ) = ∆ cos(b
k x ), i.e. frequencies ±b
k only: ∆k = 21 ∆(δk−bk + δk+bk ).
The driving frequency is b
k q. We can expand in the amplitude ∆ to write
c∆
k
a
=
L
∑e
−ik na+∆(na)
n
= c0k −
i k 2 ∆2 h
i
ik∆ h 0
c b + c0 b −
2c0k + c0 b + c0 b + O(∆3 ).
k− k
k+ k
k −2 k
k +2 k
2
8
The order ∆ term has contributions at k = b
k q = 2π/a, which can be screened in the full
ck by setting hk = +ik∆k i.e. h( x ) = −b
k∆ sin(b
k x ). What survives in ck then are contributions
2
b
at k = 0, k = ±q and k = ±q ± k:
i q ∆
q2 ∆2
ck = δk + ∑ δk±q 1 −
+ ∑[δk∓q−bk + δk∓q+bk ] ±
+ O(δk±2q , ∆2 )
4
2
±
±
All but the zero-frequency term are part of wk = (ck − δk )e−k
2
2 /2
The contributions at and k = ±q ± 2b
k will only matter at order ∆4 in S( X ).
5
, and enter (6) independently,
giving this:
i ∆q 2
2
q2 ∆2 2 h −(q−bk )2
b 2
− e
+ e−(q+ k )
+ O(∆4 ) + O(e−2q )
4
2
h
i
1 6 b4
2 8 b6
2
4 b2
− q2
2
1 + ∆ q k + q k + q k + . . . 1 + O(1/q ) + . . . + . . . (9)
= log L − e
3
45
S( X ) = log L − e−q
2
1−
As promised, the order ∆2 term comes with the same overall exponential as in (8) above.
Restoring units briefly, the expansion in round brackets makes sense only if b
k q σ2 1.3 In
terms of length scales this means
a = 2π/q
comb spacing
2π/b
k
σ
σa , or writing all the assumptions made:
σ=1
kernel width
≪
2π/b
k
perturbation wavelength
L
(10)
.
box size
Entropy density
We can think of this entropy (9) as arising from some density: S( X ) =
claim is that this density takes the form
2
2
128π 6
S( x ) = − L e−(2π ) ρ(x) 1 +
ρ ( x )4 ρ 0 ( x )2 .
3
R
dx
L S( x )
= S0 . Our
(11)
The constant term is clearly fixed by (8). To connect the kinetic term to (9), we need
ρ( x ) =
i
1
1h
0
0
2
3
1
−
∆
(
x
)
+
∆
(
x
)
+
O(
∆
)
=
a(1 + ∆0 ( x ))
a
thus ρ0 ( x ) = − 1a ∆00 ( x ) + . . . and
h
ih
i
2
2
2
e−(2π ) ρ( x) = e−q 1 + 2q2 ∆0 ( x ) + 2q4 ∆0 ( x )2 + O(∆3 ) 1 + O(1/q2 ) .
Multiplying these pieces, the order ∆1 term of S( x ) integrates to zero. We can write the
order ∆2 term in terms of Fourier coefficients (using (6), and ∆0k = ik∆k ), and we recover the
leading terms in (9). The next term there q8 b
k 6 would arise from a term ρ( x )6 ρ00 ( x )2 in the
4
density, which we neglect.
The Euler-Lagrange equations from (11) read
0 = ρ( x )4 ρ00 ( x ) + 2ρ( x )3 ρ0 ( x )2 − 4π 2 ρ( x )5 ρ0 ( x )2 +
3
ρ ( x ).
32π 4
We are interested in the large-x behaviour of a solution with boundary condition at x = 0 of
ρ = 1. Or any constant density, but this value is independent of L because the only interaction
is of scale σ = 1. This is also what we observe numerically, shown in figure 3. Making the
ansatz ρ( x ) = 1 + x η with η > 1, these four terms scale as
x5η −2 ,
x5η −2 ,
x7η −2 ,
xη ,
2η
all × e− x , x → ∞.
(12)
Clearly the first two terms are subleading to the third, and thus the last two terms must
cancel each other. We have 7η − 2 = η and thus η = 1/3. Then the total number of delta
functions in length L is
K=
3
4
Z L
0
dx ρ( x ) ∼ L4/3
(13)
In footnote 4 we confirm that this indeed holds.
The term q8 b
k 6 in S( X ) (9) corresponds to a term ρ6 (ρ00 )2 in S( x ) (11). This gives a term in the equations of motion
(12) going like x9η −4 , which goes to zero as x → ∞ with η = 1/3. Thus we are justified in dropping this.
6
Proper length L
20
20
15
15
10
10
5
5
Gaussian
0
2 e
4
6
8
10
Bernoulli
2 π 4
0
6
Proper distance θ
8
10
φ
Figure 3: Positions of delta functions in optimal priors p? (θ ) = ∑K
a =1 λ a δ ( θ − θ a ),
for various values of L. Each horizontal line represents one prior. We observe that
the second (and third...) delta functions occur at fixed proper distance from the first,
justifying the fixed boundary condition on ρ( x ).
On the right, we show similar data for √
the Bernoulli model of section 3 below, in
terms of proper distance φ. Here L = π m for m = 1, 2, 3, ...50.
establishing the result (1).
3. Extensions
The other one-dimensional example studied in [1] was Bernoulli problem, of determining
the weighting of an unfair coin given the number of heads seen after m flips:
m!
θ x (1 − θ ) m − x ,
x!(m − x )!
p( x |θ ) =
θ ∈ [0, 1],
x ∈ {0, 1, 2, 3, . . . , m}.
(14)
The Fisher metric here is
gθθ =
m
θ (1 − θ )
⇒
L=
Z 1q
0
√
gθθ dθ 2 = π m
and we define the proper parameter φ by
ds2 =
m dθ 2
= dφ2
θ (1 − θ )
⇐
θ = sin2
φ
√ ,
2 m
√
φ ∈ 0, π m .
The optimal prior found by maximising the mutual information is again discrete, and
when m → ∞ it also obeys the scaling law (1) with the same slope ζ. Numerical data showing
this is also plotted in figure 1 above. This scaling relies on the behaviour far from the ends of
the interval, where this binomial distribution can be approximated by a Gaussian:
p( x |θ ) ≈
2
2
1
√ e−(x−mθ ) /2σ ,
σ 2π
m → ∞, θ finite,
σ 2 = m θ (1 − θ ).
(15)
The agreement of these very different models suggests that the ζ = 3/4 power is in some
sense universal for nonsingular one-dimensional models.
7
Near to the ends of the interval, we observe in figure 3 that first few delta functions
again settle down to fixed proper distances. In this regime (15) is not a good approximation,
and instead the binomial (14) approaches a Poisson distribution:
p( x |θ ) ≈
e−µ µ− x
,
x!
m → ∞,
µ = mθ ≈
φ2
finite.
2
(16)
In fact the second delta function appears to be within half a percent of φ2 = π, one of several
curious numerical co-incidences. The first few positions and weights are as follows:5
φ2 ≈ 3.13 ≈ π
λ2 /λ1 ≈ 0.63 ≈ 2/π = 0.6366
φ3 ≈ 5.41
λ3 /λ1 ≈ 0.54
φ4 ≈ 7.42
λ4 /λ1 ≈ 0.49 .
(17)
This implies that the second delta function is at mean µ = π 2 /4 ≈ 2.47, skipping the first
few integers x.
More dimensions
Returning to the bulk scaling law, one obvious thing to wonder is whether this extends
to more dimensions. The trivial example is to consider the same Gaussian model (7) in
D-dimensional cube:
2
p(~x |~θ ) ∝ e−(~x−~σ) /2 ,
~θ ∈ [0, L] D , ~x ∈ RD .
(18)
This simply factorises into the same problem in each direction: (2) is the sum of D identical
mutual information terms. Thus the optimal prior is simply
p? (~θ ) =
D
K
µ =1
a1 ...a D =1
∏ p? (θµ ) = ∑
λ a1 · · · λ a D δ ( θ − θ a1 ) · · · δ ( θ − θ a D )
with the same coefficients as in (4) above. The total number of delta functions is Ktot = K D
which scales as
Ktot ∼ L D/ζ = V 1/ζ ,
ζ = 3/4, V → ∞.
(19)
We believe that this scaling law is also generic, provided the large-volume limit is taken
such that all directions expand together. If the scaling arises from repeating an experiment
√
m times, then this will always be true as all directions grow as m.
To check this in a less trivial example, we consider now the bivariate binomial problem
studied by [11]. We have two unfair coins whose weights we wish to determine, but we flip
the second coin only when the first coin comes up heads. After m throws of the first coin,
the model is
m x
m− x x
p( x, y|θ, φ) =
θ (1 − θ )
φ y (1 − φ ) x − y .
(20)
x
y
with θ, φ ∈ [0, 1] and 0 ≤ y ≤ x ≤ m ∈ Z. The Fisher information metric here is
ds2 =
5
m
mθ
dθ 2 +
dφ2
θ (1 − θ )
φ (1 − φ )
For the Gaussian model (3), instead we see θ2 ≈ e, with one more significant figure numerically. The corresponding
table reads:
θ2 ≈ 2.718
λ2 /λ1 ≈ 0.672
θ3 ≈ 4.889
λ3 /λ1 ≈ 0.582 .
8
1
Mutual information (bits)
θφ
5.0
0
0
1
θ
4.0
3.5
Bivariate Binomial
2D Gaussian
ζ=1
ζ = 0.75
θ2
13
4.5
3.0
0
0
13
θ1
16
25
36
49
64
100
Total number of atoms K
Figure 4: Scaling law in two dimensions. On the right we plot I ( X; Θ)/ log 2 against
log Ktot for the bivariate binomial model (20) and the D = 2 Gaussian model (18).
On the left we show examples of the priors, for m = 20 and L1 = L2 = 13. For
the bivariate binomial the plot axes are (θ, θφ) so as to respect the topology of the
parameter space, but the figure is not isometric to Θ.
which implies
V=
Z 1
0
dθ
Z 1
0
dφ
q
det gµν = 2πm
1
and pJ (θ, φ) = 2π
[(1 − θ )φ(1 − φ)]−1/2 . Topologically the parameter space is a triangle,
√
since at θ = 0 the φ edge is of zero length. The other three sides are each of length π m,
and so will all grow in proportion as m → ∞.
We can find the optimal priors for this numerically.6 In figure 4 we see that the mutual
information obeys the same law as (5) above: I ( X; Θ) ∼ ζ log K with ζ ≈ 0.75. Since the
Fisher volume is proportional to the number of distinguishable states I ( X; Θ) ∼ log V, this
implies (19).
Finally, suppose that instead of a square (or an equilateral triangle), a two-dimensional
Θ has one direction much longer than the other:
√
√
L1 = a1 m,
L2 = a2 m,
a1 a2 .
Then as we increase m we will pass through three regimes, according to how many of the
lengths are long enough to be in the scaling regime:
L1 , L2 > 1
L1 > 1 L2
1 L1 , L2
:
K constant
:
K ∼ L1/ζ ∝ m1/2ζ = m2/3
:
K∼L
2/ζ
=V
1/ζ
∝m
1/ζ
=m
(21)
4/3
.
The last regime is the one we discussed above. When plotting K against log m (or log L), we
expect to see a line with a series of straight segments, and an increase in slope every time
another dimension becomes relevant.7
6
See the appendix of [1] for a discussion of the algorithms used here.
7
These transitions are similar to what LaMont and Wiggins call high-temperature freeze-out [12].
9
4. Conclusion
As with [1] our conclusion must reflect surprise that this law has not been noticed before.
Discreteness in such optimal distributions has been known for a long time [4–7, 13, 14], and
while the numerical work here and in [1] could not have been done on a laptop in 1967, the
analytic derivation would have been equally simple. Maximising mutual information is the
same problem as maximising channel capacity, and the papers which discovered discreteness
[4–6] were working in this area.8 There are various algorithms for this maximisation [17–19],
and those of [18, 19] exploit this discreteness; our Appendix A looks at some data from one
of these [18] which turns out to contain a related scaling law.
Besides the intrinsic interest of a scaling law, this served a useful purpose in [1]. While a
discrete prior seems very strange at first, we are accustomed to thinking about parameters
which we can measure with good accuracy. This means that they have large L. And this
law is what ensures that the neighbouring delta functions then become indistinguishable,
leaving an effectively a continuous prior along relevant directions.
In section 3 we also studied some generalisations beyond [1]. Very near to the end of a
long parameter, the discreteness does not wash out as L → ∞. We observed that the proper
distance to the second delta function is e in the Gaussian case, and π in the binomial (Poisson)
case. At present we cannot explain these numbers, and they may not be exact. Finally we
observed that this law holds in any number of dimensions, if stated in terms of the mutual
information (5). But stated in terms of the length L, it gives a slope which depends on the
number of large dimensions (21).
Acknowledgements
We thank our co-authors Henry and Mark for an enjoyable collaboration on [1].
M.C.A. was supported by NCN grant 2012/06/A/ST2/00396. B.B.M. was supported by
a Lewis-Sigler Fellowship and by NSF PHY 0957573.
A. Chang and Davisson
Chang and Davisson [18] came close to finding this scaling law. They maximised I ( X; Θ)
for various L, and noted the number of delta functions K 0 their algorithm used to achieve
good enough accuracy. But this is not quite the optimal K: they tend to use too many delta
functions, in a way which varies with L. This results in a slightly different law:
I ( X; Θ) ∼ ζ 0 log K 0 ,
ζ 0 > 0.5 .
(22)
Their paper does not mention this, nor attempt to plot this data as in figure 5.
B. Laplace and Sunrise
The limit of fixed proper distance from the boundary (16) also controls another commonly
discussed feature of the Bernoulli model (14), namely Laplace’s rule of succession. After
observing x = 0 out of m trials, we calculate the posterior and use this to predict the result
of one more flip of the coin. Some numerical results are shown in figure 6.
8
This maximisation is also equivalent to a minimax optimisation problem [15], and discreteness was known in
other minimax problems slightly earlier [16]. We believe this amounts to 4 independent discoveries, and so it
would not surprise us to find an earlier example.
10
Mutual information (bits)
2.5
2.0
1.5
Chang & Davisson
ζ = 1 (as above)
ζ = 0.75 (as above)
ζ = 0.5
ζ = 0.4
1.0
2
4
8
16
32
64
Number of delta functions K (log scale)
Figure 5: Approximate scaling law (22) seen in the data from Table 1 of [18].
The standard priors for this problem are the beta distributions
θ α −1 (1 − θ ) β −1
,
pα,β (θ ) =
B(α, β)
Z
dθ pα,β (θ ) = 1
with α = β = 1 clearly a flat prior, and α = β = 1/2 giving Jeffreys prior. These have the
attractive property that the posterior is of the same form, simply p(θ | x, m) = pα+ x, β+m− x (θ ).
Thus it is trivial to calculate the probability of surprise, which depends only on the behaviour
of prior close to θ = 0:
psurprise =
Z
dθ θ pα, β+m (θ ) =
B(α + 1, β + m)
α
= + O(m2 )
B(α, β + m)
m
In this regard the optimal prior p? (θ ) seems to behave like α = β = 1/8.
Calculating instead for a discrete prior, again it is clearly dominated by the first few
delta functions near to φ1 = 0:
psurprise =
∑ θa p(θa |x = 0, m)
a
=
∑ a ≥2 θ a λ a (1 − θ a ) m
λ 1 + ∑ a ≥2 λ a (1 − θ a ) m
=
1
1
1
∑ a≥2 φa2 λ a e−φa /4
·
+
O
≈
+...
2
2
4m λ1 + ∑ a≥2 λ a e−φa /4
m × 7.73
m
2
On the last line we substitute in the numbers from (17). The result is not convincingly close
to an integer, unfortunately. And anyway it is far from clear to us what this this psurprise
means physically.
References
[1] Mattingly, Transtrum, Abbott & Machta, Rational Ignorance: Simpler Models learn more Information
from Finite Data, 1705.01166.
[2] Shannon, A mathematical theory of communication, Bell Sys. Tech. J. 27 (1948) 623–656.
11
Probability of surprise (log scale)
0.3
Flat prior
1/m
Jeffreys
1/2m
Optimal
1/8m
0.1
0.03
0.01
0.003
5
10
20
40
Number of repetitions m (log scale)
Figure 6: Laplace’s rule for the optimal prior, compared to Jeffreys and flat prior.
Lindley, On a Measure of the Information Provided by an Experiment, Ann. Math. Statist. 27 (1956)
986–1005.
Rényi, On some basic problems of statistics from the point of view of information theory, Proc. 5th Berkeley
Symp. on Math. Statist. and Prob. (1967) 531–543. [projecteuclid.org].
[3] Bernardo, Reference Posterior Distributions for Bayesian-Inference, J. Roy. Stat. Soc. B 41 (1979) 113–147.
[4] Färber, Die Kanalkapazität allgemeiner Übertragunskanäle bei begrenztem Signalwertbereich beliebigen
Signalübertragungszeiten sowie beliebiger Störung, Arch. Elektr. Übertr. 21 (1967) 565–574.
[5] Smith, The information capacity of amplitude-and variance-constrained scalar gaussian channels,
Information and Control 18 (1971) 203–219.
[6] Fix, Rate distortion functions for squared error distortion measures, Proc. 16th Annu. Allerton Conf.
Commun. Control Comput. (1978) 704–711.
[7] Berger, Bernardo & Mendoza, On Priors that Maximize Expected Information, Recent Developments in
Statistics and Their Applications J. Klein and J.C. Lee. Seoul Freedom Academy Publishing (1988).
[www.uv.es].
[8] Clarke & Barron, Jeffreys’ prior is asymptotically least favorable under entropy risk, J. Stat. Plan. Infer. 41
(1994) 37–60.
Scholl, Shannon optimal priors on independent identically distributed statistical experiments converge
weakly to Jeffreys’ prior, Test 7 (1998) 75–94.
[9] Jeffreys, An Invariant Form for the Prior Probability in Estimation Problems, Proc. Roy. Soc. A 186 (1946)
453–461.
[10] Gutenkunst, Brown, Myers, Brouwer & Sethna, Sloppy-model universality class and the Vandermonde
matrix., Phys. Rev. Lett. 97 (2006) 150601–4 [cond-mat/0605387].
Machta, Chachra, Transtrum & Sethna, Parameter Space Compression Underlies Emergent Theories and
Predictive Models, Science 342 (2013) 604–607 [1303.6738].
Transtrum & Qiu, Model reduction by manifold boundaries, Phys. Rev. Lett. 113 (2014) 098701–6.
Transtrum, Machta, Brown, Daniels, Myers & Sethna, Perspective: Sloppiness and emergent theories in
physics, biology, and beyond, J. Chem. Phys. 143 (2015) 010901 [1501.07668].
[11] Polson & Wasserman, Prior distributions for the bivariate binomial, Biometrika 77 (1990) 901–904.
[12] LaMont & Wiggins, A correspondence between thermodynamics and inference, 1706.01428.
[13] Zhang, Discrete Noninformative Priors. PhD thesis, Yale University, 1994. [UMI 9523257].
12
[14] Sims, Rational Inattention: Beyond the Linear-Quadratic Case, American Economic Review 96 (2006)
158–163.
Jung, Kim, Matejka & Sims, Discrete Actions in Information-Constrained Decision Problems, .
[princeton.edu/~sims/#RIDiscrete].
[15] Haussler, A general minimax result for relative entropy, IEEE Trans. Inform. Theory 43 (1997)
1276–1280.
[16] Ghosh, Uniform Approximation of Minimax Point Estimates, Ann. Math. Statist. 35 (1964) 1031–1047.
Casella & Strawderman, Estimating a Bounded Normal Mean, Ann. Statist. 9 (1981) 870–878.
Feldman, Constrained Minimax Estimation of the Mean of the Normal Distribution with Known Variance,
Ann. Statist. 19 (1991) 2259–2265.
[17] Arimoto, An algorithm for computing the capacity of arbitrary discrete memoryless channels, IEEE Trans.
Inform. Theory 18 (1972) 14–20.
Blahut, Computation of channel capacity and rate-distortion functions, IEEE Trans. Inform. Theory 18
(1972) 460–473.
[18] Chang & Davisson, On calculating the capacity of an infinite-input finite (infinite)-output channel, IEEE
Trans. Inform. Theory 34 (1988) 1004–1010.
[19] Lafferty & Wasserman, Iterative Markov chain Monte Carlo computation of reference priors and minimax
risk, Proc. 17th conf. Uncert. AI (2001) 293–300 [1301.2286].
Dauwels, Numerical computation of the capacity of continuous memoryless channels, Proc. 26th Symp. Inf.
Th. Benelux (2005). [www.dauwels.com].
13
| 10math.ST
|
Algebraic Matroids with Graph Symmetry
Franz J. Király ∗
Zvi Rosen †
Louis Theran‡
Abstract
arXiv:1312.3777v1 [math.CO] 13 Dec 2013
This paper studies the properties of two kinds of matroids: (a) algebraic matroids
and (b) finite and infinite matroids whose ground set have some canonical symmetry,
for example row and column symmetry and transposition symmetry.
For (a) algebraic matroids, we expose cryptomorphisms making them accessible to
techniques from commutative algebra. This allows us to introduce for each circuit in an
algebraic matroid an invariant called circuit polynomial, generalizing the minimal polynomial in classical Galois theory, and studying the matroid structure with multivariate
methods.
For (b) matroids with symmetries we introduce combinatorial invariants capturing
structural properties of the rank function and its limit behavior, and obtain proofs which
are purely combinatorial and do not assume algebraicity of the matroid; these imply
and generalize known results in some specific cases where the matroid is also algebraic.
These results are motivated by, and readily applicable to framework rigidity, low-rank
matrix completion and determinantal varieties, which lie in the intersection of (a) and
(b) where additional results can be derived. We study the corresponding matroids
and their associated invariants, and for selected cases, we characterize the matroidal
structure and the circuit polynomials completely.
∗
Department of Statistical Science, University College London; and MFO; [email protected]
Department of Mathematics, University of California, Berkeley, [email protected]
‡
Inst. Math., Freie Universität Berlin, [email protected]
†
1
Contents
1 Introduction
1.1 Representations of Algebraic Matroids .
1.2 Matroids with Graph Symmetry . . . . .
1.3 Multivariate Minimal Polynomials . . . .
1.4 Main Contributions . . . . . . . . . . . . .
1.5 Context and Novelty . . . . . . . . . . . .
1.6 A Reading Guide to the Main Theorems
1.7 Acknowledgements . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
4
5
5
5
6
7
9
2 Matroid Realizations and their Symmetries
2.1 Preliminaries . . . . . . . . . . . . . . . . .
2.2 Matroids of Ideals and Varieties . . . . .
2.3 Graph Matroids . . . . . . . . . . . . . . .
2.4 Symmetrization and Bipartition . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
9
11
14
17
3 Determinantal and Rigidity Matroids
3.1 The Determinantal Matroid . . . . . . .
3.2 The Symmetric Determinantal Matroid
3.3 The Rigidity Matroid . . . . . . . . . . .
3.4 The Bipartite Rigidity Matroid . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
19
19
21
22
25
.
.
.
.
.
.
.
.
27
27
28
30
33
36
37
40
45
.
.
.
.
.
.
46
46
47
49
51
52
53
6 Algebraic and Combinatorial Structure Theorems
6.1 Symmetries of Circuit Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Constructing circuits with the (t, 1)-move . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
54
56
7 A Dictionary of Circuits
7.1 Symmetries and Number of Circuits for Bipartite Graph Matroids
7.2 The Determinantal Matroid . . . . . . . . . . . . . . . . . . . . . . . .
7.3 The Bipartite Rigidity Matroid . . . . . . . . . . . . . . . . . . . . . .
7.4 Symmetries and Number of Circuits for Graph Matroids . . . . . .
58
58
59
63
65
.
.
.
.
4 Limit Matroids and their Rank Function
4.1 Direct Limits of Matroids . . . . . . . . . . . . . . . . . . . .
4.2 One-sided Bipartite Graph Matroid Limits . . . . . . . . . .
4.3 Relative Rank, Average Rank and Realization Size . . . . .
4.4 Circuits and Bases of One-Sided Bipartite Graph Limits . .
4.5 Two-sided Bipartite Graph Matroid Limits . . . . . . . . . .
4.6 Average Rank, and Realization Size for Two-Sided Limits
4.7 Circuits and Bases of Two-Sided Bipartite Graph Limits . .
4.8 Infinite Graph matroids . . . . . . . . . . . . . . . . . . . . .
5 Circuit Polynomials and Top-Degrees
5.1 Top-degree . . . . . . . . . . . . . .
5.2 Circuit polynomials . . . . . . . . .
5.3 The Multihomogenization . . . . .
5.4 MH-bases . . . . . . . . . . . . . . .
5.5 Finiteness of top-degrees . . . . .
5.6 Multivariate Galois Theory . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7.5 The Symmetric Determinantal Matroid . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.6 The Rigidity Matroid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
65
66
1. Introduction
1.1. Representations of Algebraic Matroids
Since their discovery by van der Waerden [36], algebraic matroids have been one of the central
objects of interest in matroid theory, and therefore in mathematics, as they describe algebraic dependence relations in Galois extensions with multiple elements. Their structure, the possible ways to
realize them, and their relation to linear matroids has been extensively explored.
In recent literature, a different way of viewing algebraic matroids has arisen - namely describing satisfiability of equations. One preliminary link in this direction can be found in the literature for the framework realization problem, which asks for the reconstruction of an unknown
n
set of points p1 , p2 , . . . , pn ⊂ R r given a subset of the 2 possible pairwise (squared) distances
di j = pi> pi − 2pi> p j + p>
j p j . A slight rephrasing of the question yields on of the central problems inthe field of combinatorial rigidity (see [13] for a recent review on the field), which asks the following: Given some subset di j , (i j) ∈ S of all distances, which of the other distances are determined
either uniquely, or up to a finite choice? By studying the Jacobian of the map {pi } 7→ {di j }, this
can be recast in the language of linear matroids, yielding novel insights about the rigidity problem [20, 24, 33, 35, 37].
Similar and more recent links can be found in the case of the low-rank matrix completion problem, which asks for the reconstruction of a large matrix A ∈ Cm×n of known rank r from a subset
Ai j , (i j) ∈ S of known entries, and which is also of great practical interest in engineering and computer science (see [7], or the introduction [17] for an overview). For matrix completion, a link to
linear matroids has first been seen by the authors of [31] by several conjectural generalizations of
known properties of the rigidity problem to the case of matrix completion.
In [17], it has been established for the case of matrix completion, that these relations to linear
matroids are consequences of a deep connection to algebraic matroids - the linear matroids are
concrete instances of the non-constructive realizability theorems which link algebraic matroids constructed from Galois extensions to linear matroids obtained from their differential, or Jacobian. In
particular, there is an algebraic matroid associated to the matrix completion problem which gives
rise to the matrix completion problem, which in turn explains the links to linear matroids which
have been previously observed.
In this paper, we show that to each algebraic matroid over an algebraically closed ground field,
there is a reconstruction problem of the following type:
Reconstruct a point x on an irreducible variety X ⊂ K N from a fixed set x i , i ∈ S of coordinates in a
fixed basis.
Moreover, we will show that the converse is equally true; for each such reconstruction problem,
there is an algebraic matroid from which it arises. That is, each algebraic matroid can be represented
by such a basis projection problem, properties of the projections giving rise to the independent
set of a matroid. Furthermore, if the ground field is not algebraically closed, the matroid can
be represented in terms of regular sequences in a ring. These triple correspondence, which is
reminiscent of other such hidden equivalences in matroid theory, we term algebraic cryptomorphism.
4
1.2. Matroids with Graph Symmetry
A different phenomenon which is observed in the two motivating examples is additional combinatorial structure. For matrix completion, the entries of A are naturally identified with the indices of
[n]
the matrix [m] × [n], and the pairwise distances in rigidity with unordered pairs 2 . Thus, there
is an identification between sets of possible observations, and labelled (bipartite) graphs. Since
permuting the rows and columns of A or relabeling the points pi does not affect which entries can
be reconstructed, we see that:
Whether a set of coordinates determines x depends only on an unlabelled graph.
Rephrased in terms of the associated algebraic matroid, the corresponding tower of Galois extensions is invariant under the symmetry action of the permutation groups.
What we will show that much can already be inferred purely from this combinatorial graph
symmetry of the matroids. That is, even when the initial assumption that the matroid is algebraic is
dropped, several phenomena which have been observed in both combinatorial rigidity and matrix
completion - such as asymptotic moves and limiting behaviour of dimension - are still present in
terms of matroid invariants, and can be proven to hold by combinatorial means.
1.3. Multivariate Minimal Polynomials
As briefly discussed above, the algebraic cryptomorphism allows to relate algebraic matroids to
regular sequences in a polynomial ring. Namely, the reconstruction problem is closely related to the
question
Given a prime ideal P, which sets of coordinates X i , i ∈ S give rise to a regular sequence/regular
coordinates in the integral domain K[X 1 , . . . , X n ]/P ?
Again, each algebraic matroid gives rise to such a regular coordinate problem, and each coordinate problem gives rise to an algebraic matroid. Through the ring theoretic formulation, we will
be able to derive algebraic invariants for algebraic matroids, which generalize classical concepts
from Galois theory such as minimal polynomial and extension degree to the multivariate setting of
any algebraic matroid. Namely, to each circuit of the underlying algebraic matroid, we associate a
single polynomial called circuit polynomial, which is a multivariate generator of all finite extensions
in the circuit. The collection of its degrees, the so-called top-degree, is a further invariant which
completely characterizes the degree behavior of those finite extensions.
Furthermore, if the assumption of the matroid being algebraic is added again, structural statements about algebraic invariants such as the ideal of the irreducible variety X or the possible topdegrees can be inferred. Those, in turn, allow novel insights on the matrix completion and combinatorial rigidity problem, such as on the number of possible reconstructions or the reconstruction
process.
1.4. Main Contributions
Summarizing, the main contributions in this paper are as follows:
• We introduce basis matroids, arising from a reconstruction problem as described above, and
coordinate matroids, related to regular sequences in polynomial rings. We show how algebraic
matroids can be equivalently - cryptomorphically - realized as either of those and that the
realization spaces are isomorphic.
5
• We introduce a notion of matroids with symmetries, or orbit matroids, and show how the
matroid properties extend to those. We identify the main invariants associated to those and
derive structure theorems in the case of graph symmetry.
• We exploit the equivalence of algebraic and basis matroids to obtain invariant polynomials of
algebraic matroids - the circuit polynomials - which generalize the minimal polynomial from
classical Galois theory.
• We provide further characterization, structure and finiteness results in the case the matroid is
both algebraic and graph symmetric, which partly imply known results in matrix completion
and combinatorial rigidity, and partly are novel.
• We provide a full characterization of some matrix completion and combinatorial rigidity matroids.
1.5. Context and Novelty
The results of this paper sit at the intersection of several active fields. Here is the general context.
1.5.1. Matroid Theory One of the most attractive features of matroid theory is the availability
of many superficially-different-seeming, but equivalent axiom systems [27, Section 1]. Our constructions for algebraic matroids follow this general theme, but with respect to providing different
realizations that capture the same algebraic invariances. The program here is also different than the
important questions of realizability over different fields or of understanding the different orientations of the same underlying matroid. We furthermore study matroids exhibiting certain characteristic symmetries, which is a reoccurring topic in matroid theory.
Some motives in our work have already appeared in existing work in different context: the
matroidal families studied in [29] can be seen as a specific case of our orbit matroids, compare
Remark 2.23. The circuit polynomial has briefly occurred under different name in [12].
1.5.2. Rigidity Theory In the preprint [4], Borcea formulated framework rigidity in terms of coordinate projections of the Cayley-Menger variety. The paper [3] then used relationships between
determinantal and Cayley-Menger varieties to derive enumerative results on the number of realizations of minimally rigid frameworks. We use some similar ideas to obtain upper and lower bounds
on the top-degrees, but studying circuit polynomials and top-degrees is a new contribution.
1.5.3. Matrix Completion Low-rank matrix completion has been studied extensively, with most
of the effort directed towards heuristics for the reconstruction problem. Work of Cucuringu-Singer
[32] on the decision problem anticipates some of our results, but works in only one of our algebraic
matroid regimes and applies only to the determinantal variety.
It is fair to say that the most powerful tools for this are based on convex relaxations, introduced
by Candès-Recht [7], and then developed further by a large number of authors (most notably, [8]).
Spectral methods [16] have also given essentially optimal bounds in a number of regimes. Typical
results of either type give a (a) “with high probability” reconstruction guarantee for the whole matrix,
given: (b) some analytic hypothesis about the true matrix A; and (c) a sampling hypothesis for the
observations.
By contrast, the methods here provide much more precise information: under only the assumption of genericity, we can determine exactly the entries which are reconstructable from any fixed
set of observations; moreover, top-degrees and circuit polynomials provide algebraic certificates
bounding the number of solutions for any fixed entry, making the decision version of the problem
more quantitative.
6
1.5.4. Universality versus Genericity Let us consider again the reconstruction problem above,
with πb(S) denoting the projection onto the span of {bi : i ∈ S}. It is now a very natural question to
consider the following “decision version”:
For x ∈ X, and a fixed set of coordinates S ⊆ E, determine the dimension and cardinality of π−1
(x).
b(S)
For frameworks, the decision version is the “rigidity problem”. If we insist on treating the reconstruction problem for every x ∈ X, phenomena related to so-called Mnëv-Sturmfels Universality imply that due to a complexity which is arbitrarily bad, even the decision problem has no satisfactory
answer - see [21, 25, 28, 30] for the theory and [11] for the specialization to matrix completion).
Our results show that there is only one generic behavior of the decision problem - that is, if x is
chosen from a generic/Zariski open set, the fiber dimension of πb(S) is controlled completely by the
rank of an algebraic matroid on E, which does not depend on x anymore. What this means, is that
the reconstruction problem has a combinatorial part, captured by a matroid M = BK (X, b) on E,
and an algebraic-geometric part which depend on X and b.
However, the algebraic-geometric, or universal, behaviour occurs only on a Zariski closed, and
therefore probability zero, subset of the variety X. Therefore, while universality makes statements
of the kind
Given some arbitrarily (universally) complex behaviour, there is a variety X, a projection π, and a
point x ∈ X where it occurs,
our results can be interpreted as a more optimistic converse
Given some arbitrarily complex variety X, the behaviour of an an arbitrary projection π at a generical
point x ∈ X is governed by an algebraic matroid.
Arguing that the variety X is usually fixed in the applications, and x is usually generic, we conclude
that universality does not occur in such applications, making way to efficient algorithms - such as
those discussed in [17] - that exploit genericity.
1.6. A Reading Guide to the Main Theorems
In this paper, we generalize and strengthen the results mentioned above. The determinantal matroid
in its two realizations (by an ideal and a variety) is an instance of a construction that applies to any
irreducible variety. Moreover, certain structural results derived by algebraic means in [17] result
only from the underlying symmetry, and can be proved purely combinatorially. Here is an overview
of the results.
1.6.1. Algebraic Matroids and Circuit Polynomials Our first results are about different ways of
realizing algebraic matroids. Full definitions are given in Section 2.
Theorem 2.2 (informal). If M is an algebraic matroid on ground set E, it has a realization as any
of: (a) a field and an extension by a set of transcendentals; (b) an irreducible variety and a specific
coordinate system; (c) a prime ideal and a specific choice of coordinate variables. Moreover, there are
natural constructions for converting from one type to the other two, and these constructions commute.
Theorem 2.2 says that, depending on what is more convenient for an application, we can switch
algebraic formalisms without losing any of the additional structure carried by a specific realization.
This should be compared with classical results saying, e.g., that all algebraic matroid are linear in
characteristic zero.
7
As mentioned above, we can associate to the circuits of an algebraic matroid polynomial invariants called circuit polynomials. These witness the minimal algebraic dependencies among the
circuits in M.
Lemma 5.6 (informal). Let M = CK (P, x) be an algebraic matroid realized via an ideal P. For each
circuit C ∈ C(M), there is a unique (up to scalar multiplication) circuit polynomial θx(C) (P) ∈ P. The
circuit polynomials are the minimal supports in P.
In our applications, the uniqueness of circuit polynomials implies that, for a circuit, C containing
an entry (i, j), there is exactly one way to solve for the variable x(i, j) given the variables x(C \(i, j)).
Thus, the degree of x(i, j) in the circuit polynomial of C yields information about the the fiber
cardinality of πb(C\(i, j)) at a generic point. Considering the degrees of every variable in a circuit
polynomial at once leads to the notion of top-degrees of circuit polynomials and, indeed, of any
polynomial in an algebraic matroid’s underlying ideal.
1.6.2. Bipartite Graph Matroids Now consider the general case of a matroid M defined on
[m] × [n], with the property that for all S ⊂ [m] × [n], σ ∈ S(m) and τ ∈ S(n), the set
(σ(i), τ( j)) : (i, j) ∈ S has the same rank in M. We call such a matroid a bipartite graph matroid, since the rank depends only on the unlabeled bipartite graph corresponding to S. This theory
is developed in Section 4.
Using only the symmetries we can show:
Theorem 4.3 (informal). Let M be a bipartite graph matroid on the ground set [m] × [n], with
m, n 0. Then for each m0 ≤ m, there is an n0 , depending only on m0 such that any circuit in M
supported on at most m0 rows has support on at most n0 columns.
In particular, if we regard m as fixed and n increasing to infinity, we see that the number of
isomorphism classes of circuits is finite, answering a question Bernd Sturmfels asked us. On the
other hand, when both indices grow, we construct infinite families of circuits in the determinantal
and rigidity matroids.
1.6.3. Structure of Algebraic Bipartite Graph Matroids Combining the algebraic and combinatorial theories together, we are able to give similar finiteness and non-finiteness statements in Section
6 for algebraic bipartite graph matroids. Here, the combinatorial objects of study are top-degrees of
polynomials, which are matrices recording the degree in each variable.
Theorem 5.2 (informal). Let M = CK (P, x) be an algebraic bipartite graph matroid on the ground
set [m] × [n], with m, n 0. Then for each m0 ≤ m, there is an n0 , depending only on m0 such that
any minimal top-degree in M supported on at most m0 rows has support on at most n0 columns. This,
in particular, implies that the number of solutions to any minimal top-degree is bounded by a function
of the size of its row support.
In particular, this says that, for a fixed number of rows, the complexity of the reconstruction
problem is bounded. As before, on a square ground set, the complexity increases rather rapidly,
as we will show via examples. Section 7 contains a detailed treatment of small determinantal
matroids, for which we can compute all the matroidal structures and polynomial invariants exactly.
1.6.4. Multivariate Galois Theory By combining the notion of top-degree with circuit polynomials, we obtain the following reformulation of Lemma 5.6 in terms of field extensions.
Theorem 5.4 (informal). Let L = K(α) be a field extension by transcendentals α = (α1 , . . . , αn ),
and suppose that the αi form an algebraic circuit. There is a (unique up to scalar multiplication)
polynomial θ with top degree t i = [L : K(α \ αi )] in each variable αi and θ (α) = 0.
8
Here is the interpretation. We first recall the minimal polynomial from Galois theory of an
element α of a field extension L/K; this is a univariate polynomial θ of degree [L : K], with the
property that θ (α) = 0. This is unique, up to a scalar in K × , and it encodes α, the degree of L/K,
and (via the conjugate solutions of θ ), information about the symmetries of L. The polynomial θ
provided by Theorem 5.4 is, then, the analogous object for extensions by multiple elements: the
top-degrees encode the degrees of the “one element” extensions by each variable, and symmetries
of θ mirror those of the underlying algebraic matroids.
1.7. Acknowledgements
We thank Bernd Sturmfels for helpful comments. We thank Winfried Bruns for insightful discussions
and many helpful comments.
LT is funded by the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013) / ERC grant agreement no 247029-SDModels. This research
was initiated at the Mathematisches Forschungsinstitut Oberwolfach, when LT and ZR were guests
supported by FK’s Leibniz Fellowship.
2. Matroid Realizations and their Symmetries
2.1. Preliminaries
We first set some notation and recall standard definitions and results required in the sequel. The
basic notions of matroids, and in particular, over finite sets can be found in the monograph [27].
Notation 2.1. We adopt the convention of a matroid M being an ordered pair (E, I), with E = E(M)
being the (finite) ground set of M (where E can be a multiset), and I ⊆ P(E) the collection of
independent sets of M (each element of I is a submultiset of E). However, instead of repeatedly
specifying a matroid as a pair, for readability, we will specify each matroid by its ground set E(M),
and by making the following invariants explicit, each sufficient for an equivalent definition of M:
I(M)
the set of independent sets of M
D(M)
the set of dependent sets of M
B(M)
the set of bases of M
C(M)
the set of circuits of M
rkM (·) : P(E) → N the rank function of M
I(M), D(M), B(M), C(M) are all subsets of the power set P(E). If clear from the context, we will
omit the dependence on M.
As usual, we consider two matroids isomorphic if their is an isomorphism on the ground sets
inducing an isomorphism on any of I, D, B, C or rk:
Definition 2.2. Let M and N be matroids on ground sets E and F . A morphism of matroids is a pair
(ϕ, ψ) of maps ϕ : I(M) → I(M) and ψ : E → F , such that, for all S ∈ I
ϕ(S) = {ψ(s) : s ∈ S}
We will say that ψ induces the morphism from M to N. A morphism (ϕ, ψ) is an isomorphism if ϕ
and ψ are bijections. A self-isomorphism of a matroid is called an automorphism. When it is clear
from the context, we will use the same symbol for both maps.
Therefore, a matroid M is considered as a purely combinatorial object, which does not formally
carry any information on what E, I, etc. describe.
9
Conversely, one can construct matroids which describe dependences on a certain ground set
which comes with algebraic or combinatorial structure. Obtaining a matroid M in this way is called
a realization of M, associated to the ground set with the dependence structure.
Notation 2.3. Let M = (E, I). A realization of M is some data D, such that M can be obtained from
some constructors E = E(D) and I = I(D). The construction process, if clear from the context, will
be understood. Usually, the data D also contains some (multi-)set d = {de : e ∈ E}; in this case, for
S ⊆ E, we will denote d(S) := {ds ∈ d : s ∈ S}.
We will think of a realization as a matroid which formally carries the additional structure given
by the data D and the associated construction process, as compared to the matroid alone which has
no structure or constructor associated. We illustrate this subtle difference in the well-known case of
linear matroids [27, Proposition 1.1.1]:
Definition 2.4. Let E be some finite ground set, let K be a field. Let v = {ve : e ∈ E} be a collection
of elements ve ∈ V (possibly a multiset), where V is some vector space over K. We construct a
matroid LK (v) from v and K in the following way:
E
is the ground set
S∈I
iff v(S) is linearly independent over K
S∈D
iff v(S) is linearly dependent over K
S∈B
iff v(S) is a basis of the K-vector space span v
S∈C
iff v(S) is a linear circuit over K
rk : P(E) → N, v(S) 7→ dim span v(S)
We call LK (v) (and therefore any matroid isomorphic to it) a linear matroid (over the field K). The
data (K, v) is called a linear realization of the matroid LK (v) over the field K, and LK (v) is said to
be realized by (K, v).
Another important way of realization is through field extensions in the case of algebraic matroids [27, Theorem 6.7.1]::
Definition 2.5. Let E be some finite ground set, let K be a field. Let α = {αe : e ∈ E} be a collection
of elements αe ∈ L (possibly a multiset), where L is some finite field extension of K. We construct
a matroid AK (α) from α and K in the following way:
E
is the ground set
S∈I
iff α(S) is algebraically independent over K
S∈D
iff α(S) is algebraically dependent over K
S∈B
iff α(S) is a basis for the field extension K(α)/K
S∈C
iff α(S) is an algebraic circuit over K
rk : P(E) → N, α(S) 7→ trdeg K(α)/K
We call AK (α) (and therefore any matroid isomorphic to it) an algebraic matroid (over the field K).
The data (K, α) is called an algebraic realization of the matroid AK (α) over the field K, and AK (α)
is said to be realized by (K, α).
Note that by our convention, saying that a matroid is linear or algebraic is a qualifier for a matroid, which does not formally include the specification of how - or in which ways - the matroid
can be obtained from data. Conversely, for realizations, we will always think of the matroid associated to it by canonical construction.1 These two viewpoints give rise to different flavours of
questions: namely, which matroids can be realized by a certain constructor, e.g., which matroids
are linear, algebraic etc., and if such a matroid is given, what are the possible realizations for it?
When considering more than one constructor, e.g., linear and algebraic, one might ask how being
1
An alternative way of formalizing this would be to consider the category of matroids and categories of realizations,
e.g., linear matroids or algebraic matroids. Each of the realization categories comes with a forgetful functor to matroids.
Saying that a matroid can be realized means that it is in the image of the respective forgetful functor. We refrain from
doing so in order to avoid confusion and unnecessary notational overhead.
10
realizable by the one and being realizable by the other relate, and if both concepts coincide, how
do the realizations relate?
For illustration, we phrase a well-known theorem on linear and algebraic representability [27,
Propositions 6.7.10f] in these terms:
Theorem 2.1. Let M be a matroid, let K be a field of characteristic 0. The following statements are
equivalent:
(a) M is realizable by a linear matroid over K.
(b) M is realizable by an algebraic matroid over K.
Moreover, given a presentation
M = AK (α) = LK (v)
there are constructions producing any of (K, α), (K, v), given the other.
It is a difficult and unsolved question to characterize all possible linear and algebraic realizations, and to relate them to each other - a question which will not be the main concern of this
paper. Instead, we will be interested in matroids that are realized by natural algebraic constructions - namely, matroids associated to varieties, and to prime ideals. We will be interested in how
these different kinds of realizations relate to each other, and we will use these relations to derive
structure theorems for those matroids in the presence of certain symmetries.
2.2. Matroids of Ideals and Varieties
In the following, we provide two novel classes of matroid realizations, basis and coordinate matroids. We will show later in Theorem 2.2 that realizations as algebraic, basis and coordinate
matroids can be canonically transformed into each other, therefore yield the same class of matroids. Later, when studying their structure and symmetries, it will turn out that each of the three
representation has its situational advantage.
Definition 2.6. Let E be some finite ground set, let K be a field. Let b = {be : e ∈ E} be a basis
of K n , let X ⊂ K n be an irreducible variety. We construct a matroid B(X, b) from b and X in the
following way:
E
is the ground set
S ∈ I iff the canonical projection map πb(S) : X → span b(S) is surjective
S ∈ D iff the canonical projection map πb(S) : X → span b(S) is not surjective
S ∈ B iff the canonical projection map πb(S) : X → span b(S) is finite surjective
We call B(X, b) (and therefore any matroid isomorphic to it) a basis matroid over K. The data (X, b)
is called an basis realization of the matroid B(X, b), and B(X, b) is said to be realized by (X, b).
Remark 2.7 — Because X is irreducible, any image πb(S) (X) also is for every S. Therefore, an
alternative definition for the matroid can be obtained by defining S ∈ I iff dim πb(S) (X) = #S.
Definition 2.8. Let E be some finite ground set, let K be a field. Let x = {X e : e ∈ E} be a set (not
a multiset) of coordinate variables, let P ⊆ K[X e , e ∈ E] be a prime ideal. We construct a matroid
C(P, x) from x and P in the following way:
E
is the ground set
S ∈ I iff x(S) is a regular sequence modulo P
S ∈ D iff x(S) is not a regular sequence modulo P
S ∈ B iff x(S) modulo P is a generating regular sequence for the ring K[x]/P
We call C(P, x) (and therefore any matroid isomorphic to it) a coordinate matroid over K. The data
(P, x) is called a coordinate realization of the matroid C(P, x), and C(P, x) is said to be realized by
(P, x).
11
That the definitions for the basis/coordinate matroid via any of I, D, B are equivalent follows
from properties of algebraic maps and regular sequences; that both give rise to matroids is not.
However, one can infer from the proof of the subsequent Theorem 2.2 that basis and coordinate
matroids are indeed matroids, therefore we state this explicitly to assert well-definedness:
Proposition 2.9. The ground set/independent set pairs, as defined in Definitions 2.6 and 2.8 for basis
and coordinate matroids, are matroids (in the sense of Notation 2.1).
However, even more is true. Algebraic, basis and coordinate realizations can be regarded as
different realizations of the same kind of matroid, all three existing whenever one of the three does.
Moreover, the following theorem asserts that there are canonical, commuting ways to obtain any of
these realizations from the other:
Theorem 2.2. Let M be a matroid, let K be an algebraically closed field. The following statements are
equivalent:
(a) M is algebraic over K
(b) M is a basis matroid over K.
(c) M is a coordinate matroid over K.
Moreover, given a presentation
M = AK (α) = B(X, b) = C(P, x),
there are canonical constructions producing any two of (α, K), (X, b), (P, x) from the third, such that
these constructions commute.
Proof. Without loss of generality, and for ease of notation, we can assume that the ground set of M
is [n] by isomorphism. The following paragraphs construct the realizations for M on the right side
of the arrow, assuming the left side is a realization for M. In particular, for S ⊆ [n] it is shown that
S ∈ I on the left side if and only if S ∈ I on the right side.
AK (α) → C(P, x): Let R = K[α]. Since Frac R ⊂ L exists, R is an integral domain. This implies
that the kernel of the canonical surjection K[X 1 , . . . , X n ] → R is a prime ideal P and that R =
K[X 1 , . . . , X n ]/P. By the construction, α(S) is algebraically independent if and only if x(S) is a
regular sequence on K[X 1 , . . . , X n ]/P.
C(P, x) → B(X, b): Let X = V(P). Since P is prime, the variety X is irreducible. For each i ∈ [n]
the variety Vi defined by X 1 = X 2 = · · · = X i−1 = X i+1 = · · · X n = 0 is a line. Pick bi as any non-zero
vector in Vi . Since K is algebraially closed, x(S) is a regular sequence on K[X 1 , . . . , X n ]/P if and
only if the projection πb(S) is surjective.
B(X, b) → AK (α): For each i ∈ E, let Vi := span bi . Define A := V1 × V2 × · · · × Vn . It holds
that K(Vi ) = K(X i ) for some transcendental variable X i . The canonical map ι : X ,→ A then
induces a canonical field homomorphism ψ : K(X 1 , . . . , X n ) → K(X) which is surjective since ι is
injective. Taking L = K(X) and αi = ψ(X i ) we see that, by construction, for S ⊂ [n], the set α(S)
is algebraically independent in L/K if and only if the projection πb(S) is surjective.
There are several remarks in order.
12
Remark 2.10 — If K is not algebraically closed in Theorem 2.2, e.g., the real numbers R, it can
happen that the variety X is empty, therefore the correspondence does not hold. However, the
equivalence can be made to hold for general K by replacing the variety X by the corresponding
scheme, with the proof of Theorem 2.2 holding almost verbatim.
Remark 2.11 — Theorem 2.2 is, by a subtlety, stronger than the analogue for linear versus algebraic matroids given by Theorem 2.1: the constructors producing realizations in Theorem 2.2 for
algebraic/basis/coordinate matroids commute, i.e., by converting realizations to each other in a
closed cycle, the conversion constructors will yield isomorphic data. It seems to be unclear whether
the constructors yielding algebraic from linear data, see [27, Propositions 6.7.10], and linear from
algebraic data, see [27, Propositions 6.7.11 and 6.7.12], can be chosen to commute. Therefore,
the constructors in Theorem 2.2 for the algebraic types of matroid are canonical, while the ones in
Theorem 2.1 given for linear and algebraic matroids are not, or not known to be.
Remark 2.12 — For an algebraic matroid M, the realization space is defined as
R := {(K, α) : AK (α) = M}.
The constructions used to prove Theorem 2.2 supply canonical, commuting bijections between
R and the spaces
S := {(X, b) : B(X, b) = M},
and
T := {(P, x) : C(P, x) = M}.
In particular, these spaces are canonically isomorphic to each other, and thus, the second part of
Theorem 2.2 represents a substantial strengthening of the first part.
Another way of looking at this phenomenon is via the canonical forgetful functors from the
category of the realizations to the category of matroids, the realization spaces being fibers of the
respective functors.
In light of the discussion above, the following definition is non-trivial.
Definition 2.13. Let M be a matroid, such that M = AK (α) = B(X, b) = C(P, x). The realizations
(K, α),(X, b) and (P, x) of M are called cryptomorphic realizations of M if they can be obtained
from each other via the canonical constructions of Theorem 2.2.
We want to remark that Theorem 2.2 yields an alternative proof for part of the equivalence of
linear and algebraic matroids presented earlier in Theorem 2.1, together with a structural interpretation.
Proposition 2.14. Let M be an algebraic matroid, with cryptomorphic realizations
M = AK (α) = B(X, b) = C(P, x).
Assume that K has characteristic zero. Let Ω1L/K be the K-vector space of Kähler differentials of L/K,
and let d : L → Ω1L/K be the canonical map. Furthermore, let ξ be a generic2 closed point of X,
and denote by d ξ : L → Ω1L/K ⊗K K(ξ) the canonical map which is obtained from composition with
evaluation at ξ. Then, the following matroids are isomorphic:
(i) the algebraic matroid AK (α).
(ii) the linear matroid LK 0 (v), where K 0 = K(X), and v = {dκ : κ ∈ α}.
(iii) the linear matroid LK (v), where v = {d ξ κ : κ ∈ α}.
2
chosen from a suitable open dense subset of X
13
Proof. (a)⇔ (b): Use Theorem 2.2 to pass to B(X, b), and observe that, in for polynomial rings the
canonical map is given by the usual partial derivatives. For S ⊂ [n], and characteristic zero, the
projection πb(S) is surjective exactly when its Jacobian matrix has full rank.
(b)⇔ (c): the Jacobian matrix has full rank if and only if it has full rank when evaluated on an
open dense subset of X.
Proposition 2.14 implies that algebraic matroids are linear, and in particular that a linear realization can be obtained over the same base field, compare [27, Propositions 6.7.11 and 6.7.12],
or [22]. It furthermore provides a structural statement how to obtain a linear realization from an
algebraic one. We finish this subsection by describing the rank functions.
Proposition 2.15. Consider a matroid M on the ground set E with its cryptomorphic realizations
M = AK (α) = B(X, b) = C(P, x).
Then, the following rank functions rk. : P(E) → N, defined by taking an arbitrary S ⊆ E, are equal:
(a) rkAK (α) (S) := trdeg K(α(S))/K, where trdeg denotes transcendence degree.
(b) rkB(X,b) (S) := dim πb(S) (X), where π· is the projection as in Definition 2.6, and dim denotes
Krull dimension.
(c) rkC(P,x) (S) := dim (K[x]/P ∩ K[x(S)]), where dim denotes Krull dimension.
(d) rkM (S), the rank function of M.
Proof. Recall that the rank of a set S is just the size of a maximal independent set contained in
S. The proposition now follows from chasing definitions: (a) the transcendence degree is the size
of a maximal set of algebraically independent elements of α(S); (b) follows from Remark 2.7; (c)
follows from the fact that all the rings here have the Cohen-Macaulay property.
2.3. Graph Matroids
Some ground sets come equipped with natural symmetry actions. The motivating example for this
paper is the set E = [m] × [n], with the group G = S(m) × S(n) acting freely on the factors. This
is motivated by the fact that G is the symmetry group of the complete bipartite graph Km,n (arrange
the edges in a table), and choosing a subset of E up to G-action can be viewed as choosing a set of
edges of Km,n .
Similarly, if M is a matroid on the ground set E = [m] × [n] and and G = S(m) × S(n) is a
group of automorphisms of M, the the rank of a set S ⊆ E is invariant under the G-action, thus
depending only on the isomorphism type of S as a bipartite subgraph of Km,n . This motivates our
terminology of graph matroids. Having such a graph symmetry is quite a strong property for M,
and we will see in the sequel that such matroids have many interesting properties, especially when
they are also algebraic (or, equivalently, basis/coordinate matroids).
We start by setting the notation:
Notation 2.16. Let E be a set and G a group of automorphisms of E. For S ⊆ E, we will denote by
S/G = {σS : σ ∈ G} the set of G-orbits3 of S; similarly, for the power set P(E), we will denote
by P(E)/G := {S/G : S ⊆ E}. Note that there is a canonical surjection P(E) → P(E)/G, sending
G-stable subsets I ⊆ P(E) to subsets I/G ⊆ P(E)/G. Also note, that in general P(E)/G is not the
same as P(E/G).
3
We deliberately write S/G instead of the usual notation GS or G · S, since we identify certain elements of the power
set P(E) under the G-action, therefore yielding a concept closer to a categorical quotient of M than of a categorical orbit.
14
If the ground set of a matroid has a symmetry, then whether a subset is independent only
depends on the orbit. We formalize this with the following notion:
Definition 2.17. Let M = (E, I) be a matroid, let G be a group of automorphisms of M. The
pair M/G := (E/G, I/G) is called the corresponding orbit matroid. Independent, dependent sets,
circuits, and bases of M/G are elements of P(E)/G which are images of independent, dependent
sets, circuits and bases of M under the quotient map P(E) → P(E)/G. Furthermore, we introduce
terminology for specific symmetries:
Notation
Graph
[n]
2
(Bipartite)
Graph
Ground Set E
Symmetry Group G
:= [n] × [n]/S(2)
S(n) acting diagonally on [n]×[n], i.e., σ(i, j) =
(σi, σ j).
S(m)× S(n) acting on E by the canonical product
action, i.e., (σ, τ)(i, j) = (σi, τ j).
[m] × [n]
The canonical symmetry group G for any type of matroid M in the table above will be assumed to
be a group of automorphisms, when talking about graph matroids, or bipartite graph matroids, and
will be denoted by S(M).
An orbit matroid M = (E, I) with (bipartite) graph symmetries will be called a (bipartite) graph
matroid. We will identify the ground set of a bipartite graph matroid with the complete bipartite
graph, and we will write Km,n = [m] × [n]/S(m) × S(n); similarly, we identify the ground set of
[n]
a graph matroid with the complete graph, writing Kn = 2 /S(n). Similarly, we identify subsets
of P(E)/G with the corresponding subgraphs of Km,n and Kn . As for matroids, these subgraphs will
fall into different sets:
I(M/G) = {S/G : S ∈ I(M)} the set of independent graphs of M/G
D(M/G) = {S/G : S ∈ D(M)} the set of dependent graphs of M/G
B(M/G) = {S/G : S ∈ B(M)} the set of basis graphs of M/G
C(M/G) = {S/G : S ∈ C(M)} the set of circuit graphs of M/G
That is, we will also call independent, dependent sets, circuits and bases of M/G independent,
dependent, circuit and orbit graphs. If clear from the context, we will not make a sharp distinction
between M and M/G and, for example, talk about circuit graphs of M when meaning a circuit
of M/G. Similarly, if G is the graph corresponding to some E ⊆ [m] × [n], we will also write
rk(G) := rk(E).
Conceptually, the ground set/subset structure of a matroid is replaced by a graph/subgraph
structure in the case of graph matroids, or more generally, an orbit/suborbit structure for orbit
matroids.
Remark 2.18 — While in this paper we will focus on graph and bipartite graph matroids, it is
straightforward to apply the concept of orbit matroids in order to model other combinatorial objects:
Notation
Multiset
Directed Graph
Multigraph
d-hypergraph
(d-partite)
d-hypergraph
Ground Set E
[n]×N
[n] × [n]
[n]×N
2
[n]
d
:= [n]×d /S(d)
[n1 ] × · · · × [nd ]
Symmetry Group G
S(N ) permuting the N copies of [n].
S(n) acting diagonally on [n] × [n].
S(n)× S(N ), where S(N ) permutes the N copies
[n]
of 2 .
S(n) acting diagonally on [n]×d .
S(n1 ) × · · · × S(nd ) acting on E by the canonical
product action.
The types of orbit matroids above will however not be discussed further in the paper.
15
For better readability, we now provide some notation for matroids with (bipartite) graph symmetry:
Definition 2.19. Let M = ([m] × [n], I) be a matroid with bipartite graph symmetry, let S ⊆ [m] ×
[n]. The vertex support of S is the set
vsupp S := E1 × E2 ,
with E1 = {i : (i, j) ∈ S} and E2 = { j : (i, j) ∈ S}.
The signature of S is the pair of numbers (m0 , n0 ), where
m0 = #{i : (i, j) ∈ S}
n0 = #{ j : (i, j) ∈ S}.
Similarly, for a matroid M =
[n]
,I
2
with graph symmetry, and S ⊆
[n]
,
2
we define
vsupp S := {i : (i, j) ∈ S}.
The signature of S is the number # vsupp S.
Notice that for (bipartite) graph matroids, the signature is an invariant under the group action,
therefore an invariant of the (bipartite) graph S/G.
Notation 2.20. Let M = ([m]×[n], I) be a matroid with bipartite graph symmetry, let S ⊆ [m]×[n]
with signature (m, n). We will depict S as
• a mask: the mask corresponding to S is an (m × n)-matrix with entries in {◦, •}, where ◦ in
the (i, j) position indicates that (i, j) 6∈ S, while • indicates (i, j) ∈ S.
• a bipartite graph: the corresponding graph S/G. The unlabeled vertex set corresponding to
[m] will be depicted with red color, and the vertex set corresponding to [n] will be depicted
with blue color.
As an example, we depict the mask and the graph for the subset S = {(1, 1), (1, 2), (1, 3), (2, 2), (3, 3)}
in the matroid M = ([3] × [3], I) :
S, as a mask:
• • •
◦ • ◦
◦ ◦ •
S, as a bipartite graph:
The mask can be understood as adjacency matrices of the graph (with 1 = • and 0 = ◦ in the usual
convention), therefore we will use mask and graph interchangeably when it is clear that we are
referring to a graph.
Similarly, for matroids with graph symmetry, symmetric masks and (ordinary) graphs will be
used as depictions. In analogy, the vertex support is a subset of [n], and the signature is its cardinality.
For this, we introduce some notation:
Notation 2.21. Let M be a bipartite graph matroid (or a graph matroid) with symmetry group
G∼
= S(m) × S(n). Let S ⊆ E(M). Then, we will denote the group of S, which naturally acts on
vsupp S, and which is isomorphic to the stabilizer of S in G, by
S(S) := {σ ∈ G : σ(S) = S}.
We call S(S) the automorphism group of S.
16
Remark 2.22 — Note that S(S) does not depend on the choice of m or n nor on the matroid M only on the orbit symmetry that M exhibits, and on the combinatorics of S ⊆ vsupp S. Moreover,
S(S) coincides with the usual graph theoretical definition of automorphism group of the graph S/G.
Therefore, the isomorphy type of S(S) and in particular the cardinality #S(S) are invariants of the
graph S/G. Elementary group theory implies that it holds that #G = #S(S) · #(S/G), therefore if
the signature of S is (µ, ν), it must hold that #S(S) divides µ!ν!.
We would like to make a final remark that axiomatization of orbit matroids is possible without
indirectly defining them through matroids:
Remark 2.23 — We would like to mention that graph matroids (and similarly the other orbit matroids) could be axiomatically defined, without resorting to quotients. For the case of graph matroids, the concept has been already investigated under the name of matroidal family of graphs,
see [29] for a review. Namely, staying close to the axiomatization there, a graph matroid can be
defined as a collection C of (circuit) graphs, such that
(G0) No G ∈ C has isolated vertices.
(G1) If G ∈ C, and G0 ∈ C is a proper subgraph of G, then G0 = G.
(G2) Let G be a graph with subgraphs G1 , G2 ∈ C. Let e ∈ G1 ∩ G2 be an edge. Then, there is a
subgraph G0 of G \ {e} such that G0 ∈ C.
This definition corresponds to the later definition of limit graph matroids in section 4. Finite graph
matroids, as defined earlier, are obtained by replacing axioms G0 and G2:
(G0) No G ∈ C has isolated vertices, each G ∈ C has at most N vertices.
(G2) Let G be a graph having at most N vertices, with subgraphs G1 , G2 ∈ C. Let e ∈ G1 ∩ G2 be an
edge. Then, there is a subgraph G0 of G \ {e} such that G0 ∈ C.
We want to note that the derivation of matroidal family of graphs, while it can be regarded equivalent, is sightly different from that of a graph matroid by the fact that we have constructed the latter
as a quotient. Furthermore, completely analogous definitions can be obtained for bipartite graph
matroids or the other orbit matroids, replacing the word “graph” with the respective orbit structure.
For the particular case of bipartite graph matroids, the infinite variant is completely analogous to
the graph matroid, by replacing graphs by bipartite graphs, it corresponds to the two-sided limits
from section 4. Two finite variants are obtainable by either restricting the number of vertices in one
class, or in both classes, corresponding either to the one-sided limits of bipartite graph matroids
from section 4, or to the bipartite matroids with finite ground set defined earlier.
A very interesting and still widely open question is which collections - and in particular, infinite
collections - of circuit graphs give rise to graph matroids. However, since this is not a question
we directly pursue in the present work, we do not make the alternate definition above. Examples
for such families of (bipartite) graph matroids, however, and structure statements which can be
interpreted as restrictions, can be found in section 4 and later. The same question naturally arises for
general orbit matroids. Also, the relation of the realization spaces of matroids with orbit symmetries
and the orbit matroids would be interesting to investigate.
2.4. Symmetrization and Bipartition
Let G be a graph on n vertices with adjacency matrix (mask) M . If we interpret M as the mask of
a bipartite graph of signature (n, n), we obtain a bipartite graph G̃ (the “Kronecker double cover”)
that is fixed under the involution (i, j) 7→ ( j, i) that exchanges the parts of the bipartition of Kn,n .
On the other hand, if B is a bipartite graph of signature of (m, n), it has a natural embedding
17
as a subgraph B̃ of Km+n that has a block-diagonal adjacency matrix. In this section, we explore
how algebraic graph and bipartite graph matroids behave under these combinatorial lifting and
projection operations.
We start by considering how graph matroids induce bipartite graph matroids.
[m+n]
Definition 2.24. Let M = (E, I) be a graph matroid on
. Write S(m, n) := {(i, j) ∈ [m] ×
2
[n] : (i, j + m) ∈ S} for all S ∈ E. The (m, n)-bipartition M|m,n of M is the pair (E 0 , I0 ) with
E 0 = [m] × [n], and I0 = {S(m, n) : S ∈ I}.
[m+n]
Proposition 2.25. Let M be a graph matroid with ground set E(M) =
. Then the (m, n)2
bipartition M|m,n is a bipartite graph matroid.
Proof. The bipartition construction is an instance of matroid restriction. Namely, M|m,n , is isomorphic to the restriction of M to the set {(i, j) : 1 ≤ i ≤ m, m + 1 ≤ j ≤ m + n} ⊆ E(M) and therefore
a matroid, see [27, section 1.3].
Proposition 2.25 shows well-definedness of the bipartition, as a matroid, and implies therefore:
[m+n]
Corollary 2.26. Let M be a graph matroid with ground set E =
, let S ⊆ E. Let M|m,n be the
2
0
bipartition of M, with ground set [m] × [n]. Write E := [m] × {m + 1, . . . m + n}. Then, the following
statements hold:
(i) If S is independent in M, then S(m, n) is independent in M|m,n .
(ii) If S ⊆ E 0 , then S(m, n) is independent in M if and only if S is independent in M|m,n .
(iii) If S ⊆ E 0 , then S(m, n) is a circuit of M if and only if S is a circuit of M|m,n .
While bipartition of a graph matroid can be described in a purely combinatorial manner, as
above, it also entails bipartitions for the different kinds of realizations, given the matroid is both a
graph matroid, and algebraic.
[m+n]
Proposition 2.27. Let M be a graph matroid with E(M) =
, with cryptomorphic realizations
2
M = AK (α) = B(X, b) = C(P, x).
Denote E|m,n = {(i, j) : 1 ≤ i ≤ m, m+1 ≤ j ≤ m+n} ⊆ E(M). Then, the following are cryptomorphic
realizations for M|m,n :
(a) (K, α0 ) with α0 α E|m,n
(b) (X0 , b0 ) with b0 = b E|m,n and X0 = πb0 (X).
(c) (P0 , x0 ) with x0 = x E|m,n and P0 = K[x0 ] ∩ P.
Proof. It holds that M|m,n = AK (α0 ) by substituting the two definitions. Consequently, by applying
the definitions of basis and coordinate matroids, one obtains AK (α0 ) = B(X0 , b0 ) once it is established that X0 is irreducible, and P0 is prime. Irreducibility of X0 follows from irreducibility of X,
being the image of X under the algebraic map πb0 , and primeness óf P0 follows from that of P,
being a coordinate section.
Proposition 2.27 implies that the cryptomorphic realizations of M descend to cryptomorphic
realizations of M|m,n .
Symmetrizing bipartite graph matroids, in order to obtain back a graph matroid, is a more
delicate operation that de-symmetrizing graph matroids by bipartitioning them. In particular, there
is not a natural combinatorial construction that is guaranteed to produce a graph matroid (though
it is not hard to produce a polymatroid). Thus, we will start directly with an algebraic construction.
18
Definition 2.28. Let M be an algebraic bipartite graph matroid on the ground set E = [n] × [n],
and let M = B(X, b) be a basis realization of M. Let H ⊂ K E be the linear space
H := {x ∈ K E : x i j = x ji for all (i, j) ∈ E}
and define X := X ∩ H . If Y is irreducible, we define the algebraic symmetrization of M to be the
matroid M := BK X , b , where b := ({(i, j) : i ≤ j}). (If X is not irreducible, the algebraic
symmetrization is undefined.)
Proposition 2.29. Let M be an algebraic bipartite graph matroid of E = [n] × [n] with algebraic
symmetrization M (assumed to exist). Then M is a well-defined algebraic graph matroid on the
[n]
ground set 2 .
Proof. By hypothesis, the variety X is irreducible, so the matroid M will be well-defined once we
check that H σ is minimally spanned by b0 := b({(i, j) : i ≤ j}). This also follows by construction.
Now we check that M is a graph matroid. Consider the diagonal subgroup
G = {(σ, τ) ∈ S(n) × S(n) : σ = τ} ⊆ S(n) × S(n).
Note that G stabilizes all sets of the form {(i, j), ( j, i)}, and induces an action on E = [n] × [n].
Since H σ is the set fixed by this same action on b, we see that b0 , and then M inherits S(n) =
[n]
S(n) × S(n)/G as symmetry group, and 2 = [n] × [n]/G as ground set.
Remark 2.30 — The algebraic symmetrization is different than a construction introduced by Lovász
[23] which builds a matroid N on the rank 2 flats of a matroid M by picking a generic element from
each flat.
Remark 2.31 — The cryptomorphisms of Theorem 2.2 allow us to reformulate the algebraic symmetrization in all three ¬languages. In terms of coordinate
matroids, the irreducibility of Y corre¶
sponds to the ideal P ∩ x i j − x ji : (i, j) ∈ [n] × [n] being primary.
Whether the algebraic symmetrization exists is an interesting question. Since the variety H
is irreducible, the algebraic symmetrization will always exist for generic varieties X; that is, for
an open dense set in the fixed-function Hilbert variety. However, this implies nothing for fixed
varieties of interest, and algebraic, or structural criteria for existence of the symmetrization would
be desirable.
3. Determinantal and Rigidity Matroids
Definition 2.17 is motivated primarily by two examples with graph symmetry: determinantal matroids, which control the algebraic aspects of low-rank matrix completion, and rigidity matroids
from geometry. In this section, we will introduce these matroids in detail. We will assume that K is
algebraically closed.
3.1. The Determinantal Matroid
The determinantal matroid, which we will denote by D(m × n, r), is the (purely combinatorial)
matroid describing dependence relations of the entries of a rank r matrix of rank (m × n). We give
three realizations for the matroid, their equivalence follows from Proposition 2.2; for notations,
recall Definition 2.6. We will assume that K is a field of characteristic zero.
Definition 3.1. We define the determinantal matroid, to be the matroid D(m × n, r) induced equivalently by any of the following realizations:
19
(a) As an algebraic matroid: let Uik , 1 ≤ i ≤ m, 1 ≤ k ≤ r be a collection of doubly indexed
transcendentals over K. Let Vjk , 1 ≤ j ≤ n, 1 ≤ k ≤ r be another such collection. Let α =
Pr
{X i j = k=1 Uik Vjk : 1 ≤ i ≤ m, 1 ≤ j ≤ n}. We define D(m × n, r) := AK (α) as the algebraic
matroid of the α over K.
(b) As a basis matroid: let M(m × n, r) = {A ∈ K m×n : rank A ≤ r} be the determinantal variety.
Let b = {B (i j) : 1 ≤ i ≤ m, 1 ≤ j ≤ n} be the set of standard unit matrices B (i j) = ei · ẽ>
j ,
where ei , 1 ≤ i ≤ m is the standard orthonormal basis of K m , and ẽ j , 1 ≤ j ≤ n is the standard
orthonormal basis of K n . We define D(m × n, r) := B(M(m × n, r), b) as the basis matroid of
M(m × n, r).
(c) As a coordinate matroid: let x = {X i j : 1 ≤ i ≤ m, 1 ≤ j ≤ n} be a set of doubly indexed
coordinates. Let I(m × n, r) ⊂ K[x] be the determinantal ideal, that is, the ideal in K[x]
generated by the (r+1)×(r+1) minors of the (m×n) matrix whose entries are the coordinates
X i j . We define D(m × n, r) := C(I(m × n, r), x) as the coordinate matroid of I(m × n, r).
(d) As a linear matroid: let Uik , Vjk , as in (a), and L = K(Uik , Vjk ). Let {uik , 1 ≤ i ≤ m, 1 ≤ k ≤
Pr
r} ∪ {vik , 1 ≤ j ≤ n, 1 ≤ k ≤ r} be any basis of K r(m+n) . Let v = {vi j = k=1 Uik v jk + uik Vjk :
1 ≤ i ≤ m, 1 ≤ j ≤ n} be a collection of vectors in L r(m+n) . We define D(m × n, r) := L L (v) as
the linear matroid of the v.
In particular, the pairs (K, α), (M(m × n, r), b) and (I(m × n, r), x) are cryptomorphic.
Proof of well-definedness: For well-definedness of (b), it needs to be checked that M(m × n, r)
is irreducible. For (c), it needs to be checked that I(m × n, r) is prime. These are known facts:
irreducibility of M(m × n, r) can be found in [6, Proposition 1.1 (a)], or alternatively follows from
Remark 3.2 below; primeness of I(m × n, r) follows from [6, Theorem 2.10, Remark 2.12, and
Corollary 5.17f].
Proof of equivalence of the definitions: (a) ⇔ (b) follows from the observation that the X i j
form a rank r matrix by definition. Therefore, the field K(x) is isomorphic to the fraction field of
M(m × n, r) in accordance with the canonical isomorphism exposed in Proposition 2.2. (b) ⇔ (c):
This is a direct consequence of Proposition 2.2. In addition, one has to note that: I(m × n, r) is a
prime ideal; M(m × n, r) = V(I(m × n, r)), therefore it is also irreducible. (a) ⇔ (d): This follows
from Proposition 2.14, identifying uik with dUik and v jk with dVjk , and noting that these generate
the r(m + n)-dimensional K-vector space Ω L/K .
Remark 3.2 — Some basic facts about the determinantal matroid:
(i) If r ≤ min(m, n), then dim M(m × n, r) = r · (m + n − r), see [6, section 1.C, Proposition 1.1].
Therefore, by Proposition 2.15 the rank of the determinantal matroid is as well rk D(m ×
n, r) = r · (m + n − r).
(ii) I(m × n, r) is a toric ideal if and only if r = 1 or r ≥ min(m, n).
(iii) There is a canonical surjective algebraic map
Υ :Cm×r × Cn×r −→ M(m × n, r)
(U, V ) 7→ U V >
Therefore, M(m × n, r) is irreducible.
20
Example 3.3. We will explicitly describe the determinantal matroid D(4 × 4, 2).
1. The prime ideal I(4 × 4, 2) is generated by the sixteen (3 × 3)-minors of a (4 × 4)-matrix of
formal variables.
2. The rank of D(4 × 4, 2) is r(m + n − r) = 12.
3. The circuit graphs are isomorphic to one of the following masks (with corresponding bipartite
graphs,
dotted lines
indicating excluded edges):
· · · ·
· · · ·
· · · ·
· · · ·
· · · ·
· · · ·
· · · ·
· · · ·
4. Bases of D(4 × 4, 2) are all sets of size 12 not containing the complete bipartite graph K3,3 .
There are nine non-isomorphic bipartite basis graphs (seven if we consider transpose symmetry).
The use of colored squares to define the mask is motivated by an observation: all our circuit
graphs can be constructed by sequential cancellation of minors.
3.2. The Symmetric Determinantal Matroid
The symmetric determinantal matroid D (n× n, r) is the analogue of the determinantal matroid for
symmetric (n×n) matrices. The determinantal matroid has an intrinsic Z/2-symmetry which comes
from the fact that rank is invariant under transposition of a matrix, therefore being a potential target
of symmetrization. We will show that this symmetrization is well-defined.
Definition 3.4. We define the symmetric determinantal matroid, to be the matroid D (n × n, r)
induced equivalently by any of the following realizations:
( ) As a symmetrization: define D (n × n, r) to be the symmetrization of D(n × n, r), in the sense
of Definition 2.28.
(a) As an algebraic matroid: P
let Vik , 1 ≤ i ≤ n, 1 ≤ k ≤ r be a set of doubly indexed transcendentals
r
over K. Let α = {X i j = k=1 Vik Vjk : 1 ≤ i ≤ j ≤ n}. We define D (n × n, r) := AK (α) as the
algebraic matroid of the α over K.
n×n
>
(b) As a basis matroid: let M (n ×
: rank A ≤ r,
¦ 1n, r) = {A ∈ K
© A = A} be the symmetric determinantal variety. Let b = 2 B (i j) + B ( ji) : 1 ≤ i ≤ j ≤ n , where the B (i j) are standard
n
unit matrices B (i j) = ei · e>
j , where ei , 1 ≤ i ≤ n is the standard orthonormal basis of K . We
define D (n × n, r) := B(M (n × n, r), b) as the basis matroid of M (n × n, r).
(c) As a coordinate matroid: let x = {X i j : 1 ≤ i ≤ j ≤ n} be a set of doubly indexed coordinates. Let I (n × n, r) ⊂ K[x] be the symmetric determinantal ideal, that is, the ideal in K[x]
generated by the (r + 1) × (r + 1) minors of the symmetric (n × n) matrix whose entries are
the coordinates X i j . We define D (n × n, r) := C(I (n × n, r), x) as the coordinate matroid of
I (n × n, r).
(d) As a linear matroid: let Vjk , as in (a), and L = K(Vjk ). Let {vik , 1 ≤ j ≤ n, 1 ≤ k ≤ r} be any
Pr
basis of K r n . Let v = {vi j = k=1 Vik v jk + vik Vjk : 1 ≤ i ≤ j ≤ n} be a collection of vectors in
L r n . We define D (n × n, r) := LK (v) as the linear matroid of the v.
In particular, the pairs (K, α), (M (n × n, r), b) and (I(n × n, r), x) are cryptomorphic.
21
Proof of well-definedness: For (b), we need to check that M (n × n, r) is irreducible; for (c),
we must check that I(m × n, r) is prime. Irreducibility of M (n × n, r) can be demonstrated by
the existence of the map in Remark 3.5. Primeness of I(m × n, r) follows from [19, Corollary to
Theorem 1].
Proof of equivalence of the definitions: Since M (n × n, r) is irreducible, the symmetrization
is by definition equivalent to the basis matroid of (b). The remaining equivalences follow from
Proposition 2.2 and Proposition 2.13 as in the asymmetric case.
Remark 3.5 — Some basic facts about the symmetric determinantal matroid:
(i) If r ≤ n, then dim M (n × n, r) = r · n − 12 r(r − 1) , see Therefore, by Proposition 2.15 the
rank of the symmetric determinantal matroid is as well rk D (n × n, r) = r · n − 21 r(r − 1) .
(ii) There is a canonical surjective algebraic map
Υ :Cn×r −→ M (n × n, r)
V 7→ V V >
Therefore, M (n × n, r) is irreducible.
Example 3.6. We will explicitly describe the symmetric determinantal matroid D (4 × 4, 2).
1. The prime ideal I (4 × 4, 2) is generated by the ten (3 × 3)-minors of a symmetric (4 × 4)matrix of formal variables.
2. The rank of D (4 × 4, 2) is r · n − 21 r(r − 1) = 7.
3. The circuit graphs are isomorphic to one of the following graphs labeled with top-degree, and
depicted with a corresponding bipartite mask: Again, we include colored squares to indicate
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
the minors generating each circuit polynomial.
4. Bases of D(4 × 4, 2) are the sets of size 7 not containing the rightmost circuit above. There
are nine such basis graphs.
3.3. The Rigidity Matroid
The other family of examples we consider are rigidity matroids, which capture the degrees of freedom in bar-joint frameworks. These are, intuitively, mechanical structures made of fixed-length bars,
connected by joints. It is assumed that the allowed continuous motions of the joints preserve the
22
lengths and connectivity of the bars. A framework is called (locally) rigid if all such continuous
motions extend to Euclidean motions of the ambient d-dimensional space.
In rigidity theory, the position of the joints are usually modelled by points p1 , . . . , pn ∈ R r , and
the lengths of the bars are given by the pairwise Euclidean distances kpi − p j k2 . If the number of
joints n is large enough, not all of these distances can be chosen independently; the rigidity matroid
models the dependence relations between these distances. In order to transform this scenario into
an algebraic problem, we replace the Euclidean distances by the pairwise squared distances
di j = kpi − p j k22 = pi> pi − 2pi> p j + p>
j pj,
and consider this as an equation over an arbitrary field K of characteristic zero. The corresponding
algebraic matroid is then straightforwardly obtained as follows:
Definition 3.7. We define the rigidity matroid, or Cayley-Menger matroid, to be the matroid CM (n×
n, r) induced by the following algebraic realization:
Let Pik : 1 ≤ i ≤ n, 1 ≤ k ≤ r be a set of doubly indexed transcendentals over K. Let α = {Di j =
2
k=1 (Pik − P jk ) : 1 ≤ i ≤ j ≤ n}. We define CM (n × n, r) := AK (α) as the algebraic matroid of
the α over K.
Pr
The equivalent realizations as basis and coordinate matroid whose existence is guaranteed by
Theorem 2.2 is slightly less straightforward, and considerably less well studied than in the case of
the determinantal matroids. Therefore, before stating the cryptomorphisms explicitly, we define the
corresponding objects and collect some useful results on those.
The first relevant object is the variety for the basis representation:
Definition 3.8. The Cayley-Menger variety is the variety of
CM (n × n, r) = {d ∈ K n×n : di j = pi> pi − 2pi> p j + p>j p j for some pi ∈ C r , 1 ≤ i
j ≤ n}.
Remark 3.9 — Some basic facts about the Cayley-Menger variety:
(i) The dimension of the Cayley-Menger variety is
dim CM (n × n, r) = r n −
r +1
2
.
This follows from [4, Theorem 4.3].
(ii) By definition of CM (n×n, r), there is a canonical surjective algebraic map, the so-called length
map,
` :Cn×r −→ CM (n × n, r)
r
X
(pik − p jk )2 .
p 7→ d, where di j =
k=1
In particular, CM (n × n, r) is an irreducible variety.
The second object is the ideal for the coordinate representation:
23
Definition 3.10. Let x = {X i j , 1 ≤ i
j ≤ n} be a collection of formal coordinates. The CayleyMenger matrix associated to the X i j is the matrix
0 1
1
···
1
1 0 X
···
X 1n
12
..
.
.
.
.
CM(n) :=
1 X 12 0
.
..
..
..
..
.
.
.
.
X (n−1)n
1 X 1n · · · X (n−1)n
0
The Cayley-Menger ideal is the ideal C (n, r) ⊂ K[x] generated by all the (r + 3) × (r + 3)-minors
of the matrix CM(n).
Since it apperas to be unknown what the ideal of the Cayley-Menger variety is, we briefly prove
that it coincides with the Cayley-Menger ideal.
Theorem 3.1. The Cayley-Menger ideal is the prime ideal corresponding to the Cayley-Menger variety,
i.e.,
C (n, r) = I(CM (n × n, r))
Proof. Results of Menger, which can be found as [10, Proposition 6.2.11] and [10, Theorem 6.2.13],
imply that variables X i j are realizable as the set of pairwise distances between points in R r if and
only if the matrix CM(n) has rank at most r +2, and a certain additional semi-definiteness condition
holds. In particular, this shows that CM (n × n, r) is the algebraic
closure of the set of squaredp
distance matrices of r-dimensional point configurations, so C (n, r) is the vanishing ideal of
CM (n × n, r).
In [10, Equation 6.2.9], the Cayley-Menger determinant is identified with the determinant of
a symmetric matrix with linear entries; therefore, the Cayley-Menger ideal is a generic symmetric
determinantal ideal under a change of coordinates. Therefore, by the same result in [19, Corollary
to Theorem 1], C (n, r) is prime, and therefore radical. By the Nullstellensatz, it is therefore
identical to I(CM (n × n, r)).
With these observations, Theorem 2.2 and Proposition 2.14 immediately yields:
Proposition 3.11. The following are equivalent realizations of CM (n × n, r):
(a) The algebraic matroid AK (α), as defined in 3.7.
(b) The basis matroid B(CM (n × n, r), b) of the Cayley-Menger variety (Definition 3.8), where b =
{B (i j) : 1 ≤ i j ≤ n} is the set of standard unit matrices B (i j) = ei · ẽ>
j , and ei , 1 ≤ i ≤ n is the
standard orthonormal basis of K n
(c) The coordinate matroid C(C (n × n, r), x), keeping the notation of Definition 3.10.
(d) The linear matroid associated to the differential of the length map ` given in Remark 3.9: let P jk ,
as in Definition 3.7, and L = K(P jk ). Let {dik , 1 ≤ j ≤ n, 1 ≤ k ≤ r} be any basis of K r n . Let
Pr
v = {vi j = k=1 (pik − p jk )(Pik − P jk ) : 1 ≤ i ≤ j ≤ n} be a collection of vectors in L r n . The
linear matroid in question is LK (v).
In particular, the pairs (K, α), (CM (n × n, r), b) and (C (n × n, r), x) are cryptomorphic.
Informally, Proposition 3.11 says that three definitions of the rigidity matroid are equivalent: (a)
via the intrinsic property of the measured lengths being generically realizable in dimension d; (b)
via the geometry of the Cayley-Menger variety; (c) via the extrinsic relations between the lengths;
(d) in terms of the invertibility of a specific map, which can be analyzed via its differential (the
“rigidity matrix” of combinatorial rigidity theory).
24
Example 3.12. We describe the rigidity matroid CM (5 × 5, 2).
1. The prime ideal C (5, 2) is generated by the 36 (5 × 5)-minors of the (6 × 6) Cayley-Menger
matrix.
r+1
2. The rank of CM (5 × 5, 2) is r n − 2 = 7.
3. CM (5 × 5, 2) has exactly two circuit graphs:
4. CM (5 × 5, 2) has three basis graphs:
3.4. The Bipartite Rigidity Matroid
The bipartite rigidity matroid is the non-symmetric version of the rigidity matroid; it can be interpreted as the matroid describing (squares of) distances kpi − q j k2 , where p1 , . . . , pm , q1 , . . . , qn ∈ R r
are two distinct sets of joints. In order to be as concise as possible, and for convenience of the
reader, we define the bipartite rigidity matroid as bipartition of the rigidity matroid.
Definition 3.13. The bipartite rigidity matroid, or the bipartite Cayley-Menger matroid, denoted by
CM(m × n, r) is defined as the bipartition of the rigidity matroid CM (m + n × m + n, r).
Definition 3.14. The bipartite Cayley-Menger variety is the variety of
CM(m × n, r) = {d ∈ K m×n : di j = pi> pi − 2pi> q j + q>j q j for some pi , q j ∈ C r , 1 ≤ i ≤ m, 1 ≤ j ≤ n}.
The second object is the ideal for the coordinate representation:
Definition 3.15. Let x = {X i j , 1 ≤ i ≤ m, 1 ≤ j ≤ n} be a collection of formal coordinates. The
Cayley-Menger matrix associated to the X i j is the matrix
0 1
1 X
11
CM(m × n) := 1 X 21
..
..
.
.
1 X n1
1
X 12
X 22
..
.
···
···
···
..
.
..
.
X n(n−1)
1
X 1n
..
.
.
X (n−1)n
X nn
The bipartite Cayley-Menger ideal is the ideal C(m×n, r) ⊂ K[x] generated by all the (r +3)×(r +3)minors of the matrix CM(m × n).
Again, we need to prove what the prime ideal for the bipartite Cayley-Menger variety is:
Theorem 3.2. The Cayley-Menger ideal is the prime ideal corresponding to the Cayley-Menger variety,
i.e.,
C(m × n, r) = I(CM(m × n, r))
25
Proof. Since C(m × n, r) is a section of the determinantal ideal I(m × n, r + 2) it holds that V(C(m ×
n, r)) = CM(m × n, r). By [19, Corollary to Theorem 1], C(n, r) is prime, and therefore identical to
I(CM(m × n, r)) by the Nullstellensatz. Thm:CMIsym
As a direct consequence of Proposition 2.27, we obtain the cryptomorphic realizations of CM(m×
n, r):
Corollary 3.16. The bipartite rigidity matroid CM(m × n, r) is, as a matroid, isomorphic to the following realizations:
(a) The algebraic matroid AK (α), defined as follows: let Pik : 1 ≤ i ≤ m, 1 ≤ k ≤ r be a set of
doubly indexed transcendentals over K. Let Q jk : 1 ≤ j ≤ n, 1 ≤ k ≤ r be another such set. Let
Pr
α = {Di j = k=1 (Pik − Q jk )2 : 1 ≤ i ≤ m, 1 ≤ j ≤ n}.
(b) The basis matroid B(CM(n × n, r), b) of the Cayley-Menger variety (Definition 3.14), where b =
{B (i j) : 1 ≤ i ≤ m, 1 ≤ j ≤ n} is the set of standard unit matrices B (i j) = ei · ẽ>
j , with
ei , 1 ≤ i ≤ m being the standard orthonormal basis of K m , and ẽ j , 1 ≤ j ≤ n the standard
orthonormal basis of K n .
(c) The coordinate matroid C(C (n × n, r), x), keeping the notation of Definition 3.15.
(d) The linear matroid LK (v), defined as follows: let P jk , Q ik as in (A), and L = K(P jk , Q ik ). Let
{pik , 1 ≤ j ≤ m, 1 ≤Pk ≤ r} be any basis of K r m , and {qik , 1 ≤ j ≤ n, 1 ≤ k ≤ r} be any basis of
r
K r n . Let v = {vi j = k=1 (pik − q jk )(Pik −Q jk ) : 1 ≤ i ≤ m, 1 ≤ j ≤ n} be a collection of vectors
in L r(m+n) .
In particular, the pairs (K, α), (CM(m × n, r), b) and (C(m × n, r), x) are cryptomorphic. Note that
α, b, x depend on m, n, and r, while b, x depend only on m, n.
Remark 3.17 — Some basic facts about the bipartite rigidity matroid:
(i) The dimension of the bipartite Cayley-Menger variety is
dim CM(m × n, r) = r · (m + n) −
1
1
r(r + 1) + (m + n − k) − ma x(0, r(r + 3) + 1 − k),
2
2
where k = m · 1(m ≥ r + 1) + n · 1(n ≥ r + 1), and 1(x) is the indicator function, being 1
when x is true and 0 if x is false. This follows from [2, Theorem 10].
(ii) In complete analogy to the case of the rigidity matroid, there is a canonical surjective algebraic
map, the so-called length map,
` :Cm×r × Cn×r −→ CM(m × n, r)
r
X
(pik − q jk )2 .
(p, q) 7→ d, where di j =
k=1
In particular, CM(m × n, r) is an irreducible variety.
Example 3.18. We describe the bipartite rigidity matroid CM(4 × 4, 2).
1. The prime ideal C(4×4, 2) is generated by the seventy 5×5-minors of the 9×9 Cayley-Menger
matrix whose only diagonal entry is CM(8)1,1 .
26
2. Because m = n = 4 ≥ 3, the rank of CM(4 × 4, 2) reduces to
r +1
1
r(m + n) −
− max(0, r(r + 3) + 1 − m − n)) = 13.
2
2
3. A subset of [4] × [4] of size 13 is a basis of CM(4 × 4, 2) if and only if the corresponding
basis graph does not contain the complete bipartite graph K3,4 . There are three such bipartite
graphs, represented by the following masks:
•
•
•
◦
•
•
◦
•
•
◦
•
•
•
•
•
•
•
•
•
◦
•
•
•
◦
•
•
◦
•
•
•
•
•
•
•
◦
◦
4. The circuit graphs are isomorphic to one of the
masks:
following
· · · ·
· · · ·
· · · ·
· ·
· ·
· ·
· ·
·
·
·
·
•
◦
•
•
· ·
· ·
· ·
· ·
•
•
•
•
·
·
·
·
•
•
•
•
·
·
·
·
4. Limit Matroids and their Rank Function
4.1. Direct Limits of Matroids
One basic observation is that a circuit graph C of signature (k, `) of D(m×n, r) is also a circuit graph
of D(m0 × n0 , r), provided that m0 ≥ k and n0 ≥ `. In this section, we show how to construct limits
of graph matroid sequences that allows us to speak of a limiting determinantal matroid D(N × N, r),
or, in general, of limits of matroids with intrinsic symmetries.
We start with a general construction of direct limit for matroids.
Definition 4.1. A collection of matroids Mν = (Eν , Iν ), ν ∈ N, such that Eν ⊆ Eν+1 , and Iν =
{S ∩ Eν : S ⊆ Iν+1 } for all n ∈ N, is called an injective sequence of matroids, and will be denoted
by (Mν ). If furthermore the matroids Mν come with groups of automorphisms Gν , we call (Mν )
compatible with the Gν -action if Gν ⊆ Gν+1 .
We will now introduce what a direct limit of such an injective sequence should mean, but we
will not develop the theory of direct limit matroids in detail. Instead, we will just state by definition
what such a direct limit should be.
Definition 4.2. Let (Mν ), ν ∈ N be an injective sequence of matroids. We define the direct limit
matroid to be the the pair
M := lim Mν := (E(M), I(M)) ,
−→
ν∈N
for which we define
S
E(M) := Sν∈N E(Mν )
C(M) := ν∈N C(Mν )
D(M) := {S ⊆ E(M) : ∃C ∈ C(M) s.t. C ⊆ S}
I(M) := P(E(M)) \ D(M)
Note that we have refrained from defining a set of bases for M; this is possible in different ways,
but we avoid it since we will not need to make use of bases of any kind for M. We also want to
note that, while it is possible to define different kinds of direct limits for matroids, in this paper, the
the ground set
the set of circuits
the set of dependent sets
the set of independent sets
27
direct limits will always come from injective sequences. Furthermore, by the recent results of [5],
the direct limit matroids defined as above lie in the class of “finitary” type infinite matroids. Before
proceeding, we explicitly describe some functorial properties of the direct limit inherited from the
injective sequence:
M
Remark 4.3 — Let (Mν ) with Eν = E(Mν ) be an injective sequence of matroids, let M = lim
−→ν∈N ν
its direct limit, having ground set E. By definiton, and elementary matroid properties, it holds that:
(i) S ⊆ E is a circuit in M if and only if S ∩ Eν is a circuit in Mν for all ν ∈ N such that S ⊆ Eν
(ii) S ⊆ E is dependent in M if and only if S ∩ Eν is dependent in Mν for some ν ∈ N.
(iii) S ⊆ E is independent in M if and only if S ∩ Eν is independent in Mν for all ν ∈ N.
By the injectivity of the limit sequence, the rank function extends as well to the direct limit
matroid:
M be a direct limit matroid, with ground sets E = E(M), Eν = E(Mν ).
Lemma 4.4. Let M = lim
−→ν∈N ν
Then there is a unique function rkM : P∗ (E) → N such that
rkM (S) = rkMν (S)
for all S ⊆ Eν and ν ∈ N, where ∗ (E) is the set of finite subsets of E.
Proof. This follows directly from Remark 4.3 and any characterization of the rank function in terms
of dependent/independent sets, see e.g. [27, Theorem 1.3.2].
By the injectivity of the limit sequence, the group action extends as well:
Lemma 4.5. Let M = lim
M be a direct limit matroid, compatible with a sequence of groups of
−→ν∈N ν
automorphisms Gν . Let G = lim
G be the corresponding direct limit of groups. Then, G is a group
−→ν∈N ν
of automorphisms of M in the sense that G acts on E(M), and
σI(M) = I(M), σD(M) = D(M), σC(M) = C(M)
for all σ ∈ G.
Proof. This follows directly from Remark 4.3, the fact that for each σ ∈ G, there is n such that
σ ∈ Gν , and the properties of automorphisms.
The notion of Lemmata 4.4 and 4.5 allows to talk about orbits in Mν /Gν and M/G. Finally, we
would like to remark that the limit construction in Definitions 4.1 and 4.2 can be iterated without
problems a countable number of times. That is, one can take Mν in both definitions to be direct
limit matroids, obtaining a new limit matroid.
4.2. One-sided Bipartite Graph Matroid Limits
For graph and bipartite graph matroids, we will want to impose some additional structure. For now,
we will concentrate on bipartite graph matroids, deferring the case of graph matroids to Section
4.8.
Definition 4.6. Let (Mν ) be an injective sequence of bipartite graph matroids, such that Mν has
ground set E(Mν ) = [m] × [ν]; the sequence is compatible with the graph symmetry. The sequence
(Mν ) is called a one-sided bipartite graph matroid sequence, and the limit M = lim
M a bipartite
−→ν∈N ν
graph matroid limit, or the limit of (Mν ).
28
Note that, as defined above, M has ground set [m] × N. When dealing with one-sided limits,
we will reserve and fix m, the number of rows in the ground set.
Remark 4.7 — The ground sets Eν can be interpreted as the edges of complete bipartite graphs
Km,ν of a one-sided bipartite graph matroid sequence (Mν /Gν ). The sequence adds one vertex at a
time, always in the same part of the bipartition (i.e., with the same color) which we call the growing
part. In the limit, one obtains the graph Km,∞ in the graph matroid M/G. From now on, we will
switch freely between the formalisms.
Lemma 4.8. Let (Mν ), ν ∈ N be a one-sided bipartite graph matroid sequence with limit matroid
M. Consider the automorphism groups Gν = S(m) × S(ν) acting on Mν . Then, the group G =
S(m) × S(N) acts canonically on M and its ground set [m] × N. Moreover, for a graph G, it holds:
(i) G is a circuit graph in M/G if and only if G is a circuit graph in Mν /Gν for all ν ∈ N such that
G ⊆ Km,ν .
(ii) G is a dependent graph in M/G if and only if G is a dependent graph in Mν /Gν for all ν ∈ N
such that G ⊆ Km,ν .
(iii) G is an independent graph in M/G if and only if G ∩ Km,ν is an independent graph in Mν /Gν for
all ν ∈ N, where G ∩ Km,ν is any edge intersection under an arbitrary (disjoint) identification of
vertices with same color.
(iv) For a graph G in M/G, it holds that rkM (G) = rkMν (G) for all ν ∈ N such that G ⊆ Km,ν .
Proof. The first part follows from the fact that the permutation group S(N) is the injective limit
of the finite permutation groups S(ν), therefore S(m) × S(N) is the injective limit of the Gν =
S(m) × S(ν), and the action on the ground sets Eν = [m] × [ν] of Mν is compatible with the
group inclusion Gν ⊆ Gν+1 . The statements on the graphs G are then direct consequences of
Lemma 4.5.
Lemma 4.8 allows us to consider circuit, dependent graphs, independent graphs, and ranks of
graphs in terms of only the limit of a one-sided sequence of bipartite graph matroids.
To illustrate the concepts in this section, we will use two running examples.
Definition 4.9. Let m ∈ N and r ∈ N be fixed. The injective sequence of determinantal matroids
D(m × ν, r) forms a one-sided bipartite graph matroid sequence in the running parameter ν. We
denote its limit as
D(m × N, r) := lim D(m × ν, r).
−→
ν∈N
Similarly, the injective sequence of bipartite rigidity matroids CM(m × ν, r) forms a one-sided bipartite graph matroid sequence. We denote its limit as
CM(m × N, r) := lim CM(m × ν, r).
−→
ν∈N
The realizations, and their cryptomorphisms in Theorem 2.2 naturally extend to cryptomorphisms of the direct limits.
Proposition 4.10. Let (Mν ), ν ∈ N be a one-sided bipartite graph matroid sequence with limit matroid
M, such that Mν is algebraic, realized by cryptomorphic pairs (a) (K, αν ), (b) (Xν , bν ), (c) (Iν , xν ),
as described in Definition 3.1. Then, the following pairs are cryptomorphic:
S∞
(a) (α, K), where α = ν=1 αν
29
(b) (X, b), where X is the inverse limit ofSthe projection morphisms Xν → X0ν for ν 0 ≤ ν which project
∞
to the the span of the b0ν , and b = ν=1 bν
¶
¬S∞
S∞
⊆ K[x], with K[x] being the polynomial ring in
(c) (I, x), where x = ν=1 xν , and I =
I
ν
n=1
the infinitely many variables x
That is, the matroids4 AK (α), B(X, b) and C(I, x) are canonically isomorphic, and furthermore, they
are isomorphic to the matroid M as given in Definition 4.9.
Proof. The statement follows from noting that taking n as countable infinity in Definition 2.6 does
not affect well-definedness of the respective matroids, nor the correctness of the proof of Theorem 2.2.
Note that the variety X in Proposition 4.10 (b) is an irreducible sub-variety of the span of b, as
an inverse limit of irreducible varieties; furthermore, the ideal I in Proposition 4.10 (c) is prime, as
a direct limit of prime ideals.
4.3. Relative Rank, Average Rank and Realization Size
We introduce some tools for analyzing the asymptotic growth of the rank along a directed system
of matroids.
Definition 4.11. Let M be a matroid, with ground set E and rank function rk. The relative rank
function is the function
rk(·|·) :E × E → N
(T, S) 7→ rk(T ∪ S) − rk(S),
The number rk(T |S) is called the relative rank of T , given S.
Proposition 4.12. For any matroid M on a finite ground set E, the relative rank function satisfies:
(i) #T ≥ rk(T |S) ≥ 0 for all T, S ⊆ E
(ii) rk(A|S) ≥ rk(A|T ) for any A ⊆ E and S ⊆ T ⊆ E.
(iii) rk(A|S ∪ B) ≤ rk(A|S) for any A, B, S ⊆ E
(iv) rk(σT |σS) = rk(T |S) for any S, T ⊆ E, and any automorphism σ of M
(v) The set S ∪ T contains a basis with rk(T |S) elements from T for all T, S ⊆ E.
Proof. (iv) follows from the fact that rk(σS) = rk(S) for any S, then substituting the definition of
the relative rank.
(v) follows from the fact that any independent set can be extended to a basis, and that the rank
of a set is the size of a basis.
All the other statements follow from submodularity of the rank function, see e.g. [27, Corollary1.3.4 (R3)]. Namely, an equivalent characterization of rk being submodular is the statement
rk(S ∪ A) − rk(S) ≥ rk(T ∪ A) − rk(T ) for all S ⊂ T ⊂ E, A ⊂ E
The inequality (ii) then follows from inserting the definition of the relative rank; and (i) follows
from taking once T = ∅, once S = ∅; (iii) from taking T = S ∪ B in (ii).
4
as in Definition 2.6 by taking n as countable infinity
30
Definition 4.13. Let M be a one-sided bipartite graph matroid limit, on the ground set E(M) =
[m] × N. We define the growth function of the sequence (Mν ) as
δrkM (·) : N → N, ν 7→ rk Km,ν+1 |Km,ν .
If clear from the context, we will just write δrk(·), omitting the dependence on M.
Note that due to Lemma 4.4, the relative rank can be taken in any Mk for k ≥ ν + 1. For a
one-sided bipartite graph matroid sequence, the growth function is monotone:
Proposition 4.14. Let M be a one-sided bipartite graph matroid limit. Then, the growth function
δrk(·) of M is non-negative and non-increasing. That is, 0 ≤ δrk(ν + 1) ≤ δrk(ν) for all ν ∈ N.
Proof. Non-negativity follows at once from Proposition 4.12 (i). For the increasing property, write
∆i = Ei+1 \ Ei . Then, for ν ∈ N, we have
δrk(ν) =
=
rk(∆ν |Eν )
rk(∆ν+1 |Eν )
by Proposition 4.12 (iv)
≥ rk(∆ν+1 |Eν+1 ) by Proposition 4.12 (ii)
=
δrk(ν + 1)
Note that up to here, the S(m)-symmetry on the first component of E has not been used,
therefore Proposition 4.14 also holds in the somewhat more general setting where the symmetry
group Gν of Mν is only S(ν).
Proposition 4.15. Let M be an one-sided bipartite graph matroid limit, with ground set [m] × N.
Denote Eν := [m] × [ν], and ∆ν := Eν+1 \ Eν , that is, ∆ν = [m] × {ν + 1}. Then, for all ν ∈ N with
S ⊂ Eν and T ⊂ ∆ν such that #T ≤ δrk(ν), it holds that
rk(T |S) = #T
Proof. Assume, for the moment, that #T = δrk(ν). By Proposition 4.12 (v), there is some T 0 ⊂ ∆ν
such that rk(T 0 |Eν ) = δrk(ν) = #T . We then have the sequence of relations
#T ≥
rk(T |S)
by Proposition 4.12 (i)
≥
rk(T |Eν )
by Proposition 4.12 (iii)
0
= rk(T |Eν ) by row symmetry and Proposition 4.12 (iv)
=
#T
which imply that the inequalities are all tight. The case in which #T < δrk(ν) follows from monotonicity and normalization of matroid rank functions.
Remark 4.16 — Proposition 4.15 is sharp in the sense that for each ν, in the sense that there exist
sets S such that rk(S) < rk(S ∩ Eν ) + #(S ∩ ∆ν ) if #(S ∩ ∆ν ) > δrk(ν + 1). Any S with the property
that #(S ∩ Eν ) is a basis of Mν has this property, from the definition of the growth function, but this
is not the only type of obstruction.
Proposition 4.15 implies that any one-sided bipartite graph matroid limit has a special basis
from which we can read off properties. This will be useful in the next section, where we analyze
two-sided matroid limits.
31
Definition 4.17. Let M be an one-sided bipartite graph matroid limit, with ground set [m]×N. The
staircase basis of M is defined as the union of the (column) rectangles [ν] × [δrk(ν)] for all ν ∈ N.
We define the numbers d1 , d2 , . . . , d t to be the jump sequence of indices ν such that δrk(ν)
δrk(ν + 1). A corner of a staircase basis is an index (µ, ν) such that (µ − 1, ν) and (µ, ν − 1) are
both in the staircase basis, but (µ, ν) is not; in particular, all the corners have column indices of the
form di + 1.
From Proposition 4.14, the growth function δrk is lower bounded and non-increasing. Therefore, its limit exists:
Definition 4.18. Let M be a one-sided bipartite graph matroid limit. The average rank ρ(M) of M
is defined to be limν→∞ δrk(ν).
Remark 4.19 — The terminology average rank can be interpreted in several ways for a one-sided
M , if we denote the average vertex degree of a graph G
bipartite limit graph matroid M = lim
−→ν∈N ν
by ρ(G):
(i) It holds that ρ(M) = limν→∞
1
ν
rk Mν .
(ii) Choose arbitrary basis graphs Gν in Mν . Then ρ(M) = limν→∞ ρ Gν .
(iii) It holds that ρ(M) = sup{ρ(G) : G ∈ I(M)}.
Proposition 4.15 has a useful specialization in terms of the average rank.
Lemma 4.20. Let M be an one-sided bipartite graph matroid limit with average rank ρ, and let G be
any finite bipartite graph with a vertex v of degree d ≤ ρ. Then any edge incident on v is a bridge in
G.
Proof. The graph G has some representative set S ⊂ Eν , with Eν = [m] × [ν], and without loss
of generality, we may assume that the vertex v is mapped to the column ν. Then by Proposition
4.14, δrk(ν) ≥ ρ ≥ d. If rk(G) = rk(G \ {v j}) for some edge v j incident on v, then we have a
contradiction to Proposition 4.15.
The average rank also controls the asymptotic behavior of rk(Mν ) in a one-sided bipartite graph
matroid sequence.
Proposition 4.21. Let M be a one-sided bipartite graph matroid limit on the ground set E(M) =
[m] × N, with average rank ρ. Then, there are unique numbers α, κ ∈ N, such that
rk(Km,ν ) = (ν − κ) · ρ + α for all ν ≥ κ
rk(Km,ν )
(ν − κ) · ρ + α for all ν
κ.
Moreover, α = rk(Km,κ ).
Proof. Write Eν := [m]
P × [ν]. By definition, rk(Km,ν+1 ) − rk(Km,ν ) = rk(Eν+1 |Eν ) = δrk(ν),
therefore, rk(Km,ν ) = n ν δrk(n). Furthermore, by Proposition 4.14, δrk(·) is non-negative nonincreasing, so there is a unique number κ such that δrk(n) = ρ for all n κ and δrk(n) ρ for all
n ≤ κ. Summing up these inequalities for δrk(n) for all n ν, one obtains the claimed inequalities.
Moreover, comparing coefficients yields α = rk(Km,κ ).
Definition 4.22. For a one-sided bipartite graph matroid limit M as in 4.21, the number κ(M) := κ
is called realization size, and the number α(M) := α the realization rank.
32
4.4. Circuits and Bases of One-Sided Bipartite Graph Limits
We will now characterize one-sided bipartite graph limits and their matroid structure through average rank and realization size.
We start with the simple situation when all matroids in question are free.
Proposition 4.23. Let (Mν ) be a one-sided bipartite graph matroid sequence with limit M, such that
M has ground set [m] × N. Then, the following are equivalent:
(i) For all ν, the matroid Mν is free.
(ii) rk(Km,ν ) = mν.
(iii) ρ = m.
In all cases, κ = α = 0.
Proof. (i)⇔(ii) follows from the definition of a free matroid, (ii)⇒(iii) from Proposition 4.21.
(iii)⇒(ii) is obtained from an inductive application of Proposition 4.15 to S = Km,ν and T being
one vertex of degree m added to the growing side.
It can happen that M is not free even if κ = α = 0, e.g. if all elements in Mν are loops. The
realization size and average rank are linked by the existence of a special circuit.
Proposition 4.24. Let M be a one-sided bipartite graph matroid limit on the ground set [m] × N, with
average rank ρ = ρ(M) < m and realization size κ(M) = κ. Then Kρ+1,κ+1 is a circuit graph in M.
Proof. Proposition 4.21 that for large enough `, (ρ +1)` > rk(M` ), and so there is some finite ` ∈ N
for which Kρ+1,` is dependent. Lemma 4.20 implies that any circuit C in Kρ+1,` must be a complete
bipartite graph, since vertices in the growing part of degree at most ρ are incident only on bridges.
For ν > κ + 1, Proposition 4.21 implies that rk(Kρ+1,ν ) < ν(ρ + 1) − 1. The only possibility, then,
for the circuit C is Kρ+1,κ+1 .
Proposition 4.24 implies that this next definition is sensible.
Definition 4.25. Let M be a one-sided bipartite graph matroid limit on the ground set [m] × N,
with average rank ρ = ρ(M) < m and realization size κ(M) = κ. The complete graph Kρ+1,κ+1 is
defined to be the elementary circuit of M.
The existence of the elementary circuit implies several combinatorial characterizations of average rank and realization size:
Theorem 4.1. Let M be a one-sided bipartite graph matroid limit on the ground set [m] × N, with
average rank ρ = ρ(M) < m. Then, the following are equal:
(i) The average rank ρ.
(ii) The largest k ∈ N such that Kk,` is independent for all ` ∈ N.
(iii) The largest k ∈ N such that if G is an independent graph of signature (i, j), then adding a vertex
of degree k to G always results in an independent graph of signature (i, j + 1).
(iv) The largest k ∈ N such that: if C is a circuit, and (c, d) ∈ C, then #{i : (i, d) ∈ C}
(v) The smallest k ∈ N such that Kk+1,` is a circuit for some ` ∈ N.
(vi) The smallest k ∈ N such that there is a circuit of signature (k + 1, `) for some ` ∈ N.
33
k.
Proof. We prove each statement in turn.
(ii): By Lemma 4.20, for k ≤ ρ every edge in Kk,` is a bridge, and hence Kk,` is independent.
On the other hand, the existence of the elementary circuit implies that Kρ+1,κ+1 is dependent.
(iii):By Lemma 4.20, for k ≤ ρ the edges incident on vertices of degree at most k are bridges,
and thus never in any circuit. For k > ρ, Remark 4.16 implies that the conclusion of (iii) is false.
(iv): In the elementary circuit, every vertex in the growing part has degree ρ + 1, and so there
exist circuits with minimum degree ρ + 1 in the growing part. On the other hand, a circuit with a
vertex of degree ρ in the growing part contradicts Lemma 4.20.
(v)–(vi): These follows from (ii) and the existence of the elementary circuit.
Theorem 4.2. Let M be a one-sided bipartite graph matroid limit, with realization size κ = κ(M)
Then, the following are the same:
0.
(i) The realization size κ.
(ii) The smallest ` ∈ N such that Kρ+1,`+1 is dependent.
(iii) The biggest ` ∈ N such that Kρ+1,` is independent.
(iv) The unique ` ∈ N such that Kρ+1,` is a circuit.
Proof. The equality of (i)–(iv) is immediate from the elementary circuit being minimally dependent.
As a direct consequence of Theorems 4.1, we can also derive the following result bounding the
size of all circuits in a bipartite graph limit:
Theorem 4.3. Let M be a one-sided bipartite graph limit, with average rank ρ, realization size κ and
realization rank α. Let C be a circuit of M, with signature (k, `). Then,
` ≤ α − ρ · κ + 1.
Proof. The statement is certainly true if M is a free matroid, since then a circuit C does not exist.
Therefore, by Proposition 4.23, we may suppose that κ 0 By Theorem 4.1 (iv), it holds that
#C ≥ ` · (ρ + 1).
On the other hand, for any e ∈ C, the set C \ {e} is independent, therefore by Proposition 4.21, it
holds that
#C − 1 ≤ rk(Km,` ) = α + (` − κ) · ρ.
Combining the inequalities, one obtains
` · (ρ + 1) ≤ #C ≤ α + (` − κ) · ρ + 1,
an elementary computation then yields the claimed inequality.
As an important corollary, we obtain a finiteness result for the number of circuit graphs:
Corollary 4.26. Let M be a one-sided bipartite graph limit. Then, there are only a finite number of
circuit graphs in M.
Proof. Let [m] × N be the ground set of M, let ρ = ρ(M), κ = κ(M), α = α(M). By Theorem 4.3,
any circuit graph G must be contained in Km,` , where ` = α − ρ · κ + 1. Since there is only a finite
number of graphs contained in Km,` there is only a finite number of circuit graphs of M.
We now return to our running examples.
34
Determinantal matroids We can interpret structural results on determinantal matroids from [17]
in terms of the parameters in this section.
Proposition 4.27. Consider the infinite determinantal matroid D(m × N, r), as obtained in Definition 4.9, with r m, average rank ρ, realization size κ and realization rank α. Then,
ρ = r,
κ = r,
α = r · m.
Proof. The dimension formula in Remark 3.2 (i) implies that
rk(D(m × n, r)) = r · (m + n − r)
(recalling the usual assumption that r ≤ m). The equalities then follow from the uniquness implied
by Proposition 4.21, and comparing coefficients.
Proposition 4.28. Let C be a circuit of D(m × N, r), with signature (k, `). Then,
` ≤ r(k − r) + 1.
Proof. Note that if C is a circuit of D(m × N, r), it is a circuit of D(k × N, r) as well, by Lemma 4.8.
The statement then follows from substituting the numbers computed in Proposition 4.27 into Theorem 4.3.
Bipartite rigidity matroids
plicated.
The behavior of the bipartite rigidity matroids is slightly more com-
Proposition 4.29. Consider the infinite bipartite rigidity matroid CM(m × N, r), as obtained in Definition 4.9, with r m, average rank ρ, realization size κ and realization rank α. Then,
r +1
r +1
ρ = r, κ =
, α= r ·m+3
.
2
3
Proof. The dimension formula in Remark 3.17 (i) implies the statement, together with uniquness
implied by Proposition 4.21, and comparing coefficients.
Remark 4.30 — The geometric intuition behind the value of κ is that for k < κ(CM(m × N, r)), the
dimension formula in Remark 3.17 (i) implies that the complete graph K r+1,k is generically flexible.
If, instead of measuring distances from r generic points on a rigid body, we measure distances to r
joints of a generic, flexible framework, the new point is no longer determined up to a discrete set
of choices.
Proposition 4.31. Let C be a circuit of CM(m × N, r), with signature (k, `). Then,
r +1
`≤ r·k−
+ 1.
2
Proof. Note that if C is a circuit of CM(m×N, r), it is a circuit of CM(k ×N, r) as well, by Lemma 4.8.
The statement then follows from substituting the numbers computed in Proposition 4.29 into Theorem 4.3.
35
4.5. Two-sided Bipartite Graph Matroid Limits
Until now, we have considered matrix-shaped ground sets which are allowed to grow, or to be
infinite, in one direction. Now we will consider both direction growing, or being infinite, at the
same time. For this, we directly introduce a corresponding sequence of bipartite matroids:
Definition 4.32. Let (Mµν ), with µ, ν ∈ N, and Mµν = (Eµν , Iµν ) be a collection of matroids.
(Mµν ) is called two-sided injective complex if Eµν ⊆ Eµ0 ν 0 and Iµν = {S ∩ Eµν : S ⊆ Iµ0 ν 0 } for all
ν ≤ ν 0 , µ ≤ µ0 . If a group Gµν is a group of automorphisms of (Mµν ), the injective complex is called
compatible with the Gµν -action if Gµν ⊆ Gµ0 ν 0 for all ν ≤ ν 0 , µ ≤ µ0 .
Definition 4.33. Let (Mµν ) be an two-sided injective complex of matroids, such that Mµν has
ground set Eµν = [µ] × [ν]. The limit
M = lim lim Mµν
−→ −→
µ∈N ν∈N
is called a two-sided matroid limit, or the limit of (Mµν ). If the Mµν are furthermore bipartite graph
limits with ground set E(Mµν ) = [µ] × [ν], such that the complex is compatible with the graph
symmetry, the sequence (Mµν ) is called atwo-sided bipartite graph matroid complex, and M is called
a two-sided bipartite graph matroid limit.
Note that the limits in Definition 4.33 exist, as iterating the direct limit construction in Definition 4.2 twice is unproblematic, as it has also been remarked at the end of section 4.1. We
also introduce notation for passing back to the one-sided limits, obtained by fixing one of the two
indices:
Definition 4.34. Let (Mµν ) be a two-sided injective complex of matroids, with limit M. We will
denote
Mµ∗ := lim Mµν and M∗ν := lim Mµν ,
−→
−→
ν∈N
µ∈N
noting that for any fixed µ ∈ N, or any fixed ν ∈ N, the Mµν form a one-sided injective sequence of
matroids. We will call the Mµ∗ and M∗ν , informally, slices of M.
Therefore, we directly obtain an analogue of Lemma 4.8 for the two-sided case:
Lemma 4.35. Let (Mµν ), µ, ν ∈ N be a two-sided bipartite graph matroid complex with limit matroid
M, consider the automorphism groups Gµν = S(µ) × S(ν) acting on Mµν . Then, the group G =
S(N) × S(N) acts canonically on M and its ground set N × N. Moreover, for a graph G, it holds:
(i) G is a circuit graph in M/G if and only if G is a circuit graph in Mµν /Gµν for all µ, ν ∈ N such
that G ⊆ Kµ,ν .
(ii) G is a dependent graph in M/G if and only if G is a dependent graph in Mµν /Gµν for all µ, ν ∈ N
such that G ⊆ Kµν .
(iii) G is an independent graph in M/G if and only if G ∩Kµν is an independent graph in Mµν /Gµν for
all µ, ν ∈ N, where G ∩ Kµν is any edge intersection under an arbitrary (disjoint) identification
of vertices with same color.
(iv) For a graph G in M/G, it holds that rkM (G) = rkMµν (G) for all µ, ν ∈ N such that G ⊆ Kµ,ν .
Proof. Iterate the proof of Lemma 4.8; once for the limit in ν, then for the limit in µ.
Therefore, the notion of circuits, dependent graphs, independent graphs, and ranks of graphs,
are again independent of the matroid in which they are considered.
We return to our two running examples.
36
Definition 4.36. Let r ∈ N be fixed. The injective complex of determinantal matroids D(µ × ν, r)
forms a two-sided bipartite graph matroid sequence in the running parameters µ, ν. We denote its
limit
D(N × N, r) := lim lim D(µ × ν, r) = lim D(µ × N, r).
−→ −→
−→
µ∈N ν∈N
µ∈N
Similarly, the injective complex of bipartite rigidity matroids CM(µ×ν, r) forms a two-sided bipartite
graph matroid sequence. We denote its limit by
CM(N × N, r) := lim lim CM(µ × ν, r) = lim CM(µ × N, r).
−→ −→
−→
µ∈N ν∈N
µ∈N
4.6. Average Rank, and Realization Size for Two-Sided Limits
The invariants from the one-sided case have analogues in the two-sided case, which we now develop.
Definition 4.37. Let (Mµν ), µ, ν ∈ N be an injective complex of matroids, with Eµ,ν = E(Mµν ). We
define the growth function of the sequence (Mν ) as
δrk(Mµν ) (·, ·) : N × N → N × N, (µ, ν) 7→ rk(Eµ+1,ν |Eµ,ν ), rk(Eµ,ν+1 |Eµ,ν ) .
If clear from the context, we will just write δrk(·, ·), omitting the dependence on the complex (Mµ,ν )
for readability. In this case, we will write δrk(µ, ν) =: δrk1 (µ, ν), δrk2 (µ, ν) for the components.
If M is the limit of (Mµ,ν ), we will talk about δrk being the growth function of M.
The average ranks (and realization sizes, which we will define later) of a two-sided graph
matroid limit will be fundamentally linked to those of its slices. Therefore, we introduce some
notation for this:
Notation 4.38. Let (Mµν ), µ, ν ∈ N be a two-sided injective complex of matroids. We will denote
(2)
ρm
(M) := ρ(Mm,∗ ),
κ(2)
m (M) := κ(Mm,∗ ),
for average rank and realization size of the slices.
ρn(1) (M) := ρ(M∗,n ),
κ(1)
n (M) := κ(M∗,n )
As in the one-sided case, we will need to distinguish between limits of free matroids and nontrivial sequences.
Definition 4.39. Let M be a two-sided bipartite graph matroid limit. We call an index (µ, ν) ∈ N2
such that Kµ,ν is a dependent graph in M realizing. If there exists a realizing index for M, we call
M realizing.
Remark 4.40 — Let M be a two-sided bipartite graph matroid. Then, the following are equivalent:
(i) M is not realizing.
(ii) M is a limit of free matroids.
Remark 4.41 — Let M be a two-sided bipartite graph matroid. Then, since supersets of dependent
sets are dependent, the following statement is true: If N ∈ N2 is a realizing index for M, then
N 0 ∈ N2 is a realizing index for all N 0 ≥ N ; here, as usual, inequality is taken component-wise.
37
Proposition 4.42. Let M be a two-sided bipartite graph matroid limit with growth function δrk(., .).
Then:
(i) For all µ, ν ∈ N2 , it holds that δrk(µ, ν) ≥ (0, 0)
(ii) For all µ, ν ∈ N2 , the following inequalities hold:
δrk1 (µ + 1, ν) ≤ δrk1 (µ, ν),
δrk2 (µ, ν + 1) ≤ δrk2 (µ, ν).
(iii) For all realizing indices N ∈ N2 , it holds that δrk(N 0 ) ≤ δrk(N ) for all N 0 ≥ N .
(iv) If M is realizing, the limit ρ(M) = limν→∞ δrk(ν, ν) exists.
By usual convention, inequalities between vectors/elements of N2 above are component-wise.
Proof. We prove each statement in turn:
(i): This is directly implied by Proposition 4.12 (i).
(ii): The statement follows Proposition 4.14, and from the fact that for fixed µ, ν ∈ N, it holds
that δrk(µ, ν) = (δrk1 (µ), δrk2 (ν)), where δrk1 is the growth function of the injective sequence
(M∗ν ), and δrk2 is the growth function of the injective sequence (Mµ∗ ).
(iii): It suffices to prove: for realizing N = (µ, ν) ∈ N2 , it holds that
δrk2 (µ + 1, ν) ≤ δrk2 (µ, ν),
δrk1 (µ, ν + 1) ≤ δrk1 (µ, ν).
The statement then follows from (ii) and double induction on µ, ν. We prove the inequality for
δrk2 . Write λ := δrk2 (µ, ν), Ek,` = [k] × [`] for all k, ` ∈ N, and ∆ = [λ] × {ν + 1}. Since N is
realizing, if holds that λ
ν. Therefore, rk(Ei,ν+1 |Eµ,ν ∪ ∆) = 0 for all integers 1 ≤ i ≤ µ. By
Proposition 4.12 (iii), this implies that rk(Ei,ν+1 |Eµ+1,ν ∪ ∆) = 0 for all 1 ≤ i ≤ µ. By using that
the rank is invariant under the S(µ + 1)-action on rows, this implies that rk(Ei,ν+1 |Eµ+1,ν ∪ ∆) = 0
for all 1 ≤ i ≤ µ + 1. Therefore, δrk2 (µ + 1, ν) ≤ λ = δrk2 (µ, ν). The inequality for δrk1 follows in
complete analogy, obtained by exchanging the indices in the above.
(iv): Since M is realizing, there is a realizing index N ∈ N2 . By Remark 4.41, there is n ∈ N such
that (n, n) is realizing. By (i) and (iii), it follows that 0 ≤ δrk(ν +1, ν +1) ≤ δrk(ν +1, ν) ≤ δrk(ν, ν),
for all ν ≥ n, which implies the statement.
Similarly as in the one-sided case, Proposition 4.42 makes the average rank of a two-sided
sequence well-defined:
Definition 4.43. Let M be a realizing two-sided bipartite graph matroid limit. The average rank
ρ(M) of M is defined as the limit limν→∞ δrk(ν, ν) from Proposition 4.42 (vi). We will usually
denote the components of the average rank by ρ(M) =: (ρ1 (M), ρ2 (M)).
If M were invariant under exchanging indices, the realization size could be defined, as for the
one-sided case, as the smallest ν for which δrk(ν, ν) achieves its limit. However, as the behavior
for the first index can be different from the second one, we make a definition which is slightly more
complicated:
Definition 4.44. The realization size κ(M) of M is defined as the unique number (k, `) ∈ N2 such
that: Under the assumption that µ ≥ k, whether (µ, ν) is realizing is independent of µ, and k is the
smallest such integer. Under the assumption that ν ≥ `, whether (µ, ν) is realizing is independent
of ν, and ` is the smallest such integer.
This definition is justified by the following result which states that average rank and realization
size are compatible with taking slices:
38
Proposition 4.45. Let M be a two-sided injective complex of matroids, having average rank (ρ1 , ρ2 ),
and realization size (κ1 , κ2 ). Assume that M is realizing. Then:
(2)
(i) For all m ≥ ρ2 , it holds that ρm
= ρ2 ; for all n ≥ ρ1 , it holds that ρn(1) = ρ1 .
ρ1 , it holds that κ(1)
n = κ1 .
Proof. (i) is a direct consequence of Proposition 4.42, observing that δrk(Mµν ) = δrk(M∗n ) , δrk(Mm∗ ) .
(ii) For all m
ρ2 , it holds that κ(2)
m = κ2 ; for all n
For (ii), observe that (µ, ν) ∈ N2 is realizing if and only if the bipartite graph Km,n is dependent in
M, and that is the case if and only if Km,n is dependent in Mk,∗ , or M∗,` , for any k ≥ m and ` ≥ n.
The statement then follows from Theorem 4.2 (ii).
Remark 4.46 — Proposition 4.45 (i) implies an asympotitic behavior for the average rank:
ρ1 (M) =
ρ2 (M) =
(1)
lim ρm
(M)
m→∞
lim ρn(2) (M)
n→∞
Proposition 4.45 (ii) gives the analogue for the realization size:
κ1 (M) =
κ2 (M) =
lim κ(1)
m (M)
m→∞
lim κ(2)
n (M)
n→∞
A more quantitative statement is given by the following two-sided analogue to the analytic
formula in Proposition 4.21:
Proposition 4.47. Let M be a two-sided bipartite graph matroid limit, assumed to be realizing, with
average rank ρ(M) = (ρ1 , ρ2 ) and realization site κ(M) = (κ1 , κ2 ). Then, there is a unique α ∈ N,
such that
rk(Kµ,ν ) = (µ − κ1 ) · ρ1 + (ν − κ2 ) · ρ2 + α for all µ ≥ κ1 , ν ≥ ρ1 and for all µ ≥ ρ2 , ν ≥ κ2
rk(Kµ,ν )
(µ − κ1 ) · ρ1 + (ν − κ2 ) · ρ2 + α for some ν ≥ ρ1 if µ
κ1 ,
rk(Kµ,ν )
(µ − κ1 ) · ρ1 + (ν − κ2 ) · ρ2 + α for some µ ≥ ρ2 if ν
κ2 .
Moreover, it holds that α = rk(Kκ1 ,κ2 ).
Proof. Note that the three inequality statements are equivalent to stating δrk(µ, ν) = ρ(M) for all
µ ≥ κ1 , ν ≥ ρ1 and for all µ ≥ ρ2 , ν ≥ κ2 , and δrk(µ, ν) ρ(M) for some ν ≥ ρ1 if µ κ1 and
some µ ≥ ρ2 if ν κ2 .
We again make use of the fact that δrk(µ, ν) = (δrk1 (µ), δrk2 (ν)), where δrk1 is the growth
function of M∗ν , and δrk2 is the growth function of Mµ∗ .
By Proposition 4.21, or alternatively, Proposition 4.42 (i) to (iii), for fixed ν 0 ≥ ρ1 , there is
a unique number κ1 (ν 0 ) ∈ N such that δrk1 (µ, ν 0 ) = ρ1 for µ ≥ κ1 and δrk1 (µ, ν 0 )
ρ1 for
µ
κ1 (ν 0 ). Similarly, for each fixed µ0 ≥ ρ2 , there is a unique number κ2 (µ0 ) ∈ N such that
δrk2 (µ, ν) = ρ2 for ν ≥ κ2 (µ0 ) and δrk2 (µ, ν)
ρ2 for ν
κ2 (µ0 ). Note that the ρi could be
0 0
a-priori different for each µ , ν , but they in fact are not due to the fact that dependence and
independence of Kµ,ν does not depend on µ0 or ν 0 , see Lemma 4.35.
By Proposition 4.45, κ1 (ν 0 ) and κ2 (µ0 ) are equal to κ1 and κ2 , under these conditions. Therefore, the formulae follow.
Substituting (µ, ν) = (κ1 , κ2 ), one obtains α = rk(Kκ1 ,κ2 ).
Definition 4.48. Let M be a two-sided bipartite graph matroid limit, assumed to be realizing. The
number α(M) = rk(Kκ1 ,κ2 ) will be called the realization rank of M.
39
4.7. Circuits and Bases of Two-Sided Bipartite Graph Limits
In this section, we study the possible properties of circuits in two-sided limits.
From the proof that the average rank and realization sizes are well-defined, we directly obtain
the existence of elementary circuits in the two-sided case.
Corollary 4.49. Let M be a two-sided bipartite graph matroid limit, assumed to be realizing, with
average rank (ρ1 , ρ2 ) and realization size (κ1 , κ2 ). Then, Kκ1 +1,ρ1 +1 and Kρ2 +1,κ2 +1 are circuit graphs
of M.
Proof. This follows directly from combining (i) and (ii) in Proposition 4.45 with Theorem 4.2 (iv).
Definition 4.50. Let M be a two-sided bipartite graph matroid limit, assumed to be realizing, with
average rank (ρ1 , ρ2 ) and realization size (κ1 , κ2 ). Then, [κ1 + 1] × [ρ1 + 1] is called the 1elementary circuit, and [ρ2 + 1] × [κ2 + 1] is called the 2-elementary circuit of M. The circuit graph
Kκ1 +1,ρ1 +1 is called the 1-elementary circuit graph, and Kρ2 +1,κ2 +1 the 2-elementary circuit graph of
M.
As in the one-sided case, we can prove minimum degree bounds on the vertices of circuit graphs.
For that, we introduce a combinatorial definition.
Definition 4.51. Let (k, `) ∈ N2 , and let G = (V, W, E) be a bipartite graph. The (k, `)-core of G is
the inclusion-wise maximal vertex-induced subgraph G0 = (V 0 , W 0 , E 0 ) of G with the property that
every vertex in V 0 has degree at least k, and every vertex in W 0 has degree at least `
Theorem 4.4. Let M be the limit of a two-sided bipartite graph matroid limit. Let N = (µ, ν) ∈ N2 be
a realizing index, let τ1 = δrk1 (µ − 1, ν) and τ2 = δrk2 (µ, ν − 1). Then:
(i) Let C = (V, W, E) be a circuit graph of M, with signature N . A vertex in V has vertex degree at
least τ1 + 1, and a vertex in W has vertex degree at least τ2 + 1.
(ii) For each basis graph B = (V, W, E) with signature N , a vertex in V has minimum vertex degree
τ1 and a vertex in W has minimum vertex degree τ2 .
(iii) For any finite graph G with signature N , any edge outside of the (τ1 , τ2 )-core of G is a bridge.
Note that (τ1 , τ2 ) ≥ ρ(M), therefore the bounds are at least the average rank of M.
Proof. This is straightforward from Lemma 4.20.
In what follows, elements of N2 are compared with the usual Pareto (component-wise) partial
order on vectors.
Definition 4.52. Let M be a two-sided bipartite graph matroid. We will denote R(M) := {N ∈
N2 : N is realizing for M} ⊆ N2 . Further, we will denote its two projections on N by R1 (M) :=
{m : ∃n, (m, n) ∈ R(M)} and R2 (M) := {n : ∃m, (m, n) ∈ R(M)}. We call an element of
R(M), minimal with respect to the inequality partial order on N2 a minimally realizing index for M.
The set of minimally realizing indices of M will be denoted by ∂ R(M). As before, we will write
∂ R1 (M) := {m : ∃n, (m, n) ∈ ∂ R(M)} and ∂ R2 (M) := {n : ∃m, (m, n) ∈ ∂ R(M)}.
Theorem 4.5. Let M be a realizing two-sided bipartite graph matroid limit, with average rank ρ.
Then, the following are the same:
(i) The average rank ρ.
(ii)
min R2 (M) − 1, min R1 (M) − 1 .
40
(iii)
min ∂ R2 (M) − 1, min ∂ R1 (M) − 1 .
Proof. All equivalences follow from the existence of the two elementary circuits and the definition
of Ri , considering Theorems 4.1, 4.2, and Proposition 4.45.
Theorem 4.6. Let M be a realizing two-sided bipartite graph matroid limit, with realization size κ.
Then, the following are the same:
(i) The realization size κ.
(ii)
max ∂ R1 (M) − 1, max ∂ R2 (M) − 1 .
Proof. All equivalences follow from the existence of the two elementary circuits and the definition
of Ri , considering Theorem 4.2, and Proposition 4.45.
}
For reference, the structural invariants developed in this section are indicated in the following
diagram:
κ2
κ1
{
}
ρ2
∂R1
∂R
R1
}
ρ1
R2
∂R2
The black area indicates the complement of the realizing indices (see Proposition 4.56) and the
red squares indicate minimally realizing indices.
Recall that, in the one-sided case, the average rank, realization size, elementary circuit, and the
staircase basis give the asymptotic structure of the limit matroid. The two-sided case is more subtle.
We start by introducing the analogue of the staircase basis.
Definition 4.53. Let M be a two-sided bipartite graph matroid limit that is realizing. The staircase basis sequences Sm∗ and S∗n are the sequences of staircase bases of the slices Mm∗ and S∗n ,
respectively.
Lemma 4.54. Let M be a two-sided bipartite graph matroid limit that is realizing. Then the staircase
basis sequences Sm∗ and S∗n converge.
(m)
Proof. We prove this for Sm∗ , since the other case is nearly identical. Denote by di
sequence of Sm∗ . Then, by Proposition 4.42, for every m ∈ N, all the
(m+1)
di
≤
(m)
di .
(m)
di
the jump
are at most κ2 , and
It follows that the jump sequences, and thus the sequence of staircase bases, con-
verge.
We introduce notation for the limit staircase bases.
41
Definition 4.55. Let M be a two-sided bipartite graph matroid limit that is realizing. We define
S2 := lim Sm∗ and S1 = lim S∗n .
The main difference between the one-sided and two-sided cases is that, to understand the realizing indices, we need the whole staircase basis sequence.
Proposition 4.56. Let M be a two-sided bipartite graph matroid limit that is realizing. The set of
realizing indices R(M) is exactly
[
[
N2 \
Sm∗ = N2 \
S∗n
m∈N
n∈N
We defer the proof of Proposition 4.56 for the moment, since it will follow from two auxiliary
results.
Lemma 4.57. Let M be a two-sided bipartite graph matroid limit that is realizing. The set of realizing
indices R(M) is a finite union of ideals
[
R(M) =
{r 0 ∈ N : r 0 ≥ r}
r∈∂ R(M)
Proof. That ∂ R(M) is finite is a straightforward application of Dickson’s Lemma. The compatibility
condition of the complex underlying M then implies that if r ∈ N is realizing, so is the ideal
{r 0 ∈ N : r 0 ≥ r}.
Lemma 4.58. Let M be a two-sided bipartite graph matroid limit that is realizing, and let (µ, ν) ∈
∂ R(M) be a minimally realizing index. The ν − 1 appears in the jump sequence of Sµ∗ and µ − 1
appears in the jump sequence of S∗ν .
Proof. By definition, the graphs Kµ,ν−1 and Kµ−1,ν are independent but Kµ,ν is not.
Proposition 4.56 is now straightforward.
Proof of Proposition 4.56. Clearly none of the staircase bases Sm∗ and S∗n can intersect R(M). On
the other hand Lemma 4.58 says that every element of ∂ R(M) appears as a corner of at least one
Sm∗ and at least one S∗m . According to Lemma 4.57, this is sufficient to conclude that any index
that is not realizing is covered by some element of each staircase basis sequence.
For two-sided sequences, we can give a more detailed description of the small circuits. To do
this, we study the evolution of the staircase basis sequences more closely. The following diagram,
which corresponds to the 3-dimensional bipartite rigidity matroid, shows the process.
Sm∗ for m = 1, . . . , 7
42
S∗n for n = 1, . . . , 7
Observe that the union of the staircase bases from either sequences gives the same set, as
show in Proposition 4.56. The realizing circuits from Theorem 4.7 below correspond to the first
appearance of each corner.
We are now in a position to analyze circuits associated with corners in the staircase basis sequence.
Theorem 4.7. Let M be a realizing, two-sided bipartite graph matroid limit. Let m ∈ N be the smallest
number such that (µ, ν) is a corner of the staircase basis Sm∗ . Then the set X = (Sm∗ ∩ ([m] × [ν −
1])) ∪ ([µ] × {ν}) contains a circuit C ⊂ X in M, such that: X and C have the same signature, and
S(X ) fixes C.
A similar statement holds for corners in the sequence S∗n .
Proof. Let X = (Sm∗ ∩ ([m] × [ν − 1])) ∪ ([µ] × {ν}). By construction, X is dependent, and thus it
contains a circuit C. Suppose, for a contradiction, that C ( X is a proper subset.
If C has smaller signature than X then it is independent: if C spans fewer columns, then it is a
subset of Sm∗ ; if it spans fewer rows, then, by the way m was selected, it will be a subset of Sm−1∗ .
The resulting contradiction implies that C and X have the same signature.
Now suppose that X \ C is not empty, and let (i, j) ∈ X \ C. If there is an element σ of S(X ) such
that (i, j) ∈ σ(C), then C ∪ σ(C) is the union of two circuits. By Proposition 4.15, both of these
contain all the entries in the set [µ] × {ν}. Eliminating any of these implies, by Proposition 4.15, a
circuit that is a proper subset of X \ [µ] × {ν}, which is a contradiction.
Corollary 4.59. In the notations of Theorem 4.7, if the set X is of the form
(i) [m] × [n]
(ii) [m] × [n] ∪ [m + 1] × [τ], with τ < n
then for the circuit it holds that C = X .
Proof. In case (i), the symmetry group of X acts transitively, so if X \ C is non-empty, then by
Theorem 4.7, then C is empty, a contradiction.
For case (ii), we have a case distinction. Let (i, j) ∈ X \ C. If (i, j) ≤ (m, τ), then Theorem 4.7
implies that the entire rectangle [m] × [τ] is in X \ C. It then follows that C contains vertices of
degree 1, contradicting Lemma 4.20. Otherwise, by similar reasoning, either a row or column of X
is missing form C. Thus, C has smaller signature than X , which is also a contradiction.
Definition 4.60. We define the circuits appearing in the conclusion of Theorem 4.7 to be the realizing circuits of a two-sided bipartite graph matroid limit M. (Observe that the elementary circuits
are included in this class.)
We give characterizations of special realizing circuits:
Theorem 4.8. Let M be a two-sided bipartite graph matroid limit, with growth function δrk. Let
(µ, ν) ∈ N2 . Let (τ1 , τ2 ) = δrk(µ−1, ν −1)+(1, 1), ς1 = δrk1 (µ−1, ν)+1, and ς2 = δrk2 (µ, ν −1)+1.
Consider the following statements:
(i) (µ, ν) ∈ ∂ R(M), that is, (µ, ν) is a minimally realizing index for M.
(ii) τ1
ς1 , and Kµ−1,ν is independent.
(iii) τ2
ς2 , and Kµ,ν−1 is independent.
(iv) It holds that τ2 ν, the set [µ − 1] × [ν] ∪ [µ] × [τ2 ] is a circuit of M, the corresponding graph
Kµ−1,ν ∪ Kµ,τ2 is a circuit graph of M.
43
(v) It holds that τ1 µ, the set [τ1 ] × [ν] ∪ [µ] × [ν − 1] is a circuit of M, the corresponding graph
Kτ1 ,ν ∪ Kµ,ν−1 is a circuit graph of M.
Then, (ii) implies (iv), and (iii) implies (v). Moreover, (i) is equivalent to the conjunction “(iv) and
(v)”.
Proof. (iv) and (v) ⇒ (i): Since τ2
ν, the graph Kµ,ν is dependent, so (µ, ν) is a realizing index. Since Kµ−1,ν ∪ Kµ,τ2 is a circuit, all Kk,` with k
µ, ` ≤ ν are proper subgraphs, therefore
independent. Similarly, since Kτ1 ,ν ∪ Kµ,ν−1 is a circuit, all Kk,` with k ≤ µ, ` ν are independent.
Therefore, (µ, ν) is a minimally realizing index.
(i) ⇒ (iv), (v) can be obtained from Corollary 4.59. Similarly, (ii) ⇒ (iv),(v) are consequences
of Corollary 4.59.
We close with bounding the signature of all circuits, generalizing Theorem 4.3 to the two-sided
case:
Theorem 4.9. Let M be a two-sided bipartite graph matroid limit, with average rank (ρ1 , ρ2 ), realization size (κ1 , κ2 ) and realisation rank α. Let C be a circuit of M, with signature (k, `). Then,
k ≤ α − ρ1 · κ1 + ρ2 · (` − κ2 ) + 1
` ≤ α − ρ2 · κ2 + ρ1 · (k − κ1 ) + 1
k≥
`≥
(κ1 · ρ1 + κ2 · ρ2 − α − 1)(ρ2 + 1)
ρ1 · ρ2 − 1
(κ1 · ρ1 + κ2 · ρ2 − α − 1)(ρ1 + 1)
ρ1 · ρ2 − 1
In particular, if ρ1 = ρ2 , κ1 = κ2 , e.g. when transposition is an isomorphism of M, one has
k ≤ α + ρ · (` − 2κ) + 1
` ≤ α + ρ · (k − 2κ) + 1
2κ · ρ − α − 1
,
k, ` ≥
ρ−1
where writing ρ = ρ1 = ρ2 and κ = κ1 = κ2 by abuse of notation.
Proof. The first two equations follow from applying Proposition 4.45 to compare M to the limit
matroids on the ground sets [k] × N and N × [`], then applying Theorem 4.3. The second two
equations are obtained from substituting the first two into each other, then solving once for k, once
for `. The last three equations are obtained from substituting ρ = ρ1 = ρ2 and κ = κ1 = κ2 into
the first four.
We return to our running examples:
Determinantal matroids
Proposition 4.61. Consider the infinite determinantal matroid D(N × N, r), as obtained in Definition 4.36, with average rank (ρ1 , ρ2 ), realization size (κ1 , κ2 ) and realization rank α. Then,
ρ1 = ρ2 = r,
κ1 = κ2 = r,
α = r 2.
Both the 1- and 2-elementary circuit graphs are equal to K r+1,r+1 .
44
Proof. This can be deduced from the dimension formula in Remark 3.2 (i), and Proposition 4.21,
or, alternatively Propositions 4.45 and 4.27.
Since transposition is an automorphism of the the determinantal matroid, application of Theorem 4.9 yields no improvmenet over Proposition 4.28.
Proposition 4.62. The set of minimally realizing indices for the matroid D(N × N, r) is ∂ R = {(r +
1, r + 1)}. All realizing circuit graphs coincide with the elementary circuit graph K r+1,r+1 .
Proof. D(N × N, r) is a realizing matroid, see for example Proposition 4.61. By substituting the
numbers from Proposition 4.61 into Theorems 4.5 and 4.6, one obtains (r + 1, r + 1) ≤ N ≤
(r + 1, r + 1) for all N ∈ ∂ R, and ∂ R 6= ∅, which implies the statement.
Bipartite rigidity matroids
Proposition 4.63. Consider the infinite bipartite rigidity matroid CM(N × N, r), as obtained in Definition 4.36, with average rank (ρ1 , ρ2 ), realization size (κ1 , κ2 ) and realization rank α. Then,
r +1
r
ρ1 = ρ2 = r, κ1 = κ2 =
, α = r · (m + n) −
.
2
2
The 1-elementary circuit graph is K(r+1)+1,r+1 , while the 2-elementary circuit graph is K r+1,(r+1)+1 .
2
2
Proof. This can be deduced from the dimension formula in Remark 3.17 (i), and Proposition 4.21,
or, alternatively Propositions 4.45 and 4.29.
Since transposition is an automorphism of the the determinantal matroid, application of Theorem 4.9 yields no improvement over Proposition 4.31.
Proposition 4.64. The set of minimally realizing indices for the matroid CM(N × N, 1) is ∂ R =
{(2, 2)}, and all realizing circuit graphs coincide with K2,2 . For r ≥ 2, the set of minimally realizing
¦ r+1
©
r+1
indices for the matroid CM(N × N, r) is ∂ R = ( 2 + 1, r + 1), (r + 2, r + 2), (r + 1, 2 + 1) .
The 1-realizing circuit graphs and the 2-realizing circuit graphs obtained from ∂ R coincide: they are
K(r+1)+1,r+1 , K r+2,r+2 and K r+1,(r+1)+1 . The remaining realizing circuit graphs are Kk,r+1 ∪ K r+1,` with
2
2
r+1
k, ` r + 1 and k + ` = 2 + r + 2.
Proof. CM(N × N, r) is a realizing matroid, see for example Proposition 4.61. By substituting the
numbers from Proposition 4.63 into Theorems 4.5 and 4.6, one obtains (r + 1, r + 1) ≤ N ≤
(r + 1, r + 1) for all N ∈ ∂ R, and ∂ R 6= ∅, which implies the statement for the realizing circuits
obtained from ∂ R. The remaining realizing circuits are obtained from the dimension formula in
Remark 3.17 (i).
4.8. Infinite Graph matroids
In this section, we briefly sketch how the results of the previous section extend to graph matroids.
The proofs of the statements can be obtained in an analogous way to the previously presented ones.
Some of the results on one-sided limits can be re-obtained by bipartition from the graph setting.
Definition 4.65. Let (Mν ) be an injective sequence of graph matroids, such that Mν has ground set
[ν]
Eν = 2 . The sequence (Mν ) is called graph matroid sequence, and the limit M = lim
M a
−→ν∈N ν
graph matroid limit, or the limit of (Mν ).
45
Definition 4.66. Let M be a graph matroid limit. We define the growth function of M as
δrkM (·) : N → N,
ν 7→ rk(Kν+1 |Kν ).
If clear from the context, we will just write δrk(·), omitting the dependence on M.
Proposition 4.67. Let M be a graph matroid limit. Then, the growth function δrk(·) of M is nonnegative and non-increasing. That is, 0 ≤ δrk(ν + 1) ≤ δrk(ν) for all ν ∈ N.
Definition 4.68. Let M be a graph matroid limit with growth function δrk. We define the average
rank of M to be ρ(M) = limν→∞ δrk(ν).
N
Proposition 4.69. Let M be a graph matroid limit on the ground set 2 , with average rank ρ. Then,
there are unique numbers α, κ ∈ N, such that
rk(Kν ) = (ν − κ) · ρ + α for all ν ≥ κ
rk(Kν )
(ν − κ) · ρ + α for all ν
κ.
Moreover, α = rk(Kκ ).
Note however that Kκ will in general not be a circuit, nor the symmetrization Kρ+1,κ+1 which is
obtained by gluing vertices.
5. Circuit Polynomials and Top-Degrees
By the unique correspondence between algebraic and coordinate representations of matroids, given
in Theorem 2.2, circuits in algebraic matroids can be studied by means of commutative algebra.
While the viewpoint of coordinate matroids is most useful for this section, all of what follows will
also be valid for the equivalent cryptomorphic realizations by linear, basis, or algebraic matroids.
5.1. Top-degree
The top-degree is a multivariate generalization of the the degree of a polynomial which has been
introduced in [18, Tutorial 48]. We slightly generalize it to the case of potentially infinite sets of
variables, in order make it applicable later to circuits with potentially infinite symmetries. This short
exposition on top-degree has mostly notational purpose in order to capture the circuit invariants,
which will follow in the next sections, in a concise way.
Notation 5.1. Let I be an index set, let x = {X i : i ∈ I} be a set of coordinates. We will denote by
K[x] the corresponding polynomial ring in potentially infinitely many variables.
Notation 5.2. Let n ⊆ I be a finite sub-multiset of I, i.e., n is a subset of I where elements can occur
with multiplicities.5 Then, by convention, we will write
Y
X n :=
Xi.
i∈n
Also, we will denote inverses by
X −n = (X n )−1 .
5
We adopt the usual multiset convention that a sub-multisets n of I is a subset of I with possible multiplicities; i.e.,
a function I → N. For i ∈ n, we denote by n(i) the multiplicity of i in n, and by ∪ and ∩ the usual multiset union and
intersection, and by ] the multiset sum.
46
Furthermore, for a multiset n and k ∈ N, we will denote
k ∗ n :=
k
[
n.
i=1
Definition 5.3. Let f ∈ S, assume
f =
X
an X n
with an ∈ K,
n⊆I
where by definition of K[x] only finitely many an are non-zero. Then, the multiset
[
>deg f =
n
an 6=0
is called top-degree of f . We denote the set associated to the multiset >deg f by supp f , called the
support of f .
We want to remark that the top-degree is, in essence, a compact way to write the collection of
single degrees:
Lemma 5.4. Let f ∈ K[x]. For i ∈ I, denote by di := degX i f the degree of f considered in the single
variable X i . Then,
[
>deg f =
di ∗ {i}.
i∈I
Proof. This follows from the definitions of degree, and top-degree, observing that taking unions of
multisets is equivalent to taking the maximum of degX i .
Definition 5.5. Let I ⊆ R. Then the set of all top-degrees >deg(I) = {>deg g : g ∈ I} is called
the exponent set of I. Similarly, the set of all supports supp(I) = {supp g : g ∈ I} is called the
support set of I. Both >deg(I) and supp(I) have the natural inclusion partial order attached to its
elements.
5.2. Circuit polynomials
Every circuit in a coordinate matroid has a unique polynomial attached to it, the circuit polynomial:
Lemma 5.6. Consider a coordinate matroid M = C(P, x), with fixed realization (P, x). Let C be a
circuit of M. Then, there is (up to multiplication with a unit in K) a unique polynomial θC such that
P ∩ K[x(C)] = θC · K[x(C)]
Proof. Since C is a circuit of M, the ideal P ∩ K[x(C)] is of height 1 in K[x(C)]. Therefore, for
example by [26, I.§7, Proposition 4], P ∩K[x(C)] is principal from which the statement follows.
Lemma 5.6 states that θC is an invariant of the circuit C, making the following well-defined:
Definition 5.7. Keep the notation of Lemma 5.6. Then, we will call θC circuit polynomial of the
circuit C, associated to the realization (P, x). Understanding that there are in general more than one
circuit polynomial, all constant multiples of each other, we will say that θC is the circuit polynomial
if the subsequent statement does not depend on the choice of the particular multiple.
Note again that by the cryptomorphisms given in Theorem 2.2, circuits in algebraic or basis
matroids as well have unique circuit polynomials attached to them as invariants.
47
Remark 5.8 — Existence and uniqueness of the circuit polynomial, as given in Lemma 5.6, has been
already proven as Lemma 1.0 and Corollary 1.1 of [12]. However, no further structural statements
have been made there, as the purpose of the discussion there is different.
Lemma 5.9. Let M = C(P, x) be some coordinate matroid. Then, supp(P) = I(M)C .
Proof. One has z ∈ supp(P) if and only if there is a non-zero polynomial P ∈ P ∩ K[z] such that
P(z) = 0. This is equivalent to z being algebraically dependent, which is true if and only if z = x(S)
for S 6∈ I(M).
An important property of circuit polynomials is that they have minimal top-degree and minimal
support:
Proposition 5.10. Let M = C(P, x) be some coordinate matroid. Then:
(i) The set of circuits C(M) is exactly the set of minima of supp(P) w.r.t. inclusion partial order.
(ii) Let C be a circuit of M. Then, >deg θC is minimal in >deg(P), and supp θC is minimal in supp(P).
Proof. (i) follows from the definition/fact that a circuit is an inclusion-minimal set in I(M)C , and
Lemma 5.9.
(ii) θC is (up to constant multiple) the unique element in P with support supp θz (P). Therefore,
it suffices to prove the statement for the supports, since inclusions are preserved under passing
from multisets to the underlying sets, thus from top-degree to support. But the latter follows from
Lemma 5.9, or directly from (i).
The top-degrees of the circuit polynomials are further invariants associated to any matroid.
While the circuits specify the dependence structure, describing solvability, the top-degrees yields
insight on the degree structure, describing the number of solutions. Naive generalizations or converses of Proposition 5.10 fail, as the multigraded part associated to minimal top-degrees does not
need to be a principal ideal. For matroids with symmetries, such as graph matroids, we introduce
notation to depict the circuit and the degree of its polynomial simultaneously:
Notation 5.11. Let M = C(P, x) be some coordinate matroid on the ground set [m] × [n], so that
the coordinates are indexed as x = {X i j : 1 ≤ i ≤ m, 1 ≤ j ≤ n}. Let C be a circuit in M. We
will depict C as a top-degree mask: the mask corresponding to C is a matrix A in Nm×n , with entries
Ai j = degX i j θx (C)(P).
If, in addition, M is a bipartite graph matroid, we can also depict C as a weighted bipartite
graph; the graph is the one with the top-degree mask as its adjacency matrix and the edge-weights
are the variable degrees. The symmetric top-degree mask and the weighted graph for circuits in
graph matroids are defined in analogous manner.
Note that when zeroes in the top-degree mask are replaced with ◦, and all other entries with
•, one obtains the ordinary mask corresponding to the circuit C. Similarly, the (bipartite) graph
associated with C is obtained by removing the weights.
Example 5.12. We return to Examples 3.3 and 3.6. In the case of D(4, 4, 2), the top-degree masks
and weighted graphs do not carry much new information. With only one exception, all of the
appearing variables have degree 1, as pictured below:
1
1
1
·
1 1 1 ·
1 · 1 1
1 1 1 ·
1 1 2 1
1 1 1 ·
2
· 1 1 1
· · · ·
48
On the other hand, the symmetric case D (4, 4, 2) has top-degree mask and weighted graph
carrying much more information about the circuit polynomials than the unlabeled counterparts, as
depicted in Table 1.
1
2
2
1
1
2
2
2
1
1
4
1
4
1
2
4
2
2
2
2
2
1
2
2
·
2
·
2
2
2
2
2
2
·
2
2
1
1
1
1
1
1
2
1
·
1
1
4
1
1
1
·
1
2
1
1
1
2
2
2
2 · 4
· 2 4
4 4 2
4 4 ·
4
4
·
2
1
1
1 2 2
2 1 2
2 2 1
· · ·
·
·
·
·
Table 1: Top-degree Masks and weighted graphs for D (4, 4, 2)
5.3. The Multihomogenization
The concepts of multihomogeneity in this section are an extension of the definitions by Kreuzer and
Robbiano in [18] to the infinite case.
For multihomogenization, we introduce a copy Yi for every variable X i in the original set of
coordinates.
Notation 5.13. Let I be an index set, let x = {X i : i ∈ I} be a set of coordinates. Let y = {Yi : i ∈ I}
be another such set. We will denote the ideal 〈y〉 ⊆ S which is generated by all Yi by Y. For
abbreviation, we will write R = K[x] and S = R[y].
The top-degree is extended to polynomials in both sets of variables as follows:
Notation 5.14. Let f ∈ S, let f 0 ∈ R be the polynomial obtained by the substitution Yi = X i for all
i ∈ I. Then, by convention, we define
>deg f := >deg f 0
supp f := supp f 0 .
Definition 5.15. Let f ∈ R. The polynomial f ∈ S defined as
X
X
fÍ=
an · X n · Y >deg f \n =
an · X n · Y >deg f · Y −n
Í
n⊆I
n⊆I
is called multihomogenization of f .
Conversely, g ∈ S is called multihomogeneous if >deg(g) = >deg(M ) for each (non-zero) monomial M occurring in g.
Remark 5.16 — A more intrinsic but less constructive way to define the multihomogenization of
f ∈ R is as the result of homogenizing f subsequently with respect to each X i -grading with a
49
new homogenizing variable. Since only finitely many X i occur in f , there are only finitely many
homogenizations to perform. The top-degree notion on S then also canonically arises from the
top-degree on R.
Lemma 5.17. Let f ∈ R. Then, f Í is multihomogeneous, of (ordinary) degree #>deg f , and >deg f Í =
>deg f .
Proof. The statements are elementary to check and follow from the definitions.
Lemma 5.18. Let f , g ∈ S be multihomogeneous. Then:
>deg( f + g) ⊆ >deg f ∪ >deg g
and
>deg( f · g) = >deg f ] >deg g.
Moreover, if >deg f 6= >deg g, then
>deg( f + g) = >deg f ∪ >deg g.
Proof. The statements are elementary to check and follow from the definitions.
Corollary 5.19. Let C ⊆ P(I). Assume all elements of C are minimal with respect to the inclusion
partial order on multisets. Let fn ∈ S, n ∈ C be a family of multihomogenous polynomials such that
>deg fn = n. Then, the strict inclusion
〈 fn ; n ∈ C \ {c}〉 ( 〈 fn ; n ∈ C〉
of ideals in S holds for all c ∈ C.
Proof. We prove the statement by contradiction. I.e., assume that there is a c ∈ C such that
〈 fn ; n ∈ C \ {c}〉 = 〈 fn ; n ∈ C〉.
This implies that there is a finite subset C ∈ C \ {c} such that
X
fc =
fn gn with homogenous gn ∈ S \ {0}.
n∈C
Lemma 5.18 implies, together with the fact that the elements of C are minimal, that
[
c⊇
n,
n∈C
which implies n ⊆ c for any n ∈ C, which is a contradiction to minimality.
Definition 5.20. Let I ⊆ R be an ideal. We define
IÍ = 〈 f Í ; f ∈ I〉
to be the ideal generated by all multihomogenizations of elements in I.
Definition 5.21. The canonical quotient morphism of rings
.Ï : S → S/〈Yi − 1 ; i ∈ I〉 = R
where the identification on the right hand side is the canonical one, will be denoted by the superscript []Ï . That is, for an element f ∈ S or an ideal I ⊆ S, we will denote the canonical image in R
by f Ï resp. IÏ .
Lemma 5.22. For f ∈ R and any multihomogenous polynomial h ∈ R, it holds that
f = f ÍÏ
and
(h · f )Í = h · f Í .
For ideals I ⊆ R and homogenous J ⊆ S, it holds that
I = IÍÏ
and (J : Y) = JÏÍ .
Proof. The statements are elementary to check and follow from the definitions.
50
5.4. MH-bases
An MH-basis is a generating set in which no variable cancellations need to occur to generate the
ideal:
Definition 5.23. Let I ⊆ R be an ideal. Then, a generating set { f j : j ∈ J} is called MH-basis of I if
for every h ∈ I, there exists a finite subset J 0 ⊆ J and g i ∈ S, i ∈ J 0 such that
X
h=
f i · g i and >deg f i ] >deg g i ⊆ >deg h for all i.
i∈J 0
The definition of MH-bases does not use homogenization. However, there is a concise characterization of MH-bases in terms of homogenization which can serve as alternate definition:
Proposition 5.24. Let I ⊆ R. The following are equivalent:
(i) { f1 , f2 , . . . } is an MH-basis of I.
(ii) { f1Í , f2Í , . . . } is a generating set for IÍ .
Proof. For readability, we prove the statement in the case of a finite MH-basis { f1 , f2 , . . . , f k }. The
proof of the general statement is in complete analogy.
First note that S and IÍ decompose as direct sums
M
M
S=
Sn and IÍ =
IÍ
n,
n⊆I
n⊆I
where the direct sum runs over all sub-multisets of I, and where Sn , IÍ
n denote the sub-K-vector
Í
spaces spanned by the multihomogenous top-degree-n-elements of S, I . For m ⊆ I, we will denote
by
M
M
S≤m =
Sn and I≤m =
IÍ
n
n⊆m
n⊆m
(ii)⇒ (i): By definition, for every homogenous h ∈ I4 , there exist g1 , . . . , g k ∈ S such that
h=
k
X
f iÍ · g i .
i=1
Let mi = >deg h − >deg f i . By the direct sum decompositions above, one can choose the g i in S≤mi .
Dehomogenizing, one obtains
hÏ =
k
X
f iÍÏ · g iÏ =
i=1
k
X
f i · g iÏ .
i=1
Now .ÍÏ is surjective as well as .Í , thus for each ` ∈ I there exists h such that ` = hÏ . Furthermore,
>deg g iÏ ⊆ >deg g i . Therefore, { f1 , . . . , f k } is an MH-basis of I.
(i)⇒ (ii): By definition, for every h ∈ I, there exist g1 , . . . , g k ∈ R such that
h=
k
X
fi · gi
and
>deg f i ] >deg g i ⊆ >deg h for all i.
i=1
51
Homogenizing, one obtains that
h =
Í
k
X
f iÍ · g iÍ · w i
for some w i ∈ Y.
i=1
Since any ` ∈ IÍ can be written as a finite sum of such hÍ , the set { f1Í , . . . , f kÍ } is a generating set
for IÍ .
Proposition 5.25. Let I ⊆ R, let { f j : j ∈ J} be an MH-basis of I. Let n be a minimum of >deg(I).
Then, there is a j such that >deg f j = n.
Proof. This is a direct consequence of Corollary 5.19 and Proposition 5.24.
Naive converses of Proposition 5.25 - i.e., that a set of elements having all minimal top-degrees
is an MH-basis, or that all elements of a MH-basis have minimal top-degrees - are false. One
important corollary is however the following:
Corollary 5.26. Let P ⊆ R be a prime ideal, let { f j : j ∈ J} be an MH-basis of I. Then, an MH-basis
{ f j } for P contains all circuit polynomials θC for circuits C in C(P, x).
5.5. Finiteness of top-degrees
In Theorem 4.3, if has been shown by purely combinatorial arguments that for one-sided bipartite
graph limit M, there are a finite number of circuit graphs. If M is, in addition, algebraic, a similar
finiteness statement holds for the set of minimal top-degrees.
For this, we will consider the following concept of finiteness, which is the algebraic counterpart
to a finite set of graphs:
Definition 5.27. Let x = {X i j : i ∈ [m], j ∈ N}, consider the canonical S(N)-action on the second
index of X i j which extends canonically to a group action on K[x]. An ideal P ⊆ K[x] is called
S(N)-stable, if σ(P) = P for all σ ∈ S(N). Furthermore, a set S ⊆∈ K[x], such that the set of
orbits S/S(N) is finite, is called S(N)-finite. The ideal P is called S(N)-finitely generated if it has a
S(N)-finite set of generators.
We will make use of the following finiteness theorem from [15] to the coordinate realization of
said graph limit:
Theorem 5.1 ([15]). Let x = {X i j : i ∈ [m], j ∈ N}, consider the polynomial ring K[x] in the
infinitely many variables x. Then, K[x] is a Noetherian K[x][S(N)]-module.
In particular, if P ⊆ K[x] is a S(N)-stable ideal, then, P is S(N)-finitely generated.
Proof. This is a reformulation of Theorem 1.1 in [1].
Corollary 5.28. Under the assumptions of 5.1, the ideal P has an S(N)-finite MH-basis.
Proof. The S(N)-action canonically extends to the ring K[x, y] in which the multihomogenization is considered, compare Notation 5.13. In particular, the multihomogenization PÍ is again
S(N)-stable. Therefore, by Theorem 5.1, there exists a S(N)-finite set of generators S of PÍ . By
Proposition 5.24, the dehomogenization S Ï is a S(N)-finite MH-basis of P.
The finiteness theorem can be directly derived from Corollary 5.28:
52
Theorem 5.2. Let M be an algebraic one-sided bipartite graph matroid limit, with coordinate realization (P, x). Then, up to the graph symmetry, there set of inclusion-minimal top-degrees in >deg(P) is
S(M)-finite. Differently phrased, there is only a finite number of minimal top-degrees in P, up to the
canonical graph symmetry.
Proof. By Corollary 5.28, the ideal P has a S(N)-finite, therefore S(M)-finite MH-basis, which by
Proposition 5.25 contains all the inclusion-minimal elements of >deg(P).
5.6. Multivariate Galois Theory
As the results in the previous sections concern algebraic matroids, they can be translated into
statements about complexes of field extensions, in particular their minimal polynomials. Namely,
the concept of circuit polynomial generalizes the concept of minimal polynomial for a finite set of
field extensions. We briefly recapitulate the main result on the minimal polynomial:
Theorem 5.3. Let K be a field and α algebraic6 over K. Write L = K(α).
Then, there is an irreducible polynomial m ∈ K[X ], of degree7 deg m = [L : K],
satisfying m(α) = 0.
Moreover, such an m is unique up to multiplication with a unit in K × .
Furthermore, for all f ∈ K[X ] such that f (α) = 0, it holds that deg f ≥ deg m, and m divides f .
Since m is unique up to multiplication with a constant, one usually takes the minimal polynomial
of α over K to be the unique m with leading coefficient 1. Since this somewhat arbitrary convention
hinders the analogy slightly, we make a surrogate definition:
Definition 5.29. Let K be a field, α algebraic over K. Any m ∈ K[X ] of degree deg m = dimK L,
satisfying m(α) = 0 is called an infimal polynomial of α over K.
The existence of a circuit polynomial, guaranteed by Lemma 5.6, implies an analogue for multiple field extensions, which is a strict generalization:
Theorem 5.4. Let K be a field and α := {α1 , . . . , αn } be a collection of elements over K. Write L :=
K(α) and L i := K(α \ {αi }). Assume that the αi form an algebraic circuit; that is, trdeg L i /K = n − 1
for all i, and L/L i is algebraic for all i.
Sn
Then, there is an irreducible8 polynomial θ ∈ K[X 1 , . . . , X n ], of top-degree >deg θ = i=1 [L :
L i ] ∗ {i}, such that θ (α) = 0.
Moreover, such a θ is unique up to multiplication with a unit in K × .
Furthermore, for all f ∈ K[X 1 , . . . , X n ] such that f (α) = 0, it holds (component-wise) that
>deg f ≥ >deg θ , and θ divides f .
Finally, for all i, the polynomial θ (α1 , . . . , αi−1 , X i , αi+1 , . . . , αn ) ∈ K[X i ] is an infimal polynomial
for αi over L i .
Proof. Existence of θ , and uniqueness up to a multiplicative unit, follows from taking the coordinate realization cryptomorphic to the algebraic realization AK (α), then applying Lemma 5.6 to it,
noticing that by assumption [n] is a single circuit.
The fact that for all f ∈ K[X 1 , . . . , X n ] such that f (α) = 0 one has that θ divides f follows from
the fact that the ideal K[X ]∩P in the proof of Lemma 5.6 is principal. That implies >deg f ≥ >deg θ
by Lemma 5.18. Furthermore, it implies that θ is irreducible.
For the remaining statements, we note that θi := θ α1 , . . . , αi−1 , X i , αi+1 , . . . , αn is an infimal
polynomial by Theorem 5.3. Therefore, for any i, it holds that θi = mi · hi for some infimal polynomial mi of αi over L i , and some hi ∈ L i [X i ]. Since the infimal polynomial mi is unique up to
that is, there exists a polynomial P ∈ K[X ] such that P(α) = 0
as usual, we denote by [L : K] the dimension dimK L, where L is considered as a K-vector space
8
that is, irreducible over K
6
7
53
multiplication in L i , we can clear denominators and assume that hi , mi ∈ K[α \ {αi }][X i ]. Since
by assumption trdeg L i /K = n − 1, we can replace α j by the corresponding variables X j , to obtain
polynomials hi , mi ∈ K[X 1 , . . . , X n ] such that hi (α) · mi (α) = 0. Since θi = mi · hi by assumption,
θ must therefore divide hi · mi . Since θ = mi · hi by assumption, θ must be equal to exactly one of
hi , mi . Since hi was arbitrary, and therefore hi also is, it follows that θ = mi , implying that θi is an
infimal polynomial for αi over L i and therefore degX i θ = [L : L i ], the latter implying the statement
about the top-degree by Lemma 5.4.
First note that Theorem 5.3 is implied by Theorem 5.4, by taking n = 1. On the other hand,
Theorem 5.4 is strictly stronger, since it states that the infimal polynomials all uniquely lift to the
circuit polynomials - up to multiplicative constant, which is now in K × instead of L i× . That is a
statement which cannot be inferred from standard Galois theory.
Furthermore, the assumption in Theorem 5.4 that the αi form an algebraic circuit is not a huge
restriction, since from matroid theory, in a set of αi which is algebraically dependent, one can
always pick subsets that are algebraic circuits. However, it is a restriction in the sense that a similar
lifting will not occur in general if there are more than one way to pick a subset which is an algebraic
circuit, as the following example illustrates:
Example 5.30 (8.15 in [34]). Let I ⊆ C[a, b, c, d] be given by:
I = 〈ad − bc, ac 4 − b3 d 2 , a3 c 2 − b5 , b2 d 3 − c 5 , a2 c 3 − b4 d〉.
The first generator has minimal top-degree (1, 1, 1, 1) despite not being a circuit.
6. Algebraic and Combinatorial Structure Theorems
In this section, we will consider matroids which are both (bipartite) graph matroids, and algebraic.
Bringing together the concept of circuit polynomial for algebraic matroids, and the symmetry statments for graph matroids, we obtain several structural statements on the circuit polynomials, and
inductive relations between circuits.
6.1. Symmetries of Circuit Polynomials
In this section we analyze how symmetries of a circuit translate into symmetries of the corresponding circuit polynomial. Namely, uniqueness of the circuit polynomial implies that the automorphism
group S(S) acts on the unit group of the ground field K:
Proposition 6.1. Let M be a bipartite graph matroid, having a coordinate realization M = C(P, x)
over the ground field K. Let C be a circuit of M, and let θC ∈ K[x(C)] be the corresponding circuit
polynomial. Then, the map
σ θC
×
S(C) → K , σ 7→
θC
is a group homomorphism. Furthermore, there are a unique minimal n ∈ N, and a group homomorphism ν : S(C) → Z/nZ, such that
σθC = ζν(σ)
θC ,
n
where ζn ∈ C is an n-th root of unity. Moreover, if char K = 0, there exists a set of monomials
S ⊂ K[x(C)], such that
X
X
θC =
ζν(σ)
σ(M ),
n
σ∈S(C)
M ∈S
which is unique up to σ-action on each element of S if S is chosen with minimal cardinality.
54
Proof. Let σ ∈ S(C) arbitrary. By definition of S(C), it holds that σ(C) = C, so by construction,
the polynomial σθC is a circuit polynomial of C. Since circuit polynomials are unique up to a
multiplicative constant in K × , as it follows from Lemma 5.6, this gives rise to the claimed group
homomorphism ϕ : S(C) → K × . Since S(C) is finite, the image ϕ(S(C)) must be finite as well,
therefore contained in the some multiplicative group generated by ζn for some n ∈ N, which can be
chosen uniquely minimal. The remaining considerations follow from substitutions.
Corollary 6.2. In the situation of Proposition 6.1, >deg θC is invariant under the action of S(C).
That is, the top-degree mask of C is invariant under the action of S(C).
Proof. This follows from the last statement in Proposition 6.1.
Therefore, top-degrees can be associated with the corresponding edges of the bipartite graph.
The circuit polynomial associated with rectangular circuits is particularly symmetric:
Proposition 6.3. Let C = [k] × [`] be a circuit of an algebraic bipartite graph matroid M = C(P, x)
over the ground field K. Let θC ∈ K[x(C)] be the corresponding circuit polynomial, and consider
S(C) = S(k) × S(`). Then, there are two cases:
Case 1: If k, ` 6= 4, then there is n ∈ N, and a group homomorphism ν : S(C) → Z/2Z, such that
σθx = (−1)ν(σ) θx = γ(σ)θx ,
and γ(σ1 , σ2 ) = sgn(σ1 )a ·sgn(σ2 ) b , with a, b ∈ {0, 1} and sgn : S(·) → {−1, 1} the usual sign/parity
function.
Moreover, if char K = 0, there exists a set of monomials S ⊂ K[x(C)], such that
X
X
X
X
σ(M ) =
γ(σ)
σ(M ),
θC =
(−1)ν(σ)
σ∈S(C)
σ∈S(C)
M ∈S
M ∈S
which is unique up to σ-action on each element of S if S is chosen with minimal cardinality.
Case 2: If one of k, ` is equal to 4, there is n ∈ N, and a group homomorphism ν : G → Z/6Z,
such that
ν(σ)
σθC = ζ6 θx = γ(σ)θx ,
where ζ6 ∈ C is a sixth root of unity, and γ(σ1 , σ2 ) = f (σ1 )a · g(σ2 ) b , with a, b ∈ {0, 1, 2}, and each
of f , g is either the sign/parity function S(·) → {−1, 1}, or the triparity function S(4) → {1, ζ3 , ζ23 }.
Moreover, if char K = 0, there exists a set of monomials S ⊂ K[x(C)], such that
X
X
X
X
ν(σ)
σ(M ) =
γ(σ)
σ(M ),
θC =
ζ6
σ∈S(C)
σ∈S(C)
M ∈S
M ∈S
which is unique up to σ-action on each element of S if S is chosen with minimal cardinality.
Proof. Existence of a homomorphism ϕ : S(C) → K × follows from Proposition 6.1, therefore the
image ϕ(S(C)) must be a quotient S(k)/M × S(`)/N by normal divisors M and N . If k is not
divisible by 4, then M is either the identity group, or the alternating group, if k is divisible by
4, then M can additionally be the Klein four group; the analogue statement holds for ` and n.
Therefore, the image ϕ(S(C)) = ϕ(S(k) × S(`)) must be isomorphic to a quotient of (Z/2Z)2 if
k, ` 6= 4, to one of the groups (Z/2Z)2 , Z/2Z × Z/3Z if exactly one of k, ` is 4, and to one of the
groups (Z/2Z)2 , Z/2Z×Z/3Z, Z/3Z2 if k = ` = 4. Since the image must also be a finite subgroup of
K × , therefore a subgroup of the cyclotomic subgroup of K × , it can only be isomorphic to a subgroup
of Z/2Z if k, ` 6= 4, and to a subgroup of Z/6Z otherwise.
55
Corollary 6.4. As in Proposition 6.3, let C = [k] × [`] be a rectangular circuit. Then, >deg θC is the
multiset d ∗ C for some d ∈ N. Phrased differently, the non-zero entries in the top-degree-mask of C are
all equal.
Proof. This follows from Corollary 6.2 and the fact that for each c1 , c2 ∈ C there is σ ∈ S(C) =
S(k) × S(`) such that σc1 = c2 .
Remark 6.5 — In the situation of Proposition 6.3, one obtains a determinantal or permanental
formula if (σ, σ)(M ) = M for all M ∈ S and σ ∈ S(m), since then
X
X
X
X
(−1)ν(σ)
σ(M ) =
(sgn σ)a+b
σ(M )
σ∈G
M ∈S
σ∈S(n)
M ∈S
6.2. Constructing circuits with the (t, 1)-move
In this section, we study how classes of circuits that can be generated by the following inductive
move:
Definition 6.6. Let B be a bipartite graph including the edge i j and let t ∈ N be a parameter with
t ≤ r. The (t, 1)-move transforms B as follows:
Remove the edge i j. Add a copy of K1,t on new vertices, and connect each new vertex to B so
that:
1. The graph remains bipartite.
2. Each new vertex has degree r + 1.
3. There is a pair of vertices m, n such that (m j) and (in) are edges, and the sets of neighbors of
m and n are properly contained in the sets of neighbors of i and j respectively.
We define a partial (t, 1)-move as a typical (t, 1)-move, without the removal of the edge i j.
Edges and vertices contained in a graph before performing a move will be called old, while
edges added in a move will be called new.
The main result of this section is:
Proposition 6.7. Let C be a circuit graph in D(m × n, r) that is spanned by a basis graph. Then
applying a (t, 1)-move over any edge of C produces a circuit graph in D((m + 1) × (n + t), r) that is
spanned by a basis graph.
We defer the proof until we have established some required intermediate results.
Lemma 6.8. Let B be a basis graph in D(m × n, r) and let C be the graph obtained by performing
the partial (t, 1)-move on B. Then, C contains a unique circuit graph D. Additionally, the graph D
contains all the new edges.
Proof. C must contain some circuit, since a partial (t, 1)-move adds r + r t + 1 edges while the rank
rises only r + r t; therefore, we have defect −1.
Any circuit in C must contain a new edge, since the set of old edges is independent. However, if
any new edge is included, all new edges must be included in order to ensure that every new vertex
has degree ≥ r + 1. Otherwise, this circuit would contradict Theorem 4.1 (iv).
If two distinct circuits were contained in C, both would contain all new edges. Via the circuit
elimination axiom, we could then eliminate a new edge, finding a new circuit in C without all the
new edges, leading us to a contradiction.
56
Lemma 6.9. Let B be a basis graph of D(m× n, r), and let B 0 be the graph resulting from a (t, 1)-move
on B. Then, B 0 is a basis graph of D((m + 1) × (n + t), r).
Proof. A partial (t, 1)-move produces a unique circuit D, by Lemma 6.8. Let PD (X ) denote the
associated circuit polynomial.
If the edge i j is in the circuit, then removing it as part of the (t, 1)-move will result in an
independent set, which will be a basis graph by cardinality.
Suppose the edge i j is not in D. We refer to P
the parametric definition of the determinantal
r
matroid in Definition 3.1 (a). In this setting, X i j = k=1 Uik Vjk is excluded from the circuit.
The (t, 1)-move introduces a new set of transcendentals, call them {Um , Vn , . . . , Vn+t−1 }. Assume
m, n are the vertices stipulated by Definition 6.6.3. Specialize Um = Ui and Vn = Vj ; any polynomial
identities will also be true when specialized to a particular value. Now one of the new edges, known
to be in the circuit, takes the same value as X i j . Therefore, PD (X ) can be considered to include X i j
and exclude X mn . Because the neighbors of m, n are also neighbors of i, j respectively, the variables
in PD (X ) are contained in B. This gives us a circuit excluding a new edge, a contradiction.
Proof of Proposition 6.7. Let D denote the new graph. We need to show that:
1.
2.
3.
4.
Defect is preserved.
D is spanned by a basis graph.
D contains exactly one circuit C 0 .
All edges of D are contained in C 0
For (1), the change in rank is r + r t, while the move adds a net (r + 1)t + (r + 1 − t) − 1 = r + r t
edges; therefore, defect is preserved.
For (2), B = C \ i j is a basis graph of D(m × n, r). By Lemma 6.9, the (t, 1)-move produces a
basis graph from B; therefore, D is spanned by a basis graph of D((m + 1) × (n + t), r).
For (3), we reason as in the proof of Lemma 6.8. First, all new edges must be in any circuit
contained in D. If there is more than one such circuit, eliminating an edge between them produces
a circuit supported on a proper subgraph of C, which is a contradiction.
Finally, for (4), we have already shown that D contains all the new edges. We suppose by
contradiction that some old edge k` is excluded from C 0 . Then C \k` would be a basis of D(m×n, r).
By Lemma 6.9, the (t, 1)-move constructs a basis, in contradiction with the hypothesis that D \ k`
contains the circuit.
Remark 6.10 — One may attempt to generalize the (t, 1)-move to a (s, t)-move adding a copy of
Ks,t and connecting to the circuit graph giving each new vertex degree r + 1. However, for s, t > 1,
the rank changes by r(s + t) while a net (r +1)(s + t)−st −1 edges are added; so the defect changes
by s + t − st − 1. For s, t > 1, the defect is not preserved as s + t − st − 1 is strictly negative. An
example of the failure is given below:
•
•
•
•
◦
◦
•
•
•
◦
• • • ◦ •
• • • ◦ •
• • • ◦ ◦
• • • ◦ ◦
• ◦ ◦ • •
• ◦ ◦ • •
◦ • ◦ • •
◦ • ◦ • •
While the upper left 3 × 3 minor is a circuit of D(3 × 3, 2), the constructed graph is a basis of
D(5 × 5, 2).
57
Remark 6.11 — Observe that starting with an elementary circuit and applying the (r, 1)-move k
times creates a circuit of signature (r + k, r k + 1). For k = m − r, this is equal to (m, r(m − r) + 1),
which is precisely the bound given in Proposition 4.28, proving it to be strict for the case of the
determinantal matroid.
7. A Dictionary of Circuits
In this section, we aim to describe the set of circuits of algebraic graph matroids in as much detail as
possible. In particular, the different circuit graphs and their symmetries, and the number of circuits
in the algebraic matroid corresponding to each graph. Even in cases where a full combinatorial
understanding is inaccessible, we will be interested in the total number of circuits.
7.1. Symmetries and Number of Circuits for Bipartite Graph Matroids
We introduce some notation for counting circuits, for the two different ways of counting - as sets,
and as graphs.
Notation 7.1. Let M be a bipartite graph matroid limit with canonical symmetric group G ∼
= S(N ×
N). We denote by
Ck,` (M) := {C/G : C ∈ C(M), # vsupp C = (k, `)}
the set of (non-isomorphic) circuit graphs of M with signature (k, `). We will denote its cardinality
by ck,` (M) := #Ck,` (M). Where clear from the context, the dependence on M will be omitted.
Proposition 7.2. Let Mµν be an injective complex of bipartite graph matroids, with bipartite graph
matroid limit M, having symmetric group G ∼
= S(N × N). Then, the following formulae hold:
(i)
#C(Mm,n /G) =
m X
n
X
ck,` (M).
k=1 `=1
(ii)
m X
n
X
#C(Mmn ) =
k=1 `=1
1
X
C∈Ck,` (M)
#S(C)
m! · n!
(m − k)!(n − `)!
.
Proof. (i) follows from the fact that C(Mm,n /G) is, by definition, a disjoint union of all Ck,` (M) with
k ≤ m, ` ≤ n.
(ii) Pick a circuit C ∈ C(Mm,n ) with signature (k, `), let Gmn be the canonical symmetry group of
Mmn . Consider the stabilizer subgroup Stab(C) of Gmn . By definition of support, σ vsupp(C) ∈
vsupp(C) for all σ ∈ Stab(C). Therefore, Stab(C) = S1 × S2 , with S2 ∼
= S([m − k] × [n − `), and
S1 is a subgroup of S(vsupp(C)) isomorphic to the automorphism group S(C/G). In particular,
# Stab(C) = #S1 · #S2 = #S(C/G) · (m − k)!(n − `)! Applying the orbit-stabilizer theorem to the
Gmn -action, we obtain
#(C/Gmn ) =
#Gmn
# Stab(C)
=
1
·
m!n!
S(C/G) (m − k)!(n − `)!
.
Noting that Ck,` (M) contains exactly orbits C/Gmn of signature (k, `), a summation over all circuit
graphs and all signatures yields the formula.
Remark 7.3 — From the theory in the previous sections, it follows that many of the Ck,` (M) are
empty, therefore the summations in Proposition 7.2 can be restricted. Namely:
58
(i) By Theorem 4.1 (ii), there are no circuits of signature (k, `) with k ≤ ρ1 or ` ≤ ρ2 .
(ii) By Theorem 4.9,
k ≤ α − ρ1 · κ1 + ρ2 · (` − κ2 ) + 1
` ≤ α − ρ2 · κ2 + ρ1 · (k − κ1 ) + 1.
In our examples, we will often be able to put tighter bounds on the summation.
Furthermore, the summation of Proposition 7.2 (ii) can be seen as in analogy to Proposition 7.2 (i). The combinatorics is completely captured in the coefficient in square brackets, which
can be interpreted as the average fraction a circuit graph contributes to the number of circuits of a
completely unsymmetric graph. Therefore, we introduce a notational abbreviation:
Notation 7.4. Let M be a bipartite graph matroid limit. We will write
1
X
βk,` (M) :=
C∈Ck,` (M)
#S(C)
.
Where clear from the context, the dependence on M will be omitted.
Remark 7.5 — With this notation, formula (ii) of Proposition 7.2 becomes
#C(Mmn ) =
m X
n
X
βk,` ·
k=1 `=1
m! · n!
(m − k)!(n − `)!
and therefore the graphless analogue of formula (i), since the factor
as coming from the lost graph symmetry.
,
m!·n!
(m−k)!(n−`)!
can be interpreted
7.2. The Determinantal Matroid
We determine some circuits and their associated invariants of the determinantal matroids. For rank
one, we can completely characterize those:
Theorem 7.1. Consider the bipartite graph matroid M = D(N × N, 1). Denote by C2n , n ∈ N the
(bipartite) cycle of length 2n, containing n vertices of each of the two classes. Then:
(i) C(M) = {C2n : n ∈ N}
(ii) Ck,` (M) = ∅ if k 6= `
(iii) Ck,k (M) = {C2k }
(iv) ck,` (M) = δk,`
(v) βk,` (M) = δk,` · (4k)−1
The top-degree is all ones for the variables in the circuit support.
Proof. (i) follows from Proposition 2.6.45 (i) in [17].
(ii) and (iii) follow from the fact that the signature of C2k is (k, k).
(iv) follows from (ii) and (iii).
(v) follows from the fact that S(C2k ) is isomorphic to the dihedral group D2k , whose cardinality is
4k.
59
Proposition 7.6 (Rank 2). The list below contains all circuit graphs with signature k ≤ l ≤ 5, along
with S(X ). Some circuits are the transpose of circuits in this list; this is indicated in the chart. As
particular consequences,
n
n
n
n
#C(D(3×n, 2)) =
,
#C(D(4×n, 2)) = 4
+96
+840
,
#C(D(5×5, 2)) = 65, 650.
3
3
4
5
60
Signature
(3, 3)
(4, 4)
(4, 5)
Mask
1
1
1
1 1 1
1 1 1
1 1 1 ·
1 · 1 1
1 1 2 1
· 1 1 1
1
1
1
·
·
1 1 · 1 1
1 1 1 1 1
· · 1 1 1
(5, 5)
S(X ) = S(3) × S(3)
deg(θX ) = 3
S(X ) = S(3)
deg(θX ) = 5
S(X ) = (Z/2Z)3
deg(θX ) = 5
*and transpose
1
1
1
1
·
1
1
1
1
·
1
1
·
·
1
·
·
1
1
1
·
·
1
1
1
S(X ) = (Z/2Z)4
deg(θX ) = 5
1
1
1
·
·
1
·
1
1
·
1
1
3
1
1
·
1
1
·
1
·
·
1
1
1
S(X ) = Z/4Z × Z/2Z
deg(θX ) = 7
1
1
1
·
·
1
·
1
·
1
2
1
2
1
1
·
1
1
1
·
1
·
·
1
1
S(X ) = Z/2Z
deg(θX ) = 7
1
1
1
·
·
1
2
1
1
·
1
2
·
2
1
·
1
·
1
1
·
·
1
1
1
S(X ) = 1
deg(θX ) = 7
2
2
2
·
·
2
·
2
2
2
2
2
·
2
2
·
2
2
2
·
·
2
2
·
2
S(X ) = Z/3Z × (Z/2Z)2
deg(θX ) = 12
*and transpose
2
2
2
·
2
2
2
·
1
1
·
1
1
·
1
1
·
1
1
1
S(X ) = Z/2Z × S(3)
deg(θX ) = 7
1
1
1
·
·
1
1
1
·
·
1
1
·
1
1
·
·
1
1
1
·
·
1
1
1
S(X ) = (Z/2Z)5
deg(θX ) = 5
1
1
1
·
·
1
2
·
1
1
1
2
·
1
1
·
1
1
1
·
·
1
1
·
1
S(X ) = Z/3Z × (Z/2Z)2
deg(θX ) = 7
1
2
2
·
1
1
2
2
·
1
1
1
·
1
·
·
·
1
1
1
·
1
1
1
·
S(X ) = (Z/2Z)2
deg(θX ) = 7
1
1
1
·
·
1
·
1
1
·
1
1
·
2
2
·
1
2
3
2
·
·
2
2
2
S(X ) = 1
deg(θX ) = 9
*and transpose
*and transpose
*and transpose
Table 2: Circuits of the matroid D(5 × 5, 2). *Degree of the circuit polynomial is included for reference.
Proof. For k = 3, the maximal minor is the only circuit by Theorem 4.4. As for k = 4, the same
argument limits us to 4 or 5 columns with one element in each column, after which an explicit
check yields the circuits above.
Let X C be the complement of a circuit X in Km,n . By Theorem 4.4, all vertices of X have degree
at least 3, so the vertices of X C have degree at most 2; therefore, X C must be a union of paths and
cycles. Rank of the matroid implies that X can contain at most r(m + n − r) + 1 = 17 elements,
which means X C has at least 8 edges. All that remains is to list all graphs X C fitting these criteria
and manually check whether X is a circuit using Macaulay2 [14].
61
To compute S(X ), we compute the automorphism group of the complement, which has factors:
1. (Z/kZ × Z/2Z) for every 2k-cycle,
2. (Z/2Z) for every even path.
3. S(k) for each set of k identical components (identical, including the same coloring of vertices).
We note that the graphs listed have transpose symmetry if and only if the number of 2k-paths with
row vertex endpoints is the same as the number of 2k-paths with column vertex endpoints, for all
k. The graphs marked with *and transpose are those where (i) the signature is asymmetric, or (ii)
the graph fails the condition on 2k-paths.
We can use similar techniques to compute the circuits for a few small signatures with r = 3.
Proposition 7.7 (Rank 3). The list below contains all circuit graphs X with signature k ≤ 5, along
with S(X ). As particular consequences,
n
n
n
n
n
#C(D(4 × n, 3)) =
, #C(D(5 × n, 3)) = 5
+ 600
+ 13, 320
+ 65, 100
.
4
4
5
6
7
Signature
Mask
(4, 4)
(5, 5)
(5, 6)
1
1
1
1
·
(5, 7)
1
1
1
1
1
1
1
1
1
·
1
1
1
1
1
1
1
1
1
2
2
1
1
1
1
1
S(X ) = S(4) × S(4)
deg(θX ) = 4
·
1
1
1
1
1
1
2
2
1
S(X ) = S(3) × (Z/2Z)3
deg(θX ) = 7
S(X ) = (Z/2Z)4
deg(θX ) = 7
S(X ) = S(5)
deg(θX ) = 15
1
1
1
1
·
1
1
1
1
·
1
·
1
1
1
1
1
2
2
1
·
1
1
1
1
·
1
1
1
1
3
3
3
3
3
2
2
2
2
·
2
2
2
·
2
2
2
·
2
2
2
·
2
2
2
·
2
2
2
2
S(X ) = S(3) × (Z/2Z)4
deg(θX ) = 7
S(X ) = S(3) × (Z/2Z)2
deg(θX ) = 10
1
1
1
1
·
1
1
1
1
·
1
·
1
1
1
1
·
1
1
1
·
1
1
1
1
·
1
1
1
1
·
1
1
1
1
1
1
1
1
·
1
1
1
1
·
1
1
·
1
1
1
·
1
1
1
·
2
2
2
2
·
2
2
2
2
·
2
2
2
2
62
1
2
2
2
3
1
1
1
·
1
1
1
·
1
1
·
2
2
2
2
1
·
1
1
1
·
2
2
2
2
S(X ) = Z/2Z × S(3)
deg(θX ) = 10
1
1
1
1
·
1
1
1
1
·
1
1
1
1
·
1
·
1
1
1
·
1
1
1
1
·
1
1
1
1
·
1
1
1
1
S(X ) = (S(3))2 × (Z/2Z)2
deg(θX ) = 7
1
1
1
1
·
1
1
1
·
1
1
1
·
1
1
1
·
1
1
1
·
3
3
3
3
·
3
3
3
3
·
3
3
3
3
S(X ) = S(4) × S(3)
deg(θX ) = 13
1
1
1
1
·
2
2
·
2
2
2
2
·
2
2
2
·
2
2
2
2
·
2
2
2
·
2
2
2
2
·
2
2
2
2
S(X ) = S(3) × (Z/2Z)3
deg(θX ) = 13
2
2
2
2
·
2
2
2
·
2
2
2
·
2
2
3
·
3
3
3
3
·
3
3
3
·
3
3
3
3
·
3
3
3
3
S(X ) = S(3) × (Z/2Z)3
deg(θX ) = 18
Table 3: Circuits of the matroid D(5 × n, 3). *Degree of the circuit polynomial is included for reference.
Proof. Again, we let X C be the complement of a circuit X in K5,n . All vertices of X have degree at
least 4, so the column vertices of X C have degree at most 1. This means that X C is a union of star
graphs, with central vertex a row vertex, and isolated vertices. Rank of the matroid implies that X
can contain at most r(m + n − r) + 1 = 3n + 7 elements, which means X C has at least 2n − 7 edges.
Checking suitable graphs amounts to checking partitions of at most 7, with at most 5 nonzero parts.
The graphs that are verified by Macaulay2 [14] as circuits are listed in the table.
For S(X ), we again compute S(X C ). As mentioned above, the unions of star graphs map canona
a
ically to partitions, by sending each star to the number of its edges or leaves. Let λ = λ11 . . . λnn be
the partition corresponding to X C . The stabilizer of X C has factors:
Qn
ai
1.
i=1 (S(λi )) × S(ai ), permuting the leaves in each star and permuting the stars of the same
size.
2. S(k) × S(l), where k is the number of isolated row vertices of X C and l is the number of
isolated column vertices.
Remark 7.8 — We record ck,l and βk,l computed for the matroid D(N × N, r), for some r.
For r = 1: ck,l = δk,l and βk,l = δk,l /4k.
For r = 2 and 3, the following tables list the first few values:
(k, l)
(3, 3)
(4, 4)
r =2:
(4, 5)
(5, 4)
(5, 5)
ck,l
1
1
2
2
12
βk,l
1/36
1/6
5/24
5/24
127/32
(k, l)
(4, 4)
r = 3 : (5, 5)
(5, 6)
(5, 7)
For all r: c r+1,r+1 = 1 and β r+1,r+1 = 1/(r + 1)!2
c r+2,r+2 = 1 and β r+2,r+2 = 1/(3! · (r − 1)!2 ).
ck,l
1
1
3
6
βk,l
1/576
1/24
37/240
31/288
7.3. The Bipartite Rigidity Matroid
Proposition 7.9. Let M(n) = CM(m × n, r). The formula from Proposition 7.2 counts circuits of M(n)
with the tighter summation bounds of:
1
r +1
r +1
r + 1 ≤ k ≤ m;
+ k − 1 ≤ l ≤ rk −
+ 1.
r
2
2
Proof. The lower bound on k – CM(m × n, r) has ρ = r, from Proposition 4.29.
The upper bound on l is from Proposition 4.31. The lower bound again uses the transpose
symmetry, switching k and l in the same inequality.
The defining ideal for the bipartite rigidity matroid involves many more variables; as such, it
is computationally much heavier. We computed circuits via the linear realization for some small
examples; therefore, top-degrees are excluded.
63
Proposition 7.10 (Rank 1). The matroid CM(m× n, 1) is the graphic matroid on Km,n . The top degree
for a cycle of size k has all appearing coordinates bounded below by 2k−1 .
Proof. No proper subset is dependent, by Proposition 4.29. The rank of the matroid is 2k − 1, for
k ≥ 2 on the induced subgraph by Remark 3.17 the set itself is dependent, thus a circuit.
By the same token, any circuit has vertices of degree ≥ 2, implying that it contains a cycle, which
we already know to be a circuit. This means the cycles are the only circuits. The lower bound on the
top-degree can be given by the number of generic solutions of a 1-dimensional rigidity framework:
after laying down the first edge, we have k − 1 vertices from which the next edge can point left or
right: This gives us 2k−1 real solutions generically, thus bounding each coordinate of the top-degree
below by 2k−1 .
Proposition 7.11 (Rank 2). The circuit set of CM(m × n, 2) is the set of bipartite graphs constructed
from K4 via edge-splits and circuit gluing. In particular, we find the following masks for CM(4 × n, 2),
along with automorphism groups:
Signature
Mask and Stabilizer
· · · ·
· · · ·
· · · ·
(3, 4)
S(X ) = S(3) × S(4)
(4, 4)
·
·
·
·
·
·
·
·
· ·
· ·
· ·
· ·
·
·
·
·
·
·
·
·
·
·
·
·
S(X ) = (S(2))2
· ·
· ·
· ·
· ·
(4, 6)
·
·
·
·
S(X ) = (S(2))3
(4, 5)
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
· ·
· ·
· ·
· ·
·
·
·
·
·
·
·
·
·
·
·
·
· ·
· ·
· ·
· ·
S(X ) = (S(2))4
· ·
· ·
· ·
· ·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
S(X ) = (S(3))2 × (S(2))2
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
S(X ) = (S(2))4
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
S(X ) = (S(2))2
S(X ) = (S(3))2
·
·
·
·
·
·
·
·
·
·
·
·
S(X ) = S(4)
· ·
· ·
· ·
· ·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
S(X ) = (S(2))3 × S(3)
Table 4: Circuits of the matroid CM(4 × n, 2).
Proof. Degree considerations restrict our attention to star graphs as circuit complements, as in the
proof of Proposition 7.7. Computation of stabilizers also follows the logic there.
64
7.4. Symmetries and Number of Circuits for Graph Matroids
The invariants introduced in section 7.1 for bipartite graph matroids can be readily adapted to the
graph matroid case:
Notation 7.12. Let M be a graph matroid limit with canonical symmetric group G ∼
= S(N × N). We
denote by
Ck (M) := {C/G : : C ∈ C(M), # vsupp C = (k)}
the set of (non-isomorphic) circuit graphs of M with signature k. We will denote its cardinality by
ck (M) := #Ck (M). We will furthermore write
βk (M) :=
1
X
C∈Ck (M)
#S(C)
.
Where clear from the context, the dependence on M will be omitted.
Proposition 7.13. Let Mν be an injective complex of bipartite graph matroids, with bipartite graph
matroid limit M, having symmetric group G ∼
= S(N). Then, the following formulae hold:
(i)
#C(Mm,n /G) =
n
X
ck
k=1
.
(ii)
#C(Mn ) =
n
X
k=1
βk ·
n!
(n − `)!
Proof. The proof is completely analogous to that of Proposition 7.2.
Remark 7.14 — There are two major differences when enumerating the circuits of a graph matroid
sequence as opposed to a bipartite graph sequence:
1. The symmetry group. We have one set of vertices to permute, as opposed to two sets that
we permute separately. Instead of S(N × N) we now use S(N).
2. Bound on circuit support. In the asymmetric case, when m is fixed, the signature of any
circuit in the sequence is bounded by a constant. Here, there may be no such bound, so
specific values of n may be accessible while no general formula exists.
7.5. The Symmetric Determinantal Matroid
Proposition 7.15 (Rank 1). The circuits of the matroid D (5 × 5, 1) are given by the graphs in the
following table, recorded with stabilizer, degree and a circuit in the asymmetric determinantal matroid
with the desired symmetrization.
Signature
2
Graph
Mask of Asymmetric Circuit
1 1
1 1
1
1
2
65
Stabilizer & Degree
S(X ) = Z/2Z
deg(θX ) = 2
1
1 1
1 ·
· 1
· 1
· 1
· ·
1
2
2
3
1
1
1
1
1 1 · ·
1 · 1 ·
· 1 · 1
· · 1 1
· · 1 1
· · 1 1
· · · ·
· · · ·
1
1
·
·
1 · · 1
· 1 · 1
· · · ·
1
1
·
·
·
1 · 1 · ·
· 1 · 1 ·
· · 1 · 1
· · · 1 1
· · 1 · 1
· · · · ·
· 1 · · 1
· 1 1 · ·
· · · · ·
1 1 · · ·
1 · 1 · ·
· 1 · 1 ·
· · · · ·
· · 1 1 ·
1
·
1
·
·
1 · · 1 ·
· · · · ·
· · · · ·
· · 1 1 ·
1
1
2
4
2
2
1
1
1
1
1
1
2
1
1
1
1
2
2
2
2
5
1
1
1
1
1
1
1
1
2
2
1
1
1
1
1
1
1
·
1
1
1
1
·
1
S(X ) = Z/2Z
deg(θX ) = 3
S(X ) = Z/2Z
deg(θX ) = 2
S(X ) = Z/2Z
deg(θX ) = 4
S(X ) = D4
deg(θX ) = 2
S(X ) = Z/2Z
deg(θX ) = 3
S(X ) = (Z/2Z)3
deg(θX ) = 5
S(X ) = Z/2Z
deg(θX ) = 3
S(X ) = Z/2Z
deg(θX ) = 4
S(X ) = Z/2Z
deg(θX ) = 3
Table 5: Circuits of the matroid D (5 × 5, 1).
One may observe that each of these graphs may be constructed inductively from the three graphs
that are symmetrizations of 2 × 2 minors. The top-degree can also be described (in these cases) as
the size of the pre-image of each variable under symmetrization of a circuit in D(m × n, 1).
7.6. The Rigidity Matroid
Proposition 7.16 (Rank 1). The matroid CM (n × n, 1) is the graphic matroid on Kn . The top degree
for a cycle of size k has all appearing coordinates equal to 2k−1 .
Proof. The proof follows as in the bipartite case; the only difference is the ground set.
66
The bases of the matroid CM (n × n, 2) are characterized by Laman’s theorem [20]. The set of
circuits were more recently described in [9], as being constructed from K4 via a finite set of moves.
We use these moves, along with computation of the matroid in its linear realization, to list some
circuits for rank 2:
Proposition 7.17 (Rank 2). The set of circuits of CM (6×6, 2) is given in the table below. In particular,
#C(CM (6 × 6, 2)) = 642.
Signature
Graph and Stabilizer
4
S(X ) = S(4)
5
S(X ) = D4
6
S(X ) = S(3) × S(2)
S(X ) = S(2) × D4
S(X ) = S(2)
S(X ) = D5
Table 6: Circuits of the matroid D (5 × 5, 1).
References
[1] Matthias Aschenbrenner and Christopher J. Hillar.
Finite generation of symmetric ideals.
Trans. Amer. Math. Soc., 359(11):5171–5192, 2007.
ISSN 00029947. doi: 10.1090/S0002-9947-07-04116-5. URL http://dx.doi.org/10.1090/
S0002-9947-07-04116-5.
[2] E. D. Bolker and B. Roth. When is a bipartite graph a rigid framework? Pacific J. Math., 90
(1):27–44, 1980. ISSN 0030-8730. URL http://projecteuclid.org/getRecord?id=
euclid.pjm/1102779115.
[3] Ciprian Borcea and Ileana Streinu.
The number of embeddings of minimally rigid
graphs. Discrete Comput. Geom., 31(2):287–303, 2004. ISSN 0179-5376. doi: 10.1007/
s00454-003-2902-0. URL http://dx.doi.org/10.1007/s00454-003-2902-0.
[4] Ciprian S. Borcea. Point configurations and Cayley-Menger varieties. Preprint, arXiv:0207110,
2002. URL http://arxiv.org/abs/math/0207110.
[5] Henning Bruhn, Reinhard Diestel, Matthias Kriesell, Rudi Pendavingh, and Paul Wollan.
67
Axioms for infinite matroids. Adv. Math., 239:18–46, 2013. ISSN 0001-8708. doi:
10.1016/j.aim.2013.01.011. URL http://dx.doi.org/10.1016/j.aim.2013.01.011.
[6] Winfried Bruns and Udo Vetter. Determinantal Rings. Springer-Verlag New York, Inc., 1988.
[7] Emmanuel J. Candès and Benjamin Recht. Exact matrix completion via convex optimization. Found. Comput. Math., 9(6):717–772, 2009. ISSN 1615-3375. doi: 10.1007/
s10208-009-9045-5. URL http://dx.doi.org/10.1007/s10208-009-9045-5.
[8] Emmanuel J. Candès and Terence Tao. The power of convex relaxation: near-optimal matrix
completion. IEEE Trans. Inform. Theory, 56(5):2053–2080, 2010.
[9] L. Chavez, L. Moshe, and W. Whiteley. Bases and circuits for 2-rigidity: constructions via tree
coverings. Preprint, Department of Mathematics and Statistics, York University, 2003, 2003.
URL http://wiki.math.yorku.ca/images/b/b8/Paper2.pdf.
[10] Michel Marie Deza and Monique Laurent. Geometry of cuts and metrics, volume 15 of Algorithms and Combinatorics. Springer-Verlag, Berlin, 1997. ISBN 3-540-61611-X.
[11] Michael Dobbins. Realizability of polytopes as a low rank matrix completion problem.
Preprint, arXiv:1012.3905, 2011. URL http://arxiv.org/abs/1012.3905.
[12] A. Dress and L. Lovász. On some combinatorial properties of algebraic matroids. Combinatorica, 7(1):39–48, 1987.
[13] Jack Graver, Brigitte Servatius, and Herman Servatius. Combinatorial rigidity, volume 2 of
Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 1993. ISBN
0-8218-3801-6.
[14] Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system for research in
algebraic geometry. Available at http://www.math.uiuc.edu/Macaulay2/.
[15] Christopher J. Hillar and Seth Sullivant. Finite Gröbner bases in infinite dimensional polynomial rings and applications. Adv. Math., 229(1):1–25, 2012. ISSN 0001-8708. doi:
10.1016/j.aim.2011.08.009. URL http://dx.doi.org/10.1016/j.aim.2011.08.009.
[16] Raghunandan H. Keshavan, Andrea Montanari, and Sewoong Oh. Matrix completion from a
few entries. IEEE Trans. Inform. Theory, 56(6):2980–2998, 2010.
[17] Franz J. Király, Louis Theran, Ryota Tomioka, and Takeaki Uno. The algebraic combinatorial
approach for low-rank matrix completion. CoRR, abs/1211.4116, 2012.
[18] Martin Kreuzer and Lorenzo Robbiano. Computational commutative algebra. 2. SpringerVerlag, Berlin, 2005. ISBN 978-3-540-25527-7; 3-540-25527-3.
[19] Ronald E. Kutz. Cohen-Macaulay rings and ideal theory in rings of invariants of algebraic
groups. Trans. Amer. Math. Soc., 194:115–129, 1974. ISSN 0002-9947.
[20] G. Laman. On graphs and rigidity of plane skeletal structures. J. Engrg. Math., 4:331–340,
1970. ISSN 0022-0833.
[21] Seok Hyeong Lee and Ravi Vakil. Mnëv-Sturmfels universality for schemes.
arXiv:1202.3934, 2012. URL http://arxiv.org/abs/1202.3934.
68
Preprint,
[22] B. Lindström. Matroids algebraic over F (t) are algebraic over F . Combinatorica, 9(1):107–
109, 1989. ISSN 0209-9683. doi: 10.1007/BF02122691. URL http://dx.doi.org/10.
1007/BF02122691.
[23] L. Lovász. Flats in matroids and geometric graphs. In Combinatorial surveys (Proc. Sixth British
Combinatorial Conf., Royal Holloway Coll., Egham, 1977), pages 45–86. Academic Press, London, 1977.
[24] L. Lovász and Y. Yemini. On generic rigidity in the plane. SIAM J. Algebraic Discrete Methods,
3(1):91–98, 1982. ISSN 0196-5212. doi: 10.1137/0603009. URL http://dx.doi.org/
10.1137/0603009.
[25] N. E. Mnëv. The universality theorems on the classification problem of configuration varieties
and convex polytopes varieties. In Topology and geometry—Rohlin Seminar, volume 1346 of
Lecture Notes in Math., pages 527–543. Springer, Berlin, 1988. doi: 10.1007/BFb0082792.
URL http://dx.doi.org/10.1007/BFb0082792.
[26] David Mumford. The Red Book of Varieties and Schemes. Lecture Notes in Mathematics.
Springer-Verlag Berlin Heidelberg, 1999. ISBN 054063293X.
[27] James Oxley. Matroid theory, volume 21 of Oxford Graduate Texts in Mathematics. Oxford
University Press, Oxford, second edition, 2011. ISBN 978-0-19-960339-8.
[28] Jürgen Richter-Gebert. Realization spaces of polytopes, volume 1643 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1996. ISBN 3-540-62084-2.
[29] Rüdiger Schmidt. On the existence of uncountably many matroidal families. Discrete Mathematics, 27(1):93 – 97, 1979.
[30] Peter W. Shor. Stretchability of pseudolines is NP-hard. In Applied geometry and discrete
mathematics, volume 4 of DIMACS Ser. Discrete Math. Theoret. Comput. Sci., pages 531–554.
Amer. Math. Soc., Providence, RI, 1991.
[31] Amit Singer and Mihai Cucuringu. Uniqueness of low-rank matrix completion by rigidity
theory. SIAM J. Matrix Anal. Appl., 31(4):1621–1641, 2009/10. ISSN 0895-4798. doi: 10.
1137/090750688. URL http://dx.doi.org/10.1137/090750688.
[32] Amit Singer and Mihai Cucuringu. Uniqueness of low-rank matrix completion by rigidity
theory. Siam J Matrix Anal Appl, 31(4):1621–1641, 2010.
[33] Ileana Streinu and Louis Theran. Slider-pinning rigidity: a Maxwell-Laman-type theorem. Discrete Comput. Geom., 44(4):812–837, 2010. ISSN 0179-5376. doi: 10.1007/
s00454-010-9283-y. URL http://dx.doi.org/10.1007/s00454-010-9283-y.
[34] Bernd Sturmfels. Solving systems of polynomial equations, volume 97 of CBMS Regional Conference Series in Mathematics. Published for the Conference Board of the Mathematical Sciences,
Washington, DC, 2002. ISBN 0-8218-3251-4.
[35] Shin-ichi Tanigawa. Generic rigidity matroids with Dilworth truncations. SIAM J. Discrete
Math., 26(3):1412–1439, 2012. ISSN 0895-4801. doi: 10.1137/100819473. URL http:
//dx.doi.org/10.1137/100819473.
[36] Bartel L. van der Waerden. Modern Algebra, volume 1. Julius Springer, Berlin, second edition,
1937.
69
[37] Walter Whiteley. Some matroids from discrete applied geometry. In Matroid theory (Seattle,
WA, 1995), volume 197 of Contemp. Math., pages 171–311. Amer. Math. Soc., Providence, RI,
1996. doi: 10.1090/conm/197/02540. URL http://dx.doi.org/10.1090/conm/197/
02540.
70
| 0math.AC
|
A General and Yet Efficient Scheme for Sub-Nyquist
Radar Processing
Shengyao Chen, Feng Xi, Zhong Liu*
Department of Electronic Engineering, Nanjing University of Science and Technology, China
Abstract—We study the target parameter estimation for sub-Nyquist pulse-Doppler radar. Several past works have
addressed this problem but either have low estimation accuracy for off-grid targets, take large computation load, or lack
versatility for analog-to-information conversion (AIC) systems. To overcome these difficulties, we present a general
and efficient estimation scheme. The scheme first formulates a general model in the sense that it is applicable to all
AICs regardless of whether the targets are on or off the grids. The estimation of Doppler shifts and delays is performed
sequentially, in which the Doppler estimation is formulated into a spatial spectrum estimation problem and the delay
estimation is decomposed into a series of compressive parameter estimation problems with each corresponding to an
estimated Doppler shift. By the sequential and decomposed processing, the computational complexity is substantially
reduced, and by the parametric estimation techniques, the high accurate estimation is obtained. Theoretical analyses
and numerical experiments show the effectiveness and the correctness of the proposed scheme.
Keywords: Compressive sampling, Pulse-Doppler radar, Delay-Doppler estimation, Off-grid target.
I. INTRODUCTION
We consider target parameter estimation in a sub-Nyquist pulse-Doppler radar via analog-to-information conversion
(AIC) systems. The problem has received extensive attention in recent years and several methods have been reported
[1-6]. Among these reported methods, some estimate the target parameters simultaneously by sparse reconstruction
[1-3], some perform sequential estimation of the delays and Doppler shifts [4-6], some assume that the targets are
located on the predefined delay-Doppler grids [1, 2, 4, 5], and some are developed on specific AIC systems [4-6].
These methods either have low estimation accuracy for off-grid targets, take large computation load, or lack versatility
for AIC systems (See [7] for a review). In general, the sequential methods have computational advantages and are more
suitable to practical radar applications.
Recently, we developed a general sequential delay-Doppler (GeSeDD) estimation scheme in [7]. GeSeDD is
versatile in suiting for any AIC system and has the advantages of high estimation accuracy, high resolution and
moderate-size computational complexity. This paper continues our study on the development of general sequential
*
Corresponding author. Email address: [email protected]
1
estimation method with the aim to reduce the computational overhead. Instead of grouping the targets into classes with
each corresponding to a distinct time delay [7], we classify the targets into groups with each corresponding to a distinct
Doppler shift. For the universal applicability, we formulate a compressive measurement model regardless of AIC
systems and on-grid and/or off-grid targets. By the new model, we find that the Doppler shifts can be estimated
independently of the time delays by parametric direction-of-arrival (DOA) estimation techniques [8]. With the
estimated Doppler shifts, we establish a tactics that further decomposes the estimation of time delays into a series of
compressive parameter estimation problems with each corresponding to a distinct Doppler shift. The compressive
parameter estimation techniques [3, 9-11] can be taken to estimate the time delays. By the sequential and decomposed
processing, combined with parametric estimation techniques, the novel estimation scheme has high computational
efficiency and high estimation accuracy. Theoretical analyses and simulation results confirm our assertions.
The rest of this paper is organized as follows. The problem formulation is defined in Section II. The estimation
scheme of Doppler shifts and delays is introduced in Section III. Section IV presents the analyses on computational
complexity. The experiments are simulated in Section V. Conclusion is given in Section VI.
II. PROBLEM FORMULATION
The data model we will develop is similar to that in [7], except that we classify the targets into groups with each
corresponding to a distinct Doppler shift. Consider a radar scene consisting of K non-fluctuating moving point targets.
Assume that there are K v ( K v L ) targets having distinct Doppler shifts and each Doppler shift vi is associated with
K v ,i different delays ij ( j 1, , K v ,i ). For a co-located pulse-Doppler radar transmitting L pulses in a coherent
processing interval (CPI), the complex baseband echo corresponding to the l -th pulse ( 0 l L 1 ) is given by
Kv
K v ,i
Kv
K v ,i
i 1
j 1
i 1
j 1
r l t e j2 i t ij g t lT ij e j2 i lT ij g t lT ij ,
t lT ,(l 1)T
(1)
where T is the pulse repetition interval (PRI), g t is the complex envelop of transmitting pulses with bandwidth
B and pulse width Tp ( Tp T ) and ij is the complex gain of the target with the delay-Doppler pair
ij
, i .
The approximation follows from the stop-and-hop assumption [12]. For unambiguous estimation, the delays and
Doppler shifts satisfy ij 0, T Tp and vi 1 2T ,1 2T .
2
Under parametric waveform dictionary [13] with the delay ij ,
, t , t g t , i 1,, K ,
ij
ij
ij
v
j 1, , K v ,i , the echo signal r l t can be expressed as
Kv
K v ,i
i 1
j 1
r l t e j2 i lT ij ij , t lT ,
t lT 0, T
(2)
In the presence of noise, the received baseband signal is given by
Kv
K v ,i
i 1
j 1
r1l t e j2 i lT ij ij , t lT n l t ,
t lT 0, T
(3)
where n l t is assumed to be a lowpass complex Gaussian process with power spectral density N 0 and bandwidth
B , and all of the n l t ( 1 l L ) are independently and identically distributed (i.i.d.).
For sub-Nyquist radars, the received signal r1l t is sampled by an AIC system to attain compressive
measurements [14]. Denote r l r l lT , r l lT Tnyq , , r l lT ( N 1)Tnyq N 1 and nl n l lT , n l lT Tnyq , ,
T
n l lT ( N 1)Tnyq N 1 as the Nyquist-rate sampling vectors of r l t and n l t in a PRI t lT ,(l 1)T ,
T
respectively, where the superscript
T
represents transposition. The compressive measurement vector slcs M 1 of
the signal r1l t obtained from any AIC system can be represented as
slcs M r l nl Mr l nlcs
(4)
T
where M M N ( M N ) is the measurement matrix and nlcs Mnl ncsl [1], ncsl [2], , ncsl [ M ] M 1 is the
compressive measurement vector of the noise n l t . The ncsl [m] ( 1 m M ) is an i.i.d. complex Gaussian process
with zero-mean and variance NN 0 B M [15]. In the formulation of (4), we are not confined to any specific AIC system.
Then the matrix M depends on the assumed AIC system. In this sense, (4) is a general compressive measurement
model for any sub-Nyquist radar.
Let
ψ( ij ) ij ,0 , ij , Tnyq , , ij ,( N 1)Tnyq N 1
T
be the Nyquist-rate sampling vector of the
parametric atom ij , t . The compressive vector slcs can be explicitly described by
K
K v ,i
K
i 1
j 1
i 1
slcs e j 2 i lT ij Mψ( ij ) n lcs e j 2 i lT Mψi nlcs
where ψi j 1 ij ψ( ij ) N 1 . Define Ψ ψ1 , ψ2 , , ψ K N K
K v ,i
v
can rewrite (5) as
3
v
(5)
and θl e j 2 lT , e j 2 lT , , e
1
2
j 2 K lT
T
Kv 1 . We
slcs MΨθl nlcs
Concatenating all
L
compressive measurement vectors
slcs
(6)
and
nlcs
as
S s0cs , s1cs , , scsL 1 M L
and
N n0cs , n1cs , , n csL 1 M L , respectively, we have
S MΨΘ N
(7)
with Θ θ0 , θ1 ,, θ L 1 K L .
Our goal is to accurately and efficiently estimate the target parameters ij , ij and i ( i 1, 2,, K v ,
j 1, 2, , K v ,i ) from the compressive measurement data matrix S . As discussed in [7], estimating these parameters can
be performed by solving a high-dimensional nonlinear optimization, which will take large computational load.
III. ESTIMATION OF TARGET PARAMETERS
In the proposed scheme, we first estimate the Doppler shifts and then the delays and gains.
A. Estimation of the Doppler shifts
The estimation of Doppler shifts is performed independently of the delays and gains. To see this, we take the
transpose operation on both sides of (7) and have
ST ΘT Ψ T M T N T
(8)
where ΘT a 1 , a 2 ,, a K LK and a i 1, e j2 T , , e j2 L 1T L1 . If we take ΘT as the steering
T
i
i
matrix and ΨT M T as the source matrix consisting of K v sources, we can see that (8) is exactly the same as the model
of classical DOA estimation problem for an L -element uniform linear array [8]. The i -th row of ΨT M T represents
the arriving source with spatial frequency f i viT 1 2 ,1 2 , and the number of snapshots equals to that of
compressive samples M in one PRI. The target parameters, gains and delays, are contained in Ψ which does not
affect the independent estimation of the Doppler shifts. Then the Doppler shifts can be estimated by the parametric
DOA estimation methods, such as MUSIC [16] and ESPRIT [17].
In practical radar scenarios, there may be several targets with the same delays but having different Doppler shifts.
In such case, the source signals resulting from the radar echoes become coherent and the resulting matrix ΨT M T is not
of full-row rank. If the case occurs, we can take the coherent DOA techniques [18] to estimate the Doppler shifts.
4
The parametric DOA estimation idea has been taken in our previous work [7] which formulates the estimation of
the time delays as a DOA estimation problem. However, the proposed method is superior to the GeSeDD method when
the coherent sources appear in the corresponding DOA estimation problem. In GeSeDD, the estimation of time delays
is targeted as a beamspace DOA estimation problem, where classical coherent DOA estimation techniques cannot be
exploited directly [8] and advanced techniques, such as interpolated array technique [19], should be utilized. Then the
computational load increases and the estimation accuracy is also affected. In contrast, the proposed method can figure
out the coherent sources by classical techniques, since the DOA estimation problem in (8) is of element-space.
When the Doppler shifts are all on the Nyquist grids, the proposed method reduces to the compressive sampling
pulse-Doppler (CoSaPD) method [4] if we take the discrete Fourier transform (DFT) on (8) to estimate the Doppler
shifts. In this case, the coherent sources have no effect on the estimation of Doppler shifts.
B. Estimation of the delays and gains
With the estimated Doppler shifts, a straightforward way to estimate the delays ij and gains ij ( i 1, 2, , K v ,
j 1, 2, , K v ,i ) is to solve the following optimization problem
arg min S MΨΘ
ij , ij
By expressing MΨΘ as MΨΘ i 1 Mψi a vi
K
T
2
(9)
F
and column-wise vectorizing the equation (7), we can translate the
problem (9) into
arg min vec S i 1 a vi M j v1,i ij ψ( ij )
K
K
ij , ij
2
(10)
2
where vec S ML1 is the column-wise vectorized S , denotes the Kronecker product and
a vi M j v1,i ij ψ( ij ) a vi Mψi vec Mψi a vi
K
T
The problem (10) is a special nonlinear least square with the unknown K v ,i , and can be resolved by the compressive
parameter estimation techniques [3, 9-11].
Directly computing (10) is time-consuming because of its high observation dimension ML in vec S ML1 . We
find that the problem (10) can be further decomposed into K v sub-problems with each corresponding to a Doppler
shift vi and having low observation dimension M .
5
Note that Θ is a Vandermonde matrix having full-row rank. Then ΘΘ† I with Θ† ΘH ΘΘ H , the
1
Moore-Penrose pseudo-inverse of Θ . Post-multiplying the both sides of the equation (7) by Θ† , we have
SΘ† MΨ NΘ† Mψ1 , Mψ2 , , Mψ Kv NΘ†
(11)
It is seen that the i -th column of SΘ† is the noisy version of the measurement vector Mψi corresponding to the
Doppler shift i . Denoting s and n respectively as the i -th column of SȆ and NȆ , we have
i
i
s i Mψi n i ,
i 1, 2, , K v
(12)
Then directly solving (10) is equivalent to solving the K v sub-problems in which each is confined to find the delays
ij and gains ij corresponding to the Doppler vi ,
2
arg min s i M j v1,i ij ψ( ij ) ,
K
ij , ij
i 1, 2, , K v
2
(13)
In comparison with (10), the observation dimension of (13) is reduced L times and thus the computational load is
greatly alleviated.
In the CoSaPD method [4], a vector similar to s is generated by the row-wise DFT of S . For the targets whose
i
Doppler shifts are on the Nyquist grids, the matrix Θ is a partial Fourier matrix and the DFT operation simply
classifies the targets into groups with each corresponding to a distinct Doppler shift. For the formulation (7), the
difficulty is that the matrix Θ is not a partial Fourier matrix. However, noting the Vandermonde structure of Θ , we
devise a similar decomposition as described in (11). This decomposition greatly enhances computational efficiency.
In addition to its generality to deal with the targets of the on- and/or off-grids, the vector s generated by (12)
i
has the improved signal-to-noise ratio (SNR) with respect to the signal vector in (5). Let b i be the i -th column of
Θ† . Then for each distinct Doppler shift i , since ΘΘ† I , we have
a i T b i 1
T
a i b i 0,
i i
(14)
Let Pi I N Θi ΘiH Θi ΘiH which is an orthogonal projection matrix with Θi a 1 , , a i 1 , a i 1 , , a K
1
L K 1
. With (14), we can derive
b i
Pi *a i
*
a i Pi *a i
T
6
*
(15)
By the definition of n , we have
i
NPi *a i
*
n i
whose power is given by n lcs
Pi a i
2
2
T
a i Pi *a i
2
T
2
(16)
a i Pi *a i
*
2
*
. Note that the signal power in (5) is the same as in (12) for
a distinct Doppler target. Then after the decomposed processing, the SNR in (12) will be improved by
* 2
a i Pi *a i
T
Pi a i 2 L . When all the Doppler shifts are on the Nyquist grids, all a i ( i 1, 2, , K v ) are
2
* 2
orthogonal to each other and Pi is an identity matrix. Then a i Pi*a i
T
Pi a i 2 L and the SNR for the
2
on-grid targets is enhanced by L times, which is consistent with the analyses on CoSaPD in [4]. For CoSaPD, the
row-wise DFT operation performs as matched filtering processing by a i in Doppler dimension. In this sense, the
*
decomposed processing to generate the signal vector s has the similar function. However, for the off-grid targets, an
i
extra pre-processing Pi* a i Pi *a i
T
*
is introduced, which cause a loss on the SNR improvement.
In classic pulse-Doppler processing, we usually perform the matched filtering in time domain and coherent
integration in Doppler dimension to enhance the target strength for further processing. One may argue if there is SNR
loss for the proposed scheme because the matched filtering processing in time domain does not explicitly appear, as
pointed out by one of reviewers. As discussed in other works [4], this is a difference of the radar signal processing in
compressed domain from the classical processing. In fact, the matched filtering processing is implied in the estimation
process by (13), which increases the input SNR by MBTp N times. Then the system processing gain after the delay
T
* 2
estimation will be improved to be a i Pi*a i
2
Pi a i 2 MBTp N . For the on-grid targets, the processing gain
equals to LMBTp N , which is same as that in CoSaPD.
The proposed scheme performs the estimation of the Doppler shifts and the delays sequentially by (8) and (13).
Although the SNR is greatly enhanced at each processing stage, the system processing gain by the proposed sequential
processing is less than that by the classic pulse-Doppler processing because of the noise-folding phenomena caused by
the sub-Nyquist sampling. Then the optimal estimation accuracy with the compressed data in (7) is lower than that by
the Nyquist data as revealed by the Cramer-Rao bound (CRB) analyses in [20]. The simulations in Section V confirms
7
the theoretical analyses.
IV. ANALYSIS OF COMPUTATIONAL COMPLEXITY
This section analyzes the computational complexity of the proposed scheme. We utilize ESPRIT algorithm [17] to
estimate the Doppler shifts because of its high estimation accuracy and computational efficiency, and exploit the
parameter perturbed orthogonal matching pursuit (PPOMP) algorithm [3] to estimate the delays and gains for its high
estimation accuracy.
As is known, estimating the Doppler shifts from S by ESPRIT algorithm takes ML2 L3 operations. To
estimate the delays and gains, we take KLM K 2 L K 3 operations to generate all the vectors s from S and
i
PMN operations to solve the problem (13) for each s i where P is the maximum number of iterations for
nonlinear optimization sub-problems in PPOMP [3]. In the worst case where the delays and Doppler shifts are distinct
between any two targets, we have to solve the problem (13) by K times. Then we have the total computational load
of the proposed scheme as ML2 L3 + KLM K 2 L K 3 + KPMN . For sub-Nyquist radar, it is known that
K L P M N . Then the computational complexity of our scheme is approximated as KPMN . In [7], we have
derived that the complexities of GeSeDD via beamspace root MUSIC (abbreviated as GeSeDD-1) and GeSeDD via
beamspace spectral MUSIC (abbreviated as GeSeDD-2) are respectively approximated as N 3 and DMN 2 ,
where D is the ratio between the Nyquist grid space and the grid-search space and generally satisfies K D L . It is
also argued in [7] that the computational complexity of GeSeDD is comparable to that of CoSaPD method and Doppler
focus (DF) method in [5]. Then the proposed scheme has the smallest computational complexity among these methods.
V. SIMULATION RESULTS
This section simulates some experiments to show the estimation accuracy and the computational advantages of the
proposed scheme. The transmitted signal is assumed to be a linear frequency modulation pulsed signal with bandwidth
B 100 MHz and pulse width Tp 10 μs . The PRI is T 100 μs and the number of pulses is L 100 . The delay cell
size 0 1 B is 0.01 μs and the Doppler cell size v0 1 ( LT ) is 0.1 KHz . Without special statements, the delay and
the Doppler shift of all the targets are randomly chosen with a uniform distribution from the intervals 0,10 μs and
8
-5,5 KHz . The gain amplitudes and phases of the targets are uniformly distributed in 0.1,1 and 0, 2 ,
respectively. The SNR slcs n lcs
2
2
nlcs
2
2
is assumed to be identical in each pulse interval. The estimated delays and
Doppler shifts are determined by the estimated targets having the K largest amplitudes of gains.
The Xampling is taken as a sample AIC with one-fifth Nyquist rate, where the measured Fourier coefficients are
randomly selected. For performance comparisons, simulation results by GeSeDD-1 and GeSeDD-2 [7], CoSaPD [4]
and DF [5] are demonstrated. As a benchmark, a simultaneous method, the PPOMP in [3], is also shown for its high
estimation accuracy. For GeSeDD-2, the grid space is set as one-tenth delay cell size for grid search in beamspace
spectral MUSIC. For the DF method, the grid spaces of delay and Doppler shift are both set as a half of cell size.
We now evaluate the delay/Doppler estimation performance of all methods in the noisy case. The relative
root-mean-square error (RRMSE) is utilized as a metric for the performance evaluation. Let the set τ 1 , 2 , , K
and v v1 , v2 , , vK be the delay and Doppler shift estimate values of K targets. The RRMSEs of the time delay
and the Doppler shift are defined respectively as
RRMSE
1
0
1
K
K
k 1
k
k
2
and RRMSE v
1
v0
1
K
K
v
k 1
k
vk
2
In the simulation, the delay and Doppler separation between any two targets are respectively set to be at least 2 0
and 2v0 to alleviate interaction between different targets. Fig. 1 displays the curves of the RRMSEs versus the SNRs
with ten off-grid targets for the noisy compressed data. The CRBs of delays and Doppler shifts are provided as a
benchmark, which can be derived from the vectorized representation of (7). It is seen that when SNR 10 dB , the
proposed scheme has the same accuracy as GeSeDD-1 and is superior to the other sequential methods. However, both
the proposed scheme and GeSeDD have a sharp increase in RRMSE when SNR 10 dB . This is because both
schemes exploit subspace DOA estimation techniques to find the Doppler shifts or the delays. As we all know, the
subspace techniques suffer a threshold effect [7], i.e., the estimation performance is sharply decreased when the SNR is
below a threshold value. For a small SNR ( SNR 10dB ), PPOMP offers the highest estimation accuracy because of its
robustness to noise [7]. In spite of the SNR, both CoSaPD and DF have poor estimation accuracy due to the basis
9
mismatch. As analyzed in [7], GeSeDD-1 has the best estimation accuracy among the simulated sequential methods.
This simulation shows that the proposed scheme has high estimation accuracy and is comparable to any other methods.
The proposed scheme is also applicable to the Nyquist data for the estimation of the off-grid targets. We simulate
the RRMSE performance of the classic pule-Doppler processing [12]. A quadratic interpolation technique [12] is
exploited to improve the estimation accuracy of the delay and Doppler shift parameters, where the Nyquist data is
windowed by a hamming window. Fig.2 shows the estimation performance along with CRBs from the Nyquist data.
We see that the proposed scheme has the best estimation accuracy which approaches to the CRBs for the higher SNRs.
In the classic processing, the interpolation techniques can improve the estimation accuracy. However the improvement
is limited as discussed in [12]. The CoSaPD method has the worst performance because it works only for the on-grid
targets. This further shows the efficiency of the proposed parametric scheme. In comparison with Fig.1, we find that the
classic processing performs better among all the simulated schemes with sub-Nyquist data at low SNRs ( SNR 15 dB )
because of its higher processing gain. Moreover, we can find that the CRB from the Nyquist data is lower than that
from the compressed data, which is consistent with the analyses in [20]. For the proposed scheme, the estimation
accuracy from the compressed data is also lower than that from the Nyquist data. Then the pulse-Doppler processing in
the compressed domain reduces the estimation accuracy to a certain extent. However, the reduction is negligible, as
shown in Fig. 1 and Fig. 2.
Another advantage of the proposed scheme is that it takes less computational time. We take the CPU time to
illustrate the performance. The simulation is performed in MATLAB 2011b 64-bit environment on a PC with 3.6 GHz
Intel core i7-4790 processor and 16 GB RAM. Fig. 3 displays the dependence of the CPU time on the number of
targets K with SNR=20dB. It is clear that the proposed scheme has the least computational cost. This further verifies
our analyses in Section IV. Note that the CPU time of the proposed method increases along with larger K . This is
because the computational complexity is dominated by the delays and gains estimation and is linearly proportional to
K , as analyzed in Section IV. Similarly, the DF and CoSaPD methods lie on K , and the DF has stronger dependence
on K . On the contrary, the CPU time of both GeSeDD methods is independent on K . The reason is that the
10
computational cost is controlled by the beamspace DOA estimation problem, which is not affected by K [7].
VI. CONCLUSION
In this paper, we have presented a new general sequential estimation scheme for estimating the sub-Nyquist radar
parameters. Different from our previous contributions [7], we first estimate the Doppler shifts and then the delays and
gains of the targets. With the new mechanism, the proposed scheme has the less computational complexity and the high
estimation accuracy among all methods reported. Simulations validate our theoretical development. This sequential
estimation idea has been used in the development of the CoSaPD method [4]. The difference is that, unlike [4] in which
the targets are assumed to be on the predefined grids, the new scheme developed in this paper removes this constraint.
In this sense, the proposed scheme is a gridless CoSaPD.
We are currently working on the development of the target detection techniques incorporated into the proposed
sequential scheme.
ACKNOWLEDGEMENTS
The authors would like to thank the anonymous reviewers whose constructive comments improved the content
and the presentation of this paper. This research is supported partially by the National Natural Science Foundation of
China (Nos. 61401210, 61571228 and 61671245).
REFERENCES
[1] M. A. Herman and T. Strohmer, High-resolution radar via compressed sensing, IEEE Trans. Signal Process. 57 (6) (2009)
2275-2284.
[2] J. Zhang, D. Zhu, and G. Zhang, Adaptive compressed sensing radar oriented toward cognitive detection in dynamic sparse
target scene, IEEE Trans. Signal Process. 60 (4) (2012) 1718-1729.
[3] O. Teke, A. C. Gurbuz, and O. Arikan, A robust compressive sensing based technique for reconstruction of sparse radar scenes,
Digital Signal Process. 27 (4) (2014) 23-32.
[4] C. Liu, F. Xi, S. Chen, Y. D. Zhang, and Z. Liu, Pulse-doppler signal processing with quadrature compressive sampling, IEEE
Trans. Aerospace and Electronic Systems 51 (2) (2015) 1216-1230.
[5] O. Bar-Ilan and Y. C. Eldar, Sub-Nyquist radar via Doppler focusing, IEEE Trans. Signal Process. 62 (7) (2014) 1796-1811.
[6] W. U. Bajwa, K. Gedalyahu, and Y. C. Eldar, Identification of parametric underspread linear systems and super-resolution
radar, IEEE Trans. Signal Process. 59 (6) (2011) 2548-2561.
[7] S. Chen, F. Xi, and Z. Liu, A general sequential delay-Doppler estimation scheme for sub-Nyquist pulse-Doppler radar, Signal
Process. 135 (2017) 210-217.
[8] P. Stoica and R. L. Moses, Spectral Analysis of Signals. Upper Saddle River, NJ, USA: Prentice-Hall, 2005.
11
[9] C. Ekanadham, D. Tranchina, and E. P. Simoncelli, Recovery of sparse translation-invariant signals with continuous basis
pursuit, IEEE Trans. Signal Process. 59 (10) (2011) 4735-4744.
[10] K. Fyhn, M. F. Duarte, and S. H. Jensen, Compressive parameter estimation for sparse translation-invariant signals using polar
interpolation, IEEE Trans. Signal Process. 63 (4) (2015) 870-881.
[11] A. Fannjiang and W. Liao, Coherence pattern-guided compressive sensing with unresolved grids, SIAM J. Imaging Sci. 5 (1)
(2012) 179-202.
[12] M. A. Richards, Fundamentals of Radar Signal Processing. New York, NY, USA: McGraw-Hill, 2005.
[13] F. Xi, S. Chen, and Z. Liu, Quadrature compressive sampling for radar signals, IEEE Trans. Signal Process. 62 (11) (2014)
2787-2802.
[14] R. G. Baraniuk and P. Steeghs, Compressive radar imaging, in Proc. IEEE Radar Conf., 2007, pp. 128-133.
[15] F. Xi, S. Y. Chen, and Z. Liu, Quardrature compressive sampling for radar signals: Output noise and robust reconstruction,
2014 IEEE China Summit and International Conference on Signal and Information Processing, 2014, pp. 790-794.
[16] R. O. Schmidt, Multiple emitter location and signal parameter estimation, IEEE Trans. on Antennas Propagat. 34 (3) (1986)
276-280.
[17] R. Roy and T. Kailath, ESPRIT-estimation of signal parameters via rotational invariance techniques, IEEE Trans. on Acoust.,
Speech, Signal Process. 37 (7) (1989) 984-995.
[18] S. U. Pillai and B. H. Kwon, Forward/backward spatial smoothing techniques for coherent signal identification, IEEE Trans.
on Acoust., Speech, Signal Process. 37 (1) (1989) 8-15.
[19] F. Xi, S. Chen, Y. D. Zhang and Z. Liu, Gridless quadrature compressive sampling with interpolated array technique, Signal
Process. 133 (2017) 1-12.
[20] P. Pakrooh, A. Pezeshki, L. L. Scharf, D. Cochran, and S. D. Howard, Analysis of Fisher information and the Cramer-Rao
bound for nonlinear parameter estimation after compressed sensing, IEEE Trans. Signal Process. 63 (23) (2015) 6423-6428.
3
3
10
10
1
1
10
10
RRMSEv
RRMSE
-1
10
-3
10
-5
10
-20
Proposed
GeSeDD-1
GeSeDD-2
CoSaPD
-10
0
DF
PPOMP
CRB
10
SNR (dB)
20
-1
10
-3
10
-5
10
30
40
-20
Proposed
GeSeDD-1
GeSeDD-2
CoSaPD
-10
0
DF
PPOMP
CRB
10
SNR (dB)
(a) time-delay
(b) Doppler shift
Fig. 1 The RRMSEs versus the SNR for noisy compressed data (K=10)
12
20
30
40
3
3
10
10
1
1
10
10
RRMSEv
-3
10
-5
10
-20
Proposed
CoSaPD
Classic
CRB
-10
0
-1
10
-3
10
-5
10
10
SNR (dB)
20
30
40
-20
Proposed
CoSaPD
Classic
CRB
-10
0
10
SNR (dB)
(a) time-delay
(b) Doppler shift
Fig. 2 The RRMSEs versus the SNR for noisy Nyquist data (K=10)
2
10
0
10
Time (s)
RRMSE
-1
10
-2
10
-4
10
1
Proposed
GeSeDD-1
GeSeDD-2
5
CoSaPD
DF
PPOMP
9
13
17
Number of targets K
21
25
Fig. 3 The CPU time versus the number of targets with SNR=20dB
13
20
30
40
| 7cs.IT
|
arXiv:1602.01661v1 [math.GR] 4 Feb 2016
On finite groups with automorphisms
whose fixed points are Engel
Pavel Shumyatsky and Danilo Sanção da Silveira
For Evgeny Khukhro on his 60th birthday
Abstract. The main result of the paper is the following theorem.
Let q be a prime, n a positive integer and A an elementary abelian
group of order q 2 . Suppose that A acts coprimely on a finite group
G and assume that for each a ∈ A# every element of CG (a) is nEngel in G. Then the group G is k-Engel for some {n, q}-bounded
number k.
1. Introduction
Let A be a finite group acting on a finite group G. Many wellknown results show that the structure of the centralizer CG (A) (the
fixed-point subgroup) of A has influence over the structure of G. The
influence is especially strong if (|A|, |G|) = 1, that is, the action of A
on G is coprime. We will give examples illustrating this phenomenon.
Thompson proved that if A is of prime order and CG (A) = 1, then
G is nilpotent [15]. Higman proved that any nilpotent group admitting
a fixed-point-free automorphism of prime order q has nilpotency class
bounded by some function h(q) depending on q alone [4]. The reader
can find in [7] and [8] an account on the more recent developments
related to these results. The next result is a consequence of the classification of finite simple groups [18]: If A is a group of automorphisms
of G whose order is coprime to that of G and CG (A) is nilpotent or
has odd order, then G is soluble. Once the group G is known to be
soluble, there is a wealth of results bounding the Fitting height of G in
1991 Mathematics Subject Classification. 20D45, 20F40, 20F45.
Key words and phrases. Finite groups, Automorphisms, Centralizers, Engel
elements.
This work was supported by CNPq and FAPDF, Brazil.
1
2
PAVEL SHUMYATSKY AND DANILO SANÇÃO DA SILVEIRA
terms of the order of A and the Fitting height of CG (A). This direction
of research was started by Thompson in [16]. The proofs mostly use
representation theory in the spirit of the Hall-Higman work [3]. A general discussion of these methods and their use in numerous fixed-point
theorems can be found in Turull [17].
Following the solution of the restricted Burnside problem it was
discovered that the exponent of CG (A) may have strong impact over
the exponent of G. Remind that a group G is said to have exponent
n if xn = 1 for every x ∈ G and n is the minimal positive integer with
this property. The next theorem was obtained in [9].
Theorem 1.1. Let q be a prime, n a positive integer and A an
elementary abelian group of order q 2 . Suppose that A acts coprimely
on a finite group G and assume that CG (a) has exponent dividing n for
each a ∈ A# . Then the exponent of G is {n, q}-bounded.
Here and throughout the paper A# denotes the set of nontrivial
elements of A. The proof of the above result involves a number of deep
ideas. In particular, Zelmanov’s techniques that led to the solution
of the restricted Burnside problem [19, 20] are combined with the
Lubotzky–Mann theory of powerful p-groups [12], Lazard’s criterion
for a pro-p group to be p-adic analytic [10], and a theorem of Bakhturin
and Zaicev on Lie algebras admitting a group of automorphisms whose
fixed-point subalgebra is PI [1].
Let n be a positive integer and let x, y be elements of a group G.
The commutators [x,n y] are defined inductively by the rule
[x,0 y] = x,
[x,n y] = [[x,n−1 y], y].
An element x is called a (left) n-Engel element if for any g ∈ G we have
[g,n x] = 1. A group G is called n-Engel if all elements of G are n-Engel.
The main result of the present article is the following theorem.
Theorem 1.2. Let q be a prime, n a positive integer and A an
elementary abelian group of order q 2 . Suppose that A acts coprimely
on a finite group G and assume that for each a ∈ A# every element of
CG (a) is n-Engel in G. Then the group G is k-Engel for some {n, q}bounded number k.
The methods employed in the proof of Theorem 1.2 are roughly
the same as those in the proof of Theorem 1.1. There are however
some rather significant differences. The main difficulty is that Theorem
1.2 does not reduce as quickly as Theorem 1.1 to questions about Lie
algebras over field. Instead, we have to work with Lie rings. This hurdle
ON GROUPS WITH AUTOMORPHISMS
3
is overcome via Proposition 2.6 obtained in the next section and some
other tools.
Throughout the paper we use without special references the wellknown properties of coprime actions:
If α is an automorphism of a finite group G of coprime order,
(|α|, |G|) = 1, then CG/N (α) = CG (α)N/N for any α-invariant normal subgroup N.
If A is a noncyclic abelian group acting coprimely on a finite group
G, then G is generated by the subgroups CG (B), where A/B is cyclic.
We use the expression “{a, b, . . . }-bounded” to abbreviate “bounded
from above in terms of a, b, . . . only”.
2. About Lie Rings and Lie Algebras
Let X be a subset of a Lie algebra L. By a commutator in elements
of X we mean any element of L that can be obtained as a Lie product
of elements of X with some system of brackets. If x1 , . . . , xk , x, y are
elements of L, we define inductively
[x1 ] = x1 ; [x1 , . . . , xk ] = [[x1 , . . . , xk−1 ], xk ]
and [x,0 y] = x; [x,m y] = [[x,m−1 y], y], for all positve integers k, m. As
usual, we say that an element a ∈ L is ad-nilpotent if there exists a
positive integer n such that [x,n a] = 0 for all x ∈ L. If n is the least
integer with the above property then we say that a is ad-nilpotent of
index n.
The next theorem is a deep result of Zelmanov [20].
Theorem 2.1. Let L be a Lie algebra over a field and suppose that
L satisfies a polynomial identity. If L can be generated by a finite set
X such that every commutator in elements of X is ad-nilpotent, then
L is nilpotent.
An important criterion for a Lie algebra to satisfy a polynomial
identity is provided by the next theorem. It was proved by Bakhturin
and Zaicev for soluble groups A [1] and later extended by Linchenko
to the general case [11].
Theorem 2.2. Let L be a Lie algebra over a field K. Assume
that a finite group A acts on L by automorphisms in such a manner
that CL (A) satisfies a polynomial identity. Assume further that the
characteristic of K is either 0 or prime to the order of A. Then L
satisfies a polynomial identity.
Both Theorem 2.1 and Theorem 2.2 admit respective quantitative
versions (see for example [14]).
4
PAVEL SHUMYATSKY AND DANILO SANÇÃO DA SILVEIRA
Theorem 2.3. Let L be a Lie algebra over a field K generated by
a1 , . . . , am . Suppose that L satisfies a polynomial identity f ≡ 0 and
each commutator in a1 , . . . , am is ad-nilpotent of index at most n. Then
L is nilpotent of {f, K, m, n}-bounded class.
Theorem 2.4. Let L be as in Theorem 2.2 and assume that CL (A)
satisfies a polynomial identity f ≡ 0. Then L satisfies a polynomial
identity of {|A|, f, K}-bounded degree.
By combining the above results we obtain the following corollary.
Corollary 2.5. Let L be a Lie algebra over a field K and A a finite
group of automorphisms of L such that CL (A) satisfies the polynomial
identity f ≡ 0. Suppose that the characterisitic of K is either 0 or
prime to the order of A. Assume that L is generated by an A-invariant
set of m elements in which every commutator is ad-nilpotent of index
at most n. Then L is nilpotent of {|A|, f, K, m, n}-bounded class.
We will need a similar result for Lie rings. As usual, γi (L) denotes
the ith term of the lower central series of L.
Proposition 2.6. Let L be a Lie ring and A a finite group of
automorphisms of L such that CL (A) satisfies the polynomial identity
f ≡ 0. Further, assume that L is generated by an A-invariant set of m
elements such that every commutator in the generators is ad-nilpotent
of index at most n. Then there exist positive integers e and c, depending
only on |A|, f, m and n, such that eγc (L) = 0.
Proof. Since the set of generators of L is A-invariant, every automorphism α ∈ A induces a permutation of the generators. Let LQ
denote the free Lie algebra on free generators x1 , . . . , xm over the field
Q of rational numbers and LZ the Lie subring of LQ generated by
x1 , . . . , xm . For each α ∈ A let φα be the automorphism of LQ (or of
LZ ) such that xφi α = xπ(i) where π is the permutation that α induces
on the generators of L. The mapping that takes α ∈ A to φα induces
a natural action of A on LQ (or on LZ ) by automorphisms.
Let I be the ideal of LZ generated by all values of f on elements of
CLZ (A) and by all elements of the form [u, n v], where u ranges through
LZ and v ranges through the set of all commutators in the generators x1 , . . . , xm . By Corollary 2.5 LQ /QI is nilpotent of {|A|, f, m, n}bounded class, say c − 1.
If y1 , . . . , yc are not necessarily distinct elements of the set x1 , . . . , xm ,
it follows that the commutator [y1 , . . . , yc ] lies in QI and therefore it
ON GROUPS WITH AUTOMORPHISMS
5
can be written in the form
[y1 , . . . , yc ] =
X
qi fi wi ,
i
where qi are rational numbers, fi are either elements of the form [u, n v]
as above or values of f in elements of CLZ (A) and fi wi are elements of
LZ obtained by multiplying (several times) fi with products in the free
generators x1 , . . . , xm .
Let e be the least common multiple of the denominators of the coefficients qi taken over all possible choices of y1 , . . . , yc in {x1 , . . . , xm }.
Then e[y1 , . . . , yc ] ∈ I for any choice of y1 , . . . , yc . Therefore eγc (LZ ) ≤
I.
We remark that there exists a natural homomorphism of LZ /I onto
L under which each automorphism φα induces the automorphism α of
L. Thus, eγc (L) = 0, as required.
We will now quote a useful lemma from [9].
Lemma 2.7. Let L be a Lie ring and H a subring of L generated by
m elements h1 , . . . , hm such that all commutators in hi are ad-nilpotent
in L of index at most n. If H is nilpotent of class c, then for some
{c, m, n}-bounded number u we have [L, H, . . . , H ] = 0.
| {z }
u
Recall that the identity
X
[y, xσ(1) , . . . , xσ(n) ] ≡ 0
σ∈Sn
is called linearized n-Engel identity. In general, Theorem 2.1 cannot be
extended to the case where L is just a Lie ring (rather than Lie algebra
over a field). However such an extension does hold in the particular
case where the polynomial identity f ≡ 0 is a linearized Engel identity.
Theorem 2.8. Let f be a Lie polynomial of degree n all of whose
coefficients equal 1 or -1. Let L be a Lie ring generated by finitely
many elements a1 , . . . , am such that all commutators in the generators
are ad-nilpotent of index at most n. Assume that L satisfies the identity
f ≡ 0. Then L is nilpotent of {f, m, n}-bounded class.
The deduction of Theorem 2.8 from Theorem 2.3 uses just standard
arguments. For the reader’s convenience we sketch out a proof.
Proof. By Proposition 2.6 applied with A = 1, there exist positive
integers e and c, depending only on f, m and n, such that eγc (L) = 0.
We set M = eL and notice that M is a nilpotent ideal of L whose
nilpotency class is bounded in terms of f, m and n. Let d be the
6
PAVEL SHUMYATSKY AND DANILO SANÇÃO DA SILVEIRA
derived length of M and use induction on d. Suppose first that M =
0. Then L is a direct sum of its primary components Lp , where p
ranges through prime divisors of e. It is sufficient to show that the
nilpotency class of each Lp is bounded in terms of f, m and n and so
without loss of generality we can assume that e = pk for a prime p.
We note that k here is {f, m, n}-bounded. The quotient L/pL can
be regarded as a Lie algebra over the field with p elements and so it
is nilpotent of bounded class by Theorem 2.3. Hence, there exists an
{f, m, n}-bounded number c1 such that γc1 (L) ≤ pL. This implies that
γkc1 (L) ≤ pk L = 0 and we are done.
Assume now that M 6= 0 and let D be the last nontrivial term
of the derived series of M. By induction L̄ = L/D is nilpotent of
bounded class. The ring L̄ naturally acts on D by derivations and
Lemma 2.7 shows that for some {f, m, n}-bounded number u we have
[D, L̄, . . . , L̄] = 0. It follows that L is nilpotent of {f, m, n}-bounded
| {z }
u
class, as required.
3. On associated Lie rings
There are several well-known ways to associate a Lie ring to a group
(see [6, 7, 14]). For the reader’s convenience we will briefly describe
the construction that we are using in the present paper.
Let G be a group. A series of subgroups
G = G1 ≥ G2 ≥ . . .
(∗)
is called an N-series if it satisfies [Gi , Gj ] ≤ Gi+j for all i, j. Obviously
any N-series is central, i.e. Gi /Gi+1 ≤ Z(G/Gi+1 ) for any i. Given
an N-series (∗), let L∗ (G) be the direct sum of the abelian groups
L∗i = Gi /Gi+1 , written additively. Commutation in G induces a binary
operation [, ] in L∗ (G). For homogeneous elements xGi+1 ∈ L∗i , yGj+1 ∈
L∗j the operation is defined by
[xGi+1 , yGj+1] = [x, y]Gi+j+1 ∈ L∗i+j
and extended to arbitrary elements of L∗ (G) by linearity. It is easy
to check that the operation is well-defined and that L∗ (G) with the
operations + and [, ] is a Lie ring. If all quotients Gi /Gi+1 of an Nseries (∗) have prime exponent p then L∗ (G) can be viewed as a Lie
algebra over the field with p elements. Any automorphism of G in the
natural way induces an automorphism of L∗ (G). If G is finite and α
is an automorphism of G such that (|α|, |G|) = 1, then the subring of
fixed points of α in L∗ (G) is isomorphic with the Lie ring associated to
ON GROUPS WITH AUTOMORPHISMS
7
the group CG (α) via the series formed by intersections of CG (α) with
the series (∗).
In the case where the series (∗) is just the lower central series of
G we write L(G) for the associated Lie ring. In the case where the
series (∗) is the p-dimension central series of G we write Lp (G) for the
subalgebra generated by the first homogeneous component G1 /G2 in
the associated Lie algebra over the field with p elements.
Proposition 3.1. Let G be an m-generated group satisfying the
hypothesis of Theorem 1.2. Let p be a prime number that divides the
order of G. Then
(1) Lp (G) is nilpotent of {p, q, m, n}-bounded class.
(2) There exists positive integers e, c depending only on m, n and
q, such that eγc (L(G)) = 0.
The proofs of the statements (1) and (2) are similar. We will give
a detailed proof of the second statement only. The proof of (1) can
be obtained simply by replacing every appeal to Proposition 2.6 in the
proof of (2) by an appeal to Corollary 2.5. An important observation
that is used in the proof is that if G is a finite p-group that can be
generated by d elements, then we can choose d generators from any
subset that generates G. This follows from the well-known Burnside
Basis Theorem [5].
Proof. Since G is generated by the centralizers CG (a), where a ∈
A# , and since all elements in the centralizers CG (a) are Engel, the Baer
Theorem [[5], III,6.14] implies that the group G is nilpotent. Therefore
G is a direct product of its Sylow subgroups. Hence, without loss
of generality we can assume that G is a p-group. Let γj = γj (G),
Lj = γj /γj+1 and L = L(G) = ⊕j≥1 Lj be the Lie ring associated with
the group G. Let A1 , . . . , Aq+1 be the distinct maximal subgroups of
A. Set Lij = CLj (Ai ). We know that any A-invariant subgroup is
generated by the centralizers of Ai . Therefore for any j we have
Lj =
q+1
X
Lij .
i=1
Further, for any l ∈ Lij there exists x ∈ γj ∩CG (Ai ) such that l = xγj+1.
Since x is n-Engel in G, it follows that l is ad-nilpotent of index at most
n. Thus,
(1)
any element in Lij is ad-nilpotent in L of index at most n.
8
PAVEL SHUMYATSKY AND DANILO SANÇÃO DA SILVEIRA
Since G is generated by m elements, the additive group L1 is generated
by m elements. It follows that the Lie ring L is generated by at most
m ad-nilpotent elements, each from Li1 for some i.
Let ω be a primitive qth root of unity and consider the tensor
product L = L ⊗ Z[ω]. Set Lj = Lj ⊗ Z[ω] for j = 1, 2, . . . . We regard
L as a Lie ring and so L = L1 . Since the additive subgroup L1 is
generated by m elements, it follows that the additive subgroup L1 is
generated by (q −1)m elements. We also remark that there is a natural
embedding of the ring L into the ring L.
The group A acts on L in the natural way and we have Lij =
CLj (Ai ) where Lij = Lij ⊗ Z[ω]. We will now establish the following
claim.
(2)
Any element in Lij is ad-nilpotent in L with {n, q}-bounded index.
Indeed choose y ∈ Lij and write
y = x0 + ωx1 + ω 2x2 + · · · + ω q−2xq−2
for suitable xs ∈ Lij . In view of (1) it is easy to see that each of the
summands ω s xs is ad-nilpotent in L of index at most n.
Let H = hx0 , ωx1 , ω 2 x2 , . . . , ω q−2xq−2 i be the subring of L generated
by x0 , ωx1 , ω 2x2 , . . . , ω q−2xq−2 . We wish to show that H is nilpotent of
{n, q}-bounded class.
Note that H ≤ CL (Ai ) since ω s xs ∈ CL (Ai ) for all s. A commutator
of weight t in the elements x0 , ωx1 , ω 2 x2 , . . . , ω q−2xq−2 has the form ω t x,
for some x ∈ Lij0 , where j0 = tj. By (1), the element x is ad-nilpotent
of index at most n in L and so we deduce that ω t x is ad-nilpotent of
index at most n in L.
Since CG (Ai ) is n-Engel, we conclude that CL (Ai ) satisfies the linearized n-Engel identity. This identity is multilinear and so it is also
satisfied by CL (Ai ) ⊗Z[ω] = CL (Ai ). Since H ≤ CL (Ai ), it follows that
the identity is satisfied by H. Hence, by Theorem 2.8, H is nilpotent
of {n, q}-bounded class.
By Lemma 2.7, there exists a positive integer v, depending only on
n and q, such that [L,v H] = 0. Since y ∈ H, we conclude that y is
ad-nilpotent in L with {n, q}-bounded index. This proves Claim (2).
An element x ∈ L will be called a common “eigenvector” for A if for
any a ∈ A there exists s such that xa = ω s x. Since (|A|, |G|) = 1, the
additive group of the Lie ring L is generated by common eigenvectors
for A (see for example [7, Lemma 4.1.1]).
ON GROUPS WITH AUTOMORPHISMS
9
We have already remarked that the additive subgroup L1 is generated by (q − 1)m elements. It follows that the Lie ring L is generated
by at most (q − 1)m common eigenvectors for A. Certainly, the Lie
ring L is generated by an A-invariant set of at most (q − 1)qm common
eigenvectors for A.
Since A is noncyclic, any common eingenvector for A is contained in
the centralizer CL (Ai ) for some i. Further, any commutator in common
eigenvectors is again a common eigenvector. Therefore if l1 , . . . , ls ∈ L1
are common eigenvectors for A generating L, then any commutator in
these generators belongs to some Lij and therefore, by Claim (2), is
ad-nilpotent of {n, q}-bounded index.
As we have mentioned earlier, for any i the subalgebra CL (Ai ) satisfies the linearized n-Engel identity. Thus, by Proposition 2.6, there
exist positive integers e, c depending only on m, n and q such that
eγc (L) = 0. Since L embeds into L, we also have eγc (L) = 0. The
proof is complete.
4. Proof of the main theorem
A well-known theorem of Gruenberg says that a soluble group generated by finitely many Engel elements is nilpotent (see [13, 12.3.3]).
We will require a quantitative version of this theorem.
Lemma 4.1. Let G be a group generated by m n-Engel elements and
suppose that G is soluble with derived length d. Then G is nilpotent of
{d, m, n}-bounded class.
Proof. Suppose that the lemma is false. Then for each positive
integer i we can choose a group Gi satisfying the hypothesis of the
lemma and having nilpotency class at least i. In each group Gi we fix
generators gi1 , gi2 , . . . , gim that are n-Engel elements.
In the Cartesian product of the groups Gi consider the subgroup D
generated by m elements
g1 = (g11 , g21 , . . .), . . . , gm = (g1m , g2m . . .).
Thus, D is a soluble group generated by m elements that are n-Engel.
By Gruenberg’s theorem D is nilpotent, say of class c. We remark that
each of the groups Gi is isomorphic to a quotient D. Hence each of the
groups Gi is nilpotent of class at most c. A contradiction.
Given a group H and an integer i, in what follows we write H i for
the subgroup of H generated by all ith powers and γi (H) for the ith
term of the lower central series of H. We will also require the concept
of powerful p-groups. These were introduced by Lubotzky and Mann
10
PAVEL SHUMYATSKY AND DANILO SANÇÃO DA SILVEIRA
in [12]: a finite p-group G is powerful if and only if Gp ≥ [G, G] for
p 6= 2 (or G4 ≥ [G, G] for p = 2). The reader can consult books [2] or
[8] for more information on these groups.
Powerful p-groups have many nice properties. In particular, if P
is a powerful p-group, the subgroups γi (P ) and P i are also powerful.
Moreover, if n1 , . . . , ns are positive integers, it follows by repeated applications of [12, Props. 1.6 and 4.1.6] that
[P n1 , . . . , P ns ] ≤ γs (P )n1 ···ns .
If P is generated by d elements, then any subgroup of P can be generated by at most d elements and P is a product of d cyclic subgroups
(see Theorem 1.11 and Theorem 1.12 in [12]).
We are now ready to prove Theorem 1.2.
Proof of Theorem 1.2. Let G be as in Theorem 1.2. We wish
to show that G is k-Engel for some {n, q}-bounded number k. Since
G is generated by the centralizers CG (a) with a ∈ A# and since all elements in the centralizers CG (a) are Engel, the Baer Theorem implies
that the group G is nilpotent. Choose arbitrarily x, y ∈ G. It is sufficient to show that the subgroup hx, yi is nilpotent of {n, q}-bounded
class. Without loss of generality we can assume that no proper Ainvariant subgroup of G contains both x and y. In other words, we will
assume that G = hxA , y A i. Therefore the group G can be generated
by 2q 2 elements. Every Sylow subgroup of G satisfies these hypotheses and therefore without loss of generality we can assume that G is a
p-group for some prime p 6= q. Using the fact that G is generated by
the centralizers CG (a) with a ∈ A# and combining it with the Burnside Basis Theorem we can assume further that G is generated by at
most 2q 2 elements which are n-Engel. In view of Lemma 4.1 it is now
sufficient to show that the derived length of G is {n, q}-bounded. Let
L = L(G). By Proposition 3.1 (2), there exist positive integers e, c that
depend only on n and q such that eγc (L) = 0. If p is not a divisor of
e, we have γc (L) = 0 and so the group G is nilpotent of class at most
c − 1. In that case the proof is complete and so we assume that p is
a divisor of e. Consider first the particular case where G is a powerful
p-group. Set R = Ge and assume that R 6= 1.
We have
2
2
[R, R] ≤ [G, G]e ≤ Gpe = Rpe .
Let L1 = L(R). By Proposition 3.1 (2), eγc (L1 ) = 0. The fact that
eγc (L1 ) = 0 means that γc (R)e ≤ γc+1(R). Taking into account that R
is powerful, we write
γc (R)e ≤ γc+1(R) = [R′ , c−1R] ≤ [Rpe , c−1 R] ≤ γc (R)pe .
ON GROUPS WITH AUTOMORPHISMS
11
Hence, γc (R)e = 1. Since γc (R) is powerful and generated by at most
2q 2 elements, we conclude that γc (R) is a product of at most 2q 2 cyclic
2
subgroups. Hence the order of γc (R) is at most e2q . It follows that
the derived length of R is {n, q}-bounded. Recall that G is a powerful
p-group and R = Ge . It follows that the derived length of G is {n, q}bounded. Thus, we proved the desired result in the case where G is a
powerful p-group.
Let us now drop the assumption that G is powerful. By Proposition
3.1 (1) the algebra Lp (G) is nilpotent with bounded nilpotency class.
Proposition 1 of [9] now tells us that G has a characteristic powerful
subgroup of bounded index. We already know that the derived length
of the powerful subgroup is {n, q}-bounded. Hence, the derived length
of G is {n, q}-bounded as well. The proof is now complete.
References
[1] Y. A. Bakhturin, M. V. Zaicev, Identities of graded algebras, J. Algebra, 205
(1998), 1–12.
[2] J.D. Dixon, M.P.F. du Sautoy, A. Mann, and D. Segal, Analytic Pro-p Groups,
Cambridge University Press, Cambridge, 1991.
[3] P. Hall, G. Higman, The p-length of a p-soluble group and reduction theorems
for Burnside’s problem, Proc. London Math. Soc., (3) 6 (1956), 1–42.
[4] G. Higman, Groups and Lie rings having automorphisms without non-trivial
fixed points, J. London Math. Soc., 32 (1957), 321–334.
[5] B. Huppert, Endliche Gruppen I, Springer-Verlag, 1967.
[6] B. Huppert, N. Blackburn, Finite groups II. Springer Verlang, Berlin, 1982.
[7] E.I. Khukhro, Nilpotent Groups and their Automorphisms, de Gruyter–Verlag,
Berlin, 1993.
[8] E.I. Khukhro, p-Automorphisms of Finite p-Groups, London Math. Soc. Lecture Note Series 246, Cambridge University Press, Cambridge, 1998.
[9] E.I. Khukhro, P. Shumyatsky, Bounding the exponent of a finite group with
automorphisms, J. Algebra, 212 (1999), 363–374.
[10] M. Lazard, Groupes Analytiques p-Adiques, Publ. Math. Inst. Hautes Études
Sci., 26, Paris, (1965), 389–603.
[11] V. Linchenko, Identities of Lie algebras with actions of Hopf algebras, Comm.
Algebra, 25 (1997), 3179–3187.
[12] A. Lubotzky, A. Mann, Powerful p-groups. I: finite groups, J. Algebra, 105
(1987), 484–505.
[13] D.J.S. Robinson, A course in the theory of groups, 2nd ed. Springer-Verlag,
New York, 1996.
[14] P. Shumyatsky, Applications of Lie ring methods to group theory, in Nonassociative Algebra and Its Applications, (Eds R. Costa et al.), Marcel Dekker,
New York, (2000), 373–395.
12
PAVEL SHUMYATSKY AND DANILO SANÇÃO DA SILVEIRA
[15] J.G. Thompson, Finite groups with fixed-point-free automorphisms of prime
order, Proc. Nat. Acad. Sci. USA, 45 (1959), 578–581.
[16] J.G. Thompson, Automorphisms of solvable groups, J. Algebra, 1 (1964), 259–
267.
[17] A. Turull, Character theory and length problems, in Finite and Locally Finite
Groups, Kluwer Academic Publ., NATO ASI Series 471, Dordrecht-BostonLondon, 1995, 377–400.
[18] Y.M. Wang, Z.M. Chen, Solubility of finite groups admitting a coprime order
operator group, Boll. Un. Mat. Ital. A, 7 (1993), 325–331.
[19] E. Zelmanov, Lie methods in the theory of nilpotent groups, in Groups ’93
Galaway/ St Andrews, Cambridge University Press, Cambridge, (1995), 567–
585.
[20] E. Zelmanov, Nil Rings and Periodic Groups, The Korean Math. Soc. Lecture
Notes in Math., Seoul, (1992).
Department of Mathematics, University of Brasilia, 70910-900 Brası́lia
DF, Brazil
E-mail address: [email protected]
Department of Mathematics, University of Brasilia, 70910-900 Brası́lia
DF, Brazil
E-mail address: [email protected]
| 4math.GR
|
Uniform measures on braid monoids
and dual braid monoids
arXiv:1607.00565v2 [math.GR] 16 Nov 2016
Samy Abbes
Sébastien Gouëzel
Jean Mairesse
Vincent Jugé
Abstract
We aim at studying the asymptotic properties of typical positive braids, respectively
positive dual braids. Denoting by µk the uniform distribution on positive (dual) braids
of length k, we prove that the sequence (µk )k converges to a unique probability measure
µ∞ on infinite positive (dual) braids. The key point is that the limiting measure µ∞
has a Markovian structure which can be described explicitly using the combinatorial
properties of braids encapsulated in the Möbius polynomial. As a by-product, we settle
a conjecture by Gebhardt and Tawn (J. Algebra, 2014) on the shape of the Garside
normal form of large uniform braids.
MSC (2010): 20F36, 05A16, 60C05
1 Introduction
Consider a given number of strands, say n, and the associated positive braid monoid Bn+ defined
by the following monoid presentation, known as the Artin presentation:
Bn+
=
σ1 , . . . , σn−1
σi σj = σj σi
σi σj σi = σj σi σj
for |i − j| ≥ 2
for |i − j| = 1
+
(1.1)
.
The elements of Bn+ , the positive braids, are therefore equivalence classes of words over the alphabet
Σ = {σ1 , . . . , σn−1 }. Alternatively, going back to the original geometric intuition, positive braids
can be viewed as isotopy classes of positive braid diagrams, that is, braid diagrams in which the
bottom strand always goes on top in a crossing, see Figure 1.
We want to address the following question:
What does a typical complicated positive braid look like?
4
3
2
1
4
3
2
1
σ1
σ2
σ3
σ2
σ3
σ1
σ2
σ3
Figure 1: Two isotopic braid diagrams representing a positive braid on 4 strands. Left: diagram corresponding to the word σ1 σ2 σ3 σ2 . Right: diagram corresponding to the word σ3 σ1 σ2 σ3 .
1
To make the question more precise, we need to clarify the meaning of “complicated” and “typical”.
First, let the complexity of a positive braid be measured by the length (number of letters) of any
representative word. This is natural since it corresponds to the number of crossings between strings
in any representative braid diagram. Therefore, a positive braid is “complicated” if its length is
large.
Second, let us define a “typical” braid as a braid being picked at random according to some
probability measure. The two natural candidates for such a probability measure are as follows.
Fix a positive integer k.
• The first option consists in running a simple random walk on Bn+ : pick a sequence of random
elements xi , i ≥ 1, independently and uniformly among the generators Σ = {σ1 , . . . , σn−1 },
and consider the “typical” braid X = x1 · x2 · · · · · xk . It corresponds to drawing a word
uniformly in Σk and then considering the braid it induces.
• The second option consists in picking a “typical” braid of length k uniformly at random
among all braids of length k.
The two approaches differ since the number of representative words varies among positive braids
of the same length. For instance, in B3+ and for the length 3, the braid σ1 · σ2 · σ1 (= σ2 · σ1 · σ2 )
will be picked with probability 2/8 in the first approach, and with probability 1/7 in the second
one, while all the other braids of length 3 will be picked respectively with probabilities 1/8 and
1/7 in the two approaches. The random walk approach has been studied for instance in [26, 30]; it
is a special instance of random walks on (semi)groups, see [32]. In this paper, our focus is on the
second approach, that is, on uniform measures on positive braids.
Let µk be the uniform probability measure on positive braids of Bn+ of length k. The general
question stated above can now be rephrased as follows : study µk for large k. Let us say that we
are interested in some specific property, say, the number of occurrences of the Garside element ∆ in
a large random braid. To study it, a first approach consists in performing a numerical evaluation.
To that purpose, the key ingredient is to have a sampling algorithm, that is, a random procedure
which takes as input k and returns as output a random braid of distribution µk . Another, more
intrinsic, approach consists first in defining a probability measure µ∞ on infinite positive braids,
encapsulating all the measures µk , and then in studying the asymptotics of the property via µ∞ .
None of these two paths is easy to follow. The difficulty is that the probability measures (µk )k are
not consistent with one another. For instance, in B3+ , we have:
1/4 = µ2 (σ1 · σ1 ) 6= µ3 (σ1 · σ1 · σ1 ) + µ3 (σ1 · σ1 · σ2 ) = 2/7 .
(1.2)
Therefore, there is no obvious way to design a dynamic process to sample braids. As another
consequence, the Kolmogorov consistency theorem does not apply, and there is no simple way to
define a uniform probability measure on infinite positive braids. This is in sharp contrast with the
simpler picture for the random walk approach described above.
To overcome the difficulties, the rich combinatorics of positive braids has to enter the scene.
Going back to Garside [22] and Thurston [20], it is known that positive braids admit a normal
form, that is a selection of a unique representative word for each braid, which is regular, that is
recognized by a finite automaton. This so called Garside normal form enables to count positive
braids in an effective way, see for instance Brazil [11], but a non-efficient one since the associated
automaton has a large number of states, exponential in the number of strands n, see Dehornoy [16].
A breakthrough is provided by Bronfman [12] (see also [5]) who obtains, using an inclusion-exclusion
principle, a simple recursive formula for counting positive braids. Based on this formula, a sampling
algorithm whose time and space complexities are polynomial in both the number of strands n and
the length k is proposed by Gebhardt and Gonzales-Meneses in [23]. Using the sampling procedure,
extensive numerical evaluations are performed by Gebhardt and Tawn in [24], leading to the stable
region conjecture on the shape of the Garside normal form of large uniform braids.
In the present paper, we complete the picture by proving the existence of a natural uniform
probability measure µ∞ on infinite positive braids. The measure induced by µ∞ on braids of
length k is not equal to µk , which is in line with the non-consistency illustrated in (1.2), but
the sequence (µk )k does converge weakly to µ∞ . The remarkable point is that the measure µ∞
2
has a Markovian structure which can be described explicitly. It makes it possible to get precise
information on µk for large k by using the limit µ∞ . For instance, we prove that the number of ∆
in a random braid of Bn+ is asymptotically geometric of parameter q n(n−1)/2 where q is the unique
root of smallest modulus of the Möbius polynomial of Bn+ . As another by-product of our results,
we settle the stable region conjecture, proving one of the two statements in the conjecture, and
refuting the other one. Our different results are achieved by strongly relying on refined properties
of the combinatorics of positive braids, some of them new.
Mutatis mutandis, the results also hold in the Birman-Ko-Lee dual braid monoid [10]. We present
the results in a unified way, with notations and conventions allowing to cover the braids and the
dual braids at the same time. The prerequisites on these two monoids are recalled in Section 2,
and the needed results on the combinatorics of braids are presented in Section 3. The main results
are proved in Section 4, with applications in Section 5, including the clarification of the stable
region conjecture. In Section 6, we provide explicit computations of the uniform measure µ∞ for
the braid monoid and the dual braid monoid on 4 strands. At last, analogs and extensions are
evoked in Section 7. Indeed, our results on braid monoids form a counterpart to the results on
trace monoids in [2, 3], and, in a forthcoming paper [1], we plan to prove results in the same spirit
for Artin-Tits monoids, a family encompassing both braids and traces.
2 Positive and dual positive braid monoids
In this section we introduce some basics on the monoid of positive braids and the monoid of positive
dual braids. We recall the notions of simple braids for these monoids, as well of combinatorial
representations of them.
2.1 Two distinct braid monoids
2.1.1 The braid group and two of its submonoids
For each integer n ≥ 2, the braid group Bn is the group with the following group presentation:
σ σ = σj σi
for |i − j| ≥ 2
Bn = σ1 , . . . , σn−1 i j
.
(2.1)
σi σj σi = σj σi σj for |i − j| = 1
Elements of Bn are called braids. Let e and “·” denote respectively the unit element and
the concatenation operation in Bn . It is well known since the work of Artin that elements of
Bn correspond to isotopy classes of braid diagrams on n strands, as illustrated in Figure 1; the
elementary move where strand i crosses strand i + 1 from above corresponds to generator σi , and
the move where strand i crosses strand i + 1 from behind to σi−1 .
We will be interested in two submonoids of Bn . The positive braids monoid Bn+ is the submonoid
of Bn generated by {σ1 , . . . , σn−1 } ; and the positive dual braid monoid Bn+∗ is the submonoid of
Bn generated by {σi,j | 1 ≤ i < j ≤ n} , where σi,j is defined by:
σi,j = σi ,
σi,j =
−1 −1
σi σi+1 . . . σj−1 σj−2
σj−3
for 1 ≤ i < n and j = i + 1,
. . . σi−1
for 1 ≤ i < n − 1 and i + 2 ≤ j ≤ n .
,
Observe the inclusion Bn+ ⊆ Bn+∗ , since each generator σi of Bn+ belongs to Bn+∗ . The elements
σi,j are often called Birman-Ko-Lee generators in the literature, while the elements σi are called
Artin generators.
Running examples for n = 3. Throughout the paper, we shall illustrate the notions and results
on the most simple, yet non trivial examples of braid monoids, namely on B3+ and on B3+∗ :
B3+ = hσ1 , σ2 i+ ,
B3+∗ = hσ1,2 , σ2,3 , σ1,3 i+
with σ1,2 = σ1 ,
3
σ2,3 = σ2 ,
σ1,3 = σ1 · σ2 · σ1−1 .
2
2
0
1
11
00
00
11
0
1
1 1
00
11
0
0000001
111111
000000
111111
1
0
000000
111111
1 111
0
000000
111111
000
000
111
000
111
3
3
1
0
1
0
0
1
00
11
0
1
00
11
4
4
Figure 2: Representation of σ1,3 in B4+∗
2.1.2 Presentations of the monoids
Defining Bn+ and Bn+∗ as submonoids of Bn , as we just did, is one way of introducing them.
Another one is through generators and relations.
First, Bn+ is isomorphic to the monoid with the monoid presentation (1.1), that is, the same
presentation as Bn but viewed as a monoid presentation instead of a group presentation. Second,
Bn+∗ is isomorphic to the monoid with n(n − 1)/2 generators σi,j for 1 ≤ i < j ≤ n and the
following relations, provided that the convention σj,i = σi,j for i < j is in force:
σi,j σj,k = σj,k σk,i = σk,i σi,j for 1 ≤ i < j < k ≤ n
(2.2)
σi,j σk,` = σk,` σi,j
for 1 ≤ i < j < k < ` ≤ n
σi,j σk,` = σk,` σi,j
for 1 ≤ i < k < ` < j ≤ n .
Elements of Bn+ are called positive braids, they correspond to isotopy classes of braid diagrams
involving only crossing of strands in the same direction, see Figure 1. Elements of Bn+∗ are called
dual positive braids. They correspond to isotopy classes of chord diagrams [10]. This time, there
are still n strands but they are arranged along a cylinder; the element σi,j corresponds to a crossing
of strands i and j. See Figure 2.
The inclusion Bn+ ⊆ Bn+∗ comes with the definition of Bn+ and Bn+∗ as submonoids of the braid
group Bn . It can be obtained as follows when considering Bn+ and Bn+∗ as abstract monoids with
generators and relations. Let ι : {σ1 , . . . , σn } → Bn+∗ be defined by ι(σi ) = σi,i+1 , and keep the
notation ι to denote the natural extension on the free monoid ι : {σ1 , . . . , σn }∗ → Bn+∗ . It is clear
that ι is constant on congruence classes of positive braids, whence ι factors through ι : Bn+ → Bn+∗ .
It can then be proved that this morphism is injective [22, 10].
Remark 2.1. — We emphasize that all the notions that we are about to define on Bn+ and on Bn+∗
may or may not coincide on Bn+ ∩ Bn+∗ = Bn+ . Henceforth, it is probably clearer to keep in mind
the point of view on these monoids through generators and relations, rather than as submonoids
of Bn .
Running examples for n = 3. The presentations of the monoids B3+ and B3+∗ are the following:
B3+ = hσ1 , σ2 | σ1 σ2 σ1 = σ2 σ1 σ2 i+
B3+∗ = hσ1,2 , σ2,3 , σ1,3 | σ1,2 σ2,3 = σ2,3 σ1,3 = σ1,3 σ1,2 i+
2.1.3 A common notation
We will consider simultaneously the monoids Bn+ and Bn+∗ . Henceforth, we will denote by Bn? a
monoid which, unless stated otherwise, may be either the monoid Bn+ or Bn+∗ . The statements
that we will prove for the monoid Bn? will then hold for both monoids Bn+ and Bn+∗ .
In addition, we will denote by Σ the set of generators of Bn? , hence Σ = {σi : 1 ≤ i ≤ n − 1} if
?
Bn = Bn+ and Σ = {σi,j : 1 ≤ i < j ≤ n} if Bn? = Bn+∗ .
4
2.1.4 Length and division relations. Mirror mapping
The above presentations (1.1) and (2.2) of Bn? are homogeneous, meaning that the relations involve
words of the same lengths. Hence, the length of x ∈ Bn? , denoted by |x|, is the length of any word
in the equivalence class x, with respect to the congruence defining Bn? .
Remark 2.2. — The length is an example of a quantity which is defined on both Bn+ and on Bn+∗ ,
and which is invariant on Bn+ . That is to say, the length |x| of a positive braid x ∈ Bn+ is invariant
whether x is considered as an element of Bn+ or as an element of Bn+∗ . Indeed, if x has length
k as an element of Bn+ , then it necessarily writes as a product x = σϕ(1) · . . . · σϕ(k) for some
function ϕ : {1, . . . , k} → {1, . . . , n − 1} . This entails that x, as an element of Bn+∗ , writes as
x = σϕ(1),ϕ(1)+1 · . . . · σϕ(k),ϕ(k)+1 , and thus x also has length k as an element of Bn+∗ .
The monoid Bn? is equipped with the left and with the right divisibility relations, denoted respectively ≤l and ≤r , which are both partial orders on Bn? , and are defined by:
x ≤l y ⇐⇒ ∃z ∈ Bn?
y = x · z,
x ≤r y ⇐⇒ ∃z ∈ Bn?
y = z · x.
The mirror mapping, defined on words by a1 . . . ak 7→ ak . . . a1 , factorizes through Bn? and
induces thus a mirror mapping on braids, denoted by x ∈ Bn? 7→ x∗ ∈ Bn? . It is an involutive antiisomorphism of monoids, it preserves the length of braids and swaps the left and right divisibility
relations:
∀x ∈ Bn?
|x∗ | = |x| ,
∀x, y ∈ Bn?
x ≤r y ⇐⇒ x∗ ≤l y ∗ .
The mirror mapping being an isomorphism of partial orders (Bn? , ≤l ) → (Bn? , ≤r ), we shall focus
on the left divisibility relation ≤l only.
Remark 2.3. — Following Remark 2.2, it is clear that the left divisibility is also invariant on Bn+ :
if x, y ∈ Bn+ are such that x ≤l y in Bn+ , then x ≤l y also holds in Bn+∗ . Observe however that the
converse is not true. For instance, consider the case n = 3 and set x = σ2 and y = σ1 · σ2 . In B3+ ,
clearly, x ≤l y does not hold. However, in B3+∗ , we have x = σ2,3 and y = σ1,2 · σ2,3 = σ2,3 · σ1,3 ,
therefore x ≤l y does hold.
2.2 Garside structure and simple braids
2.2.1 Garside structure
The monoid Bn? is known to be a Garside monoid [4, 6, 10]; that is to say:
(1) Bn? is a cancellative monoid;
(2) Bn? contains a Garside element, that is to say, an element whose set of left divisors coincides
with its set of right divisors and contains the generating set Σ ;
bound in (Bn? , ≤l ), and a greatest lower bound
(3) Every finite subset X of Bn? has
W a least upper
V
if X 6= ∅, respectively denoted l X and l X.
W
?
?
Let r X denote the least upper
W boundWin (Bn , ≤r ) of a subset X ⊆ Bn . Then, if X is a subset
of Σ, it is known [10, 22] that l X and r X coincide. We introduce therefore the notation ∆X
for:
_
_
∆X = l X = r X ,
for X ⊆ Σ .
Moreover, ∆X has the same left divisors and right divisors in Bn? : {x ∈ Bn? : x ≤l ∆X } = {x ∈
Bn? : x ≤r ∆X } .
The element ∆X one obtains when considering X = Σ plays a special role in Garside theory.
Indeed, based on the above remarks, it is not difficult to see that ∆Σ is a Garside element of Bn? ,
and is moreover the smallest Garside element of Bn? . Defining the elements ∆n ∈ Bn+ and δn ∈ Bn+∗
by:
∆n = (σ1 · . . . · σn−1 ) · (σ1 · . . . · σn−2 ) · . . . · (σ1 · σ2 ) · σ1 ,
δn = σ1,2 · σ2,3 · . . . · σn−1,n ,
we have ∆Σ = ∆n if Bn? = Bn+ and ∆Σ = δn if Bn? = Bn+∗ . We adopt the single notation ∆ = ∆Σ
to denote either ∆n or δn according to which monoid we consider.
5
∆3
δ3
σ2 · σ1
σ1 · σ2
σ1,2
σ1
σ2,3
σ1,3
σ2
e
e
Figure 3: Hasse diagrams of (S3 , ≤l ) for B3+ at left and for B3+∗ at right. Elements of D3 are
framed.
2.2.2 Definition of simple braids
The set of all divisors of ∆ is denoted by Sn , and its elements are called simple
is a
W braids. It V
bounded subset of Bn? , with e as minimum and ∆ as maximum, closed under l and under l .
With the induced partial order, (Sn , ≤l ) is thus a lattice.
Consider the mapping Φ : P(Σ) → Sn , X 7→ ∆X , and its image:
Dn = {∆X : X ⊆ Σ} .
Then Φ : P(Σ), ⊆ → (Sn , ≤l ) is a lattice homomorphism, and Dn is thus a sub-lattice of (Sn , ≤l ).
If Bn? = Bn+ , the mapping Φ is injective, but not onto Sn , and so (Dn , ≤l ) is isomorphic to
P(Σ), ⊆ . If Bn? = Bn+∗ , the mapping Φ is not injective, but it is onto Sn , hence Dn = Sn .
Remark 2.4. — Contrasting with the length discussed in Remark 2.2, the notion of simplicity
is not invariant on Bn+ . For instance, the braid ∆n is simple in Bn+ , but it is not simple as an
element of Bn+∗ as soon as n ≥ 3. Indeed, since its length is |∆n | = n(n − 1)/2 , it cannot be a
divisor of δn which is of length |δn | = n − 1.
Running examples for n = 3.
The Garside elements of B3+ and of B3+∗ are:
δ3 = σ1,2 · σ2,3 ,
∆3 = σ 1 · σ 2 · σ 1 .
The Hasse diagrams of (S3 , ≤l ) are pictured in Figure 3. For B3+ , the lattice (D3 , ≤l ) consists
of the following four elements:
∆∅ = e
∆σ 1 = σ 1
∆σ2 = σ2
∆σ1 ,σ2 = σ1 · σ2 · σ1 = ∆
whereas the lattice (S3 , ≤l ) contains the two additional elements σ1 · σ2 and σ2 · σ1 . For B3+∗ , the
elements of D3 and S3 are:
∆∅ = e
∆{σ1,2 } = σ1,2
∆{σ2,3 } = σ2,3
∆{σ1,3 } = σ1,3
∆{σ1,2 ,σ2,3 } = ∆{σ2,3 ,σ1,3 } = ∆{σ1,2 ,σ1,3 } = ∆{σ1,2 ,σ2,3 ,σ1,3 } = δ3
2.2.3 Combinatorial representations of simple braids
The natural map that sends each generator σi of the braid group Bn to the transposition (i, i + 1)
induces a morphism from Bn to Sn , the set of permutations of n elements. Hence, this map also
induces morphisms from Bn+ and from Bn+∗ to Sn .
In the case of the braid monoid Bn+ , this morphism induces a bijection from Sn to Sn . Thus,
Sn has cardinality n! . The element e corresponds to the identity permutation, and the element
∆n to the mirror permutation i 7→ n + 1 − i.
6
3
•
2
•
4•
3
•
•1
•
5
2
•
4•
•
6
3
•
•1
•
5
•
6
2
•
4•
•1
•
5
•
6
Figure 4: Left, representation of σ2,3 · σ4,5 · σ5,6 in B6+∗ as a non-crossing partition of {1, . . . , 6}.
Middle, representation of σ1,5 · σ5,6 . Right, representation of their least upper bound
(σ2,3 · σ4,5 · σ5,6 ) ∨l (σ1,5 · σ5,6 ) = σ2,3 · σ1,4 · σ4,5 · σ5,6 .
From the point of view of braid diagrams, such as the one pictured in Figure 1, simple braids
correspond to braids such that in any representative braid diagram, any two strands cross at most
once.
In the case of the dual braid monoid Bn+∗ , this morphism only induces an injection from Sn
to Sn . It is thus customary to consider instead the following alternative representation. Recall that
a partition {T 1 , . . . , T m } of {1, . . . , n} is called non-crossing if the sets {exp(2ikπ/n) : k ∈ T i }
have pairwise disjoint convex hulls in the complex plane. For instance, on the left of Figure 4, is
illustrated the fact that {{1}, {2, 3}, {4, 5, 6}} is a non-crossing partition of {1, 2, . . . , 6}.
Now, for each subset T of {1, . . . , n}, let xT denote the dual braid σt1 ,t2 ·σt2 ,t3 ·. . .·σtk−1 ,tk , where
t1 < t2 < . . . < tk are the elements of T . Then, for each non-crossing partition T = {T 1 , . . . , T m }
of {1, . . . , n}, we denote by xT the (commutative) product xT 1 · . . . · xT m . It is known [10, 7]
that the mapping T 7→ xT is a lattice isomorphism from the set of non-crossing partitions
of
2n
1
{1, . . . , n} onto Sn . Thus in particular, the cardinality of Sn is the Catalan number n+1
n . In
this representation, e corresponds to the finest partition {{1}, . . . , {n}}, and δn to the coarsest
partition {{1, . . . , n}}. See Figure 4.
Running examples for n = 3. Let us consider the case n = 3. For B3+ , the correspondence
between simple braids and permutations of {1, 2, 3} is the following:
e = Id
σ1 · σ2 = (1, 2, 3)
σ1 = (1, 2)
σ2 · σ1 = (1, 3, 2)
σ2 = (2, 3)
∆ = (3, 1)
Simple braids of B3+∗ correspond to non-crossing partitions of {1, 2, 3}, which in this case are
simply all the partitions of {1, 2, 3}. The correspondence is the following, where singletons are
omitted:
e=
σ1,2 = {1, 2}
σ2,3 = {2, 3}
σ1,3 = {1, 3}
δ3 = {1, 2, 3}
3 Garside normal form and combinatorics of braids
Braids are known to admit normal forms, that is to say, a unique combinatorial representation for
each braid. Normal forms are the standard tool for several algorithmic problems related to braids,
for instance the word problem to cite one of the most emblematic [17]. Among the several normal
forms for braids introduced in the literature, we shall focus in this work on the Garside normal
form which derives from the Garside structure attached to our braid monoids, as recalled above.
3.1 Garside normal form of braids
In the monoid Bn? , and regardless on whether Bn? = Bn+ or Bn? = Bn+∗ , a sequence (x1 , . . . , xk ) of
simple braids is said to be normal if xj = ∆∧l (xj ·. . .·xk ) holds for all j = 1, . . . , k. Intuitively, this
is a maximality property, meaning that no left divisor of xj+1 · · · xk could be moved to xj while
7
remaining in the world of simple braids. We recall the two following well known facts concerning
normal sequences of braids:
1. For x, y ∈ Sn , let x → y denote the relation R(x) ⊇ L(y), where the sets R(x) and L(y) are
defined as follows:
R(x) = σ ∈ Σ : x · σ ∈
/ Sn } ,
L(y) = σ ∈ Σ : σ ≤l y .
(3.1)
Then a sequence (x1 , . . . , xk ) is normal if and only if xj → xj+1 holds for all j = 1, . . . , k − 1,
again meaning that left divisors of xj+1 are already present in xj , and therefore cannot be
pushed into xj while keeping it simple.
2. For every non-unit braid x ∈ Bn? , there exists a unique normal sequence (x1 , . . . , xk ) of nonunit simple braids such that x = x1 · . . . · xk . This sequence is called the Garside normal
form or decomposition of x.
In this work, the integer k is called the height of x (it is also called the supremum of x in the
literature [20]). We denote it by τ (x).
Regarding the special elements e and ∆, the following dual relations hold, meaning that e is
“final” whereas ∆ is “initial”:
∀x ∈ Sn
x→e
∀x ∈ Sn
∆→x
∀x ∈ Sn
∀x ∈ Sn
e → x ⇐⇒ x = e
x → ∆ ⇐⇒ x = ∆
Therefore, the Garside decomposition starts with a finite (possibly zero) number of ∆s, and then
is given by a finite path in the finite directed graph (Sn \ {∆, e}, →). By convention, we define
the Garside normal form of the unit braid e as the sequence (e), and we put τ (e) = 1. (It might
seem that τ (e) = 0 would be a more natural convention, but it turns out that taking τ (e) = 1 is
the good choice for convenient formulation of several results below as it encompasses the fact that,
in normal forms, e can not be followed by any letter. For instance, Lemma 4.3 would not hold
otherwise.) Then, it is a well-known property of Garside monoids that τ (x) is the least positive
integer k such that x is a product of k simple braids.
Moreover, it will be convenient to complete the normal form of a braid with infinitely many
factors e. We call the infinite sequence (xk )k≥1 of simple braids thus obtained the extended
Garside decomposition of the braid. The directed graph (Sn , →) is then their accepting graph:
extended Garside decompositions of braids correspond bijectively to infinite paths in (Sn , →) that
eventually hit e, and then necessarily stay in e forever.
Remark 3.1. — Following up on Remark 2.4, just as simplicity was observed not to be invariant
on Bn+ , the height and the Garside normal forms are not invariant on Bn+ . For instance, the braid
∆n has Garside normal form the sequence (∆n ) itself in Bn+ , and the sequence (δn , δn−1 , . . . , δ2 )
in Bn+∗ . Its height is 1 in Bn+ and n − 1 in Bn+∗ .
We gather in the following proposition some well-known properties of Garside normal forms [18]
that we shall use later.
Proposition 3.2. For all braids x, y ∈ Bn? :
1. the height τ (x) is the smallest integer k ≥ 1 such that x ≤l ∆k ;
2. if (x1 , . . . , xk ) is the normal form of x, then x1 · . . . · xj = x ∧l ∆j for all j ∈ {1, . . . , k} ;
3. x ≤l y ⇐⇒ x ≤ y ∧l ∆τ (x) .
Remark 3.3. — We should stress that, while the normal form is very convenient to enumerate
braids, it behaves poorly with respect to multiplication: consider x with height k and Garside
decomposition (x1 , . . . , xk ), and let σ be a generator. Then, y = x · σ has height in {k, k + 1}, but
if it has height k then the normal form y = (y1 , . . . , yk ) might be completely different from that of
x (in the sense that y1 6= x1 , . . . , yk 6= xk ), although it is algorithmically computable.
8
L(x) x ∈ S3
∅
e
{σ1 }
σ1
{σ2 }
σ2
{σ1 } σ1 · σ2
{σ2 } σ2 · σ1
Σ
∆3
R(x)
∅
{σ1 }
{σ2 }
{σ2 }
{σ1 }
Σ
{y ∈ S3 : x → y}
{e}
{e, σ1 , σ1 · σ2 }
{e, σ2 , σ2 · σ1 }
{e, σ2 , σ2 · σ1 }
{e, σ1 , σ1 · σ2 }
S3
/ σ1 · σ2
O
6
σ1
? Y
?
*
/e
4 A _
∆
σ2
(
/ σ2 · σ1
_
Figure 5: Left hand side: subsets L(x) and R(x) for simple braids x ∈ B3+ . Right hand side: the
resulting accepting graph (S3 , →) for B3+
L(x) x ∈ S3
∅
e
{σ1,2 }
σ1,2
{σ2,3 }
σ2,3
{σ1,3 }
σ1,3
Σ
δ3
R(x)
{y ∈ S3
∅
{e}
{σ1,2 , σ1,3 } {e, σ1,2 ,
{σ2,3 , σ1,2 } {e, σ2,3 ,
{σ1,3 , σ2,3 } {e, σ1,3 ,
Σ
S3
e ho
O X
: x → y}
σ1,3 }
σ1,2 }
σ2,3 }
σ2,3
H O
v
σ1,2
/ σ1,3
_
?
δ3
_
Figure 6: Left hand side: subsets L(x) and R(x) for simple braids x ∈ B3+∗ . Right hand side: the
resulting accepting graph (S3 , →) for B3+∗
9
Running examples for n = 3. Let us describe explicitly the accepting graphs (S3 , →) for B3+ and
for B3+∗ . Consider first the case of B3+ . The subsets L(x) and R(x) are easily computed through
their definition (3.1), from which the relation → is derived. The results of these computations are
depicted in Figure 5. The analogous computations for B3+∗ result in the data pictured in Figure 6.
3.2 Combinatorics of braids
How many braids x ∈ Bn? of length k are there? Is there either an exact or an approximate formula?
The aim of this subsection is to recall the classical answers to these questions, which we will do by
analyzing the ordinary generating function, or growth series, of Bn? . This series turns out to be
rational, and the Garside normal form is an efficient tool for the study of its coefficients.
3.2.1 Growth series of braid monoids and Möbius polynomial
Let Gn (t) be the growth series of Bn? . It is defined by:
X
Gn (t) =
t|x| .
?
x∈Bn
According to a well-known result [19, 15, 12], the growth function of Bn? is rational, inverse of a
polynomial:
X
1
Gn (t) =
,
with Hn (t) =
(−1)|X| t|∆X | .
(3.2)
Hn (t)
X⊆Σ
There exist explicit or recursive formulæ allowing to compute effectively Hn (t):
Hn (t) =
n
X
(−1)k+1 t
k(k−1)
2
Hn−k (t)
if Bn? = Bn+
k=1
Hn (t) =
n−1
X
(−1)k
k=0
(n − 1 + k)!
tk
(n − 1 − k)!k!(k + 1)!
if Bn? = Bn+∗
For reasons that will appear more clearly in a moment (see Subsection 3.3), the polynomial
Hn (t) deserves the name of Möbius polynomial of Bn? .
Running examples for n = 3.
as follows:
For B3+ , the computation of the Möbius polynomial may be done
H3 (t) = 1 − t|σ1 | − t|σ2 | + t|σ1 ∨σ2 | = 1 − 2t + t3
since σ1 ∨ σ2 = ∆3
and similarly for B3+∗ :
H3 (t) = 1 − t|σ1,2 | − t|σ2,3 | − t|σ1,3 | + t|σ1,2 ∨σ1,3 | + t|σ1,2 ∨σ2,3 | + t|σ2,3 ∨σ1,3 | − t|δ3 |
= 1 − 3t + 2t2
since σ1,2 ∨ σ1,3 = σ1,2 ∨ σ2,3 = σ2,3 ∨ σ1,3 = δ3
3.2.2 Connectivity of the Charney graph
The growth series Gn (t) is a rational series with non-negative coefficients and with a finite positive
radius of convergence, say qn , which, by the Pringsheim theorem [21], is necessarily itself a pole
of Gn (t). Since Gn (t) = 1/Hn (t) as recalled in (3.2), it follows that qn is a root of minimal modulus
of the polynomial Hn (t). In order to evaluate the coefficients of Gn (t), we shall prove that Gn (t)
has no other pole of modulus qn , or equivalently, that Hn (t) has no other root of modulus qn .
This is stated in Corollary 3.5 below.
To this end, we first study the connectivity of the Charney graph, which is the directed graph
G = (V, E) with set of vertices V = Sn \ {e, ∆} and set of edges E = {(x, y) ∈ V 2 : x → y}. The
connectivity of G is well known for Bn? = Bn+ [8, 14, 25], and actually the same result also holds for
Bn? = Bn+∗ , although it does not seem to be found in the literature. We obtain thus the following
result.
10
Proposition 3.4. For n ≥ 3, the Charney graph of Bn? is strongly connected and contains loops.
Proof. First, observe that the graph G contains the loop σ → σ for every generator σ ∈ Σ. Since
proofs of the strong connectivity of G are found in the literature when Bn? = Bn+ , we focus on
proving that G is strongly connected when Bn? = Bn+∗ .
Recall that simple braids are in bijection with non-crossing partitions of {1, . . . , n}. For each
subset T of {1, . . . , n}, we denote by xT the braid σt1 ,t2 ·σt2 ,t3 ·. . .·σtk−1 ,tk , where t1 < t2 < . . . < tk
are the elements of T . Then, for each non-crossing partition T = {T 1 , . . . , T m } of {1, . . . , n}, we
denote by xT the (commutative) product xT 1 · . . . · xT m . It is known [10] that (1) the mapping
T 7→ xT is a bijection from the set of non-crossing partitions of {1, . . . , n} to Sn , as mentioned in
Subsection 2.2; (2) the set L(xT ) is equal to {σu,v : ∃T ∈ T u, v ∈ T }; (3) the set R(xT ) is equal
to
σu,v : ∃T ∈ T T ∩ {u + 1, . . . , v} =
6 ∅ and T ∩ {1, . . . , u, v + 1, . . . , n} =
6 ∅ .
Hence, consider two braids y, z ∈ Sn \ {e, ∆}, and let m = b n2 c, as well as the set A =
{σ1,2 , . . . , σn−1,n , σn,1 }. Since z <l ∆, we know that z = xZ where Z is a partition of {1, . . . , n} in
at least two subsets. It follows that Z is a refinement of a non-crossing partition V of {1, . . . , n}
in exactly two subsets. The map σi,j 7→ σi+1,j+1 induces an automorphism
of the dual braid
monoid Bn+∗ , hence we assume without loss of generality that V = {1, . . . , i, n}, {i+1, . . . , n−1}
for some integer i ∈ {0, . . . , m − 1}.
Finally, consider some generator σa,b ∈ L(y), with a < b. Since L(y) ⊆ R(y), it comes that that
y → σa,b → σ1,a → σ1,n → σ2,n if 1 < a, or y → σ1,b → σ1,n → σ2,n if a = 1. Since 2 ≤ m + 1 < n,
we then observe that σ2,n → σ1,m+1 → xT → xU → xV → xZ = z, where
(
{u, n + 1 − u} : 1 ≤ u ≤ m
if n is even,
T=
{u, n + 1 − u} : 1 ≤ u ≤ m ∪ {m + 1}
if n is odd;
U = {1, . . . , m}, {m + 1, . . . , n} .
This completes the proof.
The connectivity of the Charney graph stated in the above proposition has several consequences
on the combinatorics of braids, which we gather in the following corollary, the result of which seems
to have been unnoticed so far.
Corollary 3.5. The Möbius polynomial Hn (t) has a unique root of smallest modulus. This root,
say qn , is real and lies in (0, 1) and it is simple. It coincides with the radius of convergence of the
growth series Gn (t).
Furthermore, for each integer k ≥ 0, put λn (k) = #{x ∈ Bn? : |x| = k} . Then, for n ≥ 3, the
following asymptotics hold for some constant Cn > 0:
λn (k) ∼k→∞ Cn qn−k .
(3.3)
Proof. Recall that we have already defined qn , at the beginning of Subsection 3.2.2, as the radius
of convergence of Gn (t), and we know that qn is a root of smallest modulus of Hn (t). We will now
derive the two statements of the corollary through an application of Perron-Frobenius theory for
primitive matrices (see, e.g., [28]).
Let I∆ and I¬∆ denote the sets
I∆ = {(i, ∆) : 1 ≤ i ≤ |∆|} ,
I¬∆ = {(i, σ) : σ ∈ Sn \ {e, ∆} and 1 ≤ i ≤ |σ|} ,
and let I denote the disjoint union I∆ ∪ I¬∆ . Let M = (Mx,y )x,y∈I be the non-negative matrix
defined as follows:
1, if j = i + 1 and σ = τ ,
M(i,σ),(j,τ ) = 1, if i = |σ| and j = 1 and σ → τ ,
0, otherwise.
By construction, M is a block triangular matrix M = ( A0 B
C ) where A, B and C are the restrictions of M to the respective sets of indices I∆ × I∆ , I∆ × I¬∆ and I¬∆ × I¬∆ .
11
Since the Charney graph is strongly connected and contains loops according to Proposition 3.4,
and since it contains at least n − 1 vertices (the elements of Σ), it follows that C is a primitive
matrix with Perron eigenvalue ρ > 1. By construction, we know that A|∆| = Id|∆| , hence that the
eigenvalues of A have modulus 1. Consequently, ρ is a simple eigenvalue of M , and has a strictly
greater modulus than all other eigenvalues of M . Hence, there exist left and right eigenvectors l and
r of M for the eigenvalue ρ with non-negative entries, whose restrictions (lx )x∈I¬∆ and (rx )x∈I¬∆
only have positive entries, and such that l · r = 1.
Then, observe that λn (k) = u · M k−1 · v for all k ≥ 1, where u is the row vector defined by
u(i,σ) = 1(i = 1) and v is the column vector defined by v(i,σ) = 1(i = |σ|). Indeed, this follows
at once from the existence and uniqueness of the Garside normal form for braids, and from the
construction of the matrix M .
Both vectors u and v have some non-zero entries in I¬∆ , and therefore
λn (k) = u · M k−1 · v ∼k→∞ ρk−1 (u · r)(l · v)
(3.4)
Hence, ρ−1 is the radius of convergence of the generating series Gn (t) = k≥0 λn (k), and thus
ρ−1 = qn .
To complete the proof, consider the decomposition of Hn (t) as a product of the form:
P
Hn (t) = (1 − t/qn ) · (1 − t/r1 ) · . . . · (1 − t/ri ) · (1 − t/a1 ) · . . . · (1 − t/aj ) ,
where r1 , . . . , ri are the other complex roots of Hn (t) of modulus qn , including qn if its multiplicity
is > 1, and a1 , . . . , aj are the remaining complex roots of Hn (t), hence of modulus greater than qn .
Since we know that Gn (t) = 1/Hn (t), and considering the series expansion of 1/Hn (t), one sees
that the equivalent found in (3.4) for the coefficients λn (·) of Gn (t) cannot hold if i > 0, whence
the result.
3.3 Two Möbius transforms
This last subsection is devoted to the study of Möbius transforms. In § 3.3.1, we particularize
to our braid monoids Bn? the classical Möbius transform, as defined for general classes of partial
orders [27, 29]. We prove the Möbius inversion formula for our particular case, although it could be
derived from more general results. Next, we introduce in § 3.3.2 a variant, called graded Möbius
transform, which will prove to be most useful later for the probabilistic analysis.
We will use extensively the notation 1(A) for the characteristic function of A, equal to 1 if A is
true and to 0 is A is false.
3.3.1 The standard Möbius transform
In the framework of braid monoids, the usual Möbius transform is defined as follows and leads to
the next proposition.
Definition 3.6. Given a real-valued function f : Bn? 7→ R, its Möbius transform is the function
h : Bn? 7→ R defined by:
X
h(x) =
(−1)|X| f (x · ∆X ) for all x ∈ Bn? .
(3.5)
X⊆Σ
P
Proposition 3.7. Let f, h : Bn? → R be two functions such that the series
? |f (x)| and
x∈Bn
P
|h(x)|
are
convergent.
Then
h
is
the
Möbius
transform
of
f
if
and
only
if
?
x∈Bn
∀x ∈ Bn?
f (x) =
X
h(x · y) .
(3.6)
?
y∈Bn
Proof. For every non-unit braid z ∈ Bn? , the sets L(z) and L(z ∗ ) are non-empty, hence theWpowersets
PL(z) and PL(z ∗ ) are non-trivial Boolean lattices. It follows from the equality ∆X = l X that:
X
X
(−1)|X| 1(∆X ≤l z) =
(−1)|X| = 0 .
X⊆Σ
X⊆L(z)
12
W
And, similarly, it follows from the equality ∆X = r X that:
X
X
(−1)|X| 1(∆X ≤r z) =
(−1)|X| = 0 .
X⊆L(z ∗ )
X⊆Σ
P
P
Consider now two functions f and h such that the series x∈B ? |f (x)| and x∈B ? |h(x)| are
n
n
convergent. Assume first that h is the Möbius transform of f . Using the change of variable
v = y · ∆X , we derive from (3.5) the following:
X X
f (x) =
(−1)|X| 1(∆X ≤r v) f (x · v)
? X⊆Σ
v∈Bn
=
X X
X
(−1)|X| f (x · y · ∆X ) =
X⊆Σ
?
y∈Bn
h(x · y),
?
y∈Bn
proving (3.6).
Conversely, if (3.6) holds, we use the change of variable u = ∆X · y to derive:
X X
(−1)|X| 1(∆X ≤l u) h(x · u)
h(x) =
? X⊆Σ
u∈Bn
=
X X
?
y∈Bn
(−1)|X| h(x · ∆X · y) =
X⊆Σ
X
(−1)|X| f (x · ∆X ).
X⊆Σ
This shows that h is the Möbius transform of f , completing the proof.
In particular, observe that, if a function f has support in Sn , then so does its Möbius transform h.
Hence, we also define the notion of Möbius transform of real-valued functions f : Sn → R in a
natural way. In that narrower context, Proposition 3.7 formulates as follows.
Corollary 3.8. Let f, h : Sn → R be two functions. Then the two statements:
X
∀x ∈ Sn f (x) =
1(x · y ∈ Sn )h(x · y)
(3.7)
y∈Sn
∀x ∈ Sn
h(x) =
X
(−1)|X| 1(x · ∆X ∈ Sn )f (x · ∆X )
(3.8)
X⊆Σ
are equivalent.
In particular, by comparing the expressions (3.2) of Hn and (3.8) of the Möbius transform of
f : Sn → R, we observe that if p is a real number, and if f : Sn → R is defined by f (x) = p|x| ,
then its Möbius transform h satisfies:
h(e) = Hn (p).
(3.9)
Running examples for n = 3. We tabulate in Table 1 the values of the Möbius transform of the
function p|x| defined on S3 , for B3+ and for B3+∗ . It is easily computed based on the elements
found in Figures 5 and 6 respectively.
3.3.2 The graded Möbius transform
The above relation between real-valued functions f : Bn? 7→ R and their Möbius transforms works
only when the Möbius transform is summable. In order to deal with all functions defined on Bn? ,
we introduce a variant of those transforms, which is the notion of graded Möbius transform. To
this end, for each finite braid x ∈ Bn? , we define Bn? [x] as the following subset:
Bn? [x] = {y ∈ Bn? : τ (x · y) = τ (x)} ,
13
x ∈ S3
e
σ1
σ2
σ1 · σ2
σ2 · σ1
∆3
x ∈ S3
e
σ1,2
σ2,3
σ1,3
δ3
h(x)
1 − 2p + p3 = H3 (p)
p − p2
p − p2
p2 − p3
p2 − p3
p3
h(x)
1 − 3p + 2p2 = H3 (p)
p − p2
p − p2
p − p2
p2
Table 1: Values of the Möbius transform h : S3 → R of the function f : S3 → R defined by
f (x) = p|x| for B3+ (left hand side) and for B3+∗ (right hand side)
Definition 3.9. Given a real-valued function f : Bn? 7→ R, its graded Möbius transform is the
function h : Bn? 7→ R defined by:
X
∀x ∈ Bn? h(x) =
(−1)|X| 1(∆X ∈ Bn? [x])f (x · ∆X ) .
(3.10)
X⊆Σ
For functions that vanish outside of Sn , the notions of Möbius transform and graded Möbius
transform coincide, while this is not the case in general.
The generalization of the summation formula (3.6) stands in the following result.
Theorem 3.10. Let f, h : Bn? → R be two functions. Then h is the graded Möbius transform of f
if and only if
X
∀x ∈ Bn? f (x) =
h(x · y) .
(3.11)
? [x]
y∈Bn
Note that, in formula (3.11), the braids y ∈ Bn? [x] may have normal forms that differ completely
from that of x. This relates with Remark 3.3.
Proof. For a generic braid x ∈ Bn? of height k = τ (x), we denote by (x1 , . . . , xk ) the Garside
decomposition of x. Observe that, for all x, y, z ∈ Bn? , we have:
y ∈ Bn? [x] ∧ z ∈ Bn? [x · y] ⇐⇒ y · z ∈ Bn? [x].
(3.12)
Indeed, y ∈ Bn? [x] ∧ z ∈ Bn? [x · y] ⇐⇒ τ (x) = τ (x · y) = τ (x · y · z) ⇐⇒ τ (x) = τ (x · y · z) ⇐⇒
y · z ∈ Bn? [x].
Hence, if h is the graded Möbius transform of f , then:
X
X X
h(x · y) =
(−1)|X| 1(y ∈ Bn? [x])1(∆X ∈ Bn? [x · y])f (x · y · ∆X ) by (3.10)
? X⊆Σ
y∈Bn
? [x]
y∈Bn
=
X X X
(−1)|X| 1(v ∈ Bn? [x])1(v = y∆X )f (x · v)
? y∈B ? X⊆Σ
v∈Bn
n
=
X X
?
v∈Bn
=
(−1)|X| 1(v ∈ Bn? [x])1(∆X ≤r v)f (x · v)
X⊆Σ
X X
(−1)|X| 1(v ∈ Bn? [x])f (x · v)
? X⊆L(v ∗ )
v∈Bn
= f (x).
14
by (3.12) with z = ∆X
Conversely, if (3.11) holds, then:
X
(−1)|X| 1(∆X ∈ Bn? [x])f (x · ∆X )
X⊆Σ
=
X X
(−1)|X| 1(∆X ∈ Bn? [x])1(z ∈ Bn? [x · ∆X ])h(x · ∆X · z)
?
X⊆Σ z∈Bn
=
X X X
(−1)|X| 1(u ∈ Bn? [x])1(u = ∆X · z)h(x · u)
? X⊆Σ z∈B ?
u∈Bn
n
X
=
X
(−1)|X| 1(∆X ≤l u)h(x · u)
? [x] X⊆Σ
u∈Bn
X X
=
(−1)|X| h(x · u)
? [x] X⊆L(u)
u∈Bn
= h(x).
This completes the proof.
3.3.3 Additional properties of Möbius transforms
Finally, we state in this subsection a couple of lemmas which we will use in next section for the
probabilistic study.
Lemma 3.11. For p a real number, let f : Sn → R be defined by f (x) = p|x| , and let h : Sn → R
be the Möbius transform of f . Let also g : Sn → R be defined by:
X
g(x) =
1(x → y)h(y) .
(3.13)
y∈Sn
Then the identity h(x) = f (x)g(x) holds for all x ∈ Sn .
Proof. Let P = P(Σ), and consider the two functions F, G : P → R defined, for A ∈ P , by:
X
X
F (A) =
(−1)|I| 1(I ⊆ L(∆Σ\A ))f (∆I ) ,
G(A) =
1(L(y) ∩ L(∆Σ\A ) = ∅)h(y) .
I∈P
y∈Sn
Then we claim that F = G.
Let us prove the claim. For every I ∈ P and for every y ∈ Sn , we have:
I ⊆ L(y) ⇐⇒ ∆I ≤l y ⇐⇒ L(∆I ) ⊆ L(y).
Therefore, according to the Möbius summation formula (3.6), we have:
X
f (∆I ) =
1(I ⊆ L(y))h(y).
y∈Sn
Reporting the right hand member above in the sum defining F (A), and inverting the order of
summations, yields:
X X
X
F (A) =
(−1)|I| 1(I ⊆ L(∆Σ\A ) ∩ L(y)) h(y) =
1(L(∆Σ\A ) ∩ L(y) = ∅)h(y) = G(A),
y∈Sn I∈P
y∈Sn
which proves the claim.
Now observe that, for every x ∈ Sn , we have
_
_
x · ∆Σ\R(x) = l {x · σ : σ ∈ Σ \ R(x)} ≤ l Sn = ∆Σ ,
and therefore L(∆Σ\R(x) ) = Σ \ R(x). This proves that
x · ∆I ∈ Sn ⇐⇒
x → y ⇐⇒
I ∩ R(x) = ∅ ⇐⇒ I ⊆ L(∆Σ\R(x) )
(3.14)
L(y) ⊆ R(x) ⇐⇒ L(y) ∩ L(∆Σ\R(x) ) = ∅.
(3.15)
15
Hence, using the multiplicativity of f , we have simultaneously
X
h(x) =
(−1)|I| 1(I ⊆ L(∆Σ\R(x) ))f (x · ∆I ) = f (x)F (R(x))
I∈P
X
and g(x) =
1(L(y) ∩ L(∆Σ\R(x) = ∅)h(y) = G(R(x))
y∈Sn
Since F = G, it implies h(x) = f (x)g(x), which completes the proof of the lemma.
Lemma 3.12. Let (x1 , . . . , xk ) be the Garside decomposition of a braid x ∈ Bn? and let X be a
subset of Σ. We have:
∆X ∈ Bn? [x] ⇐⇒ ∆X ∈ Bn? [xk ].
(3.16)
Proof. The result is immediate if x = e. Moreover, if x 6= e and if ∆X ∈ Bn? [xk ], we observe that
xk ∆X is a simple braid, and therefore that x1 · . . . · xk−1 · (xk ∆X ) is a factorization of x into k
simple braids, whence τ (x∆X ) ≤ τ (x). Since the Garside length is non-decreasing for ≤l , it follows
that ∆X ∈ Bn? [x].
Conversely, if x 6= e and if ∆X ∈
/ Bn? [xk ], since the set Sn is closed under ∨l , there must exist
?
some generator σ ∈ X \ Bn [xk ]. Hence, we have x1 → . . . xk → σ, and therefore τ (x · ∆X ) ≥
τ (x · σ) = k + 1, i.e. ∆X ∈
/ Bn? [x].
Corollary 3.13. Let f : Bn? → R be the function defined by f (x) = p|x| . Then the graded Möbius
transform h : Bn? → R of f satisfies the following property:
h(x) = p|x1 |+...+|xk−1 | h(xk ) ,
(3.17)
where (x1 , . . . , xk ) is the Garside decomposition of x.
Proof. It follows directly from the definition (3.5) of the graded Möbius transform, together with
Lemma 3.12.
4 Uniform measures on braid monoids
We are now equipped with adequate tools to study uniform measures on braids. Consider the
following (vague) questions: how can we pick a braid uniformly at random? How can we pick a
large braid uniformly at random? What are the characteristics of such random braids?
Since there are countably many braids, these questions cannot be given immediately a consistent
meaning. However, for each fixed integer k ≥ 0, there are finitely many braids of size k, and it
is thus meaningful to pick uniformly a braid of size k at random. Please notice the difference
between picking a braid of size k uniformly at random, and picking a word uniformly in Σk and
then considering the braid it induces. The later corresponds to the uniform random walk on Bn? ,
but not the former.
A possible way of picking a braid at random is thus the following: first pick the size k at random,
and then pick a braid uniformly among those of size k. The problem remains of how to draw k in
a “natural” way. It is the topic of this section to demonstrate that there is indeed a natural family,
indexed by a real parameter p, of conducting this random procedure. Furthermore, the parameter
p is bound to vary in the interval (0, qn ), where qn is the root of Hn (t) introduced earlier; and
letting p tend to qn by inferior values, the distributions induced on braids weight large braids more
and more, such that at the limit we obtain a natural uniform measure on “infinite braids”. In turn,
we shall derive in the next section information on large random braids, that is to say, on random
braids of size k when k is large enough, based on the notion of uniform measure on infinite braids.
4.1 Generalized braids
Considering the extended Garside decomposition of braids, one sees that elements of Bn? are in
bijection with infinite paths in (Sn , →) that eventually hit e. Therefore, it is natural to define a
16
∗
compactification Bn? as the set of all infinite paths in this graph. As a subset of SnN , it is endowed
with a canonical topology, for which it is compact. Moreover, the restriction of this topology to
Bn? is the discrete topology, and Bn? is the closure of Bn? . This is the set of generalized braids. We
endow the set Bn? with its Borel σ-algebra. All measures we shall consider on Bn? will be finite and
Borelian.
We may refer to elements of Bn? as finite braids, to emphasize their status as elements of Bn? .
We define the boundary ∂Bn? by:
∂Bn? = Bn? \ Bn? .
Elements in ∂Bn+ correspond to infinite paths in (Sn , →) that never hit e, we may thus think of
them as infinite braids.
If (x1 , . . . , xp ) is a finite path in the graph (Sn , →), the corresponding cylinder set D(x1 ,...,xp )
is defined as the set of paths starting with vertices (x1 , . . . , xp ). Cylinder sets are both open and
closed, and they generate the topology on Bn? .
Definition 4.1. For x ∈ Bn? of Garside decomposition (x1 , . . . , xp ), we define the Garside cylinder,
and we denote by Cx , the cylinder subset of Bn? given by Cx = D(x1 ,...,xp ) .
Garside cylinders only reach those cylinders sets of the form D = D(x1 ,...,xp ) with xp 6= e . But
if xp = e , then D reduces to the singleton {x}, with x = x1 · . . . · xp . And then, denoting by q the
greatest integer with xq 6= e and writing y = x1 · . . . · xq , one has:
[
Cy·z .
{x} = Cy \
z∈Sn \{e} : xq →z
It follows that Garside cylinders generate the topology on Bn? , which implies the following result.
Proposition 4.2. Any finite measure on the space Bn? of generalized braids is entirely determined
by its values on Garside cylinders. In other words, if ν and ν 0 are two finite measures on Bn? such
that ν(Cx ) = ν 0 (Cx ) for all x ∈ Bn? , then ν = ν 0 .
Proof. We have already seen that Garside cylinders generate the topology, and thus the Borel
σ-algebra of Bn? . The collection of Garside cylinders, augmented with the empty set, is obviously
stable by intersection:
∀x, y ∈ Bn?
either Cx ⊆ Cy or Cx ⊇ Cy or Cx ∩ Cy = ∅ ,
and forms thus a so-called π-system. The result follows from classical measure theory [9, Th. 3.3].
Garside cylinders are very natural from the point of view of the normal forms, however they
are somewhat unnatural from the algebraic point of view as they discard most of the divisibility
information (cf. Remark 3.3). A more natural notion is that of visual cylinder, which corresponds,
for a given finite braid x ∈ Bn? , to the subset of those generalized braids which are “left divisible”
by x. It will be useful to differentiate between generalized braids and infinite braids, therefore we
introduce both the full visual cylinder ⇑ x and the visual cylinder ↑ x, as follows:
⇑ x = Closure {x · z : z ∈ Bn? } ,
↑ x = ⇑ x ∩ ∂Bn? ,
where Closure(A) denotes the topological closure of the set A.
The relationship between Garside cylinders and visual cylinders is given by the following result.
Lemma 4.3. For each finite braid x ∈ Bn? , the full visual cylinder ⇑ x is the following disjoint
union of Garside cylinders:
[
⇑x=
Cx·y .
(4.1)
? [x]
y∈Bn
Proof. We first observe that ⇑ x ∩ Bn? = y∈Bn? [x] (Bn? ∩ Cx·y ). Indeed, the ⊇ inclusion is obvious,
while
the converse one is a consequence of points 2 and 3 of Proposition
3.2. Since ⇑ x and
S
S
?
?
?
C
are
the
respective
topological
closures
of
⇑
x
∩
B
and
of
?
?
x·y
n
y∈Bn [x]
y∈Bn [x] (Bn ∩ Cx·y ) in Bn ,
the result follows.
S
Hence, ⇑ x, as a finite union of Garside cylinders, is also open and closed in Bn? . In the same
way, ↑ x is open and closed in ∂Bn? .
17
4.2 Studying finite measures on generalized braids via the graded Möbius
transform
In this subsection, we study finite measures on the set Bn? of generalized braids.
Assume that ν is some finite measure on Bn? . Then for practical purposes, we are mostly
interested in the values of ν on Garside cylinders ν(Cx ). However, most natural measures will
enjoy good properties with respect to the full visual cylinders ⇑ x, which is not surprising as these
sets are most natural from the point of view of divisibility properties. For instance, the limits ν
of uniform measures on the set of braids of length k will satisfy ν( ⇑ x) = p|x| for some p, see
Definition 4.6 and Theorem 5.1 below.
Henceforth, to understand these measures, we need to relate ν(Cx ) and ν( ⇑ x) in an explicit
way, and this is where the graded Möbius transform of Subsection 3.3.2 plays a key role, as shown
by Proposition 4.4 below. In turn, Proposition 4.4 provides a nice probabilistic interpretation of
the graded Möbius transform.
Proposition 4.4. Let ν be a finite measure on Bn? . Let f : Bn? 7→ R be defined by f (x) = ν( ⇑ x) ,
and let h : Bn? 7→ R be the graded Möbius transform of f . Then, for every integer k ≥ 1 and for
every finite braid y of height k, holds:
ν(Cy ) = h(y).
(4.2)
Proof. The decomposition (4.1) of a full visual cylinder as a disjoint union of Garside cylinders
shows that
X
ν( ⇑ x) =
ν(Cx·y ).
? [x]
y∈Bn
Thus, the characterization (3.11) of the graded Möbius transforms shows that y 7→ ν(Cy ) is the
graded Möbius transform of x 7→ ν( ⇑ x), as claimed.
Corollary 4.5. A finite measure ν on Bn? is entirely determined by its values ν( ⇑ x) on the
countable collection of full visual cylinders.
Proof. According to Proposition 4.2, a finite measure ν is entirely determined by its values on
Garside cylinders. Hence the result follows from Proposition 4.4.
4.3 Uniform measures
Our ultimate goal is to understand the uniform measure µn,k on braids in Bn? of a given length k,
when k tends to infinity. We will see below in Theorem 5.1 that this sequence of measures converges
to a measure on ∂Bn? which behaves nicely on the visual cylinders ↑ x (this is not surprising as
these are the natural objects from the point of view of the monoid structure on Bn? ).
Therefore, it is good methodology to study the general class of measures which do behave
nicely on full visual cylinders. Our usual conventions and notations are in force throughout this
subsection, in particular concerning Bn? which may be either Bn+ or Bn+∗ .
Definition 4.6. A uniform measure for braids of parameter p ≥ 0 is a probability measure νp on
Bn? satisfying:
∀x ∈ Bn?
νp ( ⇑ x) = p|x| .
Although not apparent from this definition, we will see in Theorem 4.8 below that such a
measure either weights Bn? or ∂Bn? , but not both. Theorem 4.8 will describe quite precisely all
uniform measures. It will allow us to define the uniform measure at infinity as the unique non
trivial uniform measure supported by the boundary ∂Bn+ , see Definition 4.9. A realization result
for uniform measures will be the topic of Subsection 4.4.
Before coming to the theorem, we state a key lemma.
18
Lemma 4.7. Let ν be a uniform measure of parameter p < 1. Assume that ν is concentrated at
infinity, i.e., ν(∂Bn? ) = 1. Then Hn (p) = 0.
Furthermore, let B = (Bx,x0 ) be the non-negative matrix indexed by pairs of simple braids (x, x0 )
such that x, x0 ∈ Sn \ {e, ∆}, and defined by:
0
Bx,x0 = 1(x → x0 )p|x | .
Then B is a primitive matrix of spectral radius 1. The Perron right eigenvector of B is the
restriction to Sn \ {e, ∆} of the vector g defined in (3.13).
Proof. Let f (x) = p|x| , and let h : Sn → R be the graded Möbius transform of f .
According to Proposition 4.4, we have h(e) = ν(Ce ) = ν({e}). Since it is assumed that ν(Bn? ) =
0, it follows that h(e) = 0. But Hn (p) = h(e), as previously stated in (3.9), hence Hn (p) = 0,
proving the first claim of the lemma.
Let g be defined on Sn as in (3.13), and let ge be the restriction of g to Sn \ {e, ∆}. It follows
from Lemma 3.11 that h(x) = p|x| g(x) holds on Sn . Therefore the computation of Be
g goes as
follows, for x ∈ Sn \ {e, ∆}:
X
X
(Be
g )x =
1(x → y)p|y| g(y) =
1(x → y)h(y).
y∈Sn \{e,∆}
y∈Sn \{e,∆}
But h(e) = 0 on the one hand; and on the other hand, x → ∆ does not hold since x 6= ∆. Hence
the above equality rewrites as:
X
(Be
g )x =
1(x → y)h(y) = ge(x).
y∈Sn
We have proved that ge is right invariant for B. Let us prove that ge is non identically zero.
Observe that h is non-negative, as a consequence of Proposition 4.4. Therefore g is non-negative
as well. If ge were identically zero on Sn \{e, ∆}, so would be h on Sn \{∆}. The Möbius summation
formula (3.7) would imply that f is constant, equal to f (∆) on Sn , which is not the case since we
assumed p 6= 1. Hence ge is not identically zero.
But B is also aperiodic and irreducible, hence primitive. Therefore Perron-Frobenius theory [28,
Chapter 1] implies that ge is actually the right Perron eigenvector of B, and B is thus of spectral
radius 1.
Theorem 4.8. For each braid monoid Bn? , uniform measures νp on Bn? are parametrized by the
parameter p ranging exactly over the closed set of reals [0, qn ] ∪ {1}.
1. For p = 0, ν0 is the Dirac measure at e.
2. For p = 1, ν1 is the Dirac measure on the element ∆∞ defined by its infinite Garside
decomposition: (∆ · ∆ · . . .) .
3. For p ∈ (0, qn ), the support of νp is Bn? , and it is equivalently characterized by:
νp {x} = Hn (p) · p|x|
or
νp ( ⇑ x) = p|x|
(4.3)
for x ranging over Bn? .
4. For p = qn , the support of νqn is ∂Bn? , and it is characterized by:
∀x ∈ Bn?
νqn ( ↑ x) = qn|x| .
(4.4)
It follows from this statement that, except for the degenerated measure ν1 , there exists a unique
uniform measure on the boundary ∂Bn? . It is thus natural to introduce the following definition.
|x|
Definition 4.9. The uniform measure on ∂Bn? which is characterized by νqn ( ↑ x) = qn
x ∈ Bn? , is called the uniform measure at infinity.
19
for
Proof of Theorem 4.8. The statements 1–4 contains actually three parts: the existence of νp for
each p ∈ [0, qn ] ∪ {1}, the uniqueness of the measures satisfying the stated characterizations, and
that [0, qn ] ∪ {1} is the only possible range for p.
Existence and uniqueness of νp for p ∈ [0, qn ] ∪ {1}. The cases p = 0 and p = 1 (points 1 and 2)
are trivial. For p ∈ (0, qn ) (point 3), let νp be the discrete distribution on Bn? defined by the left
hand side of (4.3). Since p < qn , the series Gn (p) is convergent, and it implies that the following
formula is valid in the field of real numbers:
Gn (p) · Hn (p) = 1 .
It implies in particular that Hn (p) > 0, and thus:
X
νp {x} = 1 ,
?
x∈Bn
and therefore νp is a probability distribution on Bn? .
It remains to prove that νp is indeed uniform with parameter p. Since Bn? is left cancellative,
we notice that, for each x ∈ Bn? , the mapping y ∈ Bn? 7→ x · y is a bijection of Bn? onto ⇑ x ∩ Bn? .
Whence:
X
νp ( ⇑ x) = Hn (p) · p|x| ·
p|y| = p|x| .
?
y∈Bn
Conversely, if ν is a probability measure on Bn? such that ν( ⇑ x) = p|x| , then ν and νp agree on
full visual cylinders, hence ν = νp according to Corollary 4.5.
We now treat the case of point 4, corresponding to p = qn . For this, let (pj )j≥1 be any sequence
of reals pj < qn such that limj→∞ pj = qn , and such that (νpj )j≥1 is a weakly convergent sequence
of probability measures. Such a sequence exists since Bn? is a compact metric space. Let ν be the
weak limit of (νpj )j≥1 .
Obviously, for each braid x fixed:
lim νpj ( ⇑ x) = qn|x| .
j→∞
But ⇑ x is both open and closed in Bn? , it has thus an empty topological boundary. The Porte|x|
manteau theorem [9] implies that the above limit coincides with ν( ⇑ x) , hence ν( ⇑ x) = qn for
?
?
all x ∈ Bn . The same reasoning applied to every singleton {x}, for x ∈ Bn , yields:
|x|
pj
= 0,
j→∞ Gn (pj )
ν({x}) = lim νpj ({x}) = lim
j→∞
the later equality since limt→qn− Gn (t) = +∞ . Since Bn? is countable, it follows that ν puts weight
on ∂Bn? only, and thus:
∀x ∈ Bn? ν( ↑ x) = ν( ⇑ x) = qn|x| .
|x|
If ν 0 is a probability measure on Bn? satisfying ν 0 ( ↑ x) = qn for every x ∈ Bn? , then we observe
first that ν 0 is concentrated on the boundary, since ν 0 (∂Bn? ) = ν 0 ( ↑ e) = 1. And since ν and ν 0
coincide on all visual cylinders ↑ x, for x ranging over Bn? , it follows from Corollary 4.5 that ν = ν 0 .
Range of p. It remains only to prove that, if ν is a uniform probability measure on Bn? of
parameter p, then p = 1 or p ≤ qn . Seeking a contradiction, assume on the contrary that p > qn
and p 6= 1 holds.
We first show the following claim:
ν(∂Bn? ) = 1 .
(4.5)
Assume on the contrary ν(∂Bn? ) < 1, hence ν(Bn? ) > 0. Then we claim that the inclusionexclusion principle yields:
∀x ∈ Bn? ν {x} = Hn (p) · p|x| .
(4.6)
20
Indeed, for any braid x ∈ Bn? , the singleton {x} decomposes as:
[
{x} = ⇑ x \
⇑ (x · σ)
σ∈Σ
and therefore:
ν ({x}) =
X
(−1)|I| ν
I⊆Σ
\
X
⇑ (x · σ) =
(−1)|I| ν ⇑ (x · ∆I ) .
σ∈I
I⊆Σ
Note that the above equality is valid for any finite measure on Bn? . Since ν is assumed to be
uniform of parameter p, it specializes to the following:
X
ν ({x}) =
(−1)|I| p|x|+|∆I | = p|x| · Hn (p) ,
I⊆Σ
given the form (3.2) for Hn (p). This proves our claim (4.6).
Together with ν(Bn? ) > 0, it implies Hn (p) > 0. Consequently, summing up ν({x}) for x ranging
over Bn? yields Gn (p) < ∞. Hence p < qn , which is a contradiction since we assumed p > qn . This
proves the claim (4.5).
Next, consider the two matrices B and B 0 indexed by all braids x ∈ Sn \ {e, ∆} and defined by:
0
Bx,x0 = 1(x → x0 )p|x |
0
0
0 |x |
and Bx,x
.
0 = 1(x → x )qn
They are both non-negative and primitive, and of spectral radius 1 according to Lemma 4.7 (which
applies since p 6= 1 by assumption). According to Perron-Frobenius theory [28, Chapter 1], there
cannot exist a strict ordering relation between them. Yet, this is implied by p > qn , hence the
latter is impossible. The proof is complete.
Remark 4.10 (Multiplicative measures). — Since the length of braids is additive, any uniform measure is multiplicative, i.e., it satisfies: νp ( ⇑ (x · y)) = νp ( ⇑ x) · νp ( ⇑ y).
Conversely, assume that ν is a multiplicative probability measure on Bn? . Then ν is entirely
determined by the values pσ = ν( ⇑ σ) for σ ∈ Σ.
If Bn? = Bn+ , let us write pi instead of pσi . The braid relations σi · σi+1 · σi = σi+1 · σi · σi+1
entail: pi pi+1 (pi − pi+1 ) = 0. Hence, if any two consecutive pi , pi+1 are non zero, they must be
equal. Removing the generators σi for which pi = 0, the braid monoid splits into a direct product
of sub-braid monoids, each one equipped with a uniform measure.
Similarly, if Bn? = Bn+∗ , let us write pi,j instead of pσi ,σj . Then the dual braid relations σi,j ·σj,k =
σj,k · σk,i = σk,i · σi,j (if i < j < k) yield the following three relations: pj,k (pi,j − pk,i ) = 0,
pi,k (pi,j − pj,k ) = 0 and pi,j (pj,k − pi,k ) = 0. Therefore the following implication holds for all
i < j < k: (pi,j > 0 ∧ pj,k > 0) =⇒ pi,j = pj,k = pi,k . Removing the generators σi,j for which
pi,j = 0, the dual braid monoid splits thus into a direct product of sub-dual braid monoids, each
one equipped with a uniform measure.
Therefore, without loss of generality, the study of multiplicative measures for braid monoids
reduces to the study of uniform measures. This contrasts with other kinds of monoids, such as
heap monoids, see [3] and the discussion in Section 7.
4.4 Markov chain realization of uniform measures
Recall that generalized braids ξ ∈ Bn? are given by infinite sequences of linked vertices in the graph
(Sn , →). For each integer k ≥ 1, let Xk (ξ) denote the k th vertex appearing in an infinite path
ξ ∈ Bn? . This defines a sequence of measurable mappings
Xk : Bn? → Sn ,
which we may interpret as random variables when equipping Bn? with a probability measure, say
for instance a uniform measure νp .
It turns out that, under any uniform measure νp , the process (Xk )k≥1 has a quite simple form,
namely that of a Markov chain. This realization result is the topic of the following theorem (the
trivial cases p = 0 and p = 1 are excluded from the discussion).
21
Theorem 4.11. Let p ∈ (0, qn ], and let νp be the uniform measure of parameter p on Bn? . Let
h : Sn → R be the Möbius transform of x ∈ Sn 7→ p|x| .
1. Under νp , the process (Xk )k≥1 of simple braids is a Markov chain, taking its values in Sn if
p < qn , and in Sn \ {e} if p = qn .
2. The initial measure of the chain coincides with h, which is a probability distribution on Sn .
The initial distribution puts positive weight on every non unit simple braid, and on the unit
e if and only if p < qn .
3. The transition matrix P of the chain (Xk )k≥1 is the following:
Px,x0 = 1(x → x0 )p|x|
h(x0 )
,
h(x)
(4.7)
where x and x0 range over Sn for p < qn or over Sn \ {e} for p = qn .
Proof. Let f : Bn? → R be defined by f (x) = p|x| .
We first show that h > 0 on Sn if p < qn , and that h > 0 on Sn \ {e} if p = qn . Obviously, it
follows from Proposition 4.4 that h is non-negative on Sn , and even on Bn? .
1. Case p < qn . Then Hn (p) > 0 and therefore, according to Theorem 4.8 and Proposition 4.4,
we obtain:
h(x) = νp (Cx ) ≥ νp ({x}) = Hn (p) · p|x| > 0 for all x ∈ Sn ,
which was to be shown.
2. Case p = qn . Consider the matrix B indexed by pairs (x, x0 ) of simple braids distinct from e
|x0 |
and from ∆ , and defined by Bx,x0 = 1(x → x0 )qn . According to Lemma 4.7, the restriction
of g to Sn \ {e, ∆} is the Perron right eigenvector of B, where g has been defined in (3.13).
|x|
Therefore g > 0 on Sn \ {e, ∆}. But h(x) = qn g(x) holds on Sn according to Lemma 3.11,
therefore h > 0 on Sn \ {e, ∆}. As for ∆ , one has h(∆) = p|∆| > 0. Hence h > 0 on Sn \ {e},
as claimed.
It follows in particular from the above discussion that the matrix P defined in the statement
of the theorem is well defined. Now, let (x1 , . . . , xk ) be any sequence of simple braids (including
maybe the unit braid). Let δ and δ 0 denote the following quantities:
δ = νp (X1 = x1 , . . . , Xk = xk )
δ 0 = h(x1 ) · Px1 ,x2 · . . . · Pxk−1 ,xk .
We prove that δ = δ 0 .
We observe first that both δ and δ 0 are zero if the sequence (x1 , . . . , xk ) is not normal. Hence,
without loss of generality, we restrict our analysis to the case where (x1 , . . . , xk ) is a normal
sequence of simple braids.
Consider the braid y = x1 · . . . · xk . By the uniqueness of the Garside normal form, the following
equality holds:
{X1 = x1 , . . . , Xk = xk } = {X1 · . . . · Xk = y} .
Applying successively Proposition 4.4 and Corollary 3.13, we have thus:
δ = h(y) = p|x1 |+...+|xk−1 | h(xk ) .
On the other hand, we have:
δ 0 = h(x1 ) · p|x1 |
h(x2 )
h(xk )
· . . . · p|xk−1 |
= p|x1 |+...+|xk−1 | h(xk )
h(x1 )
h(xk−1 )
which completes the proof of the equality δ = δ 0 . It follows that (Xk )k≥1 is indeed a Markov chain
with the specified initial distribution and transition matrix.
If p = qn , then we have already observed in the proof of Lemma 4.7 that h(e) = 0. It implies
both ν(X1 = e) = 0 and Px,e = 0 for all x ∈ Sn \ {e}. We conclude that (Xk )k≥1 does never
reach e, which completes the proof of point 1, and of the theorem.
22
∆3
σ1
σ2
σ1 · σ2
σ2 · σ1
√
5−2
0
0
0
0
√
√
5−2
√ 5−2
( 5 − 1)/2 √ 0
0
(√5 − 1)/2
0
( 5 − 1)/2
√
( 5 − 1)/2
0
√
(7 − 3√ 5)/2
(3 − 5)/2
0
0
√
(3 − 5)/2
√
(7 − 3 5)/2
0
√
(3 − √5)/2
(3 − 5)/2
0
δ3
1/4 1/4 1/4 1/4
σ1,2
0 1/2 0 1/2
σ2,3
0 1/2 1/2 0
σ1,3
0
0 1/2 1/2
Table 2: Transition matrix of the Markov chain on simple braids induced by the uniform measure
at infinity for B3+ (up) and for B3+∗ (down). The initial distribution of the chain can be
read on the first line of each matrix.
Running examples for n = 3. We characterize the uniform measure at infinity both for B3+ and
for B3+∗ . For this, we first determine the root√of smallest modulus of the Möbius polynomial,
which we determined in Subsection 3.2.1: q3 = ( 5 − 1)/2 for B3+ and q3 = 1/2 for B3+∗ .
The Markov chain of simple braids induced by the uniform measure at infinity takes its values
in S3 \ {e}, which has 5 elements for B3+ and 4 elements for B3+∗ . Since the Möbius transform
|x|
of q3 is tabulated in Table 1, we are in the position to compute both the initial distribution and
the transition matrix of the chain by an application of Theorem 4.11, yielding the results given in
Table 2.
5 Applications to finite uniform distributions
5.1 Weak convergence of finite uniform distributions
The following result states a relationship between the finite uniform distributions, and the uniform
measure at infinity. If one were only interested in finite uniform distributions, that would be a
justification for studying uniform measures as defined previously.
Theorem 5.1. The uniform measure at infinity νqn is the weak limit of the sequence (µn,k )k≥0
as k → ∞ , where µn,k is for each integer k ≥ 0 the uniform distribution on the finite set Bn? (k)
defined by:
Bn? (k) = {x ∈ Bn? : |x| = k} .
Proof. Recall that Bn? (k), as a subset of Bn? , is identified with its image in Bn? , and thus µn,k
identifies with a discrete probability distribution on Bn? . We denote λn (k) = #Bn? (k). For a fixed
braid x ∈ Bn? , the map y 7→ x · y is a bijection between Bn? and ⇑ x ∩ Bn? , a fact that we already
used in the proof of Theorem 4.8, point 4. Hence, for any k ≥ |x|, and using the asymptotics found
in Corollary 3.5, one has:
µn,k ( ⇑ x) =
λn (k − |x|)
→k→∞ qn|x| .
λn (k)
Invoking the Portemanteau theorem as in the proof of Theorem 4.8, we deduce that any weak
|x|
cluster value ν of (µn,k )k≥0 satisfies ν( ⇑ x) = qn for any full visual cylinder ⇑ x . Theorem 4.8
implies ν = νqn . By compactness of Bn? , it follows that (µn,k )k≥0 converges toward νqn .
A practical interest of Theorem 5.1 lies in the following corollary. Define Xi : Bn? → Sn by
Xi (ξ) = xi , where (xk )k≥1 is the extended Garside normal form of ξ.
23
4
3
2
1
σ3
σ3
σ2
σ1
σ1
σ2
σ2
σ2
σ2
σ2
σ1
σ3
σ3
σ2
σ2
σ1
σ3
σ3
σ2
Figure 7: A random uniform braid on four strands. The first elements of its Garside decomposition
are: (σ3 ) · (σ3 σ2 σ1 ) · (σ1 σ2 ) · (σ2 ) · (σ2 ) · (σ2 ) · (σ2 σ1 σ3 ) · (σ3 σ2 ) · (σ2 σ1 σ3 ) · (σ3 σ2 )
Corollary 5.2. Let j ≥ 1 be an integer. Then the joint distribution of the first j simple braids
appearing in the extended Garside decomposition of a uniformly random braid of size k converges,
as k → ∞, toward the joint distribution of (X1 , . . . , Xj ) under the uniform measure at infinity.
Proof. By definition of the topology on Bn? , the mapping ξ ∈ Bn? 7→ (X1 , . . . , Xj ) is continuous for
each integer j ≥ 1. The result follows thus from Theorem 5.1.
Example for n = 4. In anticipation of the computations to be performed in Section 6, we depict
in Figure 7 the beginning of a “truly random infinite braid” on n = 4 strands, up to height k = 10.
These are “typical 10 first elements” in the decomposition of a large random braid on four strands.
Observe the absence of ∆; the numerical values found in next subsection make it quite likely.
5.2 A geometric number of ∆s.
The ∆ element only appears at the beginning of normal sequences of simple braids. Accordingly,
under the uniform probability measure νp , the occurrences of ∆ in the Markov chain (Xk )k≥1 are
only observed in the first indices, if any, and their number is geometrically distributed.
This behavior is quite easy to quantify, as the probabilistic parameters associated with ∆ have
simple expressions:
( n(n−1)
if Bn? = Bn+
p 2
|∆|
,
P∆,∆ = p|∆| .
νp (X1 = ∆) = h(∆) = p =
n−1
p
if Bn? = Bn+∗
It follows that the number of ∆s appearing in the normal form of a random braid, possibly
infinite and distributed according to a uniform measure of parameter p ∈ (0, qn ] , is geometric of
parameter p|∆| .
As a consequence of Theorem 5.1, we obtain this corollary.
Corollary 5.3. Let Tk : Bn? (k) → N denote the number of ∆s in the Garside decomposition of a
random braid of size k. Then (Tk )k≥0 converges in distribution, as k → ∞ , toward a geometric
n(n−1)
2
distribution of parameter qn
if Bn? = Bn+ , or qnn−1 if Bn? = Bn+∗ .
Authors are sometimes only interested by the elements of the Garside decomposition of a “large”
braid that appear after the last occurrence of ∆. The notion of uniform measure at infinity allows
also to derive information on these, as we shall see next.
Examples for n = 3 and n = 4. Exact or numerical values for the parameter of the geometrical
distribution are easily computed for n = 3 and for n = 4, based on our previous computations for
n = 3 and on the computations of Section 6 for n = 4; see Table 3.
5.3 On a conjecture by Gebhardt and Tawn
This subsection only deals with positive braids monoids Bn+ . In their Stable region Conjecture, the
authors of [24] suggest the following, based on a thorough experimental analysis. For each integer
i ≥ 1, let λi∗ (µk ) be the distribution of the ith factor in the extended Garside normal that occurs
after the last ∆ , for braids drawn at random uniformly among braids of length k. Then two facts
are suspected to hold, according to [24, Conjecture 3.1]:
24
n=3
n=4
monoid Bn+
parameter
prob. of occ. of ∆
√ 3
1
5
−2 + 2
≈ 0.236
≈ .309
≈ 0.0121
≈ 0.0122
monoid Bn+∗
parameter
prob. of occ. of ∆
1
4
1
2
√
5
10
−
3
1
3
≈ 0.021
≈ 0.022
Table 3: For each monoid, we tabulate on the left hand side the parameter of the geometric distribution of the number of ∆s appearing in a random infinite braid. On the right hand
side, we tabulate the associated probability of occurrence of at least one ∆.
Presumed Fact 1. For each integer i ≥ 1, the sequence (λi∗ (µk ))k≥1 is convergent as k → ∞ .
Presumed Fact 2. There exists a probability measure, say µ on Sn \ {e, ∆} , and a constant C > 0
such that holds:
λi∗ (µk ) → µ as k → ∞ .
∀i > C
(5.1)
Theorems 5.1 and 4.11 translate the problem of the limiting behavior of factors of the normal
form within the familiar field of Markov chains with a finite number of states. This brings a simple
way of determining the status of the above conjecture.
It follows from Theorem 5.1 that the distribution of the k th element in the extended Garside
decomposition of a random braid (including all the starting ∆s), distributed uniformly among
finite braids of size k, converges toward the distribution of the k th element in the extended Garside
decomposition of an infinite braid, distributed according to the unique uniform measure at infinity.
And, according to Theorem 4.11, this is the distribution of a Markov chain at time k, with the
prescribed initial distribution and transition matrix.
As for λi∗ (µk ) it converges thus toward the distribution of the same chain, i steps after having
left the state ∆ . Hence we can affirm the veracity of Fact 1. Using the notations of Theorem 4.11,
we may also describe the limit, say λi∗ = limk→∞ λi∗ (µk ) , as follows:
∀s ∈ Sn \ {e, ∆}
λ1∗ (s) =
h(s)
1 − qn
n(n−1)
2
,
where the denominator comes from the conditioning on s 6= ∆ . The next values for λi∗ are
obtained recursively:
∀i ≥ 1
λi∗ = λ1∗ P i−1 ,
(5.2)
where P is the transition matrix of the chain described in Theorem 4.11.
On the contrary, Fact 2 is incorrect. Indeed, keeping the notation λi∗ = limk→∞ λi∗ (µk ), if (5.1)
was true, then µ = λi∗ for i large enough, would be the invariant measure of the Markov chain
according to (5.2). But that would imply that the chain is stationary. We prove below that this is
not the case when n > 3. What is true however, is that (λi∗ )i≥1 converges toward the stationary
measure of the chain when i → ∞ .
Assume, seeking a contradiction, that the chain (Xk )k≥1 is stationary. That would imply that
|x|
the Möbius transform h of the function f (x) = qn , identified with a vector indexed by Sn \{e, ∆} ,
is left invariant for the transition matrix P . Hence, for y ∈ Sn \ {e, ∆}:
X
h(y) = (hP )y =
h(x)
x∈Sn \{e,∆} : x→y
f (x)h(y)
.
h(x)
We deduce, since h > 0 on Sn \ {e} :
X
∀y ∈ Sn \ {e, ∆}
x∈Sn \{e,∆} : x→y
25
qn|x| = 1 .
(5.3)
∆4
σ1 ·σ2 ·σ1 ·σ3 ·σ2 σ1 ·σ2 ·σ3 ·σ2 ·σ1 σ2 ·σ1 ·σ3 ·σ2 ·σ1
σ1 ·σ2 ·σ1 ·σ3
σ1 ·σ2 ·σ3
σ1 ·σ2 ·σ3 ·σ2
σ1 ·σ2 ·σ1
σ1 ·σ2
σ2 ·σ1 ·σ3 ·σ2
σ1 ·σ3 ·σ2
σ1 ·σ3 ·σ2 ·σ1
σ2 ·σ1 ·σ3
σ2 ·σ3 ·σ2 ·σ1
σ2 ·σ3 ·σ2
σ2 ·σ1
σ1 ·σ3
σ2 ·σ3
σ1
σ2
σ3
σ3 ·σ2 ·σ1
σ3 ·σ2
e
Figure 8: Hasse diagram of S4 for B4+ . Elements of D4 are circled
Consider y = σ1 and y 0 = σ1 · σ2 · σ1 . One has:
L(y) = {σ1 } ( L(y 0 ) = {σ1 , σ2 } ,
which entails:
{x ∈ Sn : (x 6= e, ∆) ∧ x → y} ) {x ∈ Sn : (x 6= e, ∆) ∧ x → y 0 } .
It follows that the equality stated in (5.3) cannot hold both for y and for y 0 , which is the sought
contradiction.
6 Explicit computations
We gather in this section the computations needed, for n = 4, to characterize the uniform measure
at infinity, both for B4+ and for B4+∗ .
6.1 Computations for B4+
The monoid B4+ has the following presentation:
B4+ = σ1 , σ2 , σ3
σ1 σ3 = σ3 σ1 , σ1 σ2 σ1 = σ2 σ1 σ1 , σ2 σ3 σ2 = σ3 σ2 σ3
+
.
In order to shorten notations, we denote a product of generators σi simply by the corresponding
sequence of indices. So for instance, the Garside element is denoted: ∆4 = 123121 .
The lattice D4 = {∆X | X ⊆ Σ4 } has 23 = 8 elements, and is isomorphic to the lattice of subsets
of {1, 2, 3}, whereas S4 has 4! = 24 elements. The Hasse diagram of S4 is depicted in Figure 8.
In order to compute the Möbius transform h of the function f (x) = p|x| defined on S4 , we refer
to the expression (3.8):
X
h(x) =
(−1)|X| f (x · ∆X )
X⊆Σ : x·∆X ∈S4
Furthermore, recalling the property x · ∆X ∈ S4 ⇐⇒ X ⊆ L(∆Σ\R(x) ) proved earlier in (3.14),
the range of those X ⊆ Σ such that x · ∆X ∈ S4 is directly derived from the knowledge of the sets
L(y) and R(y). All these elements are gathered in Table 4.
The Möbius polynomial H4 (t) can be obtained, for instance, by evaluating on e the Möbius
transform of the function x 7→ t|x| . From the first line of Table 4, we read:
H4 (t) = (1 − t)(1 − 2t − t2 + t3 + t4 + t5 )
26
L(x) x ∈ S4
∅
e
1
1
2
2
3
3
1
12
3, 1
13
2
21
2
23
3
32
2, 1
121
1
123
3, 1
132
2
213
3, 2
232
3
321
2, 1 1213
3, 1 1232
3, 1 1321
2 2132
3, 2 2321
3, 1 12321
3, 2 21321
2, 1 21323
3, 2, 1
∆4
R(x)
∅
1
2
3
2
1, 3
1
3
2
1, 2
3
2
1, 3
2, 3
1
1, 3
2, 3
1, 2
2
1, 3
1, 3
1, 2
2, 3
1, 2, 3
{y ∈ S4 : x → y}
e
1, 12, 123
2, 21, 23, 213, 2132
3, 32, 321
2, 21, 23, 213, 2132
1, 3, 12, 13, 32, 123, 132, 321, 1232, 1321, 12321
1, 12, 123
3, 32, 321
2, 21, 23, 213, 2132
1, 2, 12, 21, 23, 121, 123, 213, 1213, 2132, 21323
3, 32, 321
2, 21, 23, 213, 2132
1, 3, 12, 13, 32, 123, 132, 321, 1232, 1321, 12321
2, 3, 21, 23, 32, 213, 232, 321, 2132, 2321, 21321
1, 12, 123
1, 3, 12, 13, 32, 123, 132, 321, 1232, 1321, 12321
2, 3, 21, 23, 32, 213, 232, 321, 2132, 2321, 21321
1, 2, 12, 21, 23, 121, 123, 213, 1213, 2132, 21323
2, 21, 23, 213, 2132
1, 3, 12, 13, 32, 123, 132, 321, 1232, 1321, 12321
1, 3, 12, 13, 32, 123, 132, 321, 1232, 1321, 12321
1, 2, 12, 21, 23, 121, 123, 213, 1213, 2132, 21323
2, 3, 21, 23, 32, 213, 232, 321, 2132, 2321, 21321
S4
h(x)
1 − 3p + p2 + 2p3 − p6
p − 2p2 + p4
p − 2p2 + p3
p − 2p2 + p4
p2 − 2p3 + p4
p2 − p3
p2 − 2p3 + p5
p2 − 2p3 + p5
p2 − 2p3 + p4
p3 − p4
p3 − 2p4 + p6
p3 − 2p4 + p5
p3 − p4
p3 − p4
p3 − 2p4 + p6
p4 − p5
p4 − p5
p4 − p5
p4 − 2p5 + p6
p4 − p5
p5 − p6
p5 − p6
p5 − p6
p6
Table 4: Characteristic elements for B4+ . Elements of D4 are framed. The column h(x) tabulates
the Möbius transform of the function f (x) = p|x| on S4 .
27
Let p = q4 be the smallest root of H4 (t). We illustrate on the example x = 12 the computation
of a line Px,• of the transition matrix corresponding to the uniform measure at infinity. From
Table 4, we read the list of non zero values of the corresponding line of the matrix, which are for
this case: 2, 21, 23, 213 and 2132. According to Theorem 4.11, for x fixed, the entries Px,y of the
matrix are proportional to h(y), and the normalization factor is p−|x| h(x). Reading the values of h
in Table 4, we use the relation 1 − 2p − p2 + p3 + p4 + p5 = 0 to write the coefficients as polynomials
in p, yielding:
P12,2 = p ,
P12,213 = −1 + p + 2p2 + 2p3 + p4 ,
P12,21 = P12,23 = p(1 − 2p2 − 2p3 − p4 ) ,
P12,2132 = p4 .
6.2 Computations for B4+∗
We now treat the case of the B4+∗ . In order to simplify the notations, we write (ij) for the
generator σi,j , so for instance: δ4 = (12) · (23) · (34) . The monoid B4+∗ has the six generators (ij)
for 1 ≤ i < j ≤ 4, subject to the following relations:
(12) · (23) = (23) · (13) = (13) · (12)
(12) · (24) = (24) · (14) = (14) · (12)
(13) · (34) = (34) · (14) = (14) · (13)
(23) · (34) = (34) · (24) = (24) · (23)
(12) · (34) = (34) · (12)
(23) · (14) = (14) · (23)
The set of simple braids S4 has 14 elements, which we organize below according to the type
of partition of the integer 4 that the associated non-crossing partition of {1, 2, 3, 4} induces (see
Subsection 2.2.3):
S4 = e,
1 partition of type 1 + 1 + 1 + 1
1 partition of type 4
δ4 ,
(12), (13), (14), (23), (24), (34),
6 partitions of type 2 + 1 + 1
(12) · (23), (12) · (24), (13) · (34), (23) · (34),
4 partitions of type 3 + 1
(13) · (24), (12) · (34)
2 partitions of type 2 + 2
Following the same scheme as for B3+∗ , we gather in Table 5 the characteristic elements for B4+∗ .
In particular, the first line gives the Möbius polynomial, from which the characteristic value q4 is
derived:
√
5
1
2
.
H4 (t) = (1 − t)(1 − 5t + 5t ) ,
q4 = −
2
10
The computation of the transition matrix of the chain of simple braids induced by the uniform
measure at infinity yields the values reported in Table 6, where the line corresponding to δ4 has
been omitted. This line, which also corresponds to the initial measure of the chain, is given by the
function ρ(x) tabulated in Table 5.
7 Extensions
There are various other questions of interest concerning the asymptotic behavior of random braids,
uniformly distributed among braids of length k. For instance, what is the asymptotic value of the
height of a large braid? In other words, how do the Garside length and the Artin length compare
to each other for large braids?
The height of braids gives rise to a sequence of integer random variables τk : Bn? (k) → N, indexed
by k, where Bn? (k) = {x ∈ Bn? : |x| = k} is equipped with the uniform distribution. Since the
ratios height over length are uniformly bounded and bounded away from zero, performing the
28
x ∈ S4
{y ∈ S4 : x → y}
e
e
δ4
S4
(12)
(12), (13), (14)
(13)
(13), (14), (23), (24), (14) · (23)
(14)
(14), (24), (34)
(23)
(12), (23), (24)
(24)
(12), (13), (24), (34), (12) · (34)
(34)
(13),(23), (34)
(12) · (23) S4 \ δ4 , (34), (23) · (34), (13) · (34),
(12) · (24) S4 \ δ4 , (23), (12) · (23), (23) · (34),
(23) · (34) S4 \ δ4 , (14), (12) · (24), (13) · (34),
(13) · (34) S4 \ δ4 , (12), (12) · (23), (12) · (24),
(14) · (23) S4 \ δ4 , (13), (12) · (23), (13) · (34)
(12) · (34) S4 \ δ4 , (24), (12) · (24), (23) · (34)
(12) · (34)
(14) · (23)
(14) · (23)
(12) · (34)
h(x)
1 − 6p + 10p2 − 5p3
p3
p(1 − 3p + 2p2 )
p(1 − 2p + p2 )
p(1 − 3p + 2p2 )
p(1 − 3p + 2p2 )
p(1 − 2p + p2 )
p(1 − 3p + 2p2 )
p2 (1 − p)
p2 (1 − p)
p2 (1 − p)
p2 (1 − p)
p2 (1 − p)
p2 (1 − p)
ρ(x)
0
√
1/5
√ − 2 5/25
5/25 √
1/10
√ + 5/50
√5/25
5/25 √
1/10
√ + 5/50
5/25 √
1/10 − √5/50
1/10 − √5/50
1/10 − √5/50
1/10 − √5/50
1/10 − √5/50
1/10 − 5/50
Table 5: Characteristic elements for the dual braid monoid B4+∗ . The column h(x) gives the Möbius
transform on S4 of the function f (x) = p|x| . The column ρ(x) evaluates the same quantity
for the particular case p = q4 .
(12)
(13)
(14)
(23)
(24)
(34)
(12) · (23)
(12) · (24)
(23) · (34)
(13) · (34)
(14) · (23)
(12) · (34)
(12)
(13)
(14)
(23)
(24)
(34)
···
1/2 − θ
0
0
1/2 − θ
−1/2 + 3θ
0
−1/10 + θ
−1/10 + θ
−1/10 + θ
0
−1/10 + θ
−1/10 + θ
2θ
1/2 − θ
0
0
1/2 − θ
2θ
1/5
1/5
1/5
1/5
0
1/5
1/2 − θ
−1/2 + 3θ
1/2 − θ
0
0
0
−1/10 + θ
−1/10 + θ
0
−1/10 + θ
−1/10 + θ
−1/10 + θ
0
−1/2 + 3θ
0
1/2 − θ
0
1/2 − θ
−1/10 + θ
0
−1/10 + θ
−1/10 + θ
−1/10 + θ
−1/10 + θ
0
1/2 − θ
2θ
2θ
1/2 − θ
0
1/5
1/5
1/5
1/5
1/5
0
0
0
1/2 − θ
0
−1/2 + 3θ
1/2 − θ
0
−1/10 + θ
−1/10 + θ
−1/10 + θ
−1/10 + θ
−1/10 + θ
···
···
···
···
···
···
···
···
···
···
···
···
···
(12) · (23)
(12) · (24)
(23) · (34)
(13) · (34)
(14) · (23)
(12) · (34)
···
···
···
···
···
···
···
···
···
···
···
···
0
0
0
0
0
0
3/10 − θ
0
3/10 − θ
0
0
3/10 − θ
0
0
0
0
0
0
3/10 − θ
3/10 − θ
0
0
3/10 − θ
0
0
0
0
0
0
0
0
0
3/10 − θ
3/10 − θ
3/10 − θ
0
0
0
0
0
0
0
0
3/10 − θ
0
3/10 − θ
0
3/10 − θ
0
1 − 4θ
0
0
0
0
3/10 − θ
0
0
3/10 − θ
3/10 − θ
3/10 − θ
0
0
0
0
1 − 4θ
0
0
3/10 − θ
3/10 − θ
0
3/10 − θ
3/10 − θ
(12)
(13)
(14)
(23)
(24)
(34)
(12) · (23)
(12) · (24)
(23) · (34)
(13) · (34)
(14) · (23)
(12) · (34)
Table 6: Transition matrix for the uniform measure at infinity for B√4+∗ , restricted to its unique
ergodic component S4 \ {e, δ4 }, and where we have put θ = 5/10.
29
correct normalization leads to considering the sequence of real random variables ρk : Bn? (k) → R
defined by:
∀x ∈ Bn? (k)
ρk (x) =
τ (x)
τ (x)
=
,
|x|
k
which takes values in the fixed interval [1/|∆|, 1]. Since all these random variables are defined on
different probability spaces, the natural way of studying their asymptotic behavior is by studying
their convergence in distribution.
A first result one may wish to establish is a concentration result: one aims to prove that (ρk )k≥1
converges in distribution toward a single value, say ρ. Hence one expects a convergence in distribution of the following form, where δρ denotes the Dirac probability measure on the singleton {ρ}:
τ (·)
d
−−−−→ δρ ,
k k→∞
(7.1)
where ρ is some real number in the open interval (1/|∆|, 1). The number ρ would appear as a limit
average rate: most of braids of Artin size k would have, for k large enough, a Garside size close
to ρk. If ρ can furthermore be simply related to the quantities we have introduced earlier, such as
the characteristic parameter qn , it is reasonable to expect that ρ would be an algebraic number.
Once this would have been established, the next step would consist in studying a Central Limit
Theorem: upon normalization, is the distribution of ρk Gaussian around its limit value ρ? Hence,
one expects a convergence in distribution of the following form, for some constant σn2 > 0 and
where N (0, σ 2 ) denotes the Normal distribution of zero mean and variance σ 2 :
√ τ (·)
d
− ρ −−−−→ N (0, σn2 )
k
k→∞
k
(7.2)
It turns out that both results (7.1) and (7.2) hold indeed. Because of space constraints, we
postpone their proofs to a forthcoming work [1].
This concerned an extension of the results established in this paper. Generalizations to other
monoids are also possible, which we intend to expose in [1]. Braid monoids fall into the wider class
of Artin-Tits monoids, investigated by several authors since the 1960’s, including Tits, Deligne,
Sato, Brieskorn, Garside, Charney, Dehornoy. Several results established in this paper for braid
monoids admit generalizations to Artin-Tits monoids, and analogues of the convergences (7.1)
and (7.2) also hold.
Among Artin-Tits monoids, one class in particular has retained the attention of the authors:
the class of trace monoids, also called partially commutative monoids [13]. In trace monoids, the
only relations between generators are commutativity relations (there is no braid relations); they
correspond to Viennot’s heap monoids [31]. Trace monoids differ from braid monoids for several
reasons, for instance there is no lattice structure and their associated Coxeter group is not finite.
From the point of view adopted in this paper, the main difference lies in the existence of a continuum
of multiplicative measures, among which the uniform measure is a particular case. Recall that we
have observed in Remark 4.10 that the uniform measure for infinite braids is the only instance of
multiplicative measures, so the situation for braids presents a sharp contrast with trace monoids.
The investigation of multiplicative measures for trace monoids has been the topic of [2].
We shall prove in [1] that, from this perspective, there are essentially only two types of ArtinTits monoids: the trace type and the braid type, corresponding respectively to the type with a
continuum of multiplicative measures, and the type where multiplicative measures reduce to the
uniform measure only. For the trace type, multiplicative measures are parametrized by a submanifold of Rm , diffeomorphic to the standard (m − 1)-simplex, where m is the minimal number
of generators of the monoid.
References
[1]
S. Abbes, S. Gouëzel, V. Jugé, and J. Mairesse. “Uniform and multiplicative measures at
infinity for Artin-Tits monoids and applications”. In preparation.
30
[2]
S. Abbes and J. Mairesse. “Uniform and Bernoulli measures on the boundary of trace
monoids”. In: J. Combin. Theory Ser. A 135 (2015), pp. 201–236.
[3]
S. Abbes and J. Mairesse. “Uniform generation in trace monoids”. In: Math. Found. Comput.
Sc. 2015 (MFCS 2015), part 1. Ed. by G. Italiano, G. Pighizzini, and D. Sannella. Vol. 9234.
Lecture Notes in Comput. Sci. Springer, 2015, pp. 63–75.
[4]
S.I. Adian. “Fragments of the word ∆ in the braid group”. In: Mat. Zametki 36.1 (1984),
pp. 25–34.
[5]
M. Albenque and P. Nadeau. “Growth function for a class of monoids”. In: 21st Internat.
Conf. Formal Power Series Algebraic Combin. (FPSAC 2009). DMTCS Proceedings, 2009,
pp. 25–38.
[6]
D. Bessis. “The dual braid monoid”. In: Ann. Sci. Éc. Norm. Supér. (4). Vol. 36. 5. 2003,
pp. 647–683.
[7]
D. Bessis, F. Digne, and J. Michel. “Springer theory in braid groups and the Birman-Ko-Lee
monoid”. In: Pacific J. Math. 205.2 (2002), pp. 287–309.
[8]
M. Bestvina. “Non-positively curved aspects of Artin groups of finite type”. In: Geom. Topol.
3.1 (1999), pp. 269–302.
[9]
P. Billingsley. Probability and Measure, 3rd edition. Wiley, 1995.
[10]
J. Birman, K.H. Ko, and S.J. Lee. “A new approach to the word and conjugacy problems in
the braid groups”. In: Adv. Math. 139.2 (1998), pp. 322–353.
[11]
M. Brazil. “Monoid growth functions for braid groups”. In: Internat. J. Algebra Comput. 1.2
(1991), pp. 201–205.
[12]
A. Bronfman. “Growth functions of a class of monoids”. Preprint. 2001.
[13]
P. Cartier and D. Foata. Problèmes combinatoires de commutation et réarrangements. Vol. 85.
Lecture Notes in Math. Springer, 1969.
[14]
S. Caruso. “On the genericity of pseudo-Anosov braids I: rigid braids”. arXiv preprint
arXiv:1306.3757. 2013.
[15]
R. Charney. “Geodesic automation and growth functions for Artin groups of finite type”. In:
Math. Ann. 301 (1995), pp. 307–324.
[16]
P. Dehornoy. “Combinatorics of normal sequences of braids”. In: J. Combin. Theory Ser. A
114.3 (2007), pp. 389–409.
[17]
P. Dehornoy. “Efficient solutions to the braid isotopy problem”. In: Disc. Appl. Math 156
(2008), pp. 3094–3112.
[18]
P. Dehornoy, F. Digne, E. Godelle, D. Krammer, and J. Michel. “Foundations of Garside
theory”. arXiv preprint arXiv:1309.0796. 2013.
[19]
P. Deligne. “Les immeubles des groupes de tresses généralisés”. In: Inv. Math. 17.4 (1972),
pp. 273–302.
[20]
D. Epstein et al. Word processing in groups. Boston: Jones and Bartlett, 1992.
[21]
P. Flajolet and R. Sedgewick. Analytic combinatorics. Cambridge Univ. Press, 2009.
[22]
F. Garside. “The braid group and other groups”. In: Quarterly J. Math. 20.1 (1969), pp. 235–
254.
[23]
V. Gebhardt and J. Gonzáles-Meneses. “Generating random braids”. In: J. Combin. Theory
Ser. A 120.1 (2013), pp. 111–128.
[24]
V. Gebhardt and S. Tawn. “Normal forms of random braids”. In: J. Algebra 408 (2014),
pp. 115–137.
[25]
V. Gebhardt and S. Tawn. “On the penetration distance in Garside monoids”. In: J. Algebra
451 (2016), pp. 544–576.
[26]
J. Mairesse and F. Mathéus. “Randomly growing braid on three strands and the manta ray”.
In: Ann. Appl. Probab. 17.2 (2007), pp. 502–536.
31
[27]
G.-C. Rota. “On the foundations of combinatorial theory I. Theory of Möbius functions”. In:
Z. Wahrscheinlichkeitstheorie 2 (1964), pp. 340–368.
[28]
E. Seneta. Non-negative Matrices and Markov Chains. Revised printing. Springer, 1981.
[29]
R.P. Stanley. Enumerative combinatorics, vol. 1. Cambridge Univ. Press, 1997.
[30]
A. Vershik, S. Nechaev, and R. Bikbov. “Statistical properties of locally free groups with
applications to braid groups and growth of random heaps”. In: Comm. Math. Physics 212.2
(2000), pp. 469–501.
[31]
X. Viennot. “Heaps of pieces, I : basic definitions and combinatorial lemmas”. In: Combinatoire énumérative. Vol. 1234. Lecture Notes in Math. Springer, 1986, pp. 321–350.
[32]
W. Woess. Random walks on infinite graphs and groups. Cambridge Tracts in Mathematics
138. Cambridge University Press, 2000.
32
| 4math.GR
|
A scalable approach for tree segmentation within small-footprint airborne
LiDAR data
Hamid Hamraza*, Marco A. Contrerasb, Jun Zhanga
a: Department of Computer Science, b: Department of Forestry
University of Kentucky, Lexington, KY 40506, USA
[email protected], [email protected], [email protected]
* Corresponding Author:
[email protected] +1 (859) 489 1261
Abstract.
This paper presents a distributed approach that scales up to segment tree crowns within a LiDAR point cloud
representing an arbitrarily large forested area. The approach uses a single-processor tree segmentation algorithm as
a building block in order to process the data delivered in the shape of tiles in parallel. The distributed processing is
performed in a master-slave manner, in which the master maintains the global map of the tiles and coordinates the
slaves that segment tree crowns within and across the boundaries of the tiles. A minimal bias was introduced to the
number of detected trees because of trees lying across the tile boundaries, which was quantified and adjusted for.
Theoretical and experimental analyses of the runtime of the approach revealed a near linear speedup. The estimated
number of trees categorized by crown class and the associated error margins as well as the height distribution of the
detected trees aligned well with field estimations, verifying that the distributed approach works correctly. The
approach enables providing information of individual tree locations and point cloud segments for a forest-level area
in a timely manner, which can be used to create detailed remotely sensed forest inventories. Although the approach
was presented for tree segmentation within LiDAR point clouds, the idea can also be generalized to scale up
processing other big spatial datasets.
Keywords: distributed computing, big spatial data, remote sensing, remote forest inventory, individual tree
information.
1
1
Introduction
Individual tree information is increasingly becoming the preferred data precision level to accurately and
efficiently monitor, assess, and manage forest and natural resources (Chen et al., 2006; Koch et al., 2006;
Schardt et al., 2002). In the last two decades, airborne light detection and ranging (LiDAR) technology
has brought drastic changes to forest data acquisition and management by providing inventory data at
unprecedented spatial and temporal resolutions (Ackermann, 1999; Maltamo et al., 2014; Shao and
Reynolds, 2006; Swatantran et al., 2016; Wehr and Lohr, 1999). However, to obtain accurate tree level
attributes such as crown width and tree height as well as derivative estimates such as diameter at breast
height (DBH), volume, and biomass, accurate and automated tree segmentation approaches are required
(Schardt et al., 2002).
Numerous methods for tree segmentation within LiDAR data have been proposed (Duncanson et
al., 2014; Hamraz et al., 2016; Hu et al., 2014; Jing et al., 2012; Li et al., 2012; Persson et al., 2002;
Popescu and Wynne, 2004; Véga and Durrieu, 2011; Véga et al., 2014; Wang et al., 2008). Nevertheless,
these methods have only been experimented for small forested areas and none of them have thoroughly
considered scalability; LiDAR data covering an entire forest is much more voluminous than the memory
of a typical workstation and may also take an unacceptably long time to be sequentially processed. Also,
given the continuous advancements of the sensor technology (Swatantran et al., 2016), the LiDAR point
clouds will be acquired with less costs and greater resolutions, which in turn increases the need for more
efficient and scalable processing schemes.
A few studies have considered processing LiDAR data (Thiemann et al., 2013; Zhou and
Neumann, 2009) using streaming algorithms (Pajarola, 2005), where the spatial locality of the LiDAR
data is used to construct out-of-core algorithms. However, streaming algorithms are unable to reduce the
time required for processing because of their inherently sequential processing scheme. A number of
recent studies have considered leveraging the power of multicore and/or GPU (shared memory)
platforms for processing LiDAR data for efficient DEM modeling (Guan and Wu, 2010; Oryspayev et al.,
2
2012; Sten et al., 2016; Wu et al., 2011), or for 3D visualization (Bernardin et al., 2011; Li et al., 2013;
Mateo Lázaro et al., 2014), although shared-memory platforms are also bounded in the amount of
memory and the number of processing units.
On the other hand, processing geospatial data such as LiDAR data can be parallelized by
partitioning the data into tiles (commonly used for data delivery purposes) and distributing the tiles to
different processors on a distributed architecture. Huang et al. (2011) proposed a master-slave distributed
method for parallelizing inverse distance weighting interpolation algorithm. Guan et al. (2013) designed
a cloud -based process virtualization platform to process vast quantities of LiDAR data. Barnes (2016)
parallelized Priority-Flood depression-filling algorithm by subdividing a DEM into tiles. However, the
above distributed approaches were designed and used for perfectly parallel problems while, in case of
non-perfectly parallel problems, dealing with the data near the boundaries of the tiles is not trivial and
should be elaborated according to the specifics of the application (Werder and Krüger, 2009).
Accounting for the data near the tile boundaries, a distributed density-based clustering for spatial
data (Ester et al., 1996) was presented by Xu et al. (2002). The authors proposed a master-slave scheme
in which the master spawns a number of slaves to perform the clustering and return the result back to the
master, who then combines the results. The scheme relies on a data placement strategy for load balancing
in which the master partitions the data and distributes the portions among the slaves for processing, hence
the runtime is determined by the last slave that finishes its job. Distributing the data and merging the
results by the master are also sequential procedures and may yield performance bottlenecks. A more
recent work (He et al., 2011) has presented a version of the density-based clustering tailored to run on a
MapReduce infrastructure (Dean and Ghemawat, 2008) performing four stages of MapReduce for
indexing, clustering, as well as identifying and merging boundary data. The MapReduce infrastructure,
although constraining the programming model, has the advantage of built-in simplicity, scalability, and
fault tolerance. Thiemann et al. (2013) have presented a framework for distributed processing of
geospatial data, where partitioning the data to tiles with overlapping areas near the borders is their core
3
solution. The overlapping area should be at least as big as the required neighborhood for processing a
local entity and the produced overlapping result may require special treatment to be unified. The authors
used the map phase of the Hadoop MapReduce infrastructure (White, 2012) for clustering buildings of
large urban areas and the overlapping result was unified separately afterwards.
Although there are various methods proposed for tree segmentation, only few studies have
considered scalable processing of large geospatial data – there is specifically no study considering forestlevel datasets. This is increasingly important when obtaining tree-level information for areas other than
small-scale plots, which is often the case when obtaining LiDAR data. This paper presents and analyzes a
distributed approach that accounts for the data near the tile boundaries and uses a tree segmentation
algorithm as a building block in order to efficiently segment trees from LiDAR point clouds representing
an entire forest. For experimentation, the approach was implemented using message passing interface
(MPI) (Walker, 1994).
2
Materials and methods
2.1 LiDAR data
We used LiDAR data acquired over the University of Kentucky Robinson Forest (Lat. 37.4611, Long. 83.1555), which covers an aggregated area of 7,441.5 ha in the rugged eastern section of the Cumberland
Plateau region of southeastern Kentucky in Breathitt, Perry, and Knott counties (37°28′23″N
83°08′36″W) (Overstreet, 1984). The LiDAR data is a combination of two datasets collected with the
same LiDAR system (Leica ALS60 at 200 kHz flown with an average speed of 105 knots) by the same
vendor. One dataset was low density (~1.5 pt/m2) collected in the spring of 2013 during leaf-off season
(average altitude of 3,096 m above the ground). The second dataset was high density (~25 pt/m2)
collected in the summer of 2013 during leaf-on season (average altitude of 196 m above the ground). The
combined dataset has a nominal pulse spacing (NPS) of 0.2 m and was delivered in 801 square (304.8 m
4
side ~ 9.3 ha area) tiles (Figure 1), each containing about 5 million LiDAR points on average and
occupying about 400 MB of disk space. The entire LiDAR dataset contains over 4 billion points and
occupies 320 GB of disk space.
0
1
2
4 Kilometers
1:60,000
Figure 1. LiDAR tile map of Robinson Forest consisting of 801 9.3-ha tiles.
2.2 Distributed processing
In a distributed processing environment, the LiDAR data representing tree crowns located across tile
boundaries is split into two or more pieces that are processed by different processing units. Identifying
such crown pieces, unifying them, and efficiently managing the distributed resources to run with a
reasonable speedup are the main challenges of a distributed approach. We propose a master-slave
distributed approach, where the master is in charge of maintaining the global tile map and coordinating
how to process individual tiles and their boundary data while the slaves perform the actual tree
segmentation.
Tile boundary data (solid/striped colored regions in Figure 2) likely represent tree crowns located
between two tiles (light-colored) – hereafter referred to as edge data – or among three or four tiles (dark5
colored) – hereafter referred to as corner data. After segmenting a tile, all segmented crowns that have at
least one LiDAR point within a horizontal distance of 2×NPS from a tile edge form part of the boundary
data. The crowns that are adjacent to only one edge (solid light colored) are regarded as a part of the
associated edge data and those that are adjacent to exactly two edges (solid dark colored) are regarded as
a part of the associated corner data.
Figure 2 A schematic of a tile with the two types of boundary data. The solid-colored tree crown pieces
inside the tile should be unified with the corresponding stripe-colored parts outside.
Figure 3 and 4 show the flowcharts of the master and the slave processes. It is assumed that all
processes can independently input tiles data and output results. Such an assumption can reasonably be
fulfilled by using a supercomputing infrastructure with a unified file system (typically designed to
efficiently support all existent physical processing cores), by maintaining the tiles and the results on a
scalable distributed file system such as the Hadoop file system (White, 2012), or by using a specialized
distributed spatial data organization/retrieval system (Aji et al., 2013; Hongchao and Wang, 2011). The
6
master initializes the work by loading the tile map and assigning each slave to process a unique tile via a
process tile (PT) message carrying the associated tile ID. Upon receiving a PT message, a slave loads and
segments the tile and identifies the boundary data inside the tile consisting of eight disjoint sets (four
edges and four corners). The slave outputs the segmented non-boundary trees, notifies the master via a
tile complete (TC) message carrying the boundary sets, and waits for the master for a new assignment.
The master then updates the tile map and inspects all of the eight boundary sets it received from the slave
to determine if any of the associated edge/corner data is ready to be unified. Edge data is ready when
both tiles sharing the edge are segmented and corner data is ready when all four tiles sharing the corner
are segmented. The master then unifies all edge/corner data that are ready and re-assigns the waiting
slave to re-segment the unified boundary data, which is conveyed by a process boundary (PB) message to
the slave. The slave process, upon receiving the PB message, segments the boundary data conveyed by
the message, outputs the result trees, and notifies the master via a boundary complete (BC) message. The
master then re-assigns a new tile (chosen on an arbitrary order) via a PT message to the slave. If the
master cannot locate any ready boundary data of the tile when it receives the TC message, it proceeds
with re-assigning the waiting slave to segment a new tile via a PT message. If all tiles are segmented, the
master terminates the slave process by sending a finalize (FIN) message. The master process continues
until all slaves are finalized, implying that all tiles and their boundary data were processed.
7
Master Flowchart
START
Load tile_map.
Send PT(tile_ID) to all
slaves.
Listen for message
from any slave.
message
type?
BC
TC(boundary_data
Update tile_map.
Save
boundary_data.
Any
ready_boun
dary_data?
Yes
Send
PB(ready_boundary_
data) to slave.
No
Any
unsegmented
tile?
No
Yes
Send PT(tile_ID)
to slave.
All slaves
finalized?
Send FIN to
slave.
No
Yes
END
Figure 3. Flowchart of the master responsible for maintaining the tile map globally and coordinating the slaves.
8
Slave Flowchart
START
Listen for message from master.
FIN
message
type?
PB(ready_boundary_data)
PT(tile_ID)
End
Load tile with tile_ID.
Segment tile.
Output
non_boundary_trees.
Segment
ready_boundary_data
conveyed by message.
Output boundary_trees.
Send TC(boundary_data)
to master.
Send BC to master.
Figure 4. Flowchart of a slave segmenting tiles and boundary data as directed by the master.
In the presented distributed approach, all tile boundaries are guaranteed to be processed; once all
tiles sharing each specific edge or corner are segmented, the edge/corner data is assigned to be processed
by the slave that completed the last tile. Also, assuming that the amount of processing incurred by the
master does not affect its responsiveness (theoretical limits are derived in the next section), the slaves
keep working all the time resulting in an efficient distributed processing scheme.
9
2.3 Theoretical analysis
We assume that the entire LiDAR data consists of N points, which is arranged in tiles of n points on
average, and LiDAR data representing each tree consist of t points (n ≫ t). We assume that the singleprocessor tile segmentation algorithm has an asymptotic runtime complexity of Ts(n). To illustrate, we
assume that p processors can be allocated for processing N/n tiles (N/n > p).
The number of trees within a tile is proportional to the area of the tile while the number of trees
along a tile edge is proportional to the edge length. Hence, given the average number of trees within a tile
is n/t, the number of trees along one edge of the tile is the square root of it (n1/2/t1/2). Multiplying the
number of trees along the edge by t results in t1/2.n1/2 LiDAR points per edge data. Therefore, the
asymptotic runtime of re-segmenting the boundary data of a tile is Ts(n1/2.t1/2). Also, the communication
of the boundary data between the master and a slave takes O(n1/2.t1/2). Each slave also needs to wait for
the master to receive its boundary data, update its internal tile map, and re-assign the slave. Assuming the
responsiveness of the master, this wait time is also bounded by O(n1/2.t1/2) because the master processes
all of the LiDAR points it communicates with the slave. Aggregating the required time for resegmenting, communicating data, and waiting for the master, the overhead for processing the boundary
data is Ts(n1/2.t1/2) + O(n1/2.t1/2). Therefore, the efficiency of a single slave when segmenting a tile in the
distributed approach presented above is given by:
es
1
T (n)
Ts (n) Ts (n .t ) O(n1/2 .t1/2 )
s
1/2 1/2
where es denotes the efficiency of the slave; the numerator is the effective work; and the denominator is
the total work including the effective work and the overhead.
Because the master does not perform segmentation, the entire segmentation that is performed by
all of the p-1 slaves is sped up by a factor of (p-1)es. Between the time when the first and the last slaves
10
are finalized, the remaining workload of each active slave is bounded by n LiDAR points because each of
them has at most one tile to complete. As soon as the first slave is finalized, a non-parallelizable workload
is introduced to the distributed scheme. Between the time the first and the second slaves are finalized, the
active slaves process with a missing fraction of the entire slaves’ power, i.e., 1/(p-1) of the power was
already finalized. This results in n/(p-1) non-parallelizable workload. Similarly, between the time the (i1)th and ith slaves are finalized, (i-1)n/(p-1) non-parallelizable workload is introduced. Therefore, the total
non-parallelizable workload is:
p 1
ws
i 2
i 1
p2
n
n
p 1
2
2
where ws denotes the non-parallelizable (serial) workload of the entire distributed processing (the
initialization workload performed by the master is a negligible constant. Hence, the ratio (P) of the
parallelizable (total minus serial) workload to the total workload is:
P
N
p2
n
2
N
3
Finally, the speedup of the entire distributed approach denoted by Sp according to Gustafson-Barsis law
(McCool et al., 2012) is:
S p 1 P P( p 1)es
4
The time the master requires to devote per tile is proportional to the number of LiDAR points it
deals with, which is O(n1/2.t1/2), while the time a slave requires to devote per tile is Ts(n) + Ts(n1/2.t1/2) +
11
O(n1/2.t1/2). Thus, in order for the master to remain responsive for p-1 slaves so that the above equations
hold, we should have:
p 1
3
Ts (n) Ts (n1/2 .t1/2 ) O(n1/2 .t1/2 )
O(n1/2 .t1/2 )
5
Results and discussions
3.1 Runtime and scalability
We adopted the tree segmentation algorithm presented by Hamraz et al. (2016) as the single-processor
building block to empirically assess the proposed distributed processing approach. The tree segmentation
algorithm can efficiently be implemented such that Ts(n) = O(n) (see Appendix A). We implemented the
master-slave scheme using the MPI and ran it on the University of Kentucky Lipscomb cluster, which has
256 symmetric basic nodes (Dell C6220 Server, 4 nodes per 2U chassis), each with 16 cores (dual Intel
E5-2670 8 Core – Sandy Bridge) at 2.6 GHz and 64 GB of RAM at 1,600 MHz. The nodes are interconnected via Mellanox Fourteen Data Rate InfiniBand (2:1 over-subscription, 14.0625 Gbit/s) and
equipped with a global file system (DDN GridScaler SFA12K storage appliance with the IBM GPFS –
Read: 25 GB/s throughput and 780,000 IO/S, Write: 22 GB/s throughput and 690,000 IO/S) (University
of Kentucky Analytics & Technologies). We experimented with four contiguous loads of data: the first
200 (Figure 1 – counting row-wise starting from the top leftmost tile toward right and then down), 400,
and 600 tiles, as well as all 801 tiles. For each load, we ran the distributed segmentation approach using
1–12 computing nodes (i.e., 16, 32, …, 192 processing cores), and measured the experimental speedups
by dividing the observed single-processor runtime by the observed distributed processing runtimes. The
observed single-processor runtime equals the number of tiles multiplied by average observed runtime of a
12
tile, which equaled 31 minutes and 8 seconds (2.8% loading from disk, 94.8% computation, and 2.4%
writing to disk) averaged for a sample of 128 tiles.
Figure 5 shows the experimental speedups overlaying the equivalent theoretical speedups using
Equation 4 for which t = 1,350 and n = 5×106 as measured in the dataset. In order to calculate the exact
value of es using Equation 1, the constant coefficients of the asymptotic functions in the numerator and
the denominator need to be measured on the specific runtime platform. According to our measurement,
the ratio of the constant coefficient of the numerator (Ts(n) – equals to O(n) here) to the constant
coefficient of O(n1/2.t1/2) appeared in the denominator is about 150. In other words, the time required for
the segmentation of a LiDAR point cloud is approximately 150 times greater than the time required for
two-way inter-process communication (from a slave to the master and back) of the same size point cloud
on our runtime platform. Substituting the values of t, n, and the ratio of the constant coefficients in
Equation 1, the efficiency of a slave (es) equals 0.9837. Similarly, using Equation 5, having p-1 ≤ 9,279
renders the master to remain responsive.
Figure 5 Experimental speedups shown by symbols, which overlay corresponding continuously drawn theoretical
speedups for different loads of data.
13
As shown in Figure 5, processing the entire tiles using 192 processing cores resulted in a practical
speedup of 167.04 (compared to 165.70 of theory), meaning that we reduced the expected singleprocessor runtime of over 17 days to 2 hours and 29 minutes. Although a few weeks of processing time
might be acceptable for forest inventory to be performed annually, it is infeasible for potential real-time
applications, e.g., more accurate aerial monitoring of wildfire using LiDAR (Arroyo et al., 2008;
Contreras, 2010). After all, natural forests may be several times greater than Robinson Forest and be
recorded with greater point densities (to become affordable given the advancements of the sensor
technology) yielding much larger datasets, which even more justifies the need for distributed processing.
The small differences between the empirical and the theoretical speedups (Figure 5) are likely due to
natural variabilities in the dataset as well as small differences in the runtime environment from the
theoretical assumptions. These results show that the distributed segmentation approach can achieve
nearly linear speedup using a reasonable number of processing cores and given a sufficiently large dataset
(at least two times more tiles than the number of cores). Because the number of tiles is typically large for
forest-level data and the number of cores is limited, scalability of the approach to arbitrarily large datasets
is fulfilled.
As the distributed approach does not assume a fixed number of slave processes, it can also be
implemented on a grid environment in which the master can be in charge of initiating new slave processes
and rescheduling tasks in case of node failure. In case Equation 5 is violated (the master is overloaded),
the straightforward solution is to increase the size of tiles to make the slaves perform proportionately
more work per each tile assignment. A more flexible solution is to augment the distributed scheme to
accommodate multiple masters in a hierarchical fashion. An additional improvement might consider
slaves not sending boundary data to the master. Instead, they can set aside the data in a buffer and send it
later on directly to the slave who would eventually process the boundary data. In this case, the master
14
should be in charge of coordinating the interactions between the slaves and would not need to deal with
receiving and sending boundary data, which decreases the master’s workload and make it independent of
the tile size. Such an improvement would not affect the asymptotic calculations of speedup presented
above, even though it may help to reduce the runtime in practice specially if the master is overloaded
and/or the inter-process communication on the runtime platform is costly. Lastly, the master can employ
any strategy for choosing a new tile to assign next without affecting the final result and the processing
time in theory, although assigning contiguous tiles makes boundary data become ready earlier and results
in freeing up memory earlier, which may become invaluable depending on the circumstances.
Tailoring the proposed approach to run under the Hadoop MapReduce infrastructure in a single
stage can also be accomplished as follows. Loading and segmentation of an individual tile should be
defined as the map phase, in which the non-boundary trees should be output to the file system and each of
the eight boundary data are assigned a unique key for the reduce phase. The unique key of each specific
edge/corner data should be the same across all the map tasks that share the specific edge/corner. The
reduce phase should be defined to unify all of the data it is given (edge/corner data portions having an
identical unique key), re-segment the data, and output the result to the file system. There would not be an
explicitly defined master process because the underlying map-reduce infrastructure is responsible for
coordination between the map and the reduce tasks, as well as scalability and fault tolerance of the entire
ecosystem. In contrast, the MPI implementation using a global scalable file system generally runs faster
because slaves barely idle, while reduce phase cannot start processing until map phase finishes. This
performance advantage is achieved because of having explicit control over the inter-process
communications enabling design of a flexible scheduling scheme using MPI, although it generally
requires more effort and expertise to design and program desired features for a distributed application.
15
3.2 Global forest parameters
Although tile size does not affect the segmentation result of the distributed approach in theory, depending
on the underlying single-processor segmentation algorithm, it may introduce slight biases in practice.
Such biases have a direct correlation with the total length of the shared edges of the tiles because the
boundary data along those edges are indeed the only places that are not processed exactly the same
compared to a single-processor run. In order to quantify the biases in terms of number of trees, we
processed five sample square (1.524 Km side ~ 232.5 ha area) blocks (each composed of 5×5 tiles) in a
single-processor manner as well as using the distributed approach. We partitioned each block to uniform
grids of 2×2, 3×3, …, 15×15 sub-blocks and ran the distributed approach for each of the grid patterns.
Single-processor execution detected an average of 62,005 trees in a block. Figure 6 shows the average
number of trees detected per block as a function of the total length of the shared edges of sub-blocks,
which equals 2 × (nsb-1) multiplied by the block side length where nsb denotes number of sub-blocks
along a block side. As expected, additional number of trees compared with single-processor run shows a
linear relation with the total shared edge length: an average of 96 additional trees (false positives) were
detected per 1 Km of shared edge, which is a small value given that more than 26,000 trees were detected
per 1 Km2.
Figure 6. Number of trees detected in a block for different partitioning patterns.
16
When applied to the entire Robinson Forest, the distributed tree segmentation approach detected a grand
total of 1,994,970 trees over the area covered by the LiDAR data. The total length of shared edges in the
tile map (Figure 1) is 446.23 Km, which results in 42,833 potential false positives (2.15%) be introduced
across the tile edges. When the number of false positives is subtracted, the grand total of detected trees
becomes 1,952,137.
Due to imperfectness of the single-processor algorithm, a portion of grand total number of detected trees
was associated with over-segmentations, and a portion of existing trees in the forest was undetected. In
order to account for the over-segmentations/undetected trees, we used the accuracy result of the singleprocessor segmentation algorithm on the LiDAR point clouds (taken from the same dataset) of a fieldsurveyed sample of 23 (0.1 ac) circular plots placed across Robinson Forest (Hamraz et al., 2016). The
accuracy result included the number of detected trees (bearing over-segmentations) and the number of
existing trees (bearing undetected trees) per four crown classes (dominant, co-dominant, intermediate, and
overtopped). Within each of the 23 plots, we calculated a fraction per crown class: the existing trees of
that crown class divided by the grand total (all crown classes) of detected trees. Table 1 shows the mean
and 95% T-confidence bounds of the fractions across the 23 plots. It also shows the adjusted estimates of
number of existing trees, which were calculated by multiplying the grand total number of detected trees
using the distributed approach to the corresponding fractions. Considering a 95% T-confidence interval,
the total number of existing trees in the 7,441.5-ha forested area is estimated to be 2,495,170 (±13.52%),
which results in an average of 335.30 trees per ha.
17
Table 1 Estimated number of trees categorized based on tree crown class.
Crown Class
Dominant
Co-dominant
Intermediate
Overtopped
Dead
All
1
Fraction of existing to grand total detected
mean
95TCB1
0.0785
±75.50%
0.3069
±23.07%
0.5376
±17.84%
0.2928
±43.29%
0.0625
±104.7%
1.2782
±13.52%
Estimated number of existing trees
entire forest
per ha
153,178
20.59
599,106
80.50
1,049,446
141.32
571,522
76.80
121,917
16.38
2,495,170
335.30
95% T-Confidence Bounds (DF=22)
For verification, we compared our tree number estimates (Table 1) with equivalent estimates
based on field measurements of another sample of 23 plots from the Robinson Forest (Figure 7). The
estimates for total number of trees differ by about 3% and the estimates of number of dominant trees
differ by about 30%. However, the large overlap between the 95% T-confidence interval errors indicate
that there is no statistically significant difference between the estimates using LiDAR and the field
collected measurements.
Figure 7. Estimated number of trees using LiDAR compared to field-collected along with the 95% T-confidence
intervals.
18
Figure 8 shows the height distribution of all detected trees (heights above 5 m) by the approach.
The height distribution follows a bimodal pattern, which can be attributed to multistory structure of
deciduous natural forests, in which the dominant and co-dominant trees form the over-story and
intermediate and overtopped trees form the mid-story. We fitted a normal mixture model to the bimodal
distribution: the larger lump on the right (associated with over-story trees) has a mean height of 26.9 m
and a standard deviation of 6.6 m, and the smaller lump (associated with mid-story trees) has a mean
height of 9.4 m and a standard deviation of 2.6 m.
Figure 8 Height distribution of 1,994,970 trees detected in Robinson Forest superimposed with estimated normal
mixture model.
We compared the LiDAR-derived mean tree height estimates with those obtained using a field
sample from the forest (371 over-story and 826 mid-story trees). The sample mean height of the overstory trees was 25.4 m with a standard deviation of 5.3 m, and the sample mean height of mid-story trees
19
was 17.0 m with a standard deviation of 4.1 m. Considering that the LiDAR-detected tree heights are in
fact biased by presence of falsely detected trees and absence of undetected trees, the field estimates are
close to the LiDAR-detected estimates for over-story trees. However, the field estimates for mid-story
trees are remarkably larger than the LiDAR-detected estimates, which can be justified as follows.
Airborne LiDAR provides considerably less information about the mid-story trees due to decreased
penetration of LiDAR points toward bottom canopy layers (Maguya et al., 2014; Reutebuch et al., 2003),
hence detected tree rate is lower for mid-story trees (Duncanson et al., 2014; Hamraz et al., 2016). Also,
the detected mid-story trees are likely biased to be smaller within the population of all existing mid-story
trees because they are easier to detect when there is less canopy closure, which is associated with stand
age and is minimal when stand is young and in general has smaller trees (Jules et al., 2008). So, detecting
relatively fewer mid-story trees that are also likely biased to be smaller leads to capturing a distribution
with smaller mean and standard deviation. After all, the only information used to fit the normal mixture
model was the heights of the trees while height may not be sufficient for classification, i.e., a moderately
tall tree can in fact be mid-story if situated in a taller stand while the mixture model always probabilizes it
strongly as over-story according to its height, and vice versa. Thus, the procedure of fitting the normal
mixture model likely separates the two tree classes more distantly with respect to height.
In addition to providing number of trees and height distributions (compared here to field surveys
for validation), the distributed approach enables identification of individual tree locations and attributes
(tree height and crown widths) as well as the point cloud segments representing tree crowns for large
forested areas in a timely manner, which in turn enables building a detailed (at the individual tree level)
forest model and performing a myriad of more accurate analyses. For instance, tree attributes can be used
to develop allometric equations to estimate other important tree metrics such as DBH and volume, and the
point cloud segments can be used to construct the 3D geometric shape of each individual tree crown to
develop mode detail estimates such as crown volume, biomass, and carbon content.
20
3.3 Approach application to other spatial datasets
As mentioned earlier, the approach uses a single-processor tree segmentation algorithm as a building
block and does not require any knowledge on how the algorithm functions. So, the approach may be used
to straightforwardly adopt any other single-processor object identification/segmentation algorithm in
order to scale up processing arbitrarily big spatial and geospatial datasets, such as remotely sensed
buildings, cars, planets, etc. The only caveat is that the objects may not be greater than the tiles, i.e., they
may not touch more than two adjacent edges of a tile.
Moreover, generalization of the approach to process 3D spatial data can be accomplished similarly as
follows. Instead of tiles that are representing surfaces, cubes representing volumes will be the data units
for 3D data. Boundary data in this case would be surface (shared between two cubes), edge (shared
among four cubes, and corner (shared among eight cubes) that can be handled for distributed processing
using the master-slave processing scheme presented in Section 2.2. The theoretical runtime analysis for
3D data would be slightly different. The average number of the entire objects within the cube is
proportional to the cube volume while the number of boundary objects (those touching a cube surface) is
proportional to the cube surface area. Hence, the number of boundary objects equals the number of
objects within the cube raised to 2/3 power, which changes the master/slave overheads and Equations 1
and 5 need to be updated accordingly.
4
Conclusions
Obtaining tree-level information over large forested areas is increasingly important for accurate
assessment, monitoring and management of forests and natural resources. Several automated tree
segmentation methods have been developed, but these methods have only been applied to small forested
areas for accuracy assessment. Although these methods can in theory be applied to larger areas, such
applications is not straightforward because LiDAR data covering forest-level data far exceeds the
memory of desktop computers and may also take unacceptably long time to be processed sequentially.
21
Here we presented and analyzed a scalable distributed approach that was applied to segment trees within a
LiDAR point cloud covering an entire forest. The distributed approach segmented trees within the tiles
and across the tile boundaries, and introduced a minimal bias compared with the single-processor
algorithm that was also quantified in this work. Comparison of the estimated number of trees and the tree
height distribution with the field surveys validated sound operation of the approach. We presented the
distributed processing approach and the associated analysis in a platform-independent manner so as the
implementation can be accomplished on different distributed platforms with minor modifications.
The presented approach enables obtaining individual tree locations and point cloud segments
representing the tree crowns for entire forested areas in a timely manner. The resulting detailed, tree-level
information has the potential to increase the accuracy of forest level information by creating remotely
sensed forest inventories for more efficient management of forest and natural resources. Although the
distributed approach was presented within the context of tree segmentation from LiDAR point clouds, it
can straightforwardly be applied to segment/identify objects within other large-scale datasets.
Acknowledgments
This work was supported by: 1) the University of Kentucky Forestry Department and the
McIntire-Stennis project KY009026 Accession 1001477, ii) the Kentucky Science and
Engineering Foundation under the grant KSEF-3405-RDE-018, and iii) the University of
Kentucky Center for Computational Sciences.
References
Ackermann, F., 1999. Airborne laser scanning—present status and future expectations. ISPRS Journal of
Photogrammetry and Remote Sensing 54, 64-67.
Aji, A., Wang, F., Vo, H., Lee, R., Liu, Q., Zhang, X., Saltz, J., 2013. Hadoop GIS: a high performance
spatial data warehousing system over mapreduce. Proceedings of the VLDB Endowment 6, 1009-1020.
22
Arroyo, L.A., Pascual, C., Manzanera, J.A., 2008. Fire models and methods to map fuel types: the role of
remote sensing. Forest Ecology and Management 256, 1239-1252.
Barnes, R., 2016. Parallel Priority-Flood depression filling for trillion cell digital elevation models on
desktops or clusters. Computers & Geosciences 96, 56-68.
Bernardin, T., Cowgill, E., Kreylos, O., Bowles, C., Gold, P., Hamann, B., Kellogg, L., 2011. Crusta: A
new virtual globe for real-time visualization of sub-meter digital topography at planetary scales.
Computers & Geosciences 37, 75-85.
Chen, Q., Baldocchi, D., Gong, P., Kelly, M., 2006. Isolating individual trees in a savanna woodland
using small-footprint LiDAR data. Photogrammetric Engineering and Remote Sensing 72, 923-932.
Contreras, M.A., 2010. Spatio-temporal optimization of tree removal to efficiently minimize crown fire
potential.
Dean, J., Ghemawat, S., 2008. MapReduce: simplified data processing on large clusters. Communications
of the ACM 51, 107-113.
Duncanson, L., Cook, B., Hurtt, G., Dubayah, R., 2014. An efficient, multi-layered crown delineation
algorithm for mapping individual tree structure across multiple ecosystems. Remote Sensing of
Environment 154, 378-386.
Ester, M., Kriegel, H.-P., Sander, J., Xu, X., 1996. A density-based algorithm for discovering clusters in
large spatial databases with noise, Kdd, pp. 226-231.
Guan, H., Li, J., Zhong, L., Yongtao, Y., Chapman, M., 2013. Process virtualization of large-scale lidar
data in a cloud computing environment. Computers & Geosciences 60, 109-116.
Guan, X., Wu, H., 2010. Leveraging the power of multi-core platforms for large-scale geospatial data
processing: Exemplified by generating DEM from massive LiDAR point clouds. Computers &
Geosciences 36, 1276-1282.
Hamraz, H., Contreras, M.A., Zhang, J., 2016. A robust approach for tree segmentation in deciduous
forests using small-footprint airborne LiDAR data. International Journal of Applied Earth Observation
and Geoinformation 52, 532-541.
23
He, Y., Tan, H., Luo, W., Mao, H., Ma, D., Feng, S., Fan, J., 2011. Mr-dbscan: an efficient parallel
density-based clustering algorithm using mapreduce, Parallel and Distributed Systems (ICPADS), 2011
IEEE 17th International Conference on. IEEE, pp. 473-480.
Hongchao, M., Wang, Z., 2011. Distributed data organization and parallel data retrieval methods for huge
laser scanner point clouds. Computers & Geosciences 37, 193-201.
Hu, B., Li, J., Jing, L., Judah, A., 2014. Improving the efficiency and accuracy of individual tree crown
delineation from high-density LiDAR data. International Journal of Applied Earth Observation and
Geoinformation 26, 145-155.
Huang, F., Liu, D., Tan, X., Wang, J., Chen, Y., He, B., 2011. Explorations of the implementation of a
parallel IDW interpolation algorithm in a Linux cluster-based parallel GIS. Computers & Geosciences 37,
426-434.
Jing, L., Hu, B., Li, J., Noland, T., 2012. Automated delineation of individual tree crowns from LiDAR
data by multi-scale analysis and segmentation. Photogrammetric Engineering and Remote Sensing 78,
1275-1284.
Jules, M.J., Sawyer, J.O., Jules, E.S., 2008. Assessing the relationships between stand development and
understory vegetation using a 420-year chronosequence. Forest Ecology and Management 255, 23842393.
Koch, B., Heyder, U., Weinacker, H., 2006. Detection of individual tree crowns in airborne LiDAR data.
Photogrammetric Engineering & Remote Sensing 72, 357-363.
Li, J., Jiang, Y., Yang, C., Huang, Q., Rice, M., 2013. Visualizing 3D/4D environmental data using
many-core graphics processing units (GPUs) and multi-core central processing units (CPUs). Computers
& Geosciences 59, 78-89.
Li, W., Guo, Q., Jakubowski, M.K., Kelly, M., 2012. A new method for segmenting individual trees from
the LiDAR point cloud. Photogrammetric Engineering & Remote Sensing 78, 75-84.
Maguya, A.S., Junttila, V., Kauranne, T., 2014. Algorithm for extracting digital terrain models under
forest canopy from airborne LiDAR data. Remote Sensing 6, 6524-6548.
Maltamo, M., Næsset, E., Vauhkonen, J., 2014. Forestry applications of airborne laser scanning. Concepts
and case studies. Manag For Ecosys 27, 2014.
24
Mateo Lázaro, J., Sánchez Navarro, J.Á., García Gil, A., Edo Romero, V., 2014. 3D-geological structures
with digital elevation models using GPU programming. Computers & Geosciences 70, 138-146.
McCool, M.D., Robison, A.D., Reinders, J., 2012. Structured Parallel Programming: patterns for efficient
computation. Elsevier.
Oryspayev, D., Sugumaran, R., DeGroote, J., Gray, P., 2012. LiDAR data reduction using vertex
decimation and processing with GPGPU and multicore CPU technology. Computers & Geosciences 43,
118-125.
Overstreet, J., 1984. Robinson Forest inventory. Department of Forestry, University of Kentucky,
Lexington, Kentucky.
Pajarola, R., 2005. Stream-processing points, Visualization, 2005. VIS 05. IEEE. IEEE, pp. 239-246.
Persson, A., Holmgren, J., Söderman, U., 2002. Detecting and measuring individual trees using an
airborne laser scanner. Photogrammetric Engineering and Remote Sensing 68, 925-932.
Popescu, S.C., Wynne, R.H., 2004. Seeing the trees in the forest. Photogrammetric Engineering &
Remote Sensing 70, 589-604.
Reutebuch, S.E., McGaughey, R.J., Andersen, H.-E., Carson, W.W., 2003. Accuracy of a high-resolution
LiDAR terrain model under a conifer forest canopy. Canadian Journal of Remote Sensing 29, 527-535.
Schardt, M., Ziegler, M., Wimmer, A., Wack, R., Hyyppa, J., 2002. Assessment of forest parameters by
means of laser scanning. International archives of photogrammetry remote sensing and spatial
information sciences 34, 302-309.
Shao, G., Reynolds, K.M., 2006. Computer applications in sustainable forest management: Including
perspectives on collaboration and integration. Springer Science & Business Media.
Sten, J., Lilja, H., Hyväluoma, J., Westerholm, J., Aspnäs, M., 2016. Parallel flow accumulation
algorithms for graphical processing units with application to RUSLE model. Computers & Geosciences
89, 88-95.
Swatantran, A., Tang, H., Barrett, T., DeCola, P., Dubayah, R., 2016. Rapid, High-Resolution Forest
Structure and Terrain Mapping over Large Areas using Single Photon Lidar. Scientific reports 6.
25
Thiemann, F., Werder, S., Globig, T., Sester, M., 2013. Investigations into partitioning of generalization
processes in a distributed processing framework, 26th International Cartographic Conference, Dresden,
Germany.
University of Kentucky Analytics & Technologies, High performance computing hardware. University of
Kentucky.
Véga, C., Durrieu, S., 2011. Multi-level filtering segmentation to measure individual tree parameters
based on Lidar data: Application to a mountainous forest with heterogeneous stands. International Journal
of Applied Earth Observation and Geoinformation 13, 646-656.
Véga, C., Hamrouni, A., El Mokhtari, S., Morel, J., Bock, J., Renaud, J.-P., Bouvier, M., Durrieu, S.,
2014. PTrees: A point-based approach to forest tree extraction from lidar data. International Journal of
Applied Earth Observation and Geoinformation 33, 98-108.
Walker, D.W., 1994. The design of a standard message passing interface for distributed memory
concurrent computers. Parallel Computing 20, 657-673.
Wang, Y., Weinacker, H., Koch, B., 2008. A lidar point cloud based procedure for vertical canopy
structure analysis and 3D single tree modelling in forest. Sensors 8, 3938-3951.
Wehr, A., Lohr, U., 1999. Airborne laser scanning—an introduction and overview. ISPRS Journal of
Photogrammetry and Remote Sensing 54, 68-82.
Werder, S., Krüger, A., 2009. Parallelizing geospatial tasks in grid computing. GIS Science 3, 71-76.
White, T., 2012. Hadoop: The definitive guide. " O'Reilly Media, Inc.".
Wu, H., Guan, X., Gong, J., 2011. ParaStream: a parallel streaming Delaunay triangulation algorithm for
LiDAR points on multicore architectures. Computers & Geosciences 37, 1355-1363.
Xu, X., Jäger, J., Kriegel, H.-P., 2002. A fast parallel clustering algorithm for large spatial databases,
High Performance Data Mining. Springer, pp. 263-290.
Zhou, Q.-Y., Neumann, U., 2009. A streaming framework for seamless building reconstruction from
large-scale aerial lidar data, Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE
Conference on. IEEE, pp. 2759-2766.
26
Appendix A: Efficient implementation of segmentation algorithm and runtime
analysis
The tree segmentation algorithm we used as a building block in this study (Hamraz et al., 2016) consists
of a pre-processing step including homogenizing the point cloud, removing non-surface points,
smoothing, and then a loop over the five major steps outlined below until the entire point cloud is
clustered: 1) locate the non-clustered highest point - global maximum (GMX); 2) generate vertical
profiles originating from the GMX with a length of maximum tree crown radius; 3) For each profile,
identify the LiDAR point along the profile that represents the crown boundary; 4) create a convex hull of
the identified boundary points; and 5) cluster all LiDAR points encompassed within the convex hull as
the highest tree crown.
For an efficient implementation, the point cloud should be indexed in a 2D horizontal grid.
Indexing and the pre-processing step takes O(n) where n is the number of points. We assume that the
main loop iterates m times. Naively locating the GMX (step 1) takes O(n) per iteration. Instead, we create
a descendingly sorted list of all of the grid cells according to the height of the point they contain and mark
all cells as unvisited. The sorting procedure takes O(n.㏒n). The grid cells are marked as visited when
they are clustered in step 5. To locate the non-clustered GMX, the sorted list is traversed from the position
of the previous GMX forward, which on average takes O(n/m) per iteration. Once the GMX is located,
clustering the highest tree (steps 2–5) has a runtime independent of n and m and is proportional to the tree
size, which is bounded and can be assumed as a constant. So, the aggregate runtime of each iteration of
the loop is O(m/n), hence the total runtime of the loop becomes O(n). Aggregating the pre-processing and
the sorting times:
Ts (n) O(n) O(n.log n)
A.1
27
where Ts(n) is the total runtime of the algorithm; the first term on the right-hand side corresponds to the
runtime of the main loop and the pre-processing step; and the second term corresponds to the runtime of
the sorting procedure before the loop.
We ran the implementation above on a workstation of 3.4 GHz CPU speed and 8 GB of RAM for
25 loads of data. Figure A.1 shows the log-log plot of the runtime of the segmentation versus the number
of points.
Figure A.1. Log-log plot of the Segmentation runtime versus the number of LiDAR points in the point
cloud. Each symbol corresponds to average across 15 strata.
The slope of the best fit line to the square symbols is 1.03, which concurs with the linear term of Equation
A.1. Also, the triangle symbols show a slightly super-linear pattern concurring with the non-linear term.
We measured the constant coefficients of both terms by dividing the execution times associated with the
terms by n and n.㏒n respectively. The ratio of the linear coefficient to the non-linear one is platform-
independent and is about 7,800 according to our measurement. This yields that n should be greater than
27,800 in order for the non-linear term to start dominating the linear term, which corresponds to a LiDAR
28
point cloud covering over 3e+2,331 times surface area of the earth. So, we may safely replace ㏒n in the
non-linear term with an upper bound constant, which reduces Equation A.1 to:
Ts (n) O(n)
A.2
29
| 5cs.CE
|
ABSTRACTIONS OF VARYING DECENTRALIZATION DEGREE FOR
COUPLED MULTI-AGENT SYSTEMS
arXiv:1603.04780v2 [cs.SY] 8 Oct 2016
D. BOSKOS AND D. V. DIMAROGONAS
Abstract. In this report, we aim at the development of a decentralized abstraction framework for multi-agent systems under coupled constraints, with the possibility for a varying
degree of decentralization. The methodology is based on the analysis employed in our recent
work, where decentralized abstractions based exclusively on the information of each agent’s
neighbors were derived. In the first part of this report, we define the notion each agent’s
m-neighbor set, which constitutes a measure for the employed degree of decentralization.
Then, sufficient conditions are provided on the space and time discretization that provides
the abstract system’s model, which guarantee the extraction of a transition system with
quantifiable transition possibilities.
1. Introduction
The analysis and control of multi-agent systems constitutes an active area of research with
numerous applications, ranging from the analysis of power networks to the automatic deployment of robotic teams [8]. Of central interest in this field is the problem of high level task
planning by exploiting tools from formal verification [18]. In order to follow this approach for
dynamic multi-agent systems it is required to provide an abstract model of the system which
can serve as a tool for analysis and control, as well as high level planning. In particular, the
use of a suitable discrete representation of the system allows the automatic synthesis of discrete
plans that guarantee satisfaction of the high level specifications. Then, under appropriate relations between the continuous system and its discrete analogue, these plans can be converted to
low level primitives such as sequences of feedback controllers, and hence, enable the continuous
system to implement the corresponding tasks.
The need for a formal approach to the aforementioned control synthesis problem has lead to
a considerable research effort for the extraction of discrete state symbolic models, also called
abstractions, which capture properties of interest of continuous state dynamical and control
systems, while ignoring detail. Results in this direction for the nonlinear single plant case have
been obtained in the papers [22] and [28], which are based on partitioning and exploit approximate simulation and bisimulation relations. Symbolic models for piecewise affine systems on
simplices and rectangles were introduced in [14] and have been further studied in [6]. Closer
related to the control framework that we adopt here for the abstraction, are the papers [15],
[16] which build on the notion of In-Block Controllability [7]. Other abstraction techniques for
nonlinear systems include [25], where discrete time systems are studied in a behavioral framework and [1], where box abstractions are studied for polynomial and other classes of systems.
It is also noted that certain of the aforementioned approaches have been extended to switched
systems [11], [12]. Furthermore, abstractions for interconnected systems have been recently
developed in [27], [24], [23], [26], [21], [9] and rely mainly on compositional approaches based
Key words and phrases. hybrid systems, multi-agent systems, abstractions, transition systems.
1
on small gain arguments. Finally, in [20], a compositional approach with a varying selection of
subsystems for the abstraction is exploited, providing a tunable tradeoff between complexity
reduction and model accuracy.
In this framework, we focus on multi-agent systems and assume that the agents’ dynamics
consist of feedback interconnection terms and additional bounded input terms, which we call
free inputs and provide the ability for motion planning under the coupled constraints. We
generalize the corresponding results of our recent work [4], where each agent’s abstract model
has been based on the knowledge of the discrete positions of its neighbors, by allowing the
agent to have this information for all members of the network up to a certain distance in the
communication graph. The latter provides an improved estimate of the potential evolution of its
neighbors and allows for more accurate discrete agent models, due to the reduction of the part of
the available control which is required for the manipulation of the coupling terms. In addition,
the derived abstractions are coarser than the ones in [4] and can reduce the computational
complexity of high level task verification. Finally, we note that this report includes the proofs
of its companion submitted conference version [3], which have been completely omitted therein
due to space constraints, as well as certain generalizations and additional material.
The rest of the report is organized as follows. Basic notation and preliminaries are introduced
in Section 2. In Section 3, we define well posed abstractions for single integrator multi-agent
systems and prove that the latter provide solutions consistent with the design requirement on
the systems’ free inputs. Section 4 is devoted to the study of the control laws that realize
the transitions of the proposed discrete system’s model. In Section 5 we quantify space and
time discretizations which guarantee well posed transitions with motion planning capabilities.
The framework is illustrated through an example with simulation results in Section 6 and we
conclude in Section 7.
2. Preliminaries and Notation
We use the notation |x| for the Euclidean norm of a vector x ∈ Rn and int(S) for the interior
of a set S ⊂ Rn . Given R > 0 and x ∈ Rn , we denote by B(R) the closed ball with center 0 ∈ Rn
and radius R, namely B(R) := {x ∈ Rn : |x| ≤ R} and B(x; R) := {y ∈ Rn : |x − y| ≤ R}.
Consider a multi-agent system with N agents. For each agent i ∈ N := {1, . . . , N } we use
the notation Ni ⊂ N \{i} for the set of its neighbors and Ni for its cardinality. We also consider
an ordering of the agent’s neighbors which is denoted by j1 , . . . , jNi and define the Ni -tuple
j(i) = (j1 (i), . . . , jNi (i)). Whenever it is clear from the context, the argument i will be omitted
from the latter notation. The agents’ network is represented by a directed graph G := (N , E),
with vertex set N the agents’ index set and edge set E the ordered pairs (`, i) with i, ` ∈ N
and ` ∈ Ni . The sequence i0 i1 · · · im with (iκ−1 , iκ ) ∈ E, κ = 1, . . . , m, namely, consisting of m
consecutive edges in G, forms a path of length m in G. We will use the notation Nim to denote
for each m ≥ 1 the set of agents from which i is reachable through a path of length m and
not by a shorter one, excluding also the possibility to reach itself throughSa cycle. Notice that
m
Ni1 = Ni . We also define Ni0 := {i} and for each m ≥ 1 the set N̄im := `=0 Ni` , namely, the
set of all agents from which i is reachable by a path of length at most m, including i. With
some abuse of language, we will use the terminology m-neighbor set of agent i for the set N̄im ,
since it always contains the agent itself and will also refer to the rest of the agents in N̄im as
the m-neighbors of i. Finally, we denote by N̄im and Nim the cardinality of the sets N̄im and
Nim , respectively.
2
Definition 2.1. For each agent i ∈ N , we consider a strict total order ≺ on each set N̄im that
0
00
satisfies i0 ≺ i00 for each i0 ∈ Nim , i00 ∈ Nim , with 0 ≤ m0 < m00 ≤ m.
Example 2.2. In this example, we consider a network of 8 agents as depicted in Figure 1
and illustrate the sets Nim , N̄im for agents 1 and 5 up to paths of length m = 3 as well as a
candidate order for the corresponding sets N̄im , in accordance to Definition 2.1.
N̄11 = {1, 2, 6},
N11 = {2, 6},
N̄12
N̄13
N̄51
N̄52
N̄53
N12
N13
N51
N52
N53
= {1, 2, 3, 5, 6, 7},
= {1, 2, 3, 4, 5, 6, 7, 8},
= {3, 5},
= {2, 3, 4, 5},
= {2, 3, 4, 5},
1 ≺ 6 ≺ 2,
= {3, 5, 7}, 1 ≺ 6 ≺ 2 ≺ 5 ≺ 3 ≺ 7,
= {4, 8},
1 ≺ 6 ≺ 2 ≺ 5 ≺ 3 ≺ 7 ≺ 4 ≺ 8,
= {3},
5 ≺ 3,
= {2, 4},
5 ≺ 3 ≺ 2 ≺ 4,
= ∅,
5 ≺ 3 ≺ 2 ≺ 4.
Ag.7
Ag.8
Ag.6
Ag.5
Ag.1
Ag.4
Ag.3
Ag.2
Figure 1. Illustration of the network topology for Example 2.2.
Given an index set I, an agent i ∈ N , its m-neighbor set N̄im , and a strict total ordering
of N̄im as in Definition 2.1, it follows that the elements of N̄im are ordered as i ≺ j11 (i) ≺
1
2
m
m
m0
m0
· · · ≺ jN
(i) ≺ j12 (i) ≺ · · · ≺ jN
for all
2 (i) · · · ≺ j1 (i) · · · ≺ jN m (i) with jκ (i) ∈ Ni
i
i
i
0
m0 ∈ {1, . . . , m} and κ ∈ {1, . . . , Nim }. Whenever it is clear from the context we will remove
the argument i from the latter ordered elements. In addition, we define the mapping pri :
m
I N → I N̄i which assigns to each N -tuple (l1 , . . . , lN ) ∈ I N the N̄im -tuple (li , lj11 , . . . , ljN1 ,
i
m
lj12 , . . . , lj 2 2 , . . . , lj1m , . . . , ljNmm ) ∈ I N̄i , i.e., the indices of agent i and its m-neighbor set in
N
i
i
accordance to the ordering.
Definition 2.3. A transition system is a tuple T S := (Q, Act, −→), where:
• Q is a set of states.
• Act is a set of actions.
• −→ is a transition relation with −→⊂ Q × Act × Q.
3
The transition system is said to be finite, if Q and Act are finite sets. We also use the (standard)
a
notation q −→ q 0 to denote an element (q, a, q 0 ) ∈−→. For every q ∈ Q and a ∈ Act we use the
notation Post(q; a) := {q 0 ∈ Q : (q, a, q 0 ) ∈−→}.
3. Abstractions for Multi-Agent Systems
We consider multi-agent systems with single integrator dynamics
ẋi = fi (xi , xj ) + vi , i ∈ N ,
(3.1)
that are governed by decentralized control laws consisting of two terms, a feedback term fi (·)
which depends on the states of i and its neighbors, which we compactly denote by xj (= xj(i) ) :=
(xj1 , . . . , xjNi ) ∈ RNi n (see Section 2 for the notation j(i)), and an extra input term vi , which
we call free input. The dynamics fi (x, xj ) are encountered in a large set of multi-agent protocols
[19], including consensus, connectivity maintenance, collision avoidance and formation control.
In addition they may represent internal dynamics of the system as for instance in the case of
smart buildings (see e.g., [2]) , where the temperature Ti , i ∈ N of each room evolves according
P
to Ṫi = j∈Ni aij (Tj − Ti ) + vi , with aij representing the heat conductivity between rooms i
and j and vi the heating/cooling capabilities of the room.
In what follows, we consider a cell decomposition of the state space Rn and adopt a modification of the corresponding definition from [13, p 129].
Definition 3.1. Let D be a domain of Rn . A cell decomposition S = {Sl }l∈I of D, where I is
a finite or countable index set, is a family of uniformly bounded and connected sets Sl , l ∈ I,
such that int(Sl ) ∩ int(Sl̂ ) = ∅ for all l 6= ˆl and ∪l∈I Sl = D.
Throughout the report, we consider a fixed m ∈ N which specifies the m-neighbor set of each
agent and will refer to it as the degree of decentralization. Also, given a cell decomposition S :=
m
{Sl }l∈I of Rn , we use the notation li = (li , lj11 , . . . , ljN1 , lj12 , . . . , lj 2 2 , . . . , lj1m , . . . , ljNmm ) ∈ I N̄i
i
N
i
i
to denote the indices of the cells where agent i and its m-neighbors belong at a certain time
instant and call it the m-cell configuration of agent i. We will also use the shorter notation
li = (li , lj11 , . . . , ljNmm ), or just li . Similarly, we use the notation l = (l1 , . . . , lN ) ∈ I N to specify
i
the indices of the cells where all the N agents belong at a given time instant and call it the cell
configuration (of all agents). Thus, given a cell configuration l, it is possible to determine the
cell configuration of agent i as li = pri (l) (see Section 2 for the definition of pri (·)).
Our aim is to derive finite or countable abstractions for each individual agent in the coupled
system (3.1), through the selection of a cell decomposition and a time discretization step δt > 0.
These will be based on the knowledge of each agent’s neighbors discrete positions up to a certain
distance in the network graph, with the latter as specified by the degree of decentralization.
Informally, we would like to consider for each agent i the transition system with states the
possible cells of the state partition, actions all the possible cells of the agents in its m-neighbor
set and transition relation defined as follows. A final cell is reachable from an initial one, if for
all states in the initial cell there is a free input such that the trajectory of i will reach the final cell
at time δt for all possible initial states of its m-neighbors in their cells and their corresponding
free inputs. For planning purposes we will require each individual transition system to be well
posed, in the sense that for each initial cell it is possible to perform a transition to at least one
final cell.
We next illustrate the concept of a well posed space-time discretization, namely, a discretization which generates for each agent a meaningful transition system as discussed above. Consider
4
a cell decomposition as depicted in Fig. 2 and a time step δt. The tips of the arrows in the
figure are the endpoints of agent i’s trajectories at time δt. In both cases in the figure we focus
on agent i and consider the same 2-cell configuration for the 2-neighbor set N̄i2 = {i, j1 , j2 } of
i, where Ni = {j1 } and Nj1 = {j2 }, as depicted by the straight arrows in the figure. However,
we consider different dynamics for Cases (i) and (ii). In Case (i), we observe that for the three
distinct initial positions in cell Sli , it is possible to drive agent i to cell Sli0 at time δt. We
assume that this is possible for all initial conditions in this cell and irrespectively of the initial
conditions of j1 and j2 in their cells and the inputs they choose. We also assume that this
property holds for all possible 2-cell configurations of i and for all the agents of the system.
Thus we have a well posed discretization for System (i). On the other hand, for the same cell
configuration and System (ii), we observe the following. For three distinct initial conditions of
i the corresponding reachable sets at δt, which are enclosed in the dashed circles, lie in different
cells. Thus, it is not possible given this cell configuration of i to find a cell in the decomposition
which is reachable from every point in the initial cell and we conclude that the discretization
is not well posed for System (ii).
xi (δt)
xj2
Sli0
Sli
xi (δt)
xi
xj2
Slj2
xj1 Slj1
Sli
System (i):
ẋi =fi,(i) (xi ,xj1 ) + vi,(i)
ẋj1 =fj1 ,(i) (xj1 ,xj2 ) + vj1 ,(i)
xi
Slj2
xj1 Slj1
System (ii):
ẋi =fi,(ii) (xi ,xj1 ) + vi,(ii)
ẋj1 =fj1 ,(ii) (xj1 ,xj2 ) + vj1 ,(ii)
Figure 2. Illustration of a space-time discretization which is well posed for
System (i) but non-well posed for System (ii).
In order to provide meaningful decentralized abstractions we follow parts of the approach
employed in [4] and design appropriate hybrid feedback laws in place of the vi ’s in order to
guarantee well posed transitions. Before proceeding to the necessary definitions related to the
problem formulation, we provide some bounds on the dynamics of the multi-agent system. We
assume that there exists a constant M > 0 such that
|fi (xi , xj )| ≤ M, ∀(xi , xj ) ∈ R(Ni +1)n , i ∈ N
(3.2)
and that fi (·) are globally Lipschitz functions. Thus, there exist Lipschitz constants L1 > 0
and L2 > 0, such that
|fi (xi , xj ) − fi (xi , yj )| ≤ L1 |(xi , xj ) − (xi , yj )|,
(3.3)
|fi (xi , xj ) − fi (yi , xj )| ≤ L2 |(xi , xj ) − (yi , xj )|,
(3.4)
n
∀xi , yi ∈ R , xj , yj ∈ R
5
Ni n
,i ∈ N.
Furthermore, we assume that each input vi , i ∈ N is piecewise continuous and satisfies the
bound
|vi (t)| ≤ vmax , ∀t ≥ 0.
(3.5)
where vmax < M . Finally, based on the uniform bound on the diameters of the cells in the
decomposition of the workspace, we can define the diameter dmax of the cell decomposition as
dmax := inf{R > 0 : ∀l ∈ I, ∃x ∈ Sl with Sl ⊂ B(x; R2 )}.
(3.6)
This definition for the diameter of the cell decomposition is in general different from the one
adopted in [4]. In particular, according to (3.6) it is defined as the “minimum” among the
diameters of the balls that can cover each cell in the decomposition, whereas in [4], it is given as
the “minimum” among the diameters of the cells. However, for certain types of decompositions
as for instance hyper-rectangles in Rn and hexagons in R2 , these two definitions coincide.
Given a cell decomposition we will consider a fixed selection of a reference point xl,G ∈ Sl for
each cell Sl , l ∈ I. For each agent i and m-cell configuration of i, the corresponding reference
points of the cells in the configuration will provide a trajectory which is indicative of the agent’s
reachability capabilities over the interval of the time discretization. In particular, for appropriate space-time discretizations the agent will be capable of reaching all points inside a ball with
center the endpoint of this trajectory at the end of the time step. Furthermore, the selected
reference points provide an estimate of the corresponding trajectories of the agent’s neighbors,
for m-cell configurations of its neighbors that are consistent with the cell configuration of i,
namely for which the “common agents” belong to the same cells, as formally given in Definition
3.3 below. Before proceeding to Definition 3.3 we provide a lemma which establishes certain
useful properties of the agents’ m-neighbor sets.
Lemma 3.2. (i) For each agent i ∈ N , neighbor ` ∈ Ni of i and m ≥ 1, it holds N̄`m−1 ⊂ N̄im .
(ii) For each i ∈ N , m ≥ 1 and ` ∈ N̄im it holds N` ⊂ N̄im+1 .
(iii) Assume that for certain i ∈ N and m ≥ 1 it holds Nim+1 = ∅. Then, for each ` ∈ N̄im it
holds N` ⊂ N̄im .
κ
0
Proof. For the proof of Part (i) let any i0 ∈ N̄`m−1 . Then, since N̄`m−1 = ∪m−1
κ=0 N` , either i = `,
0
which implies that i0 ∈ Ni and hence also i0 ∈ N̄im , or i0 ∈ N`m for certain m0 ∈ {1, . . . , m}.
In addition, if i0 = i, then i0 ∈ N̄im and hence, it remains to consider the case where i0 6= i and
0
i0 ∈ N`m for some m0 ∈ {1, . . . , m}. The latter implies that there exists a shortest path i0 . . . im0
with i0 = i0 and im0 = ` from i0 to `. Then, since im0 = ` ∈ Ni , it follows that i0 . . . im0 i is
a path of length m0 + 1 ≤ m from i0 to i. Thus, either it is a shortest path, implying that
0
i0 ∈ Nim +1 ⊂ N̄im , or, since i0 6= i, there exist 0 < m00 < m0 + 1 and a shortest path of length
0
m00 joining i0 and i. In the latter case, it follows that i0 ∈ Nim and hence again that i0 ∈ N̄im .
The proof of Part (i) is now complete.
For the proof of Part (ii) let ` ∈ N̄im and i0 ∈ N` . If i0 = i then i0 ∈ N̄im+1 . Otherwise, since
` ∈ N̄im there exists 1 ≤ m0 ≤ m and a path i0 . . . im0 of length m0 with i0 = ` and im0 = i,
implying that i0 i0 . . . im0 is a path of length m0 + 1 ≤ m + 1 from i0 to i. Thus, it follows that
i0 ∈ N̄im+1 .
For the proof of Part (iii) let ` ∈ N̄im and i0 ∈ N` . If ` = i or i0 = i, then the result
follows directly from the facts that Ni ⊂ N̄im and i ∈ N̄im , respectively. Otherwise, there exists
0
1 ≤ m0 ≤ m such that ` ∈ Nim , implying that there exists a path i0 . . . im0 of length m0 with
i0 = ` and im0 = i. Thus, since i0 ∈ N` , we get that i0 i0 . . . im0 is a path of length m0 + 1
from i0 to i. If it is a shortest path, then it follows that m0 < m, because otherwise, since
6
i0 6= i, we would have a shortest path of length m + 1 joining i0 and i, implying that Nim+1 6= ∅
0
and contradicting the hypothesis of Part (iii). Hence, it holds in this case that i0 ∈ Nim +1 for
certain 1 ≤ m0 < m and thus, that i0 ∈ N̄im . Finally, if i0 i0 . . . im0 is not a shortest path, then
00
there exists a shortest path of length 1 ≤ m00 < m joining i0 and i, which implies that i0 ∈ Nim
and hence again, that i0 ∈ N̄im .
Definition 3.3. Consider an agent i ∈ N , a neighbor ` ∈ Ni of i and m-cell configurations
li = (li , lj11 , . . . , ljNmm ) and l` = (¯l` , ¯lj(`)11 , . . . , ¯lj(`)m
) of i and `, respectively. We say that l` is
Nm
i
`
m
m
consistent with li if for all κ ∈ N̄ ∩ N̄ it holds lκ = ¯lκ . /
i
`
The following definition provides for each agent i its reference trajectory and the estimates
of its neighbors’ reference trajectories, based on i’s m-cell configuration.
Definition 3.4. Given a cell decomposition S = {Sl }l∈I of Rn , a reference point xl,G ∈ Sl
for each l ∈ I, a time step δt and a nonempty subset W of Rn , consider an agent i ∈ N ,
its m-neighbor set N̄im and an m-cell configuration li = (li , lj11 , . . . , ljNmm ) of i. We define the
i
functions χi (t), χj (t) := (χj1 (t), . . . , χjNi (t)), t ≥ 0, through the solution of the following initial
value problem, specified by Cases (i) and (ii) below:
Case (i). Nim+1 = ∅. Then we have the initial value problem
χ̇` (t) =f` (χ` (t), χj(`) (t)) = f` (χ` (t), χj(`)1 (t), . . . , χj(`)N` (t)), t ≥ 0, ` ∈ N̄im ,
χ` (0) =xl` ,G , ∀` ∈ N̄im ,
(3.7)
where j(`)1 , . . . , j(`)N` denote the corresponding neighbors of each agent ` ∈ N̄im .
Case (ii). Nim+1 6= ∅. Then we have the initial value problem
χ̇` (t) =f` (χ` (t), χ` (t)) = f` (χ` (t), χj(`)1 (t), . . . , χj(`)Ni (t)), t ≥ 0, ` ∈ N̄im−1 ,
χ` (0) =xl` ,G , ∀` ∈ N̄im−1 ,
(3.8)
with the terms χ` (·), ` ∈ Nim defined as
χ` (t) := xl` ,G , ∀t ≥ 0, ` ∈ Nim . /
(3.9)
Remark 3.5. (i) Notice, that in Case (i) for the initial value problem in Definition 3.4, the
requirement Nim+1 = ∅ implies by vitue of Lemma 3.2(iii) for each agent ` ∈ N̄im its neighbors
j(`)1 , . . . , j(`)N` also belong to N̄im . Hence, the subsystem formed by the agents in N̄im is
decoupled from the other agents in the system and the initial value problem (3.7) is well defined.
(ii) In Case (ii), the subsystem formed by the agents in N̄im is not decoupled from the other
agents in the system. However, by considering the agents in Nim fixed at their reference points
as imposed by (3.9), it follows that the initial value problem (3.8)-(3.9) is also well defined.
[i]
(iii) Apart from the notation χi (·) and χj (·) above, we will use the notation χ` (·) for the
[i]
trajectory of each agent ` ∈ N̄im (including the functions χ` (·) as defined by (3.9) when
Nim+1 6= ∅), as specified by the initial value problem corresponding to the m-cell configuration
[i]
of i in Definition 3.4. We will refer to χi (·) ≡ χ` (·) as the reference trajectory of agent i and
[i]
to each χ` (·) with ` ∈ N̄im \ {i} as the estimate of `’s reference trajectory by i.
Example 3.6. In this example we demonstrate the IVPs of Definition 3.4 for m = 3, as
specified by Cases (i) and (ii) for agents 5 and 1, respectively. The topology of the network is
7
the same as in Example 2.2 and the agents involved in the solution of the correposnding initial
value problems are depicted in Fig. 3.
χ̇5 (t) = f5 (χ5 (t), χ3 (t))
χ̇3 (t) = f3 (χ3 (t), χ2 (t), χ4 (t))
χ̇2 (t) = f2 (χ2 (t), χ3 (t))
Agent 5 - Case (i)
χ̇4 (t) = f4 (χ4 (t))
χκ (0) = xκ,G , κ = 5, 3, 2, 4
χ̇1 (t) = f1 (χ1 (t), χ6 (t), χ2 (t))
χ̇6 (t) = f6 (χ6 (t), χ1 (t), χ2 (t), χ5 (t), χ7 (t))
χ̇2 (t) = f2 (χ2 (t), χ3 (t))
χ̇5 (t) = f5 (χ5 (t), χ3 (t))
Agent 1 - Case (ii)
χ̇3 (t) = f3 (χ3 (t), χ2 (t), x4,G )
χ̇7 (t) = f7 (χ7 (t), χ6 (t), x8,G )
χκ (0) = xκ,G , κ = 1, 6, 2, 5, 3, 7
Ag.7
Ag.8
Ag.6
Ag.5
Ag.1
Ag.4
Ag.2
Ag.3
Figure 3. The neighbors on the right of the dotted line are considered fixed
for the IVP of the 3-cell configuration of agent 1.
We next show that under certain conditions on the structure of the network graph in a
neighborhood of each agent i, the reference trajectories of agent i’s neighbors coincide with
their estimates by i, for consistent cell configurations. The following lemma provides this
result.
Lemma 3.7. Assume that for agent i ∈ N it holds Nim+1 = ∅ and let li be an m-cell configuration of i. Then, for every agent ` ∈ Ni with N`m+1 = ∅ and m-cell configuration l` of `
[`]
[i]
[`]
[i]
consistent with li , it holds χ` (t) = χ` (t), for all t ≥ 0, with χ` (·) and χ` (·) as determined by
the initial value problem of Definition 3.4 for the m-cell configurations l` and li , respectively.
Proof. Since Nim+1 = ∅, it follows that the initial value problem which corresponds to the
[i]
m-cell configuration of agent i and specifies the trajectory χ` (·) of ` is provided by Case (i) of
Definition 3.4, namely, by (3.7). We rewrite (3.7) in the compact form
Ẋ = F (X); F = (fκ1 , . . . , fκN̄ m ), X = (χκ1 , . . . , χκN̄ m ),
i
8
i
(3.10)
with initial condition χκν (0) = xlκν ,G , ν = 1, . . . , N̄im , and κν being the ν-th index of N̄im
according to the total order of Definition 2.1. Next, since N`m+1 = ∅, we similarly obtain that
the initial value problem which corresponds to the m-cell configuration of agent ` and specifies
[`]
its reference trajectory χ` (·) is given as
Ẋ1 = F1 (X1 ); F1 = (fκ01 , . . . , fκ0N̄ m ), X1 = (χκ01 , . . . , χκ0N̄ m ),
`
(3.11)
`
with initial condition χκ0ν (0) = xlκ0 ν ,G , ν = 1, . . . , N̄`m , and κ0ν being the corresponding νth index of N̄`m . By taking into account that Nim+1 = ∅, we get from Lemma 3.2(i) that
N̄`m ⊂ N̄im+1 = N̄im ∪ Nim+1 = N̄im . By assuming that without any loss of generality the
inclusion is strict, we obtain from (3.10), (3.11) and Remark 3.5(i) for ` (namely, that the
subsystem formed by the agents in N̄`m is decoupled from the other agents), that by a reordering
of the components χκν , ν = 1, . . . , N̄im , (3.10) can be cast in the form
Ẋ1 = F1 (X1 )
Ẋ2 = F2 (X1 , X2 ),
(3.12)
with X1 and F1 (·) as in (3.11), X2 = (χκ0N̄ m +1 , . . . , χκ0N̄ m ), F2 = (fκ0N̄ m +1 , . . . , fκ0N̄ m ), and the
i
`
i
`
same initial condition for the X1 part as (3.11), due to the consistency of l` with li . Hence,
since the X1 part of the solution in (3.12) is independent of X2 , it follows that the reference
[`]
[i]
trajectory χ` (·) of agent ` as given by (3.11) and its corresponding estimate χ` (·) obtained
from the first (decoupled) subsystem in (3.12) coincide.
We next check the conditions of Lemma 3.7 for the graphs depicted in Fig. 4 below. Consider
first the graph of Case I, and assume that (i, `) ∈
/ E, where the interconnection between i and `
is depicted through the non-dashed arrow. Then, by selecting m = 3 and m = 4, respectively,
it turns out that Ni4 = Ni5 = ∅. However, we have N`4 = {j4 } and N`5 = {j5 }, respectively,
which implies that the conditions of Lemma 3.7 are not fulfilled. The same observation holds
also for the graph of Case II when (i, `) ∈
/ E. On the other hand, when (i, `) ∈ E it follows in
Case I that for both m = 3 and m = 4 it holds Ni4 = Ni5 = ∅ and N`4 = N`5 = ∅, namely the
requirements of Lemma 3.7 are satisfied. However, for Case II (again when (i, `) ∈ E), these
are only satisfied for m = 4 where Ni5 = ∅ and N`5 = ∅, since for m = 3 it holds Ni4 = ∅ and
N`4 = {j5 } =
6 ∅.
Ag.j1
Ag.`
Ag.j1
Ag.j2
Ag.`
Ag.j2
Ag.j5
Ag.i
Ag.j5
Ag.j4
Ag.j3
Case I.
Ag.i
Ag.j3
Ag.j4
Case II.
Figure 4. Illustration of graph topologies that satisfy the conditions of
Lemma 3.8 for m = 4.
9
We note that in both cases discussed above, the fact that when (i, `) ∈ E, the condition
Ni4 = ∅ implies that N`5 = ∅ is justified by the following lemma, which provides sufficient
conditions for the assumptions of Lemma 3.7 to hold.
Lemma 3.8. Assume that for agent i it holds Nim = ∅ for certain m ≥ 2. (i) Then, for every
agent ` ∈ Ni it holds
N̄`m+κ ⊂ N̄im , ∀κ ≥ 0.
(3.13)
(ii) Furthermore, if i ∈ N` , then it also holds
N`m+1 = ∅.
(3.14)
Nim+κ = ∅, ∀κ ≥ 1.
(3.15)
Proof. For Part (i), we first show that
Indeed, assume on the contrary that i0 ∈ Nim+κ 6= ∅ for certain κ ≥ 1 and thus, there exists a
shortest path i0 i1 . . . im−1+κ im+κ of length m+κ in G with i0 = i0 and im+κ = i. Then, it follows
that iκ iκ+1 . . . im−1+κ im+κ is a shortest path of length m from iκ to i, because otherwise, if there
were a shorter one iκ i0κ+1 . . . i0m0 −1+κ im+κ with m0 < m, i0 i1 . . . iκ i0κ+1 . . . i0m0 −1+κ im+κ would
be a path of length m0 +κ < m+κ from i0 to i, contradicting minimality of i0 i1 . . . im−1+κ im+κ .
Thus, we get that iκ ∈ Nim , which contradicts the assumption that Nim 6= ∅ and we conclude
that (3.15) is fulfilled. By exploiting the latter, we get that for each κ ≥ 0 (3.13) holds, since
Sκ+1
0
by virtue of Lemma 3.2(i) we have N̄`m+κ ⊂ N̄im+κ+1 = N̄im ∪ κ0 =1 Nim+κ = N̄im .
For the proof of Part (ii), we proceed again by contradiction. Hence, consider a shortest
path i0 i1 . . . im im+1 with im+1 = `, and notice that by virtue of (3.13) with κ = 1, it holds
i0 ∈ N̄im . We consider two cases. If i0 = i, then by virtue of the fact that i ∈ N` , which implies
that (i, `) ∈ E, we have that i0 ` is a path of length 1 joining i0 and `, contradicting minimality
of i0 i1 . . . im im+1 . Now consider the case where i0 6= i. Then, it follows from the facts that
i0 ∈ N̄im , i0 6= i and the hypothesis Nim = ∅, that there exists a shortest path i0 i01 . . . i0m0 −1 i of
length m0 < m, joining i0 and i. Thus, since (i, `) ∈ E we obtain that i0 i01 . . . i0m0 −1 i` is a path
of length m0 + 1 ≤ m joining i0 and `, contradicting minimality of i0 i1 . . . im im+1 . The proof
is now complete.
Despite the result of Lemma 3.7, in principle, the actual trajectory of each agent’s neighbor
and its estimate, based on the solution of the initial value problem for the reference trajectory of
the specific agent do not coincide. Explicit bounds for the deviation between these trajectories
and their estimates are given in Proposition 3.10 below, whose proof requires the following
auxiliary lemma.
Lemma 3.9. Consider the agent i ∈ N and let li be an m-cell configuration of i. Also, pick
` ∈ Ni and any m-cell configuration l` of ` consistent with li . Finally, let t∗ be the unique
positive solution of the equation
L22
L 2 t∗
e
− L2 + √
t∗ − 1 = 0,
(3.16)
L1 Nmax
with
Nmax := max{Ni : i ∈ N }.
Then, for each κ ∈
N̄`m−1
(recall that
N̄`m−1
⊂
N̄im
by Lemma 3.2(i)) it holds:
[`]
∗
|χ[i]
κ (t) − χκ (t)| ≤ M t, ∀t ∈ [0, t ],
10
(3.17)
(3.18)
[i]
[`]
where χκ (·) and χκ (·) are determined by the initial value problem of Definition 3.4 for the
m-cell configurations li and l` , respectively.
Proof. The proof of the lemma is given in the Appendix.
Proposition 3.10. Consider the agent i ∈ N and let li be an m-cell configuration of i. Also,
[i]
pick ` ∈ Ni and any m-cell configuration l` of ` consistent with li . Then the difference |χ` (·) −
[`]
χ` (·)| satisfies the bound:
[i]
[`]
|χ` (t) − χ` (t)| ≤ Hm (t), ∀t ∈ [0, t∗ ],
where t∗ is given in (3.16), the functions Hκ (·), κ ≥ 1, are defined recursively as
Z t
p
H1 (t) := M t, t ≥ 0; Hκ (t) :=
eL2 (t−s) L1 Nmax Hκ−1 (s)ds, t ≥ 0
(3.19)
(3.20)
0
[i]
[`]
and χ` (·), χ` (·) are determined by the initial value problem of Definition 3.4 for the m-cell
configurations li and l` , respectively.
Proof. The proof is carried out by induction and is based on the result of Lemma 3.9. We will
prove the following induction hypothesis:
0
IH. For each m0 ∈ {1, . . . , m} and ι ∈ N̄`m−m it holds
[`]
∗
|χ[i]
ι (t) − χι (t)| ≤ Hm0 (t), ∀t ∈ [0, t ].
(3.21)
Notice that for m0 = m the Induction Hypothesis implies the desired (3.19). Also, by virtue
of Lemma 3.9, IH is valid for m0 = 1. In order to prove the general step, assume that IH is
m−(m0 +1)
fulfilled for certain m0 ∈ {1, . . . , m − 1} and consider any ι ∈ N̄`
. Notice first, that
[i]
[`]
since m − (m0 + 1) ≤ m − 2, both χj(ι) (·) and χj(ι) (·) are well defined and the differences
[i]
[`]
|χν (·) − χν (·)|, ν ∈ Nj(ι) of their respective components satisfy (3.21) with m0 . It then follows
1
P
√
[`]
[i]
[`]
[i]
2 2
(t)|
(t)
−
χ
that |χj(ι) (t)−χj(ι) (t)| =
|χ
≤ Nι Hm0 (t) for all t ∈ (0, t∗ ). Thus,
ν
ν
ν∈Nj(i)
[i]
[`]
by evaluating |χι (·) − χι (·)| as in (8.4) in the Appendix, we obtain that
Z t p
Z t
[`]
[`]
0 (s)ds +
|χ[i]
(t)
−
χ
(t)|
≤
L
N
H
L2 |χ[i]
1
ι
m
ι
ι
ι (s) − χι (s)|ds
0
0
By exploiting Fact I used in the proof of Lemma 3.9 in the Appendix, we obtain from (3.21)and
(3.20) that
Z t
p
[`]
|χ[i]
(t)
−
χ
(t)|
≤
eL2 (t−s) L1 Nι Hm0 (s)ds
ι
ι
0
Z t
p
≤
eL2 (t−s) L1 Nmax Hm0 (s)ds = Hm0 +1 (t), ∀t ∈ [0, t∗ ],
0
which establishes the general step of the induction procedure. The proof is now complete.
Remark 3.11. Explicit formulas for the functions Hκ (·) in (3.20) are provided in the Appendix.
In addition, we provide some linear upper bounds for them in the following section, which are
then further utilized for the explicit derivation of acceptable space-time discretizations.
11
In order to provide the definition of well posed transitions for the individual agents, we will
consider for each agent i ∈ N the following system with disturbances:
ẋi = fi (xi , dj ) + vi ,
(3.22)
where dj1 , . . . , djNi : [0, ∞) → Rn (also denoted d` , ` ∈ Ni ) are continuous functions. Also,
before defining the notion of a well posed space-time discretization we provide a class of hybrid
feedback laws which are assigned to the free inputs vi in order to obtain meaningful discrete
transitions. For each agent, these control laws are parameterized by the agent’s initial conditions
and a set of auxiliary parameters belonging to a nonempty subset W of Rn . These parameters,
as will be clarified in the next section, are exploited for motion planning. In addition, for each
agent i, the feedback laws in the following definition depend on the selection of the cells where
i and its m-neighbors belong.
Definition 3.12. Given a cell decomposition S = {Sl }l∈I of Rn and a nonempty subset W of
Rn , consider an agent i ∈ N and an initial cell configuration li of i. For each xi0 ∈ Sli and
wi ∈ W , consider the mapping ki,li (·, ·, ·; xi0 , wi ) : [0, ∞) × R(Ni +1)n → Rn , parameterized by
xi0 ∈ Sli and wi ∈ W . We say that ki,li (·) satisfies Property (P), if the following conditions are
satisfied.
(P1) The mapping ki,li (t, xi , xj ; xi0 , wi ) is continuous on [0, ∞) × R(Ni +1)n × Sli × W .
(P2) The mapping ki,li (t, ·, ·; xi0 , wi ) is globally Lipschitz continuous on (xi , xj ) (uniformly with
respect to t ∈ [0, ∞), xi0 ∈ Sli and wi ∈ W ). /
The next definition characterizes the bounds on the deviation between the reference trajectory
of each agent’s neighbor and its estimate obtained from the solution of the initial value problem
for the specific agent.
Definition 3.13. Consider an agent i ∈ N . We say that a continuous function αi : [0, δt] →
R≥0 satisfies the neighbor reference trajectory deviation bound, if for each cell configuration li
of i, neighbor ` ∈ Ni of i and cell configuration of ` consistent with li it holds:
[i]
[`]
|χ` (t) − χ` (t)| ≤ αi (t), ∀t ∈ [0, δt]. /
(3.23)
We are now in position to formalize our requirement which describes the possibility for an
agent to perform a discrete transition, based on the knowledge of its m-cell configuration. The
corresponding definition below includes certain bounds on the evolution of the agent, which we
sharpen, by requiring the reference points xl,G , l ∈ I of the cell decomposition to satisfy
|xl,G − x| ≤
dmax
, ∀x ∈ Sl , l ∈ I.
2
(3.24)
Due to (3.6), the latter is always possible. The definition exploits the auxiliary system with
disturbances (3.22), which is inspired by the approach adopted in [10], where a nonlinear system
is modeled by means of a piecewise affine system with disturbances.
Definition 3.14. Consider a cell decomposition S = {Sl }l∈I of Rn , a time step δt, a nonempty
subset W of Rn , and a continuous function β : [0, δt] → R≥0 satisfying
dmax
≤ β(0); β(δt) ≤ vmax δt.
2
(3.25)
Also, consider an agent i ∈ N , a continuous function αi : [0, δt] → R≥0 satisfying the neighbor
reference trajectory deviation bound (3.23), an m-cell configuration li of i and the solution of
12
the initial value problem of Definition 3.4. Then, given a control law
vi = ki,li (t, xi , xj ; xi0 , wi )
(3.26)
as in Definition 3.12, that satisfies Property (P), a vector wi ∈ W , and a cell index l0 ∈ I, we say
that the Consistency Condition is satisfied if the following hold. The set int(B(χi (δt); β(δt))) ∩
Sli0 is nonempty, and there exists a point x0i ∈ int(B(χi (δt); β(δt))) ∩ Sli0 , such that for each
initial condition xi0 ∈ int(Sli ) and selection of continuous functions d` : R≥0 → Rn , ` ∈ Ni
satisfying
[i]
|d` (t) − χ` (t)| ≤ αi (t) + β(t), ∀t ∈ [0, δt],
(3.27)
[i]
where χ` (·), ` ∈ Ni correspond to i’s estimates of its neighbors’ reference trajectories, the
solution xi (·) of the system with disturbances (3.22) with vi = ki,li (t, xi , dj ; xi0 , wi ), satisfies
[i]
|xi (t) − χi (t)| < β(t), ∀t ∈ [0, δt].
(3.28)
xi (δt) = x0i ∈ Sli0
(3.29)
Furthermore, it holds
and
|ki,li (t, xi (t), dj (t); xi0 , wi )| ≤ vmax , ∀t ∈ [0, δt].
/
(3.30)
Notice, that when the Consistency Condition is satisfied, agent i can be driven to cell Sli0
precisely in time δt under the feedback law ki,li (·) corresponding to the given parameter wi in
the definition. The latter is possible for all disturbances which satisfy (3.27) and capture the
possibilities for the evolution of i’s neighbors over the time interval [0, δt], given the knowledge
of the m-cell configuration of i.
We next provide some intuition behind the Consistency Condition and the particular selection
of the neighbor reference trajectory deviation bounds αi (·) and the function β(·) introduced
therein, through an example with four agents as illustrated in Fig. 5. The agents i, `, j1 and
j2 form a path graph, and their neighbor sets are given as Ni = {`}, N` = {j1 }, Nj1 = {j2 }
and Nj2 = ∅, respectively. The degree of decentralization is assumed to be m = 3 and we focus
primarily on agent i. Agent i is depicted at the left of the figure together with its reference
[i]
trajectory χi (·) = χi (·) initiated from the reference point xli ,G . The latter constitutes the
center of the ball with minimal radius that encloses the cell Sli where agent i is contained,
as specified by (3.24) and depicted with the dashed circle in the figure. The (green) area
enclosing agent i is the set ∪t∈[0,δt] B(χi (t); β(t)), namely, the union of all possible positions
of the agent whose distance from the reference trajectory at each time t does not exceed the
bound β(t), or equivalently satisfy (3.28). This is the restriction that the Consistency Condition
imposes on agent i, and implicitly through the acceptable disturbances that satisfy (3.27), to its
[i]
neighbor `. The latter is depicted through the larger (red) area ∪t∈[0,δt] B(χ` (t); αi (t) + β(t)),
which encloses the reference point xl` ,G of agent `. The interior (darker red) part of this area
[i]
is ∪t∈[0,δt] B(χ` (t); αi (t)), namely, the union of the points with distance from the (dashed)
[i]
reference trajectory χ` (t) of ` as estimated by i is no more than the reference trajectory
deviation bound ai (t) at each t. The exterior part depicts the inflation ot the interior one by
β(·). Thus, when the Consistency Condition is applied to agent `, namely, when its motion
[`]
over [0, δt] with respect to its own reference trajectory χ` (t) is bounded at each t by β(t), as
depicted with the dotted area in the figure, agent ` will remain within the larger area enclosing
its reference point xl` ,G . Therefore, in order to capture the behaviour of agent `, it is assumed
that the disturbances d` (·) which model the possible trajectories of ` satisfy (3.27) and thus
13
belong to the lager (red) area containing the reference point xl` ,G . The reason for this hypothesis
[i]
comes from the fact that the m-cell configuration of i allows only for the computation of χ` (·)
[`]
and not the actual reference trajectory χ` (·) of `.
On the right hand side of the figure we also illustrate the trajectories of the reference solutions
[j ]
[`]
[i]
χj11 (·), χj1 (·) and χj1 (·) for agent j1 , as evaluated by the agent itself and agents ` and i,
[j ]
respectively. As was the case for agent `, the reference trajectory χj11 (t) of j1 and its estimate
[`]
χj1 (t) by ` do not exceed the reference trajectory deviation bound ai (t) at each time t. However,
[i]
as shown in the figure, the same does not necessarily hold for the corresponding estimate χj1 (·)
[j ]
by i, since χj11 (·) lies outside the dotted area depicting the reference trajectory deviation bound
[i]
around χj1 (·). Finally, it is noted that according to the Consistency Condition, agent i can be
assigned a feedback law in order to reach the final cell Sli0 from each point inside the initial cell
Sli and for any disturbance satisfying the left hand side of (3.28). Such a path of agent i and
a corresponding disturbance are depicted in the figure. More details on the specific feedback
laws that are applied for this purpose are provided in the next section.
Sl0i
[i]
χi (δt)
d` (δt)
β(δt)
[`]
χ` (δt)
[i]
χ` (δt)
αi (δt)
[j ]
χj (δt)
β(δt)
[`]
χj (δt)
[i]
χj (δt)
dmax
β(0)
Sli
xli ,G
β(0)
xlj ,G
xl` ,G
xlj ,G
Figure 5. Illustration of the estimates on agent’s i reachable set over the
interval [0, δt] and the reference trajectory deviation bound.
We next provide the definition of a well posed space-time discretization. This definition
formalizes the acceptable space and time discretizations through the possibility to assign a
feedback law to each agent, in order to enable a meaningful transition from an initial to a final
cell in accordance to the Consistency Condition above.
Definition 3.15. Consider a cell decomposition S = {Sl }l∈I of Rn , a time step δt and a
nonempty subset W of Rn .
14
(i) Given a continuous function β : [0, δt] → R≥0 that satisfies (3.25), an agent i ∈ N , a
continuous function αi : [0, δt] → R≥0 satisfying (3.23), an initial m-cell configuration li of
l
i
li0 is well posed with respect to
i, and a cell index li0 ∈ I, we say that the transition li −→
the space-time discretization S − δt, if there exist a feedback law vi = ki,li (·, ·, ·; xi0 , wi ) as in
Definition 3.12 that satisfies Property (P), and a vector wi ∈ W , such that the Consistency
Condition of Definition 3.14 is fulfilled.
(ii) We say that the space-time discretization S − δt is well posed, if there exists a continuous
function β : [0, δt] → R≥0 that satisfies (3.25), such that for each agent i ∈ N there exists a
continuous function αi : [0, δt] → R≥0 satisfying (3.23), in a way that for each cell configuration
l
i
li0 is well posed with respect
li of i, there exists a cell index li0 ∈ I such that the transition li −→
to S − δt.
Given a space-time discretization S − δt and based on Definition 3.15(i), it is now possible to
provide an exact definition of the discrete transition system which serves as an abstract model
for the behaviour of each agent.
Definition 3.16. For each agent i, its individual transition system T Si := (Qi , Acti , −→i ) is
defined as follows:
• Qi := I (the indices of the cell decomposition)
m
• Acti := I N̄i (the set of all m-cell configurations of i)
l
l
i
i
0
0
0
1 , l j 2 , . . . , lj 2 ,
• li −→
i li iff li −→ li is well posed for each li , li ∈ Qi and li = (li , lj11 , . . . , ljN
1
2
i
m
N
i
. . . , lj1m , . . . , ljNmm ) ∈ I N̄i . /
i
Remark 3.17. Given a well posed space-time discretization S − δt and an initial cell configuration l = (l1 , . . . , lN ) ∈ I N , it follows from Definitions 3.15 and 3.16 that for each agent i ∈ N
it holds Posti (li ; pri (l)) 6= ∅ (Posti (·) refers to the transition system T Si of each agent).
According to Definition 3.15, a well posed space-time discretization requires the existence of
a well posed transition for each agent i and m-cell configuration of i, and the latter reduces to
the selection of an appropriate feedback controller for i, which also satisfies Property (P) and
guarantees that the auxiliary system with disturbances (3.22) satisfies the Consistency Condition. We next show, that given an initial cell configuration and a well posed transition for each
agent, it is possible to choose a local feedback law for each agent, so that the resulting closedloop system will guarantee all these well posed transitions (for all possible initial conditions in
the cell configuration). At the same time, it will follow that the magnitude of the agents’ hybrid
control laws evaluated along the solutions of the system does not exceed the maximum allowed
magnitude vmax of the free inputs on [0, δt], and hence, establishes consistency with (3.5).
Proposition 3.18. Consider system (3.1), let l = (l1 , . . . , lN ) ∈ I N be an initial cell configuration and assume that the space-time discretization S − δt is well posed, which according
to Remark 3.17 implies that for all i ∈ N it holds that Posti (li ; pri (l)) 6= ∅. Then, for every
0
final cell configuration l0 = (l10 , . . . , lN
) ∈ Post1 (l1 ; pr1 (l)) × · · · × PostN (lN ; prN (l)), there exist
feedback laws
vi = ki,pri (l) (t, xi , xj ; xi0 , wi ), i ∈ N ,
(3.31)
satisfying Property (P), w1 , . . . , wN ∈ W and x01 ∈ Sl1 , . . . , x0N ∈ SlN , such that the solution of
the closed-loop system (3.1), (3.31) (with vκ = kκ,prκ (l) , κ ∈ N ) is well defined on [0, δt], and
for each i ∈ N , its i-th component satisfies
xi (δt, x(0)) = x0i ∈ Sli0 , ∀x(0) ∈ RN n : xκ (0) = xκ0 ∈ Slκ , κ ∈ N .
15
(3.32)
Furthermore, it follows that each control law ki,li evaluated along the corresponding solution of
the system satisfies
|ki,pri (l) (t, xi (t), xj (t); xi0 , wi )| ≤ vmax , ∀t ∈ [0, δt], i ∈ N ,
(3.33)
which provides the desired consistency with the design requirement (3.5) on the vi ’s.
0
Proof. Indeed, consider a final cell configuration l0 = (l10 , . . . , lN
) as in the statement of the
proposition. By the definitions of the operators Posti (·), i ∈ N and the transition relation of
each corresponding agent’s individual transition system, it follows that there exist continuous
functions β : [0, δt] → R≥0 satisfying (3.25) and αi : [0, δt] → R≥0 , i ∈ N satisfying (3.23),
pr (l)
i
such that each transition li −→
li0 is well posed in the sense of Definition 3.15(i). Hence, we
can pick for each agent i ∈ N a control law ki,pri (l) (·) that satisfies Property (P) and vectors
wi ∈ W , x0i ∈ Sl0i , such that the requirements of the Consistency Condition are fulfilled.
After the selection of ki,pri (l) (·), wi and x0i , we pick for each agent i an initial condition
xi0 ∈ int(Sli ). Notice, that by virtue of Property (P2) of the control laws, the solution of the
closed loop system is defined for all t ≥ 0. Furthermore, by recalling that for each i ∈ N it
holds xi0 ∈ int(Sli ), we obtain from (3.24) that |xi0 − xli ,G | < dmax
for all i ∈ N . Hence, by
2
continuity of the solution of the closed-loop system (3.1), (3.31) we deduce from (3.25) that
there exists δ ∈ (0, δt], such that for all i ∈ N it holds
[i]
|xi (t) − χi (t)| < β(t), ∀t ∈ [0, δ],
(3.34)
[i]
where xi (·) is the i-th component of the solution and χi (·) is the reference trajectory of i
[i]
corresponding to the m-cell configuration pri (l) of i, with initial condition χi (0) = xli ,G . We
claim that for each i ∈ N , (3.34) holds with δ = δt. Indeed, suppose on the contrary that there
exist an agent ` ∈ N and a time T ∈ (0, δt], such that
[`]
|x` (T ) − χ` (T )| ≥ β(T ).
(3.35)
Next, we define
[i]
τ := sup{t̄ ∈ (0, δt] : |xi (t) − χi (t)| < β(t), ∀t ∈ [0, t̄], i ∈ N }
(3.36)
Then, it follows from (3.34), (3.35) and (3.36), that τ is well defined and satisfies
0 < τ ≤ δt,
and that there exists ` ∈ N such that
[`]
|x` (τ ) − χ` (τ )| = β(τ ).
(3.37)
Next, notice that by the definition of τ , it holds
|xκ (t) − χκ[κ] (t)| ≤ β(t), ∀t ∈ [0, τ ], κ ∈ N` .
(3.38)
Also, since for each κ ∈ N` the m-cell configuration prκ (l) of κ is consistent with the configuration pr` (l) of `, it follows from (3.23) and the fact that τ < δt, that
[κ]
|χ[`]
κ (t) − χκ (t)| ≤ α` (t), ∀t ∈ [0, τ ], κ ∈ N` .
(3.39)
|xκ (t) − χ[`]
κ (t)| ≤ β(t) + α` (t), ∀t ∈ [0, τ ], κ ∈ N` .
(3.40)
Hence, we obtain that
16
By setting dκ (t) := xκ (t), t ≥ 0, κ ∈ N` , it follows from standard uniqueness results from
ODE theory, that x` (·) is also the solution of the system with disturbances (3.22), with i = `,
vi = ki,li (t, xi , dj ; xi0 , wi ), and initial condition x`0 ∈ int(Sl` ). Thus, by exploiting causality of
(3.22) with respect to the disturbances and the fact that due to (3.40) the disturbances satisfy
[`]
(3.27) for all t ∈ [0, τ ), it follows from (3.28) that |x` (τ ) − χ` (τ )| < β(τ ), which contradicts
(3.37). Hence, we conclude that (3.34) holds for δ = δt.
Next, by using the same arguments as above, we can deduce that for each agent i, the i-th
component of the solution of the closed loop system (3.1), (3.31), is the same as the solution
of system (3.22) for i, with disturbances dκ (·), κ ∈ Ni being the components xκ (·), κ ∈ Ni
of the solution corresponding to i’s neighbors. Furthermore, it follows that the disturbances
satisfy (3.27). Hence, from the Consistency Condition, and the fact that the components of
the solution of the closed loop system and the corresponding solutions of the systems with
disturbances are identical, we obtain that (3.32) and (3.33) are satisfied.
For the general case where xi0 ∈ Sli (not necessarily the interior of Sli ) for all i ∈ N , we
can exploit for each i continuity of the i-th component of the solution xi (·) with respect to
initial conditions and parameters, in order to prove validity of the proposition. In particular,
xi (t, x0 ; x0 , w) depends by virtue of Property (P1) continuously on x0 (both as the initial
condition and as a parameter) and on the parameters w = (w1 , . . . , wN ) ∈ W N . Thus, by
selecting a sequence {x0,ν }ν∈N with x0,ν → x0 and x0i,ν ∈ int(Sli ), ∀i ∈ N , ν ∈ N, it follows
that the components xi (·, x0,ν ; x0,ν , w) of the solution of the closed loop system (3.1), (3.31)
satisfy
xi (t, x0,ν ; x0,ν , w) → xi (t, x0 ; x0 , w), ∀t ∈ [0, δt], i ∈ N ,
(3.41)
with the vector w of the parameters wi , i ∈ N as selected at the beginning of the proof. In
addition, since x0i,ν ∈ int(Sli ), ∀i ∈ N , ν ∈ N, it follows from the first part of the proof that
the functions xi (·, x0,ν ; x0,ν , w) satisfy (3.32) and (3.33), namely, it holds
xi (δt, x0,ν ; x0,ν , w) = x0i ∈ Sli0 , ∀i ∈ N ,
and
|ki,pri (l) (t, xi (t, x0,ν ; x0,ν , w), xj (t, x0,ν ; x0,ν , w); xi0 , wi )| ≤ vmax , ∀t ∈ [0, δt], i ∈ N .
Hence, we deduce by virtue of (3.41) that xi (·, x0 ; x0 , w), i ∈ N also satisfy (3.32) and (3.33),
which establishes the desired result for the general case. The proof is now complete.
4. Design of the Hybrid Control Laws
Consider again system (3.1). We want to determine sufficient conditions which guarantee that
the space-time discretization S −δt is well posed. According to Definition 3.15, establishment of
a well posed discretization is based on the design of appropriate feedback laws which guarantee
well posed transitions for all agents and their possible cell configurations. We postpone the
derivation of acceptable dmax and δt that fulfill well posedness for the next section and proceed
by defining the control laws that are exploited in order to derive well posed discretizations.
Consider a cell decomposition S = {Sl }l∈I of Rn , a time step δt and a selection of a reference
point xl,G for each cell, satisfying (3.24). For each agent i and m-cell configuration li of i, we
define the family of feedback laws ki,li : [0, ∞) × R(Ni +1)n → Rn parameterized by xi0 ∈ Sli
and wi ∈ W as
ki,li (t, xi , xj ; xi0 , wi ) := ki,li ,1 (t, xi , xj ) + ki,li ,2 (xi0 ) + ki,li ,3 (t; wi ),
17
(4.1)
where
W := B(vmax ) ⊂ Rn ,
(4.2)
and
ki,li ,1 (t, xi , xj ) := fi (χi (t), χj (t)) − fi (xi , xj ),
1
ki,li ,2 (xi0 ) := (xli ,G − xi0 ),
δt
ki,li ,3 (t; wi ) := ζ(t)wi ,
t ∈ [0, ∞), (xi , xj ) ∈ R
(Ni +1)n
(4.3)
(4.4)
(4.5)
, xi0 ∈ Sli , wi ∈ W,
ζi : R≥0 → [λ, λ̄], 0 ≤ λ ≤ λ̄ < 1.
¯
¯
(4.6)
The functions χi (·) and in χj (·) in (4.3) are defined for all t ≥ 0 through the solution of the
initial value problem of Definition 3.4, with xl` ,G satisfying (3.24) for all ` ∈ N̄im . In particular,
χi (·) constitutes a reference trajectory, whose endpoint agent i should reach at time δt, when
the agent’s initial condition lies in Sli and the feedback ki,li (·) is applied with wi = 0 in (4.5).
We also note that the feedback laws ki,li (·) depend on the cell of agent i and specifically on its
m-cell configuration li , through the reference point xli ,G in (4.4) and the trajectories χi (·) and
χj (·) in (4.3), as provided by the initial value problem of Definition 3.4. The parameters λ and λ̄
¯
in (4.6) stand for the minimum and maximum portion of the free input, respectively, that can be
further exploited for motion planning. In particular, for each wi ∈ W in (4.2), the vector ζi (t)wi
provides the “velocity” of a motion that we superpose to the reference trajectory χi (·) of agent
i at time t ∈ [0, δt]. The latter allows the agent to reach all points inside a ball with center the
Rt
position of the reference trajectory at time δt by following the curve x̄i (t) := xi (t)+wi 0 ζi (s)ds,
as depicted in Fig. 6 below. Recall that the reference trajectory of i, starting from xli ,G , is
obtained from the initial value problem of Definition 3.4, by considering the unforced solution
of the subsystem formed by agent’s i m-neighbor set. Then, the feedback term ki,li ,1 (·) enforces
the agent to move in parallel to its reference trajectory. Also, by selecting a vector wi in W
and a function ζi (·) as in (4.6), we can exploit the extra terms ki,li ,2 (·) and ki,li ,3 (·) to navigate
the agent to the point x inside the ball B(χi (δt); ri ) (the dashed circle in Figure 1) at time δt
from any initial state xi0 ∈ Sli . In a similar way, it is possible to reach any point inside this
ball by a different selection of wi . This ball has radius
Z
ri :=
0
δt
ζi (s)dsvmax ≥ λδtvmax ,
¯
(4.7)
namely, the distance that the agent can cross in time δt by exploiting ki,li ,3 (·), which corresponds
to the part of the free input that is available for planning. Hence, it is possible to perform a
well posed transition to any cell which has a nonempty intersection with B(χi (δt); ri ) (the cyan
cells in Fig. 6). Notice that due to the assumption vmax < M below (3.5), it is in principle not
possible to cancel the interconnection terms. An example of a cooperative controller motivating
this assumption can be found in our recent paper [5], where appropriate coupling terms ensure
robust connectivity of the multi-agent network for bounded free inputs.
18
x̄i (t)
xi (t)
χi (δt)
Sli
xli ,G
B(χi (δt); ri )
xi (δt) = x
xi0
χi (δt) + twi
Sli0
Figure 6. Consider any point x inside the ball with center χi (δt). Then, for
each initial condition xi0 in the cell Sli , the endpoint of agent’s i trajectory
xi (·) coincides with the endpoint of the curve x̄i (·), which is precisely x, and
lies in Sli0 , namely, xi (δt) = x̄i (δt) = x ∈ Sli0 .
In order to verify the Consistency Condition for the derivation of well posed discretizations,
we will select the function β(·) in Definition 3.14 as
β(t) :=
dmax (δt − t)
+ λ̄vmax t, t ∈ [0, δt]
2δt
(4.8)
Furthermore, we select a constant c̄ ∈ (0, 1), which provides for each agent a measure of the
deviation between each reference trajectory of its neighbors and its estimate by the agent (for
corresponding consistent cell configurations). By defining
L2
t̄ := sup t > 0 : eL2 t − L2 + c̄ √ 2
t−1<0 ,
(4.9)
L1 Nmax
it follows that
0 < t̄ < t∗ ,
(4.10)
∗
where t is defined in (3.16), and that the function Hm (·) in (3.20) satisfies
Hm (t) ≤ c̄m−1 M t, ∀t ∈ [0, t̄]
(4.11)
The proof of both (4.10) and (4.11) is based on elementary calculations and is provided in the
Appendix. Hence, it follows from (4.11) and the result of Proposition 3.10, namely, (3.19), that
if we select the functions αi (·) in Definition 3.13 as αi (·) ≡ α(·), ∀i ∈ N , with
α(t) := cM t, ∀t ∈ [0, δt];
c := c̄m−1 ,
(4.12)
then the neighbor reference trajectory deviation bound (3.23) is satisfied for all 0 < δt ≤ t̄.
Remark 4.1. It follows from (4.12) that for a fixed constant c̄ ∈ (0, 1), the neighbor reference
trajectory deviation bound decreases exponentially with respect to the degree of decentralization.
5. Well Posed Space-Time Discretizations with Motion Planning Capabilities
In this section, we exploit the controllers introduced in Section 4 to provide sufficient conditions for well posed space-time discretizations. By exploiting the result of Proposition 3.18 this
framework can be applied for motion planning, by specifying different transition possibilities
for each agent through modifying its controller appropriately. As in the previous section, we
consider the system (3.1), a cell decomposition S = {Sl }l∈I of Rn , a time step δt and a selection of reference points xl,G , l ∈ I as in (3.24). For each agent i ∈ N and m-cell configuration
19
m
li ∈ I N̄i of i, consider the family of feedback laws given in (4.3), (4.4), (4.5), and parameterized by xi0 ∈ Sli and wi ∈ W . As in the previous sections, χi (·) is the reference solution
of the initial value problem corresponding to the m-cell configuration of i as in Definition 3.4.
Recall that the parameters λ and λ̄ in (4.6) provide the lower and upper portion of the free
¯
input that is exploited for planning. Thus, they can be regarded as a measure for the degree
of control freedom that is chosen for the abstraction. We proceed by providing the desired
sufficient conditions for a space-time discretization to be suitable for motion planning.
Theorem 5.1. Consider a cell decomposition S of Rn with diameter dmax , a time step δt, the
constant ri defined in (4.7) and the parameters λ, λ̄ in (4.6). We assume that dmax and δt
¯
satisfy the following restrictions:
(1 − λ)vmax
¯
0, min t̄, √
L1 Nmax (cM + λ̄vmax ) + λL2 vmax
¯
2(1 − λ)vmax δt
√¯
∈ 0, min
, 2(1 − λ)vmax δt
¯
1 + (L1 Nmax + L2 )δt
oi
p
−2(L1 Nmax (cM + λ̄vmax ) + λL2 vmax )δt2 ,
¯
δt ∈
dmax
(5.1)
(5.2)
with L1 , L2 , M , vmax , c and t̄ as given in (3.3), (3.4), (3.2), (3.5), (4.12) and (4.9), respectively.
Then, for each agent i ∈ N and cell configuration li of i we have Posti (li ; li ) 6= ∅, namely, the
space-time discretization is well posed for the multi-agent system (3.1). In particular, it holds
Posti (li ; li ) ⊃ {l ∈ I : Sl ∩ B(χi (δt); ri ) 6= ∅},
(5.3)
where ri is defined in (4.7) with
ζi (t) := λ.
¯
(5.4)
Proof. In order to prove that the discretization is well posed, we will specify, according to
Definition 3.15(ii), continuous functions β(·) and αi (·), i ∈ N , satisfying (3.25) and (3.23),
m
respectively, in such a way that (5.13) holds for all i ∈ N and li ∈ I N̄i . We pick β(·) as in
(4.8) and αi (·) ≡ α(·), for all i ∈ N , with α(·) as given in (4.12). Notice first that β(·) satisfies
(3.25). In addition, due to the requirement that δt ≤ t̄ in (5.1) and the discussion below (4.12),
the functions αi (·) satisfy the reference trajectory deviation bound (3.23). Thus, it follows that
the requirements of a well posed discretization on the mapping β(·) and αi (·) are fulfilled.
m
Next, let i ∈ N and li ∈ I N̄i . For the derivation of (5.13), namely, that each transition
l
i
li −→
li0 , with Sli0 ∩ B(χi (δt); ri ) 6= ∅ is well posed, it suffices to show that for each x ∈
l
i
B(χi (δt); ri ) and li0 ∈ I such that x ∈ Sli0 , the transition li −→
li0 is well posed. Thus, for each
0
x ∈ B(χi (δt); ri ) and li ∈ I such that x ∈ Sli0 , we need according to Definition 3.15(i) to find
a feedback law (3.26) satisfying Property (P) and a vector wi ∈ W , in such a way that the
Consistency Condition is fulfilled.
Let x ∈ B(χi (δt); ri ) and define
x − χ(δt)
.
(5.5)
λδt
¯
ri
with λ as in (4.6). Then, it follows from (4.7) and (5.4) that |wi | ≤ λδt
≤ vmax and hence, by
¯
¯
virtue of (4.2) that wi ∈ W . We now select the feedback law ki,li (·) as given by (4.1), (4.3),
(4.4), (4.5) and with wi as defined in (5.5), and we will show that for all li0 ∈ I such that
x ∈ Sli0 the Consistency Condition is satisfied. Notice first that ki,li (·) satisfies Property (P).
wi :=
20
In order to show the Consistency Condition we pick xi0 ∈ int(Sli ), x0i := x with x as selected
above and prove that the solution xi (·) of (3.22) with vi = ki,li (t, xi , dj ; xi0 , wi ) satisfies (3.28),
(3.29) and (3.30), for any continuous dj1 , . . . , djNi : R≥0 → Rn that satisfy (3.27). We break
the subsequent proof in the following steps.
STEP 1: Proof of (3.28) and (3.29). By taking into account (3.22), (4.1), (4.3)-(4.5) and (5.4)
we obtain for any continuous dj1 , . . . , djNi : R≥0 → Rn the solution xi (·) of (3.22) with vi = ki,li
as
Z
t
(fi (xi (s), dj (s)) + ki,li (s, xi (s), dj (s); xi0 , wi ))ds
Z t
1
fi (χi (s), χj (s))ds + (xli ,G − xi,0 ) + λwi ds
= xi0 +
¯
δt
0
t
δt − t
= xi0 + χi (t) − xli ,G + (xli ,G − xi,0 ) + tλwi = χi (t) +
(xi0 − xli ,G ) + tλwi , t ≥ 0
¯
¯
δt
δt
xi (t) = xi0 +
0
Hence, from the fact that xi0 ∈ int(Sli ), we deduce from (3.24) that
|xi (t) − χi (t)| ≤
δt − t
(δt − t)dmax
(xi0 − xli ,G )| + tλwi <
+ tλvmax , ∀t ∈ [0, δt],
¯
¯
δt
2δt
(5.6)
which by virtue of (4.8) and (4.6) establishes validity of (3.28). Furthermore, we get that
xi (δt) = χi (δt) + δtλwi = x = x0i and thus, (3.29) also holds.
¯
STEP 2: Estimation of bounds on ki,li ,1 (·), ki,li ,2 (·) and ki,li ,3 (·) along the solution xi (·) of
(3.22) with vi = ki,li and dj1 , . . . , djNi satisfying (3.27). Pick any continuous disturbances
[i]
dj1 , . . . , djNi that satisfy |djκ (t) − χjκ (t)| ≤ αi (t) + β(t), for all t ∈ [0, δt]. We first show that
dmax (δt − t)
+ (cM + λ̄vmax )t
2δt
(δt − t)dmax
+ L2
+ λvmax t , ∀t ∈ [0, δt].
¯
2δt
|ki,li ,1 (t, xi (t), dj (t))| ≤ L1
p
Nmax
(5.7)
Indeed, notice that by virtue of (4.3), we have
ki,li ,1 (t, xi (t), dj (t)) = [fi (χi (t), χj (t)) − fi (xi (t), χj (t))] + [fi (xi (t), χj (t)) − fi (xi (t), dj (t))].
(5.8)
For the second difference on the right hand side of (5.8), we obtain from (3.3), (3.17), (3.28),
(4.8) and (4.12) that
[i]
[i]
|fi (xi (t), χj (t)) − fi (xi (t), dj (t))| ≤L1 |(dj1 (t) − χj1 (t), . . . , djNi (t) − χjN (t))|
i
! 21
N
i
X
≤L1
(α(t) + β(t))2
κ=1
p
≤L1 Nmax (α(t) + β(t))
p
dmax (δt − t)
≤L1 Nmax
+ (cM + λ̄vmax )t .
2δt
21
For the other difference in (5.8), it follows from (3.4) and the obtained trajectory xi (·) in Step
1, that
|fi (xi (t), χj (t)) − fi (χi (t), χj (t))| ≤L2 χi (t) + tλwi + 1 − δtt (xi0 − xli ,G ) − χi (t)
¯
(δt − t)dmax
≤L2
+ λvmax t .
¯
2δt
Hence, it follows from the evaluated bounds on the differences of the right hand side of (5.8)
that (5.7) holds. Next, by recalling that xli ,G satisfies (3.24), it follows directly from (4.4) that
|ki,li ,2 (xi0 )| =
1
dmax
|xi0 − xli ,G | ≤
, ∀xi0 ∈ Sli .
δt
2δt
(5.9)
Finally, for ki,li ,3 (·) we get from (4.5), (4.2) and (5.4) that
|ki,li ,3 (t; wi )| = |λwi | ≤ λvmax , ∀t ∈ [0, δt], wi ∈ W.
(5.10)
¯
¯
STEP 3: Verification of (3.30). In this step we exploit the bounds obtained in Step 2 in order
to show (3.30) for any dj1 , . . . , djNi satisfying (3.27). By taking into account (4.1), (5.7), (5.9)
and (5.10) we want to prove that
L1
p
dmax (δt − t)
dmax
+ (cM + λ̄vmax )t +
2δt
2δt
(δt − t)dmax
+ λvmax t + λvmax ≤ vmax , ∀t ∈ [0, δt].
+L2
¯
¯
2δt
Nmax
(5.11)
Due to the linearity of the left hand side of (5.11) with respect to t, it suffices to verify it for
t = 0 and t = δt. For t = 0 we obtain that
p
dmax
dmax
dmax
L1 Nmax
+
+ L2
+ λvmax ≤ vmax ⇐⇒
¯
2
2δt
2
p
L1 Nmax δtdmax + dmax + L2 δtdmax ≤ 2(1 − λ)vmax δt
¯
whose validity is guaranteed by (5.2). For the case where t = δt, we have
p
dmax
+ L2 λvmax δt + λvmax ≤ vmax ⇐⇒
L1 Nmax (cM + λ̄vmax )δt +
¯
¯
p 2δt
dmax + 2(L2 λvmax δt2 + L1 Nmax (cM + λ̄vmax )δt2 ) ≤ 2(1 − λ)vmax δt
¯
¯
which also holds because of (5.2). Hence, we deduce that (3.30) is also fulfilled and the proof
is complete.
Remark 5.2. In Fig. 7 we depict the bounds on the feedback term ki,li in order to provide
some additional intuition behind the selection of dmax and δt. By rearranging terms in (5.11)
we get
p
p
dmax
dmax
+
(L1 Nmax + L2 )(δt − t) + L1 Nmax (cM + λ̄vmax ) + L2 λvmax )t
¯
2δt
2δt
≤ (1 − λ)vmax , ∀t ∈ [0, δt].
(5.12)
¯
√
The third term of the sum in (5.12) is maximized at δt, namely, becomes L1 Nmax (cM +
λ̄vmax ) + L2 λvmax )δt and is independent of the decomposition diameter dmax . This term is
¯
(1−λ)v
responsible for the requirement that δt is upper bounded by L √N (cM +¯λ̄vmax )+λL v
in
1
max
max
2 max
¯
(5.1), since it cannot exceed (1 − λ)vmax . Hence, assuming that we have selected a time
¯
22
step δt which satisfies this constraint, we can exploit the remaining part of the input for the
manipulation of the first two terms in (5.12). Notice that both these terms are proportional
to dmax which implies that there will always exist a sufficiently small selection of dmax which
guarantees consistency with the bound on the available control. In particular, dmax can be
selected between zero and the largest value that will result in the full exploitation of the control
with magnitude (1 − λ)vmax , either at time 0 or at time δt as depicted in Fig. 7 right and
¯
left, respectively. The later holds because of the linearity of the right hand side of (5.12) with
respect to t and provides the two possible upper bounds on dmax through the min operator in
(5.2).
(1 − λ)vmax
¯
(1 − λ)vmax
¯
√
dmax
2δt (L1 Nmax +L2 )(δt−t)
√
dmax
2δt (L1 Nmax +L2 )(δt−t)
dmax
2δt
dmax
2δt
(L1
√
Nmax (cM +λ̄vmax )
+L2 λvmax )t
¯
(L1
√
Nmax (cM +λ̄vmax )
+L2 λvmax )t
¯
δt
δt
Figure 7. Illustration of the worst case bounds on the feedback terms as given
in (5.12).
Once a cell decomposition diameter dmax and a time step δt which satisfy the restrictions
(5.1) and (5.2) of Theorem 5.1 have been selected, it is possible to improve the reachability
properties of the abstraction by exploiting the remaining available part of the control. This
is established through a modification of the function ζi (·) which was selected as a constant in
(5.4). The corresponding result is provided in Corollary 5.3 below.
Corollary 5.3. Assume that the hypotheses of Theorem 5.1 are satisfied and let any dmax and
δt that satisfy (5.1) and (5.2). Then, for each agent i ∈ N and cell configuration li of i it holds
Posti (li ; li ) ⊃ {l ∈ I : Sl ∩ B(χi (δt); ri ) 6= ∅},
(5.13)
where ri is defined in (4.7) with
ζi (t) :=λ + ξi (t), t ∈ [0, δt],
¯
AL (δt − t) + AR
i t
ξi (t) := min λ̄ − λ,
¯
δtvmax (1 + t)
23
(5.14)
(5.15)
and the constants AL , AR
i ≥ 0 are given as
p
dmax
dmax
dmax
AL :=(1 − λ)vmax − L1 Nmax
−
− L2
,
¯
2
2δt
2
√
max
− L2 λvmax δt − L1 Nmax λ̄vmax δt,
(1 − λ)vmax − d2δt
¯
m+1
m+1 ¯
if
N
=
∅
and
N
= ∅ for all ` ∈√Ni ,
i
`
AR
:=
i
dmax
(1 − λ)vmax − 2δt − L2 λvmax δt − L1 Nmax (cM + λ̄vmax )δt,
¯
¯
otherwise.
2(1−λ)v
(5.16)
(5.17)
δt
max
Proof. Notice first that due to (5.2), namely that dmax ≤ 1+(L √N
and dmax ≤ 2(1 −
¯
1
max +L2 )δt
√
λ)vmax δt − 2(L1 Nmax (cM + λ̄vmax ) + λL2 vmax )δt2 , it follows that AL ≥ 0 and AR
i ≥ 0 for all
¯
¯
i ∈ N , respectively. The rest of the proof follows the same arguments with those of the proof
of Theorem 5.1 which certain modifications that are provided below. First, the parameter wi
in (5.5) is chosen as
x − χi (δt)
.
wi = R δt
ζi (s)ds
0
In addition, when for certain i ∈ N it holds Nim+1 = ∅ and N`m+1 = ∅ for all ` ∈ Ni , it
follows from Lemma 3.7 that for corresponding consistent cell configurations li and l` it holds
[i]
[`]
χ` (t) = χ` (t), ∀t ≥ 0. Thus, we select the neighbor reference trajectory deviation bound
αi (·) ≡ 0 in that case and as in (4.12) otherwise.
Due to the above modifications, the bounds on ki,li ,1 (·) and ki,li ,3 (·) are now given as
√
(δt−t)
max
+ λ̄vmax )t) + L2 ( (δt−t)d
+ (λ + ξi (t))vmax t),
|ki,li ,1 (t, xi (t), dj (t))| ≤ L1 Nmax ( dmax2δt
2δt
¯
m+1
m+1
∀t ∈ [0, δt] when Ni
= ∅ and N`
= ∅ for all ` ∈ Ni , or |ki,li ,1 (t, xi (t), dj (t))| ≤
√
(δt−t)
max
L1 Nmax ( dmax2δt
+ (cM + λ̄vmax )t) + L2 ( (δt−t)d
+ (λ + ξi (t))vmax t), ∀t ∈ [0, δt] oth2δt
¯
erwise, and |ki,li ,3 (t, wi )| ≤ (λ + ξi (t))vmax ,∀t ∈ [0, δt], wi ∈ W , respectively. Hence, instead of
¯
(5.11), we need to verify that
p
dmax
dmax (δt − t)
L1 Nmax
+ (cM + λ̄vmax )t +
2δt
2δt
(δt − t)dmax
+ (λ + ξ(t))vmax t + (λ + ξ(t))vmax ≤ vmax , ∀t ∈ [0, δt],
(5.18)
+L2
¯
¯
2δt
when Nim+1 = ∅ and N`m+1 = ∅ for all ` ∈ Ni , and
p
dmax (δt − t)
dmax
L1 Nmax
+ λ̄vmax t +
2δt
2δt
(δt − t)dmax
+L2
+ (λ + ξ(t))vmax t + (λ + ξ(t))vmax ≤ vmax , ∀t ∈ [0, δt],
¯
¯
2δt
(5.19)
otherwise. By taking into account (5.14), (5.15), (5.16) and (5.17), it follows that both (5.18)
and (5.19) are fulfilled and the proof is complete.
Remark 5.4. (i) The exploitation of the remaining available control in Corollary 5.3, after
a specific space-time discretization has been selected, is illustrated in Fig. 8. Specifically,
Theorem 5.1 provides for each agent i sufficient conditions on the space-time discretization,
which guarantee that the agent can reach a ball of radius λvmax δt, given that its neighbors
¯
may exploit the larger part λ̄vmax of their free input for reachability purposes. Since the same
reasoning is applied for all the agents, once a discretization has been chosen, the reachable ball
of each agent can be increased by exploiting the remaining part (λ̄ − λ)vmax of the free input,
¯
24
as long the latter does not violate consistency with the total magnitude of the feedback laws
(to be bounded by vmax ). Thus, the lengths of the left and right bases of the upper trapezoid
in Fig. 8(left) are AL vmax and AR
i vmax , respectively, and determine the linearly dependent on
t available control over [0, δt], as depicted in Fig. 8(bottom right). Based on this available
control, ζ(·) is then selected under the conditions that its effect in the control modification does
not exceed this threshold, and that it is bounded by λ̄vmax , as illustrated in Fig. 8(up right).
(ii) Assume that heterogeneous Lipschitz constants L1 (i), L2 (i) and M (i) are given for each
agent i ∈ N , which are upper bounded by L1 , L2 and M , respectively, as the latter are
considered in the report. Then the result of Corollary 5.3 remains valid with the constants L1 ,
L2 , M and Nmax in (5.16) and (5.17) being replaced by L1 (i), L2 (i), M (i) and Ni , respectively.
Hence, it is possible to obtain improved reachability properties for agents with weaker couplings
in their dynamics.
ζi (0)vmax
λ̄vmax
vmax
(1 − λ)vmax
¯
λvmax
¯
ξi (t)vmax
λvmax
¯
δt
√
dmax
2δt (L1 Nmax +L2 )
×(δt−t)
dmax
2δt
(L1
AL vmax
√
Nmax (cM +λ̄vmax )
+L2 λvmax )t
√
¯
or (L1 Nmax λ̄vmax +L2 λvmax )t
¯
AR
i vmax
δt
δt
Figure 8. Illustration of the bounds on the feedback terms and the exploitation of the remaining available control through the modification of ζi in (5.14)(5.17).
We next provide an improved version of Theorem 5.1, for the case where the conditions of
Lemma 3.7 are satisfied for all agents, namely, when for every possible m-cell configuration of
each agent, the estimate of its neighbors’ reference trajectories and their reference trajectories
corresponding to consistent configurations coincide.
25
Theorem 5.5. Assume that Nim+1 = ∅ holds for all i ∈ N . Then, the result of Theorem 5.1
remains valid for any δt and dmax satisfying
(1 − λ)vmax
¯
(5.20)
δt ∈ 0, √
L1 Nmax λ̄vmax + λL2 vmax
¯
2(1 − λ)vmax δt
√¯
dmax ∈ 0, min
, 2(1 − λ)vmax δt
¯
1 + (L1 Nmax + L2 )δt
oi
p
−2(L1 Nmax λ̄vmax + λL2 vmax )δt2 .
(5.21)
¯
Proof. By exploiting Lemma 3.7 and the fact that according to the hypothesis of the theorem
it holds Nim+1 = ∅, N`m+1 = ∅ for each pair of agents i ∈ N , ` ∈ Ni , it follows that for
[i]
[`]
corresponding consistent cell configurations li and l` it holds χ` (t) = χ` (t), ∀t ≥ 0. Thus, we
can select for each agent i ∈ N the neighbor reference trajectory deviation bound αi (·) ≡ 0.
The rest of the proof follows the same arguments employed for the proof of Theorem 5.1 and
is therefore omitted.
As in Corollary 5.3, the following corollary presents the analogous improved reachability
result when dmax and δt are selected by Theorem 5.5.
Corollary 5.6. Assume that the hypotheses of Theorem 5.5 are fulfilled and let any δt and
R
R
dmax that satisfy (5.20) and (5.21). Then,
√ the result of Corollary 5.3 is valid with Ai = A :=
dmax
(1 − λ)vmax − 2δt − L2 λvmax δt − L1 Nmax λ̄vmax δt.
¯
¯
6. Example and Simulation Results
As an illustrative example we consider a system of four agents in R2 . Their dynamics are
given as
ẋ1 = satρ (x2 − x1 ) + v1 ,
ẋ2 = v2
ẋ3 = satρ (x2 − x3 ) + v3
ẋ4 = satρ (x3 − x4 ) + v4
where the saturation function satρ : R2 → R2 is defined as
x,
if x < ρ
, x ∈ R2 .
satρ (x) :=
ρ
x,
if
x≥ρ
|x|
The agents’ neighbors’ sets in this example are N1 = {2}, N2 = ∅, N3 = {2}, N4 = {3} and
specify the corresponding network topology. The constant ρ > 0 above represents a bound on
the distance between agents 1, 2, and agents 2, 3, that we will require the system to satisfy
during its evolution. It is not hard to show that this is possible if we select vmax := ρ2 . Hence
we obtain the dynamics bounds and Lipschitz constants M := ρ, L1 := 1 and L2 := 1 for all
agents. By selecting the degree of decentralization m = 2, it follows that the conditions of
Theorem 5.5 are satisfied. By choosing the constant λ̄ = 1 in (4.6) we obtain from (5.21) and
(5.20) (notice also that Nmax = 1) that
0 < δt <
26
1−λ
¯
1+λ
¯
and
0 < dmax ≤ min
By picking δt =
1−λ
2(1+¯λ)
¯
(1 − λ)δt
2
¯ ρ, ((1 − λ)δt − (1 + λ)δt )ρ .
¯
¯
1 + 2δt
we obtain that
(1 − λ)2
¯ ρ.
4(1 + λ)
¯
For the simulation results we pick λ = 0.4, ρ = 10 and evaluate the corresponding value of
¯
δt above and the maximum acceptable value for the diameter dmax of the cells. We partition
the workspace [−10, 10] × [−10, 10] ⊂ R2 into squares and select the initial conditions of the
agents 1, 2, 3 and 4 at (9, 4), (4, 4), (−6, 6) and (−9, −4), respectively, which respect the desired
initial bound on the distance between agents 1, 2, and agents 2, 3. We assume that agent 2,
which is unaffected by the coupled constraints has constant velocity v2 = (−1, −4) and study
reachability properties of the system over the time interval [0, 2]. The sampled trajectory of
agent 2 is visualized with the circles in the figure below, and the blue, magenta cells indicate the
union of agent’s 3 and 1 reachable cells over the time interval, respectively, given the trajectory
of 2. Finally, given the trajectory of 2 and selecting the discrete trajectory of 3 which is depicted
with the red cells in the figure, we obtain with yellow the corresponding reachable cells of agent
4. The simulation results have been implemented in MATLAB with a running time of the order
of a few seconds, on a PC with an Intel(R) Core(TM) i7-4600U CPU @ 2.10GHz processor.
0 < dmax ≤
Figure 9. Reachable cells of agents 1, 3 and 4, given the trajectory of agent
2 in green.
27
7. Conclusions
We have provided an abstraction framework for multi-agent systems which is based on a
varying degree of decentralization for the information of each agent with respect to the graph
topology of the network. Well posed transitions have been modeled by exploiting a system
with disturbances in order to model the possible evolution of the agents neighbors. Sufficient
conditions on the space and time discretization have been quantified in order to capture the
reachability properties of the symbolic models through well defined transitions. The latter
are realized by means of hybrid feedback control laws which take into account the coupled
constraints and navigate the agents to their successor cells.
Ongoing work includes the improvement of the agents’ reachability properties, based on
their local dynamics bounds and Lipschitz constants which will enable the exploitation of a
larger part of the free inputs for the transitions to successor cells. In addition we aim at the
formulation of an online abstraction framework for heterogeneous agents with updated choices
of the discretization and planning parameters.
8. Appendix
8.1. Proof of Lemma 3.9.
Proof. For the proof, it suffices to show that (3.18) holds for all agents κ ∈ N̄im ∩ N̄`m , since by
Lemma 3.2(i) we have N̄`m−1 ⊂ N̄im . We distinguish the following cases.
Case (i). Nim+1 6= ∅ and N`m+1 6= ∅. For Case (i) we consider the following subcases for each
agent κ ∈ N̄im ∩ N̄`m .
Case (ia). κ ∈ (N̄im ∩ N̄`m ) ∩ (Nim ∪ N`m ). In this case, it follows from (3.9) that either
[`]
[i]
χκ (·) ≡ xlκ ,G or χκ (·) ≡ xlκ ,G . Without any loss of generality we assume that κ ∈ N̄`m−1 and
[i]
thus, that χκ (·) ≡ xlκ ,G . Then, we get from (3.2), the initial value problem which specifies
[`]
[i]
[`]
χκ (·), as given by (3.8), and the consistency of l` with li , which implies χκ (0) = χκ (0) = xlκ ,G ,
that
Z t
[`]
[i]
[`]
[`]
|χκ (t) − χκ (t)| = |xlκ ,G − χκ (t)| ≤
|fκ (χ[`]
(8.1)
κ (s), χj(κ) (s))|ds ≤ M t, ∀t ≥ 0.
0
Case (ib). κ ∈ (N̄im ∩ N̄`m ) \ (Nim ∪ N`m ). Notice first, that (N̄im ∩ N̄`m ) \ (Nim ∪ N`m ) ⊂
N̄im−1 ∩ N̄`m−1 and thus, for each agent κ ∈ (N̄im ∩ N̄`m ) \ (Nim ∪ N`m ) we have that κ ∈ N̄im−1
and κ ∈ N̄`m−1 . Hence, we obtain from Lemma 3.2(ii) that Nκ ⊂ N̄im and Nκ ⊂ N̄`m ,
respectively, implying that Nκ ⊂ N̄im ∩ N̄`m . Consequently, it follows from Definition 3.4 that
[i]
[`]
both χj(κ) (·) and χj(κ) (·) are well defined. In order to show (3.18) for all κ of Case (ib) we will
prove the following claim.
Claim I. There exists a δ ∈ (0, t∗ ), such that (3.18) holds for all t ∈ [0, δ] and all κ of Case (ib).
In order to show Claim I, we select
δ := min
1
ln 2
√
,
4L1 Nmax L2
(8.2)
[i]
[`]
and any κ ∈ (N̄im ∩ N̄`m ) \ (Nim ∪ N`m ). By recalling that χj(κ) (·) and χj(κ) (·) are well defined
and exploiting the dynamics bound (3.2), (3.17), (8.2), and that l` is consistent with li , which
28
[`]
[i]
by virtue of the fact that Nκ ⊂ N̄im ∩ N̄`m implies that χν (0) = χν (0) = xlν ,G for all ν ∈ Nκ ,
we deduce that
! 21
! 12
X
X
[i]
[`]
2
[i]
[`]
2
(2M t)
≤
|χν (t) − χν (t)|
|χj(κ) (t) − χj(κ) (t)| =
ν∈Nκ
ν∈Nκ
p
M
≤ 2M Nmax t ≤
, ∀t ∈ [0, δ].
2L1
(8.3)
Next, by virtue of (3.3) and (3.4) we evaluate the difference
Z t
[i]
[`]
[`]
|χκ[i] (t) − χ[`]
(t)|
≤
|fκ (χ[i]
κ
κ (s), χj(κ) (s)) − fκ (χκ (s), χj(κ) (s))|ds
0
Z t
[i]
[i]
[`]
≤
(|fκ (χ[i]
κ (s), χj(κ) (s)) − fκ (χκ (s), χj(κ) (s))|
0
[i]
[`]
[`]
+ |fκ (χ[`]
κ (s), χj(κ) (s)) − fκ (χκ (s), χj(κ) (s))|)ds
Z t
[i]
[`]
≤
(L2 |χκ[i] (s) − χ[`]
κ (s)| + L1 |χj(κ) (s) − χj(κ) (s)|)ds.
(8.4)
0
Hence, from (3.18) and (8.3) we obtain that for all t ∈ [0, δ] it holds
Z t
Mt
[`]
[i]
[`]
+
L2 |χ[i]
|χκ (t) − χκ (t)| ≤
κ (s) − χκ (s)|ds.
2
0
(8.5)
[`]
[i]
In order to derive a bound for |χκ (·) − χκ (·)| from the integral inequality above, we will use
the following version of the Gronwall Lemma, whose proof is given later in the Appendix.
Fact I. Let λ : [a, b] → R be a continuously differentiable function satisfying λ(a) = 0 and µ a
nonnegative constant. If a continuous function y(·) satisfies
Z t
y(t) ≤ λ(t) +
µy(s)ds,
a
then, on the same interval it holds
Z
y(t) ≤
t
eµ(t−s) λ̇(s)ds.
/
a
By exploiting Fact I, we obtain from (8.5) and (8.2) that
Z t
M
M L2 t
M L2 lnL 2
2 t = M t, ∀t ∈ [0, δ].
|χκ[i] (t) − χ[`]
(t)|
≤
eL2 (t−s) ds ≤
e t≤
e
κ
2
2
2
0
(8.6)
Hence, we have proved Claim I. /
We proceed by showing that (3.18) also holds for all t ∈ [0, t∗ ] and κ of Case (ib). Assume on
the contrary that there exist κ0 ∈ (N̄im ∩ N̄`m ) \ (Nim ∪ N`m ) and T ∈ (0, t∗ ] such that
[i]
[`]
|χκ0 (T ) − χκ0 (T )| > M T.
(8.7)
[i]
By continuity of the functions χκ (·), we can define
m
m
m
m
τ := max{t̄ ∈ [0, T ] : |χκ[i] (t) − χ[`]
κ (t)| ≤ M t, ∀t ∈ [0, t̄], κ ∈ (N̄i ∩ N̄` ) \ (Ni ∪ N` )}. (8.8)
29
[i]
[`]
Then, it follows from (8.6), (8.7) and continuity of the functions χκ (·), χκ (·), that there exists
κ00 ∈ (N̄im ∩ N̄`m ) \ (Nim ∪ N`m ) such that
[i]
[`]
|χκ00 (τ ) − χκ00 (τ )| = M τ.
(8.9)
Also, from Claim I and (8.8) we get that
0 < τ < t∗ .
(8.10)
Next, by recalling that since κ00 ∈ (N̄im ∩ N̄`m ) \ (Nim ∪ N`m ), it holds Nκ00 ⊂ N̄im ∩ N̄`m ,
it follows that for each neighbor ν ∈ Nκ00 of κ00 either ν ∈ (N̄im ∩ N̄`m ) ∩ (Nim ∪ N`m ), or
ν ∈ (N̄im ∩ N̄`m ) \ (Nim ∪ N`m ). Thus, we deduce from (8.1) and (8.8), respectively, that
|χν[i] (t) − χ[`]
ν (t)| ≤ M t, ∀t ∈ [0, τ ], ν ∈ Nκ00 .
(8.11)
It then follows from (8.11) (by similar calculations as in (8.4)) that
Z τ
[i]
[`]
[`]
[`]
[i]
[i]
(L2 |χκ00 (s) − χκ00 (s)| + L1 |χj(κ00 ) (s) − χj(κ00 ) (s)|)ds
|χκ00 (τ ) − χκ00 (τ )| ≤
0
12
Z τ
X
[i]
[`]
[`]
2
=
|χ[i]
ds
(L2 |χκ00 (s) − χκ00 (s)| + L1
ν (s) − χν (s)|
0
Z
≤
0
ν∈Nκ00
τ
[i]
Z
[`]
L2 |χκ00 (s) − χκ00 (s)|ds +
τ
L1 M
p
Nκ00 sds.
0
Hence, we obtain from Fact I and (3.17) that
Z τ
p
[i]
[`]
|χκ00 (τ ) − χκ00 (τ )| ≤
eL2 (τ −s) L1 M Nκ00 sds
Z0 τ
p
≤
eL2 (τ −s) L1 M Nmax sds
0
L2 τ
L1 p
1
e
=
M Nmax
−τ −
.
L2
L2
L2
It can then be checked by elementary calculations that
L2
eL2 t − L2 + √ 2
t − 1 < 0, ∀t ∈ (0, t∗ ),
L1 Nmax
(8.12)
(8.13)
with t∗ as specified in (3.16). Also, from (8.9) and (8.12) we have that
L2 τ
L1 p
e
1
Mτ ≤
M Nmax
−τ −
⇐⇒
L2
L2
L2
L2
√2
τ ≤ eL2 τ − L2 τ − 1 ⇐⇒
L1 Nmax
L22
L2 τ
e
− L2 + √
τ − 1 ≥ 0,
L1 Nmax
which contradicts (8.13), since by (8.10), it holds 0 < τ < t∗ . Thus, we have shown (3.18) for
Case (ib).
Case (ii). Nim+1 6= ∅ and N`m+1 = ∅. For Case (ii) we also consider the following subcases for
each agent κ ∈ N̄im ∩ N̄`m .
30
[i]
Case (iia). κ ∈ (N̄im ∩ N̄`m ) ∩ Nim . In this case, it follows from (3.9) that χκ (·) ≡ xlκ ,G and
thus, by using similar arguments with Case (ia) that (8.1) is fulfilled.
Case (iib). κ ∈ (N̄im ∩ N̄`m ) \ Nim . Notice that (N̄im ∩ N̄`m ) \ Nim ⊂ N̄im−1 ∩ N̄`m and thus, for
each agent κ ∈ (N̄im ∩ N̄`m ) \ Nim we have that κ ∈ N̄im−1 and κ ∈ N̄`m . Hence, we obtain from
Lemma 3.2(ii) and the fact that N`m+1 = ∅, that Nκ ⊂ N̄im and Nκ ⊂ N̄`m+1 = N̄`m ∪ N`m+1 =
N̄`m , respectively, implying that Nκ ⊂ N̄im ∩ N̄`m . The remaining proof for this case follows
similar arguments with the proof of Case (ib) and is omitted.
Case (iii). Nim+1 = ∅ and N`m+1 6= ∅. We consider again the following subcases for each agent
κ ∈ N̄im ∩ N̄`m .
[`]
Case (iiia). κ ∈ (N̄im ∩ N̄`m ) ∩ N`m . In this case, it follows from (3.9) that χκ (·) ≡ xlκ ,G and
thus, by using again similar arguments with Case (ia) that (8.1) is fulfilled.
Case (iiib). κ ∈ (N̄im ∩ N̄`m ) \ N`m . Notice that (N̄im ∩ N̄`m ) \ N`m ⊂ N̄im ∩ N̄`m−1 and thus,
for each agent κ ∈ (N̄im ∩ N̄`m ) \ N`m we have that κ ∈ N̄im and κ ∈ N̄`m−1 . Hence, we obtain
from Lemma 3.2(ii) and the fact that Nim+1 = ∅, that Nκ ⊂ N̄im+1 = N̄im ∪ Nim+1 = N̄im and
Nκ ⊂ N̄`m , respectively, implying that Nκ ⊂ N̄im ∩ N̄`m . The remaining proof for Case (iiib)
follows again similar arguments with the proof of Case (ib) and is omitted.
Case (iv). Nim+1 = ∅ and N`m+1 = ∅. In this case the result follows from the proof Lemma 3.7,
[`]
[i]
which implies that the trajectories χκ (·) and χκ (·) coincide for all κ ∈ N̄im ∩ N̄`m . The proof
is now complete.
8.2. Proof of Fact I. We provide the proof of Fact I.
Proof. Indeed, from the classical version of the Gronwall Lemma (see for instance [17]) we have
that
Z t
y(t) ≤ λ(t) +
λ(s)µeµ(t−s) ds
a
Z t
d
= λ(t) +
λ(s) (−eµ(t−s) )ds
ds
a
Z t
= λ(t) − λ(t) + λ(a)eµ(t−a) +
eµ(t−s) λ̇(s)ds
a
Z t
=
eµ(t−s) λ̇(s)ds.
a
8.3. Evaluation of explicit expressions for the functions Hm (·), m ≥ 2 in (3.20). For
m = 2 and m = 3 the functions H2 (·) and H3 (·) are given as
L2 t
e
1
L1 p
Nmax M
−t−
L2
L2
L2
p
2
L1
2 L2 t
2
L2 t
H3 (t) =
Nmax M e t −
e
+t+
.
L2
L2
L2
H2 (t) =
31
For each m ≥ 4, it follows by induction that
p
m−1
m−2
m−2
m−2−j
X
L1
t
t
Hm (t) =
Nmax
M eL2 t Lm−3
+
(−1)j (j + 1)L2m−3−j
2
L2
(m − 2)!
(m
−
2
−
j)!
j=1
m−1
+(−1)m−1 t +
.
L2
8.4.
(4.10)
and (4.11) in Section 4. In order to show (4.10), let g(t) := eL2 t −
Proof of
L2
L22
t − 1, whose derivative ġ(t) = L2 eL2 t − L2 − c̄ L √N2
satisfies ġ(0) =
L2 + c̄ L √N
1
max
1
max
L2
< 0. Hence, since g(0) = 0 and limt→∞ g(t) = ∞, it holds t̄ ∈ R>0 from (4.9).
−c̄ L √N2
1
max
Next, by recalling that t∗ is the unique solution of (3.16), we obtain that
∗
L2
t∗ − 1
eL2 t − L2 + c̄ √ 2
L1 Nmax
∗
L2
(1 − c̄)L22 ∗
(1 − c̄)L22 ∗
√
= eL2 t − L2 + √ 2
t∗ − 1 + √
t =
t > 0,
L1 Nmax
L1 Nmax
L1 Nmax
which by virtue of (4.9) implies (4.10). We next show (4.11) by induction on κ ∈ {1, . . . , m}.
Notice that by (3.20), (4.11) holds for κ = 1. Assuming that it is valid for certain κ ∈ {1, . . . , m}
we show that also holds for κ = κ + 1. Hence, by exploiting (3.20), (4.9) and (4.11) with m = κ
we obtain
Z t
p
Hκ+1 (t) − c̄κ M t =
eL2 (t−s) L1 Nmax Hκ (s)ds − c̄κ M t
0
Z t
p
c̄t
≤ c̄κ−1 L1 Nmax
eL2 (t−s) sds − √
L1 Nmax
0
L2 t
p
1
e
1
c̄t
= c̄κ−1 L1 Nmax
−t−
− √
L2
L2
L2
L1 Nmax
2
p
c̄L
t − 1 ≤ 0, ∀t ∈ [0, t̄],
= c̄κ−1 L1 Nmax L22 eL2 t − L2 + √ 2
L1 Nmax
which establishes (4.11) with m = k + 1.
9. Acknowledgements
This work was supported by the H2020 ERC Starting Grant BUCOPHSYS, the Knut and
Alice Wallenberg Foundation, the Swedish Foundation for Strategic Research (SSF), and the
Swedish Research Council (VR).
References
[1] A. Abate, A. Tiwari, and S. Sastry. Box invariance in biologically-inspired dynamical systems. 7(45):1601–
1610, 2009.
[2] M. Andreasson, D. V. Dimarogonas, H. Sandberg, and K. H. Johansson. Distributed control of networked
dynamical systems: Static feedback, integral action and consensus. IEEE Transactions on Automatic Control, 59(7):1750–1764, 2014.
[3] D. Boskos and D. V. Dimarogonas. Abstractions of varying decentralization degree for coupled multi-agent
systems. accepted at the 55th IEEE Conference on Decision and Control.
[4] D. Boskos and D. V. Dimarogonas. Decentralized abstractions for feedback interconnected multi-agent
systems. in Proceedings of the 54th IEEE Conference on Decision and Control, pages 282–287, 2015.
32
[5] D. Boskos and D. V. Dimarogonas. Robust connectivity analysis for multi-agent systems. in Proceedings of
the 54th IEEE Conference on Decision and Control, pages 6767–6772, 2015.
[6] M. E. Brouke and M. Gannes. Reach control on simplices by piecewise affine feedback. SIAM Journal on
Control and Optimization, 5(52):3261–3286, 2014.
[7] P. E. Caines and Y. J. Wei. The hierarchical lattices of a finite machine. Systems and Control Letters,
(25):257–263, 1995.
[8] Y. Chen, X. C. Ding, A. Stefanescu, and C. Belta. Formal approach to the deployment of distributed robotic
teams. IEEE Transactions on Robotics, 28(1):158–171, 2012.
[9] E. Dallal and P. Tabuada. On compositional symbolic controller synthesis inspired by small-gain theorems.
in Proceedings of the 54th IEEE Conference on Decision and Control, pages 6133–6138, 2015.
[10] A. Girard and S. Martin. Synthesis for constrained nonlinear systems using hybridization and robust controllers on simplices. IEEE Transactions on Automatic Control, 57(4):1046–1051, 2012.
[11] A. Girard, G. Pola, and P. Tabuada. Approximately bisimilar symbolic models for incrementally stable
switched systems. IEEE Transactions on Automatic Control, 55(4):116–126, 2010.
[12] E. A. Gol, X. Ding, M. Lazar, and C. Belta. Approximately bisimilar symbolic models for incrementally
stable switched systems. IEEE Transactions on Automatic Control, 59(12):3122–3134, 2014.
[13] L. Grüne. Asymptotic Behavior of Dynamical and Control Systems under Perturbation and Discretization,
volume 1783. Springer-Verlag, Berlin, Germany, 2002.
[14] L. C. Habets and J. H. van Schuppen. Control of piecewise-linear hybrid systems on simplices and rectangles.
In International Workshop on Hybrid Systems: Computation and Control, pages 261–274. Springer, 2001.
[15] M. K. Helwa and P. E. Caines. In-block controllability of affine systems on polytopes. in Proceedings of the
53rd IEEE Conference on Decision and Control, pages 3936–3942, 2014.
[16] M. K. Helwa and P. E. Caines. Epsilon controllability of nonlinear systems on polytopes. in Proceedings of
the 54th IEEE Conference on Decision and Control, pages 252–257, 2015.
[17] H. Khalil. Nonlinear systems. Pentice Hall, 2002.
[18] S. G. Loizou and K. J. Kyriakopoulos. Automatic synthesis of multi-agent motion tasks based on ltl specifications. in Proceedings of the 43rd IEEE Conference on Decision and Control, pages 153–158, 2004.
[19] M. Mesbahi and M. Egerstedt. Graph Theoretic Methods for Multiagent Networks. Princeton University
Press, 2010.
[20] P. Meyer. Invariance and symbolic control of cooperative systems for temperature regulation in intelligent
buildings. PhD thesis, Université Grenoble-Alpes, 2015.
[21] P. Meyer, A. Girard, and E. Witrant. Safety control with performance guarantees of cooperative systems
using compositional abstractions. in Proceedings of the 5th IFAC Conference on Analysis and Design of
Hybrid Systems, pages 317–322, 2015.
[22] G. Pola, A. Girard, and P. Tabuada. Approximately bisimilar symbolic models for nonlinear control systems.
Automatica, 44(10):2508–2516, 2008.
[23] G. Pola, P. Pepe, and M. D. di Benedetto. Symbolic models for networks of control systems. IEEE Transactions on Automatic Control, TAC.2016.2528046, 2011.
[24] G. Pola, P. Pepe, and M. D. di Benedetto. Symbolic models for networks of discrete-time nonlinear control
systems. in Proceedings of the American Control Conference, pages 1787–1792, 2014.
[25] G. Reissig. Computing abstractions of nonlinear systems. IEEE Transactions on Automatic Control,
56(11):2583–2598, 2011.
[26] M. Rungger and M. Zamani. Compositional construction of approximate abstractions. in Proceedings of
the 18th International Conference on Hybrid Systems: Computation and Control, pages 68–77, 2015.
[27] Y. Tazaki and J.-i. Imura. Bisimilar finite abstractions of interconnected systems. In International Workshop
on Hybrid Systems: Computation and Control, pages 514–527. Springer, 2008.
[28] M. Zamani, G. Pola, M. Mazo, and P. Tabuada. Symbolic models for nonlinear control systems without
stability assumptions. IEEE Transactions on Automatic Control, 57(7):1804–1809, 2012.
Department of Automatic Control, School of Electrical Engineering, KTH Royal Institute of
Technology, Osquldas väg 10, 10044, Stockholm, Sweden
E-mail address: [email protected]
Department of Automatic Control, School of Electrical Engineering, KTH Royal Institute of
Technology, Osquldas väg 10, 10044, Stockholm, Sweden
E-mail address: [email protected]
33
| 3cs.SY
|
arXiv:1706.08351v1 [math.GR] 26 Jun 2017
The automorphism group of a nonsplit metacyclic
2-group
Haimiao Chen
∗
Mathematics, Beijing Technology and Business University, Beijing, China
Abstract
We determine the structure of automorphism group or each nonsplit metacyclic 2-group.
This completes the work on automorphism groups of metacyclic p-groups.
Keywords: automorphism group, nonsplit, metacyclic 2-group.
MSC 2010: 20D45.
1
Introduction
A metacyclic group is an extension of a cyclic group by another cyclic group. As is well-known
(see Section 3.7 of [8]), each metacyclic group can be presented as
H(n, m; t, r) = hα, β | αn = 1, β m = αt , βαβ −1 = αr i.
(1)
for some positive integers n, m, r, t with r m − 1 ≡ t(r − 1) ≡ 0 (mod n).
In recent years, people have been studying automorphism groups of metacyclic p-groups.
The order and the structure of Aut(G) when G is a split metacyclic p-group were found in [1,4];
when G is a nonsplit metacyclic p-group with p 6= 2, Aut(G) was determined in [5].
In this paper, we deal with nonsplit metacyclic 2-groups, based on [3] in which we derived
explicit formulas for automorphisms of a general metacyclic group. According to Theorem 3.2
of [7], each nonsplit metacyclic 2-group is isomorphic to one of the following:
(I) HI (a, b, c, d) := H(2a , 2b ; 2c , 2d + 1) for a unique quadruple (a, b, c, d) with
d>1
and
max{d, a − d − 1} < c < min{a, b},
(2)
(II) HII (a, b, e) := H(2a , 2b ; 2a−1 , 2e − 1) for a unique triple (a, b, e) with
max{1, a − b} < e < min{a, b},
(III) HIII (a) := H(2a , 2; 2a−1 , 2a − 1) for a unique a > 1.
∗
Email: [email protected]
1
(3)
We do not deal with case (III), in which G is a generalized quaternion group, and the automorphism group was determined in [9]. Denote r = 2d +1 in case (I), and r = 2e −1, c = a−1, d = 1
in case (II). Denote H(2a , 2b ; 2c , r) simply as H whenever there is no ambiguity.
Notation 1.1. For an integer N > 0, let ZN denote Z/N Z, and regard it as a quotient ring
of Z. Let Z×
N denote the multiplicative group of units. For u ∈ Z, denote its image under the
quotient Z ։ ZN also by u.
For integers u, s > 0, let [u; s] = 1 + s + · · · + su−1 and [0; s] = 0.
For any k, v ∈ Z, by r k we mean r ǩ , where ǩ ∈ {0, 1, . . . , 2b − 1} denotes the remainder
when dividing k by 2b ; by [v; r k ] we mean [v̌; r ǩ ]. For w ∈ Z2a , by αw we mean αw̃ where w̃ is
any integer mapped to w by Z ։ Z2a ; similarly, β w for w ∈ Z2b makes sense.
For an integer u 6= 0, let kuk denote the largest integer s with 2s | u; set k0k = +∞.
Use expα (u) to denote αu when the expression for u is too long.
Throughout the paper, we abbreviate A ≡ B (mod 2a ) to A ≡ B as often as possible.
2
Preparation
It follows from (1) that
v1
(αu1 β v1 )(αu2 β v2 ) = αu1 +u2 r β v1 +v2 ,
u v k
(4)
u[k;r] vk
(α β ) = α
β .
(5)
For the following two lemmas, see [3] Lemma 2.1–2.4.
Lemma 2.1. If s > 1 with ks − 1k = ℓ > 1 and x > 0 with kxk = u > 0, then
[x; s] ≡ (1 + 2ℓ−1 )x
sx − 1 ≡ (s − 1 + 22ℓ−1 )x
(mod 2ℓ+u ),
(mod 22ℓ+u ).
Lemma 2.2. There exists σ ∈ Aut(H) with σ(α) = αx1 β y1 , σ(β) = αx2 β y2 if and only if
(r y1
2b−d | y1
2c y1
x2 [2b ; r y2 ] + 2c y2 − x1 [2c ; r y1 ] − b 2c ≡ 0,
2
(r − 1)y1 c
y1
y2
− 1)x2 + ([r; r ] − r )x1 +
2 ≡ 0,
2b
2 ∤ x1 y 2 − x2 y 1 .
(6)
(7)
(8)
(9)
Due to (6), we may write y1 = 2b−d y. The following two lemmas are special cases of [3]
Lemma 2.8 (iii) and (ii), respectively. Here we reprove them in a relatively succinct way.
Lemma 2.3. When H = HI (a, b, c, d), the conditions (7) and (8) hold if and only if
x1 ≡ 1 + 2b−c x2
c−d
y2 ≡ 1 + (2
(mod 2a−c ),
c−1
+2
2
)y
(10)
a−d
(mod 2
).
(11)
Proof. Now that ky1 k ≥ b − d > c − d ≥ a − 2d, applying Lemma 2.1, we obtain r y1 ≡ 1 + 2b y,
[r]ry1 ≡ r, [2c ]ry1 ≡ 2c , [2b ]ry2 ≡ 2b , hence (7) and (8) become
2b x2 + 2c (y2 − x1 − 2c−d y) ≡ 0,
(12)
2b x2 y + (r − r y2 )x1 + 2c y ≡ 0,
(13)
respectively. It follows from (13) that
d + ky2 − 1k = kr y2 −1 − 1k ≥ c ≥ a − d;
(14)
by Lemma 2.1, r y2 −1 − 1 ≡ 2d (1 + 2d−1 )(y2 − 1), hence
r y2 − r = (1 + 2d )(r y2 −1 − 1) ≡ (2d + 22d−1 )(y2 − 1).
So (13) can be converted into (1 + 2d−1 )(y2 − 1)x1 ≡ (2b−d x2 + 2c−d )y (mod 2a−d ); multiplying
1 − 2d−1 and using (14), we obtain
x1 (y2 − 1) ≡ (2b−d x2 + 2c−d − 2c−1 )y
(mod 2a−d ).
Then
(x1 + 2c−d y)(y2 − 1 − 2c−d y) = x1 (y2 − 1) + 2c−d (y2 − x1 − 2c−d y − 1)y
≡ (2b−d x2 − 2c−1 + 2c−d (y2 − x1 − 2c−d y))y
≡ −2c−1 y
(mod 2a−d )
(mod 2a−d ),
where in the last line, (12) is used. Thanks to 2 ∤ x1 + 2c−d y and c ≥ a − d, we can deduce
(11), then (10) follows from (11), (12) and the condition c − 1 ≥ d ≥ a − c.
Conversely, it can be verified that (10), (11) indeed imply (12), (13).
Lemma 2.4. When H = HII (a, b, e), the conditions (7) and (8) hold if and only if one of the
following occurs:
• 2 | y and y2 ≡ 1 (mod 2a−e );
• e ≤ a − 2, 2 ∤ y and ky2 − 1k = a − e − 1.
Proof. For any z with kzk = u ≥ 1, applying Lemma 2.1 to s = r 2 , x = z, we obtain
[2z; r] = (1 + r)[z; r 2 ] ≡ 2e (1 + 2e )z ≡ 2e z
r 2z − 1 = (r 2 )z − 1 ≡ (r 2 − 1 + 22e+1 )z
(mod 2e+u+1 ),
(mod 22e+2+u );
(15)
these are still true when r is replaced by r y2 , as 2 ∤ y2 so that kr y2 ± 1k = kr ± 1k. In particular,
[2b ; r y2 ] ≡ 2e+b−1 ≡ 0,
r y1 − 1 ≡ 0.
(16)
Note that 22c−b y1 ≡ 2c (y2 − x1 ) ≡ 0, hence (7) holds for free. The condition (8) becomes
− 1 ≡ 2a−1 y, which is, by (15), equivalent to
r y2 −1
e + ky2 − 1k ≥ a
if
2 | y,
e + ky2 − 1k = a − 1 if
by (9), the second possibility occurs only when e ≤ a − 2.
3
2 ∤ y;
3
The structure of automorphism group
3.1
H = HI (a, b, c, d)
Let Ξ denote the set of (x1 , x2 ) ∈ Z2a × Z2a satisfying (10), and let Ω denote the set of
(y, y2 ) ∈ Z2a+d−c × Z2a+b−c satisfying (11). By Lemma 2.1 and (2), we have
b−d y
r2
b−d y
= (1 + 2d )2
≡ 1 + 2b y.
Thus each automorphism of HI (a, b, c, d) can be expressed as
σx1 ,x2 ;y,y2 : αu β v 7→ expα (x1 [u; 1 + 2b y] + (1 + 2b yu)x2 [v; r y2 ]) expβ (2b−d yu + y2 v)
(17)
for some quadruple (x1 , x2 , y, y2 ) ∈ Ξ × Ω, and σx1 ,x2 ;y,y2 = σx′1 ,x′2 ;y′ ,y2′ if and only if
y′ ≡ y
(mod 2d ),
y2′ ≡ y2
x′1 − x1 + 2c−d (y ′ − y) ≡ 0,
(18)
y2′
(19)
x′2 − x2 +
(mod 2b ),
− y2 c
2 ≡ 0.
2b
Let
φx1 ,x2 = σx1 ,x2 ;0,1 ,
ψy,y2 = σ1,0;y,y2 .
(20)
The following can be verified using (4), (5):
φx′1 ,x′2 ◦ φx1 ,x2 = φx′1 x1 ,x′1 x2 +x′2 ,
(21)
ψy′ ,y2′ ◦ ψy,y2 = ψy′ +y2′ y,y2′ y2 ,
(22)
φx1 ,x2 ◦ ψy,y2 = σx1 +x2 [2b−d y;r],x2[y2 ;r];y,y2 ,
(23)
ψy,y2 ◦ φx1 ,x2 = σ[x1 ;r′ ],[x2;r′ ];yx1,2b−d yx2 +y2 ,
r ′ = 1 + 2b y.
with
(24)
Let
X = {φx1 ,x2 : (x1 , x2 ) ∈ Ξ},
Y = {ψy1 ,y2 : (y, y2 ) ∈ Ω}.
(25)
Then X, Y are subgroups of Aut(H), and there is a decomposition
σx1 ,x2 ;y,y2 = φx̃1 ,x̃2 ◦ ψy,y2 ,
with
b−d
x̃1 = x1 − x2 [2
(26)
−1
y; r][y2 ; r]
,
−1
x̃2 = x2 [y2 ; r]
.
(27)
As special cases of (18) and (19), φx1 ,x2 = ψy,y2 if and only if
kyk ≥ d,
ky2 − 1k ≥ b,
x1 ≡ 1 + 2c−d y,
x2 ≡
y2 − 1 c
2.
2b
(28)
It follows that
X ∩ Y = hφ1+2c ,0 , φ1,2c i = hψ2d ,1 , ψ0,1+2b i ∼
= Z2a−c × Z2a−c .
4
(29)
Let
(
−1,
z=
0,
f = min{a, b},
φ0 = φ1,2a−f ,
φ1 = φ1−2f −c ,z ,
b < a,
b ≥ a,
ψ0 = ψ2a−c ,1 ,
w = −(1 + 2d−1 )−1 ,
(30)
ψ1 = ψw,1−2c−d .
(31)
Consider the homomorphism ρ : X → Z2a sending φx1 ,x2 to x1 . It is easy to see that
ker(ρ) = {φ1,x2 : kx2 k ≥ a − f } = hφ0 i ∼
= Z2f ,
Im(ρ) = {x : kx − 1k ≥ f − c} ≤
Z×
2a .
(32)
(33)
If f ≥ c + 2, then Im(ρ) = h1 − 2f −c i ∼
= Z2a+c−f , hence
f
a+c−f
X = hφ0 , φ1 | φ20 , φ21
f −c −1
2
, φ1 φ0 φ−1
1 φ0
i∼
= Z2f ⋊ Z2a+c−f ;
(34)
∼
if f = c + 1, then Im(ρ) = Z×
2a = h−1, 5i = Z2 × Z2a−2 by Theorem 2’ on Page 43 of [6]. Hence
f
a−2
X = hφ0 , φ1 , φ2 | φ20 , φ21 , φ22
−5
∼
, (φ1 φ0 )2 , φ2 φ0 φ−1
2 φ0 , [φ1 , φ2 ]i = Z2f ⋊ (Z2 × Z2a−2 ),
with
φ2 = φ5,−2z .
(35)
(36)
Consider the homomorphism µ : Y → Z2b+a−c sending ψy1 ,y2 to y2 . Clearly
ker(µ) = {ψy,1 : y ∈ Z2a+d−c : kyk ≥ a − c} = hψ0 i ∼
= Z2d ,
Im(µ) = {y2 : ky2 − 1k ≥ c − d} ≤
.
Z×
2a+b−c
(37)
(38)
If c ≥ d + 2, then Im(µ) = h1 − 2c−d i ∼
= Z2a+b+d−2c , hence
d
a+b−d−2c
Y = hψ0 , ψ1 | ψ02 , ψ12
c−d −1
, ψ1 ψ0 ψ1−1 φ20
i∼
= Z2d ⋊ Z2a+b+d−2c ;
(39)
= h−1, 5i ∼
if c = d + 1, then Im(µ) = Z×
= Z2 × Z2a+b−c−2 , hence
2b+a−c
d
a+b−c−2
Y = hψ0 , ψ1 , ψ2 | ψ02 , ψ12 , ψ22
, (ψ1 ψ0 )2 , ψ2 ψ0 ψ2−1 ψ0−5 , [ψ1 , ψ2 ]i ∼
= Z2d ⋊ (Z2 × Z2a+b−c−2 ),
(40)
with
ψ2 = ψ−2w,5 .
The above can be summarized as
Theorem 3.1. For H = HI (a, b, c, d), we have Aut(H) = XY such that
(
Z2f ⋊ Z2a+c−f ,
f ≥ c + 2,
X∼
=
Z2f ⋊ (Z2 × Z2a−2 ), f = c + 1,
(
Z2d ⋊ Z2a+b+d−2c ,
c ≥ d + 2,
Y ∼
=
Z2d ⋊ (Z2 × Z2a+b−c−2 ), c = d + 1,
X ∩Y ∼
= Z2a−c × Z2a−c .
Consequently, the order of Aut(H) is 2b+c+2d .
5
(41)
Remark 3.2. In principle, we are able to obtain a presentation for Aut(H). The generators
are those of X, Y , and the relator set can be divided into RX ∪ RY ∪ RX,Y ∪ R′ , where
• RX (resp. RY ) consists of relators in the presentations for X (resp. Y ),
• RX,Y consists of two elements corresponding to the two generators of X ∩Y , i.e., φ1+2c ,0 =
ψ2d ,1 and φ1,2c = ψ0,1+2b ,
• relators in R′ have the form ϑj θi = γδ, where ϑj (resp. θi ) is a generator for Y (resp.
X), and γ (resp. δ) is a product of generators for X (resp. Y ).
As an example of element in R′ ,
ψ0 ◦ φ0 = σ1,2a−f ;2a−c ,h = φh−1 ,2a−f h−1 ◦ ψ2a−c ,h ,
with
h = 1 + 22a+b−c−d−f ;
one can further write φh−1 ,2a−f h−1 (resp. ψ2a−c ,h ) as a product of generators for X (resp. Y ).
However, the computations are so complicated that we choose not to write down explicitly.
3.2
H = HII (a, b, e)
Recall that r = 2e − 1, c = a − 1 and d = 1.
Let Ξ denote the set of (x1 , x2 ) ∈ Z2a × Z2a with 2 ∤ x1 , and let Ω denote the set of
(y, y2 ) ∈ Z4 × Z2b+1 satisfying the conditions in Lemma 2.4. Each automorphism of HII (a, b, e)
can be expressed as (recalling r y1 ≡ 1, as in (16))
σx1 ,x2 ;y,y2 : αu β v 7→ expα (x1 u + x2 [v; r y2 ]) expβ (2b−1 yu + y2 v)
(42)
for some quadruple (x1 , x2 , y, y2 ) ∈ Ξ × Ω, and σx1 ,x2 ;y,y2 = σx′1 ,x′2 ;y′ ,y2′ if and only if
y′ ≡ y
y2′ ≡ y2
(mod 2),
x′1 − x1 + 2a−2 (y ′ − y) ≡ 0,
(mod 2b ),
x′2 − x2 +
y2′
− y2 a−1
2
≡ 0.
2b
(43)
(44)
Let
φx1 ,x2 = σx1 ,x2 ;0,1 ,
ψy,y2 = σ1,0;y,y2 .
(45)
We have
φx′1 ,x′2 ◦ φx1 ,x2 = φx′1 x1 ,x′1 x2 +x′2 ,
(46)
ψy′ ,y2′ ◦ ψy,y2 = ψy′ +y2′ y,y2′ y2 ,
(47)
φx1 ,x2 ◦ ψy,y2 = σx1 +x2 [2b−1 y;r],x2[y2 ;r];y,y2 ,
(48)
ψy,y2 ◦ φx1 ,x2 = σx1 ,x2 ;yx1 ,2b−1 yx2 +y2 .
(49)
Let
X = {φx1 ,x2 : (x1 , x2 ) ∈ Ξ},
Y = {ψy,y2 : (y, y2 ) ∈ Ω}.
6
(50)
Thus X, Y are subgroups of Aut(H), and there is a decomposition
σx1 ,x2 ;y,y2 = φx̃1 ,x̃2 ◦ ψy,y2 ,
(51)
x̃1 = x1 − x2 [2b−1 y; r][y2 ; r]−1 ,
with
x̃2 = x2 [y2 ; r]−1 .
(52)
As special cases of (43), (44), φx1 ,x2 = ψy,y2 if and only if
y2 ∈ {1, 1 + 2b },
y ∈ {0, 2},
x1 ≡ 1 + 2a−2 y,
x2 ≡
y2 − 1 a−1
2 .
2b
(53)
Hence
X ∩ Y = {ψ0,1 , ψ0,1+2b , ψ2,1 , ψ2,1+2b } ∼
= Z2 × Z2 .
(54)
Let
φ0 = φ1,1 ,
φ1 = φ−1,0 ,
φ2 = φ5,0 ,
ψ0 = ψ2,1 ,
ψ1 = ψ1,1−2a−e−1 .
(55)
Similarly as in the previous subsection, we can obtain
a
a−2
X = hφ0 , φ1 , φ2 | φ20 , φ21 , φ22
−1 −5
∼
, φ1 φ0 φ−1
1 φ0 , φ2 φ0 φ2 φ0 , [φ1 , φ2 ]i = Z2a ⋊ (Z2 × Z2a−2 ). (56)
If e ≤ a − 3, then
b+e−a+2
Y = hψ0 , ψ1 | ψ02 , ψ12
, [ψ0 , ψ1 ]i ∼
= Z2 × Z2b+e−a+2 .
(57)
If e = a − 2, then
b−1
Y = hψ1 , ψ2 | ψ12 , ψ22
, [[ψ1 , ψ2 ], ψ1 ], [[ψ1 , ψ2 ], ψ2 ], [ψ1 , ψ2 ]2 i,
with
ψ2 = ψ1,5 .
(58)
In the notations of [2] Chapter IV, Section 3, choose a map s : Z2 × Z2b−1 → Y , (u, v) 7→ ψ1u ψ2v ,
so that s(0, 0) = 1, then it can be computed that
s(u1 , v1 )s(u2 , v2 )s(u1 + u2 , v1 + v2 )−1 = [ψ1 , ψ2 ]u2 v1 .
Hence Y is isomorphic to the central extension of Z2 × Z2b−1 by Z2 determined by the 2-cocycle
f : (Z2 × Z2b−1 )2 → Z2 ,
(u1 , v1 , u2 , v2 ) → u2 v1 ;
(59)
, [ψ0 , ψ̃1 ], [ψ0 , ψ̃2 ], [ψ̃1 , ψ̃2 ]i ∼
= Z2 × Z2 × Z2b−1 ,
(60)
denote this group by Ef .
If e = a − 1, then we find directly that
b−1
Y = hψ0 , ψ̃1 , φ̃2 | ψ02 , ψ̃12 , ψ̃22
with
ψ̃1 = ψ0,−1 ,
ψ̃2 = ψ0,5 .
Theorem 3.3. For H = HII (a, b, e), we have Aut(H) = XY such that
X∼
= Z2a ⋊ (Z2 × Z2a−2 ),
e ≤ a − 3,
Z2 × Z2b+e−a+2 ,
Y ∼
e = a − 2,
= Ef ,
Z2 × Z2 × Z2b−1 , e = a − 1,
∼ Z2 × Z2 .
X ∩Y =
Consequently, the order of Aut(H) is 2a+b+min{a−2,e} .
7
(61)
References
[1] J.N.S. Bidwell and M.J. Curran, The automorphism group of a split metacyclic p-group.
Arch. Math. 87 (2006), 488–497.
[2] K.S. Brown, Cohomology of groups. Graduate Texts in Mathematics Vol 87, SpringerVerlag, New York, Heidelberg, Berlin, 1982.
[3] H.-M. Chen, Y.-S. Xiong, Z.-J. Zhu, Automorphism of metacyclic groups. arXiv:1506.02234,
accepted by Czechoslovak J. Math.
[4] M.J. Curran, The automorphism group of a split metacyclic 2-group. Arch. Math. 89 (2007),
10–23.
[5] M.J. Curran, The automorphism group of a nonsplit metacyclic p-group. Arch. Math. 90
(2008), 483–489.
[6] K. Ireland, M. Rosen, A Classical Introduction to Modern Number Theory, Graduate Texts
in Mathematics Vol 84, 2nd edition, Springer-Verlag, New York, 1990.
[7] B.W. King, Presentations of metacyclic groups. Bull. Austral. Math. Soc. 8 (1973), 103–131.
[8] H.J. Zassenhaus, The theory of groups. Second edition, Chelsea, New York, 1956.
[9] D.-G. Zhu and G.-X. Zuo, The Automorphism Group and Holomorph of Quaternion
Group(in generalized sense). Acta Mathematiea Scientia, 25 (2005), no. 1, 79–83.
8
| 4math.GR
|
SUBMITTED TO THE IEEE TRANSACTIONS ON INFORMATION THEORY, DECEMBER 2017
1
Low SNR Asymptotic Rates of Vector Channels
with One-Bit Outputs
Amine Mezghani, Member, IEEE, Josef A. Nossek, Life Fellow, IEEE, A. Lee Swindelhurst Fellow, IEEE
arXiv:1712.05728v1 [cs.IT] 15 Dec 2017
Abstract
We analyze the performance of multiple-input multiple-output (MIMO) links with one-bit output quantization in terms of
achievable rates and characterize their performance loss compared to unquantized systems for general channel statistical models
and general channel state information (CSI) at the receiver. One-bit ADCs are particularly suitable for large-scale millimeter
wave MIMO Communications (massive MIMO) to reduce the hardware complexity. In such applications, the signal-to-noise ratio
per antenna is rather low due to the propagation loss. Thus, it is crucial to analyze the performance of MIMO systems in this
regime by means of information theoretical methods. Since an exact and general information-theoretic analysis is not possible,
we resort to the derivation of a general asymptotic expression for the mutual information in terms of a second order expansion
around zero SNR. We show that up to second order in the SNR, the mutual information of a system with two-level (sign) output
signals incorporates only a power penalty factor of π/2 (1.96 dB) compared to system with infinite resolution for all channels
of practical interest with perfect or statistical CSI. An essential aspect of the derivation is that we do not rely on the common
pseudo-quantization noise model.
Index Terms
Massive MIMO communication, Broadband regime, One-bit Quantization, Mutual information, Optimal input distribution,
Ergodic capacity, Millimeter-wave Communications.
I. I NTRODUCTION
In this paper, we investigate the theoretically achievable rates under one-bit analog-to-digital conversion (ADC) at the receiver
for a wide class of channel models. To this end, we consider general multi-antenna communication channels with coarsely
quantized outputs and general communication scenarios, e.g. correlated fading, full and statistical channel state information
(CSI) at the transmitter and the receiver, etc.. Since exact capacity formulas are intractable in such quantized channels, we
resort to a low signal-to-noise ratio (SNR) approximation and lower bounds on the channel capacity to perform the analysis.
Such mutual information asymptotics can be utilized to evaluate the performance of quantized output channels or design and
optimize the system in practice. Additionally, the low SNR analysis under coarse quantization is useful in the context of large
scale (or massive) multiple-input multiple-output (MIMO) [4], [5], [6] and millimeter-wave (mmwave) communications [7], [8],
[9], [10], [11], considered as key enablers to achieve higher data rates in future wireless networks. In fact, due to high antenna
gains possible with massive MIMO and the significant path-loss at mmwave frequencies, such systems will likely operate at
rather low SNR values at each antenna, while preferably utilizing low cost hardware and low resolution ADCs, in order to
access all available dimensions even at low precision. Our asymptotic analysis demonstrates that the capacity degradation due
to quantized sampling is surprisingly small in the low SNR regime for most cases of practical interest.
A. Less precision for more dimensions: The motivation for coarse quantization
The use of low resolution (e.g., one-bit) ADCs and DACs is a potential approach to significantly reducing cost and
power consumption in massive MIMO wireless transceivers. It was proposed as early as 2006 by [12]-[15] in the context
of conventional MIMO. In the last three years however, the topic has gained significantly increased interest by the research
community [16]-[50] as an attractive low cost solution for large vector channels. In the extreme case, a one-bit ADC consists of
a simple comparator and consumes negligible power. One-bit ADCs do not require an automatic gain control and the complexity
and power consumption of the gain stages required prior to them are substantially reduced [51]. Ultimately, one-bit conversion
is, in view of current CMOS technology, the only conceivable option for a direct mmwave bandpass sampling implementation
close to the antenna, eliminating the need for power intensive radio-frequency (RF) components such as mixers and oscillators.
In addition, the use of one-bit ADCs not only simplifies the interface to the antennas by relaxing the RF requirements but also
A. Mezghani was with the University of California, Irvine, USA, and is now with the University of Texas at Austin, USA. J. A. Nossek is with the
Technical University of Munich Germany, and the Federal University of Ceará, Fortaleza, Brazil. A. L. Swindelhurst is with the Department of EECS,
University of California, Irvine, USA, and is also a Hans Fischer Senior Fellow of the Institute for Advanced Study at the Technical University of Munich.
E-mail: [email protected].
This work was supported by the National Science Foundation (NSF) under ECCS-1547155 and by the German Research Foundation (DFG) under the
priority program SPP1202.
Parts of this work was published in [1], [2], [3], where some proofs were missing, and in particular, the main theorem in Section III was not stated and
the application to massive MIMO was not investigated.
Manuscript received December xx, xxxx; revised January xx, xxxx.
2
SUBMITTED TO THE IEEE TRANSACTIONS ON INFORMATION THEORY, DECEMBER 2017
simplifies the interface between the converters and the digital processing unit (DSP/FPGA). For instance, the use of 10-bit
converters running at 1 Gsps for 100 antennas would require a data pipeline of 1 Tbit/s to the FPGAs and a very complex
and power consuming interconnect. By using only one-bit quantization the speed and complexity are reduced by a factor of
10. Sampling with one-bit ADCs might therefore qualify as a fundamental research area in communication theory.
Even though use of only a single quantization bit, i.e., simply the sign of the sampled signal, is a severe nonlinearity, initial
research has shown that the theoretical “best-case” performance loss that results with a one-bit quantizer is not as significant
as might be expected, at least at low SNRs where mmwave massive MIMO is expected to operate, prior to the beamforming
gain, which can still be fully exploited. This is also very encouraging in the context of low-cost and low-power IoT devices
which will also likely operate in relatively low SNR regimes. Figure 1 shows how the theoretical spectral efficiency versus
energy efficiency (Eb /N0 ) of a one-bit transceiver that uses QPSK symbols in an additive white Gaussian noise (AWGN)
channel compares with that of an infinite-precision ADC using a Gaussian input, i.e. the Shannon limit log2 (1 + SNR). In
fact, the capacity of the one-bit output AWGN channel is achieved by QPSK signals and reads as [52], [21]
√
C1−bit = 2 1 − Hb (Φ( SNR)) ,
(1)
where we make use of the binary entropy function Hb (p) = −p · log2 p − (1 − p) · log2 (1 − p) and the cumulative Gaussian
distribution Φ(z). Surprisingly, at low SNR the loss due to one-bit quantization is approximately equal to only π/2 (1.96dB)
[52], [53] and actually decreases to roughly 1.75dB at a spectral efficiency of about 1 bit per complex dimension, which
corresponds to the spectral efficiency of today’s 3G systems.
3
1−bit AWGN (QPSK input)
∞−bit AWGN (Gaussian input)
Spectral Efficiency C
2.5
2
1.5
≈ 1.75 dB
1
0.5
PSfrag replacements
0
−2
≈ 1.96 dB
−1
0
1
2
3
4
5
6
Energy Efficiency Eb /N0 = SNR/C in dB
Fig. 1.
Spectral efficiency versus energy efficiency for One- and Infinite-Bit Quantization in AWGN channels.
Even if a system is physically equipped with higher resolution converters and high performance RF-chains, situations may
arise where the processing of desired signals must be performed at much lower resolution, due for instance to the presence
of a strong interferer or a jammer with greater dynamic range than the signals of interest. In fact, after subtracting or zeroforcing the strong interferer, the residual effective number of bits available for the processing of other signals of interest
is reduced substantially. Since future wireless systems must operate reliably even under severe conditions in safety-critical
applications such as autonomous driving, investigating communication theory and signal processing under coarse quantization
of the observations is crucial.
B. Related Work and Contributions
Many contributions have studied MIMO channels operating in Rayleigh fading environments in the unquantized (infinite
resolution) case, for both the low SNR [53]-[58] and high SNR [59] regimes. Such asymptotic analyses are very useful since
characterizing the achievable rate for the whole SNR regime is in general intractable. This issue becomes even more difficult
in the context of one-bit quantization at the receiver side, apart from very special cases. In the works [15], [1], the effects of
quantization were studied from an information theoretic point of view for MIMO systems, where the channel is perfectly known
at the receiver. These works demonstrated that the loss in channel capacity due to coarse quantization is surprisingly small
at low to moderate SNR. In [2], [3], the block fading single-input single-output (SISO) non-coherent channel was studied in
detail. The work of [27] provided a general capacity lower bound for quantized MIMO and general bit resolutions that can be
applied for several channel models with perfect CSI, particularly with correlated noise. The achievable capacity for the AWGN
channel with output quantization has been extensively investigated in [20], [21], and the optimal input distribution was shown to
MEZGHANI, NOSSEK and SWINDLEHURST: LOW SNR ASYMPTOTIC RATES OF VECTOR CHANNELS WITH SINGLE BIT OUTPUTS
3
be discrete. The authors of [19] studied the one-bit case in the context of an AWGN channel and showed that the capacity loss
can be fully recovered when using asymmetric quantizers. This is however only possible at extremely low SNR, which might
not be useful in practice. In [60], it was shown that, as expected, oversampling can also reduce the quantization loss in the
context of band limited AWGN channels. In [23], [28], non-regular quantizer designs for maximizing the information rate are
studied for intersymbol-interference channels. More recently, [31] studied bounds on the achievable rates of MIMO channels
with one-bit ADCs and perfect channel state information at the transmitter and the receiver, particularly for the multiple-input
single-output (MISO) channel. The recent work of [50] analyzes the sum capacity of the two-user multiple access SISO AWGN
channel, which turn to be achievable with time division and power control.
Motivated by these works, we aim to study and characterize the communication performance of point-to-point MIMO
channels with general assumptions about the channel state information at the receiver taking into account the 1-bit quantization
as a deterministic operation. In particular, we derive asymptotics for the mutual information up to the second order in the
SNR and study the impact of quantization. We show that up to second order in SNR for all channels of practical interest, the
mutual information of a system with two-level (1-bit sign operation) output signals incorporates only a power penalty of π2
(-1.96 dB) compared to a system with infinite resolution. Alternatively, to achieve the same rate with the same power up to
the second order as in the ideal case, the number of one-bit output dimensions has to be increased by a factor of π/2 for the
case of perfect CSI and at least by π 2 /4 for the statistical CSI case, while essentially no increase in the number of transmit
dimensions is required. We also characterize analytically the compensation of the quantization effects by increasing the number
of 1-bit receive dimensions to approach the ideal case.
This paper is organized as follows: Section II describes the system model. Then, Section III provides the main theorem
consisting of a second order asymptotic approximation of the entropy of one-bit quantized vector signals. In Section IV, we
provide a general expression for the mutual information between the inputs and the quantized outputs of the MIMO system with
perfect channel state information, then we expand that into a Taylor series up to the second order in the SNR. In Section V, we
extend these results to elaborate on the asymptotic capacity of 1-bit MIMO systems with statistical channel state information
including Rayleigh flat-fading environments with delay spread and receive antenna correlation.
C. Notation
Vectors and matrices are denoted by lower and upper case italic bold letters. The operators (•)T , (•)H , tr(•), (•)∗ , Re(•)
and Im(•) stand for transpose, Hermitian (conjugate transpose), matrix trace, complex conjugate, real and imaginary parts of a
complex number, respectively. The terms 0M and 1M denote the M -dimensional vectors of all zeros and all ones, respectively,
while I M represents the identity matrix of size M . The vector xi is the i-th column of matrix X and xi,j denotes the (ith,
jth) element, while xi is the i-th element of the vector x. The operator E[•] stands for expectation with respect to all random
variables, while the operator Es|q [•] stands for the expectation with respect to the random variable s given q. In addition,
C x = E[xxH ] − E[x]E[xH ] represents the covariance matrix of x and C xy denotes E[xy H ]. The functions P (s) and P (s|q)
symbolize the joint probability mass function (pmf) and the conditional pmf of s and q, respectively. Additionally, diag(A)
denotes a diagonal matrix containing only the diagonal elements of A and nondiag(A) = A − diag(A). Finally, we represent
element-wise multiplication and the Kronecker product of vectors and matrices by the operators ”◦” and ”⊗”, respectively.
II. S YSTEM M ODEL
We consider a point-to-point quantized MIMO channel with M transmit dimensions (e.g. antennas or, more generally, spatial
and temporal dimensions) and N dimensions at the receiver. Fig. 2 shows the general form of a quantized MIMO system,
where H ∈ CN ×M is the channel matrix, whose distribution is known at the receiver side. The channel realizations are in
general unknown to both the transmitter and receiver, except for the ideal perfect CSI case. The vector x ∈ CM comprises the
M transmitted symbols, assumed to be subjected to an average power constraint E[ x2 ] ≤ PTr . The vector η represents the
additive noise, whose entries are i.i.d. and distributed as CN (0, ση2 ). The quantized channel output r ∈ CN is thus represented
as
r = Q(y) = Q(Hx + η).
(2)
In a one-bit system, the real parts yi,R and the imaginary parts yi,I of the unquantized receive signals yi , 1 ≤ i ≤ N , are each
quantized by a symmetric one-bit quantizer. Thus, the resulting quantized signals read as
+1 if yi,c ≥ 0
ri,c = Q(yi,c ) = sign(yi,c ) =
, for c ∈ {R, I}, 1 ≤ i ≤ N,
(3)
−1 if yi,c < 0
The operator Q(y) will also be denoted as sign(y) and represents the one-bit symmetric scalar quantization process in each
real dimension. The restriction to one-bit symmetric quantization is motivated by its simple implementation. Since all of the
real and imaginary components of the receiver noise η are statistically independent with variance ση2 , we can express each of
4
SUBMITTED TO THE IEEE TRANSACTIONS ON INFORMATION THEORY, DECEMBER 2017
the conditional probabilities as the product of the conditional probabilities on each receiver dimension
P (r = sign(y)|x, H) =
Y
N
Y
Y
N
Y
c∈{R,I} i=1
P (ri,c |x)
r
[Hx]
i,c
i,c
Φ q
=
,
2
ση /2
c∈{R,I} i=1
PSfrag replacements
(4)
G
Rx
t2
x̂ cumulative normal distribution function. We first state the main theorem used throughout
where Φ(x) = √12π −∞ e− 2 dt is the
the paper, and then provide the asymptotics of the mutual information for several channel models up to second order in the
SNR.
H
y
Q(•)
x
√
Fig. 2.
M
SNR
r
N
η
Quantized MIMO System
III. M AIN THEOREM FOR THE ASYMPTOTIC ENTROPY OF ONE - BIT QUANTIZED VECTOR SIGNALS
We provide a theorem that can be used for deriving the second order approximation of the mutual information. It considers
the 1-bit signal r = sign(εx + η), where x is a random vector with a certain distribution and η is random with i.i.d. Gaussian
entries and unit variance, while ε is a signal scaling parameter.
Theorem 1: Assuming x ∈ CN is a proper complex random vector (E[xxT ] = E[x]E[xT ]) satisfying Ex [kxk4+ε
4 ] < δ for
some finite constants ε, δ > 0 and η is i.i.d. Gaussian with unit variance, then the following entropy approximation holds up
to the second order in ε2
2
2
2
2
tr
(nondiag(C
))
H(sign(εx + η)) =2N ln 2 − ε2 kE[x]k2 − ε4
x
π
π2
(5)
4
4
4
1
4
kE[x] ◦ E[x ◦ x ◦ x]k1 + 2 kE[x]k4 + o(ε ),
−
3π
3π
P
p
where kakp = i,c |ai,c |p and [a ◦ b]i = ai,R bi,R + jai,I bi,I , while the expectation is taken with respect to x and C x =
E[xxH ] − E[x]E[xH ] is the covariance matrix of x.
Proof: See appendix A.
From this theorem, we can deduce some useful lemmas.
4+ε
Lemma 1: For any possibly non-deterministic function g(x) satisfying Ex [kg(x)k4 ] < δ and E[g(x)g(x)T ] = E[g(x)]E[g(xT )]
for some finite constants ε, δ > 0, we have to the second order in ε2
2
2
2
H(sign(εg(x) + η)) ≈2N ln 2 − ε2 kE[g(x)]k22 − ε4
tr nondiag(C g(x) ))
2
π
π
4
4
4
1
kE[g(x)] ◦ E[g(x) ◦ g(x) ◦ g(x)]k1 + 2 kE[g(x)]k4 .
−
3π
3π
T
T
Lemma 2: For any function g(x) satisfying Ex [kg(x)k4+ε
4 ] < δ and E[g(x)g(x) |x] = E[x|x]E[g(x)(x) |x], we have
the following second order approximation of the conditional entropy
2
2 2
2
2
4
H(sign(εg(x) + η)|x) ≈2N ln 2 − ε Ex [kE[g(x)|x]k2 ] − ε Ex 2 tr nondiag(C g(x)|x )
π
π
4
4
4
1
−
kE[g(x)|x] ◦ E[g(x) ◦ g(x) ◦ g(x)|x]k1 + 2 kE[g(x)|x]k4 .
3π
3π
Proof: Lemma 1 is a direct result of Theorem 1, where we just replace the random vector x by g(x) if the stated conditions
are fulfilled for g(x). For Lemma 2 we just perform the expectation in Theorem 1 first conditioned on x, to get the entropy
for a given x and then we take the average with respect to x, again if the stated assumptions regarding the distribution of
g(x) are fulfilled. These results will be used to derive a second order approximation of the mutual information of quantized
MIMO systems for the case of perfect as well as statistical channel state information.
MEZGHANI, NOSSEK and SWINDLEHURST: LOW SNR ASYMPTOTIC RATES OF VECTOR CHANNELS WITH SINGLE BIT OUTPUTS
IV. M UTUAL I NFORMATION
AND
C APACITY
WITH FULL
5
CSI
When the channel H is perfectly known at the receiver, the mutual information (in nats/s/Hz) between the channel input
and the quantized output in Fig. 2 reads as [61]
"
#
X
P (r|x, H)
P (r|x, H)ln
,
(6)
I(x, r) = Ex
P (r|H)
r
with P (r|H) = Ex [P (r|x, H)] and Ex [·] is the expectation taken with respect to x. For large N , the computation of the
mutual information has intractable complexity due to the summation over all possible r, except for low dimensional outputs
(see [31] for the single output case), which is not relevant for the massive MIMO case. Therefore, we resort to a low SNR
approximation to perform the analysis on the achievable rates.
A. Second-order Expansion of the Mutual Information with 1-bit Receivers
In this section, we will elaborate on the second-order expansion of the input-output mutual information (6) of the considered
system in Fig. 2 as the signal-to-noise ratio goes to zero.
Theorem 2: Consider the one-bit quantized MIMO system in Fig. 2 under a zero-mean input distribution p(x) with covariance
4+α
matrix C x , satisfying p(x) = p(jx), ∀x ∈ CM (zero-mean proper complex distribution)1 and Ex [kHxk4 ] < γ for some
finite constants α, γ > 0. Then, to the second order, the mutual information (in nats) between the inputs and the quantized
outputs with perfect CSI is given by:
2
1
2
I(x, r) = tr(HC x H H ) 2 − 2 tr((nondiag(HC x H H ))2 )
π
ση
π
1
1
4
(7)
4
+ (1 − )Ex [kHxk4 ] 4 + ∆I(x, r),
3π
π
ση | {z }
o( σ14 )
η
where
4
kak4
is the 4-norm of a taken to the power 4:
P
4
i,c ai,c .
Proof: We start with the definition of the mutual information [61]
I(x, r) = H(r) − H(r|x)
= H(sign(Hx + η)|H) − H(sign(Hx + η)|x, H),
(8)
1
ση
and g(x) = Hx to get the following asymptotic expression:
2
2 1
1
2
2
I(x, r) = 2N ln 2 −
tr (nondiag(C Hx ))
kE[Hx]k
−
2
2
4
2
π ση
ση π
4
4
4
1
kE[Hx] ◦ E[Hx ◦ Hx ◦ Hx]k1 + 2 kE[Hx]k4
−
3π
3π
2
2 1
2
1
2
− 2N ln 2 +
tr
E
[kE[Hx|x]k
]
+
E
nondiag(C
)
x
x
Hx|x
2
π ση2
ση4
π2
1
4
4
4
1
kE[Hx|x] ◦ E[Hx ◦ Hx ◦ Hx|x]k1 + 2 kE[Hx|x]k4 + o( 4 )
−
3π
3π
ση
2
2
1
2 1
2
H
tr
nondiag(HC
H
)
E
k[H(x
−
E[x])]k
−
=
x
2
π ση2
ση4 π 2
4
4
kE[Hx] ◦ E[Hx ◦ Hx ◦ Hx]k11 + 2 E[kHE[x]k44 ]
−
3π
3π
4
4
1
4
4
+ E[kHxk4 ] − 2 E[kHxk4 ] + o( 4 ).
3π
3π
ση
then we use Lemma 1 and Lemma 2 with ε =
In the case that the distribution is zero-mean E[x] = 0, we end up exactly with the formula stated by the theorem. The
4+α
condition Ex [kHxk4 ] < γ for some finite constants α, γ > 0 is necessary, so that the remainder term of the expansion
given by
1
(9)
∆I(x, r) = Ex [o(kHxk44 4 )]
ση
1 This
restriction is simply justified by symmetry considerations.
6
SUBMITTED TO THE IEEE TRANSACTIONS ON INFORMATION THEORY, DECEMBER 2017
satisfies
lim
1
σ4
∆I(x, r)
1
ση4
→0
= 0,
(10)
as already stated by Theorem 1.
For comparison, we use the results of Prelov and Verdú [54] to express the mutual information (in nats) between the input
x and the unquantized output r with the same input distribution as in Theorem 2:
I(x, y) = tr(HC x H H )
1
1 tr((HC x H H )2 ) 1
−
+ o( 4 ).
2
4
ση
2
ση
ση
(11)
While the mutual information for the unquantized channel in (11), up to the second order, depends only on the input covariance
matrix, in the quantized case (7) it also depends on the fourth order statistics of x (the fourth mixed moments of its components).
Now, using (7) and (11), we deduce the mutual information penalty in the low SNR (or large dimension) regime incurred
by quantization
I(x, r)
2
lim
= ,
(12)
1
I(x,
y)
π
→0
σ2
η
which is independent of the channel and the chosen distribution. Since the Gaussian distribution achieves the capacity of the
unquantized channel but not necessarily for the quantized case, we obtain for the supremum of the mutual information, i.e the
capacity
C1-bit
2
≥ .
(13)
lim
1
C
π
→0
∞-bit
σ2
η
These results can be also obtained based on the pseudo-quantization noise model [16], [27] and it generalizes the result known
for the AWGN channel [52].
Fig. 3 illustrates the mutual information for a randomly generated 4×4 channel2 with QPSK signaling and total power
tr(C x ) = PTr = SNR · ση2 , computed exactly using (6), and also its first and second-order approximations from (7). For
comparison, the mutual information without quantization (using i.i.d. Gaussian input) is also plotted. Fig. 3 shows that the
ratio π2 holds for low to moderate SNR = PσTr
2 .
η
For a larger number of antennas, the inner summation in (6) may be intractable. In this case, the second-order approximation
in (7) is advantageous at low SNR to overcome the high complexity of the exact formula.
2
Mutual Information in bpcu
1.8
PSfrag replacements
1.6
Exact formula
First−order approximation
Second−order approximation
Unquantized with i.i.d. Gaussian input
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
SNR (linear)
Fig. 3. Mutual information of a 1-bit quantized 4×4 QPSK MIMO system and its first and second-order approximations. For comparison the mutual
information without quantization is also plotted.
2 The
generated entries hi,j of H are uncorrelated and hi,j ∼ CN (0, 1).
MEZGHANI, NOSSEK and SWINDLEHURST: LOW SNR ASYMPTOTIC RATES OF VECTOR CHANNELS WITH SINGLE BIT OUTPUTS
7
B. Capacity with Independent-Component Inputs
Lacking knowledge of the channel or its statistics, the transmitter assigns the power evenly over the components xi,c of
Tr
the input vector x, i.e., E[x2i,c ] = P2M
, in order to achieve good performance on average. Furthermore, let us assume these
components to be independent
of
each
other (e. g. multi-streaming scenario).3 Thus, the probability density function of the
Q
input vector x is p(x) = i,c pi,c (xi,c ).4 Now, with
X
[Hx]i,R = [hi,j,R xj,R − hi,j,I xj,I ] ,
(14)
j
µj,c
=
E[x4j,c ]
= 4M 2 E[x4j,c ],
E[x2j,c ]2
(15)
and the kurtosis of the random component xj,c defined as
κj,c = µj,c − 3,
we get
Ex [([Hx]i,R )4 ] =
(16)
X
X
1
3
µj,c h4i,j,c
h2i,j,c h2i,j ′ ,c′ +
2
4M
′ ′
j,c
j,c,j ,c
(j,c)6=(j ′ ,c′ )
=
(17)
2 X
1
H
κj,c h4i,j,c .
3
HH
+
i,i
4M 2
j,c
Similar results hold for the other components of the vector Hx. Plugging this result and C x = PMTr I into (7), we obtain an
expression for the mutual information with independent-component inputs and C x = PMTr I up to second order:
"
2
2
H PTr
ind
− 2 tr((nondiag(HH H ))2 )
I (x, r) ≈ tr(HH )
2
π
M ση
π
(18)
2
X
1
2
H 2
4 PTr
κj,c hi,j,c
.
+ (1− ) 3 tr((diag(HH )) ) +
3π
π
M ση2
i,j,c
Now, we state a theorem on the structure of the near-optimal input distribution under these assumptions.
Theorem 3: To second order, QPSK signals are capacity-achieving among all signal distributions with independent components. The achieved capacity up to second order is
"
2
2
H PTr
C1-bit ≈ tr(HH )
tr((nondiag(HH H ))2 )
−
π
M ση2
π2
(19)
2
X
2
1
P
Tr
H 2
+ (1− ) 3 tr((diag(HH )) ) − 2
.
h4i,j,c
3π
π
M ση2
i,j,c
E[x4 ]
Proof: Since E[x4i,c ] ≥ E[x2i,c ]2 , we have κi,c = E[x2i,c]2 − 3 ≥ −2, ∀i, c. Obviously, the QPSK distribution is the unique
i,c
distribution with independent-component inputs that can achieve all these lower bounds simultaneously, i.e., κi,c = κQPSK =
−2 ∀i, c, and thus maximize I ind (x, r) in (18) up to second order.
C. Ergodic Capacity under i.i.d. Rayleigh Fading Conditions
Here we assume the channel H to be ergodic with i.i.d. Gaussian components hi,j ∼ CN (0, 1). The ergodic capacity can
be written as
erg
C1-bit
= EH [C1-bit ].
(20)
3 Clearly,
this is not necessarily the capacity achieving strategy.
that the pi,c (xi,c ) have to be even functions and pi,R (xi,R ) = pi,I (xi,I ) ∀i, due to the symmetry (see Theorem 2) and convexity of the mutual
information.
4 Note
8
SUBMITTED TO THE IEEE TRANSACTIONS ON INFORMATION THEORY, DECEMBER 2017
We apply the expectation over H using the second order expansion of C1-bit in (19). By expanding the following expressions
and taking the expectation over the i.i.d. channel coefficients, we have
(21)
EH tr(HC x H H ) = N tr(C x )
2
H 2
= N (N − 1)tr(C x )
(22)
EH tr (nondiag(HC x H ))
h
i
h
i
3
4
4
=
(23)
EH kHxk4
N Ex kxk2 .
2
The ergodic mutual information over an i.i.d. channel can be obtained as
h
i 2 1 2
2 1
N
4
2
EH [I(x, r)] ≈ N tr(C x )
(N − 1)tr(C x ) + (π − 1)Ex kxk2
−
.
π ση2
2
π ση2
(24)
Next, we characterize the capacity achieving distribution up to second order in the SNR.
Theorem 4: The ergodic capacity of the 1-bit quantized i.i.d. MIMO channel is achieved asymptotically at low SNR by
QPSK signals and reads as
2
2 PTr
N (N + (π − 1)M − 1) 2 PTr
erg
.
(25)
C1-bit ≈ N
−
π ση2
2M
π ση2
i2
h
i h
2
2
4
Proof: Since tr(C 2x )/M ≥ (tr(C x )/M )2 = PTr /M with equality if C x = PMTr I, and Ex kxk2 ≥ Ex kxk2
= PTr
with equality if the input has a constant norm of 1, the ergodic capacity is achieved by a constant-norm uncorrelated input
2
(C x = PMTr I and kxk2 = PTr ), which can be obtained for instance by QPSK signals Finally, using (19) we obtain the
second-order expression for the ergodic capacity of one-bit quantized Rayleigh fading channels for the QPSK case.
Compared to the ergodic capacity in the unquantized case achieved by i.i.d. Gaussian inputs (or even by QPSK up to the
second order) [53]
2
N (N + M ) PTr
PTr
,
(26)
C erg ≈ N · 2 −
ση
2M
ση2
erg
the ergodic capacity of one-bit quantized MIMO under QPSK C1-bit
incorporates a power penalty of almost
considering only the linear term that characterizes the capacity in the limit of infinite bandwidth.
π
2
(1.96 dB), when
On the other hand, the second order term quantifies the convergence of the capacity function to the low SNR limit, i.e. the
first order term, by reducing the power or increasing the bandwidth [53]. Therefore, it can be observed from
1<
N + (π − 1)M − 1
< π − 1,
N +M
(27)
that the quantized channel converges to this limit slower than the unquantized channel. Nevertheless, for M = 1 or M ≪ N
(massive MIMO uplink scenario), this difference in the convergence behavior vanishes almost completely, since both secondorder expansions (25) and (26) become nearly the same up to the factor 2/π in SNR.
erg
In addition, the ergodic capacity of the quantized channel C1-bit in (25) increases linearly with the number of receive antennas
N and only sublinearly with the number of transmit antennas M , which holds also for C erg . For the special case of one receive
erg
antenna, N = 1, C1-bit
does not depend on the number of transmit antennas M up to the second order, contrary to C erg . On the
other hand, if one would achieve, up to the second order, the same ergodic capacity at the same power with one-bit receivers
as in the ideal case by adjusting the number of receive and transmit antennas, i.e.,
2
2
2 PTr
N (N + M ) PTr
N1−bit (N1−bit + (π − 1)M1−bit − 1) 2 PTr
PTr
=
N
−
,
(28)
N· 2 −
1−bit
ση
2M
ση2
π ση2
2M1−bit
π ση2
then we can deduce by equating coefficients that
N1−bit =
π
N,
2
M1−bit = M
πN − 2
.
πN − (π − 2)M
(29)
The one-bit receive dimension has to be increased by π/2, while the behavior for the number of transmit antennas is shown
in Fig. 4. Clearly, when N ≫ M , which corresponds to a typical massive MIMO uplink scenario, we have M1−bit ≈ M .
This means that, at the transmitter (or user) side, there is no need to increase the number of antennas up to the second order
in SNR, showing that the total increase of dimensions is moderate.
MEZGHANI, NOSSEK and SWINDLEHURST: LOW SNR ASYMPTOTIC RATES OF VECTOR CHANNELS WITH SINGLE BIT OUTPUTS
9
1.6
1.5
M1−bit /M
1.4
PSfrag replacements
1.3
1.2
N = 40
N = 30
N = 20
1.1
1
0.9
0
Fig. 4.
5
10
15
M
20
Required M1−bit /M to achieve ideal ergodic capacity up to the second order.
V. M UTUAL I NFORMATION
AND
C APACITY
WITH
S TATISTICAL CSI
AND
1- BIT R ECEIVERS
We reconsider now the extreme case of 1-bit quantized communications over MIMO Rayleigh-fading channels but assuming
that only the statistics of the channel are known at the receiver. Later, we will also treat the achievable rate for the SISO
channel case for the whole SNR range.
Generally, the mutual information (in nats/s/Hz) between the channel input and the quantized output in Fig. 2 with statistical
CSI reads as [61]
I(x, r) = H(r) − H(r|x)
#
"
X
P (r|x)
,
= Ex
P (r|x)ln
P (r)
r
(30)
where P (r) = Ex [P (r|x)] and H(·) and H(·|·) represent the entropy and the conditional entropy, respectively. If the channel
is Gaussian distributed with zero mean, then given the input x, the unquantized output y is zero-mean complex Gaussian with
covariance E[yy H |x] = ση2 · IN + EH [HxxH H H ], and thus we have
p(y|x) =
exp(−y H (ση2 IN + EH [HxxH H H ])−1 y)
π N ση2 IN + EH [HxxH H H ]
.
(31)
Thus, we can express the conditional probability of the quantized output as
P (r|x) =
Z
∞
0
=
Z
0
∞
···
···
Z
∞
0
Z
0
∞
p(y ◦ r|x)dr
exp(−(y ◦ r)H (ση2 · IN + E[HxxH H H ])−1 (y ◦ r))
π N ση2 · IN + E[HxxH H H ]
(32)
dy,
where the integration is performed over the positive orthant of the complex hyperplane.
The evaluation of this multiple integral is in general intractable. Thus, we consider first a simple lower bound involving the
mutual information under perfect channel state information at the receiver, which turns out to be tight in some cases as shown
10
SUBMITTED TO THE IEEE TRANSACTIONS ON INFORMATION THEORY, DECEMBER 2017
later. The lower bound is obtained by the chain rule and the non-negativity of the mutual information:
"
#
X
EH [P (r|x, H)]
P (r|x, H)ln
I(x, r) = Ex,H
Ex,H [P (r|x, H)]
r
= I(H, r) + I(x, r|H) − I(H, r|x)
≥ I(x, r|H) − I(H, r|x)
"
#
X
EH [P (r|x, H)]
= Ex,H
P (r|x, H)ln
Ex [P (r|x, H)]
r
#
"
X
P (r|x)
.
= Ex,H
P (r|x, H)ln
P (r|H)
r
On the other hand, an upper bound is given by the coherent assumption (channel perfectly known at the receiver)
"
#
X
P (r|x, H)
I(x, r) ≤ Ex,H
P (r|x, H) · ln
,
P (r|H)
r
(33)
(34)
where we can express each of the conditional probabilities P (r|x, H) as the product of the conditional probabilities on each
receiver dimension, since the real and imaginary components of the receiver noise η are statistically independent with power
1
2 in each real dimension:
N
Y Y
P (ri,c |x)
P (r|x, H) =
c∈{R,I} i=1
=
Y
N
Y
c∈{R,I} i=1
√
Φ ri,c [Hx]i,c 2SNR ,
(35)
Rx
t2
where Φ(x) = √12π −∞ e− 2 dt is the cumulative normal distribution function. Evaluating the lower bound in (33), even
numerically, is very difficult, except for some simple cases such as SISO block fading channels, as considered next.
A. The non-coherent block-Rayleigh fading SISO Case
Here we treat the block-Rayleigh fading SISO case in more detail, where H = h · IT , h ∼ CN (0, 1), and M = N = T is
the coherence time. For simplicity and ease of notation we assume that ση2 = 1, therefore we have without loss of generality
H
H
SNR = PσTr
is the sum of an identity matrix and a rank one matrix.
2 = PTr . The covariance matrix E[yy |x] = IT + xx
η
Then, we obtain the conditional probability of the 1-bit output as
P (r|x) = Eh [P (r|x, h)]
Z
T
Y
1
−|h|2
=
e
π C
t=1
=
=
1
π
Z
1
2π
e−|h|
C
T
2 Y
t=1
Z
Y
c∈{R,I}
√
√
Φ( 2Re(hxt )rt,R ) · Φ( 2Im(hxt )rt,I )dh
+∞Z +∞ 2 2
− u +v
2
e
−∞ −∞
√
Φ( 2([hxt ]c )rt,c ) · dh
T
Y
t=1
(36)
Φ((xt,R u − xt,I v)rt,R ) · Φ((xt,R v + xt,I u)rt,I )dudv,
Rx
t2
where Φ(x) = √12π −∞ e− 2 dt is the cumulative normal distribution function.
1) Achievable Rate with i.i.d. QPSK for the 1-bit Block Fading SISO Model: With the above formula, the achievable rate
of the one-bit quantized SISO channel with QPSK input reads as
1 XX
P (r|x)
QPSK
C1−bit
(SNR) = T
)
P (r|x) log2 (
4 x r
P (r)
(37)
X
=
P (r|x0 ) log2 (4T P (r|x0 )),
r
√
√
√
√
Here, x is drawn from all possible sequences √
of T equally likely QPSK data symbols, i.e., x ∈ { SNR, − SNR, j SNR, −j SNR}T
and x0 denotes the constant sequence x0t = SNR, ∀t. The second equality in (37) follows due to the symmetry of the QPSK
MEZGHANI, NOSSEK and SWINDLEHURST: LOW SNR ASYMPTOTIC RATES OF VECTOR CHANNELS WITH SINGLE BIT OUTPUTS
constellation and since
1 X
1
P (r) = T
P (r|x) = T
4 x
4
X
x′ ∈{±1,±j}
P (r|x′ ◦ x0 )
due to (36)
=
1
4T
X
x′ ∈{±1,±j}
P (x′ ◦ r|x0 ) =
11
1 X
1
P (r ′ |x0 ) = T . (38)
4T ′
4
r
We note that the rate expression (37) corresponds exactly to its lower bound in (33) due to the fact that in the i.i.d. QPSK
case P (r|h) = P (r) = 4−T and thus I(h, r) = 0. Furthermore, we use (36) to get a simpler expression for P (r|x0 ) as
P (r|x0 ) =
1
2π
Z
T
+∞ 2 Y
− u2
e
−∞
Z
√
Φ( SNRrt,R u)du
t=1
0
0
+∞
e−
v2
2
−∞
= P (Re(r)|x ) · P (Im(r)|x ).
T
Y
√
Φ( SNRrt,I v)dv
t=1
(39)
Then (37) simplifies to
!
!
2
Z +∞ − u2 Y
T
T
Y
√
√
e 2
e
T
√
√
=2
Φ( SNRrt u)du · log2 2
Φ( SNRrt u)du
2π t=1
2π t=1
−∞
−∞
t,rt =±1
!
!
Z +∞ − u2
Z +∞ − u2
T
X
√
√
√
√
e 2
e 2
T
T
k
T −k
k
T −k
√ Φ(− SNRu) Φ( SNRu)
√ Φ(− SNRu) Φ( SNRu)
du · log2 2
du .
=2
k
2π
2π
−∞
−∞
k=0
(40)
For the special case T = 2, we use the following closed form solution from [62] for the integral in (40)
Z
X
QPSK
C1−bit
(SNR)
1
√
2π
Z
+∞ − u2
+∞
−∞
e
2
− u2
2
Y
√
1
2
SNR
Φ( SNRrt,c u)du =
1 + arcsin(
)r1,c r2,c .
4
π
1 + SNR
t=1
(41)
Thus (40) becomes
SNR
2
QPSK
arcsin(
)
log
C1−bit,T
(SNR)
=
1
+
2 1+
=2
π
1 + SNR
2
SNR
1 − arcsin(
) log2 1 −
π
1 + SNR
2
SNR
arcsin(
) +
π
1 + SNR
2
SNR
arcsin(
) .
π
1 + SNR
(42)
And for the case T = 3, we substitute using again the corresponding formula from [62])
Z +∞ 2 Y
3
√
u
1
√
Φ( SNRrt,c u)du =
e− 2
2π −∞
t=1
1
2
SNR
SNR
SNR
1+
arcsin(
)r1,c r2,c + arcsin(
)r1,c r3,c + arcsin(
)r2,c r3,c
8
π
1 + SNR
1 + SNR
1 + SNR
(43)
in (40) and we obtain
C1−bit,T =3 (SNR) =
1
3
SNR
+ arcsin(
) log2 1 +
2 π
1 + SNR
3
SNR
3
− arcsin(
) log2 1 −
2 π
1 + SNR
6
SNR
arcsin(
) +
π
1 + SNR
2
SNR
arcsin(
) .
π
1 + SNR
(44)
A closed form solution for the integral in (39) for T > 3 is unknown and only approximations are found in the literature [62].
In Fig. 5, we plot the capacity of the one-bit SISO Rayleigh-fading channel per channel use over the SNR for two cases
T = 2 and T = 3. The coherent capacity achieved by quantized QPSK, which is an upper bound on our non-coherent capacity
(see (34)), is also shown. It is obtained assuming that the receiver knows the channel coefficient h (or T goes to infinity). The
average capacity achieved by QPSK over the quantized coherent Rayleigh-fading channel is obtained from (34):
coh
C1−bit
= Eh [H(x, r|h)]
"
#
XX
P (r|x, h) log2 P (r|x, h)
= Eh 2 +
x
y
Z ∞
√
2
1
− u2
Hb (Φ( SNR · u))du ,
e
=2 1− √
2π −∞
(45)
12
SUBMITTED TO THE IEEE TRANSACTIONS ON INFORMATION THEORY, DECEMBER 2017
where we have used the binary entropy function Hb (p) = −p · log2 p − (1 − p) · log2 (1 − p), and we substituted H(r) = 2
since all four possible outputs are equiprobable, and
!
!
Re[y]Re[hx]
Im[r]Im[hx]
p
p
P (r|x, h) = Φ
Φ
.
(46)
1/2
1/2
In Fig. 6, we plot the normalized capacity of the one-bit Rayleigh-fading SISO channel versus coherence interval T for
SNR=10dB. The coherent upper-bound is also plotted.
1.4
Coherent upper bound
QPSK Rate for T=2
QPSK Rate for T=3
C/T (bits/channel use)
1.2
1
0.8
0.6
0.4
0.2
PSfrag replacements
0
0
2
4
6
8
10
SNR (linear)
Fig. 5.
Capacity of the one-bit Rayleigh-fading SISO channel for T = 2, T = 3, T = ∞.
1.6
C/T (bits/channel use)
1.4
1.2
1
0.8
0.6
0.4
0.2
PSfrag replacements
0
1
2
3
4
5
6
7
8
9
10
T
Fig. 6.
Capacity of the one-bit Rayleigh-fading SISO channel versus coherence interval T for SNR=10dB (dashed curve is for the coherent case T → ∞).
2) Training Schemes for the 1-bit Block Fading Model: Training based schemes are attractive for communication over a
priori unknown channels, since the receiver task becomes significantly easier. Therefore, most of the wireless communication
standards use part of the transmission block for sending a pilot sequence that is known at the receiver:
xT
x=
,
(47)
xD
where xT is a fixed known training vector of length TT < T . This transmit strategy based on the separation of data and
training symbols is in general suboptimal from an information theoretical point of view, but is beneficial from a practical point
MEZGHANI, NOSSEK and SWINDLEHURST: LOW SNR ASYMPTOTIC RATES OF VECTOR CHANNELS WITH SINGLE BIT OUTPUTS
13
of view. The achievable rate with joint processing of data and training of this scheme can be written as follows:
Training
C1−bit
= H(r|xT ) − H(r|x)
= H(r) − I(xT , r) − H(r|x) = C1−bit,T − C1−bit,TT ,
(48)
where we have used the formula for the mutual information I(xT , r) = H(r) − H(r|xT ) = C1−bit,TT . The expression of the
capacities with the coherence time can be obtained from (40). In other words, we have taken the difference of the entropy of
the quantized output given the training part xT . In fact, the capacity is reduced by an amount that corresponds to the rate (40)
with coherence time TT :
!
Z +∞ − u2
TT
X
√
√
e 2
TT
k
TT −k
√ Φ(− SNRu) Φ( SNRu)
C1−bit,TT = 2
du ·
k
2π
−∞
k=0
(49)
!
Z +∞ − u2
√
√
2
e
√ Φ(− SNRu)k Φ( SNRu)TT −k du .
log2 2T
2π
−∞
It is worth mentioning that for the case where we fix one symbol in the input sequence x as a training symbol, we can get the
same capacity since CTT =1 = 0. Therefore, in contrast to the unquantized case, a single training symbol for the input sequence
x can always be used without any penalty in the channel capacity. Fig. 7 shows the achievable capacity of a SISO channel
with a coherence length of T = 10 and SNR= 10dB as a function of the training length TT . Again, we observe that only one
symbol as training will not reduce the capacity, while the curve then decreases with almost a slope of −1. Therefore, we can
see that choosing the training length to be negligible compared with the coherence time is necessary such that the penalty due
to the separation of the channel estimation from the data transmission is small enough.
1
0.9
0.8
T=10
SNR=10 dB
PSfrag replacements
H · x2
C Training /C
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
TT /T
Fig. 7.
Achievable rate as a function of the fraction of the coherence time spent on training for T = 10 and SNR= 10dB.
B. Second-order Expansion of the Mutual Information
In this section, we will consider more general channels and elaborate on the second-order expansion of the input-output
mutual information (30) of the considered system in Fig. 2 as the signal-to-noise ratio goes to zero, and where only statistical
CSI is available at the receiver. We state the main result and we prove it afterwards.
Theorem 5: Consider the one-bit quantized MIMO system in Fig. 2, where the channel matrix H is zero-mean and circularly
4+ε
distributed under an input distribution p(x) satisfying Ex,H [kHxk4 ] < δ for some finite constants ε, δ > 0. Then, to second
order, the mutual information (in nats) with statistical CSI between the inputs and the quantized outputs is given by
2 n h
o
i
1 2 1
tr E (nondiag(E[HxxH H H |x]))2 − (nondiag(E[HE[xxH ]H H ]))2
I(x, r) =
2
2 π ση
(50)
+ ∆I(x, r),
| {z }
o( σ14 )
η
where
4
kak4
is the 4-norm of a taken to the power 4 defined as
P
4
i,c ai,c .
14
SUBMITTED TO THE IEEE TRANSACTIONS ON INFORMATION THEORY, DECEMBER 2017
1) Comments on Theorem 5: As an example, Fig. 8 illustrates the rate expression (44) and the quadratic approximation (50)
computed for a block fading SISO model with a coherence interval of 3 symbol periods (H = h · I3 , h ∼ CN (0, 1)) under
QPSK signaling.
PSfrag replacements
Mutual Information in bpcu
0.5
Exact Rate
nd
2 order approx.
0.4
0.3
0.2
0.1
N
0
0
Fig. 8.
0.1
0.2
PTr 0.3
ση2
0.4
0.5
Mutual information of the one-bit block Rayleigh-faded SISO channel with block length 3.
A similar result was derived by Prelov and Verdù in [54] for the soft output y
o
i
1
1 1 n h
I(x, y) = 4 tr E (E[HxxH H H |x])2 − (E[HE[xxH ]H H ])2 + o( 4 ),
2 ση
ση
(51)
where we identify a power penalty of π2 due to quantization and we see that the diagonal elements of EH [HxxH H H ]
do not contribute to the mutual information in the hard-decision system. This is because no mutual information can be
extracted from the amplitude of each two-level received signal. That means that the channel coefficients have to be correlated,
otherwise EH [HxxH H H ] is diagonal and the mutual information is zero. Nevertheless in most systems of practical interest
correlation between the channel coefficients, whether temporal or spatial, exists even in multipath rich mobile environments,
thus EH [HxxH H H ] is a rather dense matrix whose Frobenius norm is dominated by the off-diagonal elements rather than
the diagonal entries. Thus the low SNR penalty due to the hard-decision is nearly 1.96 dB for almost all practical channels.
This confirms that low-resolution sampling in the low SNR regime performs adequately regardless of the channel model and
the kind of CSI available at the receiver, while reducing power consumption.
2) Proof of Theorem 5: We start again with the definition of the mutual information
I(x, r) = H(r) − H(r|x)
= H(sign(Hx + η)) − H(sign(Hx + η)|x),
then we use, Lemma 1 and Lemma 2 with ε =
(52)
1
ση
and g(x) = Hx to get the following asymptotic expression:
2
2 1
2
1
2
H
H
I(x, r) = 2N ln 2 −
tr
nondiag(Ex,H [Hxx H ])
kEx,H [Hx]k2 − 4
π ση2
ση π 2
4
4
4
1
kEx,H [Hx] ◦ Ex,H [Hx ◦ Hx ◦ Hx]k1 + 2 kEx,H [Hx]k4
−
3π
3π
2
2
1
2 1
2
H
H
Ex [kEH [Hx]k2 ] + 4 Ex 2 tr
nondiag(EH [Hxx H ])
− 2N ln 2 +
π ση2
ση
π
1
4
4
4
1
kEH [Hx] ◦ EH [Hx ◦ Hx ◦ Hx]k1 + 2 kEH [Hx]k4 + o( 4 )
−
3π
3π
ση
2
2
1
2
H
H
H
H
+ o( 4 ),
= 4 2 tr Ex [ nondiag(EH [Hxx H ]) ] − nondiag(Ex,H [Hxx H ])
ση π
ση
where the last step follows from the fact that the channel matrix H has zero-mean. We note that since H has a proper
distribution, then g(x) = Hx also has a zero-mean proper distribution, fulfilling p(H) = p(jHx). Therefore the conditions
of Theorem 1 are fulfilled. This yields the result stated by the theorem and completes the proof. Note that the condition
4+ε
Ex,H [kH(x)k4 ] < δ for some finite constants ε, δ > 0 is necessary, so that the remainder of the expansion is asymptotically
negligible as shown in the coherent case (see Section IV-A).
MEZGHANI, NOSSEK and SWINDLEHURST: LOW SNR ASYMPTOTIC RATES OF VECTOR CHANNELS WITH SINGLE BIT OUTPUTS
15
C. IID Block Rayleigh Fading MIMO Channels
We consider as an example a point-to-point quantized MIMO channel where the transmitter employs M ′ antennas and
the receiver has N ′ antennas. We assume a block-Rayleigh fading model [63], in which the channel propagation matrix
′
′
H ′ ∈ CM ×N remains constant for a coherence interval of length T symbols, and then changes to a new independent value.
The entries of the channel matrix are i.i.d. zero-mean complex circular Gaussian with unit variance. The channel realizations
are unknown to both the transmitter and receiver. At each coherence interval, a sequence of vectors x1 , . . . , xT are transmitted
at each time slot via the multiple antennas. The transmitted and received signal matrices are then related as follows (total
dimensions: N = T N ′ and M = T M ′ )
y1
x1
H′ 0 · · ·
0
η1
y2 0 H ′ · · ·
0
x2 η 2
· . + . .
(53)
.. =
.
..
. 0
0 .. ..
0
yT
| {z }
0
0
{z
|
y
H
···
H′
xT
} | {z }
x
ηT
| {z }
Then the expected value of the received signal conditioned on the input is given by
′H
′H
E[H ′ x1 xH
E[H ′ x1 xH
···
1H ]
2H ]
′H
′
′H
H
E[H ′ x2 xH
H
]
E[H
x
x
H
]
···
2
1
2
E[HxxH H H |x] =
.
.
..
..
..
.
′H
E[H ′ xT xH
1H ]
′H
E[H ′ xT xH
2 H ] ···
η
′H
E[H ′ x1 xH
TH ]
′
H
E[H x2 xT H ′H ]
..
.
′H
E[H ′ xT xH
TH ]
.
(54)
′H
T ∗
Since H ′ is i.i.d. distributed, it can be shown that EH ′ [H ′ xi xH
j H ] = xi xj · IN ′ . Therefore, we have
′
E[HxxH H H |x] = (X · X H ) ⊗ IN ′ ,
where the rows of the matrix X ∈ CT ×M are the vectors xT
i , for 1 ≤ i ≤ T . With this, (50) asymptotically becomes
2
o
i
n h
1
N′ 2 1
H
H 2
2
+ o( 4 ),
−
(nondiag(E[X
·
X
]))
tr
E
(nondiag(X
·
X
))
I(x, r) =
2
2 π ση
ση
while for the ideal case, we get from (51) (see also [55])
2 n h
o
i
1
N′ 1
tr E (X · X H )2 − (E[X · X H ])2 + o( 4 ).
I(x, y) =
2
2 ση
ση
Now, assuming i.i.d. Gaussian inputs with E[XX H ] = PTr · IT , then we obtain finally (c.f. expectation (22))
2
1
N ′ 2 PTr
T (T − 1) + o( 4 )
I(x, r) =
2 π ση2
ση
2
N ′ PTr
1
I(x, y) =
T 2 + o( 4 ).
2
2
ση
ση
(55)
(56)
(57)
(58)
Evidently, in order for I(x, r) and I(x, y) to be equal with the same power, we need to increase the number of one-bit receive
antennas N ′ by roughly a factor of π 2 /4 when the coherence interval satisfies T ≫ 1. The same result has been obtained in
[47] with a pilot-based scheme and using the pseudo-quantization noise model.
D. SIMO Channels with Delay Spread and Receive Correlation at low SNR
As a further example, we use the result from Theorem 5 to compute the low SNR mutual information of a frequency-selective
single input multi-output (SIMO) channel with delay spread and receive correlation both in time and space and obtain the
asymptotic achievable rate under average and peak power constraints. The quantized output of the considered model at time
k is
rk = Q(y k )
yk =
T
−1
X
t=0
′
hk [t]xk−t + η k ∈ CN ,
(59)
where the noise process {ηk } is i.i.d. in time and space, while the T fading processes {hk [t]} at each tap t are assumed to be
independent zero-mean proper complex Gaussian processes. Furthermore, we assume a separable temporal spatial correlation
model, i.e.
(60)
E[hk [t]hk′ [t′ ]H ] = C h · ch (k − k ′ )αt δ[t − t′ ].
16
SUBMITTED TO THE IEEE TRANSACTIONS ON INFORMATION THEORY, DECEMBER 2017
Here, C h denotes the receive correlation matrix, ch (k) is the autocorrelation function of the fading process, and the scalars
αi represent the power-delay profile. The correlation parameters can be normalized so that
T
−1
X
tr(C h ) = N ′ , ch (0) = 1 and
αt = 1.
(61)
t=0
In other words, the energy in each receive antenna’s impulse response equals one on average. On the other hand we assume that
the transmit signal xk is subject to an average power constraint E[kxk k2 ] ≤ PTr and a peak power constraint |xk |2 ≤ β · PTr ,
∀k, with β ≥ 1. It should be pointed out that a peak power constraint constitutes a stronger condition than necessary for the
validity of Theorem 5, involving just a fourth-order moment constraint on the input. We consider now a time interval of length
n (a block of n transmissions). Collect a vector sequence y k of length n into the vector y as
T
(62)
y(n) = y Tn−1 , . . . , y T0 ,
′
and form the block cyclic-shifted matrix H(n) ∈ C(N n)×n
hn−1 [0] · · ·
hn−1 [T −1] 0 · · ·
..
..
.
.
.
H(n) =
..
..
.
.
· · · h0 [T −1] 0
···
0 h0 [0]
With
T
x(n) = [xn−1 , . . . , x0 ] ,
(n)
(n)
(63)
(64)
(n)
and η and r defined similar to y , the following quantized space-time model may be formulated as a (loose) approximation
of (59), where the resulting dimensions are N = N ′ n and M = n
(65)
r(n) = Q(y(n) ) = Q H(n) x(n) + η(n) .
1) Asymptotic Achievable Rate: Now, we elaborate on the asymptotic information rate of this channel setting. First we
establish an upper bound and then we examine its achievability.
Proposition 1: If ch (k) is square-summable, then the mutual information of the described space-time model admits the
following asymptotic upper bound, for any distribution fulfilling the average and peak power constraints
2
1
2 PTr
(n) (n)
(66)
lim I(x , r ) ≤
U (β),
n→∞ n
π ση2
where
U (β) =
(
β · ζ + (β − 1)χ
2
β 2 (ζ+χ)
4χ
ζ = tr(C 2h )
∞
X
for β(ζ + χ) ≥ 2χ
else,
(67)
ch (k)2
(68)
k=1
and
χ=
1
tr((nondiag(C h ))2 )ch (0).
2
(69)
Proof: Due to the peak power constraints, the conditions of Theorem 5 are satisfied; thus the second order approximation
(50) is valid. A tight upper bound is obtained by looking at the maximal value that can be achieved by the expression (50) up
to second order. We do this in two steps. We first maximize the trace expression in (50) under a prescribed average power per
symbol γ. The maximum can be, in turn, upper bounded by the supremum of the first term minus the infimum of the second
term under the prescribed average power and the original peak power constraint. After that, we perform an optimization over
the parameter γ itself. That is
(
2
i
h
1
1
2
(n) (n) (n),H (n),H
2
I(x(n) , r(n) ) ≤
sup
max
tr
E
(nondiag(E[H
x
x
H
|x]))
0≤γ≤1
2 π ση2
|xk |2 ≤βPTr ,∀k
E[kx(n) k2 ]=γPTr ·n
)
1
(n)
(n),H
(n) (n),H
2
− inf tr (nondiag(E[H E[x x
]H
])) +o( 4 ).
2
σ
|xk | ≤βPTr ,∀k
η
(70)
E[kx(n) k2]=γPTr ·n
Evaluating the expectation with respect to the channel realizations, we get
E[H(n) x(n) x(n),H H(n),H ] = D ⊗ C h ,
(71)
MEZGHANI, NOSSEK and SWINDLEHURST: LOW SNR ASYMPTOTIC RATES OF VECTOR CHANNELS WITH SINGLE BIT OUTPUTS
17
with
dk,k′ = ch (k − k ′ )
Therefore
T
−1
X
t=0
αt xk−t x∗k′ −t for k, k ′ ∈ {0, . . . , n − 1}.
(72)
2
T
−1
n−1
X
X n−1
X
2
(n) (n) (n),H (n),H
∗
′
2
αt xk−t xk′ −t +
tr(C h )ch (k − k )
tr (nondiag(E[H x x
H
|x])) =
k=0
n−1
X
t=0
k′ =0
k′ 6=k
tr((nondiag(C h ))2 )ch (0)
T
−1
X
t=0
k=0
(73)
2
αt |xk−t |2 .
It turns out that the supremum of the first trace term is achieved when all xk inputs take, simultaneously during the considered
time interval, either the value zero, or the peak value β with a duty cycle of γβ −1 . On the other hand, the infimum of the
second trace expression is obtained, under the prescribed average power, when E[xxH ] = γPTr In . Calculation shows that
!"
2 TX
n−1
−1 2
X
1 (n) (n)
k
2 PTr
2
γβ
·tr(C
)
(1− )ch (k)2
I(x , r )≤ max
α
t
h
2
0≤γ≤1 π ση
n
n
t=0
k=1
|
{z
}
ζ(n)
#
1
1
2
+ γ(β − γ) tr((nondiag(C h )) )ch (0) + o( 4 ).
ση
|2
{z
}
(74)
χ
Now, the maximization over γ delivers
γopt = min{1, β
ζ(n) + χ
}.
2χ
(75)
Thus, taking the limit n → ∞ yields ζ(n) → ζ as in (68) and by the normalizations in (61), we end up with the result of the
proposition.5
We next turn to the question of whether the upper bound suggested by Proposition 1 is achievable. A closer examination of
(70) demonstrates that the upper bound could be achieved if the input distribution satisfies the following condition, for any
time instants k and l within the block of length n:
xk x∗l = aejΩ(k−l) ,
(76)
√
for some Ω 6= 0 and random a ∈ {0, β · PTr }, while having E[x(n) x(n),H ] = γPTr In as already mentioned in the proof
of Proposition 1. Clearly an on-off frequency-shift keying (OOFSK) modulation for the input, as follows, can fulfill these
conditions
(77)
xk = Z · ejkΩ , k ∈ {1, . . . , n},
√
where Z takes the value βPTr with probability γopt β −1 and zero with probability (1 − γopt β −1 ), and Ω is uniformly
2π(n−1)
distributed over the set { 2π
}. This is similar to the results of the unquantized case [57], [64].
n ,...,
n
2) Discussion: We notice from (75) that the average power constraint E[kxk k2 ] ≤ PTr is only active when β(ζ + χ) ≥ 1,
which means that it is not necessarily optimal to utilize the total available or allowed average power especially if a tight peak
power constraint is present. In addition, if we impose only a peak power constraint, i.e. β = 1, then if ζ < χ the on-off
strategy with the zero symbol (a = 0) is required to approach the capacity. We notice that on-off modulation in this case is
crucial and should not be regarded as duty cycle transmission where the link remains inactive for a certain amount of time. We
observe also from Proposition 1 that spreading the power over different taps does not affect the low SNR rate, while receiver
correlation is beneficial due to two effects. First the mutual information increases with χ defined in (69) which is related to
the norm of the off-diagonal elements of C h . Second, under the normalization tr(C h ) = N ′ , the Frobenius norm tr(C 2h )
increases with more correlation among the receive antennas, and consequently, higher rates at low SNR can be achieved due
to relation (68). In fact, both spatial and temporal correlations are extremely beneficial at low SNR, even more than in the
unquantized case. Besides we note that the achievability of the upper bound stated in Proposition 1, as discussed previously,
is obtained at the cost of burstiness in frequency which may not agree with some specifications imposed on many systems.6
5 Observe
6 Note
that ζ and χ are indicators for the temporal and spatial coherence, respectively.
that such observations hold also in the unquantized case [57].
18
SUBMITTED TO THE IEEE TRANSACTIONS ON INFORMATION THEORY, DECEMBER 2017
√
√
Therefore, it is interesting to look at the asymptotic rate of i.i.d. input symbols drawn from the set {− βPTr , 0, βPTr }. In
that case it turns out by (50) that
!2
2
T
−1
T
−1
X
X
2 PTr 2 (n)
1
IIID (x(n) , r(n) )≈ max
γ ζ
α2t + γ(β − γ)χ
αt .
(78)
2
0≤γ≤1 π ση
n
t=0
t=0
Here, we observe that,P
contrary to to the FSK-like scheme, the mutual information with i.i.d input is negatively affected by
−1 2
the delay spread since Tt=0
αt ≤ 1 by the normalization (61). Nevertheless for the case χ ≫ ζ, i.e. low temporal correlation
(high Doppler spread), the gap to the upper bound in Proposition 1 vanishes, as demonstrated in Fig. 9.
PSfrag replacements
Ratio between i.i.d. rate and upper bound
1
0.8
0.6
T =5
0.4
0.2
0
0
Fig. 9.
T =1
2
4
χ
ζ
6
8
Ratio of i.i.d. rate (78) and upper bound (66) vs. the spatial-to-temporal coherence ratio
χ
ζ
10
for β = 2 and uniform delay spread, i.e. αt =
1
T
.
VI. C ONCLUSION
Motivated by the simplicity of one-bit ADCs for sampling large dimensional signals, we provided a general second order
asymptotic analysis for the entropy of one-bit quantized vector signals. We used these results to evaluate the mutual information
around zero SNR for a wide class of channel models. We have shown that the reduction of low SNR channel capacity by a
factor of 2/π due to 1-bit quantization holds for the general MIMO case with uncorrelated noise. Additionally, the non-coherent
MIMO channel was studied in detail and a similar conclusion was made. Thereby, as one might expect, it turns out that QPSK
plays an essential role in achieving optimal performance for 1-bit quantized channels at low SNR.
A PPENDIX A
P ROOF OF T HEOREM 1
We aim at the derivation of the second order approximation (Taylor expansion) of the entropy H(r), where r = sign(y) =
sign(εx + η), with respect to ε. Thereby, we assume that the vector η is i.i.d. Gaussian distributed with unit variance and x
has a general distribution. As will be shown in the following, this second order Taylor expansion is a function of the moments
of the distribution of x, which have to fulfill certain conditions. For simplicity, we consider first the real-valued case for the
computation of the second order approximation of the entropy H(sign(εx + η)) assuming that x has a real valued distribution
and η follows a real-valued Gaussian distribution with variance 21 . The generalization to the complex case can be obtained
based on the real-valued representation of the complex channel and is performed at the end of the proof. The probability mass
function Pǫ (r) of the random vector r = sign(εx + η) ∈ {±1}N is given by
Z
Pǫ (r) = P (r ◦ y ∈ RN
)
=
E
pǫ (r ◦ y|x)dy,
(79)
x
+
RN
+
where
pǫ (y|x) =
1
π N/2
2
e−ky−εxk2 .
We also have the general expression of the entropy (in nats) of random vector r as
X
−Pε (r) ln Pε (r).
H(r) =
r∈{±1}N
(80)
(81)
MEZGHANI, NOSSEK and SWINDLEHURST: LOW SNR ASYMPTOTIC RATES OF VECTOR CHANNELS WITH SINGLE BIT OUTPUTS
19
Next, based on the derivatives of the function −z ln z, we calculate the second order expansion of the entropy with respect
to ε. For that, we first notice that the linear and cubic terms of the Taylor expansion vanish due to the fact that the entropy
function H(sign(εx + η)) is an even function with
P respect to ε since H(sign(εx + η)) = H(sign(−εx + η)). Additionally,
since Pε (r) is a probability mass function, i.e., r Pε (r) = 1, we have
X
Pε(k) (r) = 0,
(82)
r
(k)
for any k-th derivative of Pε (r) with respect to ε. Based on these observations, we get the following second order expansion
H(r) =
X
r
ε
−P0 (r) ln P0 (r) − ε2
P0′ (r)2
−
2P0 (r)
(3)
P0′ (r)2 P0′′ (r) P0′′ (r)2
P0′ (r)P0 (r)
P0′ (r)4
−
+
+
12P0 (r)3
4P0 (r)2
8P0 (r)
6P0 (r)
4
!
(83)
+ ∆H(r),
(3)
where P0′ (r), P0′′ (r) and P0′ (r)P0 (r) are the first, second and third order derivatives of Pε (r) at ε = 0, which will be
derived in the following. To this end, we utilize the following derivatives of the unquantized output distribution pǫ (y|x) =
2
1
e−ky−εxk2
π N/2
2
1
(84)
p0 (y|x) = N/2 e−kyk2 ,
π
p′0 (y|x) =
p′′0 (y|x) =
2
e−kyk2 2y T x,
(85)
2
2
e−kyk2 (2(y T x)2 − kxk2 ),
2
2
2
e−kyk2 2y T x( (y T x)2 − kxk2 ).
3
6
(3)
p0 (y|x) =
2
π N/2
1
π N/2
π N/2
(86)
(87)
Now, we calculate P0 (r)
P0 (r) =
Z
RN
+
2
π
N
2
2
e−kyk2 dy =
1
.
2N
Afterward, we evaluate the first order derivative as
Z
2 −kyk22
1 2
P0′ (r) =
(r ◦ y)T E[x]dy = N √ r T E[x],
N e
N
2
π
2
π
R+
(88)
(89)
while the second order derivative reads as
P0′′ (r)
= Ex
= Ex
1
= N
2
1
= N
2
where we used the following result
Z
RN
+
Z
Z
2
H
2
−kyk22
2((r
◦
y)
x)
−
kxk
dy
e
2
N/2
2
RN
+
π
RN
+
π N/2
2
2
e−kyk2 xT 2(r ◦ y)(r ◦ y)T − 1 x dy
(90)
4 T
E x nondiag rrT x
π
4 T
r nondiag E[xxT ] r,
π
2
e−kyk
(yi · ri )(yj · rj )dy =
π N/2
1
N
2·2
ri rj
π·2N
for i = j
else,
and the relation tr(nondiag(A)B) = tr(Anondiag(B)) for any two quadratic matrices A and B.
(91)
20
SUBMITTED TO THE IEEE TRANSACTIONS ON INFORMATION THEORY, DECEMBER 2017
(3)
In a similar way, we calculate P0 (r)
Z
6 −kyk22
2
(3)
2
T
T
2
P0 (r) = Ex
e
2((r
◦
y)
x)
((r
◦
y)
x)
−
kxk
2 dy
πN
3
RN
+
Z
8 −kyk22
12 −krk22
2
T
3
T
= Ex
e
((r ◦ y) x) − N e
((r ◦ y) x) kxk2 dy
πN
π
RN
+
Z
12
8 −kyk22
2
e
((r ◦ y)T x)3 dy − N √ r T · E[kxk2 · x]
= Ex
N
π
2
π
RN
+
X
12
48 1
1 −4 T
2
E [ri xi rj xj rl xl ]
= N √ (r E[x ◦ x ◦ x]) + N √ rT · E[kxk2 · x] + N 3
2
2 π 2 j6=i,i6=l,j6=l
π
2
π
12
2
− N √ r T · E[kxk2 · x]
2
π
X
1 −4
48 1
= N √ (r T E[x ◦ x ◦ x]) + N 3
ri rj rl E [xi xj xl ] ,
2
2 π 2 j6=i,i6=l,j6=l
π
which follows from
Z
8 −kyk22
e
((r ◦ y)T x)3 dy =
πN
Z
X
X
2 X
8
= N Ex
ri3 yi3 x3i + 3
ri2 yi2 x2i rj yj xj + 6
e−kyk2 (
ri yi xi rj yj xj rl yl xl )dy
π
RN
+
i
i,j6=i
j6=i6=l
X
X
X
1
1
1
1
1
8
1 2
ri √ x3i + 3
xi rj √ xj + 6
ri √ xi rj √ xj rl √ xl
= N Ex
2
π
2
π
π
π
π
i
i,j6=i
j6=i6=l
X
X
X
1
1
1
1
1
3
1
8
ri √ x3i +
x2 rj √ xj + 6
ri √ xi rj √ xj rl √ xl
= N Ex −
2
2 i
π
2 i,j i
π
π
π
π
Ex
(92)
RN
+
(93)
j6=i6=l
12
48 1
4
2
= − N Ex (r T E[x ◦ x ◦ x]) + N √ rT · E[kxk2 · x] + N 3
2
2
π
2 π2
X
E [ri xi rj xj rl xl ] .
j6=i,i6=l,j6=l
(3)
Now, after the derivation of the derivatives P0′ (r), P0′′ (r) and P0 (r), we have to perform a summation with respect to
r ∈ {−1, +1}N of the terms given in (83). For this, we make use of the following properties:
X
1
r T Ar = tr(A),
(94)
N
2
N
r∈{−1,+1}
and
1
2N
X
r∈{−1,+1}N
r T Ar · r T Br = tr(A · nondiag(B + B T )) + tr(A)tr(B),
(95)
which can be verified by expanding the left-hand side and identifying the non-zero terms. We start by the second order
expression of (83)
X
X
2
P0′ (r)2
=
(r T E[x]T )2
N
2P0 (r)
2 ·π
N
N
r∈{−1,+1}
r∈{−1,+1}
=
X
r∈{−1,+1}N
2
rT E[x]E[x]T r
| {z }
2N · π
(96)
A
2
2
= kxk2 .
π
Then, we consider the different terms of the fourth-order derivatives, starting with
X
X
P0′ (r)4
4
=
(r T E[x]T )4
3
N · 3π 2
12P
(r)
2
0
N
N
r∈{−1,+1}
r∈{−1,+1}
=
X
r∈{−1,+1}N
2N
4
r T E[x]E[x]T rrT E[x]E[x]T r
| {z }
| {z }
· 3π 2
A
B
4
4
= 2 tr(nondiag(E[x]E[x]T )2 ) + 2 kE[x]k44 .
π
3π
(97)
MEZGHANI, NOSSEK and SWINDLEHURST: LOW SNR ASYMPTOTIC RATES OF VECTOR CHANNELS WITH SINGLE BIT OUTPUTS
On the other hand, we have
X
P0′′ (r)2
=
8P0 (r)
N
r∈{−1,+1}
X
r∈{−1,+1}N
21
2
rT nondiag E[xxT ] rrT nondiag E[xxT ] r
2N · π 2
(98)
4
= 2 tr(nondiag(E[xxT ])2 ).
π
Next, we get
X
r∈{−1,+1}N
P0′ (r)2 P0′′ (r)
=
4P0 (r)2
X
r∈{−1,+1}N
2N
4
r T nondiag E[xxT ] rr T nondiag E[x]E[x]T r
2
·π
(99)
8
= 2 tr(nondiag(E[x]E[x]T )nondiag(E[xxT ])).
π
Further, we obtain
(3)
P0′ (r)P0 (r)
=
6P0 (r)
r∈{−1,+1}N
X
2 T
1
√ r E[x]
6 · 2N
π
N
X
X
48 1
−4
√ (r T E[x ◦ x ◦ x]) + N 3
ri rj rl E [xi xj xl ]
π
2 π 2 j6=i,i6=l,j6=l
r∈{−1,+1}
X
4
16 X
= − E[x]T · E[x ◦ x ◦ x] + 2
ri rj rl rk E[xk ]E [xi xj xl ]
3π
π
N j6=i,i6=l,j6=l,k
!
(100)
r∈{±1}
4
= − E[x]T · E[x ◦ x ◦ x] + 0.
3π
We plug the results from (96)–(100) into the expression for the entropy in (83), to get finally
2
4
2
2
H(sign(εx + η)) =2N ln 2 − ε2 kE[x]k2 − ε4
tr (nondiag(C x ))
2
π
π
4
4
4
1
kE[x] ◦ E[x ◦ x ◦ x]k1 + 2 kE[g(x)]k4 + ∆H(sign(εx + η)).
−
3π
3π
(101)
Now, for the complex valued case, we easily obtain a similar result, since any N -dimensional vector can be represented as
a real-valued 2N -dimensional vector, where we split the vector into the real part and the imaginary part and write them as:
Re{x}
r
x =
.
(102)
Im{x}
This implies that the vector norms can be defined as
kxk44 =
X
k
Re{xk }4 + Im{xk }4 .
(103)
On the other hand, the covariance matrix can be written as
C x = E[xxH ] − E[x]E[xH ].
T
(104)
T
Under the assumption that x is circularly distributed, i.e., E[xx ] = E[x][x ], then we have
1
Re{C x } Im{C x }
,
C xr = E[xr xr,T ] =
2 −Im{C x } Re{C x }
(105)
where we note that the matrix Im{C x } is a skew-symmetric matrix, i.e., Im{C T
x } = −Im{C x } and has diagonal elements
equal to zero, thus we have
1
(nondiag(Re{C x }))2 −
2
1
= (nondiag(Re{C x }))2 −
2
1
= (nondiag(C x ))2 .
2
(nondiag(C xr ))2 =
1
(Im{C x })2
2
1
(nondiag(Im{C x }))2
2
(106)
Therefore, we can conclude that the formula remains the same except that we replace (nondiag(E[xxH ]))2 by
1
H
2
2 (nondiag(E[xx ])) , and the norms retain their same definition. Now, we turn to the question of the necessary conditions
22
SUBMITTED TO THE IEEE TRANSACTIONS ON INFORMATION THEORY, DECEMBER 2017
such that the second order approximation is valid. The condition Ex [kxk4+α
] < γ for some finite constants α, γ > 0 stated
4
by the theorem is necessary, so that the remainder term of the expansion given by
4
∆H(sign(εx + η)) = Ex [o(kxk4 ε4 )]
satisfies
∆H(x, y)
= 0,
ε4
→0
lim
4
ε
(107)
(108)
since
∆H(x, r) = Ex [o(kxk44 ε4 )]
4
≤ Ex [(kxk4 ε4 )1+
′
4+α
≤ Ex [kxk4
α′
4
′
]ε4+α
4+α′
], for some α′ ∈]0, α]
′
≤ Ex [kxk4+α
] 4+α ε4+α (Hölder’s inequality)
4
≤γ
4+α′
4+α
′
ε4+α
= o(ε4 ).
This yields the result stated by the theorem and completes the proof.
R EFERENCES
[1] A. Mezghani and J. A. Nossek, “On Ultra-Wideband MIMO Systems with 1-bit Quantized Outputs: Performance Analysis and Input Optimization,”
IEEE International Symposium on Information Theory (ISIT), Nice, France, June 2007.
[2] A. Mezghani and J. A. Nossek, “Analysis of Rayleigh-Fading Channels with 1-bit Quantized Output,” IEEE International Symposium on Information
Theory (ISIT), Toronto, Canada, June 2008.
[3] A. Mezghani and J. A. Nossek, “Analysis of 1-bit Output Noncoherent Fading Channels in the Low SNR Regime,” IEEE International Symposium on
Information Theory (ISIT), Seoul, Korea, June 2009.
[4] T. L. Marzetta, “Multi-cellular wireless with base stations employing unlimited numbers of antennas,” in in UCSD Information Theory and Applications
Workshop, February 2010.
[5] T.L. Marzetta, “Noncooperative Cellular Wireless with Unlimited Numbers of Base Station Antennas,” Wireless Communications, IEEE Transactions
on, vol. 9, no. 11, pp. 3590–3600, November 2010.
[6] L. Lu, G. Y. Li, A. L. Swindlehurst, A. Ashikhmin, and R. Zhang, “An Overview of Massive MIMO: Benefits and Challenges,” IEEE Journal of
Selected Topics in Signal Processing, vol. 8, no. 5, pp. 742–758, Oct 2014.
[7] U. Madhow, “MultiGigabit millimeter wave communication: System concepts and challenges,” in in UCSD Information Theory and Applications
Workshop, January 2008, pp. 193–196.
[8] W. Roh, J.-Y. Seol, J. Park, B. Lee, J. Lee, Y. Kim, J. Cho, K. Cheun, and F. Aryanfar, “Five Disruptive Technology Directions for 5G,” IEEE Comm.
Magazine, vol. 52, no. 2, pp. 74–80, February 2014.
[9] A.L. Swindlehurst, E. Ayanoglu, P. Heydari, and F. Capolino, “Millimeter-wave massive MIMO: the next wireless revolution?,” Communications
Magazine, IEEE, vol. 52, no. 9, pp. 56–62, September 2014.
[10] W. Roh, Ji-Yun Seol, Jeongho Park, Byunghwan Lee, Jaekon Lee, Yungsoo Kim, Jaeweon Cho, Kyungwhoon Cheun, and F. Aryanfar, “Millimeter-wave
beamforming as an enabling technology for 5G cellular communications: theoretical feasibility and prototype results,” Communications Magazine, IEEE,
vol. 52, no. 2, pp. 106–113, February 2014.
[11] S. Rangan, T.S. Rappaport, and E. Erkip, “Millimeter-Wave Cellular Wireless Networks: Potentials and Challenges,” Proceedings of the IEEE, vol. 102,
no. 3, pp. 366–385, March 2014.
[12] M. T. Ivrlač and J. A. Nossek, “Challenges in Coding for Quantized MIMO Systems,” in IEEE International Symposium on Information Theory, Seattle,
U.S.A., July 2006, pp. 2114–2118.
[13] B. Murray and I. B. Collings, “AGC and Quantization Effects in a Zero-Forcing MIMO Wireless System,” in Proc. of the IEEE Int. Vehicular Technology
Conf. (VTC) - Spring, Melbourne, Australia, May 2006.
[14] B. M. Murray, “Maximizing the cutoff rate in a quantized MIMO wireless system with AGC,” in International Conference on Wireless Broadband and
Ultra Wideband Communications, Sydney, Australia, March 2006.
[15] J. A. Nossek and M. T. Ivrlač, “Capacity and coding for quantized MIMO systems,” in Intern. Wireless Commun. and Mobile Computing Conf.
(IWCMC), Vancouver, Canada, July 2006, pp. 1387–1392.
[16] A. Mezghani, M. S. Khoufi, and J. A. Nossek, “A Modified MMSE Receiver for Quantized MIMO Systems,” In Proc. ITG/IEEE WSA, Vienna, Austria,
February 2007.
[17] A. Mezghani, M. S. Khoufi, and J. A. Nossek, “Spatial MIMO Decision Feedback equalizer Operating on Quantized Data,” Accepted for publication
at to ICASSP, 2008.
[18] A. Mezghani, F. Antreich, and J. A. Nossek, “Multiple Parameter Estimation With Quantized Channel Output,” in Proceedings of the International
ITG/IEEE Workshop on Smart Antennas WSA, Bremen, Germany, February 2010.
[19] T. Koch and A. Lapidoth, “Asymmetric Quantizers are Better At Low SNR,” IEEE International Symposium on Information Theory (ISIT), Saint
Petersburg, Russia, August 2011.
[20] J. Singh, O. Dabeer, and U. Madhow, “On the limits of communication with low-precision analog-to-digital conversion at the receiver,” Communications,
IEEE Transactions on, vol. 57, no. 12, pp. 3629–3639, December 2009.
[21] J. Singh, O. Dabeer, and U. Madhow, “Capacity of the discrete-time AWGN channel under output quantization,” in IEEE Intern. Symp. Inform. Theory
(ISIT), Toronto, Canada, July 2008, pp. 1218–1222.
[22] T. Koch and A. Lapidoth, “Does output quantization really cause a loss of 2dB?,” in IEEE International Symposium on Information Theory, Saint
Petersburg, Russia, August 2011.
[23] G. Zeitler, A. C. Singer, and G. Kramer, “Low-precision A/D conversion for maximum information rate in channels with memory,” in IEEE Int. Symp.
on Information Theory, Saint Petersburg, Russia, August 2011, pp. 63–78.
[24] K. Nakamura and T. Tanaka, “Performance analysis of signal detection using quantized received signals of linear vector channel,” in Proc. Inter. Symp.
Inform. Theory and its Applications (ISITA), Auckland, New Zealand, December 2008.
MEZGHANI, NOSSEK and SWINDLEHURST: LOW SNR ASYMPTOTIC RATES OF VECTOR CHANNELS WITH SINGLE BIT OUTPUTS
23
[25] A. Mezghani and J. A. Nossek, “Belief Propagation based MIMO Detection Operating on Quantized Channel Output,” IEEE International Symposium
on Information Theory (ISIT), Seoul, South Korea, June 2010.
[26] A. Mezghani, M. S. Khoufi, and J. A. Nossek, “Maximum Likelihood Detection For Quantized MIMO Systems,” in In Proc. ITG/IEEE WSA, Darmstadt,
Germany, February 2008.
[27] A. Mezghani and J. A. Nossek, “Capacity Lower Bound of MIMO Channels with Output Quantization and Correlated Noise,” in Proceedings of 2012
IEEE International Symposium on Information Theory, July 2012, p. not available on IEEE Xplore.
[28] G. Zeitler, A.C. Singer, and G. Kramer, “Low-Precision A/D Conversion for Maximum Information Rate in Channels with Memory,” Communications,
IEEE Transactions on, vol. 60, no. 9, pp. 2511–2521, September 2012.
[29] Shengchu Wang, Yunzhou Li, and Jing Wang, “Multiuser detection for uplink large-scale MIMO under one-bit quantization,” in Communications (ICC),
2014 IEEE International Conference on, June 2014, pp. 4460–4465.
[30] Shengchu Wang, Yunzhou Li, and Jing Wang, “Multiuser Detection in Massive Spatial Modulation MIMO With Low-Resolution ADCs,” Wireless
Communications, IEEE Transactions on, vol. 14, no. 4, pp. 2156–2168, April 2015.
[31] J. Mo and R.W. Heath, “Capacity Analysis of One-Bit Quantized MIMO Systems With Transmitter Channel State Information,” Signal Processing,
IEEE Transactions on, vol. 63, no. 20, pp. 5498–5512, October 2015.
[32] C. Studer and G. Durisi, “Quantized Massive MU-MIMO-OFDM Uplink,” IEEE Transactions on Communications, vol. 64, no. 6, pp. 2387–2399, June
2016.
[33] S. Jacobsson, G. Durisi, M. Coldrey, U. Gustavsson, and C. Studer, “Throughput Analysis of Massive MIMO Uplink with Low-Resolution ADCs,” To
appear in IEEE Transactions on Wireless Communications.
[34] J. Singh, S. Ponnuru, and U. Madhow, “Multi-gigabit communication: The ADC bottleneck,” in in IEEE International Conference on Ultra-Wideband
(ICUWB), Sep. 2009, pp. 22–27.
[35] A. Mezghani, M. Rouatbi, and J. Nossek, “An iterative receiver for quantized MIMO systems,” in in IEEE Mediterranean Electrotechnical Conference
(MELECON), Hammamet, Tunisia, March 2012, pp. 1049–1052.
[36] J. Mo and R. W. Heath Jr., “High SNR capacity of millimeter wave MIMO systems with one-bit quantization,” in in Information Theory and Applications
Workshop (ITA), Feb. 2014.
[37] D. Persson C. Risi and E. G. Larsson, “Massive MIMO with 1-bit ADC,” 2014, Available: http://arxiv.org/abs/1404.7736.
[38] J. Mo P. Schniter N. González-Prelcic and R. W. Heath Jr., “Channel estimation in millimeter wave MIMO systems with one-bit quantizationn,” in in
48th Asilomar Conference on Signals, Systems and Computers, Nov. 2014, pp. 957–961.
[39] N. Liang and W. Zhang, “Mixed-ADC Massive MIMO,” IEEE Journal on Selected Areas in Communications, vol. 34, no. 4, pp. 983–997, April 2016.
[40] S. Jacobsson, G. Durisi, M. Coldrey, U. Gustavsson, and C. Studer, “One-bit massive MIMO: Channel estimation and high-order modulations,” in 2015
IEEE International Conference on Communication Workshop (ICCW), June 2015, pp. 1304–1309.
[41] J. Choi, J. Mo, and R. W. Heath, “Near Maximum-Likelihood Detector and Channel Estimator for Uplink Multiuser Massive MIMO Systems With
One-Bit ADCs,” IEEE Transactions on Communications, vol. 64, no. 5, pp. 2005–2018, May 2016.
[42] C. K. Wen, S. Jin, K. K. Wong, C. J. Wang, and G. Wu, “Joint Channel-and-Data estimation for large-MIMO systems with low-precision ADCs,” in
2015 IEEE International Symposium on Information Theory (ISIT), June 2015, pp. 1237–1241.
[43] C. Mollén, J. Choi, E. G. Larsson, and R. W. Heath, “Performance of Linear Receivers for Wideband Massive MIMO with One-Bit ADCs,” in WSA
2016; 20th International ITG Workshop on Smart Antennas, March 2016.
[44] C. Mollén, J. Choi, E. G. Larsson, and R. W. Heath, “One-bit ADCs in wideband massive MIMO systems with OFDM transmission,” in 2016 IEEE
International Conference on Acoustics, Speech and Signal Processing (ICASSP), March 2016, pp. 3386–3390.
[45] C. Mollén, J. Choi, E. G. Larsson, and R. W. Heath, “Uplink Performance of Wideband Massive MIMO With One-Bit ADCs,” IEEE Transactions on
Wireless Communications, vol. 16, no. 1, pp. 87–100, Jan 2017.
[46] Y. Li, C. Tao, L. Liu, G. Seco-Granados, and A. L. Swindlehurst, “Channel estimation and uplink achievable rates in one-bit massive MIMO systems,”
in 2016 IEEE Sensor Array and Multichannel Signal Processing Workshop (SAM), July 2016.
[47] Y. Li, C. Tao, L. Liu, A. Mezghani, and A. L. Swindlehurst, “How much training is needed in one-bit massive mimo systems at low snr?,” in 2016
IEEE Global Communications Conference (GLOBECOM), Dec 2016.
[48] Y. Li, C. Tao, G. Seco-Granados, A. Mezghani, A. L. Swindlehurst, and L. Liu, “Channel Estimation and Performance Analysis of One-Bit Massive
MIMO Systems,” IEEE Transactions on Signal Processing, vol. 65, no. 15, pp. 4075–4089, August 2017.
[49] C. Stöckle, J. Munir, A. Mezghani, and J. A. Nossek, “Channel estimation in massive MIMO systems using 1-bit quantization,” in 2016 IEEE 17th
International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), July 2016.
[50] B. Rassouli, M. Varasteh, and D. Gunduz, “Gaussian Multiple Access Channels with One-Bit Quantizer at the Receiver,” 2017, Available:
https://arxiv.org/abs/1703.02324.
[51] I. D. O’Donnell and R. W. Brodersen, “An Ultra-Wideband Transceiver Architecture for Low Power, Low Rate, Wireless systems,” IEEE Trans. on
Vehicular Technology, vol. 54, no. 5, pp. 1623–1631, September 2005.
[52] A. J. Viterbi and J. K. Omura, Principles of Digital Communication and Coding, McGraw-Hill, 1979.
[53] S. Verdú, “Spectral Efficiency in the Wideband Regime,” IEEE Trans. Inform. Theory, vol. 48, no. 6, pp. 1319–1343, June 2002.
[54] V. Prelov and S. Verdú, “Second-order Asymptotics of Mutual Information,” IEEE Trans. Inform. Theory, vol. 50, no. 8, pp. 1567–1580, August 2004.
[55] C. Rao and B. Hassibi, “Analysis of multiple antenna wireless links at low SNR,” IEEE Trans. Inform. Theory, vol. 50, no. 9, pp. 2123–2130, September
2004.
[56] X. Wu and R. Srikant, “MIMO Channels in the Low SNR Regime: Communication Rate, Error Exponent and Signal Peakiness,” in IEEE Information
theory workshop, San Antonio, TX, 2004.
[57] V. Sethuraman, Ligong Wang, B. Hajek, and A. Lapidoth, “Low-SNR Capacity of Noncoherent Fading Channels,” Information Theory, IEEE Transactions
on, vol. 55, no. 4, pp. 1555–1574, April 2009.
[58] G. Durisi, U.G. Schuster, H. Bolcskei, and S. Shamai, “Noncoherent Capacity of Underspread Fading Channels,” Information Theory, IEEE Transactions
on, vol. 56, no. 1, pp. 367–395, January 2010.
[59] L. Zheng and D. N. C. Tse, “Communication on the Grassman manifold: A geometric approach to the non-coherent multiple-antenna channel,” IEEE
Trans. Inform. Theory, vol. 48, no. 2, pp. 359–383, February 2002.
[60] T. Koch and A. Lapidoth, “Increased Capacity per Unit-Cost by Oversampling,” September 2010, Available: http://arxiv.org/abs/1008.539.
[61] T. M. Cover and J. A. Thomas, Elements of Information Theory, John Wiley and Son, New York, 1991.
[62] Ralph Hoyt Bacon, “Approximations to Multivariate Normal Orthant Probabilities,” The Annals of Mathematical Statistics, vol. 34, no. 1, pp. 191–198,
March 1963.
[63] T.L. Marzetta and B. M. Hochwald, “Capacity of a mobile multiple-antenna communication link in Rayleigh flat fading,” IEEE Trans. Inform. Theory,
vol. 45, no. 1, pp. 139–157, January 1999.
[64] I. E. Telatar and D. N. Tse, “Capacity and mutual information of wideband multipath fading channels,” IEEE Trans. Inform. Theory, vol. 46, no. 4, pp.
1384–1400, July 2000.
| 7cs.IT
|
Power Systems Data Fusion based on
Belief Propagation
Francesco Fusco, Seshu Tirupathi and Robert Gormally
arXiv:1705.08815v1 [stat.ML] 24 May 2017
IBM Research Ireland
Dublin, Ireland
{francfus,seshutir,robertgo}@ie.ibm.com
Abstract—The increasing complexity of the power grid, due
to higher penetration of distributed resources and the growing
availability of interconnected, distributed metering devices requires novel tools for providing a unified and consistent view
of the system. A computational framework for power systems
data fusion, based on probabilistic graphical models, capable
of combining heterogeneous data sources with classical state
estimation nodes and other customised computational nodes, is
proposed. The framework allows flexible extension of the notion
of grid state beyond the view of flows and injection in bus-branch
models, and an efficient, naturally distributed inference algorithm
can be derived. An application of the data fusion model to the
quantification of distributed solar energy is proposed through
numerical examples based on semi-synthetic simulations of the
standard IEEE 14-bus test case.
I. I NTRODUCTION
The electrical grid is going through a significant transformation towards a more distributed architecture for demand-supply
balancing, due to a higher penetration of distributed sources of
renewable generation, storage and demand flexibility. Internetof-Things (IOT) technologies are an integral part of the
transformation, with energy utilities availing of more and more
highly-distributed intelligent devices which produce an everincreasing amount of heterogeneous data significantly different
in terms of format, resolution and quality [1], [2].
Effective management and coordination of such increased
complexity requires scalable ingestion and fusion of all available data sources, from traditional supervisory control and data
acquisition (SCADA) systems to advanced metering infrastructure (AMI) and machine learning models based on highresolution weather data and forecasts [3], [4]. In order to make
best use of all the available information and support robust
advanced analytics and effective data-driven decision making,
one of the key challenges is to obtain a unified, consistent
view of the electrical grid system as a whole from the limited
views, often overlapping but at times inconsistent, offered by
the aforementioned set of data sources.
Traditionally, power systems state estimation is the tool
designed to provide such a unified view of the grid for
specific automation and control purposes [5], [6]. Extended
state estimation has been proposed in the past to deal with
the identification of unknown parameters of the bus-branch
model [6]. Various studies also proposed the idea of combining
different data sources, such as smart meters or machine
learning models, with the objective of providing regularization
or pseudo-measurements to overcome unobservability [7], [8].
More recently, computational methods bades on probabilistic
graphical models and belief propagation were studied to offer
a naturally distributed solution of state estimation [9]–[11].
In this context, the state variable was designed to represent
uniquely the set of voltages, flows and injections in a branchbus model of the grid. Such view is, however, becoming
obsolete as it does not provide sufficient observability into
the system for both operation and planning purposes. As an
example, system operators are more and more in need to
quantify the impact of distributed resources, such as renewable
energy or demand response, at a certain feeder or substation.
Compiling such information may require combination of data
from AMI and SCADA, and necessitates of ad-hoc heuristics
and domain expertise to resolve eventual gaps and inconsistencies between the data sources. A framework for power
systems data fusion, based on probabilistic graphical models,
is therefore proposed, where the set of state variables can
be conveniently augmented as needed and a unified estimate
based on all available data sources is obtained. A set of
diverse computational nodes for each data source, including
but not limited to classical state estimation algorithms, can
be efficiently combined in a plug and play fashion, so that
existing software modules can be reused and the system can
be extended in a modular way.
After a description of the problem, in section II, the
computational framework is detailed in section III. The main
technical contribution of the paper, compared to existing methods proposed in [9]–[11], is the derivation of a more general
Gaussian belief propagation algorithm supporting multivariate,
non-linear nodes, as described in section III-B Some numerical
examples are proposed in section IV.
II. P ROBLEM S TATEMENT
Consider a set of data sources, yi ∈ Rmi , with i = 1, . . . M ,
which may be related to different aspects of the power
grid, described by a set of state variables xi ∈ Rni , with
i = 1, . . . M , as follows:
yi = fi (xi ) + εi
i = 1, 2, . . . M.
(1)
In (1), fi (·) are generally non-linear, possibly known, relations
between the data and the state variables, while εi represents
the uncertainty in such relations or the noise in the data.
In power systems state estimation, a relation of the type
in (1) is used to relate the data available from SCADA to a
state variable, usually defined as the set of voltage magnitudes
and phase angles at the network buses in a bus-branch model,
through power flow equations [5], [6].
Recent trends have seen the growth of distributed, interconnected sensor devices, such as AMI. As a result, metering
data of electrical consumption and distributed energy resources
at individual premises, including generation but also controllable devices such as thermostats [12], are increasingly made
available. Extensions to the traditional notion of power system
state, in the sense of set of physical quantities fully characterizing the behaviour of the system, beyond the abstraction
of bus injections, can therefore be considered. For example,
the fundamental contributions of solar generation or electric
storage to a bus injection or the effects of demand response on
the voltage at a bus, can also be of interest. A relation of the
type in (1), where xi represents the distributed generation from
solar systems or the energy due to controllable thermal loads,
could therefore be available. A functional relation fi (·) can be
a simple identity where such resources are directly metered, or
more complex data-driven or physical models where necessary
(see [12] for a model of distributed thermal loads).
Yet another source of data, of the type in (1), can come
from machine learning models of energy resources (demand,
generation, etc.) based on weather, calendar and behavioural
data. The importance of statistical modelling has been widely
recognized in the context of energy forecasting for market
bidding and network operation [13]. Latest trends have seen
the need for obtaining forecasts at much finer spatial resolution due to an increased penetration of distributed resources
[3]. Statistical load models have also been proposed in the
context of load profiling or pseudo-measurement generation
for unobservable state estimation problems [7], [8].
The individual functions of state estimation, AMI data
processing and machine learning can all benefit from each
other in order to provide a more accurate, holistic view of
the system, summarized by an appropriately defined set of
state variables. By finding a mathematical relation of the
type g(x) = 0, where x = {x1 , x2 , . . . xM }, relating the
state variables seen by the different data sources, one could
formalise the problem as solution of:
y = f (x) + ε
(2)
0 = g(x) + ε0 .
(3)
The problem in (2) and (3) will be referred to as the data fusion
problem. Least-squares solutions, based on a Gauss-Newton
iterative scheme, could be used. However, scalability can be
an issue when the number of state variables or data sources
grow. In addition, such a monolithic approach is not flexible,
and the solution algorithm to the data fusion problem needs
to be changed whenever a new data source or state variable is
added or removed. By exploiting the natural factorization of
the problem, a more flexible approach based on factor graphs
and belief propagation is proposed in section III.
III. M ETHODOLOGY
A natural way to combine heterogeneous set of models like
the ones described in section II is provided by factor graphs.
As a quite general class of graphical models, factor graphs
provide a way to link potentially very different models together
in a principled fashion that respects the roles of probability
theory. An inference and learning algorithm can be obtained
by linking together the modules and associated algorithms,
in a plug and play fashion [14]. A brief introduction of
factor graphs is provided in section III-A (see [15]–[17] for
more details). A general inference algorithm based on belief
propagation is then derived in section III-B.
A. Factor Graphs
Given a set of variables, Z = {z1 , z2 , . . . , zp }, factor
graphs are a mathematical tool which can be used to express
factorizations of the form:
Φ(z1 , z2 , . . . , zp ) =
q
Y
φj (Zj ),
(4)
j=1
where φj (Zj ) is a factor defined on a subset of the variables
Zj ∈ Z. Graphically, (4) is represented by a node for each
variable zi (a variable node), and by a node for each factor φj
(a factor node). An edge between a factor node and a variable
node exists for each zi ∈ Zj .
In the particular setting described in section II, the set of
variables Z = {X , Y} can be specified by a set of n random
state variables X = {x1 , x2 , . . . , xn } and by a set of m independent sources of noisy observations Y = {y1 , y2 , . . . , ym }.
The joint probability distribution over all the variables can
then be factorized as:
p(X , Y) =
m
Y
p(yi |Xi ),
(5)
i=1
where p(·) denotes a probability density function, p(y|X )
denotes conditional probability and Xi ∈ X . With reference to
the data fusion model defined in section II, the measurement
models in (2) define the conditional densities p(yi |Xi ), while
the constraints in (3) can also be represented as conditional
probabilities where yi = 0.
The objective of the data fusion model is to obtain an
estimate of the posterior distribution of the unknown state
variables with respect to a realization of the measurement data,
which means solving the following inference problem:
p(X |Y) =
p(Y|X )p(X )
.
p(Y)
(6)
B. Inference Algorithm based on Belief Propagation
Belief propagation, or the sum-product algorithm, reduces
the inference problem in (6) to the computation of localised
messages between factor and variable nodes, along each edge
of the factor graph defined by the factorization of the density
function in (5).
The messages are real-valued functions expressing the influence between variables. Messages from variables to factors
are given by [17]:
Y
µxi →φj (xi ) =
µφk →xi (xi ),
(7)
where Ki is the set of factors fi connected to the variable xi .
Messages from factor j to variable i are, on the other hand,
calculated based on:
X jk
hfj →xi = hj −
Jj (Jxk →fj + Jjkk )−1 (hxk →fj + hkj )
φk ∈{ne(xi )\φj }
k∈Kj \i
where ne(x) denotes the set of factors connected to the variable x. Note that in (7) the dependency on y was dropped since
it will be treated as a known quantity, given that it represents
evidence (measurement data), as described in section III-A.
Messages from factors to variables are given by [17]:
Z
Y
µφj →xi (xi ) =
φj (Xj )
µxk →φj (xj ) .
w∈{Xj \xi }
xk ∈{ne(φj )\xi }
(8)
Once all the messages are updated, variables marginalization
is computed as [17]:
Y
p(xi ) =
µφk →xi (xi ).
(9)
φk ∈ne(xi )
Based on (7), (8) and (9), the inference is reduced to a sum
(integral) of products of simpler terms than the ones in the full
joint distribution, therefore the name sum-product algorithm.
Computing the integral in (8) is not feasible, in general,
for continuous distribution and it can become numerically
intractable for discrete distributions. By further assuming
that the measurement factions and the state variables in (1)
represent Gaussian densities, the factorization in (5) can be
expressed as follows:
m
Y
>
1
(10)
p(X , Y) ∝
e− 2 [yi −fi (Xi )] Ri [yi −fi (Xi )] ,
i=1
where ∝ denotes proportionality and Ri is the covariance
matrix of the error term εi in (1).
Under the factorization in (10), the messages in (7) and (8)
are also Gaussian distributions defined, net of a normalizing
constant, by a mean vector and a covariance matrix. The sumproduct algorithm is then reduced to a set of simple linear
algebra calculations.
In particular, the algorithm can be derived by exploiting
the canonical form of a Gaussian distribution and linearizing
around an operational point:
1
e− 2 [yi −fi (Xi )]
>
Ri [yi −fi (Xi )]
1
>
∝ e− 2 δXi
Ji δXi +δXi> hi
,
(11)
Fi> Ri−1 Fi ,
where Ji =
with Fi being the Jacobian of fi (Xi )
with respect to a point Xi = X i , and hi = Fi> Ri−1 (yi −
fi (X i )). Note that, in (11), the dependency from the constant
term due to yi − fi (Xi ) has been dropped.
At a given iteration, t, assuming an estimate xt is available,
then it is possible to compute the Jj and hj for every factor,
based on (11). Messages from variable xi to factor fj can then
be computed as:
X
hxi →fj =
hfk →xi
(12)
k∈Ki \j
Jxi →fj =
X
k∈Ki \j
(14)
Jfj →xi = Jj −
X
Jjjk (Jxk →fj
+
Jjkk )−1 Jjkj ,
(15)
k∈Kj \i
where Jjjk is the block of the Jj matrix with rows corresponding to the variable xj and columns corresponding to
variable xk . Similarly, hkj denotes the block of the hj vector
corresponding to the variable xj .
All messages (12) to (15) can be computed by starting from
the leaf factor nodes and iteratively updating all messages
where the required input messages have been processed. Once
messages from all incoming factors are available, variable
marginals can be updated with the following iterative scheme:
xt+1
= xti + δxi
i
(16)
!−1
δxi =
X
!
X
Jfk →xi
k∈Ki
hfk →xi
,
(17)
k∈Ki
with the corresponding covariance matrix given by:
!−1
X
t+1
Sxi =
Jfk →xi
.
(18)
k∈Ki
If the graph is not a tree, a loopy version of the proposed belief
propagation algorithm can be derived to iteratively converge
to a solution for the marginals [10]. For a full derivation of
the fundamental Gaussian belief propagation messages in the
linear case, the reader can refer to [15], [16]. The proposed
algorithm is a generalization to non-linear factors and it is
actually equivalent, in the case of trees, to a distributed
implementation of the Gauss-Newton algorithm.
In the context of power systems, the belief propagation
algorithm proposed in [9], [10] was limited to the solution
of state estimation problems. To the best of the authors
knowledge, a belief propagation algorithm for non-linear AC
power flow models was only studied in [11]. The latter,
however, involved a fully distributed algorithm on scalar factor
nodes, resulting in heavily loopy graphs and requiring unclear
heuristic rules to converge even for simple problems. The
algorithm derived here can regarded as a generalization of the
mentioned studies where complete state estimation problems
(or parts of) represent individual factor nodes, thus reducing
the number of possible loops even for very large and complex
grid topologies. Additional non-linear measurement functions
of any desired state variable can also be integrated into the
model in a plug and play fashion.
IV. R ESULTS
A. Available Data
Jfk →xi ,
(13)
Experiments were designed based on the standard 14-bus
IEEE test case, shown in Fig 1. The grid was simulated with
p(y1 |x1 )
x1
p(x1 , x2 )
x2
p(y3 |x2 )
p(y2 |x2 )
Injection
Energy Demand
Solar Generation
0.6
0.3
0
Aug 2
Load
Solar Generation
Energy Demand
100
90
Aug 4
Aug 6
Aug 8, 2015
Voltage Magnitude [p.u.]
Fig. 1. The IEEE 14 Bus Test Case [19].
Active Power [p.u.]
Fig. 3. Factor graph representation of the power systems data fusion model
used in the experiments. Squares denote factors, circles are variable nodes.
1.03
1.02
1.01
1
Aug 2
Aug 4
Aug 6
Aug 8
(a)
80
[GWh/h]
70
60
50
40
30
20
(b)
10
0
Aug 2, 2015 (Sun)
Aug 4
Aug 6
Aug 8, 2015 (Sat)
Fig. 2. Hourly Germany load and solar generation data from Aug 2, 2015 to
Aug 8, 2015 [18].
time series data obtained from the Open Power Systems Data
Platform [18]. In particular, national load and solar generation
data from Germany from Jan 1, 2014 to Aug 8, 2015,
sampled at hourly resolution were considered. Figure 2 shows
a snapshot of the last week of the data, and also a calculation
of the electrical demand from the net electrical load and the
solar photovoltaic generation data. The additional component
of the residual load, due to distributed wind generation, was
not considered, although significant in the case of Germany,
since renewable energy modelling is not the focus of the study.
Weather data were used in order to design the machine
learning models. The Weather Company’s Cleaned Observations application programming interface (API) [20] provides
cleaned, interpolated, historical observations of a wide variety
of weather variables. Version two of the API was used to
retrieve hourly historical observations of surface temperature,
diffuse horizontal radiation, direct normal irradiance, and
downward solar radiation at 52.52N, 13.61E (Berlin) from Jan
1, 2014, to Aug 8, 2015. A more rigorous choice would have
been to use a combination of points appropriately distributed
across Germany. A simple approach was however favoured as
it did not impact the objectives of the proposed experiments.
B. Experimental Setup for the Data Fusion Model
Based on the load and solar generation profiles shown in
Fig. 2, a realistic simulation of the 14-bus system is provided
as follows. The active injection at each network bus provided
Fig. 4. Simulated data sources: (a) Active power injection, energy demand,
solar generation and voltage magnitude at bus 4; (b) Forecasts of energy
demand and solar generation at bus 4, with the 95% confidence interval.
by the standard model in [19] is taken to represent the peak
load over the validation week (Aug 2 to Aug 8, 2015). The
hourly profile is then modulated according to the load profile.
It is further assumed that the solar generation is uniformly
spread across the load buses 3 to 6 and 9 to 14, such to obtain
the demand and generation components of the active load.
The data fusion model was then built based on the assumptions that data were generated from three different systems,
producing the factor graph shown in Figure 3. Details of the
factor nodes are outlined in the following.
1) State Estimation Factor Node: A set of data, y1 , is
generated based on a power flow simulation of the 14-bus
system. It is assumed that measurements are available for
all active/reactive power injections and voltage magnitudes.
Measurement data were generated by simulating the 14-bus
power flow using Matpower [21] and the available load profiles. White Gaussian noise with standard deviation of 0.01 p.u.
for the power measurements and of 0.5e−3 p.u. for the voltage
magnitudes was introduced. Figure 4(a) shows an example of
active power and voltage magnitude measurements generated
at bus 4. The state variables, x1 , were defined as the set of
bus voltages and measurement model, y1 = f1 (x1 ) + ε1 , was
based on the usual AC power flow equations [5].
2) Smart Meters Factor Node: Smart meter data, y2 , were
assumed to provide total energy demand and solar generation
at the 10 load buses of the system, buses 3 to 6 and 9 to
14. White Gaussian noise with standard deviation of 0.02 p.u.
was introduced. Figure 4(a) shows an example of the energy
D = s1 (hour, daytype) + s2 (Tm ) + s3 (Tmax ) + ζ
Solar Genertaion [p.u]
Energy Demand [p.u.]
0.5
0.45
0.4
0.35
0.1
0.3
0.25
Aug 7
0
Aug 7
Abs. Error [p.u.]
Aug 8
0.04
0.02
0
Aug 2
Aug 4
Aug 6
Aug 8
Aug 8
Raw Data
Fused Data
0.04
0.02
0
Aug 2
Aug 4
Aug 6
Aug 8
Fig. 5. Comparison between fused data and raw data of energy demand and
solar genertaion at bus 4.
(19)
where D is the energy demand at a given time instant, daytype
is a categorical variable representing each day of a week, hour
is the numerical variable representing the hour in a day, Tm is
the mean temperature for the day and Tmax is the maximum
temperature of the day. The basis functions si (u) are smooth
splines based on cubic B-splines. The solar generation forecast
model was considered to be a linear model with normal and
diffuse irradiance data as covariates.
The models were trained on the period from January 1, 2014
to August 1, 2015, and a sample output for bus 4 is shown in
Fig. 4(b). A mean absolute percentage error for the demand
forecasts of about 4.3% was observed on the validation week
between August 2 and August 8, 2015. For the solar generation
forecasts, a 10.4% error was observed (normalized on the daily
peak of generation to avoid numerical issues when generation
is near the zero). Improved demand models could be designed
by considering a more extensive set of features. Similarly,
better solar generation forecasts can be obtained, for example
by exploiting non-linead physics-based models [3]. However
the focus of the paper is not on energy forecasting and such
alternatives were not pursued.
4) Joint Factor Node: A joint factor relating the state
variables x1 and x2 was designed as:
0 = f41 (x1 ) − F42 x2 + ε4 ,
True Value
Raw Data
Fused Data
0.6
0.55
Abs. Error [p.u.]
demand and generation data simulated at bus 4. The hidden
state variable, x2 , was defined as representing the amount of
energy demand and solar generation at each bus, so that the
relation in (1) is an identify matrix, namely y2 = x2 + ε2 .
3) Machine Learning Factor Node: Forecasting models
providing data, y3 , of energy demand and solar generation at
the 10 load buses specified in the state variable x2 were also
considered, so that the same model as for the smart meter
factor node, y3 = x2 + ε3 , was utilized.
The forecasts were computed using generalized additive
models (GAMs) based on penalized iteratively re-weighted
least squares method [22], [23]. Extensive research has gone
into studying the covariates that are important to model
electricity (see [13], [22], [24] and references within). The
energy demand model considered in this study is given by:
(20)
where f41 (x1 ) represents the active injection equations of the
AC power flow model, and F42 is a matrix where each row has
a +1 and a −1 in correspondence, respectively, of the demand
and generation entries in the state variable x2 at the specific
network bus. The Gaussian random error ε4 ∼ N(0, R4 ) is
chosen such thah R4 is a diagonal matrix with very small
diagonal entries of 1e − 10.
C. Experimental results
Since the factor graph contains no loops, as noted in
section III-B, the proposed inference algorithm is equivalent
to solving the global problem (2)-(3) with the Gauss-Newton
method. However, instead of computing one matrix inversion
of dimension n1 + n2 , four matrix inversions of dimension
Fig. 6. Estimation of the active power injection at bus 4, with loss of
observability in the state estimation node on August 7. The confidence interval
before Aug 7 is not visible as much smaller than after loss of observability.
n1 and n2 are performed at every iteration, based on (15) and
(17). As the number and dimensionality of the nodes increases,
significant computational gain can be obtained.
In a first experiment, the proposed inference algorithm
was run on the simulated data over the validation week.
Convergence was achieved in less then 10 steps for all 168
hourly samples. As expected, fusion of the three data sources
allows reduction of uncertainty in estimating the true quantities
with respect to using the information from the individual
sources alone. As shown in Figure 5, the estimation of energy
demand and solar generation at the bus 4, which is only
metered directly from the smart meter data, also benefits from
fusion with the state estimation node, and shows lower error
than when using the raw data only. Although not shown here,
the improvement is consistent across all the network buses.
In a second experiment, loss of observability on the last
2 days of the validation period is considered in the state
estimation node. In particular, voltages, active/reactive power
and smart meter data are assumed unavailable from network
buses 3, 4, 9 and 10 (bottom right of network diagram in
Fig. 1). It can be shown that in this case the state estimation
node has a singularity, but the belief propagation algorithm
does not experience invertibility issues since, through (13) and
(15), implicitly applies a regularization based on information
available from the energy forecasts node. Fused information
is still obtained, although with higher uncertainty than when
all data are available, as shown in Fig. 6 for bus 4.
A final experiment shows how inconsistencies in a data
source can be detected through data fusion. It is assumed
that new solar installations cause an increase of 50% capacity
R EFERENCES
Fig. 7. Fused data and true value of solar generation at bus 4, when new
solar capacity becomes available on August 7 but is not reflected in raw data.
at bus 4 in the last 2 days of the validation period, and
that this is not reflected in the smart meter data (actually, a
typical real-world scenario). Figure 7 shows how the fused
estimate of solar generation at bus 4 is in larger than the raw
data and closer to the true value by an amount comparable
to the confidence interval. The possible inconsistency can
be formally detected using statistical testing. Obviously such
information comes from the grid data which are affected by
the increased solar generation at the bus.
V. C ONCLUSION
In this study, a novel computational framework for power
systems data fusion was proposed. Based on probabilistic
graphical models, heterogeneous data sources and related
measurement models (physical laws or data-driven machine
learning models) can be combined such that a consistent
and unified view of the system is obtained. An efficient and
naturally distributed inference algorithm based on Gaussian
belief propagation is also derived.
In a set of numerical examples it was demonstrated how the
traditional notion of state can be extended to provide visibility
into the amount of solar generation at the buses of a network
model. The fused information reduces the effect of noise in the
various data sources and occurrences of missing or erroneous
data are easily overcome and diagnosed.
While fully known measurement functions were assumed,
future work will investigate scenarios where they are partly
unknown and need to be learned from the data. Further
possibility for extensions of the state variable to include,
for example, effects of temperature variation or of demand
response programs on the active injections will also be investigated. Since the data consistency is one the key benefits of
data fusion, extensions of traditional bad data analysis to the
proposed framework is also of interest for future studies.
ACKNOWLEDGMENTS
This project has received funding from the European Research Council (ERC) under the European Union’s Horizon
2020 research and innovation programme (grant agreement no.
731232).
[1] N. Yu, S. Shah, R. Johnson, R. Sherick, M. Hong, M. Ieee, and
K. Loparo, “Big Data Analytics in Power Distribution Systems,” in
Proceedings of the Innovative Smart Grid Technologies Conference,
Washington, DC, USA, 2015.
[2] S. E. Collier, “The emerging enernet: Convergence of the smart grid with
the internet of things,” IEEE Industry Applications Magazine, vol. 23,
no. 2, 2017.
[3] V. P. A. Lonij, J.-B. Fiot, B. Chen, F. Fusco, P. Pompey, Y. Gkoufas,
M. Sinn, D. Tougas, M. Coombs, and A. Stamp, “A scalable demand
and renewable energy forecasting system for distribution grids,” IEEE
Power and Energy Society General Meeting (PESGM), 2016.
[4] F. Fusco, U. Fischer, V. Lonij, P. Pompey, J.-B. Fiot, B. Chen, Y. Gkoufas, and M. Sinn, “Data management system for energy analytics
and its application to forecasting,” Proceedings of the Workshops of the
EDBT/ICDT Joint Conference 2016, 2016.
[5] A. Abur and A. G. Exposito, Power System State Estimation: Theory
and Implementation. CRC Press, 2004.
[6] A. Monticelli, “Electric power system state estimation,” Proceedings of
the IEEE, vol. 88, no. 2, pp. 262–282, feb 2000.
[7] E. Manitsas, R. Singh, B. Pal, and G. Strbac, “Modelling of PseudoMeasurements for Distribution System State Estimation,” in CIRED
Seminar 2008: SmartGrids for Distribution, 2008.
[8] J. Wu, Y. He, and N. Jenkins, “A Robust State Estimator for Medium
Voltage Distribution Networks,” IEEE Trans. on Power Systems, 2012.
[9] Y. Hu, A. Kuh, A. Kavcic, and T. Yang, “A Belief Propagation
Based Power Distribution System State Estimator,” IEEE Computational
Intelligence Magazine, no. AUGUST, pp. 36–46, 2011.
[10] M. Cosovic and D. Vukobratovic, “State estimation in electric power
systems using belief propagation: An extended dc model,” IEEE International workshop on Signal Processing advances in Wireless Communications (SPAWC), 2016.
[11] ——, “Distributed gauss-newton method for ac state estimation: A belief
propagation approach,” IEEE International Conference on Smart Grid
Communications, 2016.
[12] D. S. Callaway and I. A. Hiskens, “Achieving controllability of electrical
loads,” Proceedings of the IEEE, 2011.
[13] R. Weron, Modeling and forecasting electricity loads and prices: A
statistical approach. John Wiley & Sons, 2007, vol. 403.
[14] B. J. Frey and N. Jojic, “A comparison of algorithms for inference
and learning in probabilistic graphical models,” IEEE Transactions on
Pattern Analysis and Machine Intelligence, vol. 27, no. 9, 2005.
[15] H.-A. Loelinger, “An introduction to factor graphs,” IEEE Signal Processing Magazine, 2004.
[16] H.-A. Loelinger, J. Dauwels, J. Hu, S. Korl, L. Ping, and F. R. Kschischang, “The factor graph approach to model-based signal processing,”
Proceedings of the IEEE, 2007.
[17] D. Barber, Bayesian Reasoning and Machine Learning. Cambridge
University Press, 2012.
[18] Open Power System Data, “Data package time series,” 2017.
[Online]. Available: http://data.open-power-system-data.org/time series/
2017-03-06/
[19] University of Washington, “Power Systems Test Case Archive,” 2013.
[Online]. Available: http://www.ee.washington.edu/research/pstca/
[20] The Weather Company, “Cleaned observations api,” 2017.
[Online]. Available: http://cleanedobservations.wsi.com/documents/
WSI%20Cleaned%20Observations%20API%20Documentation.pdf
[21] R. D. Zimmerman, C. E. Murillo-Snchez, and R. J. Thomas, “Matpower:
Steady-state operations, planning and analysis tools for power systems
research and education,” Power Systems, IEEE Transactions on, vol. 26,
no. 1, pp. 12–19, 2011.
[22] A. Ba, M. Sinn, Y. Goude, and P. Pompey, “Adaptive Learning of
Smoothing Functions : Application to Electricity Load Forecasting,”
in Proceedings of the Neural Information Processing Systems (NIPS)
Conference, Lake Tahoe, Nevada, USA, 2012.
[23] S. N. Wood, Y. Goude, and S. Shaw, “Generalized additive models
for large data sets,” Journal of the Royal Statistical Society: Series C
(Applied Statistics), vol. 64, no. 1, pp. 139–155, 2015.
[24] A. Pierrot and Y. Goude, “ Short-Term Electricity Load Forecasting With
Generalized Additive Models ,” in 16th Intelligent System Applications
to Power Systems Conference (ISAP), 2011.
| 3cs.SY
|
Resilience: A Criterion for Learning in the Presence of Arbitrary Outliers
Jacob Steinhardt∗
Moses Charikar†
Gregory Valiant‡
JSTEINHA @ STANFORD . EDU
MOSES @ STANFORD . EDU
VALIANT @ STANFORD . EDU
arXiv:1703.04940v3 [cs.LG] 27 Nov 2017
Abstract
We introduce a criterion, resilience, which allows properties of a dataset (such as its mean or best low rank
approximation) to be robustly computed, even in the presence of a large fraction of arbitrary additional
data. Resilience is a weaker condition than most other properties considered so far in the literature, and
yet enables robust estimation in a broader variety of settings. We provide new information-theoretic results
on robust distribution learning, robust estimation of stochastic block models, and robust mean estimation
under bounded kth moments. We also provide new algorithmic results on robust distribution learning, as
well as robust mean estimation in `p -norms. Among our proof techniques is a method for pruning a highdimensional distribution with bounded 1st moments to a stable “core” with bounded 2nd moments, which
may be of independent interest.
Keywords: robust learning, outliers, stochastic block models, p-norm estimation, low rank approximation
1. Introduction
What are the fundamental properties that allow one to robustly learn from a dataset, even if some fraction
of that dataset consists of arbitrarily corrupted data? While much work has been done in the setting of
noisy data, or for restricted families of outliers, it is only recently that provable algorithms for learning
in the presence of a large fraction of arbitrary (and potentially adversarial) data have been formulated in
high-dimensional settings (Klivans et al., 2009; Xu et al., 2010; Diakonikolas et al., 2016a; Lai et al., 2016;
Steinhardt et al., 2016; Charikar et al., 2017). In this work, we formulate a conceptually simple criterion
that a dataset can satisfy–resilience–which guarantees that properties such as the mean of that dataset can
be robustly estimated even if a large fraction of additional arbitrary data is inserted.
To illustrate our setting, consider the following game between Alice (the adversary) and Bob. First, a
set S ⊆ Rd of (1 − )n points is given to Alice. Alice then adds n additional points to S to create a new
set S̃, and passes S̃ to Bob. Bob wishes to output a parameter µ̂ that is as close as possible to the mean µ of
the points in the original set S, with error measured according to some norm kµ̂ − µk. The question is: how
well can Bob do, assuming that Alice is an adversary with knowledge of Bob’s algorithm?
The above game models mean estimation in the presence of arbitrary outliers; one can easily consider
other problems as well (e.g. regression) but we will focus on mean estimation in this paper.
With no asumptions on S, Bob will clearly incur arbitrarily large error in the worst case—Alice can
add points arbitrarily far away from the true mean µ, and Bob has no way of telling whether those points
actually belong to S or were added by Alice. A first pass assumption is to suppose that S has diameter
at most ρ; then by discarding points that are very far away from most other points, Bob can obtain error
O(ρ). However, in most high-dimensional settings, the diameter ρ grows polynomially with the dimension
∗
†
‡
JS was supported by a Fannie & John Hertz Foundation Fellowship, an NSF Graduate Research Fellowship, and a Future of
Life Instute grant.
MC was supported by NSF grants CCF-1617577, CCF-1302518 and a Simons Investigator Award.
GV was supported by NSF CAREER Award CCF-1351108 and a Sloan Research Fellowship.
Figure 1: Illustration of the robust mean estimation setting. First, a set of points (blue) is given to Alice,
who adds an fraction of adversarially chosen points (red). Bob’s goal is to output the mean of
the original set (indicated in green).
√
d (e.g. the d-dimensional hypercube has `2 -diameter Θ( d)). Subtler criteria are therefore needed to obtain
dimension-independent bounds in most settings of interest.
p
Recently, Diakonikolas et al. (2016a) showed that Bob can incur `2 -error O( log(1/)) when the
points in S are drawn frompa d-dimensional Gaussian, while Lai et al. (2016) concurrently showed that
Bob can incur `2 -error O( log(d)) if the points in S are drawn from a distribution with bounded 4th
moments. Since then, a considerable amount of additional work has studied high-dimensional estimation in
the presence of adversaries, which we discuss in detail below. However, in general, both Bob’s strategy and
its analysis tend to be quite complex, and specialized to particular distributional assumptions. This raises the
question—is it possible to formulate a general and simple-to-understand criterion for the set S under which
Bob has a (possibly inefficient) strategy for incurring small error?
In this paper, we provide such a criterion; we identify an assumption–resilience–on the set S, under
which Bob has a straightforward exponential-time algorithm for estimating µ accurately. This yields new
information-theoretic bounds for a number of robust learning problems, including robust learning of stochastic block models, of discrete distributions, and of distributions with bounded kth moments. We also identify
additional assumptions under which Bob has an efficient (polynomial-time) strategy for estimating µ, which
yields an efficient algorithm for robust learning of discrete distributions, as well as for robust mean estimation in `p -norms.
The resilience condition is essentially that the mean of every large subset of S must be close to the mean
of all of S. More formally, for a norm k · k, our criterion is as follows:
Definition 1 (Resilience) A set of points {xi }i∈S lying in Rd is (σ, )-resilient in a norm k · k around a
point µ if, for all subsets T ⊆ S of size at least (1 − )|S|,
1 X
(xi − µ) ≤ σ.
|T |
(1)
i∈T
More generally, a distribution p is said to be (σ, )-resilient if kE[x − µ | E]k ≤ σ for every event E of
probability at least 1 − .
In the definition above, µ need not equal the mean of S; this distinction is useful in statistical settings where
the sample mean of a finite set of points differs slightly from the true mean. However, (1) implies that µ
differs from the mean of S by at most σ.
2
Importantly, Definition 1 is satisfied with high probability by a finite sample in many settings. For
instance, samples from a distribution with kth moments bounded by σ will be (O(σ1−1/k ), )-resilient in
the `2 -norm with high probability. Resilience also holds with high probability under many other natural
distributional assumptions, discussed in more detail in Sections 1.1, 6, and 7.
Assuming that the original set S is (σ, )-resilient, Bob’s strategy is actually quite simple—find any
large (σ, )-resilient subset S 0 of the corrupted set S̃, and output the mean of S 0 . By pigeonhole, S 0 and S
have large intersection, and hence by (1) must have similar means. We establish this formally in Section 2.
Pleasingly, resilience reduces the question of whether Bob can win the game to a purely algorithmic
question—that of finding any large resilient set. Rather than wondering whether it is even informationtheoretically possible to estimate µ, we can instead focus on efficiently finding resilient subsets of S̃.
We provide one such algorithm in Section 4, assuming that the norm k · k is strongly convex and that we
can approximately solve a certain generalized eigenvalue problem in the dual norm. When specialized to
the `1 -norm, our general algorithm yields an efficient procedure for robust learning of discrete distributions.
In the remainder of this section, we will outline our main results, starting with information-theoretic results and then moving on to algorithmic results. In Section 1.1, we show that resilience is indeed informationtheoretically sufficient for robust mean estimation. In Section 1.2, we then provide finite-sample bounds
showing that resilience holds with high probability for i.i.d. samples from a distribution.
In Section 1.3, we turn our attention to algorithmic bounds. We identity a property–bounded variance
in the dual norm–under which efficient algorithms exist. We then show that, as long as the norm is strongly
convex, every resilient set has a large subset with bounded variance, thus enabling efficient algorithms. This
connection between resilience and bounded variance is the most technically non-trivial component of our
results, and may be of independent interest.
Both our information-theoretic and algorithmic bounds yield new results in concrete settings, which we
discuss in the corresponding subsections. In Section 1.4, we also discuss an extension of resilience to lowrank matrix approximation, which enables us to derive new bounds in that setting as well. In Section 1.5 we
outline the rest of the paper and point to technical highlights, and in Section 1.6 we discuss related work.
1.1. Information-Theoretic Sufficiency
First, we show that resilience is indeed information-theoretically sufficient for robust recovery of the mean
µ. In what follows, we use σ∗ () to denote the smallest σ such that S is (σ, )-resilient.
Proposition 2 Suppose that S̃ = {x1 , . . . , xn } contains a set S of size (1 − )n that is resilient around µ
(where S and µ are both unknown). Then if < 21 , it is possible to recover a µ̂ such that kµ̂−µk ≤ 2σ∗ ( 1−
).
1
More generally, if |S| ≥ αn (even if α < 2 ), it is possible to output a (random) µ̂ such that kµ̂ − µk ≤
α
α
16
σ
α ∗ ( 4 ) with probability at least 2 .
The first part says that robustness to an fraction of outliers depends on resilience to a 1−
fraction of
deletions. Thus, Bob has a good strategy as long as σ∗ ( 1− ) is small.
The second part, which is more surprising, says that Bob has a good strategy even if the majority of S̃
is controlled by Alice. Here one cannot hope for recovery in the usual sense, because if α = 12 (i.e., Alice
controls half the points) then Alice can make S̃ the disjoint union of two identical copies of S (one of which
is shifted by a large amount) and Bob has no way of determining which of the two copies is the true S.
Nevertheless, in this situation Bob can still identify S (and hence µ) with probability 12 ; more generally, the
second part of Proposition 2 says that if |S| = α|S̃| then Bob can identify µ with probability at least α2 .
The fact that estimation is possible even when α < 12 was first established by Steinhardt et al. (2016)
in a crowdsourcing setting, and later by Charikar et al. (2017) in a number of settings including mean
3
estimation. Apart from being interesting due to its unexpectedness, estimation in this regime has immediate
implications for robust estimation of mixtures of distributions (by considering each mixture component in
turn as the “good” set S) or of planted substructures in random graphs. We refer the reader to Charikar et al.
(2017) for a full elaboration of this point.
The proof of Proposition 2, given in detail in Section 2, is a pigeonhole argument. For the < 12 case,
we simply search for any large resilient set S 0 and output its mean; then S and S 0 must have large overlap,
and by resilience their means must both be close to the mean of their intersection, and hence to each other.
For the general case where |S| = α|S̃| (possibly with α < 21 ), a similar pigeonhole argument applies
0 . We can show
but we now need to consider a covering of S̃ by α2 approximately disjoint sets S10 , . . . , S2/α
that the true set S must overlap at least one of these sets by a decent amount, and so outputting the mean of
one of these sets at random gives a good approximation to the mean of S with probability α2 .
1.2. Finite-Sample Concentration
While Proposition 2 provides a deterministic condition under which robust mean estimation is possible, we
would also like a way of checking that resilience holds with high probability given samples x1 , . . . , xn from
a distribution p. First, we provide an alternate characterization of resilience which says that a distribution is
resilient if it has thin tails in every direction:
Lemma 3 Given a norm k · k, define the dual norm kvk∗ = supkxk≤1 hv, xi. For a fixed vector v, let τ (v)
denote the -quantile of hx − µ, vi: Px∼p [hx − µ, vi ≥ τ (v)] = . Then, p is (σ, )-resilient around its mean
µ if and only if
1−
Ep [hx − µ, vi | hx − µ, vi ≥ τ (v)] ≤
σ whenever kvk∗ ≤ 1.
(2)
In other words, if we project onto any unit vector v in the dual norm, the -tail of x − µ must have mean at
most 1−
Thus, for instance, a distribution with variance at most σ02 along every unit vector would have
σ.
√
σ = O(σ0 ). Note that Lemma 3 requires µ to be the mean, rather than an arbitrary vector as before.
We next provide a meta-result establishing that resilience of a population distribution p very generically
transfers to a finite set of samples from that distribution. The number of samples necessary depends on two
quantities B and log M that will be defined in detail later; for now we note that they are ways of measuring
the effective dimension of the space.
Proposition 4 Suppose that a distribution p is (σ, )-resilient around its mean µ with < 12 . Let B be such
that P[kx − µk ≥ B] ≤ /2. Also let M be the covering number of the unit ball in the dual norm k · k∗ .
Then, given n samples x1 , . . . , xn ∼ p, with probability 1 − δ − exp(−n/6)
there is a subset T of
q
log(M/δ)
(B/σ) log(M/δ)
(1 − )n of the xi such that T is (σ 0 , )-resilient with σ 0 = O σ · 1 +
+
.
2
n
n
Note that Proposition 4 only guarantees resilience on a (1 − )n-element subset of the xi , rather than all
of x1 , . . . , xn . From the perspective of robust estimation, this is sufficient, as we can simply regard the
remaining n points as part of the “bad” points controlled by Alice. This weaker requirement seems to
be actually necessary to achieve Proposition 4, and was also exploited in Charikar et al. (2017) to yield
improved bounds for a graph partitioning problem. There has been a great deal of recent interest in showing
how to “prune” samples to achieve faster rates in random matrix settings (Guédon and Vershynin, 2014;
Le et al., 2015; Rebrova and Tikhomirov, 2015; Rebrova and Vershynin, 2016), and we think the general
investigation of such pruning results is likely to be fruitful.
4
We remark that the sample complexity in Proposition 4 is suboptimal in many cases, requiring roughly
samples when d samples would suffice. At the end of the next subsection we discuss a tighter but more
specialized bound based on spectral graph sparsification.
Applications. Propositions 2 and 4 together give us a powerful tool for deriving information-theoretic
robust recovery results: one needs simply establish resilience for the population distribution p, then use
Proposition 4 to obtain finite sample bounds and Proposition 2 to obtain robust recovery guarantees. We
do this in three illustrative settings: `2 mean estimation, learning discrete distributions, and stochastic block
models. We outline the results below; formal statements and proofs are deferred to Section 7.
Mean estimation in `2 -norm. Suppose that a distribution on Rd has bounded kth moments: Ex∼p [|hx−
µ, vi|k ]1/k ≤ σkvk2 for all v for some k ≥ 2. Then p is (O(σ1−1/k ), )-resilient in the `2 -norm. Propo1.5
sitions 4 and 2 then imply that, given n ≥ d + d2 samples from p, and an -fraction of corruptions, it is
possible to recover the mean to `2 -error O(σ1−1/k ). Moreover, if only an α-fraction of points are good, the
mean can be recovered to error O(σα−1/k ) with probability Ω(α).
The d1.5 / term in the sample complexity is likely loose, and we believe the true dependence on d is at
most d log(d). This looseness comes from Proposition 4, which uses a naı̈ve covering argument and could
potentially be improved with more sophisticated tools. Nevertheless, it is interesting that resilience holds
long before the empirical kth moments concentrate, which would require dk/2 samples.
Distribution learning. Suppose that we are given k-tuples of independent samples from a discrete
distribution: p = π k , where π is a distribution on {1, . . . , m}. By taking the empirical average of the
k samples from π, we can treat a sample from p as an elementpin the m-dimensional simplex ∆m . This
distribution turns out to be resilient in the `1 -norm with σ = O( log(1/)/k), which p
allows us to estimate
p in the `1 -norm (i.e., total variation norm) and recover π̂ such that kπ̂ − πkT V = O( log(1/)/k). This
reveals a pleasing “error correction”
√ property: if we are given k samples at a time, either all or none of
which are good, then our error is k times smaller than if we only observe the samples individually.
Stochastic block models. Finally, we consider the semi-random stochastic block model studied in
Charikar et al. (2017) (described in detail in Section 7.3). For a graph on n vertices, this model posits
a subset S of αn “good” vertices, which are connected to each other with probability na and to the other
(“bad”) vertices with probability nb (where b < a); the connections among the bad vertices can be arbitrary.
The goal is to recover the set S.
We think of each row of the adjacency matrix as a vector in {0, 1}n , and show that for the good vertices
these vectors are resilient in a truncated `1 -norm kxk, defined
as the sum of the αn largest coordinates of
p
x (in absolute value). In this case, we have σ = O(α a log(2/α)) (this requires a separate argument
from Proposition 4 to get tight bounds). Applying Proposition 2, we find that we are able to recover (with
probability α2 ) a set Ŝ with
a log(2/α)
1
|S4Ŝ| = O
.
(3)
αn
(a − b)2 α2
d1.5
(a−b)2
log(2/α)
. Charikar
a
α2
(a−b)2
log(2/α)
, and remark
a
α3
In particular, we get non-trivial guarantees as long as
et al. (2017) derive
a weaker (but computationally efficient) bound when
on the similarity to
(a−b)2
a
1
,
α2
the famous Kesten-Stigum threshold
which is the conjectured threshold for computationally
efficient recovery in the classical stochastic block model (see Decelle et al. (2011) for the conjecture, and
Mossel et al. (2013); Massoulié (2014) for a proof in the two-block case). Our information-theoretic upper
bound matches the Kesten-Stigum threshold up to a log(2/α) factor. We conjecture that this upper bound is
tight; some evidence for this is given in Steinhardt (2017), which provides a nearly matching informationtheoretic lower bound when a = 1, b = 12 .
5
1.3. Strong Convexity, Second Moments, and Efficient Algorithms
Most existing algorithmic results on robust mean estimation rely on analyzing the empirical covariance of
the data in some way (see, e.g., Lai et al. (2016); Diakonikolas et al. (2016a); Balakrishnan et al. (2017)).
In this section we establish connections between bounded covariance and resilience, and show that in a very
general sense, bounded covariance is indeed sufficient to enable robust mean estimation.
Given a norm k · k, we say that a set of points x1 , . . . , xn has variance bounded by σ02 in that norm if
1 Pn
2
2
2
along
i=1 hxi − µ, vi ≤ σ0 kvk∗ (recall k · k∗ denotes the dual norm). Since this implies a tail bound
n
√
every direction, it is easy to see (c.f. Lemma 3) that a set with variance bounded by σ02 is (O(σ0 ), )resilient around its mean for all < 21 . Therefore, bounded variance implies resilience.
An important result is that the converse is also true, provided the norm is strongly convex. We say that
a norm k · k is γ-strongly convex if kx + yk2 + kx − yk2 ≥ 2(kxk2 + γkyk2 ) for all x, y ∈ Rd .1 As an
example, the `p -norm is (p − 1)-strongly convex for p ∈ (1, 2]. For strongly convex norms, we show that
any resilient set has a highly resilient “core” with bounded variance:
Theorem 5 If S is (σ, 21 )-resilient in a γ-strongly convex norm k · k, then S contains a set S0 of size at least
1 P
288σ 2
1
2
2
i∈S0 hxi − µ, vi ≤
2 |S| with bounded variance: |S0 |
γ kvk∗ for all v.
Using Lemma 3, we can show that (σ, 12 )-resilience is equivalent to having bounded 1st moments in every
direction; Theorem 5 can thus be interpreted as saying that any set with bounded 1st moments can be pruned
to have bounded 2nd moments.
We found this result quite striking—the fact that Theorem 5 can hold with no dimension-dependent
factors is far from obvious. In fact, if we replace 2nd moments with 3rd moments or take a non-stronglyconvex norm then the analog of Theorem 5 is false: we incur polynomial factors in the dimension even if
S is the standard basis of Rd (see Section A for details). The proof of Theorem 5 involves minimax duality
and Khintchine’s inequality. We can also strengthen Theorem 5 to yield S0 of size (1 − )|S|. The proofs of
both results are given in Section 3 and may be of independent interest.
Algorithmic results. Given points with bounded variance, we establish algorithmic results assuming that
one can solve the “generalized eigenvalue” problem maxkvk∗ ≤1 v > Av up to some multiplicative accuracy
κ. Specifically, we make the following assumption:
Assumption 6 (κ-Approximability) There is a convex set P of PSD matrices such that
sup v > Av ≤ sup hA, M i ≤ κ sup v > Av
kvk∗ ≤1
M ∈P
(4)
kvk∗ ≤1
for every PSD matrix A. Moreover, it is possible to optimize linear functions over P in polynomial time.
A result of Nesterov (1998) implies that this is true with κ = O(1) if k · k∗ is any “quadratically convex”
norm, which includes the `q -norms for q ∈ [2, ∞]. Also, while we do not use it in this paper, one can
sometimes exploit weaker versions of Assumption 6 that only require supM ∈P hA, M i to be small for certain
matrices A; see for instance Li (2017), which obtains an algorithm for robust sparse mean estimation even
though Assumption 6 (as well as strong convexity) fails to hold in that setting.
Our main algorithmic result is the following:
Theorem 7 Suppose that x1 , . . . , xn contains a subset S of size (1 − )n whose variance around its mean
µ is bounded by σ02 in the norm k · k. Also suppose that Assumption 6 holds for the dual norm
k · k∗ . Then,
√
1
if ≤ 4 , there is a polynomial-time algorithm whose output satisfies kµ̂ − µk = O σ0 κ .
1. In the language of Banach space theory, this is also referred to as having bounded co-type.
6
If, in addition, k · k is γ-strongly
convex, then even if S only has size αn there is a polynomial-time
√
κσ0
algorithm such that kµ̂ − µk = O √γα with probability Ω(α).
This is essentially a more restrictive, but computationally efficient version of Proposition 2. We note that for
the `2 -norm, the algorithm can be implemented as an SVD (singular value decomposition) combined with a
filtering step; for more general norms, the SVD is replaced with a semidefinite program.
√
In the small- regime, Theorem 7 is in line with existing results which typically achieve errors of O( )
in specific norms (see for instance a concurrent result of Diakonikolas et al. (2017a)). While several papers
achieve stronger rates of O(3/4 ) (Lai et al., 2016) or Õ() (Diakonikolas et al., 2016a; Balakrishnan et al.,
2017), these stronger results rely crucially on specific distributional assumptions such as Gaussianity. At the
√
time of writing of this paper, no results obtained rates better than O( ) for any general class of distributions
(even under strong assumptions such as sub-Gaussianity). After initial publication of this paper, Kothari and
√
Steinhardt (2017) surpassed and obtained rates of 1−γ for any γ > 0, for distributions satisfying the
Poincaré isoperimetric inequality.
In the small-α regime, Theorem 7 generalizes the mean estimation results of Charikar et al. (2017) to
√
norms beyond the `2 -norm. That paper achieves a better rate of 1/ α (versus the 1/α rate given here); we
√
think it is likely possible to achieve the 1/ α rate in general, but leave this for future work.
Applications. Because Assumption 6 holds for `p -norms, we can perform robust estimation in `p -norms
for any p ∈ [1, 2], as long as the data have bounded variance in the dual `q -norm (where p1 + 1q = 1);
see Corollary 27 for a formal statement. This is the first efficient algorithm for performing robust mean
estimation in any `p -norm with p 6= 2. The `1 -norm in particular is often a more meaningful metric than the
`2 -norm in discrete settings, allowing us to improve on existing results.
Indeed, as in the previous section, suppose we are given k-tuples of samples from a discrete distribution
π on {1, . . .p
, m}. Applying Theorem 7 with the `1 -norm yields an algorithm recovering a π̂pwith kπ̂ −
πkT V = Õ( /k).2 In contrast, bounds using the `2 -norm would only yield kπ̂ − πk2 = O( πmax /k),
which is substantially weaker when the maximum probability πmax is large. Our result has a similar flavor
to that of Diakonikolas et al. (2016a) on robustly estimating binary product distributions, for which directly
applying `2 mean estimation was also insufficient. We discuss our bounds in more detail in Section 8.
Finite-sample bound. To get the best sample complexity for the applications above, we provide an
additional finite-sample bound focused on showing that a set of points has bounded variance. This is a
simple but useful generalization of Proposition B.1 of Charikar et al. (2017); it shows that in a very generic
sense, given d samples from a distribution on Rd with bounded population variance, we can find a subset of
samples with bounded variance with high probability. It involves pruning the samples in a non-trivial way
based on ideas from graph sparsification (Batson et al., 2012). The formal statement is given in Section 6.2.
1.4. Low-Rank Recovery
Finally, to illustrate that the idea behind resilience is quite general and not restricted to mean estimation, we
also provide results on recovering a rank-k approximation to the data in the presence of arbitrary outliers.
Given a set of points [xi ]i∈S , let XS be the matrix whose columns are the xi . Our goal is to obtain a low-rank
matrix P such that the operator norm k(I −P )XS k2 is not much larger than σk+1 (XS ), where σk+1 denotes
the k + 1st singular value; we wish to do this even if S is corrupted to a set S̃ by adding arbitrary outliers.
As before, we start by formulating an appropriate resilience criterion:
2. The Õ notation suppresses log factors in m and ; the dependence on m can likely be removed with a more careful analysis.
7
Definition 8 (Rank-resilience) A set of points [xi ]i∈S in Rd is δ-rank-resilient if for all subsets T of size
at least (1 − δ)|S|, we have col(XT ) = col(XS ) and kXT† XS k2 ≤ 2, where † is the pseudoinverse and col
denotes column space.
Rank-resilience says that the variation in X should be sufficiently spread out: there should not be a direction
of variation that is concentrated in only a δ-fraction of the points. Under rank-resilience, we can perform
efficient rank-k recovery even in the presence of a δ-fraction of arbitrary data:
Theorem 9 Let δ ≤ 31 . If a set of n points contains a set S of size (1 − δ)n that is δ-rank-resilient, then it
is possible to efficiently recover a matrix P of rank at most 15k such that k(I − P )XS k2 = O(σk+1 (XS )).
The power of Theorem 9 comes from the fact that the error depends on σk+1 rather than e.g. σ2 , which is
what previous results yielded. This distinction is crucial in practice, since most data have a few (but more
than one) large singular values followed by many small singular values. Note that in contrast to Theorem 7,
Theorem 9 only holds when S is relatively large: at least (1 − δ)n ≥ 32 n in size.
1.5. Summary, Technical Highlights, and Roadmap
In summary, we have provided a deterministic condition on a set of points that enables robust mean estimation, and provided finite-sample bounds showing that this condition holds with high probability in many
concrete settings. This yields new results for distribution learning, stochastic block models, mean estimation
under bounded moments, and mean estimation in `p norms. We also provided an extension of our condition
that yields results for robust low-rank recovery.
Beyond the results themselves, the following technical aspects of our work may be particularly interesting: The proof of Proposition 2 (establishing that resilience is indeed sufficient for robust estimation), while
simple, is a nice pigeonhole argument that we found to be conceptually illuminating.
In addition, the proof of Theorem 5, on pruning resilient sets to obtain sets with bounded variance,
exploits strong convexity in a non-trivial way in conjunction with minimax duality; we think it reveals a
fairly non-obvious geometric structure in resilient sets, and also shows how the ability to prune points can
yield sets with meaningfully stronger properties.
Finally, in P
the proof of our algorithmic result (Theorem 7), we establish an interesting generalization of
the inequality i,j Xij2 ≤ rank(X) · kXk22 , which holds not just for the `2 -norm but for any strongly convex
norm. This is given as Lemma 18.
Roadmap. The rest of the paper is organized as follows. In Section 2, we prove our information-theoretic
recovery result for resilient sets (Proposition 2). In Section 3, we prove Theorem 5 establishing that all
resilient sets in strongly convex norms contain large subsets with bounded variance; we also prove a more
precise version of Theorem 5 in Section 3.1. In Section 4, we prove our algorithmic results, warming up with
the `2 -norm (Section 4.1) and then moving to general norms (Section 4.2). In Section 5, we prove our results
on rank-k recovery. In Section 6, we present and prove the finite-sample bounds discussed in Section 1.2. In
Sections 7 and 8, we provide applications of our information-theoretic and algorithmic results, respectively.
1.6. Related Work
A number of authors have recently studied robust estimation and learning in high-dimensional settings:
Lai et al. (2016) study mean and covariance estimation, while Diakonikolas et al. (2016a) focus on estimating Gaussian and binary product distributions, as well as mixtures thereof; note that this implies
mean/covariance estimation of the corresponding distributions. Charikar et al. (2017) recently showed that
robust estimation is possible even when the fraction α of “good” data is less than 21 . We refer to these papers
8
for an overview of the broader robust estimation literature; since those papers, a number of additional results have also been published: Diakonikolas et al. (2017a) provide a case study of various robust estimation
methods in a genomic setting, Balakrishnan et al. (2017) study sparse mean estimation, and others have
studied problems including regression, Bayes nets, planted clique, and several other settings (Diakonikolas
et al., 2016b,c, 2017b,c; Kane et al., 2017; Meister and Valiant, 2017).
Special cases of the resilience criterion are implicit in some of these earlier works; for instance, `2 resilience appears in equation (9) in Diakonikolas et al. (2016a), and resilience in a sparsity-inducing norm
appears in Theorem 4.5 of Li (2017). However, these conditions typically appear concurrently with other
stronger conditions, and the general sufficiency of resilience for information-theoretic recovery appears to
be unappreciated (for instance, Li (2017), despite already having implicitly established resilience, proves its
information-theoretic results via reduction to a tournament lemma from Diakonikolas et al. (2016a)).
Low rank estimation was studied by Lai et al. (2016), but their bounds depend on the maximum eigenvalue kΣk2 of the covariance matrix, while our bound provides robust recovery guarantees in terms of lower
singular values of Σ. (Some work, such as Diakonikolas et al. (2016a), shows how to estimate all of Σ in
e.g. Frobenius norm, but appears to require the samples to be drawn from a Gaussian.)
2. Resilience and Robustness: Information-Theoretic Sufficiency
P
Recall the definition of resilience: S is (σ, )-resilient if k |T1 | i∈T (xi − µ)k ≤ σ whenever T ⊆ S and
|T | ≥ (1 − )|S|. Here we establish Proposition 2 showing that, if we ignore computational efficiency,
resilience leads directly to an algorithm for robust mean estimation.
Proof (Proposition 2) We prove Proposition 2 via a constructive (albeit exponential-time) algorithm. To
prove the first part, suppose that the true set S is (σ, 1−
)-resilient around µ, and let S 0 be any set of
size (1 − )n that is (σ, 1− )-resilient (around some potentially different vector µ0 ). We claim that µ0 is
sufficiently close to µ.
Indeed, let T = S ∩ S 0 , which by the pigeonhole principle has size at least (1 − 2)n = 1−2
1− |S| =
(1 − 1−
)|S|. Therefore, by the definition of resilience,
1
|T |
P
i∈T (xi
− µ) ≤ σ.
(5)
P
But by the same argument, k |T1 | i∈T (xi − µ0 )k ≤ σ as well. By the triangle inequality, kµ − µ0 k ≤ 2σ,
which completes the first part of the proposition.
For the second part, we need the following simple lemma relating -resilience to (1 − )-resilience:
Lemma 10 For any 0 < < 1, a distribution/set is (σ, )-resilient around its mean µ if and only if it is
( 1−
even if µ is not the mean then the distribution/set is ( 2−
σ, 1 − )-resilient. More generally,
σ, 1 −
1 P
)-resilient. In other words, if k |T | i∈T (xi − µ)k ≤ σ for all sets T of size at least (1 − )n, then
P
0
k |T10 | i∈T 0 (xi − µ)k ≤ 2−
σ for all sets T of size at least n.
Given Lemma 10, the second part of Proposition 2 is similar to the first part, but requires us to consider
multiple resilient sets Si rather than a single S 0 . Suppose S is (σ, α4 )-resilient around µ–and thus also
( α8 σ, 1 − α4 )-resilient by Lemma 10–and let S1 , . . . , Sm be a maximal collection of subsets of [n] such that:
1. |Sj | ≥ α2 n for all j.
2. Sj is ( α8 σ, 1 − α2 )-resilient around some point µj .
3. Sj ∩ Sj 0 = ∅ for all j 6= j 0 .
9
Clearly m ≤ α2 . We claim that at least one of the µj is close to µ. By maximality of the collection {Sj }m
j=1 ,
it must be that S0 = S\(S1 ∪ · · · ∪ Sm ) cannot be added to the collection. First suppose that |S0 | ≥ α2 n.
Then S0 is ( α8 σ, 1 − α2 )-resilient (because any subset of α2 |S0 | points in S0 is a subset of at least α4 |S| points
α
in S). But this contradicts the maximality of {Sj }m
j=1 , so we must have |S0 | < 2 n.
Now, this implies that |S ∩ (S1 ∪ · · · ∪ Sm )| ≥ α2 n, so by pigeonhole we must have |S ∩ Sj | ≥ α2 |Sj |
for some j. Letting T = S ∩ Sj as before, we find that |T | ≥ α2 |Sj | ≥ α4 |S| and hence by resilience of Sj
and S we have kµ − µj k ≤ 2 · ( α8 σ) = 16
α σ. If we output one of the µj at random, we are then within the
1
desired distance of µ with probability m
≥ α2 .
3. Powering up Resilience: Finding a Core with Bounded Variance
In this section we prove Theorem 5, which says that for strongly convex norms, every resilient set contains a
core with bounded variance. Recall that this is important for enabling algorithmic applications that depend
on a bounded variance condition.
First recall the definition of resilience
(Definition 1): a set S is (σ, )-resilient if for every set T ⊆ S
P
of size (1 − )|S|, we have k |T1 | i∈T (xi − µ)k ≤ σ. For = 21 , we observe that resilience in a norm is
equivalent to having bounded first moments in the dual norm:
1
Lemma 11 Suppose that S is (σ,
2 )-resilient in a norm k · k, and let k · k∗ be the dual norm. Then S has
1 P
1st moments bounded by 3σ: |S| i∈S |hxi − µ, vi| ≤ 3σkvk∗ for all v ∈ Rd .
Conversely, if S has 1st moments bounded by σ, it is (2σ, 12 )-resilient.
The proof is routine and can be found in Section C. Supposing a set has bounded 1st moments, we will
show that it has a large core with bounded second moments. This next result is not routine:
Proposition 12 Let S be any set with 1st moments bounded by σ. Then if the norm k·k is γ-strongly convex,
2
1 P
2
there exists a core S0 of size at least 21 |S| with variance bounded by 32σ
i∈S0 |hxi −µ, vi| ≤
γ . That is, |S0 |
32σ 2
2
γ kvk∗
for all v ∈ Rd .
The assumptions seem necessary: i.e., such a core does not exist when k · k is the `p -norm with p > 2
(which is a non-strongly-convex norm), or with bounded 3rd moments for p = 2 (see Section A). The
proof of Proposition 12 uses minimax duality and Khintchine’s inequality (Khintchine, 1923). Note that
Lemma 11 and Proposition 12 together imply Theorem 5.
Proof (Proposition 12) Without loss of generality take µ = 0 and suppose that S = [n]. We can pose the
problem of finding a resilient core as an integer program:
n
1X
min
max
ci |hxi , vi|2 .
n
c∈{0,1}n ,kck1 ≥ n
kvk
≤1
∗
2
(6)
i=1
Here the variable ci indicates whether the point i lies in the core S0 . By taking a continuous relaxation and
applying a standard duality argument, we obtain the following:
P
2
Lemma 13 Suppose that for all m and all vectors v1 , . . . , vm satisfying m
j=1 kvj k∗ ≤ 1, we have
v
n uX
X
um
1
t
|hxi , vj i|2 ≤ B.
(7)
n
i=1
j=1
Then the value of (6) is at most 8B 2 .
10
The proof is straightforward and deferred to Section D. Now, to bound (7), let s1 , . . . , sm ∈ {−1, +1} be
i.i.d. random sign variables. We have
v
"√ n
#
n uX
m
X
X X
(i)
um
1
2
t
(8)
|hxi , vj i|2 ≤ Es1:m
sj hxi , vj i
n
n
i=1
j=1
i=1 j=1
"√ n
#
m
E
2X D X
= Es1:m
xi ,
(9)
sj vj
n
i=1
j=1
#
"
m
X
(ii)
√
2σ
(10)
sj vj
≤ Es1:m
≤
√
2σEs1:m
h
j=1
m
X
sj vj
j=1
∗
2 i 21
∗
.
(11)
Here (i) is Khintchine’s inequality (Haagerup, 1981) and (ii) is the assumed first moment bound. It remains
to bound (11). The key is the following inequality asserting that the dual norm k · k∗ is strongly smooth
whenever k · k is strongly convex (c.f. Lemma 17 of Shalev-Shwartz (2007)):
Lemma 14 If k · k is γ-strongly convex, then k · k∗ is (1/γ)-strongly smooth: 21 (kv + wk2∗ + kv − wk2∗ ) ≤
kvk2∗ + (1/γ)kwk2∗ .
h P
i
2
m
Applying Lemma 14 inductively to Es1:m
s
v
, we obtain
j
j
j=1
∗
Es1:m
m
h X
sj vj
j=1
2i
∗
≤
m
1
1X
kvj k2∗ ≤ .
γ
γ
(12)
j=1
p
Combining with (11), we have the bound B ≤ σ 2/γ, which yields the desired result.
3.1. Finding Resilient Cores when α ≈ 1
Lemma 11 together with Proposition 12 show that a (σ, 21 )-resilient set has a core with bounded 2nd moments. One piece of looseness is that Proposition 12 only exploits resilience for = 12 , and hence is not
sensitive to the degree of (σ, )-resilience as → 0. In particular, it only yields a core S0 of size 12 |S|, while
we might hope to find a much larger core of size (1 − )|S| for some small .
Here we tighten Proposition 12 to make use of finer-grained resilience information. Recall that we let
σ∗ () denote the resilience over sets of size (1 − )|S|. For a given , our goal is to construct a core S0 of
size (1 − )|S| with small second moments. The following key quantity will tell us how small the second
moments can be:
sZ
def
σ
e∗ () =
1/2
u−2 σ∗ (u)2 du.
(13)
/2
The following proposition, proved in Section E, says that σ
e∗ controls the 2nd moments of S0 :
Proposition 15 Let S be any resilient set in a γ-strongly-convex norm. Then for any ≤ 21 , there exists a
core S0 of size (1 − )|S| with variance bounded by O σ
e∗2 ()/γ .
11
The proof is similar to Proposition 12, but requires more careful bookkeeping.
To interpret σ
e∗ , suppose that σ∗ () = σ1−1/r for some r ∈ [1, 2), which roughly corresponds to having
R 1/2
2 2/r−1
σ2
bounded rth moments. Then σ
e∗2 () = σ 2 /2 u−2/r du ≤ 2/r−1
. If r = 1 then a core of size
2
(1 − )|S| might require second moments as large as σ ; on the other hand, as r → 2 the second moments
p
can be almost as small as σ 2 . In general, σ
e∗ () is O σ1/2−1/r if r ∈ [1, 2), is O σ log(1/) if r = 2,
and is O (σ) if r > 2.
4. Efficient Recovery Algorithms
We now turn our attention to the question of efficient algorithms. The main point of this section is to prove
Theorem 7, which yields efficient robust mean estimation for a general class of norms.
4.1. Warm-Up: Recovery in `2 -norm
We first prove a warm-up to Theorem 7 which focuses on the `2 -norm. Our warm-up result is:
Proposition 16 P
Let x1 , . . . , xn ∈ Rd , and let S be a subset of size αn with bounded variance in the `2 1
norm: λmax ( |S| i∈S (xi − µ)(xi − µ)> ) ≤ σ 2 , where µ is the mean of S. Then there is an efficient
randomized algorithm (Algorithm 1) which with probability Ω(α) outputs a parameter µ̂ such that kµ −
√
µ̂k2 = O ασ . Moreover, if α = 1 − ≥ 43 then kµ − µ̂k2 = O (σ ) with probability 1.
At the heart of Algorithm 1 is the following optimization problem:
minimize kX − XW k22
W ∈Rn×n
subject to 0 ≤ Wji ≤
1
∀i, j,
αn
X
Wji = 1 ∀i.
(14)
j
Here X ∈ Rd×n is the data matrix [x1 · · · xn ] and kX − XW k2 is the operator norm (maximum singular
value) of X − XW . Note that (14) can be expressed as a semidefinite program; however, it can actually be
solved more efficiently than this, via a singular value decomposition (see Section F for details).
The idea behind (14) is to re-construct each xi as an average of αn other xj . Note that by assumption
we can always re-construct each element of S using the mean of S, and have small error. Intuitively, any
element that cannot be re-constructed well must not lie in S, and can be safely removed. We do a soft form
of removal by maintaining weights ci on the points xi (initially all 1), and downweighting points with high
reconstruction error. We also maintain an active set A of points with ci ≥ 21 .
Informally, Algorithm 1 for estimating µ takes the following form:
1. Solve the optimization problem (14).
2. If the optimum is σ 2 n, then find the columns of X that are responsible for the optimum being
large, and downweight them.
3. Otherwise, if the optimum is O(σ 2 n), then take a low rank approximation W0 to W , and return a
randomly chosen column of XW0 .
The hope in step 3 is that the low rank projection XW0 will be close to µ for the columns belonging to S.
The choice of operator norm is crucial: it means we can actually
expect XW to be close to X (on the order
√
√
of σ n). In contrast, the Frobenius norm would scale as σ nd.
12
Algorithm 1 Algorithm for recovering the mean of a set with bounded variance in `2 -norm.
1: Initialize ci = 1 for all i = 1, . . . , n and A = {1, . . . , n}.
2: Let Y ∈ Rd×d and W ∈ RA×A be the maximizer/minimizer of the saddle point problem
X
max
min
ci (xi − XA wi )> Y (xi − XA wi ).
Y 0, 0≤Wji ≤ 4−α ,
P α(2+α)n i∈A
tr(Y )≤1
j Wji =1
LetPτi? = (xi − XA wi )> Y (xi − XA wi ).
?
2
4: if
i∈A ci τi > 4nσ then
(15)
3:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
For i ∈ A, replace ci with 1 −
τi?
τmax
ci , where τmax = maxi∈A τi? .
For all i with ci < 21 , remove i from A.
Go back to line 2.
end if
Let W1 be the result of zeroing out all singular values of W that are greater than 0.9.
Let Z = XA W0 , where W0 = (W − W1 )(I − W1 )−1 .
if rank(Z) = 1 then
Output the average of the columns of XA .
else
Output a column of Z at random.
end if
Finally, we note that kX − XW k22 is equal to
kX − XW k22 =
max
Y 0,tr(Y )≤1
n
X
(xi − Xwi )> Y (xi − Xwi ),
(16)
i=1
which is the form we use in Algorithm 1.
Proof (Proposition 16) We need to show two things: (1) that the outlier removal step removes many more
outliers than good points, and (2) that many columns of XW0 are close to µ.
Outlier removal. To analyze the outlier removal step (step 2 above, or lines 5-6 of Algorithm 1), we make
use of the following general lemma:
P
P
Lemma 17 For any scalars τi and a, suppose that i∈A ci τi ≥ 4a
τi ≤ αa. Then the
P while i∈S∩A αciP
following invariants are preserved by lines 5-6 of Algorithm 1: (i) i∈S (1 − ci ) ≤ 4 ni=1 (1 − ci ), and
(ii) |S ∩ A| ≥ α(2+α)
4−α n.
Lemma 17 says that we downweight points within S at least 4 times slower than we do overall (property
i), and in particular we never remove too many points from S (property ii). This type of lemma is not new
(cf. Lemma 4.5 of Charikar et al. (2017)) but for completeness we prove it in Section
P G.
We will show that we can take a = nσ 2 in Lemma 17, or in other words that i∈S∩A ci τi? P
≤ αnσ 2 . Let
1
τi (w) = (xi − XA w)> (xi − XA w), and note that τi? = τi (wi ) = min{τi (w) | 0 ≤ wj ≤ αn , j wj = 1}.
This is because for a fixed Y , each of the wi are optimized independently.
13
We can therefore bound τi? by substituting any feasible ŵi . We will choose Ŵji =
case XA ŵi = µ̂, where µ̂ is the average of xj over S ∩ A. Then we have
X
X
ci τi? ≤
ci τi (ŵi )
i∈S∩A
I[j∈S∩A]
|S∩A| ,
in which
(17)
i∈S∩A
X
≤
ci (xi − µ̂)> Y (xi − µ̂)
(18)
ci (xi − µ)> Y (xi − µ)
(19)
i∈S∩A
(i)
X
≤
i∈S∩A
≤
X
(xi − µ)> Y (xi − µ) ≤ αnσ 2 tr(Y ) ≤ αnσ 2
(20)
i∈S
as desired; (i) is because the covariance around the mean (µ̂) is smaller than around any other point (µ).
Analyzing XW0 . By Lemma 17, we will eventually exit the if statement and obtain Z = XA W0 . It
therefore remains to analyze Z; we will show in particular that kZA∩S −µ1> kF is small, where the subscript
indicates restricting to the columns in S ∩ A. At a high level, it suffices to show that W0 has low rank (so
that Frobenius norm is close to spectral norm) and that XW0 and X are close in spectral norm (note that X
and µ1> are close by assumption).
4−α
To bound rank(W0 ), note that the constraints in (15) imply that kW k2F ≤ α(2+α)
, and so at most
4−α
0.81α(2+α) singular values of W can be greater than 0.9. Importantly, at most 1 singular value can be greater
than 0.9 if α ≥ 43 , and at most O( α1 ) can be in general. Therefore, rank(W0 ) ≤ O( α1 ).
Next, we show that XA and Z are close in operator norm. Indeed, XA − Z = XA (I − W0 ) =
XA (I − W )(I − W1 )−1 , hence:
kXA − Zk2 = kXA (I − W )(I − W1 )−1 k2
−1
≤ kXA (I − W )k2 k(I − W1 )
(21)
k2
(22)
(i)
≤ 10kXA (I − W )k2
(iii)
√
√
≤ 10 2kXA (I − W ) diag(cA )1/2 k2 ≤ 20 2nσ.
(23)
(ii)
(24)
Here (i) is because all singular values of W1 are less than 0.9, (ii) is because diag(cA )1/2 √12 I, and
(iii) is by the condition in the if statement (line 4 of Algorithm 1), since the sum on line 4 is equal to
kXA (I − W ) diag(cA )1/2 k22 .
Combining the previous two observations, we have
X
kzi − µk22 ≤ (rank(Z) + 1)k[zi − µ]i∈S∩A k22
(25)
i∈S∩A
≤ (rank(Z) + 1) (k[zi − xi ]i∈S∩A k2 + k[xi − µ]i∈S∩A k2 )2
2
√
2
(i)
√
≤ (rank(Z) + 1) 20 2nσ + αnσ = O σα n .
(26)
(27)
Here (i) uses the preceding bound on kXA − Zk2 , together with the 2nd moment bound k[xi − µ]i∈S k2 ≤
√
αnσ. Note that rank(Z) ≤ rank(W0 ) = O( α1 ).
14
Since |S ∩ A| = Ω(αn) by Lemma 17, the average value of kzi − µk22 over S ∩ A is O
|S∩A|
2|A|
σ2
2
α
σ
α
, and hence
= Ω(α), a randomly chosen zi will be within distance O
with probability at least
of µ, which
completes the first part of Proposition 16.
For the second part, when α = 1 − ≥ 34 , recall that we have rank(W0 ) = 1, and that W0 =
(W − W1 )(I − W1 )−1 . One can then verify that√1> W0 = 1> . Therefore, W0 = u1> for some u.
Letting µ̃ = XA u, we have
1> k2 ≤ 20 2nσ by (24). In particular, A is resilient (around its
q kXA − µ̃√
2
≤ 40σ for ≤ 12 . Thus by the proof of Proposition 2 and the fact that
mean) with σ() ≤ 20σ 1−
√
5
|A| ≥ |S ∩ A| ≥ α(2+α)
4−α n ≥ (1 − 3 (1 − α))n, the mean of A is within O(σ 1 − α) of µ, as desired.
4.2. General Case
We are now ready to prove our general algorithmic result, Theorem 7 from Section 1.3. For convenience we
recall Theorem 7 here:
Theorem Suppose that x1 , . . . , xn contains a subset S of size (1−)n whose variance around its mean µ is
bounded by σ 2 in the norm k·k. Also suppose that Assumption 6 (κ-approximability) holds for the dual norm
√
k · k∗ . Then, if ≤ 41 , there is a polynomial-time algorithm whose output satisfies kµ̂ − µk = O σ κ .
If, in addition, k · k is γ-strongly
convex, then even if S only has size αn there is a polynomial-time
√
κσ
√
algorithm such that kµ̂ − µk = O γα with probability Ω(α).
Recall that bounded variance means that
1
|S|
P
i∈S hxi
− µ, vi2 ≤ σ 2 kvk2∗ for all v ∈ Rd . There are two
equivalent conditions to bounded variance that will be useful. The first is supkvk∗ ≤1 v > Σv ≤ σ 2 for all
1 P
>
v ∈ Rd , where Σ = |S|
i∈S (xi − µ)(xi − µ) ; this is useful because Assumption 6 allows us to κapproximate this supremum for any given Σ.
The second equivalent condition re-interprets σ in terms of a matrix norm. Let k · kψ denote the norm
k·k above, and for a matrix M define the induced 2 → ψ-norm kM k2→ψ as supkuk2 ≤1 kM ukψ . Then the set
p
S has variance at most σ 2 if and only if k[xi − µ]i∈S k2→ψ ≤ |S|σ. This will be useful because induced
norms satisfy helpful compositional properties such as kABk2→ψ ≤ kAk2→ψ kBk2 .
The algorithm establishing Theorem 7 is almost identical to Algorithm 1, with two changes. The first
change is that on line 4, the quantity 4nσ 2 is replaced with 4κnσ 2 , where κ is the approximation factor
in Assumption 6. The second change is that in the optimization (15), the constraint Y 0, tr(Y ) ≤ 1 is
replaced with Y ∈ P, where P is the feasible set in Assumption 6. In other words, the only difference is
that rather than finding the maximum eigenvalue, we κ-approximate the 2 → ψ norm using Assumption 6.
We therefore end up solving the saddle point problem
X
max
min
ci (xi − XA wi )> Y (xi − XA wi ).
(28)
Y ∈P 0≤Wji ≤ 4−α ,
P α(2+α)n i∈A
j Wji =1
Standard optimization algorithms such as Frank-Wolfe allow us to solve (28) to any given precision with a
polynomial number of calls to the linear optimization oracle guaranteed by Assumption 6.
While Algorithm 1 essentially minimizes the quantity kX − XW k22 , this new algorithm can be thought
of as minimizing kX − XW k22→ψ . However, for general norms computing the 2 → ψ norm is NP-hard,
and so we rely on a κ-approximate solution by optimizing over P. We are now ready to prove Theorem 7.
15
Proof (Theorem 7) The proof is similar to Proposition 16, so we only provide a sketch of the differences.
First, the condition of Lemma 17 still holds, now with a equal to κnσ 2 rather than nσ 2 due to the approximation ratio κ. (This is why we needed to change line 4.)
Next, we need to modify equations (21-24) to hold for the 2 → ψ norm rather than operator norm:
kXA − Zk2→ψ = kXA (I − W )(I − W1 )−1 k2→ψ
(29)
(i)
≤ kXA (I − W )k2→ψ k(I − W1 )−1 k2→2
(30)
≤ 10kXA (I − W )k2→ψ
√
√
≤ 10 2kXA (I − W ) diag(cA )1/2 k2→ψ ≤ 20 2κnσ.
(31)
(32)
Here (i) is from the general fact kABk2→ψ ≤ kAk2→ψ kBk2→2 , and the rest of the inequalities follow for
the same reasons as in (21-24).
We next need to modify equations (25-27). This can be done with the following inequality:
Lemma
18 For any matrix A = [a1 · · · an ] of rank r and any γ-strongly convex norm k · kψ , we have
Pn
r
2
2
i=1 kai kψ ≤ γ kAk2→ψ .
This generalizes the inequality kAk2F ≤ rank(A) · kAk22 . Using Lemma 18 (proved below), we have
X
kzi − µk2ψ ≤ rank(Z)+1
(33)
k[zi − µ]i∈S∩A k22→ψ
γ
i∈S∩A
≤
rank(Z)+1
γ
=O
κσ 2 n
αγ
(k[zi − xi ]i∈S∩A k2→ψ + k[xi − µ]i∈S∩A k2→ψ )2
(34)
.
(35)
The inequalities again follow for the same reasons
√ as
before. If we choose zi at random, with probability
σ κ
Ω(α) we will output a zi with kzi − µkψ = O α√γ . This completes the first part of the proposition.
√
For the second part, by the same reasoning as before we obtain µ̃ with kXA − µ̃1> k2→ψ = O ( nκσ),
√
which implies that A is resilient with σ∗ () = O (σ κ) for ≤ 21 . The mean of A will therefore be within
√
distance O (σ κ) of µ as before, which completes the proof.
We finish by proving Lemma 18.
Proof (Lemma 18) Let s ∈ {−1, +1}n be a uniformly random sign vector. We will compare Es [kAsk2ψ ] in
two directions. Let P be the projection onto the span of A. On the one hand, we have kAsk2ψ = kAP sk2ψ ≤
kAk22→ψ kP sk22 , and hence Es [kAsk2ψ ] ≤ Es [kP sk22 ]kAk22→ψ = rank(A)kAk22→ψ . On the other hand,
P
similarly to (12) we have Es [kAsk2ψ ] ≥ (1/γ) ni=1 kai k2ψ by the strong convexity of the norm k · kψ . Combining these yields the desired result.
5. Robust Low-Rank Recovery
In this section we present results on rank-k recovery. We first justify the definition of rank-resilience (Definition 8) by showing that it is information-theoretically sufficient for (approximately) recovering the best
rank-k subspace. Then, we provide an algorithm showing that this subspace can be recovered efficiently.
16
5.1. Information-Theoretic Sufficiency
Let XS = [xi ]i∈S . Recall that δ-rank-resilience asks that col(XT ) = col(XS ) and kXT† XS k2 ≤ 2 for
|T | ≥ (1 − δ)|S|. This is justified by the following:
δ
Proposition 19 Let S ⊆ [n] be a set of points of size (1 − δ)n that is 1−δ
-rank-resilient. Then it is possible
to output a rank-k projection matrix P such that k(I − P )XS k2 ≤ 2σk+1 (XS ).
δ
Proof Find the 1−δ
-rank-resilient set S 0 of size (1 − δ)n such that σk+1 (XS0 ) is smallest, and let P be the
projection onto the top k singular vectors of XS 0 . Then we have k(I −P )XS 0 k2 = σk+1 (XS 0 ) ≤ σk+1 (XS ).
Moreover, if we let T = S ∩ S 0 , we have k(I − P )XT k2 ≤ k(I − P )XS 0 k2 ≤ σk+1 (XS ) as well. By
δ
)|S|. Therefore col(XT ) = col(XS ), and k(I − P )XS k2 =
pigeonhole, |T | ≥ (1 − 2δ)n = (1 − 1−δ
k(I − P )XT XT† XS k2 ≤ k(I − P )XT k2 kXT† XS k2 ≤ 2σk+1 (XS ) as claimed.
5.2. Efficient Recovery
We now address the question of efficient recovery, by proving Theorem 9, whose statement we recall here
p
def
(with a few details added). For convenience let σ = σk+1 (XS )/ |S|.
Theorem Let δ ≤ 13 . If a set of n points contains a subset S of size (1 − δ)n that is δ-rank-resilient,
then there is an efficientpalgorithm (Algorithm 2) for recovering a matrix P of rank at most 15k such that
k(I − P )XS k2 = O(σ |S|) = O(σk+1 (XS )).
Algorithm 2 is quite similar to Algorithm 1, but we include it for completeness. The proof is also similar to
the proof of Proposition 16, but there are enough differences that we provide details below.
Algorithm 2 Algorithm for recovering a rank-k subspace.
2
Initialize ci = 1 for all i = 1, . . . , n and A = {1, . . . , n}. Set λ = (1−δ)nσ
.
k
2: Let Y ∈ Rd×d and Q ∈ RA×A be the maximizer/minimizer of the saddle point problem
X
max min
ci [(xi − XA qi )> Y (xi − XA qi ) + λkqi k22 ].
1:
Y 0, Q∈Rn×n
i∈A
tr(Y )≤1
3:
4:
5:
6:
7:
8:
9:
10:
(36)
LetPτi? = (xi − XA qi )> Y (xi − XA qi ) + λkqi k22 .
if i∈A ci τi? > 8nσ 2 then
τi?
For i ∈ A, replace ci with (1 − τmax
)ci , where τmax = maxi∈A τi? .
For all i with ci < 12 , remove i from A.
Go back to line 3.
end if
Let Q1 be the result of zeroing out all singular values of Q greater than 0.9.
†
Output P = XA Q0 XA
, where Q0 = (Q − Q1 )(I − Q1 )−1 .
Proof (Theorem 9) First, we
17 holds with a = 2nσ 2 . As before, the qi can be optimized
Pshow that Lemma
P
independently for each i, so i∈S∩A ci τi? ≤ i∈S∩A ci τi (q̂i ) for any matrix Q̂. Take Q̂ to be the projection
P
2
2
onto the
top
k
(right)
singular
vectors
of
X
.
Then
we
have
S∩A
i∈S∩A λkq̂i k2 ≤ λkQ̂kF = λk. We also
P
have i∈S∩A ci (xi − X q̂i )> Y (xi − X q̂i ) ≤ σk+1 (XS∩A )2 ≤ σ 2 |S| ≤ (1 − δ)nσ 2 . Together these imply
17
P
that i∈S∩A ci τi (q̂i ) ≤ (1 − δ)nσ 2 + λk = 2(1 − δ)nσ 2 , so we can indeed take a = 2nσ 2 . Consequently,
|S ∩ A| ≥ (1−δ)(3−δ)
n ≥ (1 − 5δ
3+δ
3 )n.
Now, when we get to line 9, we have kXA (I − Q) diag(cA )1/2 k22 + λkQk2F ≤ 8nσ 2 , and in particular
2
8k
= 1−δ
, and so Q0 will
each of the two terms individually is bounded by 8nσ 2 . Therefore, kQk2F ≤ 8nσ
λ
√
√
8k
have rank at most 0.81(1−δ) ≤ 15k. Moreover, kXA (I − Q)k2 ≤ 2kXA (I − Q) diag(cA )1/2 k2 ≤ 4 nσ.
We then have
†
k(I − P )XA k2 = k(I − XA Q0 XA
)XA k2
(37)
†
= kXA (I − Q0 )XA
XA k2
(38)
≤ kXA (I − Q0 )k2
(39)
= kXA (I − Q)(I − Q1 )
−1
k2
p
√
≤ 10kXA (I − Q)k2 ≤ 40 nσ = O σ |S| .
(40)
(41)
By the same argument as Proposition 19, since |S ∩ A| ≥ 1−5δ/3
1−δ |S| ≥ (1 − δ)|S|, it follows that
p
k(I − P )XS k2 ≤ 2k(I − P )XS∩A k2 = O σ |S| as well, which completes the proof.
6. Finite-Sample Concentration
In this section we provide two general finite-sample concentration results that establish resilience with high
probability. The first holds for arbitrary resilient distributions but has suboptimal sample complexity, while
the latter is specialized to distributions with bounded variance and has near-optimal sample complexity.
6.1. Concentration for Resilient Distributions
Our first result, stated as Proposition 4 in Section 1.2, applies to any (σ, )-resilient distribution p; recall that
p is (σ, )-resilient iff kE[x | E] − µk ≤ σ for any event E of probability 1 − .
We define the covering number of the unit ball in a norm k · k∗ to be the minimum M for which there
1
are vectors v1 , . . . , vM , each with kvj k∗ ≤ 1, such that maxM
j=1 hx, vj i ≥ 2 supkvk∗ ≤1 hx, vi for all vectors
v ∈ Rd . Note that log M is a measure of the effective dimension of the unit ball, i.e. log M = Θ(d) if k · k∗
is the `∞ or `2 norm, while log M = Θ(log d) for the `1 norm.
We recall Proposition 4 for convenience:
Proposition Suppose that a distribution p is (σ, )-resilient around its mean µ with < 12 . Let B be such
that P[kx − µk ≥ B] ≤ /2. Also let M be the covering number of the unit ball in the dual norm k · k∗ .
Then, given n samples x1 , . . . , xn ∼ p, with probability1 − δ − exp(−n/6) there is a subset
T of
q
(1 − )n of the xi such that T is (σ 0 , )-resilient with σ 0 = O σ 1 + log(M/δ)
+ (B/σ) log(M/δ)
.
n
2 n
Proof Let p0 be the distribution of samples from p conditioned on kx − µk ≤ B. Note that p0 is (σ, 2 )resilient since every event with probability 1 − /2 in p0 is an event of probability (1 − /2)2 ≥ 1 − in
p. Moreover, with probability 1 − exp(−n/6), at least (1 − )n of the samples from p will come from p0 .
Therefore, we can focus on establishing resilience of the n0 = (1 − )n samples from p0 .
18
With a slight abuse
of notation, let x1 , . . . , xn0 be the samples from p0 . Then to check resilience we
P
need to bound k |T1 | i∈T (xi − µ)k for all sets T of size at least (1 − )n0 . We will first use the covering
v1 , . . . , vM to obtain
1 X
1 X
M
(xi − µ) ≤ 2 max
hxi − µ, vj i.
(42)
j=1 |T |
|T |
i∈T
i∈T
The idea will be to analyze the sum over hxi − µ, vj i for a fixed vj and then union bound over the M
possibilities. For a fixed vj , we will split the sum into two components: those with small magnitude (roughly
σ/) and those with large magnitude (between σ/ and B). We can then bound the former with Hoeffding’s
inequality, and using resilience we will be able to upper-bound the second moment of the latter, after which
we can use Bernstein’s inequality.
More formally, let τ = 1−
/4 σ and define
yi = hxi − µ, vj iI[|hxi − µ, vj i| < τ ],
(43)
zi = hxi − µ, vj iI[|hxi − µ, vj i| ≥ τ ].
(44)
Clearly yi + zi = hxi − µ, vj i. Also, we have |yi | ≤ τ almost surely, and |zi | ≤ B almost surely (because
xi ∼ p0 and hence hxi − µ, vj i ≤ kxi − µk ≤ B). The threshold τ is chosen so that zi is non-zero with
probability at most /2 under p (see Lemma 3).
Now, for any set T of size at least (1 − )n0 , we have
1 X
1 X
hxi − µ, vj i =
yi + zi
|T |
|T |
i∈T
i∈T
1 X
1 X
yi +
|zi |
≤
|T |
|T |
i∈T
1
≤
|T |
n0
X
i=1
(45)
(46)
i∈T
n0
1 X
1 X
yi +
yi +
|zi |
|T |
|T |
i6∈T
(47)
i=1
0
0
n
n
X
1 X
1
1
y
+
τ
+
|zi |.
≤
i
1 − n0
1−
(1 − )n0
i=1
(48)
i=1
Pn0
P 0
The last step uses the fact that |yi | ≤ τ for all i. It thus suffices to bound | n10 i=1 yi | as well as n10 ni=1 |zi |.
For the yi term, note that by resilience kE[yi ]k ≤ σ (since yi is sampled from p conditioned on |hxi −
µ, vj i| < τ and kxi − µk ≤ B, which each occur with probabiliy at least 1 − /2). Then by Hoeffding’s
p
P 0
inequality, | n10 ni=1 yi | = O(σ + τ log(2/δ)/n0 ) with probability 1 − δ.
For the zi term, we note that E[|zi |] = E[max(zi , 0)] + E[max(−zi , 0)]. Let τ 0 be the -quantile of
hxi − µ, vj i under p, which by Lemma 3 is at most τ . Then we have
Ep [max(zi , 0)] = Ep [hxi − µ, vj iI[hxi − µ, vj i ≥ τ ]]
0
≤ Ep [hxi − µ, vj iI[hxi − µ, vj i ≥ τ ]]
(i)
≤ ·
1−
σ = (1 − )σ,
(49)
(50)
(51)
1
where (i) is again Lemma 3. Then we have Ep0 [max(zi , 0)] ≤ 1−
Ep [max(zi , 0)] ≤ σ, and hence Ep0 [|zi |] ≤
2σ (as E[max(−zi , 0)] ≤ σ by the same argument as above). Since |zi | ≤ B, we then have E[|zi |2 ] ≤ 2Bσ.
19
Therefore, by Bernstein’s inequality, with probability 1 − δ we have
0
n
1 X
|z
|
≤
O
σ+
i
n0
r
i=1
σB log(2/δ) B log(2/δ)
B log(2/δ)
+
=
O
σ
+
.
n0
n0
n0
(52)
P
Taking a union bound over the vj for both y and z, and plugging back into (48), we get that |T1 | i∈T hxi −
q
µ, vj i ≤ O σ + σ log(2M/δ)
for all T and vj with probability 1 − δ. Plugging back into
+ B log(2M/δ)
n
n
q
1 P
σ
(42), we get that k |T | i∈T (xi − µ)k ≤ O σ + log(2M/δ)
+ B log(2M/δ)
, as was to be shown.
n
n
6.2. Concentration Under Bounded Covariance
In this section we state a stronger but more restrictive finite-sample bound giving conditions under which
samples have bounded variance. It is a straightforward extension of Proposition B.1 of Charikar et al. (2017),
so we defer the proof to Section H.
Proposition 20 Suppose that a distribution p has bounded variance in a norm k · k∗ : Ex∼p [hx − µ, vi2 ] ≤
σ 2 kvk2∗ for all v ∈ Rd . Then, given n samples x1 , . . . , xn ∼ p, with probability 1 − exp(−2 n/16) there is
a subset T of (1 − )n of the points such that
1 X
hxi − µ, vi2 ≤ (σ 0 )2 kvk2∗ for all v ∈ Rd ,
|T |
(53)
i∈T
where (σ 0 )2 =
4σ 2
1+
d
(1−)n
.
Proposition 20 says that whenever a distribution p on Rd has bounded variance, if n ≥ d samples xi are
drawn from p then some large subset of the samples will have bounded variance as well.
7. Information-Theoretic Applications
Using Proposition 2 together with our finite-sample bounds, we establish information-theoretic robust recovery results in several cases of interest.
7.1. Distribution Learning
First we consider the case where we are given k-tuples of independent samples from a distribution π on
{1, . . . , m}. By taking the empirical distribution of the k samples, we get an element of the simplex ∆m ;
for instance if m = 5 and k = 3 and a given triple of samples is (2, 4, 2), then we produce the vector
(0, 32 , 0, 31 , 0) ∈ ∆5 . For general π and k, we will denote the resulting distribution over ∆m as F (π, k), and
think of both π and samples x ∼ F (π, k) as elements of Rm .
We first establish finite-sample concentration for F (π, k):
Proposition 21 Let p = F (π, k), where π is a distribution on {1, . . . , m}. Then:
• Given m samples x1 , . . . ,P
xm from p, with probability
1 − exp(−Ω(m)) there is a subset T of
the samples such that |T1 | i∈T hxi − π, vi2 ≤ O k1 kvk2∞ for all v ∈ Rd .
20
3m
4
of
√
• Given m k + m2 samples from p, with probability 1 − exp(−Ω(m)) there is a subset of (1 − )m of
p
the samples that is (σ, )-resilient in the `1 -norm, with σ = O( log(1/)/k).
As a corollary, we immediately get robust recovery bounds for π. For instance, we have:
√
Corollary 22 Suppose > 0 is sufficiently small. Then, given n = Ω( m k + m2 ) samples x1 , . . . , xn ,
(1 − )n of which are drawn from π k , with probability
p 1 − exp(−Ω(m)) it is information-theoretically
possible to recover π with total variational error O( log(1/)/k).
We omit the proof of Corollary 22, as it follows straightforwardly from Proposition 21 and Proposition 2.
Proof (Proposition 21) Consider a sample y ∼ π, interpreted
indicator vector in {0, 1}m . First, we
Pm as an
2
m
2
observe that for any vector v ∈ {−1, +1} , E[hy, vi ] = j=1 πj vj = 1, and so certainly E[hy − π, vi2 ] ≤
1. Therefore, if x is an average of k independent samples from π, E[hx − π, vi2 ] ≤ 1/k for all v with
kvk∞ ≤ 1. Invoking Proposition 20 then yields the first part of Proposition 21.
We next analyze the moment generating function E[exp(λ> (y − π))] in order to get a tail bound on
y − π. Provided kλk∞ ≤ 1, the moment generating function is bounded as
E[exp(λ> (y − π))] = exp(−λ> π)
≤ exp(−λ> π)
m
X
j=1
m
X
πj exp(λj )
(54)
πj (1 + λj + λ2j )
(55)
j=1
m
m
X
X
≤ exp(−λ> π) exp(
πj (λj + λ2j )) = exp(
πj λ2j ).
j=1
(56)
j=1
P
In particular, for any sign vector v and c ∈ [0, 1] we have E[exp(cv > (y − π))] ≤ exp(c2 m
j=1 πj ) =
2
>
exp(c ). For an average x of k independent samples from π, we therefore have E[exp(cv (x − π))] ≤
exp(c2 /k) for c ∈ [0, k]. But note that v > (x − π) ≤ 2, and so E[exp(cv > (x − π))] ≤ exp(2c2 /k) for all
c ≥ k as well. Hence E[exp(cv > (x − π))] ≤ exp(2c2 /k) for all c ≥ 0.
Now, let E be any event of probability . We have
1
log(E[exp(cv > (x − π)) | E])
c
1
1
≤ log( E[exp(cv > (x − π))])
c
1
1
≤ log( exp(2c2 /k))
c
log(1/) + 2c2 /k
=
.
c
E[v > (x − π) | E] ≤
(57)
(58)
(59)
(60)
p
Optimizing c yields E[v > (x − π) | E] p
≤ 8 log(1/)/k, andp
so (by Lemma 10) p is (σ, )-resilient around
its mean in the `1 -norm, with σ = 1−
8 log(1/)/k = O( log(1/)/k).
d
The second part of Proposition 21 now follows
p immediately upon invoking Proposition 4, with M = 2
d
(for the 2 sign vectors), B = 2 and σ = O( log(1/)/k).
21
7.2. Bounded kth Moments
We next give results for distributions with bounded kth moments.
Proposition 23 Suppose that a distribution p on Rd has bounded kth moments: E[|hx − µ, vi|k ]1/k ≤
1.5
σk kvk2 for all v ∈ Rd and some k ≥ 2. Then, given n ≥ d + d2 samples x1 , . . . , xn ∼ p, with probability
1 − exp(−Ω(d)) there is a set T of (1 − )n of the xi that is (O(σk 1−1/k ), )-resilient in the `2 -norm.
Proof We will invoke Proposition 4. Let U be the uniform distribution on the unit sphere. First note that
Ex [kx − µkk2 ] = Ex [(dEv∼U [|hx − µ, vi|2 ])k/2 ]
≤ dk/2 Ex,v [|hx − µ, vi|k ]
≤
dk/2 σkk .
(61)
(62)
(63)
√
Therefore, by Chebyshev’s
inequality, P[kx − µk2 ≥ (/2)−1/k σk d] ≤ /2. We can therefore take
√
B = O(−1/k σk d) in Proposition 4. Similarly, we can take σ = O(1−1/k σk ) in Proposition 4, and there
is a covering of the `2 ball with 9d elements, so log(M ) = O(d). Together, these imply that with probability
1 − exp(−d) there is a set T of size (1 − )n that is (O(σk 1−1/k ), )-resilient, as was to be shown.
As before, we obtain the following immediate corollary:
Corollary 24 Let p be a distribution on Rd with kth moments bounded by σk , for k ≥ 2, and suppose that
1.5
> 0 is sufficiently small. Then, given n = Ω( d + d2 ) samples x1 , . . . , xn , (1 − )n of which are drawn
from p, with probability 1 − exp(−Ω(d)) it is possible to recover µ̂ within `2 -distance O(1−1/k σk ) of the
mean of p.
7.3. Stochastic Block Model
Our final set of information-theoretic results is for the semi-random stochastic block model from Charikar
et al. (2017). In this model, we consider a graph G on n vertices, with an unknown set S of αn “good”
vertices. For simplicity we assume the graph is a directed graph. For i, j ∈ S, i is connected to j with
probability na , and for i ∈ S, j 6∈ S, i is connected to j with probability nb , where b < a. For i 6∈ S the
edges are allowed to be arbitrary. If we let A ∈ {0, 1}n×n be the adjacency matrix of G, then the rows in S
are jointly independent samples from a distribution p on {0, 1}n . The mean of this distribution is na for the
coordinates lying in S, and nb for the coordinates outside of S. We will apply our robust mean estimation
results to this distribution, and use this to robustly recover the set S itself. We let x1 , . . . , xαn denote the αn
samples from p corresponding to the elements of S.
Lemma 25 Suppose that x1 , . . . , xαn are drawn from a stochastic block model with parameters (α, a, b).
Take the norm kxk = max|J|=αn kxJ k1 , which is the maximum `1 -norm over any αn coordinates of x.
Define µj = na if j ∈ S and nb if j 6∈ S. Then, with probability 1 − exp(−Ω(αn)), the xi are (σ, α/2)resilient under k · k around µ with parameter
p
σ = O α a log(2/α) + log(2/α) .
(64)
Proof Note that we canexpress kxk as maxv∈V hx, vi, where V is the set of αn-sparse {0, +1, −1} vectors;
n
in particular, |V| = αn
2αn . By Lemma 10 and the definition of resilience, σ is equal to
D 1 X
E
α/2
max
max
xi − µ, v .
(65)
1 − α/2 T ⊆{1,...,αn},|T |= 12 α2 n v∈V |T |
i∈T
22
P
P
We will union bound over all T and v. For a fixed T and v, the inner expression is equal to α22 n i∈T nj=1 vj (xij −
µj ). Note that all of the vj (xij − µj ) are independent, zero-mean random variables with variance at most
a 2
n vj and are bounded in [−1, 1]. By Bernstein’s inequality, we have
!
n
2
XX
t
1
P
(66)
vj (xij − µj ) ≥ t ≤ exp − P a 2 1
2 i,j n vj + 3 t
i∈T j=1
!
t2
= exp − 3
,
(67)
α an + 23 t
Now, if we want
union bound to hold with probability 1 − δ, we need to set the term in (67) to
h our overall
n αn i
αn
t2
be at most δ/ α2 n/2 αn 2 , so α3 an+
2 = log(1/δ) + O (αn log(2/α)) = O (αn log(2α)) (since we
t
3
p
will take δ = exp(−Θ(αn))). Hence we can take t = O
α4 an2 log(2/α) + αn log(2/α) . Dividing
through by 12 α2 n and multiplying by
α/2
1−α/2 ,
we get
p
σ = O α a log(2/α) + log(2/α) ,
(68)
as was to be shown.
Using Lemma 25, we obtain our result on robust recovery of the set S:
Corollary 26 Under the semi-random stochastic block model with parameters α, a, and b, it is informationtheoretically possible to obtain a set Ŝ satisfying
s
!
1
a log(2/α)
|Ŝ4S| = O
.
(69)
αn
α2 (a − b)2
with probability Ω(α).
Proof By Lemma 25 and Proposition 2, we can recover a µ̂ that with probability Ω(α) satisfies
p
log(2/α)
a log(2/α) +
.
kµ̂ − µk = O
α
(70)
Note that µj = na if j ∈ S and nb if j 6∈ S, where b < a. We will accordingly define Ŝ to be the set
a−b
of coordinates j such that µ̂j ≥ a+b
2n . We then have that kµ̂ − µk = Ω( n min(|Ŝ4S|, αn)), and hence
1
1
αn |Ŝ4S| = O( α(a−b) kµ̂ − µk) whenever the right-hand-side is at most 1. Using (70), we have that
p
1
1
log(2/α)
|Ŝ4S| = O
a log(2/α) +
(71)
αn
α(a − b)
α
s
s
!
!
a log(2/α)
a log(2/α)
log(2/α)
a log(2/α)
=O
+
≤O
+
(72)
α2 (a − b)2 α2 (a − b)
α2 (a − b)2 α2 (a − b)2
a
Here the last step simply multiplies the second term by a−b
, which is at least 1. Note that the first term in
(72) dominates whenever the bound is meaningful, which yields the desired result.
23
8. Algorithmic Applications
We end by presenting applications of our algorithmic result Theorem 7. First, we show that Theorem 7
implies a deterministic recovery result for sets that are resilient in the `p -norm, for any p ∈ (1, 2]:
Corollary 27 Suppose that a set of points x1 , . . . , xn contains a set S of size αn that is (σ, 12 )-resilient
around its mean µ in `p-norm, with p ∈ (1, 2]. Then there is an efficient algorithm whose output µ̂ satisfies
kµ − µ̂kp = O
σ
α(p−1)
with probability Ω(α).
√
Moreover, if α = 1 − ≥ 0.87 and S is (σ δ, δ)-resilient for δ ∈ [ 2 , 12 ], the same algorithm has
q
e σ
kµ − µ̂kp = O
with probability 1. In particular, this holds if S has variance bounded by σ 2 .
p−1
Proof First, we observe that the `p -norm is (p − 1)-strongly convex for p ∈ (1, 2] (c.f. Lemma 17 of
Shalev-Shwartz (2007)). Therefore, if S has the assumed level of resilience, then we can find a core S0 of
σ2
size α2 with second moments bounded by O( p−1
) by Proposition 12. Moreover, if α = 1 − ≥ 0.87, then
3
2
we can find a core of size (1 − ) n ≥ 4 n by Proposition 15. In this latter case, the core will have second
2
moments bounded by O( σ log(1/)
).
p−1
In each case, we can apply Theorem 7 to this bounded-variance core. We need to check that Assumption 6 holds. This is due to the following result, which follows from Theorem 3 of Nesterov (1998):
Theorem 28 (Nesterov) Let k · kq denote the `q -norm for q ≥ 2. For a matrix A ∈ Rd×d , let f (A) =
maxkvkq ≤1 v > Av, and let g(A) = maxY 0,k diag(Y )kq/2 ≤1 hA, Y i. Then f (A) ≤ g(A) ≤ π2 f (A).
p
In particular, for the `p -norm with p ∈ [1, 2], the dual norm is the `q -norm with q = p−1
∈ [2, ∞]. Therefore,
Theorem 28 asserts that Assumption 6 holds for P = {Y | Y 0, k diag(Y )kq/2 ≤ 1} and κ = π2 = O(1).
σ
Now, applying Theorem 7 for general α, this yields a recovery guarantee of kµ̂ − µkp = O α(p−1)
with probability
Ω(α),
p
while for α = 1 − ≥ 0.87 this yields a recovery guarantee of kµ̂ − µkp =
σ
O √p−1
log(1/) with probability 1, as claimed.
We next specialize Corollary 27 to the distribution learning case and provide finite-sample bounds.
Corollary 29 For the distribution learning problem from Section 7.1, suppose that we are given n samples,
of which αn are independent k-tuples from a distribution π on {1, . . . , m}, and the remaining samples are
arbitrary. Then:
• If αn ≥ m, with probability 1 − exp(−Ω(m)) there is a randomized algorithm outputting a π̂ satis1
fying kπ̂ − πkT V = Õ( α√
) with probability Ω(α).
k
√
• If α = 1 − , where is sufficiently small, and n ≥ m k + m2 , then with probability 1 − exp(−Ω(m))
p
there is a deterministic algorithm outputting a π̂ satisfying kπ̂ − πkT V = Õ( /k).
In both of the above bounds, the Õ notation hides polylog factors in both and m.
Proof By Proposition 21, in the first case there will with high probability be a set of size 43 αn with bounded
variance under the `1 -norm. Since the `1 -norm and `1+ 1 -norm are within constant factors of each other,
log(m)
the set will also have bounded variance in the `1+
1
log(m)
then invoke Corollary 27 to obtain error O( σ log(m)
).
α
24
1
-norm, which is Ω( log(m)
)-strongly convex. We can
In thepsecond case, Proposition 21 says that we can remove |S|/2 points
p from S to make the remaining
set (O( log(1/)/k), /2)-resilient. This implies that it is also (O(δ log(1/)/k), δ)-resilient for all
δ ∈ [ 2 , 12 ], and so by Proposition 15 it has a core of size (1 − )(1 − /2)|S| = (1 − O())n with variance
at most log(m) klog(1/) (we again need to pass from `1 to `1+log(m) to apply Proposition 15, which is why we
incur the log(m) factor).
Applying Theorem 7 as in the
p
p second part of Corollary 27 then yields the desired error bound of
O( log(1/) log(m)/k) = Õ( /k).
Appendix A. Counterexample to Analogues of Proposition 12
In this section we show:
Proposition 30 Let S = {e1 , . . . , en }, where the ei are the standard basis in Rn . Then:
1. S is (2n1/p−1 , 21 )-resilient around 0 in `p -norm, for all p ∈ [1, ∞].
2. Any subset T of
n
2
elements of S has maxkvkq ≤1
1
|T |
P
i∈T
max(−1,k( p1 −1))
|hei , vi|k = (n/2)
.
Here we let k · kq denote the dual norm to k · kp , i.e. q is such that p1 + 1q = 1.
Proposition 30 implies in particular that for large n, an analog of Proposition 12 can only hold if k(1/p−
p
1) ≥ −1, which implies that k ≤ p−1
. For k = 2, this implies p ≤ 2 (so the norm is strongly convex), and
for p = 2 this implies k ≤ 2. We do not know whether an analog of Proposition 12 holds for e.g. k = 3,
p = 1.5.
Proof (Proposition 30) For the first part, note that the mean of any subset of n2 elements of S will have all
coordinates lying in [0, n2 ]. Therefore, the `p -norm of the mean is at most n2 · n1/p = 2n1/p−1 , as claimed.
For the second part, assume without loss of generality that T = {e1 , . . . , en/2 }. The optimization
then reduces to maxkvkq ≤1 n2 kvkkk , where v ∈ Rn/2 . For the optimal v, it is clear that kvkk = 1 if
1
k ≥ q, and kvkk = (n/2) k
max(−1,−k(1− p1 ))
(n/2)
− 1q
if k < q, whence
2
k
n kvkk
=
2
n
max(0,1− kq )
· (n/2)
max(−1,− kq )
= (n/2)
=
, as claimed.
Appendix B. Basic Properties of Resilience
Here we prove Lemmas 3 and 10. Recall that the first states that resilience around the mean µ is equivalent
to bounded tails, while the latter relates -resilience to (1 − )-resilience.
Proof (Lemma 10) Note that for a distribution p, resilience around µ can be computed as σ∗ () =
max{kE[X | E] − µk | E has probability 1 − }. When µ is the mean, then µ = E[X]. We can then
write
= kE[X | E] − E[X]k
(73)
= kE[X | E] − P[E]E[X | E] − P[¬E]E[X | ¬E]k
(74)
= (1 − P[E])kE[X | E] − E[X | ¬E]k.
(75)
From (75), it is clear that σ∗ (1 − ) =
1 − P[E] becomes P[E].
1−
σ∗ (),
as (75) is invariant under replacing E with ¬E, except that
25
More generally, if µ is not the mean then we still have kµ − E[X]k = σ∗ (0) ≤ σ∗ () by resilience. Then
if E is an event with probability , we have
kE[X | E] − µk ≤ kE[X | E] − E[X]k + kE[X] − µk
1−
kE[X | ¬E] − E[X]k + σ∗ ()
≤
1−
≤
(kE[X | ¬E] − µk + kµ − E[X]k) + σ∗ ()
2−
1−
· 2σ∗ () + σ∗ () =
σ∗ ().
≤
Therefore, even when µ is not the mean we have σ∗ (1 − ) ≤
2−
σ∗ (),
(76)
(77)
(78)
(79)
which completes the proof.
Proof (Lemma 3) We note that
1−
σ∗ (1 − )
(ii) 1 −
max{kE[X − µ | E]k | P[E] ≥ }
=
E
(iii) 1 −
=
max{hE[X − µ | E], vi | P[E] ≥ , kvk∗ ≤ 1}
E,v
(iv) 1 −
=
max{E[hX − µ, vi | E] | P[E] ≥ , kvk∗ ≤ 1}
E,v
(v) 1 −
=
max{E[hX − µ, vi | hX − µ, vi ≥ τ (v)] | kvk∗ ≤ 1},
v
(i)
σ∗ () =
(80)
(81)
(82)
(83)
(84)
as was to be shown. Here (i) is Lemma 10, (ii) is the definition of resilience, and (iii) is the definition of the
dual norm. Then (iv) uses linearity of expectation, while (v) explicitly maximizes E[hX − µ, vi | E] over
events E with probability at least .
Appendix C. Proof of Lemma 11
For a fixed v, let S+ = {i | hxi − µ, vi > 0}, and define S− similarly. Either S+ or S− must have size at
least 21 |S|, without loss of generality assume it is S+ . Then we have
X
X
1
1 X
|hxi − µ, vi| =
2
hxi − µ, vi −
hxi − µ, vi
(85)
|S|
|S|
i∈S
i∈S+
i∈S
X
X
1
≤
2k
(xi − µ)kkvk∗ + k
(xi − µ)kkvk∗
(86)
|S|
i∈S+
i∈S
|S+ |
≤ 2
+ 1 σkvk∗ ≤ 3σkvk∗ ,
(87)
|S|
26
which completes the first part. In the other direction, for any set T , we have
1 X
1 X
(xi − µ) = max
hxi − µ, vi
|T |
kvk∗ =1 |T |
i∈T
(88)
i∈T
1 X
|S|
≤ max
|hxi − µ, vi| ≤
σ ≤ 2σ,
|T |
kvk∗ =1 |T |
(89)
i∈S
as was to be shown.
Appendix D. Proof of Lemma 13
We start by taking a continuous relaxation of (6), asking for weights ci ∈ [0, 1] rather than {0, 1}:
n
1X
ci |hxi , vi|2 .
3n kvk ≤1 n
n
∗
c∈[0,1] ,kck1 ≥ 4
min
max
(90)
i=1
3n
4 ,
Note that we strengthened the inequality to kck1 ≥
whereas in (6) it was kck1 ≥ n2 . Given any solution
c1:n to (90), we can obtainPa solution c0 to (6) byPletting c0i = I[ci ≥ 12 ]. Then c0i ∈ {0, 1} and kc0 k1 ≥ n2 .
Moreover, c0i ≤ 2ci , so n1 ni=1 c0i |hxi , vi|2 ≤ n2 ni=1 ci |hxi , vi|2 for all v. Therefore, the value of (6) is at
most twice the value of (90).
Now, by the minimax theorem, we can swap the min and max in (90) in exchange for replacing the
single vector v with a distribution over vectors vj , thus obtaining that (90) is equal to
n
lim
max
min n
m→∞ α1 +···+αm ≤1 c∈[0,1]
α≥0,kvj k∗ ≤1 kck1 ≥ 3n
4
m
1X X
ci
αj |hxi , vj i|2 .
n
i=1
j=1
By letting vj0 = αj vj , the above is equivalent to optimizing over vj satisfying
lim
max
m→∞ kv1 k2∗ +···+kvm k2∗ ≤1
(91)
n
m
i=1
j=1
P
j
1X X
ci
min n
|hxi , vj i|2 .
c∈[0,1] n
kck1 ≥ 3n
4
kvj k2∗ ≤ 1:
(92)
For any q
v1 , . . . , vm , we will find c such that the above sum is bounded. Indeed, define B(v1:m ) to be
Pm
Pm
3n
1 Pn
2
2
2
i=1
j=1 |hxi , vj i| . Then take ci = I[ j=1 |hxi , vj i| < 16B ], which has kck1 ≥ 4 by Markov’s
n
P P
inequality, and for which i ci j |hxi , vj i|2 ≤ 4B(v1:m )2 .
Therefore, the value of (90) is bounded by maxm,v1:m 4B(v1:m )2 , and so the value of (6) is bounded by
maxm,v1:m 8B(v1:m )2 , which yields the desired result.
Appendix E. Proof of Proposition 15
The proof mirrors that of Proposition 12, but with more careful bookkeeping. Instead of (90), we consider
the minimax problem
n
1X
ci |hxi , vi|2 .
(93)
minn
max
c∈[0,1] ,
kvk∗ ≤1 n
i=1
kck1 ≥(1−/2)n
27
The only difference is that 34 n has been replaced with (1 − /2)n in the constraint on kck1 . We then end up
needing to bound
(1− 2 )n m
X
1 X
(94)
|hxik , vj i|2 .
P max2
n
kv
k
≤1
j
∗
j
k=1
j=1
Here we suppose that the indices i are arranged in order i1 , i2 , . . . , in such that the output of the inner sum
is monotonically increasing in ik (in other words, the outer sum over k is over the (1 − 2 )n smallest values
of the inner sum, which corresponds to the optimal choice of c). Now we have
v
!2
(1− 2 )n m
(1− 2 )n
n uX
m
X
X
(i) 1 X
u
1
1 X
t
(95)
|hxik , vj i|2 ≤
|hxil , vj i|2
n
n
n−k+1
k=1
j=1
k=1
(ii)
j=1
l=k
(1− 2 )n
2 X
≤
n
k=1
n D
m
E
X
X
1
Es
x il ,
sj vj
n−k+1
"
k=1
3n min(σ∗ ( n−k+1
), σ∗ ( 12 ))
n
Es
n−k+1
(1− 2 )n
18 X
≤
nγ
min(σ∗ ( n−k+1
), σ∗ ( 21 ))
n
n−k+1
n
k=1
(96)
j=1
l=k
(1− 2 )n
2 X
≤
n
(iii)
#!2
"
m
X
#!2
sj vj
j=1
(97)
∗
!2
.
(98)
Here (i) bounds the kth smallest term by the average of the n − k + 1 largest terms, (ii) is Khinchine’s
inequality, and (iii) is the following lemma which we prove later:
P
Lemma 31 If S is resilient around µ and T ⊆ S has size at most |S|, then i∈T |hxi −µ, vi| ≤ 3|S|kvk∗ ·
min(σ∗ (), σ∗ ( 21 )).
Continuing, we can bound (98) by
18
γ
Z
0
1− 2
min(σ∗ (1 − u), σ∗ ( 12 ))
1 − u − 1/n
2
18
du ≤
γ
2
1
−
2 Z
1
n
1
/2
min(σ∗ (u), σ∗ ( 12 ))
u
2
du
Z 1/2
1
−2
2
1 2
≤ O(1) · ·
u σ∗ (u) du + 2σ∗ ( 2 )
γ
/2
1
2
1 2
=O
.
· σ
e∗ () + σ∗ ( 2 )
γ
(99)
(100)
(101)
Since the σ
e∗ () term dominates, this completes the bound on (93), from which the remainder of the proof
follows identically to Proposition 12.
We end this section by proving Lemma 31.
Proof (Lemma 31) Note that the σ∗ ( 12 ) part of the bound was already established in Lemma 11, so it suffices
to show the σ∗ () part. Let µ̂ be the mean of S. Note that we must have kµ − µ̂k ≤ σ∗ () by definition. We
can thus replace µ with µ̂ in the statement of Lemma 31 at the cost of changing the left-hand-side by at most
28
|T |σ∗ ()kvk∗ . Now let T+ be the elements of T with hxi − µ̂, vi > 0 and define T− similarly. We have
X
X
X
hxi − µ, vi
(102)
hxi − µ, vi −
|hxi − µ, vi| =
i∈T
i∈T−
i∈T+
=−
X
hxi − µ, vi +
X
hxi − µ, vi
(103)
X
(104)
i∈S\T−
i∈S\T+
≤
X
(xi − µ) kvk∗ +
i∈S\T+
(xi − µ) kvk∗
i∈S\T−
≤ 2|S|σ∗ ()kvk∗ ,
(105)
which yields the desired result.
Appendix F. Fast Algorithm for `2 Mean Estimation
Recall that in Section 4.1 we wanted to solve the problem
min kX − XW k2 subject to 0 ≤ Wij ≤
W
1
, 1> W = 1> .
αn
(106)
An examination of the proof of Proposition 16 reveals that it actually suffices to replace the constraint
1
0 ≤ Wij ≤ αn
with the Frobenius norm constraint kW k2F ≤ α1 . We will show how to solve this modified
problem efficiently (via a singular value decomposition and linear-time post-processing). For r = α1 , we
want to solve
min kX − XW k2 subject to 1> W = 1> , kW k2F ≤ r.
(107)
W
11
Let P = n1 > .
satisfy > W 0 =
1
Then given any W , we can pass to the matrix W 0 = P + (I − P )W , which will always
1> . Therefore, it suffices to minimize
kX − XW 0 k22 = kX − X(P + (I − P )W )k22 = kX(I − P ) − X(I − P )W k22
(108)
with a constraint on
kW 0 k2F = kP + (I − P )W k2F = kP k2F + k(I − P )W k2F .
(109)
Now, at the optimum we have W = (I − P )W , and for all W we have k(I − P )W k2F ≤ kW k2F , so we
can replace k(I − P )W k2F with kW k2F without changing the optimization. In addition, P is a constant and
kP k2F = 1, so the constraint boils down to kW k2F ≤ r − 1. Letting X 0 = X(I − P ), it therefore suffices to
solve
min kX 0 − X 0 W k22 subject to kW k2F ≤ r − 1.
(110)
W
This problem is unitarily invariant and so can be solved efficiently via an SVD. In particular let X 0 =
U ΛV > . The optimal W lies in the span of V > , so after an appropriate change of variables it suffices to
minimize kΛ − ΛW k22 subject to kW k2F ≤ r − 1. We can also check that the optimal W is diagonal in this
basis, so we need only solve the vector minimization
X
min max λ2j (1 − wj )2 subject to
wj2 ≤ r − 1,
(111)
w
j
j
which can be done in linear time.
29
Appendix G. Proof of Lemma 17
Assuming that (i) and (ii) holdP
prior to line 5 of the algorithm,
P we need to show that they continue to hold.
that the amount that
P By assumption, we have i∈S∩A ci τi ≤ αa,
P while i∈A ci τi ≥ 4a. But note P
n
c
decreases
in
line
5
is
proportional
to
c
τ
,
while
the
amount
that
i=1 ci decreases is
i∈S i
i∈S∩A i i
P
α
proportional to i∈A ci τi . Therefore, the former decreases at most 4 times as fast as the latter, meaning
that (i) is preserved.
P
P
P
Since (i) is preserved, we have i∈S (1 − ci ) ≤ α4
i∈S (1 − ci ) +
i6∈S (1 − ci ) . Re-arranging yields
P
α(1−α)
2α(1−α)
α P
1
i∈S (1 − ci ) ≤ 4−α
i6∈S (1 − ci ) ≤ 4−α n. In particular, #{i | ci ≤ 2 } ≤
4−α n, so at most
2α(1−α)
4−α n
elements of S have been removed from A in total. This implies |S ∩ A| ≥
preserved as well.
α(2+α)
4−α n,
and so (ii) is
Appendix H. Proof of Proposition 20
This is a corollary of Proposition B.1 of Charikar et al. (2017), which states that given a distribution with
covariance Σ0 , where
Σ0 σ02 I, with probability 1 − exp(−2 n/16) it is possible to find a set T of (1 − )n
P
points such that |T1 | i∈T (xi − µ)(xi − µ)> (σ00 )2 I, where σ00 and σ0 have the same relation as σ 0 and
σ above. Applying this proposition directly to p will not give us our result, so we instead apply it to a
transformed version of p.
In particular, let Σ be the covariance of p, and given a sample xi ∼ p, let yi = Σ−1/2 (xi − µ).3 Then the
yi have identity covariance and mean 0, and so Proposition
B.1 of Charikar
et al. (2017) implies that we can
1 P
4
d
>
find (1 − )n of the yi such that |T | i∈T yi yi 1 + (1−)n I. In terms of the xi , after multiplying
by Σ1/2 on the left and right, we obtain
1 X
4
(xi − µ)(xi − µ)>
|T |
1+
i∈T
d
(1 − )n
Σ,
(112)
which implies that
1 X
4
hxi − µ, vi2 ≤
|T |
i∈T
1+
d
(1 − )n
v > Σv
4
d
=
1+
Ex∼p [hx − µ, vi2 ]
(1 − )n
4
d
≤
1+
σ 2 kvk2∗ = (σ 0 )2 kvk2∗ ,
(1 − )n
(113)
(114)
(115)
as was to be shown.
References
S. Balakrishnan, S. S. Du, J. Li, and A. Singh. Computationally efficient robust sparse estimation in high
dimensions. In Conference on Learning Theory (COLT), pages 169–212, 2017.
3. In the case that Σ is not invertible, we need to add a small multiple of the identity matrix before inverting, but we omit this part
of the argument for brevity.
30
J. Batson, D. A. Spielman, and N. Srivastava. Twice-Ramanujan sparsifiers. SIAM Journal on Computing,
41(6):1704–1721, 2012.
M. Charikar, J. Steinhardt, and G. Valiant. Learning from untrusted data. In Symposium on Theory of
Computing (STOC), 2017.
A. Decelle, F. Krzakala, C. Moore, and L. Zdeborová. Asymptotic analysis of the stochastic block model
for modular networks and its algorithmic applications. Physical Review E, 84(6), 2011.
I. Diakonikolas, G. Kamath, D. Kane, J. Li, A. Moitra, and A. Stewart. Robust estimators in high dimensions
without the computational intractability. In Foundations of Computer Science (FOCS), 2016a.
I. Diakonikolas, D. Kane, and A. Stewart. Robust learning of fixed-structure Bayesian networks. arXiv,
2016b.
I. Diakonikolas, D. M. Kane, and A. Stewart. Statistical query lower bounds for robust estimation of highdimensional Gaussians and Gaussian mixtures. arXiv, 2016c.
I. Diakonikolas, G. Kamath, D. Kane, J. Li, A. Moitra, and A. Stewart. Being robust (in high dimensions)
can be practical. arXiv, 2017a.
I. Diakonikolas, G. Kamath, D. M. Kane, J. Li, A. Moitra, and A. Stewart. Robustly learning a Gaussian:
Getting optimal error, efficiently. arXiv, 2017b.
I. Diakonikolas, D. M. Kane, and A. Stewart. Learning geometric concepts with nasty noise. arXiv, 2017c.
O. Guédon and R. Vershynin. Community detection in sparse networks via Grothendieck’s inequality. arXiv,
2014.
U. Haagerup. The best constants in the khintchine inequality. Studia Mathematica, 70(3):231–283, 1981.
D. Kane, S. Karmalkar, and E. Price. Robust polynomial regression up to the information theoretic limit.
arXiv, 2017.
A. Khintchine. Über dyadische brüche. Mathematische Zeitschrift, 18:109–116, 1923.
A. R. Klivans, P. M. Long, and R. A. Servedio. Learning halfspaces with malicious noise. Journal of
Machine Learning Research (JMLR), 10:2715–2740, 2009.
P. Kothari and J. Steinhardt. Better agnostic clustering via tensor norms. arXiv, 2017.
K. A. Lai, A. B. Rao, and S. Vempala. Agnostic estimation of mean and covariance. In Foundations of
Computer Science (FOCS), 2016.
C. M. Le, E. Levina, and R. Vershynin. Concentration and regularization of random graphs. arXiv, 2015.
J. Li. Robust sparse estimation tasks in high dimensions. arXiv, 2017.
L. Massoulié. Community detection thresholds and the weak Ramanujan property. In Symposium on Theory
of Computing (STOC), pages 694–703, 2014.
M. Meister and G. Valiant. A data prism: Semi-verified learning in the small-alpha regime. arXiv, 2017.
31
E. Mossel, J. Neeman, and A. Sly. A proof of the block model threshold conjecture. arXiv, 2013.
Y. Nesterov. Semidefinite relaxation and nonconvex quadratic optimization. Optimization methods and
software, 9:141–160, 1998.
E. Rebrova and K. Tikhomirov. Coverings of random ellipsoids, and invertibility of matrices with iid heavytailed entries. arXiv, 2015.
E. Rebrova and R. Vershynin. Norms of random matrices: local and global problems. arXiv, 2016.
S. Shalev-Shwartz. Online Learning: Theory, Algorithms, and Applications. PhD thesis, The Hebrew
University of Jerusalem, 2007.
J. Steinhardt. Does robustness imply tractability? A lower bound for planted clique in the semi-random
model. arXiv, 2017.
J. Steinhardt, G. Valiant, and M. Charikar. Avoiding imposters and delinquents: Adversarial crowdsourcing
and peer prediction. In Advances in Neural Information Processing Systems (NIPS), 2016.
H. Xu, C. Caramanis, and S. Mannor. Principal component analysis with contaminated data: The high
dimensional case. arXiv, 2010.
32
| 2cs.AI
|
Applications of position-based coding to classical
communication over quantum channels
arXiv:1704.01361v3 [quant-ph] 16 Nov 2017
Haoyu Qi∗
Qingle Wang∗†
Mark M. Wilde∗‡
November 17, 2017
Abstract
Recently, a coding technique called position-based coding has been used to establish achievability statements for various kinds of classical communication protocols that use quantum
channels. In the present paper, we apply this technique in the entanglement-assisted setting
in order to establish lower bounds for error exponents, lower bounds on the second-order coding rate, and one-shot lower bounds. We also demonstrate that position-based coding can be a
powerful tool for analyzing other communication settings. In particular, we reduce the quantum
simultaneous decoding conjecture for entanglement-assisted or unassisted communication over a
quantum multiple access channel to open questions in multiple quantum hypothesis testing. We
then determine achievable rate regions for entanglement-assisted or unassisted classical communication over a quantum multiple-access channel, when using a particular quantum simultaneous
decoder. The achievable rate regions given in this latter case are generally suboptimal, involving
differences of Rényi-2 entropies and conditional quantum entropies.
1
Introduction
Understanding optimal rates for classical communication over both point-to-point quantum channels and quantum network channels are fundamental tasks in quantum Shannon theory (see, e.g.,
[Hol12, Wat16, Wil16, Wil17b]). Early developments of quantum Shannon theory are based on the
assumption that the information is transmitted over an arbitrarily large number of independent and
identically distributed (i.i.d.) uses of a given quantum channel. By taking advantage of this assumption, general formulas have been established for capacities of various communication protocols, with
or without preshared entanglement. When a sender and receiver do not share entanglement before
communication begins, it is known that the Holevo information of a quantum channel is an achievable rate for classical communication [Hol98, SW97]. Regularizing the Holevo information leads to
a multi-letter formula that characterizes the capacity for this task. Regarding communication over
quantum network channels, an achievable rate region for classical communication over quantum
multiple-access channels was given in [Win01] and regularizing it leads to a characterization of the
∗
Hearne Institute for Theoretical Physics and Department of Physics & Astronomy, Louisiana State University,
Baton Rouge, Louisiana 70803, USA
†
State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing 100876, China
‡
Center for Computation and Technology, Louisiana State University, Baton Rouge, Louisiana 70803, USA
1
capacity region for this task. However, only inner bounds on the capacity region for general broadcast channels are known [YHD11, RSW16, SW15], except when the quantum broadcast channel
is a particular kind of degraded channel [WDW17]. When there is entanglement shared between
the communicating parties, many scenarios have also been studied, including classical communication over point-to-point quantum channels [BSST02, Hol02], quantum multiple-access channels
[HDW08, XW13], and quantum broadcast channels [DHL10, WDW17].
Although channel capacity gives a fundamental characterization of the communication capabilities of a quantum channel, many practically important properties of quantum channels are not
captured by this quantity. To close this gap, several works have focused on the study of refined
notions of capacity, including error exponents [BH98, Win99b, Hol00, Hay07, Dal13, DW14] and
second-order asymptotics [TT15, WRG16, DTW16]. The latter works build upon strong connections between hypothesis testing and coding, as considered in [Hay09a, PPV10, WR12]. The
refined characterizations of capacity are of importance for regimes of practical interest, in which a
limited number of uses of a quantum channel are available. Complementary to these developments,
to go beyond the i.i.d. assumption, many works have been dedicated to the one-shot formalism
[WR12, DRRW13, DH13, MW14] and the information-spectrum approach [HN03, Hay06, BD06],
with very few assumptions made on the structure of quantum channels.
In a recent work [AJW17b], a technique called position-based coding was developed in order
to give one-shot achievability bounds for various classical communication protocols that use entanglement assistance. This was then extended to the cases of unassisted classical communication
and private classical communication [Wil17a]. The method of position-based coding is a derivative
of the well known and long studied coding technique called pulse position modulation (see, e.g.,
[Ver90, CE03]). In pulse position modulation, a sender encodes a message by placing a pulse in
one slot and having no pulse in the other available slots. Each slot is then communicated over the
channel, one by one. The receiver can decode well if he can distinguish “pulse” from “no pulse.”
Position-based coding borrows this idea: a sender and receiver are allowed to share many copies
of a bipartite quantum state. The sender encodes a message by sending a share of one of the
bipartite states through a channel to the receiver. From the receiver’s perspective, only one share
of his systems will be correlated with the channel output (the pulse in one slot), while the others
will have no correlation (no pulse in the other slots). So if the receiver can distinguish “pulse”
from “no pulse” in this context, then he will be able to decode well, just as is the case in pulse
position modulation. The authors of [AJW17b] applied the position-based coding technique to a
number of problems that have already been addressed in the literature, including point-to-point
entanglement-assisted communication [BSST02, Hol02], entanglement-assisted coding with side information at the transmitter [Dup09], and entanglement-assisted communication over broadcast
channels [DHL10].
In the present paper, we use position-based coding to establish several new results. First, we establish lower bounds on the entanglement-assisted error exponent and on the one-shot entanglementassisted capacity. The latter improves slightly upon the result from [AJW17b] and in turn gives a
simpler proof of one of the main results of [DTW16], i.e., a lower bound on the second-order coding rate for entanglement-assisted classical communication. We then turn to communication over
quantum multiple-access channels when using a quantum simultaneous decoder, considering both
cases of entanglement assistance and no assistance. The quantum simultaneous decoding conjecture
from [FHS+ 12, Wil11] stands as one of the most important open problems in network quantum
information theory. Here we report progress on this conjecture and connect it to some open ques-
2
tions from [AM14, BHOS15] in multiple quantum hypothesis testing. At the same time, we give
new achievable rate regions for entanglement-assisted classical communication over multiple-access
channels, where the bounds on achievable rates are expressed as a difference of a Rényi entropy of
order two and a conditional quantum entropy.
This paper is organized as follows. We first summarize relevant definitions and lemmas in
Section 2. In this section, we also prove Proposition 3, which relates the hypothesis testing relative
entropy to the quantum Rényi relative entropy and is an interesting counterpart to [CMW16,
Lemma 5]. In Section 3, we consider entanglement-assisted point-to-point classical communication.
By using position-based coding, we establish a lower bound on the entanglement-assisted error
exponent. We also establish a lower bound on the one-shot entanglement-assisted capacity in
terms of hypothesis-testing mutual information and state how it is close to a previously known
upper bound from [MW14]. Based upon this one-shot bound, we then rederive a lower bound on
the second-order coding rate for entanglement-assisted communication with a proof that is arguably
simpler than that given in [DTW16]. In Section 4, we apply position-based coding to entanglementassisted classical communication over multiple-access channels and establish an explicit link to
multiple quantum hypothesis testing. We give an achievable rate region for i.i.d. channels by
using techniques from the theory of quantum typicality. We demonstrate the power of positionbased coding technique in unassisted classical communication in Section 4.2, by considering classical
communication over multiple-access channel. We explicitly show how to derandomize a randomnessassisted protocol. In Section 5, we tie open questions in multiple quantum hypothesis testing to
quantum simultaneous decoding for the quantum multiple-access channel. Finally, we summarize
our main results and discuss open questions in Section 6.
2
Preliminaries
Trace distance, fidelity, and gentle measurement. Let D(H) denote the set of density
operators (positive semi-definite operators with unit trace) acting on a Hilbert space H. The
√ trace
distance between two density operators ρ, σ ∈ D(H) is equal to kρ − σk1 , where kAk1 ≡ Tr{ A† A}.
Another quantity to measure the closeness between two quantum states is the fidelity, defined as
√ √
F (ρ, σ) ≡ k ρ σk21 [Uhl76]. Two inequalities relating trace distance and quantum measurement
operators are as follows:
Lemma 1 (Gentle operator [Win99a, ON07]) Consider a density operator ρ ∈ D(H) and a measurement operator Λ where 0 ≤ Λ ≤ I. Suppose that the measurement operator Λ detects the state
ρ with high probability Tr{Λρ} ≥ 1 − ε, where ε ∈ [0, 1]. Then
√ √
√
ρ − Λρ Λ ≤ 2 ε .
(2.1)
1
Lemma 2 Consider two quantum states ρ, σ ∈ D(H) and a measurement operator Λ where 0 ≤
Λ ≤ I. Then we have
Tr{Λρ} ≥ Tr{Λσ} − kρ − σk1 .
(2.2)
More generally, the same bound holds when ρ and σ are subnormalized, i.e., Tr{ρ}, Tr{σ} ≤ 1.
Information spectrum. The information spectrum approach [Han03, NH07, HN03, DR09]
gives one-shot bounds for operational tasks in quantum Shannon theory, with very few assumptions
3
made about the source or channel [NH07, Hay03, HN03, Hay07]. What plays an important role in
the information spectrum method is the positive
Pspectral projection of an operator. For a Hermitian
operator X with spectral decomposition X = i λi |iihi|, the associated positive spectral projection
is denoted and defined as
X
{X ≥ 0} ≡
|iihi| .
(2.3)
i : λi ≥0
Relative entropies and mutual informations. For a state ρ ∈ D(H) and a positive semidefinite operator σ, the quantum Rényi relative entropy of order α, where α ∈ [0, 1) ∪ (1, +∞) is
defined as [Pet86, TCR09]
1
Dα (ρkσ) ≡
log2 Tr{ρα σ 1−α } .
(2.4)
α−1
If α > 1 and supp(ρ) 6⊆ supp(σ), it is set to +∞. In the limit as α → 1, the above definition
reduces to the quantum relative entropy [Ume62]
D(ρkσ) ≡ Tr{ρ[log2 ρ − log2 σ]} ,
(2.5)
which is defined as above when supp(ρ) ⊆ supp(σ) and it is set to +∞ otherwise. Using the above
definition, we can define the Rényi mutual information for a bipartite state θRB as
Iα (R; B)θ ≡ Dα (θRB kθR ⊗ θB ) .
(2.6)
The ε-hypothesis testing relative entropy for a state ρ and a positive semi-definite σ is defined for
ε ∈ [0, 1] as [BD10, WR12]
ε
DH
(ρkσ) ≡ − log2 inf {Tr{Λσ} : 0 ≤ Λ ≤ I ∧ Tr{Λρ} ≥ 1 − ε} .
Λ
(2.7)
Similarly, we define the ε-hypothesis testing mutual information of a bipartite state θRB as
ε
ε
IH
(R; B)θ ≡ DH
(θRB kθR ⊗ θB ) .
(2.8)
The following proposition establishes an inequality relating hypothesis testing relative entropy
and the quantum Rényi relative entropy, and it represents a counterpart to [CMW16, Lemma 5].
We give its proof in the appendix, where we also mention how [CMW16, Lemma 5] and the following
proposition lead to a transparent proof of the quantum Stein’s lemma [HP91, ON00].
Proposition 3 Let ρ be a density operator and let σ be a positive semi-definite operator. Let
α ∈ (0, 1) and ε ∈ (0, 1). Then the following inequality holds
α
1
ε
DH (ρkσ) ≥
log2
+ Dα (ρkσ).
(2.9)
α−1
ε
The hypothesis testing relative entropy has the following second-order expansion [TH13, Li14,
DPR16]:
p
ε
DH
(ρ⊗n kσ ⊗n ) = nD(ρkσ) + nV (ρkσ)Φ−1 (ε) + O(log n) ,
(2.10)
where V (ρkσ) = Tr{ρ[log2 ρ − log2 σ]2 } − [D(ρkσ)]2 is the quantum relative entropy variance and
the function Φ(a) is the cumulative distribution function for a standard normal distribution:
Z a
1
2
Φ(a) ≡ √
dx e−x /2 .
(2.11)
2π −∞
4
Let σ be a quantum state now. The hypothesis testing relative entropy is relevant for asymmetric
hypothesis testing, in which the goal is to minimize the error probability Tr{Λσ} subject to a
constraint on the other kind of error probability Tr{(I −Λ)ρ} ≤ ε. We could also consider symmetric
hypothesis testing, in which the goal is to minimize both kinds of error probabilities simultaneously.
It is useful for us here to take the approach of [AM14] and consider general positive semi-definite
operators A and B rather than states ρ and σ. As in [AM14], we can define the error “probability”
in identifying the operators A and B as follows:
Pe∗ (A, B) ≡
inf
T : 0≤T ≤I
= Tr{A} −
Tr{(I − T )A} + Tr{T B}
sup
T : 0≤T ≤I
Tr{T (A − B)}
(2.12)
(2.13)
= Tr{A} − Tr{{A − B ≥ 0}(A − B)}
(2.14)
1
= (Tr{A + B} − kA − Bk1 ) .
(2.15)
2
The following lemma allows for bounding Pe∗ (A, B) from above, and we use it to establish bounds
on the error exponent for entanglement-assisted communication.
Lemma 4 ([ACMnT+ 07]) Let A and B be positive semi-definite operators and s ∈ [0, 1]. Then
the following inequality holds
Pe∗ (A, B) =
1
(Tr{A + B} − kA − Bk1 ) ≤ Tr{As B 1−s }.
2
(2.16)
The above lemma was first proved in [ACMnT+ 07], but the reader should note that a much simpler
proof due to N. Ozawa is presented in [JOPS12, Proposition 1.1] and [Aud14, Theorem 1].
Weak typicality. We will use results from the theory of weak typicality in some of our
achievability proofs (see, e.g., [Wil16, Wil17b] for a review). Consider a density operator ρA with
P
spectral decomposition: ρA = x pX (x)|xihx|A . The weakly δ-typical subspace TAρ,δ
n is defined as
n
the span of all unit vectors |x i ≡ |x1 i ⊗ |x2 i ⊗ · · · ⊗ |xn i such that the sample entropy H(xn ) of
their classical label is close to the true entropy H(X) = H(A)ρ of the distribution pX (x):
n
n
TAρ,δ
(2.17)
n ≡ span |x i : H(x ) − H(X) ≤ δ ,
P
where H(xn ) ≡ − n1 log2 (pX n (xn )) and H(X) ≡ − x pX (x) log2 pX (x). The δ-typical projector
Πnρ,δ onto the typical subspace of ρ is defined as
X
|xn ihxn |,
(2.18)
Πρ,δ
≡
n
A
xn ∈TδX
n
where we have used the symbol TXδ n to refer to the set of δ-typical sequences:
TXδ n ≡ xn : H(xn ) − H(X) ≤ δ .
(2.19)
Three important properties of the typical projector are as follows:
⊗n
Tr{Πρ,δ
} ≥ 1 − ε,
An ρ
2−n[H(A)+δ] Πρ,δ
An ≤
Tr{Πρ,δ
An }
⊗n ρ,δ
Πρ,δ
ΠA n
An ρ
5
n[H(A)+δ]
≤2
(2.20)
,
≤ 2−n[H(A)−δ] Πρ,δ
An ,
(2.21)
(2.22)
where the first property holds for arbitrary ε ∈ (0, 1), δ > 0, and sufficiently large n. We will also
need the following ‘projector trick’ inequality [GLM12, FHS+ 12]:
n[H(A)+δ] ⊗n
Πρ,δ
ρA ,
An ≤ 2
(2.23)
⊗n Πρ,δ =
which follows as a consequence of the leftmost inequality in (2.22) and the fact that Πρ,δ
An ρ
An
p
p
⊗n ≤ ρ⊗n . A final inequality we make use of is the following one
ρ⊗n Πρ,δ
ρ
An
⊗n −1/2
−n[H(A)−δ]/2
Πρ,δ
ρ
,
(2.24)
An ≤ 2
−1/2
which is a consequence of sandwiching the rightmost inequality of (2.22) by [ρ⊗n ]
⊗n , and operator monotonicity of the square root function.
Πρ,δ
An ≤ I
, applying
Hayashi-Nagaoka operator inequality. We repeatedly use the following operator inequality
from [HN03] when analyzing error probability:
Lemma 5 Given operators S and T such that 0 ≤ S ≤ I and T ≥ 0, the following inequality holds
for all c > 0:
I − (S + T )−1/2 S(S + T )−1/2 ≤ cI (I − S) + cII T ,
(2.25)
where cI ≡ 1 + c and cII ≡ 2 + c + c−1 .
3
Entanglement-assisted point-to-point classical communication
We begin by defining the information-processing task of point-to-point entanglement-assisted classical communication, originally considered in [BSST02, Hol02] and studied further in [DH13, MW14,
DTW16]. Before communication begins, the sender Alice and the receiver Bob share entanglement
in whatever form they wish, and we denote their shared state as ΨRA . Suppose Alice would like
to communicate some classical message m from a set M ≡ {1, . . . , M } over a quantum channel
NA0 →B , where M ∈ N denotes the cardinality of the set M. An (M, ε) entanglement-assisted
m
classical code, for ε ∈ [0, 1], consists of a collection {EA→A
0 }m of encoders and a decoding POVM
m
{ΛRB }m , such that the average error probability is bounded from above by ε:
M
1 X
m
Tr{(I − Λm
RB )NA0 →B (EA→A0 (ΨRA ))} ≤ ε,
M
(3.1)
m=1
For fixed ε, let M ∗ (N , ε) denote the largest M for which there exists an (M, ε) entanglementassisted classical communication code for the channel N . Then we define the ε-one-shot entanglementassisted classical capacity as log2 M ∗ (N , ε). We note that one could alternatively consider maximum
error probability when defining this capacity. The entanglement-assisted capacity of a channel N
is then defined as
1
CEA (N ) ≡ lim lim inf log2 M ∗ (N ⊗n , ε).
(3.2)
ε→0 n→∞ n
For fixed M , the one-shot entanglement-assisted error exponent − log2 ε∗ (N , M ) is such that ε∗
is equal to the smallest ε for which there exists an (M, ε) entanglement-assisted classical communication code. In the i.i.d. setting, the entanglement-assisted error exponent is defined for a fixed
rate R ≥ 0 as
1
∗
⊗n nR
(3.3)
EEA (N , R) ≡ lim sup − log2 ε (N , 2 ) .
n
n→∞
6
3.1
One-shot position-based coding
We now review the method of position-based coding [AJW17b], as applied to point-to-point entanglementassisted communication, and follow the review by showing how the approach leads to a lower
bound on the error exponent for entanglement-assisted communication, a lower bound for oneshot entanglement-assisted capacity, and a simple proof for a lower bound on the second-order
coding rate for entanglement-assisted communication. We note that a lower bound for one-shot
entanglement-assisted capacity using position-based coding was already given in [AJW17b], but the
lower bound given here leads to a lower bound on the entanglement-assisted second-order coding
rate that is optimal for covariant channels [DTW16].
The position-based entanglement-assisted communication protocol consists of two steps, encoding and decoding, and we follow that discussion with an error analysis of its performance.
Encoding: Before communication begins, Alice and Bob share the following state:
⊗M
θRA
≡ θR1 A1 ⊗ · · · ⊗ θRM AM ,
(3.4)
where Alice possesses the A systems and Bob has the R systems. To send message m, Alice simply
sends the mth A system through the channel. So this leads to the following state for Bob:
⊗M −m
⊗m−1
.
⊗ NA→B (θRm Am ) ⊗ θR
ρm
R M B ≡ θR
(3.5)
Decoding: Define the following measurement:
Γm
RM B ≡ IRm−1 ⊗ TRm Bm ⊗ IRM −m ,
(3.6)
where TRm Bm = TRB is a “test” or measurement operator satisfying 0 ≤ TRB ≤ IRB , which we will
specify later. For now, just think of it as corresponding to a measurement that should distinguish
well between NA→B (θRA ) and θR ⊗ NA→B (θA ). This is important for the following reason: If
message m is transmitted and the test is performed on the mth R system and the channel output
system B, then the probability of it accepting is
m
Tr{Γm
RM B ρRM B } = Tr{TRB NA→B (θRA )}.
(3.7)
If however the test is performed on the m0 th R system and B, where m0 6= m, then the probability
of it accepting is
0
m
Tr{Γm
(3.8)
RM B ρRM B } = Tr{TRB [θR ⊗ NA→B (θA )]}.
We use these facts in the forthcoming error analysis.
We use a square-root measurement to form a decoding POVM for Bob as follows:
Λm
RM B ≡
M
X
!−1/2
m0
RM B
Γm
RM B
Γ
m0 =1
M
X
!−1/2
m0
RM B
Γ
.
(3.9)
m0 =1
This is called the position-based decoder.
Error analysis: The error probability under this coding scheme is the same for each message m
(see, e.g., [AJW17b, Wil17a]) and is as follows:
m
pe (m) ≡ Tr{(IRM B − Λm
RM B )ρRM B }.
7
(3.10)
Applying Lemma 5 with S = Γm
and T =
RM B
be bounded from above as
m
Tr{(IRM B − Λm
RM B )ρRM B }
m
≤ cI Tr{(IRM B − Γm
RM B )ρRM B } + cII
m0
m0 6=m ΓRM B ,
P
0
X
m0 6=m
= cI Tr{(IRB − TRB )NA→B (θRA )} + cII
we find that this error probability can
m
Tr{Γm
RM B ρRM B }
X
m0 6=m
Tr{TRB [θR ⊗ NA→B (θA )]}
= cI Tr{(IRB − TRB )NA→B (θRA )} + cII (M − 1) Tr{TRB [θR ⊗ NA→B (θA )]}.
(3.11)
(3.12)
(3.13)
The same bound applies for both the average and the maximum error probability, due to the
symmetric construction of the code.
Our bound for a test operator TRB is thus as follows and highlights, as in [AJW17b], an
important connection between quantum hypothesis testing (i.e., the ability to distinguish the states
NA→B (θRA ) and θR ⊗ NA→B (θA )) and entanglement-assisted communication:
pe (m) ≤ cI Tr{(IRB − TRB )NA→B (θRA )} + cII (M − 1) Tr{TRB [θR ⊗ NA→B (θA )]}.
3.2
(3.14)
Lower bounds on one-shot and i.i.d. entanglement-assisted error exponents
We first prove a lower bound on the one-shot error exponent, and then a lower bound for the
entanglement-assisted error exponent in the i.i.d. case directly follows.
Theorem 6 For a quantum channel NA→B , a lower bound on the one-shot entanglement-assisted
error exponent for fixed message size M is as follows:
"
#
− log2 ε∗ (N , M ) ≥ sup (1 − s) sup Is (R; B)N (θ) − log2 M − 2 ,
(3.15)
θRA
s∈[0,1]
where θRA is a pure bipartite entangled state and Is (R; B)N (θ) is the Rényi mutual information
defined in (2.6).
Proof. Following the position-based encoding and decoding procedure described in Section 3.1
and setting c = 1 in (3.14), the error probability for each message can be bounded as
m
pe (m) = Tr{(IRM B − Λm
RM B )ρRM B }
(3.16)
≤ 4 [Tr{(IRB − TRB )NA→B (θRA )} + M Tr{TRB [θR ⊗ NA→B (θA )]}]
= 4 [Tr{NA→B (θRA )} − Tr{TRB (NA→B (θRA ) − M [θR ⊗ NA→B (θA )])}] .
(3.17)
(3.18)
To minimize the term in the last line above, it is well known that one should take the test operator
TRB as follows:
TRB = {NA→B (θRA ) − M [θR ⊗ NA→B (θA )] ≥ 0} .
(3.19)
The statement for quantum states is due to [Hel69, Hol73, Hel76] and the extension (relevant for
us) to the more general case of positive semi-definite operators appears in [AM14, Eq. (22)] (see
8
also (2.12)–(2.15)). This then leads to the following upper bound on the error probability:
pe (m) ≤ 4 [Tr{NA→B (θRA )} − Tr{TRB (NA→B (θRA ) − M [θR ⊗ NA→B (θA )])}]
h
= 2 Tr{NA→B (θRA ) + M [θR ⊗ NA→B (θA )]}
i
− kNA→B (θRA ) − M [θR ⊗ NA→B (θA )]k1
≤ 4 Tr{NA→B (θRA )s (M θR ⊗ NA→B (θA ))1−s }
1−s
1−s
s
= 4M
Tr{NA→B (θRA ) [θR ⊗ NA→B (θA )]
= 4 2−(1−s)[Is (R;B)N (θ) −log2 M ] .
(3.20)
(3.21)
(3.22)
}
(3.23)
(3.24)
The first equality is standard, using the relation of the positive part of an operator to its modulus
(see, e.g., [AM14, Eq. (23)]). The second inequality is a consequence of [ACMnT+ 07, Theorem 1],
recalled as Lemma 4 in Section 2, and holds for all s ∈ [0, 1] (see [JOPS12, Proposition 1.1] and
[Aud14, Theorem 1] for a simpler proof of [ACMnT+ 07, Theorem 1] due to N. Ozawa). The last
equality follows from the definition of Rényi mutual information in (2.6). Since this bound holds
for an arbitrary s ∈ [0, 1] and an arbitrary input state θRA , we can conclude the following bound:
h
i
Is (R;B)N (θ) −log2 M
− sups∈[0,1] (1−s) supθ
RA
pe (m) ≤ 4 2
.
(3.25)
Note that it suffices to take θRA as a pure bipartite state, due to the ability to purify a mixed θRA
and the data-processing inequality for Is (R; B)N (θ) , holding for all s ∈ [0, 1] [Pet86]. Finally taking
a negative binary logarithm of both sides of (3.25) gives (3.15).
We remark that the above proof bears some similarities to that given in [Hay07] (one can find
a related result in the later work [MD09, Lemma 3.1]). One of the results in [Hay07] concerns
a bound on the error exponent for classical communication over classical-input quantum-output
channels. The fundamental tool used in the proof of this result in [Hay07] is Lemma 4, attributed
above to [ACMnT+ 07]. Our proof above clearly follows the same approach.
⊗n
leads to the followApplying the above result in the i.i.d. case for a memoryless channel NA→B
ing:
Proposition 7 For a quantum channel NA→B , a lower bound on the entanglement-assisted error
exponent EEA (N , R) (defined in (3.3)) for fixed rate R ≥ 0 is as follows:
"
#
EEA (N , R) ≥ sup (1 − s) sup Is (R; B)N (θ) − R
,
(3.26)
θRA
s∈[0,1]
where θRA is a pure bipartite entangled state and Is (R; B)N (θ) is the Rényi mutual information
defined in (2.6).
⊗n
Proof. A proof follows by plugging in the memoryless channel NA→B
into (3.15), setting the
nR
number of messages to be M = 2 , and considering that
sup Is (Rn ; B n )N ⊗n (θ(n) ) ≥ sup Is (Rn ; B n )[N (θ)]⊗n
(3.27)
⊗n
θRA
(n)
θRn An
= n sup Is (R; B)N (θ) ,
θRA
9
(3.28)
leading to the following bound:
"
#
1
2
∗
⊗n nR
− log2 ε (N , 2 ) ≥ sup (1 − s) sup Is (R; B)N (θ) − R − .
n
n
θRA
s∈[0,1]
(3.29)
The equality in (3.28) follows from the additivity of the Rényi mutual information for tensor-power
states. Taking the large n limit then gives (3.26). Alternatively, plugging the memoryless channel
⊗n
NA→B
in to (3.25), we find that the bound on the error probability becomes
pe (m) ≤ 4 2−(1−s)n[Is (R;B)N (θ) −R] ,
(3.30)
holding for all s ∈ [0, 1] and states θRA . After taking a negative logarithm, normalizing by n, and
taking the limit as n → ∞, we arrive at (3.26).
3.3
Lower bounds on one-shot entanglement-assisted capacity and entanglementassisted second-order coding rate
By using position-based coding, here we establish a lower bound on the one-shot entanglementassisted capacity. Note that a similar lower bound was established in [AJW17b], but the theorem
below allows for an additional parameter η ∈ (0, ε), which is helpful for giving a lower bound on
the entanglement-assisted second-order coding rate.
Theorem 8 Given a quantum channel NA→B and fixed ε ∈ (0, 1), the ε-one-shot entanglementassisted capacity of NA→B is bounded as
ε−η
(R; B)N (θ) − log2 (4ε/η 2 ),
log2 M ∗ (N , ε) ≥ max IH
θRA
(3.31)
where η ∈ (0, ε) and the hypothesis testing mutual information is defined in (2.8).
Proof. The idea is to use the same coding scheme described in Section 3.1 and take the test
operator TRB in Bob’s decoder to be Υ∗RB , where Υ∗RB is the optimal measurement operator for
ε−η
(R; B)N (θ) , with η ∈ (0, ε). Then, starting from the upper bound on the error probability in
IH
(3.14), the error analysis reduces to
Tr{(IRM B − ΛRM B )ρm
RM B }
≤ cI Tr{(IRB − Υ∗RB ) [NA→B (θRA )]} + cII M Tr{Υ∗RB [θR ⊗ NA→B (θA )]}
(3.32)
≤ cI (ε − η) + cII M 2
(3.33)
ε−η
−IH
(R;B)N (θ)
.
The second inequality follows from the definition of quantum hypothesis testing relative entropy,
which gives that
Tr{Υ∗RB [NA→B (θRA )]} ≥ 1 − (ε − η),
ε−η
−IH
(R;B)N (θ)
Tr{Υ∗RB [θR ⊗ NA→B (θA )]} = 2
(3.34)
.
(3.35)
To make the error pe (m) ≤ ε, we set c = η/(2ε − η) for η ∈ (0, ε), and this leads to
ε−η
log2 M = IH
(R; B)N (θ) − log2 (4ε/η 2 ).
10
(3.36)
ε−η
The inequality in the theorem follows after maximizing IH
(R; B)N (θ) with respect to all input
states θRA .
Comparison to upper bound. The authors of [MW14] established the following upper bound
on one-shot entanglement-assisted capacity:
ε
ε
IH
(R; B)N (θ) ≥ max min DH
(NA→B (θRA )kθR ⊗ σB )
θRA
σB
≥ log2 M ∗ (N , ε).
(3.37)
(3.38)
Thus, there is a sense in which the upper bound from [MW14] is close to the lower bound in
(3.31). In particular, we could pick η = δε for any constant δ ∈ (0, 1), and the lower bound beε(1−δ)
ε(1−δ)
comes maxθRA IH
(R; B)N (θ) −log2 (4/εδ 2 ). Thus the information term maxθRA IH
(R; B)N (θ)
ε (R; B)
can become arbitrarily close to IH
by
picking
δ
smaller,
but
at
the
cost
of
the term
N (θ)
− log2 (4/εδ 2 ) becoming more negative with decreasing δ.
Lower bound on second-order coding rate. To get a lower bound on the entanglementε−η
(R; B)N (θ)
assisted second-order coding rate for an i.i.d. channel N ⊗n , evaluate the formula IH
√
⊗n
for an i.i.d. state N (θ) , pick η = 1/ n and n large enough such that ε − η > 0, and use the
ε in (2.10). We then recover one of the main results of [DTW16]:
second-order expansions for DH
q
log2 M ∗ (N ⊗n , ε) ≥ nI(R; B)N (θ) + nV (R; B)N (θ) Φ−1 (ε) + O(log n).
(3.39)
Interestingly, this is achievable for maximal error in addition to average error due to the above
analysis. Additionally, it does seem that this approach for arriving at a lower bound on the
entanglement-assisted second-order coding rate is much simpler than the previous approach developed in [DTW16].
4
Classical communication over quantum multiple-access channels
We now establish a link between communication over a quantum multiple-access channel and multiple quantum hypothesis testing. One advantage of this development is the reduction of the communication problem to a hypothesis testing problem, which is perhaps simpler to state and could also
be considered a more fundamental problem than the communication problem. Later, in Section 5,
we discuss the relation of the quantum simultaneous decoding conjecture from [FHS+ 12, Wil11] to
open questions in multiple quantum hypothesis testing from [AM14, BHOS15] (here we note that
the solution of the multiple Chernoff distance conjecture from [Li16] does not evidently allow for the
solution of the quantum simultaneous decoding conjecture). We point the reader to [Dut11, DF13]
for further discussions and variations of the quantum simultaneous decoding conjecture.
We begin by considering the case of entanglement-assisted communication and later consider
unassisted communication. We first define the information-processing task of entanglement-assisted
classical communication over quantum multiple-access channels (see also [HDW08, XW13]). Consider the scenario in which two senders Alice and Bob would like to transmit classical messages to a
receiver Charlie over a two-sender single-receiver quantum multiple-access channel NAB→C . Alice
and Bob choose their messages from message sets L and M. The cardinality of the sets L and M
11
R
Charlie
A’
Alice
L
E
B’
N
B
M
Bob
L’
A
C
D
M’
F
S
Figure 1: The coding task for entanglement-assisted classical communication over a quantum
multiple-access channel. Alice and Bob each share entanglement with the receiver Charlie, and
Charlie employs a decoding channel D to figure out which messages Alice and Bob transmitted.
are denoted as L and M , respectively. Suppose that Alice and Bob each share an arbitrary entangled state with Charlie before communication begins. Let ΦRA0 denote the state shared between
Charlie and Alice, and let ΨSB 0 denote the state shared between Charlie and Bob.
Let L, M ∈ N and ε ∈ [0, 1]. An (L, M, ε) entanglement-assisted multiple-access classical
communication code consists of a set {EAl 0 →A , FBm0 →B , Λl,m
RSC }l,m of encoders and a decoding POVM,
such that the average error probability is bounded from above by ε:
L M
1 XX
pe (l, m) ≤ ε ,
LM
(4.1)
l=1 m=1
where the error probability for each message pair is given by
l
m
pe (l, m) ≡ Tr{(IRSC − Λl,m
RSC )NAB→C (EA0 →A (ΦRA0 ) ⊗ FB 0 →B (ΨSB 0 ))} .
(4.2)
Figure 1 depicts the coding task.
4.1
One-shot position-based coding scheme
We now describe and analyze a position-based coding scheme for entanglement-assisted communication over a quantum multiple-access channel, in which the decoding POVM is a quantum
simultaneous decoder.
Encoding: Before communication begins, suppose that Alice and Charlie share L copies of the
⊗L
same bipartite state: θRA
≡ θR1 A1 ⊗ · · · ⊗ θRL AL . Similarly, suppose that Bob and Charlie share M
⊗M
copies of the same bipartite state: γSB
≡ γS1 B1 ⊗ · · · ⊗ γSM BM . To send message (l, m) ∈ L × M,
12
⊗L
⊗M
Alice sends the lth A system of θRA
and Bob sends the mth B system of θSB
over the quantum
multiple-access channel NAB→C . So this leads to the following state for Charlie:
⊗(l−1)
ρl,m
= θR
RL S M C
⊗(m−1)
⊗ γS
⊗(L−l)
⊗ NAB→C (θRl Al ⊗ γSm Bm ) ⊗ θR
⊗(M −m)
⊗ γS
.
(4.3)
Decoding: To decode the message transmitted, Charlie performs a measurement on the systems RL , S M , and the channel output C to determine the message pair (l, m) that Alice and Bob
transmitted. Consider the following measurement operator:
Γl,m
≡ TRl Sm C ,
RL S M C
(4.4)
where identity operators are implicit for all of the R and S systems besides Rl and Sm and TRSC
is a measurement operator satisfying 0 ≤ TRSC ≤ IRSC . Let us call the action of performing
the measurement {Γl,m
, I L M − Γl,m
} “checking for the message pair (l, m).” If Charlie
RL S M C R S C
RL S M C
checks for message pair (l, m) when indeed message pair (l, m) is transmitted, then the probability
of incorrectly decoding is
Tr{(I − Γl,m
)ρl,m
} = Tr{(I − TRSC )NAB→C (θRA ⊗ γSB )}.
RL S M C RL S M C
(4.5)
The equality above follows by combining (4.3) and (4.4) and applying partial traces. There are
three other kinds of error probabilities to consider. If message pair (l, m) was transmitted and
Charlie checks for message pair (l0 , m) for l0 6= l, then the probability of decoding as (l0 , m) is
0
l,m
Tr{ΓlR,m
L S M C ρRL S M C } = Tr{TRSC NAB→C (θR ⊗ θA ⊗ γSB )}.
(4.6)
If message pair (l, m) was transmitted and Charlie checks for message pair (l, m0 ) for m0 6= m, then
the probability of decoding as (l, m0 ) is
0
Tr{Γl,m
ρl,m
} = Tr{TRSC NAB→C (θRA ⊗ γS ⊗ γB )}.
RL S M C RL S M C
(4.7)
If message pair (l, m) was transmitted and Charlie checks for message pair (l0 , m0 ) for l0 6= l and
m0 6= m, then the probability of decoding as (l0 , m0 ) is
0
0
l,m
Tr{ΓlR,m
L S M C ρRL S M C } = Tr{TRSC NAB→C (θR ⊗ θA ⊗ γS ⊗ γB )}.
(4.8)
The above observations are helpful in the forthcoming error analysis.
We now take Charlie’s position-based quantum simultaneous decoder to be the following squareroot measurement:
!−1/2
!−1/2
L X
M
L X
M
X
X
0
0
0
0
l ,m
Γl,m
ΓlR,m
.
(4.9)
Λl,m
≡
ΓR
LSM C
LSM C
RL S M C
RL S M C
l0 =1 m0 =1
l0 =1 m0 =1
Error analysis: Due to the code construction, the error probability under the position-based
coding scheme is the same for each message pair (l, m):
pe (l, m) = Tr{(I − Λl,m
)ρl,m
}.
RL S M C RL S M C
13
(4.10)
Applying Lemma 5 and (4.5)–(4.8), we arrive at the following upper bound on the decoding error
probability:
X
0
0
l,m
l,m
(4.11)
pe (l, m) ≤ cI Tr{(I − Γl,m
)ρ
}
+
c
Tr{ΓlR,m
II
L
M
L
M
L S M C ρRL S M C }
R S C R S C
(l0 ,m0 )6=(l,m)
= cI Tr{(I − Γl,m
)ρl,m
} + cII
RL S M C RL S M C
+ cII
X
m0 6=m
0
Tr{Γl,m
ρl,m
}
RL S M C RL S M C
X
l0 6=l
0
l,m
Tr{ΓlR,m
L S M C ρRL S M C }
+ cII
X
l0 6=l,m0 6=m
0
0
l,m
Tr{ΓlR,m
L S M C ρRL S M C }
(4.12)
= cI Tr{(I − TRSC )NAB→C (θRA ⊗ γSB )} + cII (L − 1) Tr{TRSC NAB→C (θR ⊗ θA ⊗ γSB )}
+ (M − 1) Tr{TRSC NAB→C (θRA ⊗ γS ⊗ γB )}
+ (L − 1)(M − 1) Tr{TRSC NAB→C (θR ⊗ θA ⊗ γS ⊗ γB )} .
(4.13)
The error probability associated with cI is the probability of incorrectly decoding when Charlie
checks for message pair (l, m). The error probabilities associated with cII are the probabilities of
decoding as some other message pair when message pair (l, m) is transmitted. There are L − 1
possibilities for Charlie to erroneously decode Alice’s message and correctly decode Bob’s message,
M − 1 possibilities to erroneously decode Bob’s message and correctly decode Alice’s message, and
(M − 1)(L − 1) possibilities of incorrectly decoding both Alice and Bob’s messages.
One-shot bound for quantum simultaneous decoding. Thus, our bound on the decoding
error probability for a position-based entanglement-assisted coding scheme is as follows:
pe (l, m) ≤ cI Tr{(I − TRSC )NAB→C (θRA ⊗ γSB )} + cII (L − 1) Tr{TRSC NAB→C (θR ⊗ θA ⊗ γSB )}
+ (M − 1) Tr{TRSC NAB→C (θRA ⊗ γS ⊗ γB )}
+ (L − 1)(M − 1) Tr{TRSC NAB→C (θR ⊗ θA ⊗ γS ⊗ γB )} . (4.14)
Interestingly, this bound is the same for all message pairs, and thus holds for maximal or average
error probability. We also remark that the above inequality forges a transparent link between
communication over multiple-access channels and multiple quantum hypothesis testing, a point
that we will return to in Section 5.
Generalization to multiple senders. The above bound can be extended as follows for an
entanglement-assisted quantum multiple-access channel NA1 ···AK →C with K senders and a single
receiver:
(
!)
K
O
pe (m1 , . . . , mK ) ≤ cI Tr (I − TR1 ···RK C )NA1 ···AK →C
θRk A k
k=1
O
X Y
O
θ Rj ⊗ θA j ⊗
θRl Al . (4.15)
(Mj − 1) Tr TR1 ···RK C NA1 ···AK →C
c
+ cII
J ⊆[K]
j∈J
j∈J
l∈J
In the above, m1 , . . . , mK are the messages for senders 1 through K, respectively, chosen from
respective message sets of size M1 , . . . , MK . The states θR1 A1 , . . . , θRK AK are entangled states
shared between the receiver and senders 1 through K, with the receiver possessing all of the R
14
systems. Finally, TR1 ···RK C is a test operator satisfying 0 ≤ TR1 ···RK C ≤ IR1 ···RK C and J is a nonempty subset of [K] ≡ {1, . . . , K}. The above bound is derived by using position-based coding as
described above and a square-root measurement that generalizes (4.9). We omit the straightforward
proof.
4.2
Unassisted classical communication over multiple-access channels
The position-based coding technique is not only a powerful tool for entanglement-assisted classical
communication protocols, but also for those that do not employ entanglement assistance or any
other kind of assistance. This was shown explicitly for the single-sender, single-receiver case in
[Wil17a]. We now demonstrate this point further by considering unassisted classical communication over a classical-input quantum-output multiple-access channel NXY →C . We do so by first
considering classical communication assisted by shared randomness, such that we can employ a
position-based coding scheme, and then we derandomize the protocol to obtain a codebook for
unassisted communication.
The classical-classical-quantum channel that we consider can be written in fully quantum form
as
X
NX 0 Y 0 →C (ωX 0 Y 0 ) =
hx|X 0 hy|Y 0 ωX 0 Y 0 |xiX 0 |yiY 0 ρx,y
(4.16)
C .
x,y
Before communication begins, Alice and Charlie share randomness in the form of the following
classical–classical state:
X
ρXX 0 ≡
pX (x)|xihx|X ⊗ |xihx|X 0 .
(4.17)
x
Similarly Bob and Charlie also share randomness represented by the following classical–classical
state:
X
σY Y 0 ≡
pY (y)|yihy|Y ⊗ |yihy|Y 0 .
(4.18)
y
We demonstrate the procedure of derandomization by proving the following theorem:
Theorem 9 There exists an unassisted, simultaneous decoding protocol for classical communication
over a classical-input quantum-output quantum multiple-access channel with the following upper
bound on its average error probability, holding for all TXY C such that 0 ≤ TXY C ≤ IXY C :
1 X
pe (l, m) ≤ cI Tr{(I − TXY C )NX 0 Y 0 →C (ρX 0 X ⊗ σY Y 0 )}
LM
l,m
+ cII (L − 1) Tr{TXY C NX 0 Y 0 →C (ρX ⊗ ρX 0 ⊗ σY Y 0 )}
+ (M − 1) Tr{TXY C NX 0 Y 0 →C (ρXX 0 ⊗ σY ⊗ σY 0 )}
+ (L − 1) (M − 1) Tr{TXY C NX 0 Y 0 →C (ρX ⊗ ρX 0 ⊗ σY ⊗ σY 0 )} , (4.19)
where L is the number of messages for the first sender and M is the number of messages for the
second sender. A generalization of this statement holds for multiple senders, with an upper bound
on the average error probability given by the right-hand side of (4.15), but with all of the R and A
systems being classical.
15
Proof. The position-based coding scheme operates exactly as
states θRA and γSB replaced by ρXX 0 and σY Y 0 , respectively,
NX 0 Y 0 →C , and the test operator TRSC replaced by TXY C . The
the following bound on the error probability when decoding the
specified in Section 4.1, with the
the channel NAB→C replaced by
same error analysis then leads to
message pair (l, m):
pe (l, m) ≤ cI Tr{(I − TXY C )NX 0 Y 0 →C (ρXX 0 ⊗ σY Y 0 )}
+ cII (L − 1) Tr{TXY C NX 0 Y 0 →C (ρX ⊗ ρX 0 ⊗ σY Y 0 )}
+ (M − 1) Tr{TXY C NX 0 Y 0 →C (ρXX 0 ⊗ σY ⊗ σY 0 )}
+ (L − 1) (M − 1) Tr{TXY C NX 0 Y 0 →C (ρX ⊗ ρX 0 ⊗ σY ⊗ σY 0 )} . (4.20)
Derandomization: Extending the development in [Wil17a], first notice that we can rewrite
the four trace terms in (4.20) as follows:
X
pX (x)pY (y)|xyihxy| ⊗ ρx,y
(4.21)
Tr{TXY C NX 0 Y 0 →C (ρXX 0 ⊗ σY Y 0 )} = Tr{TXY C
C }
x,y
=
X
x,y
Tr{TXY C NX 0 Y 0 →C (ρX ⊗ ρX 0 ⊗ σY Y 0 )} =
X
Tr{TXY C NX 0 Y 0 →C (ρXX 0 ⊗ σY ⊗ σY 0 )} =
X
Tr{TXY C NX 0 Y 0 →C (ρX ⊗ ρX 0 ⊗ σY ⊗ σY 0 )} =
X
x,y
x,y
x,y
x,y
pX (x)pY (y) Tr{Qx,y
C ρC } ,
(4.22)
y
pX (x)pY (y) Tr{Qx,y
C ρ̄C } ,
(4.23)
x
pX (x)pY (y) Tr{Qx,y
C ρ̄C } ,
(4.24)
pX (x)pY (y) Tr{Qx,y
C ρ̄C } ,
(4.25)
where we define the following averaged output states:
X
ρ̄C ≡
pX (x)pY (y)ρx,y
C ,
(4.26)
x,y
ρ̄yC ≡
X
ρ̄xC ≡
X
pX (x)ρx,y
C ,
(4.27)
pY (y)ρx,y
C ,
(4.28)
x
y
and the measurement operator
Qx,y
C ≡ hx, y|XY TXY C |x, yiXY .
(4.29)
Thus, in the case that the code is randomness-assisted, it suffices to take the test operator TXY C
to have the following form:
X
TXY C =
|xihx|X ⊗ |yihy|Y ⊗ Qx,y
(4.30)
C ,
x,y
because, as we will show, the upper bound on the average error probability does not change when
doing so. Then we can rewrite the decoding POVM in (4.9) as follows:
Γl,m
≡ TXl Ym C
XLY M C
X
=
|xL ihxL |X L ⊗ |y M ihy M |Y M ⊗ QxCl ,ym ,
xL ,y M
16
(4.31)
(4.32)
P
where we use the resolution of the identity IX =
x |xihx|X to expand the implicit identity
L
operators and we employ the notation x ≡ x1 · · · xL and y M ≡ y1 · · · yM . Then this implies that
−1/2
!−1/2 L M
L X
M
X
X
X
X
0
0
x 0 ,y 0
l ,m
ΓX
=
|xL ihxL |X L ⊗ |y M ihy M |Y M ⊗ QCl m
(4.33)
LY M C
l0 =1 m0 =1
l0 =1 m0 =1 xL ,y M
=
X
|xL ihxL |X L ⊗ |y M ihy M |Y M ⊗
xL ,y M
X
=
xL ,y M
|xL ihxL |X L ⊗ |y M ihy M |Y M ⊗
L X
M
X
−1/2
x 0 ,ym0
QCl
(4.34)
l0 =1 m0 =1
L X
M
X
!−1/2
x 0 ,y 0
QCl m
. (4.35)
l0 =1 m0 =1
The last step follows from the fact that {|xi}x and {|yi}y form orthonormal bases. Therefore, the
decoding POVM for the randomness-assisted protocol can be decomposed as
!−1/2
!−1/2
L X
M
L X
M
X
X
l0 ,m0
l,m
l,m
l0 ,m0
ΓX L Y M C
ΛX L Y M C ≡
ΓX L Y M C
ΓX L Y M C
,
(4.36)
l0 =1 m0 =1
X
=
L
xL ,y M
L
M
|x ihx |X L ⊗ |y ihy |Y M ⊗
where
ΩxCl ,ym
≡
M
L X
M
X
l0 =1
!−1/2
x 0 ,y 0
QCl m
QxCl ,ym
m0 =1
l0 =1 m0 =1
ΩxCl ,ym ,
L X
M
X
l0 =1
(4.37)
!−1/2
x 0 ,y 0
QCl m
.
(4.38)
m0 =1
By definition, the output state of Charlie in (4.3), for our case of interest, can be written as
X
(4.39)
ρl,m
=
pX L (xL )pY M (y M )|xL ihxL |X L ⊗ |y M ihy M |Y M ⊗ ρxCl ,ym .
L
M
X Y C
xL ,y M
By combining (4.37) and (4.39), we find that the average error probability for the code can be
rewritten as
L M
1 XX
Tr{(IX L Y M C − Λl,m
)ρl,m
}
XLY M C XLY M C
LM
=
l=1 m=1
L X
M
X
1
LM
X
l=1 m=1 xL ,y M
pX L (xL )pY M (y M ) Tr{(IC − ΩxCl ,ym )ρxCl ,ym }
#
L M
1 XX
xl ,ym xl ,ym
pX L (x )pY M (y )
Tr{(IC − ΩC )ρC } .
LM
M
(4.40)
"
=
X
xL ,y
L
M
(4.41)
l=1 m=1
Suppose now that there exists a randomness-assisted position-based code that has an average error
probability ≤ ε. By the above equalities and since the average can never exceed the maximum, we
know there must exist a particular choice of xL , y L such that
L M
1 XX
Tr{(I − ΩxCl ,ym )ρxCl ,ym } ≤ ε .
LM
l=1 m=1
17
(4.42)
Thus for an unassisted multiple-access classical communication protocol, if we choose {xl }L
l=1 as
xl ,ym
M
Alice’s codebook and {ym }m=1 as Bob’s codebook, and the POVM {Ωc
} as Charlie’s decoder,
an upper bound on the average probability error is given by
L M
L M
1 XX
1 XX
pe (l, m) =
Tr{(I − ΩxCl ,ym )ρxCl ,ym } ≤ ε .
LM
LM
l=1 m=1
(4.43)
l=1 m=1
This proves the statement of the theorem after considering the upper bound in (4.20).
4.3
Achievable rate region for i.i.d. case
We now demonstrate rates that are achievable when using a particular quantum simultaneous
decoder. Interestingly, we show how the same quantum simultaneous decoder leads to two generally
different bounds for the achievable rate region. In the first one, the rates are bounded by terms which
consist of the difference of a Rényi entropy of order two and a conditional quantum entropy. In the
second one, the rates are bounded by terms which consist of the difference of a collision conditional
entropy and a conditional quantum entropy. Although these rates are suboptimal when compared
to what is achievable by using successive decoding [Win01, HDW08], we nevertheless think that
the following coding theorems represent progress toward finding a quantum simultaneous decoder.
Before we state the theorems, we require the following definition: A rate pair (R1 , R2 ) is achievable for communication over a quantum multiple access channel if there exists a (2n[R1 −δ] , 2n[R2 −δ] , ε)
code for communication over NA⊗n
0 B 0 →C for all ε ∈ (0, 1), δ > 0, and sufficiently large n.
Theorem 10 An achievable rate region for entanglement-assisted classical communication over a
two-sender quantum multiple-access channel NAB→C , by employing a quantum simultaneous decoder, is as follows:
˜ CR)ω ,
R1 ≤ I(S;
˜ CS)ω ,
R2 ≤ I(R;
˜
R1 + R2 ≤ I(RC;
S)ω ,
(4.44)
(4.45)
(4.46)
where ωRSC ≡ NAB→C (θRA ⊗ γSB ) and θRA and γSB are quantum states. Here we define the
following mutual-information-like quantities:
˜ CS)ω ≡ H2 (SC)ω − H(SC|R)ω ,
I(R;
˜ CR)ω ≡ H2 (RC)ω − H(RC|S)ω ,
I(S;
˜
I(RS;
C)ω ≡ H2 (C)ω − H(C|RS)ω ,
(4.47)
(4.48)
(4.49)
where H2 (A)ρ ≡ − log2 Tr{ρ2A } is the Rényi entropy of order two.
Proof. In our setting, Alice and Bob use an i.i.d. channel NA⊗n
0 B 0 →C . In order to bound the error
probability, we replace each state in (4.14) by its n-copy version. Defining ωRSC ≡ NAB→C (θRA ⊗
γSB ), we choose the test operator T to be the following ‘coated’ typical projector:
ω,δ
ω,δ
ω,δ
ω,δ
TRn S n C n ≡ (Πω,δ
Rn ⊗ ΠS n )ΠRn S n C n (ΠRn ⊗ ΠS n ) ,
18
(4.50)
ω,δ
ω,δ
where Πω,δ
Rn , ΠS n , and ΠRn S n C n are the typical projectors corresponding to the respective states
ωR , ωS , and ωRSC . Applying (4.14), we find the following upper bound on the error probability
when decoding the message pair (l, m):
⊗n
pe (l, m) ≤ cI Tr{(I − TRn S n C n )ωRSC
} + cII L Tr{TRn S n C n [NAB→C (θR ⊗ θA ⊗ γSB )]⊗n )}
+ M Tr{TRn S n C n [NAB→C (θRA ⊗ γS ⊗ γB )]⊗n }
+ LM Tr{TRn S n C n [NAB→C (θR ⊗ θA ⊗ γS ⊗ γB )]⊗n } . (4.51)
We evaluate each term sequentially. To give an upper bound on the first term, consider the following
chain of inequalities, which holds for sufficiently large n:
ω,δ
ω,δ
ω,δ
ω,δ
⊗n
Tr{(Πω,δ
Rn ⊗ ΠS n )ΠRn S n C n (ΠRn ⊗ ΠS n )ωRSC }
ω,δ
ω,δ
ω,δ
ω,δ
⊗n
⊗n
⊗n
≥ Tr{Πω,δ
Rn S n C n ωRSC } − (ΠRn ⊗ ΠS n )ωRSC (ΠRn ⊗ ΠS n ) − ωRSC
√
≥ 1 − ε − 2 2ε .
1
(4.52)
(4.53)
The first inequality follows from Lemma 2. The second inequality follows from (2.20), [HDW08,
Eq. (81)], and the application of Lemma 1. We then obtain an upper bound on the first term:
√
⊗n
} ≤ ε + 2 2ε .
Tr{(I − TRn S n C n )ωRSC
(4.54)
Now we consider the second term in (4.51):
ω,δ
ω,δ
ω,δ
ω,δ
⊗n
)}
L Tr{(Πω,δ
Rn ⊗ ΠS n )ΠRn S n C n (ΠRn ⊗ ΠS n )[NAB→C (θR ⊗ θA ⊗ γSB )]
ω,δ ⊗n
ω,δ
ω,δ
⊗n
⊗n
)}
≤ L 2n[H(RSC)ω +δ] Tr{(Πω,δ
Rn ⊗ ΠS n )ωRSC (ΠRn ⊗ ΠS n )θR ⊗ [NAB→C (θA ⊗ γSB )]
i
h
ω,δ ⊗n ω,δ
ω,δ
⊗n
⊗n
}
= L 2n[H(RSC)ω +δ] Tr{Πω,δ
S n ωRSC ΠS n (ΠRn θR ΠRn ) ⊗ [NAB→C (θA ⊗ γSB )]
ω,δ
⊗n
⊗n
n ⊗ [NAB→C (θA ⊗ γSB )]
Π
I
}
ω
≤ L 2n[H(RSC)ω +δ] 2−n[H(R)ω −δ] Tr{Πω,δ
n
n
R
RSC S
S
≤
≤
⊗n ω,δ ⊗n
L2
Tr{Πω,δ
S n ωSC ΠS n ωSC }
⊗n 2
) }
L 2n[H(SC|R)ω +2δ] Tr{(ωSC
nR1 n[H(SC|R)ω +2δ] −nH2 (SC)ω
n[H(SC|R)ω +2δ]
=2
2
=2
(4.56)
(4.57)
(4.58)
(4.59)
2
−n[H2 (SC)ω −H(SC|R)ω −R1 −2δ]
(4.55)
(4.60)
.
(4.61)
The first inequality follows from the application of the projector trick inequality from (2.23) to the
⊗n
state ωRSC
. The first equality follows from cyclicity of trace. The second inequality follows from
n
the right-hand side of (2.22), the fact that θR = ωR , and the inequality Πω,δ
Rn ≤ IR . The third
inequality follows from a partial trace over the Rn systems. The fourth inequality follows because
⊗n ω,δ ⊗n
⊗n ω,δ ⊗n
⊗n 2 ω,δ
⊗n 2
Tr{Πω,δ
S n ωSC ΠS n ωSC } ≤ Tr{ωSC ΠS n ωSC } = Tr{(ωSC ) ΠS n } ≤ Tr{(ωSC ) },
(4.62)
⊗n ω,δ ⊗n
⊗n 2
n
which is a consequence of the facts that Πω,δ
S n ≤ IS and ωSC ΠS n ωSC and (ωSC ) are positive seminR
definite. Finally, the second equality follows from the fact L = 2 1 and the definition of Rényi
entropy of order two.
Following a similar analysis, we obtain the following upper bounds for the other two terms:
M Tr{TRn S n C n [NAB→C (θRA ⊗ γS ⊗ γB )]⊗n } ≤ 2−n[H2 (RC)ω −H(RC|S)ω −R2 −2δ] ,
LM Tr{TRn S n C n [NAB→C (θR ⊗ θA ⊗ γS ⊗ γB )]⊗n } ≤ 2−n[H2 (C)ω −H(C|RS)ω −(R1 +R2 )−3δ] .
19
(4.63)
(4.64)
Taking the sum of the upper bounds for the above four terms, we find the following upper
bound on the error probability when decoding the message pair (l, m):
√
pe (l, m) ≤ ε + 2 2ε + 2−n[H2 (SC)ω −H(SC|R)ω −R1 −2δ]
+ 2−n[H2 (RC)ω −H(RC|S)ω −R2 −2δ] + 2−n[H2 (C)ω −H(C|RS)ω −(R1 +R2 )−3δ] . (4.65)
Thus, if the rate pair (R1 , R2 ) satisfies the following inequalities (related to those in the statement
of the theorem)
˜ CR)ω ,
R1 + 3δ ≤ I(S;
˜ CS)ω ,
R2 + 3δ ≤ I(R;
˜
R1 + R2 + 4δ ≤ I(RC;
S)ω ,
(4.66)
(4.67)
(4.68)
then the error probability can be made arbitrarily small with increasing n. However, since δ can
be taken arbitrarily small after the limit of large n, we conclude that the rate region given in the
statement of the theorem is achievable.
Our results can be easily extended to the multiple-sender scenario, which we state below without
explicitly writing down a proof (note that the proof is a straightforward generalization of the above
analysis for two senders).
Theorem 11 An achievable rate region for entanglement-assisted classical communication over a
K-sender multiple-access quantum channel NA1 A2 ,...AK →C , by employing a quantum simultaneous
decoder, is given by the following:
X
˜
Rj ≤ I(S(J
); CS(J c ))ω ,
for every J ⊆ [K] ,
(4.69)
j∈J
where ωS1 ···SK C ≡ NA1 ···AK →C (θS1 A1 ⊗ · · · ⊗ θSK AK ). Here we define mutual-information-like quantities
˜
I(S(J
); CS(J c ))ω ≡ H2 (S(J c )C)ω − H(S(J c )C|S(J ))ω ,
(4.70)
where H2 (B)ρ = − log2 Tr{ρ2B } is the Rényi entropy of order two.
By combining Theorems 9 and 11, we obtain the following rate region that is achievable for
unassisted classical communication over a quantum multiple-access channel when using a quantum
simultaneous decoder:
Theorem 12 The following rate region is achievable when using a quantum simultaneous decoder
for unassisted classical communication over the K-sender, classical-input quantum-output multipleaccess channel x1 , . . . , xK → ρx1 ,...,xK :
X
˜
Rj ≤ I(X(J
); CX(J c ))ω ,
for every J ⊆ [K] ,
(4.71)
j∈J
where
ωX1 ···XK C ≡
X
x1 ,...,xK
pX1 (x1 ) · · · pXK (xK )|x1 ihx1 |X1 ⊗ · · · ⊗ |xK ihxK |XK ⊗ ρx1 ,...,xK .
20
(4.72)
The following alternative achievable rate region is generally different from the one in Theorem 11:
Theorem 13 An achievable rate region for entanglement-assisted classical communication over the
K-sender quantum multiple-access channel NA1 A2 ,...AK →C , by employing a quantum simultaneous
decoder, is given by the following:
X
Rj ≤ I 0 (S(J ); CS(J c ))ω ,
for every J ⊆ [K] ,
(4.73)
j∈J
where ωS1 ···SK C ≡ NA1 ···AK →C (θS1 A1 ⊗ · · · ⊗ θSK AK ). Here we define mutual-information-like quantities
I 0 (S(J ); CS(J c ))ω ≡ H2 (C|S(J c ))ω − H(C|S1 · · · SK )ω ,
(4.74)
−1/2
where H2 (A|B)ρ = − log2 Tr{ρAB ρB
−1/2
ρAB ρB
} is the collision conditional entropy [DFW15].
Proof. We prove this theorem for the case of two senders, and then the extension to three or more
senders is straightforward. The analysis proceeds similarly as in the proof of Theorem 10, and we
pick up at (4.58). We find that
⊗n ω,δ ⊗n
L 2n[H(SC|R)ω +2δ] Tr{Πω,δ
S n ωSC ΠS n ωSC }
⊗n
⊗n
}
(ω ⊗n )−1/2 ωSC
≤ L 2n[H(SC|R)ω +2δ] 2−n[H(S)ω −δ] Tr{(ωS⊗n )−1/2 ωSC
h
in S
−1/2
−1/2
= 2nR1 2n[H(C|RS)ω +3δ] Tr{ωS ωSC ωS ωSC }
= 2−n[H2 (C|S)ω −H(C|RS)ω −R1 −3δ] .
(4.75)
(4.76)
(4.77)
The inequality follows because
⊗n ω,δ ⊗n
Tr{Πω,δ
S n ωSC ΠS n ωSC }
⊗n ω,δ ⊗n
ΠS n ωSC }
≤ 2−n[H(S)ω −δ]/2 Tr{(ωS⊗n )−1/2 ωSC
(4.78)
=
(4.79)
≤
⊗n ω,δ
⊗n
ΠS n }
(ωS⊗n )−1/2 ωSC
2
Tr{ωSC
⊗n
⊗n
(ωS⊗n )−1/2 },
(ωS⊗n )−1/2 ωSC
2−n[H(S)ω −δ]/2 2−n[H(S)ω −δ]/2 Tr{ωSC
−n[H(S)ω −δ]/2
(4.80)
⊗n
⊗n
⊗n ω,δ ⊗n
≥ 0. The
(ωS⊗n )−1/2 ωSC
where we apply (2.24) twice and the facts that ωSC
ΠS n ωSC ≥ 0 and ωSC
equality in (4.76) follows because H(S)ω = H(S|R)ω for the state ω and then because H(SC|R)ω −
H(S|R)ω = H(C|RS)ω . The equality in (4.77) follows from the definition of the collision conditional
entropy.
Following a similar analysis, we obtain the following upper bounds for the other two terms:
M Tr{TRn S n C n [NAB→C (θRA ⊗ γS ⊗ γB )]⊗n } ≤ 2−n[H2 (C|R)ω −H(C|RS)ω −R2 −3δ] ,
LM Tr{TRn S n C n [NAB→C (θR ⊗ θA ⊗ γS ⊗ γB )]⊗n } ≤ 2−n[H2 (C)ω −H(C|RS)ω −(R1 +R2 )−3δ] .
(4.81)
(4.82)
The rest of the proof is then the same as in the proof of Theorem 10.
By combining Theorems 9 and 13, we obtain the following rate region that is achievable for
unassisted classical communication over a quantum multiple-access channel when using a quantum
simultaneous decoder:
21
Theorem 14 The following rate region is achievable when using a quantum simultaneous decoder
for unassisted classical communication over the K-sender, classical-input quantum-output multipleaccess channel x1 , . . . , xK → ρx1 ,...,xK :
X
Rj ≤ I 0 (X(J ); CX(J c ))ω ,
for every J ⊆ [K] ,
(4.83)
j∈J
where
ωX1 ···XK C ≡
X
x1 ,...,xK
pX1 (x1 ) · · · pXK (xK )|x1 ihx1 |X1 ⊗ · · · ⊗ |xK ihxK |XK ⊗ ρx1 ,...,xK .
Observe that H2 (A|Y )σ = − log
P
y
p(y) Tr{σy2 } for a state σY A =
P
y
(4.84)
p(y)|yihy|Y ⊗ σy .
The rate region given in Theorem 14 is arguably an improvement over that from [FHS+ 12,
Theorem 6]. By using the coding technique from [FHS+ 12, Theorem 6], for an m-sender multiple
access channel, only m of the inequalities feature the conditional von Neumann entropy as the first
term in the entropy differences, whereas the other 2m −m−1 inequalities feature the conditional minentropy. On the other hand, all 2m − 1 inequalities in Theorem 14 feature the collision conditional
entropy (conditional Rényi entropy of order two), which is never smaller than the conditional minentropy. Thus, as the number m of senders grows larger, the volume of the achievable rate region
from Theorem 14 is generally significantly larger than the volume of the achievable rate region from
[FHS+ 12, Theorem 6].
We can also compare the rate regions from Theorems 12 and 14. By using the identity
H2 (X(J c )C)ω − H(X(J c )C|X(J ))ω
= H2 (X(J c )C)ω − H(C|X1 · · · XK )ω − H(X(J c )|X(J ))ω
c
c
= H2 (X(J )C)ω − H(C|X1 · · · XK )ω − H(X(J ))ω ,
(4.85)
(4.86)
the difference between the information-theoretic terms in the inequalities in (4.73) and (4.69) is
given by
H2 (C|X(J c ))ω − H2 (X(J c )C)ω + H(X(J c ))ω .
(4.87)
The above difference can sometimes be negative or positive, as one can find by some simple numerical tests with qubit states. Thus, the two rate regions from Theorems 12 and 14 are generally
incomparable. However, we can easily make use of both results: a standard time-sharing argument
establishes that the convex hull of the two rates regions is achievable.
Remark 15 The quantum simultaneous decoding conjecture from [FHS+ 12, Wil11] is the statement that the Rényi entropies of order two in Theorems 12 and 14 can be replaced by quantum
entropies (i.e., von Neumann entropies), while still employing a quantum simultaneous decoder.
An explicit statement of the quantum simultaneous decoding conjecture is available in [FHS+ 12,
Conjecture 4]. The conjecture has been solved in the case of two senders for unassisted classical
communication in [FHS+ 12] and for entanglement-assisted classical communication in [XW13], but
not for three or more senders in either case.
22
5
Quantum simultaneous decoding for multiple-access channels
and multiple quantum hypothesis testing
In this section, we establish explicit links between the quantum simultaneous decoding conjecture
from [FHS+ 12, Wil11] and open questions from [AM14, BHOS15] in multiple quantum hypothesis testing. Recall that the general goal of quantum hypothesis testing is to minimize the error
probability in identifying quantum states. In binary quantum hypothesis testing, one considers
two hypotheses: the null hypothesis is that a quantum system is prepared in the state ρ, and the
alternative hypothesis is that the quantum system is prepared in the state σ, where ρ, σ ∈ D(H).
Operationally, the discriminator receives the state ρ with probability p ∈ (0, 1) and the state σ
with probability 1 − p, and the task is to determine which state was prepared, by means of some
quantum measurement {T, I − T }, where the test operator T satisfies 0 ≤ T ≤ I. There are two
kinds of errors: a Type I error occurs when the state is identified as σ when in fact ρ was prepared
and a Type II error is the opposite kind of error. The error probabilities corresponding to the two
types of errors are as follows:
α(T, ρ) ≡ Tr{(I − T )ρ},
β(T, σ) ≡ Tr{T σ} .
(5.1)
(5.2)
As in information theory, quantum hypothesis testing has been studied in the asymptotic
i.i.d. setting. In the setting of symmetric hypothesis testing, we are interested in minimizing
the overall error probability
Pe∗ (pρ, (1 − p)σ) ≡
=
inf
T : 0≤T ≤I
p α(T, ρ) + (1 − p)β(T, σ)
1
(Tr{pρ + (1 − p)σ} − kpρ − (1 − p)σk1 ) .
2
(5.3)
(5.4)
In the i.i.d. setting, n quantum systems are prepared as either ρ⊗n or σ ⊗n , and the goal is to
determine the optimal exponent of the error probability, defined as
1
lim − log2 Pe∗ (pρ⊗n , (1 − p)σ ⊗n ).
n→∞ n
(5.5)
One of the landmark results in quantum hypothesis testing is that the optimal exponent is equal
to the quantum Chernoff distance [ACMnT+ 07, NS09]:
1
lim − log2 Pe∗ (pρ⊗n , (1 − p)σ ⊗n ) = C(ρ, σ) ≡ sup − log2 Tr{ρs σ 1−s }.
n
s∈[0,1]
n→∞
(5.6)
This development can be generalized to the setting in which ρ, σ, p, and 1 − p can be replaced
by positive semi-definite operators 6= 0 and positive constants. Indeed, for positive semi-definite A
and B, we can define
Pe∗ (A, B) ≡
=
inf
T : 0≤T ≤I
Tr{(I − T )A} + Tr{T B}
1
(Tr{A + B} − kA − Bk1 ) .
2
23
(5.7)
(5.8)
Then for positive constants K0 , K1 > 0, we find that
1
lim − log2 Pe∗ (K0 A⊗n , K1 B ⊗n ) = C(A, B) ≡ sup − log2 Tr{As B 1−s }.
n
s∈[0,1]
n→∞
(5.9)
In the setting of asymmetric hypothesis testing, we are interested in the optimal exponent of the
Type II error β(T, σ), under a constraint on the Type I error, i.e., α(T, σ) ≤ ε, with ε ∈ [0, 1]. That
is, we are interested in the following quantity, now known as hypothesis testing relative entropy:
ε
DH
(ρkσ) ≡ − log2 inf {β(T, σ) : 0 ≤ T ≤ I ∧ α(T, σ) ≤ ε}.
T
(5.10)
The optimal exponential decay rate in the asymmetric setting is given by the quantum Stein’s
lemma [HP91, ON00], which establishes the following for all ε ∈ (0, 1):
1 ε ⊗n ⊗n
D (ρ kσ ) = D(ρkσ),
n→∞ n H
lim
(5.11)
giving the quantum relative entropy its fundamental operational interpretation.
As we can see from [AJW17b] and our developments in Section 3, position-based coding forges
a direct connection between single-sender single-receiver communication and binary quantum hypothesis testing. Specifically, the Chernoff distance from symmetric hypothesis testing gives a lower
bound on the entanglement-assisted error exponent; while the application of the results from asymmetric hypothesis testing leads to a lower bound on the one-shot entanglement-assisted capacity
and in turn on the second-order coding rate for entanglement-assisted communication.
In what follows, we discuss the generalization of both asymmetric and symmetric hypothesis
testing to multiple quantum states and their connections to quantum simultaneous decoding.
5.1
Symmetric multiple quantum hypothesis testing and quantum simultaneous
decoding
We now tie one version of the quantum simultaneous decoding conjecture to [AM14, Conjecture 4.2],
which has to do with distinguishing one state from a set of other possible states. To recall the setting
of [AM14, Conjecture 4.2], suppose that a state ρ is prepared with probability p ∈ (0, 1) and with
probability 1 − p one state σi of r states is prepared with probability qi , where i ∈ {1, . . . , r}. The
goal is to determine whether ρ was prepared or whether one of the other states was prepared, and
the error probability in doing so is equal to
!
r
X
∗
Pe pρ, (1 − p)
qi σi .
(5.12)
i=1
The measurement operator that achieves the minimum error probability is equal to
(
)
r
X
pρ − (1 − p)
qi σi ≥ 0 .
(5.13)
i=1
As usual, we are interested in the i.i.d. case, in which ρ and σi are replaced by ρ⊗n and σi⊗n for
large n, and [AM14, Conjecture 4.2] states that
!
r
X
1
lim − log2 Pe∗ pρ⊗n , (1 − p)
qi σi⊗n = min C(ρ, σi ).
(5.14)
n→∞ n
i
i=1
24
We now propose a slight generalization of [AM14, Conjecture 4.2] and (5.9), in which ρ and σi are
replaced by positive semi-definite operators and p and 1 − p are replaced by positive constants:
Conjecture 16 Let {A, B1 , . . . , Br } be a set of positive semi-definite operators with trace strictly
greater than zero, and let K0 , . . . , Kr be strictly positive constants. Then the following equality
holds
!
r
X
1
lim − log2 Pe∗ K0 A⊗n ,
Ki Bi⊗n = min C(A, Bi ) ,
(5.15)
n→∞ n
i
i=1
where
Pe∗
is defined in (5.7)–(5.8) and C(A, Bi ) in (5.9).
To see how Conjecture 16 is connected to quantum simultaneous decoding, recall from (4.14)
our bound on the error probability when simultaneously decoding the message pair (l, m):
pe (l, m) ≤ 4 Tr{(I − TRSC )NAB→C (θRA ⊗ γSB )} + 4 L Tr{TRSC NAB→C (θR ⊗ θA ⊗ γSB )}
+ M Tr{TRSC NAB→C (θRA ⊗ γS ⊗ γB )} + LM Tr{TRSC NAB→C (θR ⊗ θA ⊗ γS ⊗ γB )} , (5.16)
where we have set c = 1 and used that L − 1 < L and M − 1 < M . Now applying this bound to
the i.i.d. case and setting L = 2nR1 and M = 2nR2 , we find that the upper bound becomes
(5.17)
pe (l, m) ≤ 4 Tr{(I − T )ρ⊗n } + Tr{T B1⊗n + B2⊗n + B3⊗n } ,
where
ρ ≡ NAB→C (θRA ⊗ γSB ),
R1
B1 ≡ 2
(5.18)
NAB→C (θR ⊗ θA ⊗ γSB ),
(5.19)
B2 ≡ 2R2 NAB→C (θRA ⊗ γS ⊗ γB ),
R1 +R2
B3 ≡ 2
(5.20)
NAB→C (θR ⊗ θA ⊗ γS ⊗ γB ).
(5.21)
To minimize the upper bound on the error probability, we should pick the test operator T as follows:
(5.22)
T ≡ ρ⊗n − B1⊗n + B2⊗n + B3⊗n ≥ 0 ,
and then the upper bound becomes
pe (l, m) ≤ 2 Tr{ρ⊗n + B1⊗n + B2⊗n + B3⊗n } − ρ⊗n − B1⊗n + B2⊗n + B3⊗n
= 4 Pe∗ ρ⊗n , B1⊗n + B2⊗n + B3⊗n .
1
(5.23)
(5.24)
The test operator T given in (5.22) was previously realized in [Wil10] and [HC17] to be relevant
as a quantum simultaneous decoder in the context of unassisted classical communication over a
quantum multiple access channel.
Now applying Conjecture 16 (provided it is true), we would find that the error probability
pe (l, m) is bounded from above as pe (l, m) . e−nE(R1 ,R2 ) , with the error exponent E(R1 , R2 ) equal
25
to
E(R1 , R2 ) = min{C(ρ, B1 ), C(ρ, B2 ), C(ρ, B3 )}
n
= min sup − log2 Tr{ρs B11−s }, sup − log2 Tr{ρs B21−s },
s∈[0,1]
(5.25)
s∈[0,1]
sup − log2 Tr{ρs B31−s }
o
(5.26)
s∈[0,1]
= min
n
h
i
h
i
sup (1 − s) Is (R; CS)ω − R1 , sup (1 − s) Is (S; CR)ω − R2 ,
s∈[0,1]
s∈[0,1]
h
i
sup (1 − s) Is (RS; C)ω − (R1 + R2 ) .
(5.27)
s∈[0,1]
Thus the rate region (R1 , R2 ) would be achievable as long as E(R1 , R2 ) > 0. Now using the fact
that, for a bipartite state ρAB
lim Is (A; B)ρ = I(A; B)ρ ,
(5.28)
s→1
we would then find that the following rate region is achievable:
R1 ≤ I(R; CS)ω ,
(5.29)
R1 + R2 ≤ I(RS; C)ω ,
(5.31)
R2 ≤ I(S; CR)ω ,
(5.30)
where ωRSC = NAB→C (θRA ⊗ γSB ) and the above approach would solve the quantum simultaneous decoding conjecture. The method clearly extends to more than two senders. We remark
that aspects of the above approach were discussed in the recent work [HC17] for the case of unassisted classical communication over a quantum multiple access channel, but there the connection
to [AM14, Conjecture 4.2] or Conjecture 16 was not realized, nor was the entanglement-assisted
case considered.
We end this section by noting that [AM14, Theorem
4.3] offers several suboptimal upper bounds
⊗n
⊗n
⊗n
⊗n
∗
on the error probability Pe ρ , B1 + B2 + B3 , which in turn could be used to establish
suboptimal achievable rate regions for the quantum multiple-access channel. However, here we
refrain from the details of what these regions would be, except to say that they would be in terms
of the negative logarithm of the fidelity, replacing Is in (5.27).
5.2
Asymmetric hypothesis testing with composite alternative hypothesis
We now tie the quantum simultaneous decoding problem to a different open question in asymmetric hypothesis testing. Recall our upper bound from (4.14) on the error probability for classical
communication over a quantum multiple-access channel, as applied for the i.i.d. case:
h
i
pe (l, m) ≤ cI Tr{(I − T )ρ⊗n } + cII Tr{T B1⊗n + B2⊗n + B3⊗n } ,
(5.32)
26
where L = 2nR1 , M = 2nR2 , and the state ρ and the positive semi-definite operators B1 , B2 , and
B3 are given by
ρ = NAB→C (θRA ⊗ γSB ) ,
(5.33)
B1 = 2R1 NAB→C (θR ⊗ θA ⊗ γSB ),
(5.34)
B3 = 2R1 +R2 NAB→C (θR ⊗ θA ⊗ γS ⊗ γB ).
(5.36)
R2
B2 = 2
NAB→C (θRA ⊗ γS ⊗ γB ),
(5.35)
Rather than try to minimize the overall error probability as we did in the previous section, we could
try to minimize all of the other error probabilities subject to a constraint on the error probability
Tr{(I − T )ρ⊗n }. Thus, we seek a test operator T which is capable of discriminating the state ρ⊗n
from the operator B1⊗n + B2⊗n + B3⊗n . This kind of task is formally called asymmetric hypothesis
testing with composite alternative hypothesis. The problem of a composite null hypothesis is that
considered in the context of the quantum Sanov theorem, which was solved in [BDK+ 05] (see also
[Hay02]), and finds application in communication over compound channels [BB09, Mos15] (see also
[DD07, Hay09b]).
The following open question, strongly related to a question from [BHOS15], is relevant for
asymmetric hypothesis testing with composite alternative hypothesis:
Question 17 Consider a quantum state ρ ∈ D(H), a positive integer r > 1, and a finite set of
positive semi-definite operators B ≡ {Bi : 1 ≤ i ≤ r}, for which supp(ρ) ⊆ supp(Bi ) ∀Bi ∈ B and
min D(ρkBi ) > 0.
i
(5.37)
What is the most general form that ρ and B can take such that the following statement is true?
For all ε ∈ (0, 1), δ > 0, and sufficiently large n, there exists a binary test {T, I − T } such that the
Type I error is bounded from above by ε:
Tr{(I − T )ρ⊗n } ≤ ε ,
(5.38)
and for all Bi ∈ B, the exponential decay rate of the Type II error is bounded from below as follows:
1
− log2 Tr{T Bi⊗n } ≥ min D(ρkBi ) − δ .
(5.39)
i
n
Below we prove the following special case:
Theorem 18 The statement at the end of Question 17 is true when the set B forms a commuting
set of operators (each of which need not commute with ρ).
Proof. To this end, we employ the notion of a relative typical projector, which was used in [BSS03]
to establish an alternate proof of the quantum
Stein lemma (see also [BLW15] for a different use
P
of relative typical projectors). Let Bi = y fYi (y)|φiy ihφiy | denote a spectral decomposition of Bi .
δ,n
For a state ρ and positive semi-definite operator Bi , define the relative typical subspace Tρ|B
for
i
δ > 0 and integer n ≥ 1 as
1
δ,n
i
i
n
Tρ|Bi ≡ span |φyn i : − log2 (fY n (y )) + Tr{ρ log2 Bi } ≤ δ ,
(5.40)
n
27
where
y n ≡ y1 · · · yn ,
n
Y
i
n
fY n (y ) ≡
fYi (yj ) ,
|φiyn i ≡
j=1
|φiy1 i
(5.41)
(5.42)
⊗ · · · ⊗ |φiyn i.
(5.43)
δ,n
Let Πnρ|Bi ,δ denote the projection operator corresponding to the relative typical subspace Tρ|B
. The
i
critical properties of the relative typical projector are as follows:
Tr{Πnρ|Bi ,δ ρ⊗n } ≥ 1 − ε,
(5.44)
2−n[− Tr{ρ log2 Bi }+δ] Πnρ|Bi ,δ ≤ Πnρ|Bi ,δ Bi⊗n Πnρ|Bi ,δ ≤ 2−n[− Tr{ρ log2 Bi }−δ] Πnρ|Bi ,δ ,
(5.45)
with the first inequality holding for all ε ∈ (0, 1), δ > 0, and sufficiently large n.
The main idea for the proof under the stated assumptions is to take the test operator T as
T = Πnρ|Br ,δ · · · Πnρ|B1 ,δ Πnρ,δ Πnρ|B1 ,δ · · · Πnρ|Br ,δ ,
(5.46)
where Πnρ,δ is the typical projector for ρ. Then we find that for all ε ∈ (0, 1), δ > 0, and sufficiently
large n
Tr{T ρ⊗n } ≥ Tr{Πnρ,δ ρ⊗n } −
√
≥ 1 − ε − 2r ε,
r
X
i=1
Πnρ|Bi ,δ ρ⊗n Πnρ|Bi ,δ − ρ⊗n
1
(5.47)
(5.48)
which follows by applying Lemmas 1 and 2 and properties of typicality and relative typicality. To
handle the other kind of error, consider that, from the assumption, all of the projectors Πnρ|Bi ,δ
commute, so that
Tr{T Bi⊗n } = Tr{Πnρ|Br ,δ · · · Πnρ|B1 ,δ Πnρ,δ Πnρ|B1 ,δ · · · Πnρ|Br ,δ Bi⊗n }
=
=
≤
(5.49)
Tr{Πnρ,δ Πnρ|B1 ,δ · · · Πnρ|Br ,δ Bi⊗n Πnρ|Br ,δ · · · Πnρ|B1 ,δ }
Tr{Πnρ,δ Πnρ|B1 ,δ · · · Πnρ|Br ,δ Πnρ|Bi ,δ Bi⊗n Πnρ|Bi ,δ Πnρ|Br ,δ · · · Πnρ|B1 ,δ }
2−n[− Tr{ρ log2 Bi }−δ] Tr{Πnρ,δ Πnρ|B1 ,δ · · · Πnρ|Br ,δ Πnρ|Bi ,δ Πnρ|Br ,δ · · · Πnρ|B1 ,δ }
≤ 2−n[− Tr{ρ log2 Bi }−δ] Tr{Πnρ,δ }
(5.50)
(5.51)
(5.52)
(5.53)
≤ 2−n[− Tr{ρ log2 Bi }−δ] 2n[H(ρ)+δ]
(5.54)
=2
(5.55)
−n[D(ρkBi )−2δ]
.
√
The statement of the theorem follows by setting ε0 ≡ ε + 2r ε and δ 0 ≡ 2δ, considering that we
have shown the existence of a test T for which
Tr{T ρ⊗n } ≥ 1 − ε0 ,
0
Tr{T Bi⊗n } ≤ 2−n[D(ρkBi )−δ ] ,
(5.56)
and it is possible to satisfy this for any choice of ε0 ∈ (0, 1) and δ 0 > 0 by taking n sufficiently large.
(Note that for the bound on the second kind of error probability to be decaying exponentially, we
28
require δ 0 > 0 to be small enough so that mini D(ρkBi ) > δ 0 .) We remark that this conclusion is
actually stronger than what is stated in Question 17 because here we conclude for all Bi ∈ B, that
1
− log2 Tr{T Bi⊗n } ≥ D(ρkBi ) − δ 0 ≥ min D(ρkBi ) − δ 0 .
(5.57)
i
n
This concludes the proof.
To see how Question 17 is related to quantum simultaneous decoding of the multiple-access
channel, consider that for ρ, B1 , . . . , B3 as defined in (5.33)–(5.36), the inequality
min D(ρkBi ) > 0
(5.58)
i
is equivalent to the following set of inequalities:
R1 < I(R; CS)ω ,
(5.59)
R2 < I(S; CR)ω ,
(5.60)
R1 + R2 < I(RS; C)ω ,
(5.61)
where ωRSC = NAB→C (θRA ⊗ γSB ). This equivalence holds because mini D(ρkBi ) > 0 is equivalent
to the following three inequalities:
D(ρkB1 ) > 0,
D(ρkB2 ) > 0,
D(ρkB3 ) > 0,
(5.62)
and
D(ρkB1 ) = D(NAB→C (θRA ⊗ γSB )k2R1 NAB→C (θR ⊗ θA ⊗ γSB ))
(5.63)
= D(NAB→C (θRA ⊗ γSB )kNAB→C (θR ⊗ θA ⊗ γSB )) − R1
(5.64)
R2
(5.66)
= I(R; CS)ω − R1 ,
D(ρkB2 ) = D(NAB→C (θRA ⊗ γSB )k2
NAB→C (θRA ⊗ γS ⊗ γB ))
= D(NAB→C (θRA ⊗ γSB )kNAB→C (θRA ⊗ γS ⊗ γB )) − R2 ,
= I(S; CR)ω − R2 ,
D(ρkB3 ) = D(NAB→C (θRA ⊗ γSB )k2
R1 +R2
NAB→C (θR ⊗ θA ⊗ γSB ))
= D(NAB→C (θRA ⊗ γSB )kNAB→C (θR ⊗ θA ⊗ γSB )) − (R1 + R2 )
= I(RS; C)ω − (R1 + R2 ).
(5.65)
(5.67)
(5.68)
(5.69)
(5.70)
(5.71)
Thus, if such a sequence of test operators existed as stated in Question 17, then the error probability
pe (l, m) when decoding a multiple-access channel could be bounded from above as
h
i
pe (l, m) ≤ cI Tr{(I − T )ρ⊗n } + cII Tr{T B1⊗n + B2⊗n + B3⊗n }
(5.72)
h
i
≤ cI ε + 3cII 2−n[mini D(ρkBi )−δ] ,
(5.73)
where ρ, B1 , . . . , B3 are defined in (5.33)–(5.36). Then by choosing the rates R1 and R2 to satisfy
R1 + 2δ ≤ I(R; CS)ω ,
(5.74)
R1 + R2 + 2δ ≤ I(RS; C)ω ,
(5.76)
R2 + 2δ ≤ I(S; CR)ω ,
29
(5.75)
which is equivalent to mini D(ρkBi ) ≥ 2δ, we would have
pe (l, m) ≤ cI ε + 3cII 2−nδ ,
(5.77)
and we could thus make the error probability as small as desired by taking n sufficiently large. Since
δ > 0 is arbitrary, we could then say that the rate region in (5.29)–(5.31) would be achievable. If
the statement at the end of Question 17 holds for the states given above, then the method would
clearly lead to a quantum simultaneous decoder for more than two senders, by a straightforward
generalization of the above approach.
The authors of [BP10] considered a similar problem in asymmetric hypothesis testing for a
specific family of states with certain permutation symmetry. We should point out that in [BP10],
the lower bound for the exponential rate of the Type II error is given by a regularized version of
mini D(ρkσi ). If a similar result held, along the lines stated in Question 17 and related to the
conjecture in [BHOS15], without the need for regularization and for the operators in (5.33)–(5.36)
(and more general ones relevant for more senders), then the developments in the present paper
would immediately give bounds for the performance of quantum simultaneous decoding for the
quantum multiple-access channel.
We end this section by remarking that our quantum simultaneous decoder in (4.50) gives a
method for distinguishing ρ⊗n from the operator B1⊗n + B2⊗n + B3⊗n with Type I error probability
bounded, for sufficiently large n, by an arbitrary ε ∈ (0, 1) and the Type II error probability
bounded, for a positive constant c, by
≈ c 2−n min{I(S;CR)ω −R1 ,
˜
˜
˜
I(R;CS)
ω −R2 , I(RC;S)ω −(R1 +R2 )}
0 (S;CR) −R ,
ω
1
I 0 (R;CS)ω −R2 , I 0 (RC;S)ω −(R1 +R2 )}
and
≈ c 2−n min{I
,
,
(5.78)
(5.79)
where ωRSC = NAB→C (θRA ⊗ γSB ). To obtain the first statement, we apply the analysis in the
proof of Theorem 10, and for the second, we apply the analysis in the proof of Theorem 13. The
above statements have clear generalizations to more systems by invoking Theorems 11 and 13.
Thus, our previous analysis in the context of communication applies for this interesting, special
case of Question 17, albeit with suboptimal rates.
6
Conclusion
In this paper, we apply position-based coding to establish bounds of various quantities for classical
communication. For entanglement-assisted classical communication over point-to-point quantum
channels, we establish lower bounds on the one-shot error exponent, the one-shot capacity, and the
second-order coding rate. We also give an achievable rate region for entanglement-assisted classical
communication over multiple-access quantum channels. Furthermore, we explicitly show how to
derandomize a randomness-assisted protocol (for multiple-access channel) to one without assistance
from any extra resources. Our results indicate that position-based coding can be a powerful tool
in achievability proofs of various communication protocols in quantum Shannon theory (for recent
applications of position-based coding to private classical communication, see [Wil17a]). We finally
tied some open questions in multiple quantum hypothesis testing to the quantum simultaneous
decoding conjecture. Thus, we have shown that open problems in multiple quantum hypothesis
30
testing are fundamental to the study of classical information transmission over quantum multipleaccess channels.
Acknowledgements. We are grateful to Anurag Anshu, Mario Berta, Fernando Brandao,
Masahito Hayashi, Rahul Jain, and Marco Tomamichel for discussions related to the topic of
this paper. HQ acknowledges support from the Air Force Office of Scientific Research, the Army
Research Office, the National Science Foundation, and the Northrop Grumman Corporation. QLW
is supported by the NFSC (Grants No. 61272057, No. 61309029 and No. 61572081) and funded by
the China Scholarship Council (Grant No. 201506470043). MMW acknowledges support from the
Office of Naval Research and the National Science Foundation.
A
Proof of Proposition 3 and quantum Stein’s lemma
Here we provide a proof of Proposition 3. After doing that, we discuss briefly how Proposition 3
and [CMW16, Lemma 5] lead to a transparent proof of the quantum Stein’s lemma [HP91, ON00].
Proof of Proposition 3. The statement of Proposition 3 is trivially true if ρσ = 0, since both
ε (ρkσ) = ∞ and D (ρkσ) = ∞ in this case. So we consider the non-trivial case when this
DH
α
equality does not hold. We exploit Lemma 4 to establish the above bound. The proof also bears
some similarities to a related proof in [AJW17a]. Recall from Lemma 4 that the following inequality
holds for positive semi-definite operators A and B and for α ∈ (0, 1):
inf
T :0≤T ≤I
1
(Tr{A + B} − kA − Bk1 )
2
≤ Tr{Aα B 1−α }.
Tr{(I − T )A} + Tr{T B} =
(A.1)
(A.2)
For p ∈ (0, 1), pick A = pρ and B = (1 − p) σ. Plugging in to the above inequality, we find that
there exists a measurement operator T ∗ = T (p, ρ, σ) such that
p Tr{(I − T ∗ )ρ} + (1 − p) Tr{T ∗ σ} ≤ pα (1 − p)1−α Tr{ρα σ 1−α }.
(A.3)
p Tr{(I − T ∗ )ρ} ≤ pα (1 − p)1−α Tr{ρα σ 1−α },
(A.4)
This implies that
and in turn that
∗
Tr{(I − T )ρ} ≤
1−p
p
1−α
Tr{ρα σ 1−α }.
(A.5)
For a given ε ∈ (0, 1) and α ∈ (0, 1), we pick p ∈ (0, 1) such that
1−p
p
1−α
Tr{ρα σ 1−α } = ε.
31
(A.6)
This is possible because we can rewrite the above equation as
ε=
⇔
1
−1
p
1−p
p
1−α
1−α
α 1−α
Tr{ρ σ
}=
1
−1
p
1−α
Tr{ρα σ 1−α }
ε
=
(A.7)
(A.8)
Tr{ρα σ 1−α }
1/(1−α)
1
ε
+1
=
p
Tr{ρα σ 1−α }
1
⇔p=
∈ (0, 1) .
α
1−α
[ε/ Tr{ρ σ
}]1/(1−α) + 1
⇔
(A.9)
(A.10)
This means that T (p, ρ, σ) with p selected as above is a measurement such that
Tr{(I − T ∗ )ρ} ≤ ε.
(A.11)
Now using the fact that the measurement T ∗∗ for the hypothesis testing relative entropy achieves
the smallest type II error probability (by definition) and the fact that
(1 − p) Tr{T ∗ σ} ≤ pα (1 − p)1−α Tr{ρα σ 1−α }
implies
∗
Tr{T σ} ≤
we find that
∗∗
Tr{T σ} ≤
Considering that
ε=
1−p
p
1−α
p
1−p
α
p
1−p
α 1−α
Tr{ρ σ
Tr{ρα σ 1−α },
(A.13)
Tr{ρα σ 1−α }.
(A.14)
α
}=
(A.12)
p
1−p
α−1
Tr{ρα σ 1−α }
(A.15)
implies that
we get that
∗∗
ε
Tr{ρα σ 1−α }
Tr{T σ} ≤
p
1−p
α
1/(α−1)
=
p
,
1−p
Tr{ρα σ 1−α }
1/(α−1) !α
ε
Tr{ρα σ 1−α }
=
Tr{ρα σ 1−α }
α/(1−α)
= εα/(α−1) Tr{ρα σ 1−α }
Tr{ρα σ 1−α }
1/(1−α)
.
= εα/(α−1) Tr{ρα σ 1−α }
32
(A.16)
(A.17)
(A.18)
(A.19)
(A.20)
Then, by taking a logarithm, we get that
1/(1−α)
− log2 Tr{T ∗∗ σ} ≥ − log2 εα/(α−1) Tr{ρα σ 1−α }
α
1
log2 (ε) +
log2 Tr{ρα σ 1−α }
α−1
α−1
α
log2 (ε) + Dα (ρkσ).
=−
α−1
=−
(A.21)
(A.22)
(A.23)
Putting everything together, we conclude the statement of Proposition 3.
We now briefly discuss how Proposition 3 and [CMW16, Lemma 5], once established, lead
to a transparent proof of the quantum Stein’s lemma [HP91, ON00] (see also [Hay07] in this
context). Before doing so, let us recall that the quantum Stein’s lemma (with strong converse) can
be summarized as the following equality holding for all ε ∈ (0, 1), states ρ, and positive semi-definite
operators σ:
1
lim Dε (ρ⊗n kσ ⊗n ) = D(ρkσ),
(A.24)
n→∞ n H
thus giving the quantum relative entropy its most fundamental operational meaning as the optimal
Type II error exponent in asymmetric quantum hypothesis testing. Before giving the transparent
proof, let us recall the sandwiched Rényi relative entropy [MLDS+ 13, WWY14], defined for α ∈
(1, ∞) as
e α (ρkσ) ≡ 1 log2 Tr{(σ (1−α)/2α ρσ (1−α)/2α )α }.
D
(A.25)
α−1
whenever supp(ρ) ⊆ supp(σ) and set to +∞ otherwise. For α ∈ (0, 1), it is defined as above. The
e α (ρkσ) =
sandwiched Rényi relative entropy obeys the following limit [MLDS+ 13, WWY14]: limα→1 D
D(ρkσ). [CMW16, Lemma 5] is the statement that the following inequality holds for all α > 1 and
ε ∈ (0, 1):
α
1
ε
e
DH (ρkσ) ≤ Dα (ρkσ) +
log2
.
(A.26)
α−1
1−ε
Employing Proposition 3 and [CMW16, Lemma 5] leads to a direct proof of the quantum Stein’s
lemma [HP91, ON00]. Applying Proposition 3, we find that the following inequality holds for all
α ∈ (0, 1) and ε ∈ (0, 1):
1 ε ⊗n ⊗n
α
1
1
DH (ρ kσ ) ≥
log2
+ Dα (ρ⊗n kσ ⊗n )
(A.27)
n
n(α − 1)
ε
n
α
1
=
log2
+ Dα (ρkσ).
(A.28)
n(α − 1)
ε
Taking the limit as n → ∞ gives the following inequality holding for all α ∈ (0, 1):
lim
n→∞
1 ε ⊗n ⊗n
D (ρ kσ ) ≥ Dα (ρkσ).
n H
(A.29)
We can then take the limit as α → 1 to get that
lim
n→∞
1 ε ⊗n ⊗n
D (ρ kσ ) ≥ D(ρkσ).
n H
33
(A.30)
Applying [CMW16, Lemma 5], we find that the following holds for all α > 1 and ε ∈ (0, 1):
1 ε ⊗n ⊗n
1e
α
1
⊗n
+ D
DH (ρ kσ ) ≤
log2
kσ ⊗n )
(A.31)
α (ρ
n
n(α − 1)
1−ε
n
1
α
e α (ρkσ).
+D
(A.32)
log2
=
n(α − 1)
1−ε
Taking the limit n → ∞, we find that the following holds for all α > 1
1 ε ⊗n ⊗n
e α (ρkσ).
D (ρ kσ ) ≤ D
n→∞ n H
lim
(A.33)
Then taking the limit as α → 1, we get that
1 ε ⊗n ⊗n
D (ρ kσ ) ≤ D(ρkσ).
n→∞ n H
lim
(A.34)
√
We note here that a slightly different approach would be to set α = 1 + 1/ n in both (A.28)
and (A.32) and then take the limit n → ∞.
References
[ACMnT+ 07] K. M. R. Audenaert, J. Calsamiglia, R. Muñoz Tapia, E. Bagan, Ll. Masanes,
A. Acin, and F. Verstraete. Discriminating states: The quantum Chernoff bound.
Physical Review Letters, 98(16):160501, April 2007. arXiv:quant-ph/0610027.
[AJW17a]
Anurag Anshu, Rahul Jain, and Naqueeb Ahmad Warsi. A one-shot achievability
result for quantum state redistribution. IEEE Transactions on Information Theory
(to appear), 2017. arXiv:1702.02396.
[AJW17b]
Anurag Anshu, Rahul Jain, and Naqueeb Ahmad Warsi. One shot entanglement
assisted classical and quantum communication over noisy quantum channels: A hypothesis testing and convex split approach. February 2017. arXiv:1702.01940.
[AM14]
Koenraad M. R. Audenaert and Milan Mosonyi. Upper bounds on the error probabilities and asymptotic error exponents in quantum multiple state discrimination.
Journal of Mathematical Physics, 55(10):102201, October 2014. arXiv:1401.7658.
[Aud14]
Koenraad M. R. Audenaert. Comparisons between quantum state distinguishability
measures. Quantum Information and Computation, 14(1-2):31–38, January 2014.
arXiv:1207.1197.
[BB09]
Igor Bjelakovic and Holger Boche. Classical capacities of compound and averaged
quantum channels. IEEE Transactions on Information Theory, 55(7):3360–3374,
July 2009. arXiv:0710.3027.
[BD06]
Garry Bowen and Nilanjana Datta. Beyond i.i.d. in quantum information theory. In
2006 IEEE International Symposium on Information Theory, pages 451–455. IEEE,
2006. arXiv:quant-ph/0604013.
34
[BD10]
Francesco Buscemi and Nilanjana Datta. The quantum capacity of channels with
arbitrarily correlated noise. IEEE Transactions on Information Theory, 56(3):1447–
1460, March 2010. arXiv:0902.0158.
[BDK+ 05]
Igor Bjelaković, Jean-Dominique Deuschel, Tyll Krüger, Ruedi Seiler, Rainer
Siegmund-Schultze, and Arleta Szkola. A quantum version of Sanov’s theorem.
Communications in mathematical physics, 260(3):659–671, 2005. arXiv:quantph/0412157.
[BH98]
M. V. Burnashev and Alexander S. Holevo. On reliability function of quantum
communication channel. Probl. Peredachi Inform., 34(2):1–13, 1998. arXiv:quantph/9703013.
[BHOS15]
Fernando G. S. L. Brandão, Aram W. Harrow, Jonathan Oppenheim, and Sergii
Strelchuk. Quantum conditional mutual information, reconstructed states, and state
redistribution. Physical Review Letters, 115(5):050501, July 2015. arXiv:1411.4921.
[BLW15]
Mario Berta, Marius Lemm, and Mark M. Wilde. Monotonicity of quantum relative entropy and recoverability. Quantum Information and Computation, 15(15 &
16):1333–1354, November 2015. arXiv:1412.4067.
[BP10]
Fernando G. S. L. Brandao and Martin B. Plenio. A generalization of quantum
Stein’s lemma. Communications in Mathematical Physics, 295(3):791–828, May
2010. arXiv:0904.0281.
[BSS03]
Igor Bjelakovic and Rainer Siegmund-Schultze. Quantum Stein’s lemma revisited,
inequalities for quantum entropies, and a concavity theorem of Lieb. July 2003.
arXiv:quant-ph/0307170.
[BSST02]
Charles H. Bennett, Peter W. Shor, John A. Smolin, and Ashish V. Thapliyal.
Entanglement-assisted capacity of a quantum channel and the reverse Shannon theorem. IEEE Transactions on Information Theory, 48(10):2637–2655, October 2002.
arXiv:quant-ph/0106052.
[CE03]
G. Cariolaro and T. Erseghe. Pulse Position Modulation. John Wiley & Sons, Inc.,
2003.
[CMW16]
Tom Cooney, Milan Mosonyi, and Mark M. Wilde. Strong converse exponents for
a quantum channel discrimination problem and quantum-feedback-assisted communication. Communications in Mathematical Physics, 344(3):797–829, June 2016.
arXiv:1408.3373.
[Dal13]
Marco Dalai. Lower bounds on the probability of error for classical and classicalquantum channels. IEEE Transactions on Information Theory, 59(12):8027–8056,
December 2013. arXiv:1201.5411.
[DD07]
Nilanjana Datta and Tony C. Dorlas. The coding theorem for a class of quantum
channels with long-term memory. Journal of Physics A: Mathematical and Theoretical, 40(28):8147, July 2007. arXiv:quant-ph/0610049.
35
[DF13]
Lukas Drescher and Omar Fawzi. On simultaneous min-entropy smoothing. In 2013
IEEE International Symposium on Information Theory, pages 161–165, July 2013.
arXiv:1312.7642.
[DFW15]
Frederic Dupuis, Omar Fawzi, and Stephanie Wehner. Entanglement sampling and
applications. IEEE Transactions on Information Theory, 61(2):1093–1112, February
2015. arXiv:1305.1316.
[DH13]
Nilanjana Datta and Min-Hsiu Hsieh. One-shot entanglement-assisted quantum and
classical communication. IEEE Transactions on Information Theory, 59(3):1929–
1939, March 2013. arXiv:1105.3321.
[DHL10]
Frédéric Dupuis, Patrick Hayden, and Ke Li. A father protocol for quantum broadcast channels. IEEE Transactions on Information Theory, 56(6):2946–2956, June
2010. arXiv:quant-ph/0612155.
[DPR16]
Nilanjana Datta, Yan Pautrat, and Cambyse Rouzé. Second-order asymptotics for
quantum hypothesis testing in settings beyond iid—quantum lattice systems and
more. Journal of Mathematical Physics, 57(6):062207, June 2016. arXiv:1510.04682.
[DR09]
Nilanjana Datta and Renato Renner. Smooth entropies and the quantum information
spectrum. IEEE Transactions on Information Theory, 55(6):2807–2815, June 2009.
arXiv:0801.0282.
[DRRW13]
Nilanjana Datta, Joseph M. Renes, Renato Renner, and Mark M. Wilde. Oneshot lossy quantum data compression. IEEE Transactions on Information Theory,
59(12):8057–8076, December 2013. arXiv:1304.2336.
[DTW16]
Nilanjana Datta, Marco Tomamichel, and Mark M. Wilde. On the second-order
asymptotics for entanglement-assisted communication. Quantum Information Processing, 15(6):2569–2591, June 2016. arXiv:1405.1797.
[Dup09]
Frederic Dupuis. The capacity of quantum channels with side information at the
transmitter. In 2009 IEEE International Symposium on Information Theory, pages
948–952, June 2009. arXiv:0805.3352.
[Dut11]
Nicolas Dutil. Multiparty quantum protocols for assisted entanglement distillation.
PhD thesis, McGill University, May 2011. arXiv:1105.4657.
[DW14]
Marco Dalai and Andreas Winter. Constant compositions in the sphere packing
bound for classical-quantum channels. In 2014 IEEE International Symposium on
Information Theory, pages 151–155, June 2014. arXiv:1401.6039.
[FHS+ 12]
Omar Fawzi, Patrick Hayden, Ivan Savov, Pranab Sen, and Mark M. Wilde. Classical communication over a quantum interference channel. IEEE Transactions on
Information Theory, 58(6):3670–3691, June 2012. arXiv:1102.2624.
[GLM12]
Vittorio Giovannetti, Seth Lloyd, and Lorenzo Maccone. Achieving the Holevo
bound via sequential measurements. Physical Review A, 85(1):012302, January 2012.
arXiv:1012.0386.
36
[Han03]
Te Sun Han. Information-spectrum methods in information theory, volume 50.
Springer Science & Business Media, 2003.
[Hay02]
Masahito Hayashi. Optimal sequence of quantum measurements in the sense of
Stein’s lemma in quantum hypothesis testing. Journal of Physics A: Mathematical
and General, 35(50):10759–10773, December 2002. arXiv:quant-ph/0208020.
[Hay03]
Masahito Hayashi. General asymptotic formulas for fixed-length quantum entanglement concentration. In Proceedings of the 2003 IEEE International Symposium on
Information Theory, page 431. IEEE, 2003.
[Hay06]
Masahito Hayashi. General formulas for fixed-length quantum entanglement concentration. IEEE Transactions on Information Theory, 52(5):1904–1921, May 2006.
arXiv:quant-ph/0206187.
[Hay07]
Masahito Hayashi. Error exponent in asymmetric quantum hypothesis testing and its
application to classical-quantum channel coding. Physical Review A, 76(6):062301,
December 2007. arXiv:quant-ph/0611013.
[Hay09a]
Masahito Hayashi. Information spectrum approach to second-order coding rate
in channel coding. IEEE Transactions on Information Theory, 55(11):4947–4966,
November 2009.
[Hay09b]
Masahito Hayashi. Universal coding for classical-quantum channel. Communications
in Mathematical Physics, 289(3):1087–1098, August 2009. arXiv:0805.4092.
[HC17]
Masahito Hayashi and Ning Cai. Exponent for classical-quantum multiple access
channel. January 2017. arXiv:1701.02939v1.
[HDW08]
Min-Hsiu Hsieh, Igor Devetak, and Andreas Winter. Entanglement-assisted capacity
of quantum multiple-access channels. IEEE Transactions on Information Theory,
54(7):3078–3090, July 2008. arXiv:quant-ph/0511228.
[Hel69]
Carl W. Helstrom. Quantum detection and estimation theory. Journal of Statistical
Physics, 1:231–252, 1969.
[Hel76]
Carl W. Helstrom. Quantum Detection and Estimation Theory. Academic, New
York, 1976.
[HN03]
Masahito Hayashi and Hiroshi Nagaoka. General formulas for capacity of classicalquantum channels. IEEE Transactions on Information Theory, 49(7):1753–1768,
July 2003. arXiv:quant-ph/0206186.
[Hol73]
Alexander S. Holevo. Statistical problems in quantum physics. In Second JapanUSSR Symposium on Probability Theory, volume 330 of Lecture Notes in Mathematics, pages 104–119. Springer Berlin / Heidelberg, 1973.
[Hol98]
Alexander S. Holevo. The capacity of the quantum channel with general signal
states. IEEE Transactions on Information Theory, 44(1):269–273, January 1998.
arXiv:quant-ph/9611023.
37
[Hol00]
Alexander S. Holevo. Reliability function of general classical-quantum channel. IEEE Transactions on Information Theory, 46(6):2256–2261, September 2000.
arXiv:quant-ph/9907087.
[Hol02]
Alexander S. Holevo. On entanglement assisted classical capacity. Journal of Mathematical Physics, 43(9):4326–4333, September 2002. arXiv:quant-ph/0106075.
[Hol12]
Alexander S. Holevo. Quantum Systems, Channels, Information. de Gruyter Studies
in Mathematical Physics (Book 16). de Gruyter, November 2012.
[HP91]
Fumio Hiai and Dénes Petz. The proper formula for relative entropy and its asymptotics in quantum probability. Communications in mathematical physics, 143(1):99–
114, 1991.
[JOPS12]
V. Jaksic, Y. Ogata, C.-A. Pillet, and R. Seiringer. Quantum hypothesis testing and non-equilibrium statistical mechanics. Reviews in Mathematical Physics,
24(06):1230002, 2012. arXiv:1109.3804.
[Li14]
Ke Li. Second order asymptotics for quantum hypothesis testing. Annals of Statistics,
42(1):171–189, February 2014. arXiv:1208.1400.
[Li16]
Ke Li. Discriminating quantum states: the multiple Chernoff distance. The Annals
of Statistics, 44(4):1661–1679, August 2016. arXiv:1508.06624.
[MD09]
Milan Mosonyi and Nilanjana Datta. Generalized relative entropies and the capacity
of classical-quantum channels. Journal of Mathematical Physics, 50(7):072104, July
2009. arXiv:0810.3478.
[MLDS+ 13]
Martin Müller-Lennert, Frédéric Dupuis, Oleg Szehr, Serge Fehr, and Marco
Tomamichel. On quantum Rényi entropies: a new definition and some properties.
Journal of Mathematical Physics, 54(12):122203, December 2013. arXiv:1306.3142.
[Mos15]
Milán Mosonyi. Coding theorems for compound problems via quantum Rényi
divergences. IEEE Transactions on Information Theory, 61(6):2997–3012, 2015.
arXiv:1310.7525.
[MW14]
William Matthews and Stephanie Wehner. Finite blocklength converse bounds for
quantum channels. IEEE Transactions on Information Theory, 60(11):7317–7329,
November 2014. arXiv:1210.4722.
[NH07]
Hiroshi Nagaoka and Masahito Hayashi. An information-spectrum approach to classical and quantum hypothesis testing for simple hypotheses. IEEE Transactions on
Information Theory, 53(2):534–549, February 2007. arXiv:quant-ph/0206185.
[NS09]
Michael Nussbaum and Arleta Szkola. The Chernoff lower bound for symmetric quantum hypothesis testing. The Annals of Statistics, pages 1040–1057, 2009.
arXiv:quant-ph/0607216.
[ON00]
Tomohiro Ogawa and Hiroshi Nagaoka. Strong converse and Stein’s lemma in quantum hypothesis testing. IEEE Transactions on Information Theory, 46(7):2428–2433,
November 2000. arXiv:quant-ph/9906090.
38
[ON07]
Tomohiro Ogawa and Hiroshi Nagaoka. Making good codes for classical-quantum
channel coding via quantum hypothesis testing. IEEE Transactions on Information
Theory, 53(6):2261–2266, June 2007.
[Pet86]
Dénes Petz. Quasi-entropies for finite quantum systems. Reports in Mathematical
Physics, 23:57–65, 1986.
[PPV10]
Yury Polyanskiy, H. Vincent Poor, and Sergio Verdú. Channel coding rate in the
finite blocklength regime. IEEE Transactions on Information Theory, 56(5):2307–
2359, May 2010.
[RSW16]
Jaikumar Radhakrishnan, Pranab Sen, and Naqueeb Warsi. One-shot Marton inner
bound for classical-quantum broadcast channel. IEEE Transactions on Information
Theory, 62(5):2836–2848, May 2016. arXiv:1410.3248.
[SW97]
Benjamin Schumacher and Michael D. Westmoreland. Sending classical information
via noisy quantum channels. Physical Review A, 56(1):131–138, July 1997.
[SW15]
Ivan Savov and Mark M. Wilde. Classical codes for quantum broadcast channels. IEEE Transactions on Information Theory, 61(12):7017–7028, December 2015.
arXiv:1111.3645.
[TCR09]
Marco Tomamichel, Roger Colbeck, and Renato Renner. A fully quantum asymptotic
equipartition property. IEEE Transactions on Information Theory, 55(12):5840—
5847, December 2009. arXiv:0811.1221.
[TH13]
Marco Tomamichel and Masahito Hayashi. A hierarchy of information quantities for
finite block length analysis of quantum tasks. IEEE Transactions on Information
Theory, 59(11):7693–7710, November 2013. arXiv:1208.1478.
[TT15]
Marco Tomamichel and Vincent Y. F. Tan. Second-order asymptotics for the classical
capacity of image-additive quantum channels. Communications in Mathematical
Physics, 338(1):103–137, August 2015. arXiv:1308.6503.
[Uhl76]
Armin Uhlmann. The transition probability in the state space of a *-algebra. Reports
on Mathematical Physics, 9(2):273–279, 1976.
[Ume62]
Hisaharu Umegaki. Conditional expectations in an operator algebra IV (entropy and
information). Kodai Mathematical Seminar Reports, 14(2):59–85, 1962.
[Ver90]
Sergio Verdu. On channel capacity per unit cost. IEEE Transactions on Information
Theory, 36(5):1019–1030, 1990.
[Wat16]
John Watrous. Theory of Quantum Information. 2016.
[WDW17]
Qingle Wang, Siddhartha Das, and Mark M Wilde. Hadamard quantum broadcast channels.
Quantum Information Processing, 16(10):248, October 2017.
arXiv:1611.07651.
[Wil10]
Mark M. Wilde. Information spectrum approach to the quantum multiple access
channel. unpublished note, November 2010.
39
[Wil11]
Mark M. Wilde. The quest for a quantum simultaneous decoder. In Difficult Problems
in Quantum Information Theory, MIT, 50 Vassar Street, Building 36, Room 36-428,
May 2011. http://xqit.mit.edu/conference2011/MarkWilde.htm.
[Wil16]
Mark M. Wilde. From classical to quantum Shannon theory.
arXiv:1106.1445v7.
[Wil17a]
Mark M. Wilde. Position-based coding and convex splitting for private communication over quantum channels. Quantum Information Processing, 16(10):264, October
2017. arXiv:1703.01733.
[Wil17b]
Mark M. Wilde. Quantum Information Theory. Cambridge University Press, second
edition, February 2017.
[Win99a]
Andreas Winter. Coding theorem and strong converse for quantum channels.
IEEE Transactions on Information Theory, 45(7):2481–2485, November 1999.
arXiv:1409.2536.
[Win99b]
Andreas Winter. Coding Theorems of Quantum Information Theory. PhD thesis,
Universität Bielefeld, July 1999. arXiv:quant-ph/9907077.
[Win01]
Andreas Winter. The capacity of the quantum multiple-access channel. IEEE
Transactions on Information Theory, 47(7):3059–3065, November 2001. arXiv:quantph/9807019.
[WR12]
Ligong Wang and Renato Renner. One-shot classical-quantum capacity and hypothesis testing. Physical Review Letters, 108(20):200501, May 2012. arXiv:1007.5456.
[WRG16]
Mark M. Wilde, Joseph M. Renes, and Saikat Guha. Second-order coding rates
for pure-loss bosonic channels. Quantum Information Processing, 15(3):1289–1308,
March 2016. arXiv:1408.5328.
[WWY14]
Mark M. Wilde, Andreas Winter, and Dong Yang. Strong converse for the classical
capacity of entanglement-breaking and Hadamard channels via a sandwiched Rényi
relative entropy. Communications in Mathematical Physics, 331(2):593–622, October
2014. arXiv:1306.1586.
[XW13]
Shen Chen Xu and Mark M. Wilde. Sequential, successive, and simultaneous decoders for entanglement-assisted classical communication. Quantum Information
Processing, 12(1):641–683, January 2013. arXiv:1107.1347.
[YHD11]
Jon Yard, Patrick Hayden, and Igor Devetak. Quantum broadcast channels. IEEE
Transactions on Information Theory, 57(10):7147–7162, October 2011. arXiv:quantph/0603098.
40
March 2016.
| 7cs.IT
|
arXiv:1708.07786v1 [cs.DS] 25 Aug 2017
Efficient Adaptive Implementation of the Serial
Schedule Generation Scheme using
Preprocessing and Bloom Filters
Daniel Karapetyan1 and Alexei Vernitski2
1
Institute for Analytics and Data Science, University of Essex, Essex, UK
[email protected]
2
Department of Mathematical Sciences, University of Essex, Essex, UK
[email protected]
Abstract. The majority of scheduling metaheuristics use indirect representation of solutions as a way to efficiently explore the search space.
Thus, a crucial part of such metaheuristics is a “schedule generation
scheme” – procedure translating the indirect solution representation into
a schedule. Schedule generation scheme is used every time a new candidate solution needs to be evaluated. Being relatively slow, it eats up most
of the running time of the metaheuristic and, thus, its speed plays significant role in performance of the metaheuristic. Despite its importance,
little attention has been paid in the literature to efficient implementation of schedule generation schemes. We give detailed description of serial
schedule generation scheme, including new improvements, and propose
a new approach for speeding it up, by using Bloom filters. The results
are further strengthened by automated control of parameters. Finally, we
employ online algorithm selection to dynamically choose which of the two
implementations to use. This hybrid approach significantly outperforms
conventional implementation on a wide range of instances.
Keywords: resource-constrained project scheduling problem, serial schedule generation scheme, Bloom filters, online algorithm selection
1
Introduction
Resource Constrained Project Scheduling Problem (RCPSP) is to schedule a set
of jobs J subject to precedence relationships and resource constraints. RCPSP
is a powerful model generalising several classic scheduling problems such as job
shop scheduling, flow shop scheduling and parallel machine scheduling.
In RCPSP, we are given a set of resources R and their capacities cr , r ∈ R.
In each time slot, cr units of resource r are available and can be shared between
jobs. Each job j ∈ J has a prescribed consumption vj,r of each resource r ∈ R.
We are also given the duration dj of a job j ∈ J. A job consumes vj,r units
of resource r in every time slot that it occupies. Once started, a job cannot be
interrupted (no preemption is allowed). Finally, each resource is assigned a set
pred j ⊂ J of jobs that need to be completed before j can start.
There exist multiple extensions of RCPSP. In the multi-mode extension, each
job can be executed in one of several modes, and then resource consumption and
duration depend on the selected mode. In some applications, resource availability
may vary with time. There could be set-up times associated with certain jobs. In
multi-project extension, several projects run in parallel sharing some but not all
resources. In this paper we focus on the basic version of RCPSP, however some
of our results can be easily generalised to many of its extensions and variations.
Most of the real-world scheduling problems, including RCPSP, are NP-hard,
and hence only problems of small size can be solved to optimality, whereas
for larger problems (meta)heuristics are commonly used. Metaheuristics usually
search in the space of feasible solutions; with a highly constrained problem such
as RCPSP, browsing the space of feasible solutions is hard. Indeed, if a schedule
is represented as a vector of job start times, then changing the start time of a
single job is likely to cause constraint violations. Usual approach is to use indirect
solution representation that could be conveniently handled by the metaheuristic
but could also be efficiently translated into the direct representation.
Two translation procedures widely used in scheduling are serial schedule
generation scheme (SSGS) and parallel schedule generation scheme [9]. Some
studies conclude that SSGS gives better performance [7], while others suggest
to employ both procedures within a metaheuristic [10]. Our research focuses on
SSGS.
With SSGS, the indirect solution representation is a permutation π of jobs.
The metaheuristic handles candidate solutions in indirect (permutation) form.
Every time a candidate solution needs to be evaluated, SSGS is executed to
translate the solution into a schedule (tj , j ∈ J), and only then the objective
value can be computed, see Figure 1.
Metaheuristic
Candidate solution (permutation π)
Objective value
SSGS
Candidate schedule (job start times tj , j ∈ J)
Objective function
Fig. 1: Classic architecture of a scheduling metaheuristic. To obtain objective
value of a candidate solution, metaheuristic uses SSGS to translate the candidate
solution into a candidate schedule, which is then used by objective function.
SSGS is a simple procedure that iterates through J in the order given by
π, and schedules one job at a time, choosing the earliest feasible slot for each
job. The only requirement for π is to respect the precedence relations; otherwise
SSGS produces a feasible schedule for any permutation of jobs. The pseudo-
Algorithm 1: Serial Schedule Generation Scheme (SSGS). Here T is the
upper bound on the makespan.
6
input : Solution π in permutation form, respecting precedence relations
output : Schedule tj , j ∈ J
At,r ← cr for every t = 1, 2, . . . , T and r ∈ R;
for i = 1, 2, . . . , |J| do
j ← π(i);
t0 ← maxj ′ ∈pred j tj ′ + dj ′ ;
tj ← find (j, t0 , A) (see Algorithm 2);
update (j, tj , A) (see Algorithm 3);
7
return tj , j ∈ J;
1
2
3
4
5
Algorithm 2: Conventional implementation of find(j, t0 , A) – a function
to find the earliest feasible slot for job j.
8
input : Job j ∈ J to be scheduled
input : Earliest start time t0 as per precedence relations
input : Current availability A of resources
output : Earliest feasible start time for job j
tj ← t0 ;
t ← tj ;
while t < tj + dj do
if At,r ≥ vj,r for every r ∈ R then
t ← t + 1;
else
tj ← t + 1;
t ← tj ;
9
return tj ;
1
2
3
4
5
6
7
code of SSGS is given in Algorithm 1, and its two subroutines find and update
in Algorithms 2 and 3.
Commonly, the objective of RCPSP is to find a schedule that minimises the
makespan, i.e. the time required to complete all jobs; however other objective
functions are also considered in the literature. We say that an objective function
of a scheduling problem is regular if advancing the start time of a job cannot
worsen the solution’s objective value. Typical objective functions of RCPSP, including makespan, are regular. If the scheduling problem has a regular objective
function, then SSGS guarantees to produce active solutions, i.e. solutions that
cannot be improved by changing tj for a single j ∈ J. Moreover, it was shown [8]
that for any active schedule S there exists a permutation π for which SSGS will
generate S. Since any optimal solution S is active, searching in the space of
feasible permutations π is sufficient to solve the problem. This is an important
property of SSGS; the parallel schedule generation scheme, mentioned above,
Algorithm 3: Procedure update(j, tj , A) to update resource availability A
after scheduling job j at time tj .
1
2
input : Job j and its start time tj
input : Resource availability A
for t ← tj , tj + 1, . . . , tj + dj − 1 do
Aj,r ← Aj,r − vj,r for every r ∈ R;
does not provide this guarantee [8] and, hence, may not in some circumstances
allow a metaheuristic finding optimal or near-optimal solutions.
The runtime of a metaheuristic is divided between its search control mechanism that modifies solutions and makes decisions such as accepting or rejecting
new solutions, and SSGS. While SSGS is a polynomial algorithm, in practice it
eats up the majority of the metaheuristic runtime (over 98% as reported in [1]).
In other words, by improving the speed of SSGS twofold, one will (almost) double
the number of iterations a metaheuristic performs within the same time budget,
and this increase in the number of iterations is likely to have a major effect on
the quality of obtained solutions.
In our opinion, not enough attention was paid to SSGS – a crucial component
of many scheduling algorithms, and this study is to close this gap. In this paper
we discuss approaches to speed up the conventional implementation of SSGS.
Main contributions of our paper are:
– A detailed description of SSGS including old and new speed-ups (Section 2).
– New implementation of SSGS employing Bloom filters for quick testing of
resource availability (Section 3).
– A hybrid control mechanism that employs intelligent learning to dynamically
select the best performing SSGS implementation (Section 4).
Empirical evaluation in Section 5 confirms that both of our implementations
of SSGS perform significantly better than the conventional SSGS, and the hybrid
control mechanism is capable of correctly choosing the best implementation while
generating only negligible overheads.
2
SSGS implementation details
Before we proceed to introducing our main new contributions in Sections 3 and 4,
we describe what state-of-the-art implementation of SSGS we use, including some
previously unpublished improvements.
2.1
Initialisation of A
The initialisation of A in line 1 of Algorithm 1 iterates through T slots, where T is
the upper bound on the makespan. It was noted in [1] that instead of initialising
A at every execution of SSGS, one can reuse this data structure between the
Algorithm 4: Enhanced implementation of find(j, t0 , A)
10
input : Job j ∈ J to be scheduled
input : Earliest start time t0 as per precedence relations
input : Current availability A of resources
output : Earliest feasible start time for job j
tj ← t0 ;
t ← tj + dj − 1;
ttest ← tj ;
while t ≥ ttest do
if At,r ≥ Vj,r for every r ∈ R then
t ← t − 1;
else
ttest ← tj + dj ;
tj ← t + 1;
t ← tj + dj − 1;
11
return tj ;
1
2
3
4
5
6
7
8
9
executions. To correctly initialise A, at the end of SSGS we restore At,r for each
r ∈ R and each slot where some job was scheduled: At,r ← cr for r ∈ R and
t = 1, 2, . . . , M , where M is the makespan of the solution. Since M ≤ T and
usually M ≪ T , this notably improves the performance of SSGS [1].
2.2
Efficient search of the earliest feasible slot for a job
The function find(j, t0 , I, A) finds the earliest slot feasible for scheduling job
j. Its conventional implementation (Algorithm 2) takes O(T |R|) time, where T
is the upper bound of the time horizon. Our enhanced implementation of find
(Algorithm 4), first proposed in [1], has the same worst case complexity but
is more efficient in practice. It is inspired by the Knuth-Morris-Pratt substring
search algorithm. Let tj be the assumed starting time of job J. To verify if it is
feasible, we need to test sufficiency of resources in slots tj , tj + 1, . . . , tj + dj − 1.
Unlike the conventional implementation, our enhanced version tests these slots
in the reversed order. The order makes no difference if the slot is feasible, but
otherwise testing in reversed order allows us to skip some slots; in particular, if
slot t is found to have insufficient resources then we know that feasible tj is at
least t + 1.
A further speed up, which was not discussed in the literature before, is to
avoid re-testing of slots with sufficient resources. Consider the point when we
find that the resources in slot t are insufficient. By that time we know that the
resources in t + 1, t + 2, . . . , tj + dj − 1 are sufficient. Our heuristic is to remember
that the earliest slot ttest to be tested in future iterations is tj + dj .
2.3
Preprocessing and Automated Parameter Control
We observe that in many applications, jobs are likely to require only a subset
of resources. For example, in construction works, to dig a hole one does not
need cranes or electricians, hence the ‘dig a hole’ job will not consume those
resources. To exploit this observation, we pre-compute vector Rj of resources
used by job j, and then iterate only through resources in Rj when testing resource
sufficiency in find (Algorithm 4, line 5) and updating resource availability in
update (Algorithm 3, line 2). Despite the simplicity of this idea, we are not
aware of anyone using or mentioning it before.
We further observe that some jobs may consume only one resource. By creating specialised implementations of find and update, we can reduce the depth of
nested loops. While this makes no difference from the theoretical point of view,
in practice this leads to considerable improvement of performance. Correct implementations of find and update are identified during preprocessing and do not
cause overheads during executions of SSGS.
Having individual vectors Rj of consumed resources for each job, we can also
intelligently learn the order in which resource availability is tested (Algorithm 4,
line 5). By doing this, we are aiming at minimising the expected number of iterations within the resource availability test. For example, if resource r is scarce
and job j requires significant amount of r, then we are likely to place r at the
beginning of Rj . More formally, we sort Rj in descending order of probability
that the resource is found to be insufficient during the search. This probability is obtained empirically by a special implementation of SSGS which we call
SSGSdata . SSGSdata is used once to count how many times each resource turned
out to be insufficient during scheduling of a job. (To avoid bias, SSGSdata tests
every resource in Rj even if the test could be terminated early.) After a single
execution of SSGSdata , vectors Rj are optimised, and in further executions default
implementation of SSGS is used.
One may notice that the data collected in the first execution of SSGS may
get outdated after some time; this problem in addressed in Section 4.
Ordering of Rj is likely to be particularly effective on instances with asymmetric use of resources, i.e. on real instances. Nevertheless, we observed improvement of runtime even on pseudo-random instances as reported in Section 5.
3
SSGS Implementation using Bloom filters
Performance bottleneck of an algorithm is usually its innermost loop. Observe
that the innermost loop of the find function is the test of resource sufficiency in
a slot, see Algorithm 4, line 5. In this section we try to reduce average runtime
of this test from O(|R|) to O(1) time. For this, we propose a novel way of using
a data structure known as Bloom filter.
Bloom filters were introduced in [2] as a way of optimising dictionary lookups,
and found many applications in computer science and electronic system engineering [3,13]. Bloom filters usually utilise pseudo-random hash functions to encode
u1,2
u1,3
u1,4
u2,1
u2,3
u3,1
u3,3
u3,4
≥2
≥3
≥4
≥1
≥3
≥1
≥3
≥4
Resource 1
Resource 2
Resource 3
Fig. 2: Example of a Bloom filter structure for a problem with 3 resources, each
having capacity 4.
data, but in some applications [6] non-hash-based approaches are used. In our
paper, we also use a non-hash-based approach, and to our knowledge, our paper
is the first in which the structure of Bloom filters is chosen dynamically according to the statistical properties of the data, with the purpose of improving the
speed of an optimisation algorithm.
In general, Bloom filters can be defined as a way of using data, and they
are characterised by two aspects: first, all data is represented by short binary
arrays of a fixed length (perhaps with a loss of accuracy); second, the process of
querying data involves only bitwise comparison of binary arrays (which makes
querying data very fast).
We represent both each job’s resource consumption and resource availability
at each time slot, by binary arrays of a fixed length; we call these binary arrays
Bloom filters. Our Bloom filters will consist of bits which we call resource level
bits. Each resource bit, denoted by ur,k , r ∈ R, k ∈ {1, 2, . . . , cr }, means “k units
of resource r” (see details below). Let U be the set of all possible resource bits. A
Bloom filter structure is an ordered subset L ⊆ U , see Figure 2 for an example.
Suppose that a certain Bloom filter structure L is fixed. Then we can introduce
B L (j), the Bloom filter of job j, and B L (t), the Bloom filter of time slot t, for
each j and t. Each B L (j) and B L (t) consists of |L| bits defined as follows: if ur,k
is the ith element of L then
(
(
1 if vj,r ≥ k,
1 if At,r ≥ k,
L
L
B (j)i =
and
B (t)i =
0 otherwise,
0 otherwise.
To query if a job j can be scheduled in a time slot t, we compare Bloom
filters B L (j) and B L (t) bitwise; then one of three situations is possible, as the
following examples show. Consider a job j and three slots, t, t′ and t′′ , with the
following resource consumption/availabilities, and Bloom filter structure as in
Figure 2:
vj,1 = 3
vj,2 = 2
vj,3 = 0
B L (j) = (110 10 000)
At,1 = 2
At,2 = 3
At,3 = 4
B L (t) = (100 11 111)
At′ ,1 = 3
At′ ,2 = 1
At′ ,3 = 4
B L (t′ ) = (110 10 111)
At′′ ,1 = 3
At′′ ,2 = 2
At′′ ,3 = 2
B L (t′′ ) = (110 10 100)
For two bit arrays of the same length, let notation ‘≤’ mean bitwise less or
equal. By observing that B L (j) 6≤ B L (t), we conclude that resources in slot t
u1,1
u1,2
u1,3
u1,4
u2,1
u2,2
u2,3
u2,4
u3,1
u3,2
u3,3
u3,4
≥1
≥2
≥3
≥4
≥1
≥2
≥3
≥4
≥1
≥2
≥3
≥4
Resource 1
Resource 2
Resource 3
Fig. 3: Example of a Bloom filter structure for a problem with 3 resources, each
having capacity 4, with L = U .
are insufficient for j; this conclusion is guaranteed to be correct. By observing
that B L (j) ≤ B L (t′ ), we conclude tentatively that resources in slot t′ may be
sufficient for j; however, further verification of the complete data related to j
and t′ (that is, vj,· and At′ ,· ) is required to get a precise answer; one can see
that vj,2 ≥ At′ ,2 , hence this is what is called a false positive. Finally, we observe
that B L (j) ≤ B L (t′′ ), and a further test (comparing vj,· and At′′ ,· ) confirms that
resources in t′′ are indeed sufficient for j.
Values of B L (j), j ∈ J, are pre-computed when L is constructed, and B L (t),
t = 1, 2, . . . , T , are maintained by the algorithm.
3.1
Optimisation of Bloom Filter Structure
The length |L| of a Bloom filter is limited to reduce space requirements and,
more importantly for our application, speed up Bloom filter tests. Note that if
|L| is small (such as 32 or 64 bits) then we can exploit efficient bitwise operators
implemented by all modern CPUs; then each Bloom filter test takes only one
CPU operation. We set |L| = 32 in our implementation. While obeying this
constraint, we aim at minimising the number of false positives, because false
positives slow down the implementation.
Our L building algorithm is as follows:
1. Start with L = U , such as in Figure 3.
2. If |L| is within the prescribed limit, stop.
3. Otherwise select ur,k ∈ L that is least important and delete it. Go to step 2.
By ‘least important’ we mean that the deletion of it is expected to have minimal
impact of the expected number of false positives. Let L = (. . . , ur,q , ur,k , ur,m , . . .).
Consider a job j such that k ≤ vj,r < m and a slot t such that q ≤ At,r < k.
With L as defined above, Bloom filters correctly identify that resources in slot
t are insufficient for job j: B L (j) 6≤ B L (t). However, without the resource level
bit ur,k we get a false positive: B L (j) ≤ B L (t). Thus, the probability of false
positives caused by deleting ur,k from L in is as follows:
! i−1
m−1
X
X
Dkr ·
Ekr ,
k=i
k=q
where Dkr is the probability that a randomly chosen job needs exactly k units
of resource r, and Ekr is the probability that a certain slot, when we examine
it for scheduling a job, has exactly k units of resource r available. The probability distribution Dr is produced from the RCPSP instance data during preprocessing.3 The probability distribution E r is obtained empirically during the
run of SSGSdata (see Section 2.2); each time resource sufficiency is tested within
SSGSdata , its availability is recorded.
3.2
Additional Speed-ups
While positive result of a Bloom filter test generally requires further verification
using full data, in some circumstances its correctness can be guaranteed. In
particular, if for some r ∈ R and j ∈ J we have ur,k ∈ L and vj,r = k, then the
Bloom filter result, whether positive or negative, does not require verification.
Another observation is that updating B L (t) in update can be done in O(|Rj |)
operations instead of O(|R|) operations. Indeed, instead of computing B L (t)
from scratch, we can exploit our structure of Bloom filters. We update each bit
related to resources r ∈ Rj , but we keep intact other bits. With some trivial
pre-processing, this requires only O(|Rj |) CPU operations.
We also note that if |Rj | = 1, i.e. job j uses only one resource, then Bloom
filters will not speed up the find function for that job and, hence, in such cases
we use the standard find function specialised for one resource (see Section 2.2).
4
Hybrid Control Mechanism
So far we have proposed two improved implementations of SSGS: one using
Bloom filters (which we denote SSGSBF ), and the other one not using Bloom
filters (which we denote SSGSNBF ). While it may look like SSGSBF should always
be superior to SSGSNBF , in practice SSGSNBF is often faster. Indeed, Bloom filters
usually speed up the find function, but they also slow down update, as in SSGSBF
we need to update not only the values At,r but also the Bloom filters B L (t)
encoding resource availability. If, for example, the RCPSP instance has tight
precedence relations and loose resource constraints then find may take only a
few iterations, and then the gain in the speed of find may be less than the loss
of speed of update. In such cases SSGSBF is likely to be slower then SSGSNBF .
In short, either of the two SSGS implementations can be superior in certain
circumstances, and which one is faster mostly depends on the problem instance.
In this section we discuss how to adaptively select the best SSGS implementation.
Automated algorithm selection is commonly used in areas such as sorting, where
multiple algorithms exist. A typical approach is then to extract easy to compute
input data features and then apply off-line learning to develop a predictor of
which algorithm is likely to perform best, see e.g. [5]. With sorting, this seems
to be the most appropriate approach as the input data may vary significantly
between executions of the algorithm. Our case is different in that the most
3
In multi-mode extension of RCPSP, this distribution depends on selected modes and
hence needs to be obtained empirically, similarly to how we obtain E r .
Restart
First 10,000 executions
data BF NBF BF NBF BF
Choosing SSGS impl.
BF
Next 10,000 executions
BF
BF data BF NBF BF NBF
Running the faster impl.
Special execution to collect data to optimise Rj and Bloom filters
Fig. 4: Stages of the Hybrid control mechanism. Each square shows one execution of SSGS, and the text inside describes which implementation of SSGS
is used. SSGSdata is always used in the first execution. Further few executions
(at most 100) alternate between SSGSBF and SSGSNBF , with each execution being timed. Once sign test shows significant difference between the SSGSBF and
SSGSNBF implementations, the faster one is used for the rest of executions. After
10,000 executions, previously collected data is erased and adaptation starts from
scratch.
crucial input data (the RCPSP instance) does not change between executions
of SSGS. Thus, during the first few executions of SSGS, we can test how each
implementation performs, and then select the faster one. This is a simple yet
effective control mechanism which we call Hybrid.
Hybrid is entirely transparent for the metaheuristic; the metaheuristic simply
calls SSGS whenever it needs to evaluate a candidate solution and/or generate
a schedule. The Hybrid control mechanism is then intelligently deciding each
time which implementation of SSGS to use based on information learnt during
previous runs. An example of how Hybrid performs is illustrated in Figure 4. In
the first execution, it uses SSGSdata to collect data required for both SSGSBF and
SSGSNBF . For the next few executions, it alternates between SSGSBF and SSGSNBF ,
measuring the time each of them takes. During this stage, Hybrid counts how
many times SSGSBF was faster than the next execution of SSGSNBF . Then we use
the sign test [4] to compare the implementations. If the difference is significant
(we use a 5% significance level for the sign test) then we stop alternating the
implementations and in future use only the faster one. Otherwise we continue
alternating the implementations, but for at most 100 executions. (Without such
a limitation, there is a danger that the alternation will never stop – if the implementations perform similarly; since there are overheads associated with the
alternation and time measurement, it is better to pick one of the implementations
and use it in future executions.)
Our decision to use the sign test is based on two considerations: first, it is very
fast, and second, it works for distributions which are not normal. This makes
our approach different from [12] where the distributions of runtimes are assumed
to be normal. (Note that in our experiments we observed that the distribution
of running times of an SSGS implementation resembles Poisson distribution.)
As pointed out in this and previous sections, optimal choices of parameters
of the SSGS implementations mostly depend on the RCPSP instance – which
does not change throughout the metaheuristic run; however solution π also affects the performance. It should be noted though that metaheuristics usually
apply only small changes to the solution at each iteration, and hence solution
properties tend to change relatively slowly over time. Consequently, we assume
that parameters chosen in one execution of SSGS are likely to remain efficient
for some further executions. Thus, Hybrid ‘restarts’ every 10,000 executions, by
which we mean that all the internal data collected by SSGS is erased, and learning starts from scratch, see Figure 4. This periodicity of restarts is a compromise
between accuracy of choices and overheads, and it was shown to be practical in
our experiments.
5
Empirical Evaluation
In this Section we evaluate the implementations of SSGS discussed above. To
replicate conditions within a metaheuristic, we designed a simplified version of
Simulated Annealing. In each iteration of our metaheuristic, current solution
is modified by moving a randomly selected job into a new randomly selected
position (within the feasible range). If the new solution is not worse than the
previous one, it is accepted. Otherwise the new solution is accepted with 50%
probability. Our metaheuristic performs 1,000,000 iterations before terminating.
We evaluate SSGSBF , SSGSNBF and Hybrid. These implementations are compared to ‘conventional’ SSGS, denoted by SSGSconv , which does not employ any
preprocessing or Bloom filters and uses conventional implementation of find (Algorithm 2).
We found that instances in the standard RCPSP benchmark set PSPLIB [11]
occupy a relatively small area of the feature space. For example, all the RCPSP
instances in PSPLIB have exactly four resources, and the maximum job duration
is always set to 10. Thus, we chose to use the PSPLIB instance generator, but
with a wider range of settings. Note that for this study, we modified the PSPLIB
instance generator by allowing jobs not to have any precedence relations. This
was necessary to extend the range of network complexity parameter (to include
instances with scarce precedence relations), and to speed up the generator, as the
original implementation would not allow us to generate large instances within
reasonable time.
All the experiments are conducted on a Windows Server 2012 machine based
on Intel Xeon E5-2690 v4 2.60 GHz CPU. No concurrency is employed in any of
the implementations or tests.
To see the effect of various instance features on SSGS performance, we select
one feature at a time and plot average SSGS performance against the values
of that feature. The rest of the features (or generator parameters) are then set
as follows: number of jobs 120, number of resources 4, maximum duration of
job 10, network complexity 1, resource factor 0.75, and resource strength 0.1.
These values correspond to some typical settings used in PSPLIB. For formal
definitions of the parameters we refer to [11].
For each combination of the instance generator settings, we produce 50 instances using different random generator seed values, and in each of our experiments the metaheuristic solves each instance once. Then the runtime of SSGS is
said to be the overall time spent on solving those 50 instances, over 50,000,000
(which is the number of SSGS executions). The metaheuristic overheads are
relatively small and are ignored.
From the results reported in Figure 5 one can see that our implementations
of SSGS are generally significantly faster than SSGSconv , but performance of each
implementation varies with the instance features. In some regions of the instance
space SSGSBF outperforms SSGSNBF , whereas in other regions SSGSNBF outperforms
SSGSBF . The difference in running times is significant, up to a factor of two in our
experiments. At the same time, Hybrid is always close to the best of SSGSBF and
SSGSNBF , which shows efficiency of our algorithm selection approach. In fact, when
SSGSBF and SSGSNBF perform similarly, Hybrid sometimes outperforms both; this
behaviour is discussed below.
Another observation is that SSGSNBF is always faster than SSGSconv (always
below the 100% mark) which is not surprising; indeed, SSGSNBF improves the performance of both find and update. In contrast, SSGSBF is sometimes slower than
SSGSconv ; on some instances, the speed-up of the find function is overweighed
by overheads in both find and update. Most important though is that Hybrid
outperforms SSGSconv in each of our experiments by 8 to 68%, averaging at 43%.
In other words, within a fixed time budget, an RCPSP metaheuristic employing Hybrid will be able to run around 1.8 times more iterations than if it used
SSGSconv .
To verify that Hybrid exhibits the adaptive behaviour and does not just stick
to whichever implementation has been chosen initially, we recorded the implementation it used in every execution for several problems, see Figure 6. For this
experiment, we produced three instances: first instance has standard parameters
except Resource Strength is 0.2; second instance has standard parameters except
Resource Factor is 0.45; third instance has standard parameters except Maximum Job Duration is 20. These parameter values were selected such that the
two SSGS implementations would be competitive and, therefore, switching between them would be a reasonable strategy. One can see that the switches occur
several times throughout the run of the metaheuristic, indicating that Hybrid
adapts to the changes of solution. For comparison, we disabled the adaptiveness
and measured the performance if only implementation chosen initially is used
throughout all iterations; the results are shown on Figure 6. We conclude that
Hybrid benefits from its adaptiveness.
6
Conclusions and Future Work
In this paper we discussed the crucial component of many scheduling metaheuristics, the serial schedule generation scheme (SSGS). SSGS eats up most
SSGSBF
SSGSNBF
Hybrid
65
Runtime, %
70
60
60
55
50
50
0
200
400
600
2
800 1,000
6
8
10
Number of resources
Number of jobs
150
Runtime, %
4
70
60
100
50
50
40
0
0.2
0.4
0.6
0.8
0.2
1
0.6
0.8
1
Resource factor
Resource strength
70
Runtime, %
0.4
100
60
80
60
50
0
0.5
1
1.5
Network complexity
2
2.5
0
10
20
30
Maximum job duration
Fig. 5: These plots show how performance of the SSGS implementations depends
on various instance features. Vertical axis gives the runtime of each implementation relative to SSGSconv . (SSGSconv graph would be a horizontal line y = 100%.)
SSGSBF
SSGSNBF
99%
96%
95%
executions
0
1 000 000
Fig. 6: This diagram shows how Hybrid switches between SSGS implementations
while solving three problem instances. The number on the right shows the time
spent by Hybrid compared with the time that would be needed if only the
implementation chosen at the start would be used for all iterations.
of the runtime in a typical scheduling metaheuristic, therefore performance of
SSGS is critical to the performance of the entire metaheuristic, and thus each
speed-up of SSGS has significant impact. We described existing and some new
speed-ups to SSGS, including preprocessing and automated parameter control.
This implementation clearly outperformed the ‘conventional’ SSGS in our experiments. We further proposed a new implementation that uses Bloom filters,
particularly efficient in certain regions of the instance space. To exploit strengths
of both implementations, we proposed a hybrid control mechanism that learns
the performance of each implementation and then adaptively chooses the SSGS
version that is best for a particular problem instance and phase of the search.
Experiments showed that this online algorithm selection mechanism is effective
and makes Hybrid our clear choice. Note that Hybrid is entirely transparent for
the metaheuristic which uses it as if it would be simple SSGS; all the learning
and adaptation is hidden inside the implementation.
The idea behind online algorithm selection used in this project can be further
developed by making the number of executions between restarts adaptable. To
determine the point when the established relation between the SSGS implementations may have got outdated, we could treat the dynamics of the implementations’ performance change as two random walks, and use the properties of these
two random walks to predict when they may intersect.
All the implementations discussed in the paper, in C++, are available for
downloading from http://csee.essex.ac.uk/staff/dkarap/rcpsp-ssgs.zip.
The implementations are transparent and straightforward to use.
While we have only discussed SSGS for the simple RCPSP, our ideas can
easily be applied in RCPSP extensions. We expect some of these ideas to work
particularly well in multi-project RCPSP, where the overall number of resources
is typically large but only a few of them are used by each job.
Acknowledgements We would like to thank Prof. Rainer Kolisch for providing
us with a C++ implementation of the PSPLIB generator. It should be noted
that, although the C++ implementation was developed to reproduce the original
Pascal implementation, the exact equivalence cannot be guaranteed; also, in our
experiments we used a modification of the provided C++ code.
References
1. Shahriar Asta, Daniel Karapetyan, Ahmed Kheiri, Ender Özcan, and Andrew J.
Parkes. Combining Monte-Carlo and hyper-heuristic methods for the multimode resource-constrained multi-project scheduling problem. Information Sciences, 373:476–498, 2016.
2. Burton H. Bloom. Space/time trade-offs in hash coding with allowable errors.
Communications of the ACM, 13(7):422–426, 1970.
3. Andrei Broder and Michael Mitzenmacher. Network applications of bloom filters:
A survey. Internet mathematics, 1(4):485–509, 2004.
4. Louis Cohen and Michael Holliday. Practical statistics for students: An introductory
text. Paul Chapman Publishing Ltd, 1996.
5. Haipeng Guo. Algorithm selection for sorting and probabilistic inference: A machine learning-based approach. PhD thesis, Kansas State University, 2003.
6. Gökçe Caylak Kayaturan and Alexei Vernitski. A way of eliminating errors when
using Bloom filters for routing in computer networks. In ICN 2016, The Fifteenth
International Conference on Networks, pages 52–57, 2016.
7. Jin-Lee Kim and Ralph D. Ellis Jr. Comparing schedule generation schemes in
resource-constrained project scheduling using elitist genetic algorithm. Journal of
construction engineering and management, 136(2):160–169, 2010.
8. Rainer Kolisch. Serial and parallel resource-constrained project scheduling methods revisited: Theory and computation. European Journal of Operational Research,
90(2):320–333, 1996.
9. Rainer Kolisch and Sönke Hartmann. Heuristic algorithms for the resourceconstrained project scheduling problem: Classification and computational analysis.
In Project scheduling, pages 147–178. Springer, 1999.
10. Rainer Kolisch and Sonke Hartmann. Experimental investigation of heuristics for
resource-constrained project scheduling: An update. European Journal of Operational Research, 174(1):23–37, 2006.
11. Rainer Kolisch and Arno Sprecher. PSPLIB – a project scheduling problem library:
OR software – ORSEP operations research software exchange program. European
Journal of Operational Research, 96(1):205–216, 1997.
12. Jeremy Lau, Matthew Arnold, Michael Hind, and Brad Calder. Online performance auditing: using hot optimizations without getting burned. SIGPLAN Not.,
41(6):239–251, 2006.
13. Sasu Tarkoma, Christian Esteve Rothenberg, and Eemil Lagerspetz. Theory and
practice of bloom filters for distributed systems. IEEE Communications Surveys
& Tutorials, 14(1):131–155, 2012.
| 8cs.DS
|
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
1
Holistic, Instance-level Human Parsing
Qizhu Li*
[email protected]
Anurag Arnab*
Department of Engineering Science
University of Oxford
Oxford, UK
[email protected]
Philip H.S. Torr
arXiv:1709.03612v1 [cs.CV] 11 Sep 2017
[email protected]
Abstract
Object parsing – the task of decomposing an object into its semantic parts – has
traditionally been formulated as a category-level segmentation problem. Consequently,
when there are multiple objects in an image, current methods cannot count the number
of objects in the scene, nor can they determine which part belongs to which object. We
address this problem by segmenting the parts of objects at an instance-level, such that
each pixel in the image is assigned a part label, as well as the identity of the object it belongs to. Moreover, we show how this approach benefits us in obtaining segmentations
at coarser granularities as well. Our proposed network is trained end-to-end given detections, and begins with a category-level segmentation module. Thereafter, a differentiable
Conditional Random Field, defined over a variable number of instances for every input
image, reasons about the identity of each part by associating it with a human detection.
In contrast to other approaches, our method can handle the varying number of people
in each image and our holistic network produces state-of-the-art results in instance-level
part and human segmentation, together with competitive results in category-level part
segmentation, all achieved by a single forward-pass through our neural network.
1 Introduction
Object parsing, the segmentation of an object into semantic parts, is naturally performed by
humans to obtain a more detailed understanding of the scene. When performed automatically
by computers, it has many practical applications, such as in human-robot interaction, human
behaviour analysis and image descriptions for the visually impaired. Furthermore, detailed
part information has been shown to be beneficial in other visual recognition tasks such as
fine-grained recognition [47], human pose estimation [13] and object detection [37]. In
this paper, we focus on the application of parsing humans as it is more commonly studied,
although our method makes no assumptions on the type of object it is segmenting.
In contrast to existing human parsing approaches [18, 29, 45], we operate at an instance
level (to our knowledge, we are the first work to do so). As shown in Fig. 1, not only do we
segment the various body parts of humans (Fig. 1b), but we associate each of these parts to
one of the humans in the scene (Fig. 1c), which is particularly important for understanding
scenes with multiple people. In contrast to existing instance segmentation work [10, 31,
* indicates equal contribution by the authors.
c 2017. The copyright of this document resides with its authors.
It may be distributed unchanged freely in print or electronic forms.
2
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
(a) Input image
(b) Category-level
Part Segmentation
(c) Instance-level
Part Segmentation
(d) Instance-level
Human Segmentation
Figure 1: Our proposed approach segments human parts at an instance level (c) (which to
our knowledge is the first work to do so) from category-level part segmentations produced
earlier in the network (b). Moreover, we can easily obtain human instance segmentations (d)
by taking the union of all pixels associated to a particular person. Therefore, our proposed
end-to-end trained neural network parses humans into semantic parts at both category and
instance level in a single forward-pass. Best viewed in colour.
34], we operate at a more detailed part level, enabling us to extract more comprehensive
information of the scene. Furthermore, with our part-level instance segmentation of humans,
we can easily recover human-level instance segmentation (by taking the union of all parts
assigned to a particular instance as shown in Fig. 1d), and we show significant improvement
over previous state-of-the-art in human instance-segmentation when doing so.
Our approach is based on a deep Convolutional Neural Network (CNN), which consists
of an initial category-level part segmentation module. Using the output of a human detector,
we are then able to associate segmented parts with detected humans in the image using a
differentiable Conditional Random Field (CRF), producing a part-level instance segmentation of the image. Our formulation is robust to false-positive detections as well as imperfect
bounding boxes which do not cover the entire human, in contrast to other instance segmentation methods based on object detectors [10, 20, 21, 26, 34]. Given object detections, our
network is trained end-to-end, given detections, with a novel loss function which allows us
to handle a variable number of human instances on every image.
We evaluate our approach on the Pascal Person-Parts [8] dataset, which contains humans
in a diverse set of poses and occlusions. We achieve state-of-the-art results on instancelevel segmentation of both body parts and humans. Moreover, our results on semantic part
segmentation (which is not-instance aware) is also competitive with current state-of-theart. All of these results are achieved with a holistic, end-to-end trained model which parses
humans at both an instance and category level, and outputs a dynamic number of instances
per image, all in a single forward-pass through the network.
2
Related Work
The problem of object parsing, which aims to decompose objects into their semantic parts,
has been addressed by numerous works [27, 29, 38, 43, 45], most of which have concentrated
on parsing humans. However, none of the aforementioned works have parsed objects at
an instance level as shown in Fig. 1, but rather category level. In fact, a lot of work on
human parsing has focussed on datasets such as Fashionista [46], ATR [27] and Deep Fashion
[35] where images typically contain only one, centred person. The notion of instance-level
segmentation only matters when more than one person is present in an image, motivating
us to evaluate our method on the Pascal Person-Parts dataset [8] where multiple people can
appear in unconstrained environments. Recent human parsing approaches have typically
been similar to semantic segmentation works using fully convolutional networks (FCNs)
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
3
[36], but trained to label parts [5, 6, 7] instead of object classes. However, methods using
only FCNs do not explicitly model the structure of a human body, and typically do not
perform as well as methods which do [29]. Structural priors of the human body have been
encoded using pictorial structures [15, 17], Conditional Random Fields (CRFs) [4, 23, 25,
43] and more recently, with LSTMs [29, 30]. The HAZN approach of [45] addressed the
problem that some parts are often very small compared to other parts and difficult to segment
with scale-variant CNNs. This scale variation was handled by a cascade of three separatelytrained FCNs, each parsing different regions of the image at different scales.
An early instance segmentation work by Winn et al. [44] predicted the parts of an object,
and then encouraged these parts to maintain a spatial ordering, characteristic of an instance,
using asymmetric pairwise potentials in a CRF. However, subsequent work has not operated
at a part level. Zhang et al. [48, 49] performed instance segmentation of vehicles using an
MRF. However, this graphical model was not trained end-to-end as done by [3, 32, 51] and
our approach. Furthermore, they assumed a maximum of 9 cars per image. Approaches using
recurrent neural networks [39, 40] can handle a variable number of instances per image by
segmenting an instance per time-step, but are currently restricted to only one object category.
Our method, on the other hand, is able to handle both an arbitrary number of objects, and
multiple object categories in the image with a single forward-pass through the network.
Various methods of instance segmentation have also involved modifying object detection
systems to output segments instead of bounding boxes [10, 20, 21, 26]. However, these
methods cannot produce a segmentation map of the image, as shown in Fig. 1, without postprocessing as they consider each detection independently. Although our method also uses an
object detector, it considers all detections in the image jointly with an initial category-level
segmentation, and produces segmentation maps naturally where one pixel cannot belong to
multiple instances in contrast to the aforementioned approaches. The idea of combining
the outputs of a category-level segmentation network and an object detector to reason about
different instances was also presented by [1]. However, that system was not trained end-toend, could not segment instances outside the detector’s bounding box, and did not operate at
a part level.
3
Proposed Approach
Our network (Fig. 2) consists of two components: a category-level part segmentation module, and an instance segmentation module. As both of these modules are differentiable, they
can be integrated into a single network and trained jointly. The instance segmentation module (Sec. 3.2) uses the output of the first category-level segmentation module (Sec. 3.1) as
well as the outputs of an object detector as its input. It associates each pixel in the categorylevel segmentation with an object detection, resulting in an instance-level segmentation of
the image. Given a H ×W × 3 input image, I, the category-level part segmentation module
produces a H ×W × (P + 1) dimensional output Q where P is the number of part classes in
the dataset and one background class. There can be a variable number, D, of human detections per image, and the output of the instance segmentation module is an H ×W × (PD + 1)
tensor denoting the probabilities, at each pixel in the image, of each of the P part classes
belonging to one of the D detections.
Two challenges of instance segmentation are the variable number of instances in every
image, and the fact that permutations of instance labels lead to identical results (in Fig. 1,
how we order the different people does not matter). Zhang et al. [48, 49] resolve these issues
by assuming a maximum number of instances and using the ground-truth depth ordering
4
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
5×𝐷
Input: 𝐻 × 𝑊 × 3
Output: 𝐻 × 𝑊 × (𝑃𝐷 + 1)
Box Consistency
Term
Human Detector
+
Fully Convolutional
Network
Instance CRF
Global Term
Category-level
Segmentation Module
𝐻 × 𝑊 × (𝑃 + 1)
Instance-level
Segmentation Module
Figure 2: Our proposed approach. An H × W × 3 image is input to a human detection
network and a body parts semantic segmentation network, producing D detections of human
and an H × W × (P + 1) dimensional feature map respectively, where (P + 1) is the size of
the semantic label space including a background class. These results are used to form the
unary potentials of an Instance CRF which performs instance segmentation by associating
labelled pixels with human detections. In the above diagram, dotted lines represent forward
only paths, and solid lines show routes where both features and gradients flow. The green
boxes form the instance-level segmentation module (Sec. 3.2). Best viewed in colour.
of instances respectively. Others have bypassed both of these issues by predicting each
instance independently [10, 20, 21, 26], but this also allows a pixel to belong to multiple
instances. Instead, we use a loss function (Sec 3.3) that is based on “matching” the prediction
to the ground-truth, allowing us to handle permutations of the ground truth. Furthermore,
weight-sharing in our instance segmentation module allows us to segment a variable number
of instances per image. As a result, we do not assume a maximum number of instances,
consider all instances jointly, and train our network end-to-end, given object detections.
3.1
Category-level part segmentation module
The part segmentation module is a fully convolutional network [36] based on ResNet-101
[22]. A common technique, presented in [6, 7], is to predict the image at three different
scales (with network weights shared among all the scales), and combine predictions together
with learned, image-dependent weights. We take a different approach of fusing information
at multiple scales – we pool the features after res5c [22] at five different resolutions (by
varying the pooling stride), upsample the features to the resolution before pooling, and then
concatenate these features before passing them to the final convolutional classifier, as proposed in [50]. As we show in Sec 4.4, this approach achieves better semantic segmentation
results than [6, 7]. We denote the output of this module by the tensor, Q, where Qi (l) is the
probability of pixel i being assigned label l ∈ {0, 1, 2, ..., P}. Further details of this module
are included in the appendix.
3.2
Instance-level segmentation module
This module creates an instance-level segmentation of the image by associating each pixel
in the input category-level segmentation, Q, with one of the D input human-detections or
the background label. Let there be D input human-detections for the image, where the i-th
detection is represented by Bi , the set of pixels lying within the four corners of its bounding
box, and si ∈ [0, 1], the detection score. We assume that the 0-th detection refers to the
background label. Furthermore, we define a multinomial random variable, Vi , at each of the
N pixels in the image, and let V = [V1 ,V2 , ...,VN ]> . This variable can take on a label from the
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
5
set {1, 2, ..., D} × {1, 2, ..., P} ∪ {(0, 0)} since each of the P part labels can be associated with
one of the D human detections, or that pixel could belong to the background label, (0, 0).
We formulate a Conditional Random Field over these V variables, where the energy of
the assignment v to all of the instance variables V consists of two unary terms, and one
pairwise term (whose weighting co-efficients are all learned via backpropagation):
N
N
i
i< j
E(V = v) = − ∑ ln [w1 ψBox (vi ) + w2 ψGlobal (vi ) + ε] + ∑ ψPairwise (vi , v j ).
(1)
The unary and pairwise potentials are computed within our neural network, differentiable
with respect to their input and parameters, and described in Sec. 3.2.1 through 3.2.3. The
Maximum-a-Posteriori (MAP) estimate of our CRF (since the energy in Eq. 1 characterises a
Gibbs distribution) is computed as the final labelling produced by our network. We perform
the iterative mean-field inference algorithm to approximately compute the MAP solution by
minimising Eq. 1. As shown by Zheng et al. [51], this can be formulated as a Recurrent
Neural Network (RNN), allowing it to be trained end-to-end as part of a larger network.
However, as our network is input a variable number of detections per image, D, the label
space of the CRF is dynamic. Therefore, unlike [51], the parameters of our CRF are not
class-specific to allow for this variable number of “channels”.
3.2.1
Box Consistency Term
We observe that in most cases, a body part belonging to a person is located inside the bounding box of the person. Based on this observation, the box consistency term is employed
to encourage pixel locations inside a human bounding box Bi to be associated with the i-th
human detection. The box term potential at spatial location k for body part j of a human i is
assigned either 0 for k ∈
/ Bi , or the product of the detection score, si , and the category-level
part segmentation confidence, Qk ( j), for k ∈ Bi . For (i, j) ∈ {1, 2, ... , D} × {1, 2, ... , P},
(
si Qk ( j) if k ∈ Bi
ψBox (Vk = (i, j)) =
(2)
0
otherwise.
Note that this potential may be robust to false-positive detections when the category-level
segmentation and human detection do not agree with each other, since Qk (l), the probability
of a pixel k taking on body-part label l, is low. Furthermore, note that we use one humandetection to reason about the identity of all parts which constitute that human.
3.2.2
Global Term
A possible shortcoming for the box consistency potential is that if some pixels belonging
to a human instance fall outside the bounding box and are consequently assigned 0 for the
box consistency term potential, they would be lost in the final instance segmentation prediction. Visually, the generated instance masks would appear truncated along the bounding box
boundaries – a problem suffered by [1, 10, 21, 26]. To overcome this undesirable effect, we
introduce the global potential: it complements the box consistency term by assuming that a
pixel is equally likely to belong to any one of the detected humans. It is expressed as
ψGlobal (Vk = (i, j)) = Qk ( j),
for (i, j) ∈ {1, 2, ..., D} × {1, 2, ..., P} ∪ {(0, 0)}.
(3)
6
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
Prediction, P
Original ground-truth, Y
“Matched” ground-truth, Y ∗
Figure 3: As different permutations of the ground-truth are equivalent in the case of instance
segmentation, we “match” the original ground-truth, Y, to our network’s prediction, P, to
obtain the “matched” ground-truth which we use to compute our loss during training.
3.2.3
Pairwise Term
Our pairwise term is composed of densely-connected Gaussian kernels [24] which are commonly used in segmentation literature [5, 51]. This pairwise potential encourages both spatial
and appearance consistency, and we find these priors to be suitable in the case of instancelevel segmentation as well. As in [51], the weighting parameters of these potentials are
learned via backpropagation, though in our case, the weights are shared among all classes.
3.3
Loss function and network training
We first pre-train the category-level segmentation part of our network, as described in the appendix. Thereafter, we add the instance segmentation module, and train with a permutationinvariant loss function which is backpropagated through both our instance- and categorylevel segmentation networks. Since all permutations of an instance segmentation have the
same qualitative result, we “match” the original ground-truth to our prediction before computing the loss, as shown in Fig. 3. This matching is based on the Intersection over Union
(IoU) [14] of a predicted and ground-truth instance, similar to [40]. Let Y = {y1 , y2 , ..., ym },
a set of m segments, denote the ground-truth labelling of an image, where each segment is
an instance and has a part label assigned to it. Similarly, let P = {p1 , p2 , ..., pn } denote our
n predicted instances, each with an associated part label. Note that m and n need not be the
same as we may predict greater or fewer instances than there actually are in the image. The
“matched” ground truth, Y ∗ is the permutation of the original ground-truth labelling which
maximises the IoU between our prediction, P and ground-truth
Y ∗ = arg max IoU(Z, P),
(4)
Z ∈π(Y )
where π(Y) denotes the set of all permutations of Y. Note that we define the IoU between
all segments of different labels to be 0. Eq. 4 can be solved efficiently using the Hungarian
algorithm as it can be formulated as a bipartite graph matching problem, and once we have
the “matched” ground-truth, Y ∗ , we can apply any loss function to it and train our network
for segmentation.
In our case, we use the standard cross-entropy loss function on the “matched” ground
truth. In addition, we employ Online Hard Example Mining (OHEM), and only compute our
loss over the top K pixels with the highest loss in the training mini-batch. We found that
during training, many pixels already had a high probability of being assigned to the correct
class. By only selecting the top K pixels with the highest loss, we are able to encourage
our network to improve on the pixels it is currently misclassifying, as opposed to increasing
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
7
the probability of a pixel it is already classifying correctly. This approach was inspired by
“bootstrapping” [12, 42] or “hard-negative mining” [16] commonly used in training object
detectors. However, these methods mined hard examples from the entire dataset. Our approach is most similar to [41], who mined hard examples online from each mini-batch in the
context of detection. Similar to the aforementioned works, we found OHEM to improve our
overall results, as shown in Sec. 4.2.
3.4
Obtaining segmentations at other granularities
Given the part instance prediction produced by our proposed network, we are able to easily
obtain human instance segmentation and semantic part segmentation. In order to achieve
human instance segmentation, we map the predicted part instance labels (i, j), i.e. part j of
person i, to i. Whereas to obtain semantic part segmentation, we map predicted part instance
labels (i, j) to j instead.
4
Experiments
We describe our dataset and experimental set-up in Sec. 4.1, before presenting results on
instance-level part segmentation (Fig. 1c), instance-level human segmentation (Fig. 1d) and
semantic part segmentation (Fig. 1b). Additional quantitative and qualitative results, failure
cases and experimental details are included in the appendix.
4.1 Experimental Set-up
We evaluate our proposed method on the Pascal Person-Part dataset [13] which contains 1716
training images, and 1817 test images. This dataset contains multiple people per image in
unconstrained poses and environments, and contains six human body part classes (Fig. 1b),
as well as the background label. As described in Sec. 3.3, we initially pre-train our categorylevel segmentation module before training for instance-level segmentation. This module is
first trained on the 21 classes of the Pascal VOC dataset [14], and then finetuned on the seven
classes of the Pascal Part training set using category-level annotations. Finally, we train for
instance segmentation with instance-level ground truth. Full details of our training process,
including all hyperparameters such as learning rate, are in the appendix. To clarify these
details, we will also release our code.
We use the standard APr metric [20] for evaluating instance-level segmentation: the
mean Average Precision of our predictions is computed where a prediction is considered
correct if its IoU with a ground-truth instance is above a certain threshold. This is similar
to the AP metric used in object detection. However, in detection, the IoU between groundtruth and predicted bounding boxes is computed, whereas here, the IoU between regions is
computed. Furthermore, in detection, an overlap threshold of 0.5 is used, whereas we vary
r which is the mean of the APr score for overlap
this threshold. Finally, we define the APvol
thresholds varying from 0.1 to 0.9 in increments of 0.1.
We use the publicly available R-FCN detection framework [11], and train a new model
with data from VOC 2012 [14] that do not overlap with any of our test sets. We train with
all object classes of VOC, and only use the output for the human class. Non-maximal suppression is performed on all detections before being fed into our network.
8
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
Table 1: Comparison of APr at various IoU thresholds for instance-level part segmentation
on the Pascal Person-Parts dataset
Method
IoU threshold
0.5
0.6
0.7
MNC [10]
38.8
28.1
19.3
36.7
Ours, piecewise trained, box term only*
Ours, piecewise trained
Ours, end-to-end trained
38.0
38.8
39.0
27.4
28.5
28.6
16.7
17.6
17.4
36.6
37.3
37.7
Ours, piecewise trained, box term only, OHEM
Ours, piecewise trained, OHEM
Ours, end-to-end trained, OHEM
38.7
39.7
40.6
28.9
29.7
30.4
17.5
18.7
19.1
36.7
37.4
38.4
r
APvol
*Model is equivalent to our reimplementation of [1]
4.2
Results on Instance-level Part Segmentation
Table 1 shows our results on part-level instance segmentation on the Pascal Person-Part
dataset. To our knowledge, we are the first work to do this, and hence we study the effects of various design choices on overall performance. We also use the publicly available
code for MNC [10], which won the MS-COCO 2016 instance segmentation challenge, and
finetune their public model trained on VOC 2011 [19] on Person-Part instances as a baseline.
We first train our model in a piecewise manner, by first optimising the parameters of the
category-level segmentation module, and then “freezing” the weights of this module and only
training the instance network. Initially, we only use the box consistency term (Sec. 3.2.1) in
the Instance CRF, resulting in an APr at 0.5 of 38.0%. Note that this model is equivalent to
our reimplementation of [1]. Adding in the global potential (Sec. 3.2.2) helps us cope with
bounding boxes which do not cover the whole human, and we see an improvement at all
IoU thresholds. Training our entire network end-to-end gives further benefits. We then train
all variants of our model with OHEM, and observe consistent improvements across all IoU
thresholds with respect to the corresponding baseline. Here, we set K = 215 , meaning that
we computed our loss over 215 or approximately 12% of the hardest pixels in each training
image (since we train at full resolution). We also employ OHEM when pre-training the
category-level segmentation module of our network, and observe minimal difference in the
final result if we use OHEM when training the category-level segmentation module but not
the instance segmentation module. Training end-to-end with OHEM achieves 2.6% higher
r over a piecewise-trained baseline model without OHEM
in APr at 0.5, and 1.8% higher APvol
and only the box term (second row), which is equivalent to the model of [1]. Furthermore,
r is 1.7% greater than the strong MNC [10] baseline. Note that although [21] also
our APvol
performed instance-level segmentation on the same dataset, their evaluation was only done
using human instance labels, which is similar to our following experiment on human instance
segmentation.
4.3 Results on Human Instance Segmentation
We can trivially obtain instance-level segmentations of humans (Fig 1d), as mentioned in
Sec. 3.4. Table 2 shows our state-of-the-art instance segmentation results for humans on the
VOC 2012 validation set [14]. We use the best model from the previous section as there is
9
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
Table 2: Comparison of APr at various IoU thresholds for instance-level human segmentation on the VOC
2012 validation set
Method
SDS [20]
Chen et al. [9]
PFN [28]
Arnab et al. [1]*
R2-IOS [31]
Arnab et al. [2]*
0.5
IoU threshold
0.6 0.7 0.8
0.9
47.9
48.3
48.4
58.6
60.4
65.6
31.8
35.6
38.0
52.6
51.2
58.0
0.1
–
0.6
–
5.9 41.3
10.7 51.8
–
–
14.6 57.4
15.7
22.6
26.5
41.1
33.2
46.7
3.3
6.5
16.5
30.4
–
33.0
r
APvol
Ours, piecewise 64.0 59.8 51.0 38.3 20.1 57.2
Ours, end-to-end 70.2 63.1 54.1 41.0 19.6 61.0
Table 3: Comparison of semantic part segmentation results on
the Pascal Person-Parts test set
Method
IoU [%]
DeepLab* [5]
Attention [7]
HAZN [45]
LG-LSTM [30]
Graph LSTM [29]
DeepLab v2 [6]
RefineNet [33]
53.0
56.4
57.5
58.0
60.2
64.9
68.6
Ours, pre-trained
Ours, final network
65.9
66.3
*Results obtained from supplementary material.
*Result reported in [45]
no overlap between the Pascal Person-Part training set, and the VOC 2012 validation set.
As Tab. 2 shows, our proposed approach outperforms previous state-of-the-art by a significant margin, particularly at high IoU thresholds. Our model receives extra supervision
in its part labels, but the fact that our network can implicitly infer relationships between
different parts whilst training may help it handle occluding instances better than other approaches, leading to better instance segmentation performance. The fact that our network is
trained with part-level annotations may also help it identify small features of humans better, leading to more precise segmentations and thus improvements at high APr thresholds.
Our APr at each IoU threshold for human instance segmentation is higher than that for part
instance segmentation (Tab. 1). This is because parts are smaller than entire humans, and
thus more difficult to localise accurately. An alternate method of performing instance-level
part segmentation may be to first obtain an instance-level human segmentation using another
method from Tab. 2, and then partition it into the various body parts of a human. However, our approach, which groups parts into instances, is validated by the fact that it achieves
state-of-the-art instance-level human segmentation performance.
4.4
Results on Category-level Part Segmentation
Finally, our model is also able to produce category-level segmentations (as shown in Fig. 1b).
This can be obtained from the output of the category-level segmentation module, or from our
instance module as described in Sec. 3.4. As shown in Tab. 3, our semantic segmentation
results are competitive with current state-of-the-art. By training our entire network consisting
of the category-level and instance-level segmentation modules jointly, and then obtaining
the semantic segmentation from the final instance segmentation output by our network, we
are able to obtain a small improvement of 0.4% in mean IoU over the output of the initial
semantic segmentation module.
5
Conclusion
Our proposed, end-to-end trained network outputs instance-level body part and human segmentations, as well as category-level part segmentations in a single forward-pass. Moreover,
10
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
Input
Semantic Segmentation
Instance Segmentation
Ground Truth
Figure 4: Some results of our system. The first column shows the input image and the input
detections we obtained from training the R-FCN detector [11]. The second and third columns
show our final semantic segmentation (Sec. 3.4) and instance-level part segmentation. First
row: our network can deal with poor bounding box localisation, as it manages to segment
the third person from the left although the bounding box only partially covers her. Second
row: our method is robust against false positive detections because of the box term. Observe
that the bowl of the rightmost person in the bottom row is falsely detected as a person, but
rejected in the final prediction. Following rows: we are able to handle overlapping bounding
boxes by reasoning globally using the Instance CRF.
we have shown how segmenting objects into their constituent parts helps us segment the object as a whole with our state-of-the-art results on instance-level segmentation of both body
parts and entire humans. Furthermore, our category-level segmentations improve after training for instance-level segmentation. Our future work is to train the object detector end-to-end
as well. Moreover, the improvement that we obtained in instance segmentation of humans as
a result of first segmenting parts motivates us to explore weakly-supervised methods which
do not require explicit object part annotations.
Acknowledgement We thank Stuart Golodetz for discussions and feedback. This work
was supported by the EPSRC, Clarendon Fund, ERC grant ERC-2012-AdG 321162-HELIOS,
EPSRC grant Seebibyte EP/M013774/1 and EPSRC/MURI grant EP/N019474/1.
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
11
References
[1] Anurag Arnab and Philip H. S. Torr. Bottom-up instance segmentation with deep higher order
crfs. In BMVC, 2016.
[2] Anurag Arnab and Philip HS Torr. Pixelwise instance segmentation with a dynamically instantiated network. In CVPR, 2017.
[3] Anurag Arnab, Sadeep Jayasumana, Shuai Zheng, and Philip H. S. Torr. Higher order conditional
random fields in deep neural networks. In ECCV, 2016.
[4] Matthieu Bray, Pushmeet Kohli, and Philip Torr. Posecut: Simultaneous segmentation and 3d
pose estimation of humans using dynamic graph-cuts. In ECCV, pages 642–655. Springer, 2006.
[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille.
Semantic image segmentation with deep convolutional nets and fully connected crfs. ICLR, 2015.
[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille.
Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and
fully connected crfs. arXiv preprint arXiv:1606.00915, 2016.
[7] Liang-Chieh Chen, Yi Yang, Jiang Wang, Wei Xu, and Alan L Yuille. Attention to scale: Scaleaware semantic image segmentation. In CVPR, 2016.
[8] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille.
Detect what you can: Detecting and representing objects using holistic models and body parts.
In CVPR, 2014.
[9] Yi-Ting Chen, Xiaokai Liu, and Ming-Hsuan Yang. Multi-instance object segmentation with
occlusion handling. In CVPR, pages 3470–3478, 2015.
[10] Jifeng Dai, Kaiming He, and Jian Sun. Instance-aware semantic segmentation via multi-task
network cascades. In CVPR, 2016.
[11] Jifeng Dai, Yi Li, Kaiming He, and Jian Sun. R-fcn: Object detection via region-based fully
convolutional networks. In NIPS, 2016.
[12] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In CVPR,
volume 1, pages 886–893. IEEE, 2005.
[13] Jian Dong, Qiang Chen, Shuicheng Yan, and Alan Yuille. Towards unified object detection and
semantic segmentation. In ECCV, pages 299–314, 2014.
[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman.
The pascal visual object classes (voc) challenge. IJCV, 2010.
[15] Pedro F Felzenszwalb and Daniel P Huttenlocher. Pictorial structures for object recognition.
IJCV, 61(1):55–79, 2005.
[16] Pedro F Felzenszwalb, Ross B Girshick, David McAllester, and Deva Ramanan. Object detection
with discriminatively trained part-based models. PAMI, 32(9):1627–1645, 2010.
[17] Martin A Fischler and Robert A Elschlager. The representation and matching of pictorial structures. IEEE Transactions on computers, 100(1):67–92, 1973.
[18] Ke Gong, Xiaodan Liang, Xiaohui Shen, and Liang Lin. Look into person: Self-supervised
structure-sensitive learning and a new benchmark for human parsing. In CVPR, 2017.
12
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
[19] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In ICCV, pages 991–998. IEEE, 2011.
[20] Bharath Hariharan, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Simultaneous detection
and segmentation. In ECCV, pages 297–312. Springer, 2014.
[21] Bharath Hariharan, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Hypercolumns for object
segmentation and fine-grained localization. CVPR, 2015.
[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image
recognition. In CVPR, 2016.
[23] Martin Kiefel and Peter Vincent Gehler. Human pose estimation with fields of parts. In ECCV,
pages 331–346. Springer, 2014.
[24] P. Krähenbühl and V. Koltun. Efficient inference in fully connected CRFs with Gaussian edge
potentials. In NIPS, 2011.
[25] Lubor Ladicky, Philip HS Torr, and Andrew Zisserman. Human pose estimation using a joint
pixel-wise and part-wise formulation. In CVPR, pages 3578–3585, 2013.
[26] Ke Li, Bharath Hariharan, and Jitendra Malik. Iterative Instance Segmentation. In CVPR, 2016.
[27] Xiaodan Liang, Si Liu, Xiaohui Shen, Jianchao Yang, Luoqi Liu, Jian Dong, Liang Lin, and
Shuicheng Yan. Deep human parsing with active template regression. PAMI, 37(12):2402–2414,
2015.
[28] Xiaodan Liang, Yunchao Wei, Xiaohui Shen, Jianchao Yang, Liang Lin, and Shuicheng Yan.
Proposal-free network for instance-level object segmentation. arXiv preprint arXiv:1509.02636,
2015.
[29] Xiaodan Liang, Xiaohui Shen, Jiashi Feng, Liang Lin, and Shuicheng Yan. Semantic object
parsing with graph lstm. In ECCV, pages 125–143. Springer, 2016.
[30] Xiaodan Liang, Xiaohui Shen, Donglai Xiang, Jiashi Feng, Liang Lin, and Shuicheng Yan. Semantic object parsing with local-global long short-term memory. In CVPR, June 2016.
[31] Xiaodan Liang, Yunchao Wei, Xiaohui Shen, Zequn Jie, Jiashi Feng, Liang Lin, and Shuicheng
Yan. Reversible recursive instance-level object segmentation. In CVPR, 2016.
[32] Guosheng Lin, Chunhua Shen, and Ian Reid. Efficient piecewise training of deep structured
models for semantic segmentation. In CVPR, 2016.
[33] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement
networks with identity mappings for high-resolution semantic segmentation. In CVPR, 2017.
[34] Shu Liu, Xiaojuan Qi, Jianping Shi, Hong Zhang, and Jiaya Jia. Multi-scale patch aggregation
(mpa) for simultaneous detection and segmentation. In CVPR, 2016.
[35] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaoou Tang. Deepfashion: Powering robust
clothes recognition and retrieval with rich annotations. In CVPR, pages 1096–1104, 2016.
[36] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic
segmentation. In CVPR, 2015.
[37] Davide Modolo and Vittorio Ferrari. Learning semantic part-based models from google images.
In arXiv preprint arXiv:1609.03140, 2016.
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
13
[38] Greg Mori, Xiaofeng Ren, Alexei A Efros, and Jitendra Malik. Recovering human body configurations: Combining segmentation and recognition. In CVPR. IEEE, 2004.
[39] Mengye Ren and Richard S Zemel. End-to-end instance segmentation and counting with recurrent attention. In CVPR, 2017.
[40] Bernardino Romera-Paredes and Philip HS Torr. Recurrent instance segmentation. In ECCV,
2016.
[41] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors
with online hard example mining. In CVPR, pages 761–769, 2016.
[42] Kah-Kay Sung. Learning and example selection for object and pattern detection. In MIT A.I.
Memo No. 1521, 1996.
[43] Peng Wang, Xiaohui Shen, Zhe Lin, Scott Cohen, Brian Price, and Alan L Yuille. Joint object
and part segmentation using deep learned potentials. In ICCV, pages 1573–1581, 2015.
[44] John Winn and Jamie Shotton. The layout consistent random field for recognizing and segmenting
partially occluded objects. In CVPR, 2006.
[45] Fangting Xia, Peng Wang, Liang-Chieh Chen, and Alan L Yuille. Zoom better to see clearer:
Human and object parsing with hierarchical auto-zoom net. In ECCV, pages 648–663. Springer,
2016.
[46] Kota Yamaguchi, M Hadi Kiapour, Luis E Ortiz, and Tamara L Berg. Parsing clothing in fashion
photographs. In CVPR, pages 3570–3577. IEEE, 2012.
[47] Ning Zhang, Jeff Donahue, Ross Girshick, and Trevor Darrell. Part-based r-cnns for fine-grained
category detection. In ECCV, pages 834–849. Springer, 2014.
[48] Ziyu Zhang, Alexander G Schwing, Sanja Fidler, and Raquel Urtasun. Monocular object instance
segmentation and depth ordering with cnns. In ICCV, pages 2614–2622, 2015.
[49] Ziyu Zhang, Sanja Fidler, and Raquel Urtasun. Instance-level segmentation for autonomous
driving with deep densely connected mrfs. In CVPR, 2016.
[50] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene
parsing network. In CVPR, 2017.
[51] Shuai Zheng, Sadeep Jayasumana, Bernardino Romera-Paredes, Vibhav Vineet, Zhizhong Su,
Dalong Du, Chang Huang, and Philip Torr. Conditional random fields as recurrent neural networks. In ICCV, 2015.
14
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
Appendix
In this appendix, we present additional results of our proposed approach in Sec. A, and provide additional training and implementation details in Sec. B (both for our model, and the strong MNC baseline [10]).
A
Additional Results
In our main paper, we reported our APr results averaged over all classes. Fig. 5 visualises the perclass results of our best model at different IoU thresholds. Fig. 6 displays the success cases of our
method, while Fig. 7 shows examples of failure cases. Furthermore, we illustrate the strengths and
weaknesses of our part instance segmentation method in comparison to MNC [10] in Fig. 8, and compare our instance-level human segmentation results, which we obtain by the simple mapping described
in Sec. 3.4 of our main paper, to MNC in Fig. 9.
Finally, we attach an additional video. We run our system offline, on a frame-by-frame basis on
the entire music video, and show how our method is able to accurately parse humans at both category
and instance level on internet data outside the Pascal dataset. Instance-level segmentation of videos
requires data association. We use a simple, greedy method which operates on a frame-by-frame basis.
Segments from one frame are associated to segments in the next frame based on the IoU, using the
same method we use for our loss function as described in Sec. 3.3 of the main paper.
Figure 5: Visualisation of per-class results for different IoU thresholds on the Pascal
Person-Parts test set. The heatmap shows the per-class APr of our best model at IoU thresholds from 0.1 to 0.9 in increments of 0.1 on the Pascal Person-Parts test set. It shows that
our method achieves best instance accuracy for the head category, and finds lower arms and
lower legs most challenging to segment correctly. This is likely because of the thin shape of
the lower limbs which is known to pose difficulty for semantic segmentation.
15
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
Input
Semantic Segmentation
Instance Segmentation
Ground Truth
Figure 6: Success cases of our method. The first column shows the input image and the
input detections we obtained from training the R-FCN detector [11]. The second column
shows our final semantic segmentation (as described in Sec. 3.4 of the main paper). Our
proposed method is able to leverage an initial category-level segmentation network and human detections to produce accurate instance-level part segmentation as shown in the third
column.
16
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
Input
Semantic Segmentation
Instance Segmentation
Ground Truth
Figure 7: Failure cases of our method. First three rows: a missing human detection confuses the instance-level segmentation module. Fourth and fifth row: overlapping detection
bounding boxes lead to incorrect instance label assignment when the overlapping region are
visually similar. Sixth row: although our method is robust against false positive detections,
two small regions on the leftmost person’s left arm and left knee are assigned to the false
positive detection.
17
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
Input
MNC [10]
Ours
Ground Truth
Figure 8: Comparison to MNC on the Pascal Person-Parts [8] test set. First row: unlike MNC which predicts for each part instance independently, our method reasons globally
and jointly. As a result, MNC predicts two instances of lower legs for the same lower leg
of the second and third person from the left. Furthermore, with a dedicated category-level
segmentation module, we are less prone to false negatives, whereas MNC misses the legs of
the rightmost person, and the lower arm of the second person from the right. Second row:
while we can handle poor bounding box localisation because of our global potential term,
MNC is unable to segment regions outside the bounding boxes it generates. Consequently,
only one lower arm of the person on the left is segmented as the other one is outside the
bounding box. The square corners of the segmented lower arm correspond to the limits imposed by the bounding box which MNC internally uses (box generation is the first stage of
the cascade [10]). Third row: By analysing an image globally and employing a differentiable CRF, our method can produce more precise boundaries. As MNC does not perform
category-level segmentation over the entire image, it has no incentive to produce a coherent
and continuous prediction. Visually, this is reflected in the gaps of “background” between
body parts of the same person. Fourth row: MNC predicts two instances of lower leg for the
second person from the right, and fails to segment any lower arms for all four people due to
the aforementioned problems.
18
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
Input
MNC [10]
Ours
Ground Truth
Figure 9: Comparison to MNC on the Pascal Person-Parts [8] test set for instance-level
human segmentation. To generate the results in the second column, we run the public
MNC model trained on VOC 2011/SBD [19] using the default parameters and extract only
its human instance predictions. In contrast with proposal-driven methods such as MNC,
our approach assigns each pixel to only one instance, is robust against non-ideal bounding
boxes, and often produces better boundaries due to the Instance CRF which is trained endto-end. First and second row: since MNC predicts instances independently, it is prone to
predicting multiple instances for a single person. Third row: due to the global potential term,
we can segment regions outside of a detection bounding box which fails to cover the entire
person, whereas MNC is unable to recover from such imperfect bounding boxes, leading to
its frequent occurrences of truncated instance predictions. Fourth row: a case where MNC
and our method show different failure modes. MNC predicts three people where there are
only two, and our method can only predict one instance due to a missing detection.
19
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
Input
MNC [10]
Ours
Ground Truth
Figure 9 (Continued): Comparison to MNC on the Pascal Person-Parts [8] test set for
instance-level human segmentation. First row: MNC is unable to recover from a false
positive detection and predicts two people. Second row: while both MNC and our method
start off with poor bounding box localisation that does not cover the whole instance, we are
able to segment the entire person, whereas MNC is bounded by its flawed region proposal.
Third row: MNC performs better in this case as it is able to segment the infant, whereas we
miss her completely due to a false negative person detection.
20
B
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
Additional information
We detail our initial category-level segmentation module and compare it to DeepLab-v2 [6] in Sec. B.1,
present our network training details in Sec. B.2, and finally describe how we train the MNC model
which serves as our baseline in Sec. B.3.
B.1
Details of the category-level segmentation module
As shown in Fig 10b, the structure of our category-level segmentation module consists of a ResNet101 backbone, and a classifier that extracts multi-scale features from the ResNet-101 output by using
average pooling with different kernel sizes. While our category-level segmentation module and the
Deeplab-v2 network (Fig. 10a) of Chen et al. [6] both attempt to exploit multi-scale information in
the image, the approach of [6] entails executing three forward passes for each image, whereas we only
need a single forward pass.
In comparison to Deeplab-v2, our network saves both memory and time, and achieves better performance. To carry out a single forward pass, our network uses 4.3GB of memory while Deeplab-v2 [6]
needs 9.5GB, 120% more than ours. Speed-wise, our network runs forward passes at 0.255 seconds
per image (3.9 fps), whereas Deeplab-v2 takes 55% longer, at 0.396 seconds per image (2.5 fps) on
average. When Deeplab-v2 adds a CRF with 10 mean-field iterations to post-process the network output, it gains a small improvement in mean IoU by 0.54% [6], but it requires 11.2GB of memory to
make a forward pass (140% of the total amount used by our full network including the instance-level
segmentation module), and takes 0.960 seconds per image (1.0 fps), almost a quater of our frame rate.
Tests are done on a single GeForce GTX Titan X (Maxwell) card. Overall, we are able to achieve better
segmentation accuracy (as shown in Tab. 3 of our main paper) and is more memory- and time-efficient
than Deeplab-v2.
B.2 Training our proposed network
B.2.1 Training the category-level segmentation module
We initialise our semantic segmentation network with the COCO pre-trained ResNet-101 weights provided by [6]. Training is first performed on the Pascal VOC 2012 training set using the extra annotations from [19], which combine to a total of 9012 training images. Care is taken to ensure that all
images from the Pascal Person-Parts test set is excluded from this training set. A polynomial learning
i p
rate policy is adopted such that the effective learning rate at iteration i is given by li = l0 (1 − imax
) ,
where the base learning rate, l0 , is set to 6.25 × 10−4 , the total number of iterations, imax , is set to 30k,
and the power, p, is set to 0.9. A batch size of 16 is used. However, due to memory constraints, we
simulate this batch size by “accumulating gradients”: We carry out 16 forward and backward passes
with one image per iteration, and only perform the weight update after completing all 16 passes. We
use a momentum of 0.9 and weight decay of 1 × 10−4 for these experiments. After 30k of iterations
are completed, we take the best performing model and finetune on the Pascal Person-Parts training set
using the same training scheme as described above. Note that the parameters of the batch normalisation
modules are kept unchanged in the whole learning process.
Online data-augmentation is performed during training to regularise the model. The training images are randomly mirrored, scaled by a ratio between 0.5 and 2, rotated by an angle between -10 and
10 degrees, translated by a random amount in the HSV colour space, and blurred with a randomly-sized
Gaussian kernel, all on-the-fly. We observe that these techniques are effective at reducing the accuracy
gap between training and testing, leading to overall higher test accuracies.
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
(a)
21
(b)
Figure 10: Comparison of the Deeplab-v2 network structure which achieves 64.9% IoU on
the Pascal Person-Parts dataset [6] and our network structure. The numbers following the
layer type denote the kernel size and number of filters. For pooling layers, only their kernel
sizes are shown as the number of filters is not applicable. The upsampling ratios can be inferred from the context. Fig. 10a: in the Deeplab-v2 architecture, a 513×513×3 input image
is downsampled by two different ratios (0.75 and 0.5) to produce multi-scale input at three
different resolutions. The three resolutions are independently processed by a ResNet-101based network using shared weights (shown by the individually coloured paths). The output
feature maps are then upsampled where appropriate, combined by taking the elementwise
maximum, and finally upsampled back to 513×513. Fig. 10b: the category-level segmentation module proposed in this paper forwards an input image of size 521×521×3 through
a ResNet-101-based CNN, producing a feature map of resolution 66×66×2048. This feature map is average-pooled with four different kernel sizes, giving us four feature maps
with spatial resolutions 1×1, 2×2, 3×3, and 6×6 respectively. Each feature map undergoes
convolution and upsampling, before being concatenated together with each other and the
66×66×2048 ResNet-101 output. This is followed by a convolution layer that reduces the
dimension of the concatenated features to 512, and a convolutional classifier that maps the
512 channels to the size of label space in the dataset. Finally, the prediction is upsampled
back to 521×521. In both Fig. 10a and 10b, the ResNet-101 backbone uses dilated convolution such that its output at res5c is at 1/8 of the input resolution, instead of 1/32 for the
original ResNet-101 [22]. The convolutional classifiers (coloured in purple) output C channels, corresponding to the number of classes in the dataset including a background class. For
the Pascal Person-Parts Dataset, C is 7. Best viewed in colour.
B.2.2
Training the instance-level segmentation module
In our model, the pairwise term of the fully-connected CRF takes the following form:
ψPairwise (vi , v j ) = µ(vi , v j )k(fi , f j )
(5)
where µ(·, ·) is a compatibility function, k(·, ·) is a kernel function, and fi is a feature vector at spatial
location i containing the 3-dimensional colour vector Ii and the 2-dimensional position vector pi [24].
22
LI, ARNAB, TORR: HOLISTIC, INSTANCE-LEVEL HUMAN PARSING
We further define the kernel as follows:
(1)
k(fi , f j ) = w
|p − p |2
|pi − p j |2 |Ii − I j |2
i
j
(2)
exp −
−
+ w exp −
2
2
2
2θα
2θβ
2θγ
(6)
where w(1) and w(2) are the linear combination weights for the bilateral term and the Gaussian term
respectively. In order to determine the initial values for the parameters in the Instance CRF to train
from, we carry out a random search. According to the search results, the best prediction accuracy is
obtained by initialising w(1) = 8, w(2) = 2, θα = 2, θβ = 8, θγ = 2. Furthermore, we use a fixed learning
rate of 1 × 10−6 , momentum of 0.9, and weight decay of 1 × 10−4 for training both the instance-level
and category-level segmentation modules jointly. Although we previously use the polynomial learning
rate policy, we find that for training the instance-level segmentation module, a fixed learning rate leads
to better results. Furthermore, our experiments show that a batch size of one works best at this training
stage. Using this scheme, we train for 175k iterations, or approximately 100 epochs.
B.3
Training Multi-task Network Cascades (MNC)
We use the publicly available Multi-task Network Cascades (MNC) framework [10], and train a new
model for instance-level part segmentation using the Pascal Person-Parts dataset. The weights are initialised with the officially released MNC model1 which has been trained on Pascal VOC 2011/SBD [19].
The base learning rate is set to 1 × 10−3 , which is reduced by 10 times after 20k iterations. A total
of 25k training iterations are carried out. A batch size of 8, momentum of 0.9 and weight decay of
5 × 10−4 are used. These settings are identical to the ones used in training the original MNC and provided in their public source code. Using these settings, we are also able to reproduce the experimental
results obtained in the original MNC paper [10], and hence we believe that the MNC model we have
trained acts as a strong baseline for our proposed approach.
1 https://github.com/daijifeng001/MNC
| 1cs.CV
|
Deep Multimodal Embedding: Manipulating Novel Objects
with Point-clouds, Language and Trajectories
I. I NTRODUCTION
Consider a robot manipulating a new appliance in a home
kitchen, e.g. the toaster in Figure 3. The robot must use
the combination of its observations of the world and natural
language instructions to infer how to manipulate objects.
Such ability to fuse information from different input modalities and map them to actions is extremely useful to many
applications of household robots [1], including assembling
furniture, cooking recipes, and many more.
Even though similar concepts might appear very differently in different sensor modalities, humans are able to
understand that they map to the same concept. For example,
when asked to “turn the knob counter-clockwise” on a
toaster, we are able to correlate the instruction language and
the appearance of a knob on a toaster with the motion to do
so. We also associate this concept more closely with a motion
which would incorrectly rotate in the opposite direction than
with, for example, the motion to press the toaster’s handle
downwards. There is strong evidence that humans are able
to correlate between different modalities through common
representations [2].
Obtaining a good common representation between different modalities is challenging for two main reasons. First,
each modality might intrinsically have very different statistical properties — for example, here our trajectory representation is inherently dense, while our representation of
language is naturally sparse. This makes it challenging to
apply algorithms designed for unimodal data. Second, even
1 Department of Computer Science, Cornell University. 2 Department
of Computer Science, Stanford University. 3 Brain Of Things, Inc.
{jysung,ianlenz,asaxena}@cs.cornell.edu
point-cloud (𝒑)
language (𝒍)
Abstract— A robot operating in a real-world environment
needs to perform reasoning over a variety of sensor modalities
such as vision, language and motion trajectories. However, it is
extremely challenging to manually design features relating such
disparate modalities. In this work, we introduce an algorithm
that learns to embed point-cloud, natural language, and manipulation trajectory data into a shared embedding space with a
deep neural network. To learn semantically meaningful spaces
throughout our network, we use a loss-based margin to bring
embeddings of relevant pairs closer together while driving lessrelevant cases from different modalities further apart. We use
this both to pre-train its lower layers and fine-tune our final
embedding space, leading to a more robust representation. We
test our algorithm on the task of manipulating novel objects
and appliances based on prior experience with other objects.
On a large dataset, we achieve significant improvements in
both accuracy and inference time over the previous state of
the art. We also perform end-to-end experiments on a PR2
robot utilizing our learned embedding space.
push
pull handle lever control
stove
cup
stop
fill towards
trajectory (𝝉)
arXiv:1509.07831v2 [cs.RO] 17 May 2017
Jaeyong Sung1,2 , Ian Lenz1 and Ashutosh Saxena3
𝒙
𝒉𝟏
𝒉𝟐
𝒉𝟑
Fig. 1. Deep Multimodal Embedding: Our deep neural network learns
to embed both point-cloud/natural language instruction combinations and
manipulation trajectories in the same semantically meaningful space, where
distance represents the relevance of embedded data.
with expert knowledge, it is extremely challenging to design
joint features between such disparate modalities. Designing
features which map different sensor inputs and actions to the
same space, as required here, is particularly challenging.
In this work, we use a deep neural network to learn
a shared embedding between the pairing of object parts
in the environment with natural language instructions, and
manipulation trajectories (Figure 1). This means that all
three modalities are projected to the same feature space. We
introduce an algorithm that learns to pull semantically similar
environment/language pairs and their corresponding trajectories to the same regions, and push environment/language
pairs away from irrelevant trajectories based on how irrelevant they are. Our algorithm allows for efficient inference
because, given a new instruction and point-cloud, we need
only find the nearest trajectory to the projection of this pair
in the learned embedding space, which can be done using
fast nearest-neighbor algorithms [3].
In the past, deep learning methods have shown impressive
results for learning features for a wide variety of domains
[4], [5], [6] and even learning cross-domain embeddings [7].
In contrast to these existing methods, here we present a new
pre-training algorithm for initializing networks to be used for
joint embedding of different modalities. Our algorithm trains
each layer to map similar cases to similar areas of its feature
space, as opposed to other methods which either perform
variational learning [8] or train for reconstruction [9].
In order to validate our approach, we test our model on
a large manipulation dataset, the Robobarista Dataset [1],
which focuses on learning to infer manipulation trajectories
for novel objects. We also present results on a series of
experiments on a PR2 robot, showing that our algorithm is
able to successfully manipulate several different objects. In
summary, the key contributions of this work are:
•
•
•
•
We present an algorithm which learns a semantically
meaningful embedding space by enforcing a varying
and loss-based margin.
We present an algorithm for unsupervised pre-training
of multi-modal features to be used for embedding which
outperforms standard pre-training algorithms [9].
We present a novel approach to manipulation via embedding multimodal data with deep neural networks.
Our approach allows for fast inference of manipulation
trajectories for novel objects, roughly 170x faster than
previous work [1] while also improving accuracy on a
large manipulation dataset [1].
II. R ELATED W ORK
A. Metric Embedding
Several works in machine learning make use of the power
of shared embedding spaces. Large margin nearest neighbors
(LMNN) [10] learns a max-margin Mahalanobis distance for
a unimodal input feature space. [11] learn linear mappings
from image and language features to a common embedding
space for automatic image annotation. [12] learn to map
songs and natural language tags to a shared embedding
space. However, these approaches learn only a shallow,
linear mapping from input features, whereas here we learn
a deep non-linear mapping which is less sensitive to input
representations.
B. Deep Learning
Multimodal data: A deep neural network has been used to
learn features of video and audio [13]. With a generative
learning, a network can be robust to missing modalities
at inference time [14]. In these works, similar to [1], a
single network takes all modalities as inputs, whereas here
we perform joint embedding of multiple modalities using
multiple networks.
Joint embedding: Several works use deep networks for joint
embedding between different feature spaces. For translation,
a joint feature space is learned from different languages [15];
for annotation and retrieval, images and natural language
“tags” are mapped to the same space [7]. We present a new
pre-training algorithm for embedding spaces and show that it
outperforms the conventional methods used in these works.
A deep network is also used as metric learning for the
face verification task [16], which enforces a constant margin
between distances among inter-class objects and among
intra-class objects, similar to LMNN [10]. In Sec. VI-A, we
show that our approach, which uses a loss-dependent variable
margin, produces better results for our problem.
C. Robotic Manipulation
Many works in robotic manipulation focus on task-specific
manipulation with known objects — for example, folding
towels [17], baking cookies [18], or planar contact manipulation [19]. Others [20], [21] focus on sequencing manipulation
tasks or choosing when to switch skills [22], assuming
manipulation primitives such as pour are available. For novel
objects, affordances are predicted and associated motions are
applied [23]. Instead, similar to [1], we skip intermediate
representations and directly generalize to novel objects.
A few recent works use deep learning approaches for
robotic manipulation. Deep networks have been used to
detect stable grasps from RGB-D data [24], [25]. [26] use a
Gaussian mixture model to learn system dynamics, then use
these to learn a manipulation policy using a deep network.
[27] use a deep network to learn system dynamics for realtime model-predictive control. Both these works focus on
learning low-level input-output controllers. Here, we instead
focus on inferring full 6-DoF trajectories, which such controllers could then be used to follow.
Sung et al. [1] perform object part-based transfer of
manipulation trajectories for novel objects and introduces
a large manipulation dataset including objects like epresso
machine and urinal. We primarily test our algorithm on this
dataset. In Sec. VI-A, we show that our approach gives better
accuracy than this prior work, while also running 171x faster.
III. OVERVIEW
The main challenge of our work is to learn a model
which maps three disparate modalities — point-clouds, natural language, and trajectories — to a single semantically
meaningful space. In particular, we focus on point-clouds
of object parts, natural language instructing manipulation of
different objects, and trajectories that would manipulate these
objects.
We introduce a method that learns a common pointcloud/language/trajectory embedding space in which the projection of a task (point-cloud/language combination) should
have a higher similarity to projections of relevant trajectories
than task-irrelevant trajectories. Among these irrelevant trajectories, some might be less relevant than others, and thus
should be pushed further away.
For example, given a door knob, that needs to be grasped
normal to the door surface, with an instruction to rotate it
clockwise, a trajectory that correctly approaches the door
knob but rotates counter-clockwise should have higher similarity to the task than one which approaches the knob from a
completely incorrect angle and does not execute any rotation.
We learn non-linear embeddings using a deep learning approach, as shown in Fig. 1, which maps raw data from these
three different modalities to a joint embedding space. Prior
to learning a full joint embedding of all three modalities, we
pre-train embeddings of subsets of the modalities to learn
semantically meaningful embeddings for these modalities.
We show in Sec. V that a learned joint embedding space
can be efficiently used for finding an appropriate manipulation trajectory for objects with natural language instructions.
A. Problem Formulation
Given tuples of a scene p ∈ P, a natural language
instruction l ∈ L and an end-effector trajectories t ∈ T , our
goal is to learn a joint embedding space and two different
mapping functions that map to this space—one from a pointcloud/language pair and the other from a trajectory.
More formally, we want to learn ΦP,L (p, l) and ΦT (τ )
which map to a joint feature space RM :
ΦP,L (p, l) : (P, L) → RM
ΦT (τ ) : T → RM
The first, ΦP,L , which maps point-clouds and languages,
is defined as a combination of two mappings. The first of
these maps to a joint point-cloud/language space RN2,pl
— ΦP (p) : P → RN2,pl and ΦL (l) : L → RN2,pl .
N2,pl represents the size of dimensions p, l are embedded
jointly. Once each is mapped to RN2,pl , this space is then
mapped to the joint space shared with trajectory information:
ΦP,L (p, l) : ((P, L) → RN2,pl ) → RM .
IV. L EARNING J OINT
P OINT- CLOUD /L ANGUAGE /T RAJECTORY M ODEL
In our joint feature space, proximity between two mapped
points should reflect how relevant two data-points are to each
other, even if they are from completely different modalities.
We train our network to bring demonstrations that manipulate a given object according to some language instruction
closer to the mapped point for that object/instruction pair,
and to push away demonstrations that would not correctly
manipulate that object. Trajectories which have no semantic
relevance to the object are pushed much further away than
trajectories that have some relevance, even if the latter would
not fully manipulate the object according to the instruction.
For every training point-cloud/language pair (pi , li ), we
have a set of demonstrations Ti and the most optimal demonstration trajectory τi∗ ∈ Ti . Using the optimal demonstration
τi∗ and a loss function ∆(τ, τ̄ ) for comparing demonstrations,
we find a set of trajectories Ti,S that are relevant (similar) to
this task and a set of trajectories Ti,D that are irrelevant
(dissimilar.) We use the DTW-MT distance function (described later in Sec. VI) as our loss function ∆(τ, τ̄ ), but
it could be replaced by any function that computes the loss
of predicting τ̄ when τ is the correct demonstration. Using a
strategy previously used for handling noise in crowd-sourced
data [1], we can use thresholds (tS , tD ) to generate two sets
from the pool of all trajectories:
Ti,S = {τ ∈ T |∆(τi∗ , τ ) < tS }
Ti,D = {τ ∈ T |∆(τi∗ , τ ) > tD }
For each pair of (pi , li ), we want all projections of τj ∈
Ti,S to have higher similarity to the projection of (pi , li ) than
τk ∈ Ti,D . A simple approach would be to train the network
to distinguish these two sets by enforcing a finite distance
(safety margin) between the similarities of these two sets
[10], which can be written in the form of a constraint:
sim(ΦP,L (pi , li ), ΦT (τj )) ≥ 1 + sim(ΦP,L (pi , li ), ΦT (τk ))
point-cloud/language embedding
trajectory embedding
𝒉𝟐,𝒑𝒍
𝒉𝟐,𝝉
𝒉𝟏
𝒉𝟏
𝒙
𝒙
point-cloud (𝒑)
language (𝒍)
trajectory (𝝉)
trajectory (𝝉′)
Fig. 2.
Pre-training lower layers: Visualization of our pre-training
approaches for h2,pl and h2,τ . For h2,pl , our algorithm pushes matching
point-clouds and instructions to be more similar. For h2,τ , our algorithm
pushes trajectories with higher DTW-MT similarity to be more similar.
Rather than simply being able to distinguish two sets, we
want to learn semantically meaningful embedding spaces
from different modalities. Recalling our earlier example
where one incorrect trajectory for manipulating a door knob
was much closer to correct than another, it is clear that our
learning algorithm should drive some incorrect trajectories to
be more dissimilar than others. The difference between the
similarities of τj and τk to the projected point-cloud/language
pair (pi , li ) should be at least the loss ∆(τj , τk ). This can
be written as a form of a constraint:
∀τj ∈ Ti,S , ∀τk ∈ Ti,D
sim(ΦP,L (pi , li ), ΦT (τj ))
≥ ∆(τj , τk ) + sim(ΦP,L (pi , li ), ΦT (τk ))
Intuitively, this forces trajectories with higher DTW-MT
distance from the ground truth to embed further than those
with lower distance. Enforcing all combinations of these
constraints could grow exponentially large. Instead, similar
to the cutting plane method for structural support vector
machines [28], we find the most violating trajectory τ 0 ∈
Ti,D for each training pair of (pi , li , τi ∈ Ti,S ) at each
iteration. The most violating trajectory has the highest value
after the similarity is augmented with the loss scaled by a
constant α:
τi0 = arg max (sim(ΦP,L (pi , li ), ΦT (τ )) + α∆(τi , τ ))
τ ∈Ti,D
The cost of our deep embedding space h3 is computed as
the hinge loss of the most violating trajectory.
Lh3 (pi , li , τi ) = |∆(τi , τi0 )+sim(ΦP,L (pi , li ), ΦT (τi0 ))
−sim(ΦP,L (pi , li ), ΦT (τi ))|+
The average cost of each minibatch is back-propagated
through all the layers of the deep neural network using the
AdaDelta [29] algorithm.
A. Pre-training Joint Point-cloud/Language Model
One major advantage of modern deep learning methods
is the use of unsupervised pre-training to initialize neural
network parameters to a good starting point before the final
supervised fine-tuning stage. Pre-training helps these highdimensional networks to avoid overfitting to the training data.
Our lower layers h2,pl and h2,τ represent features extracted exclusively from the combination of point-clouds
Segmenting Object Parts
Input
Deep Multimodal Embedding
Object
Part
Candidate
Ranking
System
2
1
language (𝒍)
point-cloud (𝒑)
Scene:
Output
Trajectory
for each step of manual
1
XY-slice
XZ-slice
pull push stop lever control
turn cup handle fill down
Object Part Candidates
XZ-slice
XY-slice
Object Part 3D Grid
Representation
trajectory (𝝉)
2
(stitched point-cloud)
Instruction:
1 Turn the adjustment knob
clockwise to the desired
darkness.
2 Push down on the right lever
to start toasting.
𝒙
𝒉𝟏
𝒉𝟐
𝒉𝟑
Fig. 3. System Overview: Given a point-cloud and a language instruction, our goal is to output a trajectory that would manipulate the object according
to the instruction. The given point-cloud scene is segmented into many parts and ranked for each step of the instruction manual. By embedding point-cloud,
language, and trajectory modalities into a joint embedding space, our algorithm selects the best trajectory to transfer to the new object.
and language, and from trajectories, respectively. Our pretraining method initializes h2,pl and h2,τ as semantically
meaningful embedding spaces similar to h3 , as shown later
in Section VI-A.
First, we pre-train the layers leading up to these layers
using sparse de-noising autoencoders [30], [31]. Then, our
process for pre-training h2,pl is similar to our approach to
fine-tuning a semantically meaningful embedding space for
h3 presented above, except now we find the most violating
language l0 while still relying on a loss over the associated
optimal trajectory:
l0 = argmax(sim(ΦP (pi ), ΦL (l)) + α∆(τi , τ ))
l∈L
Lh2,pl (pi , li , τi ) = |∆(τi , τ 0 )+sim(ΦP (pi ), ΦL (l0 ))
−sim(ΦP (pi ), ΦL (li ))|+
Notice that although we are training this embedding space
to project from point-cloud/language data, we guide learning
using trajectory information.
After the projections ΦP and ΦL are tuned, the output of
these two projections are added to form the output of layer
h2,pl in the final feed-forward network.
B. Pre-training Trajectory Model
For our task of inferring manipulation trajectories for novel
objects, it is especially important that similar trajectories
τ map to similar regions in the feature space defined by
h2,τ , so that trajectory embedding h2,τ itself is semantically
meaningful and they can in turn be mapped to similar regions
in h3 . Standard pretraining methods, such as sparse denoising autoencoder [30], [31] would only pre-train h2,τ to
reconstruct individual trajectories. Instead, we employ pretraining similar to pre-training of h2,pl above, except now
we pre-train for only a single modality — trajectory data.
As shown on right hand side of Fig. 2, the layer that
embeds to h2,τ is duplicated. These layers are treated as
if they were two different modalities, but all their weights
are shared and updated simultaneously. For every trajectory
τ ∈ Ti,S , we can again find the most violating τ 0 ∈ Ti,D and
the minimize a similar cost function as we do for h2,pl .
V. A PPLICATION : M ANIPULATING N OVEL O BJECTS
As an example application, we consider manipulating
novel appliances [1]. Our goal is to use our learned embedding space to allow the robot to infer a manipulation
trajectory when it is introduced to a new appliance with its
natural language instruction manual. For example, as shown
in Fig. 3, given a point-cloud of a scene with a toaster and
an instruction such as ‘Push down on the right lever to start
toasting,’ it should output a trajectory, representative of how
a two-fingered end-effector should move, including how to
approach, grasp, and push down on the lever. Our algorithm
allows the robot to leverage prior experience with different
appliances — for example, a trajectory which manipulates
the handles of a paper towel dispenser might be transferred
to manipulate the toaster handle.
First, in order to correctly identify a part p out of a scene
s that an instruction asks to manipulate, a point-cloud of a
scene s is segmented into many small potential candidates.
All segments are ranked for each step of the manual instruction. Multiple variations of correct segmentations and lots of
incorrect segmentation make our embedding representation
even more robust as shown later in Sec. VI-A.
Then, from a library of trajectories with prior experience,
the trajectory that gives the highest similarity to the selected
point-cloud p and language l in our embedding space RM :
argmax sim(ΦP,L (p, l), ΦT (τ ))
τ ∈T
As in [11], similarity is defined as sim(a, b) = a · b.
The previous approach to this problem [1] requires projecting a new point-cloud and natural language instruction
with every trajectory in the training set exhaustively through
the network during inference.
Instead, our approach allows us to pre-embed all candidate
trajectories into a shared embedding space. The correct
trajectory can be identified by embedding only a new pointcloud/language pair. As shown in Sec. VI-A, this significantly improves both the inference run-time and accuracy
and makes it much more scalable to a larger training dataset.
A. Segmenting Object Parts from Point-clouds
Our learning algorithm (Sec. IV) assumes that object parts
p corresponding to each instruction l have already been
segmented from the point-cloud scene s. While our focus
is on learning to manipulate these segmented parts, we also
introduce a segmentation approach which allows us to both
build an end-to-end system and augment our training data
for better unsupervised learning, as shown in Sec. VI-A.
1) Generating Object Part Candidates: We employ a
series of geometric feature based techniques to segment a
scene s into small overlapping segments {p1 , p2 , ..., pn }. We
first extract Euclidean clusters of points while limiting the
difference of normals between a local and larger region [32],
[33]. We then filter out segments which are too big for human
hands. To handle a wide variety of object parts of different
scales, we generate two sets of candidates with two different
sets of parameters, which are combined for evaluation.
2) Part Candidate Ranking Algorithm: Given a set of
segmented parts p, we must now use our training data D
to select for each instruction lj the best-matching part p∗j .
We do so by optimizing the score ψ(pj , lj ) of each segment
pi for a step lj in a manual, evaluated in three parts:
ψ(pi , lj ; D) = ψf eat (pi , lj ) ψpc (pi , lj ) + ψlang (pi , lj )
The score ψpc is based on the kp -most identical segments
from the training data D, based on cosine similarity using
our grid representation (Sec. V-B). The score is a sum
of similarity against these
Pn segments and their associated
language: ψpc (pi , lj ) = k=1 (sim(pi , pk ) + β sim(lj , lk )).
If the associated language does not exist (i.e. pk is not a
manipulable part), it is given a set similarity value. Similarly,
the score ψlang is based on the kl -most identical language
instructions in D. It is a sum of similarity against identical
language and associated expert point-cloud segmentations.
The feature score ψf eat is computed by weighted features
wT φf eat (pi , lj ) as described in Sec. V-A.3. Each score of
the segmented parts ψ(pi , lj ) is then adjusted by multiplying
by ratio of its score against the marginalized score in the
ψ(pi ,lj )
ψ(pi , lj ). For each
manual: ψ̂(pi , lj ) = P
lk ∈mnew ψ(pi ,lk )
lj ∈ mnew , an optimal segment of the scene chosen as the
segment with the maximum score: maxpi ∈snew ψ̂(pi , lj ).
3) Features: Three features are computed for each segment in the context of the original scene. we first infer where
a person would stand by detecting the ‘front’ of the object, by
a plane segmentation constrained to have a normal axis less
than 45◦ from the line between the object’s centroid and the
original sensor location, assuming that the robot is introduced
close to the ‘front’. We then compute a ‘reach’ distance
from an imaginary person 170cm tall, which is defined as
the distance from the head of the person to each segment
subtracted by the distance of the closest one. Also, because
stitched point-clouds have significant noise near their edges,
we compute the distance from the imaginary view ray, a
line defined by the centroid of the scene to the head of the
person. Lastly, objects like a soda fountain a sauce dispenser
have many identical parts, making it difficult to disambiguate
different choices (e.g. Coke/Sprite, ketchup/mustard). Thus,
for such scenarios, we also provided a 3D point as if human
is pointing at the label of the desired selection. Note that
this does not point at the actual part but rather at its label or
vicinity. A distance from this point is also used as a feature.
B. Data Representation
All three data modalities (p, l, τ ) are variable-length and
must be transformed into a fixed-length representation.
Each point-cloud segment is converted into a real-valued
3D occupancy grid where each cell’s value is proportional to
how many points fall into the cube it spans. We use a 100 ×
100 × 100 grid of cubic cells with sides of 0.25cm. Unlike
our previous work [1], each cell count is also distributed
to the neighboring cells with an exponential distribution.
This smooths out missing points and increases the amount
of information represented. The grid then is normalized to
be between 0 ∼ 1 by dividing by the maximal count.
While our approach focuses on the shape of the part
in question, the shape of the nearby scene can also have
a significant effect on how the part is manipulated. To
account for this, we assign a value of 0.2 to any cell which
only contains points which belong to the scene but not the
specific part in question, but are within some distance from
the nearest point for the given part. To fill hollow parts
behind the background, such as tables and walls, we raytrace between the starting location of the sensor and cells
filled by background points and fill these similarly.
While our segment ranking algorithm uses the full-sized
grid for each segment, our main embedding algorithm uses
two compact grids generated by taking average of cells: 10×
10 × 10 grids with cells with sides of 2.5cm and of 1cm.
Natural language l is represented by a fixed-size bag-ofwords. Trajectories are variable-length sequences of waypoints, where each waypoint contains a position, an orientation and a gripper state (‘open’/‘closed’/‘holding’) defined
in the coordinate frame of object part. All trajectories τ are
then normalized to a fixed length of 15 waypoints. For more
details on trajectory representation, please refer to [1].
C. Robotic Platform
We tested our algorithms on a PR2, a mobile robot with
two arms with seven degrees of freedom each (Fig. 5). All
software for controlling the robot is written in ROS [34], and
the embedding algorithm are written with Theano [35]. All
of the computations are done on a remote computer utilizing
a GPU for our embedding model.
D. Parameters
Through validation, we found an optimal embedding space
size M of 25 and intermediate-layer sizes N1,p , N1,l , N1,τ ,
N2,pl , and N2,τ of 250, 150, 100, 125, and 100 with the loss
scaled by α = 0.2. These relatively small layer sizes also had
the advantage of fast inference, as shown in Sec. VI-A.
VI. E XPERIMENTS
In this section, we present two sets of experiments. First,
we show offline experiments on the Robobarista dataset [1]
which test individual components of our system, showing
Models
chance
object part classifier [1]
LSSVM + kinematic [1]
similarity + weights [1]
Sung et al. [1]
LMNN-like cost [10]
Ours w/o pretraining
Ours with SDA
Ours w/o Mult. Seg
Our Model
per manual
DTW-MT
28.0 (±0.8)
17.4 (±0.9)
13.3 (±1.2)
13.0 (±1.3)
15.4 (±1.8)
13.2 (±1.4)
11.5 (±0.6)
11.0 (±0.8)
10.3 (±0.8)
per instruction
DTW-MT
Accu. (%)
27.8 (±0.6)
11.2 (±1.0)
22.9 (±2.2)
23.3 (±5.1)
17.5 (±1.6)
40.8 (±2.5)
12.5 (±1.2)
53.7 (±5.8)
12.2 (±1.1)
60.0 (±5.1)
14.7 (±1.6)
55.5 (±5.3)
12.4 (±1.0)
54.2 (±6.0)
11.1 (±0.6)
62.6 (±5.8)
10.5 (±0.7)
65.1 (±4.9)
9.9 (±0.5)
68.4 (±5.2)
improvements for each. Second, we present a series of realworld robotic experiments which show that our system is
able to produce trajectories that can successfully manipulate
objects based on natural language instructions.
Dataset. We test our model on the Robobarista dataset [1].
This dataset consists of 115 point-cloud scenes with 154
manuals, consisting of 249 expert segmented point-clouds
and 250 free-form natural language instructions. It also
contains 1225 crowd-sourced manipulation trajectories which
are demonstrated for 250 point-cloud/language pairs. The
point-clouds are collected by stitching multiple views using
Kinect Fusion. The manipulation trajectories are collected
from 71 non-experts on Amazon Mechanical Turk.
Evaluation. All algorithms are evaluated using five-fold
cross-validation, with 10% of the data kept out as a validation set. For each point-cloud/language pair in test set of each
fold, each algorithm chooses one trajectory from the training
set which best suits this pair. Since our focus is on testing
ability to reason about different modalities and transfer
trajectory, segmented parts are provided as input. To evaluate
transferred trajectories, the dataset contains a separate expert
demonstration for each point-cloud/language pair, which is
not used in the training phase [1]. Every transferred trajectory
is evaluated against these expert demonstrations.
Metrics. For evaluation of trajectories, we use dynamic time
warping for manipulation trajectories (DTW-MT) [1], which
non-linearly warps two trajectories of different lengths while
preserving weak ordering of matched trajectory waypoints.
Since its values are not intuitive, [1] also reports the percentage of transferred trajectories that have a DTW-MT value of
less than 10 from the ground-truth trajectory, which indicates
that it will most likely correctly manipulate according to the
given instruction according to an expert survey.
We report three metrics: DTW-MT per manual, DTW-MT
per instruction, and Accuracy (DTW-MT < 10) per instruction. Instruction here refers to every point-cloud/language
pair, and manual refers to list of instructions which comprises
a set of sequential tasks, which we average over.
Baselines. We compare our model against several baselines:
1) Chance: Trajectories are randomly transferred.
2) Sung et al. [1]: State-of-the-art result on this dataset
100
90
Accuracy (%)
TABLE I
R ESULT ON ROBOBARISTA DATASET WITH 5-fold cross-validation.
ROWS LIST MODELS WE TESTED INCLUDING OUR MODEL AND
BASELINES . E ACH COLUMN SHOWS A DIFFERENT METRIC USED TO
EVALUATE THE MODELS . F OR THE DTW-MT METRIC , LOWER VALUES
ARE BETTER . F OR ACCURACY, HIGHER IS BETTER .
80
70
60
50
Sung et al. [1]
LMNN [10]-like cost func.
Our Model
40
30
20
5
10 Threshold 15
20
25
Fig. 4. Accuracy-threshold graph showing results of varying thresholds
on DTW-MT measures. Our algorithm consistently outperforms the previous
approach [1] and an LMNN-like cost function [10].
that trains a neural network to predict how likely each known
trajectory matches a given point-cloud and language.
We also report several baselines from this work which rely
on more traditional approaches such as classifying pointclouds into labels like ‘handle’ and ‘lever’ (object part
classifier), or hand-designing features for multi-modal data
(LSSVM + kinematic structure, task similarity + weighting).
3) LMNN [10]-like cost function: For both fine-tuning
and pre-training, we define the cost function without loss
augmentation. Similar to LMNN [10], we give a finite margin
between similarities. For example, as cost function for h3 :
|1 + sim(ΦP,L (pi , li ), ΦT (τ 0 )) − sim(ΦP,L (pi , li ), ΦT (τi ))|+
4) Our Model without Pretraining: Our full model finetuned without any pre-training of lower layers.
5) Our Model with SDA: Instead of pre-training h2,pl and
2,τ
h
as defined in Secs. IV-A and IV-B, we pre-train each
layer as stacked de-noising autoencoders [30], [31].
6) Our Model without Multiple Segmentations: Our model
trained only with expert segmentations, without taking utilizing all candidate segmentations in auto-encoders and multiple correct segmentations of the same part during training.
A. Results
We present the results of our algorithm and the baseline
approaches in Table I. Additionally, Fig. 4 shows accuracies
obtained by varying the threshold on the DTW-MT measure.
The state-of-the-art result [1] on this dataset has a DTWMT measure of 13.0 per manual and a DTW-MT measure
and accuracy of 12.2 and 60.0% per instruction. Our full
model based on joint embedding of multimodal data achieved
10.3, 9.9, and 68.4%, respectively. This means that when a
robot encounters a new object it has never seen before, our
model gives a trajectory which would correctly manipulate
it according to a given instruction approximately 68.4%
of the time. From Fig. 4, we can see that our algorithm
consistently outperforms both prior work and an LMNN-like
cost function for all thresholds on the DTW-MT metric.
What does our learned deep embedding space represent? Fig. 6 shows a visualization of the top layer h3 ,
the joint embedding space. This visualization is created by
projecting all training data (point-cloud/language pairs and
trajectories) of one of the cross-validation folds to h3 , then
embedding them to 2-dimensional space using t-SNE [36].
One interesting result is that our system was able to
naturally learn that “nozzle” and “spout” are effectively
Fig. 5. Robotic Experiments: We test our algorithm on a PR2 robot with
three different novel objects — coffee dispenser handle, beverage dispenser
lever, and door handle.
synonyms for purposes of manipulation. It clustered these
together in the upper-left of Fig. 6 based solely on the fact
that both are associated with similar point-cloud shapes and
manipulation trajectories.
In addition to the aforementioned cluster, we see several
other logical clusters. Importantly, we can see that our embedding maps vertical and horizontal rotation operations to
very different regions of the space — roughly 2 o’clock and 8
o’clock in Fig. 6, respectively. Despite the fact that these have
nearly identical language instructions, our algorithm learns
to map them differently based on their point-clouds, mapping
nearby the appropriate manipulation trajectories.
Should the cost function be loss-augmented? When
we change the cost function for pre-training h2 and finetuning h3 to use a constant margin of 1 between relevant
Ti,S and irrelvant Ti,D demonstrations [10], performance
drops to 55.5%. This loss-augmentation is also visible in
our embedding space. Notice the purple cluster around the
12 o’clock region of Fig. 6, and the right portion of the red
cluster in the 11 o’clock region. The purple cluster represents
tasks and demonstrations related to pushing a bar, and the
lower part of the red cluster represents the task of holding a
cup below the nozzle. Although the motion required for one
task would not be replaceable by the other, the motions and
shapes are very similar, especially compared to most other
motions e.g. turning a horizontal knob.
Is pre-embedding important? As seen in Table I, without any pre-training our model gives an accuracy of only
54.2%. Pre-training the lower layers with the conventional
stacked de-noising auto-encoder (SDA) algorithm [30], [31]
increases performance to 62.6%, still significantly underperforming our pre-training algorithm, at 68.4%. This shows
that our metric embedding pre-training approach provides a
better initialization for an embedding space than SDA.
Can automatically segmented object parts be manipulated? From Table II, we see that our segmentation approach
was able to find a good segmentation for the object parts in
question in 50 of 60 robotic trials (Sec. VI-B), or 83.3% of
the time. Most failures occurred for the beverage dispenser,
which had a small lever that was difficult to segment.
When our full DME model utilizes two variations of same
part and uses all candidates as a training data for the autoencoder, our model performs at 68.4% compared to 65.1%
which only used expert segmentations.
Does embedding improve efficiency? While [1] has
749, 638 parameters to be learned, our model only has
616, 175 (and still gives better performance.)
The previous model had to compute joint point-
TABLE II
R ESULTS OF 60 EXPERIMENTS ON A PR2 ROBOT RUNNING END - TO - END
EXPERIMENTS AUTONOMOUSLY ON THREE DIFFERENT OBJECTS .
Success Rate
Dispenser Beverage
Door
of Each Step
Handle
Lever
Handle
Avg.
1) Segmentation
90.0%
65.0%
95.0%
83.3%
2) DME Traj. Inference
94.4%
100.0%
78.9%
91.1%
3) Execution of Traj.
82.4%
76.9%
100.0% 86.4%
cloud/language/trajectory features for all combinations of
the current point-cloud/language pair with each candidate
trajectory (i.e. all trajectories in the training set) to infer
an optimal trajectory. This is inefficient and does not scale
well with the number of training datapoints. However, our
model pre-computes the projection of all trajectories into h3 .
Inference in our model then requires only projecting the new
point-cloud/language combination to h3 once and finding the
trajectory with maximal similarity in this embedding.
In practice, this results in a significant improvement in
efficiency, decreasing the average time to infer a trajectory
from 2.3206ms to 0.0135ms, a speed-up of about 171x. Time
was measured on the same hardware, with a GPU (GeForce
GTX Titan X), using Theano [35]. We measured inference
times 10000 times for first test fold, which has a pool of 962
trajectories. Times to preprocess the data and load into GPU
memory were not included in this measurement.
B. Robotic Experiments
To test our framework, we performed 60 experiments on
a PR2 robot in three different scenes shown in Fig. 5. We
presented the robot with the object placed within reach from
different starting locations along with a language instruction.
Table II shows results of our robotic experiments. It
was able to successfully complete the task end-to-end autonomously 39 times. Segmentation was not as reliable for
the beverage dispenser which has a small lever. However,
when segmentation was successful, our embedding algorithm
was able to provide a correct trajectory with an accuracy
of 91.1%. Our PR2 was then able to correctly follow these
trajectories with a few occasional slips due to the relatively
large size of its hand compared to the objects.
Video of robotic experiments are available at this website:
http://www.robobarista.org/dme/
VII. C ONCLUSION
In this work, we introduce an algorithm that learns a
semantically meaningful common embedding space for three
modalities — point-cloud, natural language and trajectory.
Using a loss-augmented cost function, we learn to embed in a
joint space such that similarity of any two points in the space
reflects how relevant to each other. As one of application,
we test our algorithm on the problem of manipulating novel
objects. We empirically show on a large dataset that our
embedding-based approach significantly improve accuracy,
despite having less number of learned parameters and being
much more computationally efficient (about 171x faster) than
the state-of-the-art result. We also show via series of robotic
experiments that our segmentation algorithm and embedding
algorithm allows robot to autonomously perform the task.
Acknowledgment. This work was supported by Microsoft
Faculty Fellowship and NSF Career Award to Saxena.
Hold the milk pitcher
under the froth
wand and submerge
the nozzle of the
wand into the milk.
Push on
the bar to
fill the cup
with ice.
Rotate the
speaker
knob
clockwise
until it clicks.
Hold the
cup below
the nozzle
on the left.
Turn handle
counterclock
wise to mix
in hot water.
Hold the
bowl below
the Colossal
Crunch
nozzle.
Pull down on
the right
dispenser's
lever to fill
the cup.
To unlock the
door, rotate
the lock
clockwise.
Rotate the
Heart to
Heart knob
clockwise to
dispense.
Push the
button to
fill the cup.
Press the
button to
turn the
range hood
on.
Fig. 6.
Learned Deep Point-cloud/Language/Trajectory Embedding Space: Joint embedding space h3 after the network is fully fine-tuned,
visualized in 2d using t-SNE [36] . Inverted triangles represent projected point-cloud/language pairs, circles represent projected trajectories. The occupancy
grid representation (Sec. V-B) of object part point-clouds is shown in green in blue grids. For presentation purpose, ‘neighbor’ cells are not shown. The
legend at the bottom right shows classifications of object parts by an expert, collected for the purpose of building a baseline. As shown by result of this
baseline (object part classifier in Table I), these labels do not necessarily correlate well with the actual manipulation motion. Thus, full separation according
to the labels defined in the legend is not optimal and will not occur. These figures are best viewed in color.
R EFERENCES
[1] J. Sung, S. H. Jin, and A. Saxena, “Robobarista: Object part-based
transfer of manipulation trajectories from crowd-sourcing in 3d pointclouds,” in International Symposium on Robotics Research, 2015.
[2] G. Erdogan, I. Yildirim, and R. A. Jacobs, “Transfer of object shape
knowledge across visual and haptic modalities,” in Proceedings of the
36th Annual Conference of the Cognitive Science Society, 2014.
[3] M. Muja and D. G. Lowe, “Scalable nearest neighbor algorithms for
high dimensional data,” PAMI, 2014.
[4] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in NIPS, 2012.
[5] R. Socher, J. Pennington, E. Huang, A. Ng, and C. Manning, “Semisupervised recursive autoencoders for predicting sentiment distributions,” in EMNLP, 2011.
[6] R. Hadsell, A. Erkan, P. Sermanet, M. Scoffier, U. Muller, and
Y. LeCun, “Deep belief net learning in a long-range vision system
for autonomous off-road driving,” in IROS, 2008.
[7] N. Srivastava and R. R. Salakhutdinov, “Multimodal learning with
deep boltzmann machines,” in NIPS, 2012.
[8] G. Hinton and R. Salakhutdinov, “Reducing the dimensionality of data
with neural networks,” Science, vol. 313, no. 5786, pp. 504–507, 2006.
[9] I. Goodfellow, Q. Le, A. Saxe, H. Lee, and A. Y. Ng, “Measuring
invariances in deep networks,” in NIPS, 2009.
[10] K. Q. Weinberger, J. Blitzer, and L. K. Saul, “Distance metric learning
for large margin nearest neighbor classification,” in NIPS, 2005.
[11] J. Weston, S. Bengio, and N. Usunier, “Wsabie: Scaling up to large
vocabulary image annotation,” in IJCAI, 2011.
[12] J. Moore, S. Chen, T. Joachims, and D. Turnbull, “Learning to embed
songs and tags for playlist prediction,” in ISMIR, 2012.
[13] J. Ngiam, A. Khosla, M. Kim, J. Nam, H. Lee, and A. Y. Ng,
“Multimodal deep learning,” in ICML, 2011.
[14] K. Sohn, W. Shang, and H. Lee, “Improved multimodal deep learning
with variation of information,” in NIPS, 2014.
[15] T. Mikolov, Q. V. Le, and I. Sutskever, “Exploiting similarities among
languages for machine translation.” CoRR, 2013.
[16] J. Hu, J. Lu, and Y.-P. Tan, “Discriminative deep metric learning for
face verification in the wild,” in CVPR, 2014.
[17] S. Miller, J. Van Den Berg, M. Fritz, T. Darrell, K. Goldberg, and
P. Abbeel, “A geometric approach to robotic laundry folding,” IJRR,
2012.
[18] M. Bollini, J. Barry, and D. Rus, “Bakebot: Baking cookies with the
pr2,” in IROS PR2 Workshop, 2011.
[19] M. Koval, N. Pollard , and S. Srinivasa, “Pre- and post-contact policy
decomposition for planar contact manipulation under uncertainty,”
International Journal of Robotics Research, August 2015.
[20] J. Sung, B. Selman, and A. Saxena, “Synthesizing manipulation
sequences for under-specified tasks using unrolled markov random
fields,” in IROS, 2014.
[21] D. Misra, J. Sung, K. Lee, and A. Saxena, “Tell me dave: Contextsensitive grounding of natural language to mobile manipulation instructions,” in RSS, 2014.
[22] D. Kappler, P. Pastor, M. Kalakrishnan, M. Wuthrich, and S. Schaal,
“Data-driven online decision making for autonomous manipulation,”
in R:SS, 2015.
[23] O. Kroemer, E. Ugur, E. Oztop, and J. Peters, “A kernel-based
approach to direct action perception,” in ICRA, 2012.
[24] D. Kappler, J. Bohg, and S. Schaal, “Leveraging big data for grasp
planning,” in ICRA, 2015.
[25] I. Lenz, H. Lee, and A. Saxena, “Deep learning for detecting robotic
grasps,” RSS, 2013.
[26] S. Levine, N. Wagener, and P. Abbeel, “Learning contact-rich manipulation skills with guided policy search,” ICRA, 2015.
[27] I. Lenz, R. Knepper, and A. Saxena, “Deepmpc: Learning deep latent
features for model predictive control,” in RSS, 2015.
[28] I. Tsochantaridis, T. Joachims, T. Hofmann, Y. Altun, and Y. Singer,
“Large margin methods for structured and interdependent output
variables.” JMLR, vol. 6, no. 9, 2005.
[29] M. D. Zeiler, “Adadelta: An adaptive learning rate method,” arXiv
preprint arXiv:1212.5701, 2012.
[30] P. Vincent, H. Larochelle, Y. Bengio, et al., “Extracting and composing
robust features with denoising autoencoders,” in ICML, 2008.
[31] M. D. Zeiler, M. Ranzato, R. Monga, et al., “On rectified linear units
for speech processing,” in ICASSP, 2013.
[32] Y. Ioannou, B. Taati, R. Harrap, et al., “Difference of normals as a
multi-scale operator in unorganized point clouds,” in 3DIMPVT, 2012.
[33] R. B. Rusu, “Semantic 3d object maps for everyday manipulation in
human living environments,” KI-Künstliche Intelligenz, 2010.
[34] M. Quigley, K. Conley, B. Gerkey, et al., “Ros: an open-source robot
operating system,” in ICRA workshop on open source software, 2009.
[35] F. Bastien, P. Lamblin, R. Pascanu, et al., “Theano: new features
and speed improvements,” Deep Learning and Unsupervised Feature
Learning NIPS 2012 Workshop, 2012.
[36] L. Van der Maaten and G. Hinton, “Visualizing data using t-sne,”
JMLR, vol. 9, no. 2579-2605, p. 85, 2008.
| 2cs.AI
|
Non-parametric threshold estimation for classical risk process
perturbed by diffusion
arXiv:1606.06459v1 [math.ST] 21 Jun 2016
Chunhao Cai, Junyi Guo and Honglong You
School of mathematical science
Nankai University
[email protected]
Abstract
In this paper,we consider a macro approximation of the flow of a risk reserve, which is first introduced by [16]. The process is observed at discrete time points. Because we cannot directly
observe each jump time and size then we will make use of a technique for identifying the times
when jumps larger than a suitably defined threshold occurred. We estimate the jump size and
survival probability of our risk process from discrete observations.
Keywords: survival estimation, Integrated Squared Error, goodness-of-fit test
1. Introduction
In the field of financial and insurance mathematics, the following model is the most commonly
used:
Nt
X
γi ,
(1)
Xt = x + ct + σWt −
i=1
where x is the initial value and c > 0 is a premium rate, σ > 0 represents the diffusion volatility,
Nt is a Poisson process with rate λ counting the number of jumps up to time t > 0, γ1 , γ2 , ...
are independent and identically distributed (i.i.d.) positive random variables with distribution F ,
PNt
i=1 γi represents the aggregate jumps up to time t, Wt is a Wiener process independent of Nt
and γi . In order to avoid the almost sure ruin of (1), we assume that the premium
R ∞rate can be
written as c = (1 + θ0 )λµ where the premium loading factor θ0 is positive and µ = 0 uF (du).
This model was first introduced by [8] and further studied by many authors during the last few
years ([6, 7, 12, 9]). Many results on ruin probability and other ruin problems have been obtained by
the works mentioned above. If we define the survival probability of (1) as Φ(x) then the problem
is that, apart from some special cases, a general expression for Φ(x) does not exist. So in this
paper we will construct an estimator of Φ(x) without specifying any parametric model on F , the
distribution function of γi .
In our model we will assume that σ and λ are unknown parameters. In the classical Poisson
risk model, the estimation of ruin probability has been considered by many authors ([5, 2, 11, 14,
15, 10, 13]). Not the same observation rules as in the work of [18], we only consider a discrete
record of n + 1 observations {Xtn0 , Xtn1 , ..., Xtnn−1 , Xtnn } where tni = ihn , Tn = tnn and hn > 0. This
observation is also important since the real data is always obtained at discrete time points.
Preprint submitted to Insurance: Mathematics and Economics
April 10, 2018
When we want to estimate the distribution of γi , we have to separate the contributions of the
diffusion part with respect to the jump of the poisson process so the threshold estimation which
has been introduced in ([3, 16, 17, 20, 19]) can be used. We only accept there is a jump for the
Poisson process between the interval (tni−1 , tni ] if and only if the increment ∆i X = Xtni − Xtni−1 has
a too big absolute value. We define this threshold function ϑ(hn ) which goes to zero when hn → 0.
Recall that the estimator of the empirical distribution function of γi can be written by
F̂Nt (u) =
Nt
1 X
I{γi ≤u} ,
Nt i=1
then we can try to write the estimator as
F̂n (u) = Pn
n
X
1
i=1 I{|∆i X|>ϑ(hn )} i=1
I{I{|∆i X|>ϑ(hn )} ∆i X≤u} .
(2)
In this paper we will try to study the properties of the estimator defined in (2), such as the
strong consistency, asymptotic normality. We also obtain the weak consistency in a sense of the
integrated squared error (ISE) of the estimator of survival probability such as in [13].
The rest of the paper is organized as follows, in section 2 we will give some notations and
construct the estimators for all the unknown parameters and the distribution of γi . In section 3,
we will give the asymptotic properties of the estimators given in section 2. In section 4, we obtain
the weak consistency in a sense of ISE of the estimator of survival probability of our risk process.
2. Preliminaries
2.1. General Notation
Throughout the paper, we use the primary notations and assumptions.
P
D
• symbols −
→ and −→ stand for the convergence in probability, and in law, respectively;
• LF (respectively, lF ) is the Laplace(respectively, Stieltjes) transform for a function F : for
s>0
Z ∞
Z ∞
e−su F (u)du;
LF (s) =
e−su F (du)
lF (s) =
0
0
where in LF , F is for every function and in lF , F is the distribution function.
RK
1
• kf kK := ( 0 |f (t)|2 dt) 2 for a function f . In particular, kf k = kf k∞ . We say that
f ∈ L2 (0, K) if kf kK < ∞ and f ∈ L2 (0, ∞) if kf k < ∞, respectively.
n
• For a stochastic sequence Xn , we denote by Xn = OP (Rn ) if X
Rn is bounded in probability
P
n
→0
and Xn = oP (Rn ) if X
Rn −
• E is a compact subset of (0, ∞).
• kn > 0 is a real-value sequence.
We make the following assumptions.
A. Let σ, γ satisfy: σ < Q, 0 < Γ ≤ γ < Q, with Q > 0.
2
B. For δ ∈ (0, 21 ), p, q > 1 with p−1 + q −1 = 1 and s ∈ E,
1. αδn (kn2 ) → 0 as n → ∞;
2.
√
Tn (αδn (kn ) + ωn (s, p, q) + Γ(s, ϕn )) → 0 as n → ∞,
where
1
αδn (kn ) = kn hn2
and
−δ
,
Γ(s, ϕn ) = λ
Z
0
1+ρ
ωn (s, p, ρ) = hn q
−1
(λ
Z
∞
(Ms − ϕn ◦ Ms )(x)F (dx)
∞
0
1
|Ms (x)|p F (dx)) p .
2.2. Estimator of unknown parameters and lF
Now we will try to construct the estimator of σ 2 , λ, ρ = λµ
c and lF . Because we only consider
the case where observations are discrete, that is to say the jumps are not observable. To overcome
the problem, [3], [20] and [17] have used a jump-discriminant filter of the form
Cin (ϑ(hn )) = {ω ∈ Ω; |∆i X| > ϑ(hn )}
to discriminate between jumps and large Brownian shocks in an interval (tni−1 , tni ] in the cases of
jump-diffusions. They judged that no jump had occurred if |∆i X| ≤ ϑ(hn ) and that a single jump
had occurred if |∆i X| > ϑ(hn ) by choosing the threshold ϑ(hn ) suitably.
According to the work of [20], we can define the following estimators of σ 2 , λ, ρ and lF :
Pn
Pn
2
I{|∆i X|>ϑ(hn )}
2
i=1 |∆i X − chn | I{|∆i X|≤ϑ(hn )}
f
Pn
, λn = i=1
,
σ
fn =
hn i=1 I{|∆i X|≤ϑ(hn )}
Tn
ρf
n =
1
c
Pn
i=1 (|∆i X|)I{|∆i X|>ϑ(hn )}
Tn
,
leF n (s) =
1
where ϑ(hn ) = Lhω
n , L > 0, ω ∈ (0, 2 ) and s ∈ E.
Pn
−s|∆i X|
)I{|∆i X|>ϑ(hn )}
i=1 (e
Pn
I
i=1 {|∆i X|>ϑ(hn )}
Remark 2.1. Here, we choose ϑ(hn ) = Lhω
∈ (0, 12 ). In [3], the author
n for a constant L > 0 and ω q
proposed the jump-discriminant threshold function such as ϑ(hn ) = L hn log h1n for a constant
L > 0. In [20], the author proposed ϑ(hn ) = Lhω
(0, 12 ). First of all, it
n for a constant L > 0 and ω ∈ √
is obvious that their threshold functions satisfy the conditions: ϑ(hn ) → 0 and hn ϑ(hn )−1 → 0. If
ϑ(hn ) does not converge to zero then one cannot detect
√ small jumps even when n → ∞. Moreover,
if ϑ(hn ) converges to zero faster than the order of hn then the filter Cin (ϑ(hn )) would possibly
misjudge
the Brownian noise as a small jump sine the variation of the Brownian motion is of order
q
1
hn log hn . Actually, the threshold function ϑ(hn ) in [20] and [3] are all suited for our aim. For
the convenience of proof on our results, we choose ϑ(hn ) in [20]. For a detailed account on the
threshold functions we refer to [17].
3
3. Properties of the Estimators
3.1. Asymptotic Normality of Fn (u)
First we will try to get the asymptotic normality of F̂n (u). Before that we will define F̄ = 1 − F
where F is the distribution of γi and F̄ˆn (u) = 1 − F̂n (u). Then
n
X
1
F̄ˆn (u) = Pn
i=1 I{|∆i X|>ϑ(hn )} i=1
I{∆i XI{|∆i X|>ϑ(hn )} >u} ,
then we have the following result:
Theorem 3.1. Suppose that nhn → ∞, hn → 0 as n → ∞ and the condition A in section 2 is
satisfied, then
p
F (u)(1 − F (u))
D
).
(3)
Tn F̄ˆn (u) − F (u) −→ N(0,
λ
then we change to F̄n (u) and F (u) that is
p
F (u)(1 − F (u))
D
Tn F̂n (u) − F (u) −→ N(0,
)
λ
(4)
To prove this theorem, we need the following lemma:
Proposition 3.2. Following from the condition of Theorem 3.1, for any ǫ > 0,
lim P |∆i XI{|∆i X|>ϑ(hn )} − γτ (i) I{∆i N ≥1} | > ǫ = 0,
n→∞
(5)
where γτ (i) is the eventual jump in time interval (tni−1 , tni ].
Proof.
Due to P {|σ
The therm
R tni
tn
i−1
P |∆i XI{|∆i X|>ϑ(hn )} − γτ (i) I{∆i N ≥1} | > ǫ
ǫ
≤ P |∆i XI{|∆i X|>ϑ(hn ),∆i N =0} | >
3
ǫ
+P |∆i XI{|∆i X|>ϑ(hn )} − γτ (i) |I{∆i N =1} >
3
ǫ
.
+P |∆i XI{|∆i X|>ϑ(hn )} − γτ (i) |I{∆i N ≥2} >
3
−c2
dWs | ≥ c} ≤ 2e 2Q2 hn , the first term in (6) is dominated by
!
Z tni
ǫ
ǫ
P chn >
dWs | >
+ P |σ
6
6
tn
i−1
2
−ǫ
ǫ
+ 2e 72Q2 hn .
≤ P chn >
6
|∆i XI{|∆i X|>ϑ(hn )} − γτ (i) |I{∆i N =1}
4
(6)
is equal to
|chn + σ
Z
≤ |chn + σ
tn
i
tn
i−1
Z
dWs + γτ (i) |I{|∆i X|>ϑ(hn )} − γτ (i) I{∆i N =1}
tn
i
tn
i−1
dWs | + |γτ (i) |I{|∆i X|≤ϑ(hn ),∆i N =1} .
Thus, (6) is dominated by
−ǫ2
ǫ
2P chn >
+ 4e 72Q2 hn + P (∆i N = 1) + P (∆i N ≥ 2) .
6
(7)
Therefor, (6) tends to zero as n → ∞.
Now we will prove the Theorem 3.1. Because
"
#
n
X
p
1
I{∆i XI{|∆i X|>ϑ(hn )} >u} − F (u)
Tn Pn
i=1 I{|∆i X|>ϑ(hn )} i=1
Pn
i=1 I{∆i XI{|∆ X|>ϑ(hn )} >u} −F (u)I{|∆i X|>ϑ(hn )}
i
p
Tn
Pn
= Tn
I{|∆ X|>ϑ(h )}
i=1
i
n
Tn
=
J
.
G
where
J=
Pn
i=1
I{∆i XI{|∆i X|>ϑ(hn )} >u} − F (u)I{|∆i X|>ϑ(hn )}
√
Tn
and
G=
Pn
i=1 I{|∆i X|>ϑ(hn )}
Tn
First, we will calculate the limit of the expectation of J:
lim E[J]
n→∞
i
n h
E(I{∆i XI{|∆i X|>ϑ(hn )} >u} − F (u)I{|∆i X|>ϑ(hn )} )
= lim √
n→∞
Tn
n
= lim √ E(H)
n→∞
Tn
n
= lim √
P (∆i XI{|∆i X|>ϑ(hn )} > u) − F (u)P (|∆i X| > ϑ(hn )) ,
n→∞
Tn
where, H = I{∆i XI{|∆i X|>ϑ(hn )} >u} − F (u)I{|∆i X|>ϑ(hn )} .
5
(8)
By Proposition 3.2 and [3], we have
lim P (∆i XI{|∆i X|>ϑ(hn )} > u)
n→∞
= P (γτ (i) I{∆i N ≥1} > u)
= P (γτ (i) I{∆i N ≥1} > u|∆i N ≥ 1)P (∆i N ≥ 1)
= P (γτ (i) > u|∆i N ≥ 1)P (∆i N ≥ 1)
= P (γτ (i) > u)P (∆i N ≥ 1)
= F (u)(λhn + o(hn ))
(9)
and
lim P (|∆i X| > ϑ(hn ))
n→∞
= P (∆i N ≥ 1)
= λhn + o(hn ).
Therefor,
lim E[J] = 0.
n→∞
Now, we will calculate the limit of the variation of J:
lim V ar[J]
n→∞
i
n h
= lim √
V ar(I{∆i XI{|∆i X|>ϑ(hn )} >u} − F (u)I{|∆i X|>ϑ(hn )} )
n→∞
Tn
n
V ar(H)
= lim
n→∞ Tn
n
= lim
[E(H)2 − E 2 (H)]
n→∞ Tn
n
[E(H)2 ].
= lim
n→∞ Tn
The term E(H)2 is equal to
2
E(I{∆
i XI{|∆
i X|>ϑ(hn )}
>u}
+
2
F (u)2 I{|∆
i X|>ϑ(hn )}
−
2F (u)I{∆i XI{|∆i X|>ϑ(hn )} >u} I{|∆i X|>ϑ(hn )} ).
Due to I 2 = I and
I{∆i XI{|∆i X|>ϑ(hn )} >u} I{|∆i X|>ϑ(hn )} = I{∆i XI{|∆i X|>ϑ(hn )} >u} , E(H)2 is equal to
E(I{∆i XI{|∆i X|>ϑ(hn )} >u}
+
F (u)2 I{|∆i X|>ϑ(hn )}
−
2F (u)I{∆i XI{|∆i X|>ϑ(hn )} >u} ).
By (11) and (10),
lim V ar[J]
n→∞
=
=
n
[λhn F (u)(1 − F (u)) + o(hn )]
Tn
λF (u)(1 − F (u)).
lim
n→∞
6
(10)
Applying the central limit theorem, we have
D
J −→ N(0, λF (u)(1 − F (u)))
as n → ∞
P
From the Lemma 3.3, we have G −
→ λ and then by the Slutsky’s theorem, we have
F (u)(1 − F (u))
J
→D N(0,
)
G
λ
3.2. Study of the estimatiors in section 2.2
As is well known in the inference for discretely observed diffusions that:
P
2
f2 −
σ
n →σ ,
In addition, if nh2n → 0,
n → ∞.
√ f2
D
n(σn − σ 2 ) −→ N (0, 2σ 4 )
as n → ∞. The proof is from Theorem 3.1 in [16].
fn , ρf
f
In order to find the asymptotic nomality of λ
n and lFn , We will first define a truncation
function ϕn (x) satisfying the following conditions:
1. |ϕn (x)| ≤ βn a.e. for a real-valued sequence βn such that βn ↑ ∞ as n → ∞.
2. ϕn (x) → x a.e. as n → ∞.
Through this ϕn we will redefine the estimator of λ, ρ and lF . Moreover, we consider the
estimator of the product of λlF . They are defined as
Pn
1 i=1 (ϕn ◦ Ms1 )(|∆i X|)I{|∆i X|>ϑ(hn )}
∗
,
=
ρf
n
c
Tn
eleF ∗ (s) =
λ
n
Pn
i=1 (ϕn
◦ Ms2 )(|∆i X|)I{|∆i X|>ϑ(hn )}
,
Tn
where Ms1 (x) = x, Ms2 (x) = e−sx ,
W
W
Ms (x) if (Ms (x) sups∈E (Ms (x))′s sups∈E (Ms (x))′x ) ≤ κn ,
ϕn ◦ Ms (x) =
0
Otherwise,
and κn > 0 is a real-value sequence, we have the following results:
Lemma 3.3. Supposes nhn → ∞, hn → 0 as n → ∞, then
and
∗ P
P
fn −
→ λ,
λ
→ ρ,
ρf
n −
∗
(11)
P
fn leF (s) − λlF (s)| −
→ 0.
sup |λ
n
(12)
{s|s∈E}
R∞
In addition, if the Condition B in section 2.1 is satisfied, 0 x2 F (dx) < ∞ and nh1+β
→ 0 for
n
some β ∈ (0, 1), we have
p
D
fn − λ) −
Tn (λ
→ N (0, λ),
(13)
7
Z
p
λ ∞ 2
D
∗
Tn (f
ρn − ρ) −→ N (0, 2
x F (dx))
c 0
and
Z
p
D
fn leF ∗ (s) − λlF (s) −
→
N
(0,
λ
Tn λ
n
∞
e−2sx F (dx)),
(14)
(15)
0
as n → ∞.
Proof. First from the theorem 3.2 and 3.4 in [16], it is easy to get (11) and (12). So we only need
to check (13),(14) and (15). We define
Pn
ϕn ◦ Ms (|∆i X|)I{|∆i X|>ϑ(hn )}
νn (s) = i=1
Tn
and
ν(s) = λ
Notice that
Z
∞
Ms (x)F (dx)
0
n
X
p
Tn (νn (s) − ν(s)) =
Yn,i
i=1
where
Yn,i
1
= √
Tn
Z
ϕn ◦ Ms (|∆i X|)I{|∆i X|ϑ(hn )} − hn λ
0
∞
Ms (x)F (dx)
(16)
we can apply the central limit theorem for {Yn,i }1≤i≤n . If the following conditions (17)-(19) are
satisfied, we can obtain (13)-(15) by [1]
n
X
i=1
n
X
i=1
P
|E[Yn,i ]| −
→ 0,
(17)
Z
(18)
P
→λ
E[Yn,i ]2 −
n
X
i=1
∞
0
Ms2 F (dx),
P
E[Yn,i ]4 −
→0
(19)
The details of the proofs of (17)-(19) may refer to Theorem [17].
4. Integrated Squared Error (ISE) of the Esitimator
4.1. Weak consistency in the ISE sense
We will present our important result that states a convergence in probability of the ISE of the
estimator.
First let us recall that let τ (x) be the time of ruin with the initial reserve x: τ (x) = inf{t >
0, Xt ≤ 0}. The survival probability Φ(x) is defined as follows:
Φ(x)
= P {τ (x) = ∞}.
8
As we know, a general expression for Φ(x) does not exist, but the corresponding Laplace transform
of Φ(x) can be obtained by [18], there we take ω ≡ 1 and δ = 0 which is the Laplace transform of
our model:
Z ∞
LΦ (s) =
e−sx Φ(x)dx
0
=
s+
σ2
2c s
2
1−
λµ
c
(20)
− λc (1 − lF (s))
^
with the estimators of the parameters in (20), we define L
Φ (s) as the estimator of LΦ (s):
^
L
Φ (s) =
s+
f
2
σ
n 2
2c s
1 − ρf
n
−
∗
en
λ
c (1
∗
− leF n (s))
.
(21)
In order to estimate the original functions Φ(x), we will apply a regularized inversion of the Laplace
transform proposed by [4] to (21). The regularized inversion is defined as follows, which is available
for any L2 functions.
2
2
Definition 4.1. Let m > 0 be a constant. The regularized Laplace inversion L−1
m : L → L is
given by
Z ∞Z ∞
1
1
−1
Lm g(t) = 2
Ψm (y)y − 2 e−tvy g(v)dvdy
π 0
0
for a function g ∈ L2 and t ∈ (0, ∞), where
Z am
cosh(πx)cos(x log y)dx
Ψm (y) =
0
and am = π
−1
cosh
−1
(πm) > 0.
Remark 4.2. It is well known that the norm L−1 is generally unbounded, which causes the illposedness of the Laplace inversion. However, L−1
m is bounded for each m > 0, in particular,
√
kL−1
πm
m k ≤ mkLk =
see [4], equation (3.7).
fΦ (s) ∈
By (20) and (21), it is obvious that LΦ (s) ∈
/ L2 (0, ∞) and L
/ L2 (0, ∞). The regularized
−1
laplace transform inversion Lm of Definition 4.1 does not apply at once. In order to ensure they
fΦ (s) will be slightly modified.
are in L2 (0, ∞), these functions LΦ (s) and L
For arbitrary fixed θ > 0, we define Φθ (x) = e−θx Φ(x). It is obvious that
^
L^
Φθ (s) = LΦ (s + θ).
LΦθ (s) = LΦ (s + θ),
Let us define an estimaor of Φ(x): for given numbers θ > 0 and m(n), we denote by
fn (x)
Φ
=
eθx Φ^
θ,m(n) (x),
(22)
−1 g
where Φ^
θ,m(n) (x) = Lm(n) LΦθ (s).
Now we shall present our important result which states a convergence in probability of the ISE
on compacts.
9
Theorem 4.3. Suppose the same assumption as in Lemma 3.3 and the net profit condition: c > λµ.
Moreover, suppose that there exist a q
constant K > 0 such that 0 ≤ Φ′ (x) = g(x) ≤ K < ∞. Then,
for numbers m(n) such that m(n) =
Tn
log Tn
as n → ∞ and for any constant B > 0, we have
e m(n) − Φk2B = OP ((log Tn )−1 )
kΦ
(n → ∞).
4.2. The proof of Theorem 4.3
To prove Theorem 4.3, we need the following lemma:
The Lemma 4.4, which is essentially obtained by the proof of Theorem 3.2 in [4], shows that
L−1
n can be a Laplace inversion asymptotically in ISE sense.
R∞
1
Lemma 4.4. Suppose that, for a function f ∈ L2 with the derivative f ′ , 0 [t(t 2 f (t))′ ]2 t−1 dt < ∞,
then
1
(n → ∞).
kLn−1 Lf − f k = O (log n)− 2
Now we will prove the Theorem 4.3.
Proof. By (22), we have
e m(n) − Φk2
kΦ
B
≤
≤
=
e θ,m(n) − Φθ k2
e2θB kΦ
B
−1
2
2
g
2e2θB {kL−1
m(n) LΦθ − Lm(n) LΦθ k + kΦθ,m(n) − Φθ k }
2e2θB [kI1 + I2 k] .
In order to deal with I2 , let us write Φ′θ = gθ and note that
Z ∞
√
1
[x( xΦθ (x))′ ]2 dx
x
Z0 ∞
√
1
1
=
[x( √ Φθ (x) + x xgθ (x))]2 dx
x
2 x
Z ∞
Z0 ∞
√
1
1
1
2 [x xgθ (x)]2 dx
≤
2 [x( √ Φθ (x)]2 +
x 2 x
x
0
Z ∞
Z0 ∞
1 2
Φ (x)dx + 2
x2 gθ2 (x)dx
=
2 θ
0
0
Z ∞
Z ∞
1 −2θx
e
dx + 2
x2 [g(x)e−θx − θΦ(x)e−θx ]2 dx
≤
2
0
0
Z ∞
Z ∞
1
≤
x2 g 2 (x)e−2θx dx + 4θ2
+4
Φ2 (x)x2 e−2θx dx
4θ
0
0
Z ∞
1
2
2
2 −2θx
≤
+ 4(K + θ )
x e
dx
4θ
0
< +∞.
Therefore, by the Lemma 4.4, we may conclude that
kΦθ,m(n) − Φθ k2 = O(
10
1
).
log m(n)
(23)
Next, we consider the formula I1 . By (20) and (21), we have
2
g
kL
Φθ − Lθ k
Z ∞
∗
1 − ρf
1−ρ 2
n
| ds
|
−
e
D(s
+ θ)
D(s + θ)
0
Z ∞
∗
1 − ρf
1−ρ
1−ρ
1−ρ 2
n
|
−
+
−
| ds
e
e
e
D(s
+ θ)
D(s
+
θ)
D(s
+
θ)
D(s
+
θ)
0
!2
!2
Z ∞
∗
e + θ) − D(s + θ))
(1 − ρ)(D(s
(f
ρn − ρ)
2
+2
ds.
e + θ)D(s + θ)
e + θ)
D(s
D(s
0
Z ∞ e
Z ∞
(D(s + θ) − D(s + θ))2
ρf
1
n ∗ −ρ 2
ds
2
ds
+
2
(
∗ 2
∗)
4
(s + θ)2
(s + θ) (1 − ρf
1 − ρf
n )
n
0
0
Z Tn e
(D(s + θ) − D(s + θ))2
ds
2
∗ 2
(s + θ)4 (1 − ρf
n )
0
Z ∞ e
(D(s + θ) − D(s + θ))2
+2
ds
∗ 2
(s + θ)4 (1 − ρf
n )
Tn
Z ∞
∗
ρf
1
n −ρ 2
ds
+2
(
∗)
(s
+
θ)2
1
−
ρ
f
n
0
≤
=
≤
≤
=
and
fn (s + θ) − D(s + θ)
D
1
(s + θ)2
2
fn ) + (λ
fn leF ∗ (s + θ) − λlF (s + θ))
(λ − λ
(f
σn − σ 2 ) +
=
n
2c
c
1 f e∗
− 12
= OP (Tn ) + (λn lF n (s + θ) − λlF (s + θ)).
c
−1
2
Note that we used here the fact that σ
fn − σ 2 = oP (Tn 2 ).
Now we consider the second term in (24). By (25), we have
Z
≤
≤
=
=
=
e + θ) − D(s + θ))2
(D(s
ds
∗ 2
(s + θ)4 (1 − ρf
n )
Tn
Z ∞
Z
fn leF ∗ (s + θ) − λlF (s + θ))2
2 ∞ (λ
1
n
−1
ds
ds + 2
OP (Tn )
∗ 2
4
c Tn
(s + θ)4 (1 − ρf
n )
Tn (s + θ)
2
Z ∞
Z
fn + λ2
1
4 ∞
λ
−1
OP (Tn )
ds + 2
∗ 2 ds
4
c Tn (s + θ)4 (1 − ρf
n )
Tn (s + θ)
Z ∞
Z ∞
1
1
OP (Tn−1 )
ds
+
O
(1)
ds
P
4
4
Tn (s + θ)
Tn (s + θ)
Z ∞
1
ds
OP (1)
4
Tn (s + θ)
1
OP ( 3 ).
Tn
∞
11
(24)
(25)
By Lemma 3.3, the first term of (24) is
Z T e
(D(s + θ) − D(s + θ))2
2
ds
∗ 2
(s + θ)4 (1 − ρf
n )
0
and the last term is
2
Z
∗
∞
(
0
Therefore,
= OP (Tn−1 )
1
ρf
n −ρ 2
ds
∗)
(s + θ)2
1 − ρf
n
=
OP (Tn−1 ).
m2 (n)
).
Tn
(26)
m2 (n)
1
) + Op (
).
Tn
log m(n)
(27)
2 g
2
kL−1
m(n) k kLΦθ − LΦθ k
= Op (
Combining (23) and (26), we have
e m(n) − Φk2 = Op (
kΦ
B
With an optimal m(n) =
−1
becomes OP ((log Tn )
).
q
Tn
log Tn
balanceing the the right hand two terms in (27), the order
Remark 4.5. The explicit integral expression
uθ Z ∞ Z ∞
−1
e m(n) (u) = e
g
Φ
e−usy L
Φθ (s)Ψm(n) (y)y 2 dsdy
2
π
0
0
R am(n)
where Ψm(n) (y) = 0
cosh(πx) cos(x log(y))dx and am(n) = π −1 cosh−1 (πm(n)) > 0 and m(n) =
q
Tn
log Tn .
5. Goodness-of-fit Test for γi
In theorem 3.1 we have proved the asymptotic properties of F̂ (u), but as we can see the variance
depends on the unknown parameter λ, when we want to do the problem of goodness-of-fit test for
the distribution of γi , we want to find the estimator with the distribution free just with the null
(
hypothese: H′ : F (x) = F0 x). To achieve this goal, we need the following convergence:
v
u n
uX
D
t
I{|∆i X|>ϑ(hn )} F̂n (u) − F (u) −→ N(0, F (u)(1 − F (u))
(28)
i=1
The proof is the same as in theorem 3.1. Now we construct the Cramér-von Mises Wn2 and
Kolmogorov-Smirnov Dn statistics are
Z ∞h
i2
2
F̂n (x) − F0 (x) dF0 (x), Dn = sup F̂n (x) − F0 (x)
Wn = n
x
−∞
From the equation (28), we can easily get the limit behavior of these statistics
Z 1
√
d
2 d
W0 (s)2 ds,
→
Wn −
→ sup |W0 (s)|
nDn −
0≤s≤1
0
where {W0 (s), 0 ≤ s ≤ 1} is a Brownian bridge, the same as the i.i.d case.
12
References
[1] A. N. Shiryaev Probability Springer-Verlag,1984
[2] C. Hipp. Estimators and bootstrap confidence intervals for ruin probabilities. Astin Bulletin,
1957–70, 1989.
[3] C. Mancini. Estimation of the characteristics of the jump of a general Poisson-diffusion model.
Scandinavian Actuarial Journal, 142–52, 2004.
[4] D. E. Chauveau, A. C. M. Vanrooij, and F. H. Ruymgaart. Regularized inversion of noisy
Laplace transforms. Advances in Applied Mathematics, 15(2):186-201, 1994.
[5] EW. Frees. Nonparametric estimation of the probability of ruin. Astin Bulletin, 16:81–90,
1986.
[6] F. Dufresne and H.U. Gerber. Risk theory for the compound Poisson process that is perturbed
by diffusion. Insurance:Mathematics and Economics, 10:51–59, 1991.
[7] H.J. Furrer and H. Schimidli. Exponential inequalities for ruin probabilities of risk process
perturbed by diffusion. Insurance:Mathematics and Economics, 15:23–36, 1994.
[8] H.U. Gerber. An extension of the renewal equation and its application in the collective theory
of risk. Scandinavian Actuarial Journal, 205–210, 1970.
[9] H. Schimidli. Cramer-Lundberg approximations for ruin probabilites of risk processes perturbed by diffusion. Insurance:Mathematics and Economics, 16:135–149, 1995.
[10] K. Politis. Semiparametric estimation for non-ruin probabilities. Scandinavian Actuarial
Journal, 2003(1):75–96, 2003.
[11] K. Croux and N. Veraverbeke. Non-parametric estimators for the probability of ruin. Insurance:
Mathematics and Economics, 9:127–130, 1990.
[12] N. Veraverbeke. Asymptotic estimates for the probability of ruin in a Poisson model with
diffusion. Insurance: Mathematics and Economics, 13:57–62, 1993.
[13] R. Mnatsakanov, L. L. Ruymgaart, and F. H. Ruymgaart. Nonparametric estimation of ruin
probabilities given a random sample of claims. Mathematical Methods of Statistics, 17(1):
35–43, 2008.
[14] SM. Pitts. Nonparametric estimation of compound distributions with applications in insurance.
Annals of the Institute of Statistical Mathematics, 46(3): 537–555, 1994.
[15] V. E. Bening and V. Y. Korolev. Nonparametric estimation of the ruin probability for generalized risk processes. Theory of Probability and its Applications, 47(1):1–16, 2002.
[16] Y. Shimizu. A new aspect of a risk process and its statistical inference. Insurance: Mathematics
and Economics, 44(1):70–77, 2009.
[17] Y. Shimizu. Functional estimation for Levy measures of semimartingales with Poissonian
jumps. Journal of Multivariate Analysis, 100(6):1073-1092, 2009.
13
[18] Y. Shimizu. Non-parametric estimation of the Gerber-Shiu function for the Wiener-Poisson
risk model. Scandinavian Actuarial Journal, 2012(1):56–69, 2012.
[19] Y. Shimizu. Denstity estimation of lévy measure for discrete observed diffusion processes with
jumps. J.Japan Statist. Soc., 36(1):37–62, 2006.
[20] Y. Shimizu and N. Yoshida. Estimation of parameters for diffusion process with jumps from
discrete observations. Statistical inference for stochastic processes, 9(3):227–277, 2006.
14
| 10math.ST
|
Crop Planning using Stochastic Visual Optimization
Gunjan Sehgal*, Bindu Gupta†, Kaushal Paneri, Karamjit Singh, Geetika Sharma, Gautam Shroff‡
arXiv:1710.09077v1 [cs.CV] 25 Oct 2017
TCS Research, India
A BSTRACT
As the world population increases and arable land decreases, it
becomes vital to improve the productivity of the agricultural land
available. Given the weather and soil properties, farmers need to
take critical decisions such as which seed variety to plant and in
what proportion, in order to maximize productivity. These decisions
are irreversible and any unusual behavior of external factors, such as
weather, can have catastrophic impact on the productivity of crop. A
variety which is highly desirable to a farmer might be unavailable or
in short supply, therefore, it is very critical to evaluate which variety
or varieties are more likely to be chosen by farmers from a growing
region in order to meet demand. In this paper, we present our visual
analytics tool, ViSeed, showcased on the data given in Syngenta 2016
crop data challenge1 . This tool helps to predict optimal soybean
seed variety or mix of varieties in appropriate proportions which is
more likely to be chosen by farmers from a growing region. It also
allows to analyse solutions generated from our approach and helps in
the decision making process by providing insightful visualizations.
Index Terms: K.6.1 [Management of Computing and Information
Systems]: Project and People Management—Life Cycle; K.7.m
[The Computing Profession]: Miscellaneous—Ethics
1 I NTRODUCTION
With increasing world population and decreasing arable land, optimizing the productivity of land is the need of the hour. To meet
the growing food demand [5] farmers need to take critical decisions
like on what(soil), which (variety), and how much(proportion) in
order to maximize the yield. However, a seed variety highly desired
by a farmer may be unavailable or in short supply. Therefore, it
is critical to evaluate which variety is more likely to be chosen by
farmers from a growing region in order to meet demand. As the
production varies due to differences in weather conditions [7] and
soil quality [3], there cannot be a single seed variety that will yield
high production across regions. Hence, predicting the appropriate
mix of proportions of different seed varieties is desirable.
While a standalone solution with predicted seed variety or mix of
varieties for a region is very helpful, but in the real world, where the
number of regions and varieties both can be large, it is very crucial
to have visualization based platform to analyze all possible solutions.
In this paper, we present a machine learning and optimization based
approach to predict seed variety or mix of varieties in appropriate
proportions which is more likely to be chosen by farmers from a
growing region. We also present our visual analytics tool to understand and analyse the solutions generated by our approach. This tool
helps in the decision making process and also provides insightful
visualizations.
We use Syngenta Crop Challenge 2016 dataset which requires
computation of an optimal combination of upto five soybean varieties
* e-mail: [email protected]
† e-mail:[email protected]
‡ e-mail:[email protected]
1 https://www.ideaconnection.com/syngenta-crop-challenge/
for the given region using historical sub-region data and experiment
data from a few experiment locations. Our approach to the challenge
comprises of (a) machine learning and optimization techniques to
compute the solution, as well as (b) a geo-spatial visual analytics
tool, ViSeed, to understand the given data and analyse the solutions.
As the production of soybean varies due to differences in weather
conditions and soil quality and there is no single variety that will
yield high production across the region, we have experimented with
two possible solutions to the challenge. The first, global solution,
required by the challenge and defined as a combination of upto
five varieties to be grown over the entire region. The second, a differentiated solution, consisting of individual optimal combinations
for each sub-region. We also compute a spatial cohesion score for
each subregion which measures the similarity of the solutions for
the sub-region with those of its neighbors. Finally, we show that
the differentiated solution performs marginally better and has lower
variance than the global solution.
Rest of the paper is organized as follows. In section 2, we present
the problem statement along with out approach to predict the optimal
set of varieties and in section 3, we present our visual analytics
tool ViSeed and showcase it for the Syngenta data challenge 2016.
In section 4, we present our results of prediction accuracies and
predicted soyabean varieties and conclude in section 5.
2
2.1
M ETHODOLOGY
AND THEORY
Problem Statement:
Let {v1 , v2 , ..., vn } be the n soybean varieties and let R be a region
with k sub-regions {R1 , R2 , ..., Rk }. Each of the sub-region Ri has
different soil and weather conditions associated with it. Our goal is
to find which soybean seed variety, or mix of up to five varieties in
appropriate proportions, will best meet the demands of farmers in
each sub-region Ri and for whole region R. In order to find such mix
of seed varieties, we predict the top five varieties for each sub-region
in terms of maximum yield and then find the optimal proportion of
these five varieties in order to maximize yield and minimize variance
in the yield.
Our approach to predict the mix of soybean varieties in appropriate proportion for each subregion Ri and for whole region R consists
of three steps: 1) Prediction of Weather and Soil attributes, 2) Yield
Prediction given the soybean variety, and weather and soil conditions, and 3) Yield Optimization. We also present our visual analytic
tool to understand and analyses the solutions generated by our approach. This tool helps in decision making process and also provide
insightful visualizations.
2.2
Weather and Soil Prediction
For each subregion Ri , let {wi1 (t), wi2 (t), ..., wiN (t)} be the weather
condition and {si1 (t), si2 (t), ..., siM (t)} be the soil condition attributes
at time t. Given these attributes from initial time t0 to the current
time t, we use Deep learning based approach LSTM [6] to predict the
value of these attributes for time t + 1. For each attribute separately,
say wiN , we prepare a sequence of its values from time t0 to t − 1, and
use it as an input to train LSTM with the target of predicting value of
wiN at time t and then we use the trained LSTM to predict the value
for time t + 1.LSTM based neural networks are competitive with the
traditional methods and are considered a good alternative to forecast
The term ∑l wl ∗ norm(Varl ) is called variability. We ran
above optimization for ten thresholds of variability, i.e., ten
values of τ from 0 to 1 with the step size of 0.1. For every
value of τ, we get the optimal solution. Therefore, for a subregion Ri , we get 10 solutions and out of these 10 solutions,
we choose the solution having maximum yield and minimum
variability and it is called default solution for sub-region Ri
Figure 1: LSTM architecture used to predict weather and soil attributes
general weather conditions [9] . Figure 1 shows an architecture used
for the prediction of weather and soil attributes.
2.3 Yield Prediction
Once we predict the weather and soil attributes for the time t + 1,
we use these attributes as a feature set to predict the yield in every
sub-region Ri for each soybean variety v j , where j = 1, 2, ..., n. We
divide the yield value into r equally sized bins by taking maximum
and minimum from historical data and we treat the prediction problem as a classification problem, where our goal is to predict the bin
value of yield. More formally, for each sub-region Ri , we compute n
probability distributions {pij,1 , pij,2 , ..., pij,r } of yield, one for every
soybean variety v j , where j = 1, 2, ..., n, using Random Forest Classifier (RFC) [2]. RFC is an ensemble learning method that operates
by constructing a multitude of decision trees at training time and
outputting the class that is the mode of the classes (classification)
of the individual trees. As an output, we get the count of each class
which represents the number of trees outputting the class. Further,
we convert these counts into probabilities by diving each count by
the sum of all counts.
2.4 Yield Optimization
Given n probability distributions of the yield yi for sub-region Ri ,
we use an optimization approach to obtain weighted combination
of varieties in order to maximize yield and minimize standard variation or variability. Steps to obtain combination of varieties in an
appropriate proportion are given as follows:
1. Given n probability distributions {pij,1 , pij,2 , ..., pij,r } of yield yi
for every soybean variety v j , where j = 1, 2, ..., n, we calculate
expected value and variance of each distribution, represented
by E j , and Var j respectively.
2. We choose top k distributions out of n having maximum score
calculated as:
score j = norm(E j ) + (1 − norm(Var j ))
(1)
where norm(E j ) and norm(Var j ) are the normalized values
between 0 to 1 of E j and Var j respectively.
3. We use an optimization technique with objective function of
maximizing yield using combination of upto five varities out
of chosen top k in an appropriate proportion. The objective
function and constraints of optimization are given as follows:
Objective function: max(∑ wl ∗ El )
(2)
l
Constraints:
∑ wl ∗ norm(Varl ) < τ , wl > .10, and ∑ wl = 1
l
l
(3)
Spatial Cohesion(SC) Score: For every solution obtained from
optimization approach, we calculate SC Score which is calculated
as follows: Let for a sub-region Ri , the optimized solution at
τ = τ1 contains five varieties vi1 , vi2 , ..., vi5 in some proportion. Let
near(Ri ) = {N1i , N2i , ..., Nlii } be the set of neighboring sub-regions of
Ri which are having maximum distance of m miles from the centroid
of it. For every vij in the solution of Ri , we calculate variety score as:
k=l
vars (vij ) =
∑k=1i wij,k
li
(4)
where wij,k is the proportion with which variety vij exist in the solution of neighboring sub-region Nki at τ = τ1 . Further, SC score of
sub-region Ri is calculated as the average of all variety scores in the
solution:
∑5j=1 vars (vij )
SCi =
(5)
5
3 V ISUAL ANALYTICS USING V I S EED
We now describe, ViSeed, our visual analytics tool to understand
the given data and analyse the solution generated by our analytics methodology.This tool differentiates our work from other related works [1] [8] as it lets the retailer explore varieties performing well in local as well as global areas. As agricultural yields
vary widely around the world due to climate and the mix of crops
grown [4],ViSeed lets the farmer explore the quality of soil and climatic variations region wise intuitively to take the planting decisions
accordingly.
The main screen of ViSeed is divided into two parts. The first part
displays a map of the United States over which various sub-region
attributes can be visualized, Figure 2 (i) A. The second (right hand
panel) part, B, contains a tabbed control panel, to switch between
various visualizations of the solution data.
Getting Started A data attribute such as precipitation or solar
radiation may be visualized by selecting it from the attribute menu
and a year from the timeline as shown in Figure 2 (i). Further, as
described in the previous section, we first compute the top k varieties
for each sub-region, based on high expected yield and low variance
and from them, an optimal solution with up to five varieties. As a
variety may occur in the optimal solution of multiple sub-regions,
we compute a distribution of weights for each variety across subregions and the expected value of this distribution. The varieties are
ranked in decreasing order of expected value which is an indicator
of prevalence of the variety across sub-regions and displayed in the
right panel when the visualization tool starts up, Figure 2 (i), C.
Thus, the user is provided a starting point to begin exploring the
possible solutions.
The histogram of weights for each variety across sub-regions,
using a colour map, is shown alongside its expected value. Selection
of a range on the histogram of a variety highlights those sub-regions
for which the variety has weight or proportion in the selected range.
Clicking on a variety name highlights those sub-regions on the
map for which the variety occurs in the optimal solution. Clicking
multiple varieties highlights the regions for all, thus allowing the
user to visualize cumulative prevalence of varieties.
Common Solution A user may explore various solutions, for the
entire region, by selecting up to five varieties from this list, 1 in
(i)
(ii)
Figure 2: ViSeed: (i) Main Data Exploration Screen (ii) Common Solution Exploration
Figure 2 (ii). On pressing the query button, 2, proportions of each
of the selected varieties are computed, 3, and the total yield for the
region is predicted, 4.
Differentiated Solution Each sub-region has a precomputed default solution from among its top k varieties. Clicking on a subregion in the map, Figure 3, 1, brings up the top k varieties for that
sub-region, 2, along with their weights in the optimal solution, count
of sub-regions in which the variety is in the top k and the predicted
yield distribution for the variety. The sub-regions for which a variety
is in the top k list can be seen by clicking on the red count bar.
For every solution, we show the Average Yield and Average Standard Deviation and Average offset in % of entire region R. The
standard deviation and offset of each sub-region is calculated as:
√
∑5l=1 wl ∗ Varl
Standard deviation (S.D) =
5
Offset % = (
S.D
) ∗ 100
ExpectedYield
(6)
(7)
where Varl is the variance of variety vl and wl is its proportion in
the solution.
Changing Variability Our tool also allows the user to analyse
solutions by playing with different variability thresholds. This can
be done by first clicking on the Variance tab. The user can set a
variability threshold by moving the slider, Figure 4, 1. On pressing
the query button, the list of varieties, with variability below the
chosen threshold is displayed, along with their histograms of weights
across sub-regions and expected values as before. A new optimal
solution is computed for each sub-region and its similarity with
the solutions of neighboring sub-regions is calculated as a spatial
cohesion score. This score is visualized on the map. The user may
also interact with this list of varieties and compute a global solution
for the entire region, as described earlier.
4
Q UANTITATIVE
Figure 3: Top k Varieties for a Sub-Region
RESULTS
In this section, we present our results of LSTM and RFC model used
for weather and yield prediction. We used keras, scikit-learn and
cvxpy respectively to implement LSTM, RFC and optimization.
Available Data: We were provided with the following two
datasets:
1. Experiment Dataset: It consists of 82000 experiments, in
583 sub-regions, between 2009 and 2015 using 174 varieties.
It has three weather and three soil condition attributes for every
sub-region.
Figure 4: Changing Solution Variability
Attributes
Temperature
Percipitation
SolrRadiation
Validation Set
0.69%
0.73%
2.6%
Test Set
0.78%
0.83%
2.8%
Table 1: N-RMSE of three weather attributes on validation and test
set using LSTM
been used as a feature set in RFC and discretized yield as a target
variable. N-RMSE of yield predicted on validation and test set using
RFC is 6.01% and 6.25% respectively. We use the trained RFC
model to predict the yield for 6490 sub-regions in Region dataset
with weather and soil attributes for year 2016 as input, predicted
using LSTM(as explained above).
(i)
4.1
(ii)
Figure 5: Common Solution: (i) Areas in which Varieties are optimal
(ii) Cumulative area in which varieties are optimal
2. Region Dataset: It consists of 6490 sub-regions with the given
three soil and three weather condition attributes from the year
2000 to 2015.
We predict three attributes of weather conditions, temperature,
precipitation, and solr radiation for every sub-region of Region
dataset. For an attributes say wiN in sub-region Ri , we prepare the
sequence of 14 values from year 2000 to 2014 as an input to train
LSTM with a target of predicting value of year 2015. We validate
and test the LSTM model by dividing Region dataset into training,
validation, and test data in the ratio of 70:15:15 respectively. Table 1,
shows the normalized root mean square error (N-RMSE) for all
three weather attributes on validation and test set. Here, N-RMSE is
defined for an attribute say, wiN as
s
2
∑ki=1 (act(wiN ) − pred(wiN ))
RMSE =
(8)
k
N-RMSE =
RMSE
∗ 100
(max(wiN ) − min(wiN ))
(9)
where act(wiN ) and pred(wi ) are the actual and predicted values,
max(wiN ) and min(wi ) are maximum and minimum values, of attribute wiN for sub-region Ri . The value of k is 6490 in our case.
N-RMSE in Table 1 indicating that the prediction of weather attributes using LSTM have less than 1% error for temperature and
precipitation and less than 3% error for Solr radiation. We use this
trained LSTM model to predict weather attributes for year 2016.
Note, we did not predict soil condition parameters as they did not
change over time in experiment dataset.
We use experiment dataset to train RFC model by dividing it into
three parts train, valid and test dataset in the ratio of 70:15:15 respectively. Soil and weather condition attributes in each experiment has
Insights from ViSeed
Common Solution In order to compute the common solution, we
used ViSeed to check multiple combinations of the top ten growing
varieties based on expected value and arrived at the reported solution.
The areas in which these varieties are in the optimal differentiated
solution are shown in figure 5 (i), and the cumulative area for all five
is shown in figure 5 (ii).
Differentiated Solution Analysing the predicted yield for the
differentiated solution, we find two areas with high yield as highlighted in figure 6 (i). The spatial cohesion score is also visualised
in the same figure and observers to be high for the areas with high
yield. We validate this in figure 6 (ii), by visualizing the sub-regions
in which a variety from a high yield, high spatial cohesion score
sub-region is grown. We may conclude that high yield varieties are
localized to certain regions, and so they do not occur in the optimal
common solution.
Experiments with Variability In figure 7, we show the changes
in spatial cohesion and yield for different variability thresholds. We
find that with a low variability threshold of 0.3, there is no optimal
solution for most of the sub-regions (dark-grey areas in the map).
Increasing the threshold, results in solutions for most of the subregions along with an increase in the average yield per sub-region
and the spatial cohesion score. However, beyond a certain threshold,
the gain in yield and spatial cohesion is very small.
Our submission to the Syngenta challenge comprised of two parts.
We use (a) machine learning and stochastic optimisation to compute
solutions and (b) have developed a visual analytics tool, ViSeed,
to analyse the results. In particular, we predict weather and soil
attributes of each sub-region through time-series regression using
LSTMs. For each sub-region and seed variety, a Random Forest
classifier is trained on experiment data to predict yield distributions.
Next, we compute weights of varieties through stochastic optimization, maximizing expected yield and minimizing variance, for each
subregion, followed by visual analytics to choose an optimal global
solution based on the spread of each variety across locally optimal
combinatios.. Our combination of soybean varieties is as follows - (i)
V156774: 48.1%, (ii) V156806: 15.2%, (iii) V152312: 13.5%, (iv)
V114565: 11.8% and (v) V152322: 11.4%. We provide a second
solution to the challenge consisting of individual optimal solutions
for each sub-region which performs marginally better than the first
reported above. Our geo-spatial visual analytics tool ViSeed, is
designed to explore the raw data as well as aid in optimization. Our
entry was not among the top 5 final entries selected from 600 registered teams and as the details of the winning entries have not been
made public, we cannot compare our approach with theirs.
5
C ONCLUSION
In this paper, we propose an approach for crop planning based on
machine learning models (RFC and LSTM), stochastic optimization
and a visual analytics platform. We have given 2 different solution
(i)
(ii)
Figure 6: Differentiated Solution: (i) High Yield Areas and Spatial Cohesion (ii) Validation of Spatial Cohesion
[8] O. Marko, S. Brdar, M. Panic, P. Lugonja, and V. Crnojevic. Soybean
varieties portfolio optimisation based on yield prediction. Computers
and Electronics in Agriculture, 127:467–474, 2016.
[9] M. A. Zaytar and C. El. Sequence to sequence weather forecasting with
long short-term memory recurrent neural networks. 143:7–11, 06 2016.
Figure 7: Differentiated solutions at different Variability thresholds
sets; i) Common solution for entire region, ii) Differentiated solutions at sub-region level. We use expected yield based on a model
using sub-region wise predictions of weather and soil conditions,
standard deviation of expected yield as the criteria to select seed
varieties. We give the spatial cohesion score for each solution which
helps to find the similar solutions in the neigbouring sub-regions.
We also present a geo-spatial visual analytics tool which has the
capability of exploring raw data and helps the retailer in decision
making by allowing exploration of solutions at sub-region as well as
global level.
R EFERENCES
[1] O. Adekanmbi and O. O. Olugbara. Multiobjective optimization of
crop-mix planning using generalized differential evolution algorithm.
2015.
[2] L. Breiman. Random forests. Machine learning, 45(1):5–32, 2001.
[3] S. Drummond, K. Sudduth, A. Joshi, S. Birrell, and N. Kitchen. Statistical and neural methods for site–specific yield prediction.
[4] J. A. Foley, N. Ramankutty, K. A. Brauman, E. S. Cassidy, J. S. Gerber,
M. Johnston, N. D. Mueller, C. OConnell, D. K. Ray, P. C. West, et al.
Solutions for a cultivated planet. Nature, 478(7369):337–342, 2011.
[5] H. C. J. Godfray, J. R. Beddington, I. R. Crute, L. Haddad, D. Lawrence,
J. F. Muir, J. Pretty, S. Robinson, S. M. Thomas, and C. Toulmin. Food security: the challenge of feeding 9 billion people. science, 327(5967):812–
818, 2010.
[6] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural
computation, 9(8):1735–1780, 1997.
[7] T. Iizumi and N. Ramankutty. How do weather and climate influence
cropping area and intensity? Global Food Security, 4:46–50, 2015.
| 7cs.IT
|
1
A Decentralized Trading Algorithm for an Electricity Market
with Generation Uncertainty
arXiv:1705.02577v1 [cs.SY] 7 May 2017
?
Shahab Bahrami? and M. Hadi Amini†
Department of Electrical and Computer Engineering, The University of British Columbia, Vancouver, Canada
†
Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA, USA
email: [email protected], [email protected],
Abstract—The uncertainties of the renewable generation units
and the proliferation of price-responsive loads make it a challenge
for independent system operators (ISOs) to manage the energy
trading market in the future power systems. A centralized energy
market is not practical for the ISOs due to the high computational
burden and violating the privacy of different entities, i.e., load
aggregators and generators. In this paper, we propose a dayahead decentralized energy trading algorithm for a grid with
generation uncertainty. To address the privacy issues, the ISO
determines some control signals using the Lagrange relaxation
technique to motivate the entities towards an operating point
that jointly optimize the cost of load aggregators and profit of
the generators, as well as the risk of the generation shortage
of the renewable resources. More, specifically, we deploy the
concept of conditional-value-at-risk (CVaR) to minimize the
risk of renewable generation shortage. The performance of the
proposed algorithm is evaluated on an IEEE 30-bus test system.
Results show that the proposed decentralized algorithm converges
to the solution of the ISO’s centralized problem in 45 iterations.
It also benefits both the load aggregators by reducing their cost
by 18% and the generators by increasing their profit by 17.1%.
Keywords: price-responsive load, generation uncertainty, distributed algorithm, conditional value-at-risk, trading market.
I. I NTRODUCTION
One goal of the future smart grid is to provide the necessary infrastructure towards integration of renewable generators
(e.g., wind turbine, photovoltaic (PV) panel) as the environmentally friendly alternatives for the fossil fuel-based generators. Meanwhile, the infrastructure in smart grid facilitates
the participation of demand side in the energy management
programs such as demand response (DR) [1].
The optimal power flow (OPF) analysis is widely used by
independent system operators (ISOs) to optimally dispatch the
available generation portfolio to meet the load demand. There
are challenges to address the OPF analysis in a grid with renewable energy generators and price-responsive load demands.
First, addressing the OPF problem by the ISO can violate
the entities’ privacy in a competitive energy market, e.g., by
revealing the demand information of the load aggregators and
cost of the generators. Second, the OPF problem is nonlinear
and nonconvex problem, which is difficult to be solved. Even
by linearizing the problem (e.g., in DC OPF), it can be
computationally difficult in large networks, especially when
the number of decision variables increases by participating
the price-responsive load aggregators in the energy market.
Third, the uncertainty of the renewable generation puts the
generation-load balance at risk, especially when the ISO does
not have access to the accurate historical generation data of
the privately owned renewable generators.
There have been some efforts in the power system operation literature to tackle the above-mentioned challenges. We
divide the related works into three thread. The first thread
is concerned with privacy issues and focuses on designing
a distributed energy planning programs for a system with
multiple suppliers and multiple users using the evolutionary
game [2], Stackelberg game [3], dual decomposition method
[4], supply bidding mechanism [5], and hierarchical bidding
[6]. These approaches may not be easily implementable in
practice, since they ignore the physical constraints imposed
by the topology and operation of the power network. The
second thread is concerned with including the power flow
constraints in the decentralized system planning procedure
using different techniques such as the primal-dual algorithm
[7], convex relaxation [8]–[10], quadratic programming [11],
alternating direction method of multipliers (ADMM) [12]–
[14], and Lagrange relaxation method [15], [16]. In these
works, however, the uncertainty issues related to integrating
renewable generators remain a challenge. The third thread
is concerned with addressing the uncertainty issues by using
different techniques such as chance constrained optimization
[17]–[19], successive constraint enforcement [20], fuzzy systems [21], stochastic optimization [22], and risk management
[23], [24]. The proposed models are centralized, and cannot
address the privacy and computational complexity challenges.
In this paper, we focus on extending the works in the third
thread by designing a distributed energy trading algorithm in a
day-ahead electricity market with renewable energy generators
and active participation of the load aggregators in load management programs. The ISO, load aggregators, and generators
use the smart grid’s communication infrastructure to execute
the proposed distributed algorithm and jointly optimize the
generators’ profit and the load aggregators’ cost. The generators’ and load aggregators solves their own optimization
problems using some locally available information. Hence, the
privacy of each entity is protected. The main challenge is to
determine proper control signals among the ISO, generators,
and load aggregators that enforce the proposed distributed
algorithm converges to the optimal solution of the centralized
problem for an ISO with complete information.
The main contributions of this paper are as follows:
• Risk Evaluation: Inspired by the works in [23], [24],
we deploy a penalty based on the conditional value-atrisk (CVaR) into the objective functions of the ISO’s
centralized problem and the generators’ local problems
to address the uncertainties of the renewable generation.
It enables the generators with renewable plants to sell
electricity and gain profit, while limiting the risk of high
generation shortage within a certain confidence level. The
ISO can also reduce the risk of generation-load mismatch.
We discuss how to deal with the difference between the
2
•
•
risk aversion levels of the ISO and generators.
Distributed Algorithm Design: In order to protect the
privacy of load aggregators and generators, and to address
the computational complexity of the centralized approach,
we propose a decentralized algorithm based on the Lagrange relaxation method [15], [16] that can be executed
by the entities in the day-ahead market. A load aggregator
to manage the controllable loads of its customers. A
generator determine its risk minimizing generation level.
Meanwhile, the ISO can meet the power flow constraints
in the grid, thereby achieving a triple-win result.
Performance Evaluation: Simulation results on an IEEE
30-bus test system show that the proposed decentralized
algorithm can converge to the global optimal solution to
the centralized problem of the ISO in about 45 iterations.
The proposed algorithm also benefits both the load aggregators by reducing their cost by 18% and the generators
by increasing their profit by 17.1% and reducing the peakto-average ratio (PAR) by 15%. When compared with the
centralized approach, our algorithm has a significantly
lower computational time. when compared with a centralized method with AC power flow in different test systems,
our approach has a lower running time at the cost of 3%
to 8% error due to the DC power flow approximation.
The rest of this paper is organized as follows. Section II
introduces the models for load aggregators, generators and
ISO. In Section III we propose a decentralized algorithm to
determine the energy market equilibrium. Section IV provides
simulation results. Section V concludes the paper. Appendices
A and B can be found in the supplementary document.
II. S YSTEM M ODEL
Consider a day-ahead energy market with a set N of load
aggregators and a set M of generators. Each load aggregator
is responsible for managing the load demand on be half
of its electricity users. Each generator sells electricity to
the market. The load aggregators and generators use a twoway communication infrastructure to exchange information
with the ISO. The load aggregators and generators are also
connected to each other within the transmission network. The
ISO monitors the power flow through the transmission lines
over the trading horizon H = {1, . . . , H}, where H is the
number of time slots with equal length, e.g., one hour.
To avoid the potential of confusion, we assume that each bus
has either a load aggregator or generator. If a load aggregator
and a generator are connected to the same bus, we divide that
bus into two buses connected to each other through a line
with zero impedance. If neither load aggregator nor generator
is connected to a bus, we add a virtual load aggregator with
zero demand to that bus. It enables us to denote the set of
buses by N ∪ M and refer a load aggregator or a generator
by its bus number. We use notation L ⊆ (N ∪ M) × (N ∪ M)
to denote the set of transmission lines. In the energy market,
the ISO provides each entity with the day-ahead market price
values for trading the electricity. Let ρb,h denote the price in
time slot h ∈ H for the entity connected to bus b ∈ N ∪ M.
Let vector ρb = (ρb,h , h ∈ H) denote the day-ahead price
profile. In the rest of this section, we will provide the models
for the load aggregators, generators, and ISO.
1) Load Aggregator’s Model: Load aggregator i ∈ N is responsible for managing the load demand of its users. The load
demand li,h in bus i in time slot h consists of uncontrollable
b
c
(or baseload) demand li,h
and controllable demand li,h
. The
controllable demand comprises a set Ai of different types of
controllable loads for residential, commercial, and industrial
users connected to bus i. Controllable load a ∈ Ai has a
scheduling horizon Ha,i ⊆ H that defines the time interval,
in which the load should be scheduled. We further divide the
controllable loads into two types based on their characteristics.
A controllable load of type 1 have hard deadline. It should
be operated within the scheduling horizon and turned off in
other time slots. Examples include the household’s electric
vehicle (EV) and production line of an industry. A controllable
load of type 2 is more flexible. it can be operated in time slots
out of the scheduling horizon at the cost of high discomfort
for the user, e.g., lighting in households, packing process in
industries, and air conditioner in commercial buildings. Let
xa,i,h denote the demand of controllable load a ∈ Ai in time
slot h. Let A1i and A2i denote the set of controllable loads of
type 1 and 2 for load aggeragtor i ∈ N , respectively. We have
a ∈ A1i , h 6∈ Ha,i ,
xa,i,h = 0,
xa,i,h ≥ 0,
a∈
max
xmin
a,i,h ≤ xa,i,h ≤ xa,i,h ,
P
min
Xa,i,h ≤ h∈H xa,i,h
a∈
max
≤ Xa,i
,
A2i , h 6∈ Ha,i ,
A1i ∪ A2i , h ∈ Ha,i ,
a ∈ A1i ∪ A2i .
(1a)
(1b)
(1c)
(1d)
Scheduling the controllable loads results in a discomfort
cost for the users. The discomfort cost for type 1 loads depends
only on the total power consumption deviation from the
desirable value (e.g., the total charging level of an EV is important). For the scheduled power consumption profile xa,i =
(xa,i,h , h ∈ H) and desirable profile xdes
(xdes , h ∈ H),
a,i =
P a,i,h
the discomfort cost Ωa,i (xa,i ) = ωa,i
h∈Ha,i (xa,i,h −
2
xdes
)
with
nonnegative
constant
ω
is
a viable cana,i
a,i,h
didate. The discomfort cost for type 2 loads depends on
both the amount of power consumption and the time of
consuming
the power. The discomfort
cost Ωa,i (xa,i ) =
P
P
des
2
0
ω
(x
−
x
)
+
a,i,h
h∈Ha,i a,i,h a,i,h
h6∈Ha,i ωa,i,h xa,i,h with
0
time dependent nonnegative coefficients ωa,i,h and ωa,i,h
,
0
ωa,i,h ωa,i,h is a viable candidate.
Let xi = (xa,i , a ∈ Ai ) denote the profile of power
demands over all loads of load aggregator i. Load aggregator
agg
i aims to minimize the total
Pcost Ci (xi ), which includes the
discomfort cost Ωi (xi ) = a∈Ai Ωa,i (xa,i ) and the payment
b
to the ISO to meet
P uncontrollable demand li,h and controllable
c
demand li,h = a∈Ai xa,i,h . For i ∈ N , we have
P
b
c
Ciagg (xi ) = Ωi (xi ) + h∈H ρi,h li,h
+ li,h
.
(2)
Let Xi denote the feasible space defined by constraints
(1a)−(1d) for load aggregator i. Load aggregator i ∈ N solves
the following optimization problem.
minimize Ciagg (xi )
xi
(3a)
3
subject to xi ∈ Xi .
(3b)
2) Generator’s Model: Generator j ∈ M sells pconv
j,h of
its conventional unit’s output power to in times slot h. Let
pconv
= (pconv
j
j,h , h ∈ H) denote the conventional unit’s
generation profile for generator j. The cost of a conventional
unit in time slot h is generally an increasing convex function
of pconv
j,h [26]. The class of quadratic generation cost functions
conv conv
2
conv
Cj,h
(pj,h ) = aj2 (pconv
j,h ) +aj1 pj,h +aj0 is well-known [27].
A generator can also use renewable units, such as solar and
wind to benefit from their zero generation cost. Due to the
uncertainty, generator j offers a renewable generation profile
pren
= (pren
j
j,h , h ∈ H), such that the risk of deviation from
the actual generation profile pbjren is minimized. Specifically,
to prevent the generators from non-credible high generation
offers in the market, the ISO charges generator j with renewable units by the price θj,h , h ∈ H (cents/kW), when its actual
ren
generation pbj,h
is lower than its offer pren
j,h . A generator with
renewable units can use risk measures such as the value-at-risk
(VaR) and conditional value-at-risk (CVaR) to limit the risk
of generation shortage. Let ∆j,h (pj,h , pbj,h ) denote a function
that captures the penalty for generation shortage in time slot
h for generator j with renewable units [28]. It is defined as
ren
ren +
∆j,h (pren
bj,h
) = θj,h pren
bj,h
, h ∈ H, j ∈ M, (4)
j,h , p
j,h − p
+
ren
= max{·, 0}. ∆i,h (pren
bj,h
) is a random variable,
j,h , p
ren
actual generation pbj,h is a stochastic process. Under
where [·]
since the
a given confidence level βj ∈ (0, 1) and the offered generation
level pren
j,h in time slot h, the VaR for generator j is defined as
the minimum threshold cost αj,h , for which the probability of
generation shortage of generator j being less than αj,h is at
least βj [28]. That is,
βj
VaRj,h
pren
j,h = min {αj,h | Pr {∆i,h (·) ≤ αj,h } ≥ βj } . (5)
Due to the non-convexity, it is difficult to minimize the
VaR. The CVaR is an alternative risk measure, which is
convex and can be optimized using sampling techniques. The
CVaR for generator j with renewable units in time slot h is
defined as the expected value of the generation shortage cost
ren
∆j,h (pren
bj,h
) when only the costs that are greater than or
j,h , p
βj
equal to VaRj,h (pren
j,h ) are considered [28]. That is,
n
o
βj
βj
ren
CVaRj,h
pren
. (6)
j,h = E ∆j,h (·) ∆j,h (·) ≥ VaRj,h pj,h
It is possible to estimate the CVaR by adopting sample average
approximation (SAA) technique [28]. Samples of the random
ren
variable pbj,h
for generator j with renewable units in time slot
h can be observed from the historical record. Consider the
set K = {1, . . . , K} of K samples of the random variable
ren
ren
pbj,h
. Let pren,k
denote the k th sample of pbj,h
for generator
j,h
j in time slot h. The CVaR in (6) can be approximated by
βj
eβj ren
CVaRj,h
(pren
j,h ) ≈ min Cj,h (pj,h , αj,h ), where
αj,h
βj
Cej,h
pren
j,h , αj,h
β
j
pute Cej,h
(pren
j,h , αj,h ). For further simplification of the objective
+
ren,k
function (9), we replace the terms ∆j,h (pren
j,h , pj,h )−αj,h
k
in (7) with the auxiliary variables ηj,h
for every sample k ∈ K
k
in time slot h ∈ H. We include θj,h (pj,h − pkj,h ) − αj,h ≤ ηj,h
k
into the constraint set. We define ηj,h = (ηj,h , k ∈ K). Hence,
(7) can be rewritten as
βj
(ηj,h , αj,h ) = αj,h +
Cej,h
X
k∈K
= αj,h+
k∈K
K(1−βj )
. (7)
Under the given offer pren
j,h , we can use the historical samples
of the renewable unit’s output power in each time slot to com-
(8)
The objective of a generator is to maximize its profit πjgen ,
which is the revenue from selling electricity in the market
with prices ρj,h , h ∈ H minus the generation cost and
the financial risk CVaR in (8). We define decision vector
ψj = (pconv
, pren
j
j , (ηj,h , αj,h , h ∈ H)) for generator j. The
generator’s profit is obtained as follows.
πjgen (ψj ) =
X
ren
conv
eβj
(pconv
j,h + pj,h )ρj,h −Cj,h (pj,h )− Cj,h (ηj,h , αj,h ) . (9)
h∈H
The local optimization problem for generator j ∈ M is
maximize πjgen (ψj )
(10a)
ψj
max
subject to pmin
≤ pconv
j
j,h ≤ pj ,
θj,h (pren
j,h
−
pren,k
j,h )
− αj,h ≤
k
ηj,h
,
0 ≤ αj,h , ηj,h ,
h ∈ H,
(10b)
k ∈ K, h ∈ H,
(10c)
h ∈ H. (10d)
3) ISO’s Model: A variety of different objectives can be
considered for the ISO in the energy market. In this paper,
we consider the objective of minimizing the social cost with
the risk of generation shortage of all renewable generators
over the planning horizon. The social cost is the generation
cost of the conventional generators plus the discomfort cost
of load aggregators. Regarding the risk minimization, the ISO
can consider the CVaR similar to (4)−(8). Remind that from
(4), the penalty for the generation shortage is important for a
generator. However, from the ISO’s perspective, the value of
generation shortage of the renewable unit is important. That is,
from the ISO’s point of view, the cost of generation shortage in
ren
ren
time slot h corresponding to generator j is ∆ISO
bj,h
)=
j,h (pj,h , p
ren
ren +
[pj,h − pbj,h ] . Moreover, the ISO may consider a confidence
level βjISO different from the confidence level βj of generator j.
β ISO
ISO
ISO
Similar to (8), we can formulate Cej,hj (ηj,h
, αj,h
) as the
approximate CVaR function that the ISO assigns to generator
ISO,k
j. Here, ηj,h
is the auxiliary variable corresponding to term
ISO,k
ren
ren +
ISO
[pj,h − pbj,h ] and ηj,h
= (ηj,h
, k ∈ K) is the vector of
auxiliary variables for samples k ∈ K of the historical data
for generator j in time slot h. We have
ISO
βj
Cej,h
X
ISO
ISO
ISO
ηj,h
, αj,h
= αj,h
+
k∈K
+
ren,k
X ∆j,h (pren
j,h , pj,h )−αj,h
k
ηj,h
.
K(1−βj )
ISO,k
ηj,h
K(1−βjISO )
.
(11)
As a netural entity, the ISO uses the DC power flow model
as an acceptable framework to determine the active power flow
through the transmission lines and the generation-load balance
in transmission systems [29], [30]. Let δi,h denote the phase
angle of the voltage in bus i ∈ N in time slot h ∈ H. We
4
use vector δh = (δi,h , i ∈ N ) to denote the profile of voltage
phase angles in all buses in time slot h. We define the vector of
ISO
ISO
decision variables ψ ISO = (pconv
, pren
j
j , ηj,h , αj,h , xi , δh , h ∈
H, j ∈ J , i ∈ N ) for the ISO. The ISO’s objective function is
X X conv
βjISO
ISO
ISO
f ISO ψ ISO =
Cj,h (pj,h ) +ϑc Cej,h
ηj,h
, αj,h
j∈M
h∈H
+
X
Ωi,h (xi,h ) ,
(12)
i∈N
where ϑc is a positive weighting coefficient. Minimizing
f ISO ψ ISO will enable timely adjustment of control settings
to jointly reduce the consumers’ discomfort cost and the
generators’ generation cost, as well reducing the likelihood of
high generation shortage. Thus, it can improve the economic
efficiency of the system operation. We formulate the ISO’s
centralized problem as
minimize f ISO ψ ISO
(13a)
ψ ISO
subject to constraints (3b) and (10b), i ∈ N , j ∈ M, (13b)
X
− li,h =
bi,r (δi,h − δr,h ),
i ∈ N , h ∈ H, (13c)
(i,r)∈L
pconv
j,h
+ pren
j,h =
X
bj,r (δj,h − δr,h ), j ∈ M, h ∈ H, (13d)
(j,r)∈L
br,s (δr,h − δs,h ) ≤ pmax
r,s ,
pren
j,h
0≤
ISO
− pren,k
j,h − αj,h
ISO,k ISO,k
αj,h
, ηj,h ,
≤
ISO,k
ηj,h
,
(r, s) ∈ L, h ∈ H, (13e)
j ∈ M, k ∈ K, h ∈ H, (13f)
j ∈ M, h ∈ H. (13g)
Constraints (13c) and (13d) are the nodal power balance
equations. Constraint (13e) is the line flow limit, where pmax
r,s
is the maximum power flow limit for line (r, s) ∈ L and br,s
is the admittance of line (r, s) ∈ L. Constraint (13f) is for
ISO,k
ren +
bj,h
] .
auxiliary variable ηj,h
corresponding to term [pren
j,h − p
Problem (13) is an optimization problem with a convex
objective function (remind that CVaR is convex) and linear
constraints. By considering one bus as a slack bus (e.g., bus
1 with δ1,h = 0, h ∈ H), we have the following theorem.
Theorem 1 If the ISO’s centralized problem in (13) has a
feasible point, then the optimal solution exists and is unique.
To solve the centralized problem in (13), the ISO needs
the information about the load aggregators’ discomfort cost,
conventional generators’ cost, and renewable units’ historical
data samples. However, these information may not be available
to the ISO. Also, solving problem (13) can be computationally
difficult in large networks. Instead, we develop a decentralized
algorithm. We show that the ISO can determine the price
signals ρb,h for bus b ∈ N ∪ M and penalties θj,h , h ∈ H for
generators j ∈ M, such that the competitive equilibrium of
the market coincides with the unique solution of the ISO’s
centralized problem in (13). Note that the energy market
equilibrium corresponds to the optimal solution to problems
(3) and (10) for all entities.
III. E NERGY M ARKET I NTERACTIONS
The ISO has no direct control over the generators’ and
load aggregators’ behavior. Alternatively, the ISO provides
the entities with sufficient access to the day-ahead energy
market to determine their own generation and load demand.
The entities compete with each other to optimize their local
problems (3) and (10). The decision vector of the load aggregator i is the controllable load profile xi and the decision
vector for generator j is ψj . The ISO influences the entities
by using the nodal prices ρ = (ρb , b ∈ N ∪ M) and the
penalties θ = (θj , j ∈ M) for renewable generators. The
goal of the ISO is to determine vectors ρ and θ, such that
the market equilibrium coincides with the optimal solution
of the centralized problem (13). The idea is to formulate
the Lagrange relaxation of the ISO’s problem (13) in order
to divide problem (13) into several subproblems. Then, we
determine vectors ρ and θ such that the subproblems becomes
the same as load aggregators and generators problems in (3)
and (10). This approach is viable since Due to problem (13) is
convex and the constraints are linear. Thus, the strong duality
gap condition (Slater’s condition) is satisfied if a feasible
solution exists [31, Ch. 5].
Let λagg
i,h , i ∈ N , h ∈ H denote the Lagrange multiplier
associated with the equality constraint (13c). Let λgen
j,h , h ∈
H, j ∈ M denote the Lagrange multiplier associated with
the equality constraint (13d). Let µr,s,h and µr,s,h , h ∈
H, (r, s) ∈ L denote the Lagrange multiplier associated with
the upper and lower inequalities (13e), respectively. Also, let
ISO,k
γj,h
, j ∈ M, k ∈ K, h ∈ H denote the Lagrange multiplier
associated with constraint (13f). We define vector φISO =
gen
ISO,k
(λagg
i,h , λj,h , µr,s,h , µr,s,h , γj,h , i ∈ N , j ∈ M, (r, s) ∈
L, k ∈ K, h ∈ H). We have the following main result.
Theorem 2 The equilibrium of the energy market coincides
with the unique solution to the ISO’s centralized problem in
(13) if and only if for i ∈ N , j ∈ M, h ∈ H the ISO sets
ρi,h = −λagg
i,h ,
ρj,h =
(14b)
P
θj,h
(14a)
λgen
j,h ,
ISO,k
k∈K γj,h
1−β
j
=
.
P
ISO,k 1 − β ISO
1 − ϑc + k∈K γj,h
j
(14c)
The proof can be found in Appendix A. It suggests a decentralized algorithm to determine the solution to problem (13).
In a decentralized algorithm, the load aggregators, generators and ISO interact with each other. Let q denote the
iteration index. Our algorithm involves the initiation phase and
market trading phase. Algorithm 1 describes the interactions
among the load aggregators, generators, and ISO. Algorithm
1 is based on the projected gradient method. The step size in
iteration q of this algorithm is denoted by q .
Initiation phase: Lines 1 to 4 describe the initiation phase.
Market trading phase: The loop involving Lines 5 to 12
describes the market trading phase, in which the load aggregators, generators and ISO update their decision variables in
an iterative fashion. This phase includes the following parts:
1) Information exchange: In Line 6, each load aggregator
q
c,q
b
i sends its load profile (li,h
= li,h
+ li,h
, h ∈ H) to the ISO
via the communication network. Each generator j sends the
generation profile pconv,q
and pren,q
to the ISO as well.
j
j
2) ISO’s update: In Line 7, when the ISO receives the
5
Algorithm 1 Decentralized Energy Market Trading Algorithm.
1: Set q := 1 and ξ := 10−2 .
2: Each load aggregator i ∈ N randomly initializes its controllable
load profile x1i .
3: Each generator j ∈ M randomly initializes its generation profile
p1j . Each generator j with renewable units initializes vectors α1j
and ηj1 randomly, and provides the ISO with confidence level βj .
4: The ISO initializes the voltage angles δb1 , b ∈ N ∪ M and the
vector of Lagrange multipliers φISO,1 .
5: Repeat
q
6: Each load aggregator i and generator j, sends li,h , h ∈ H and
7:
8:
9:
10:
11:
12:
pconv,q
and pren,q
to the ISO, respectively.
j
j
q
ISO updates the voltage angles δb,h
, h ∈ H, b ∈ N ∪ M, and
the vector of Lagrange multipliers φISO,q according to (15a)
and (15b).
ISO communicates the updated nodal price ρq+1 and penalties
θ q+1 to the entities.
Each load aggregator i updates the controllable load profile xqi
by solving its local problem in (3).
Each generator j updates its decision vector ψjq by solving its
local problem in (10).
q := q + 1. The step size q is updated.
q
q−1
Until ||δb,h
− δb,h
|| ≤ ξ, h ∈ H, b ∈ N ∪ M.
information from the entities, it obtains the updated values
q+1
of the voltage angles δb,h
for b ∈ N ∪ M in time slot
h ∈ H and the vector of Lagrange multipliers φISO,q+1 using
ISO
the gradient of the Lagrangian function fLag
(ψ ISO , φISO ) (See
(S-1) in Appendix A) in iteration q as follows:
q+1
q
ISO
q
δb,h
= δb,h
+ q ∇δi,h
fLag
(ψ ISO,q , φISO,q ),
(15a)
q
(15b)
φ
ISO,q+1
=φ
ISO,q
+
ISO
∇φISO,qfLag
(ψ ISO,q , φISO,q ),
where ∇ is the gradient operator. The ISO uses the results of
Theorem 2 to compute the updated values of the nodal prices
ρq+1 and penalties θ q+1 . In Line 8, the ISO communicates
the above control signals to the corresponding load aggregator
and generator.
3) Load aggregator’s update: When load aggregator i
receives the control signals ρq+1
i,h , h ∈ H from the ISO, in
Line 9, it updates its controllable load profile xqi by solving
its local problem in (3). It is a quadratic program with linear
constraints and can be solved efficiently by the load aggregator.
Generator’s update: When generator j receives the control
q+1
signals ρq+1
j,h and θj,h , h ∈ H from the ISO, in Line 10, it
updates its decision vector ψjq by solving its local problem
in (10). It is a quadratic program with linear constraints and
can be solved efficiently by the generator using its local
information about its conventional and renewable units.
Step size update: We use a nonsummable
step
P∞ diminishing
q
size with conditions limq→∞ q = 0,
=
∞,
and
q=1
P∞
1
q 2
q
= ∞. One example is = a+bq , where a and
q=1
b are positive constant coefficients. In Line 11, the step size
is updated. In Line 12, the stopping criterion is given. For
stoppin criterion, we have used the convergence of the voltage
angles, since they depends on the generation and load level
at all buses. Hence, the convergence of the voltage angles
implies the convergence of all generators’ and load aggregators’ decision variables. The proposed Lagrange relaxation-
based algorithm converges for a nonsummable diminishing
step size [31, Ch. 5].
IV. P ERFORMANCE E VALUATION
In this section, we evaluate the performance of our proposed
decentralized algorithm on an IEEE 30-bus test system with
6 generators and 21 load aggregators.
1) Simulation setup: The data for the test system and the
generators cost functions are given in [32]. Since buses 2, 5,
and 8 have both generator and load aggregator in the original
test system, we add the new virtual buses 31, 32, and 33 for
the load aggregators as shown in Fig. S-1 in Appendix B. The
trading horizon is one day with H = 24 one-hour time slots.
The weight coefficient in the ISO’s objective function (12) is
set to ϑc = 2 (without units).
To obtain different baseload patterns for the load aggregators, we use a load pattern for about 5 million consumers
(which includes residential, commercial, and industrial consumers) from Ontario, Canada power grid database [33] from
November 1 to November 21 2016. We scale the load pattern
for each bus, such that the average baseload becomes equal to
60% of the load demand of that bus in [32]. To simulate the
controllable load demand of each bus, we randomly generate
the desirable load profile of 500 to 1000 controllable loads of
types 1 and 2 with average demand of 2 to 15 kW at each
min
bus. Limits xmax
a,i,h and xa,i,h for controllable load a in bus
i are set to ±30% of the desirable demand of that load in
max
min
time slot h. Limits Xa,i
and Xa,i
for controllable load a in
bus i are set to ±5% of the desirable total demand of that
load. The discomfort coefficients ωa,i for controllable load
a of type 1 are randomly chosen from a truncated normal
distribution, which is lower bounded by zero and has a mean
value of ωiavg = 15 cents/(kWh)2 and a standard deviation
of 5 cents/(kWh)2 . The discomfort coefficients ωa,i,h , h ∈ H
for controllable load a of type 2 in bus i are randomly chosen
from a truncated normal distribution, which is lower bounded
by zero and has a mean value of ωiavg = 15 cents/(kWh)2
and a standard deviation of 5 cents/(kWh)2 . The discomfort
0
coefficients ωa,i,h
, h ∈ H for controllable load a of type 2 in
bus i are set to 50 cents/(kWh)2 .
We assume that generators in buses 8 and 11 have PV panel
and wind turbine, respectively, in addition to their conventional
units. To obtain the samples for the output power of the wind
turbine and PV panel, we use the available historical data from
Ontario, Canada power grid database [33], from November 1
to November 21 2016. For each renewable generator, we scale
down the available historical data, such that the average output
power over the historical data becomes equal to 4 MW.
1
. For the benchmark scenario,
The step size is q = 10+0.2q
we consider a system without renewable units and DR program. We perform simulations using Matlab R2016b in a PC
with processor Intel(R) Core(TM) i7-3770K [email protected] GHz.
2) Load aggregators: Each load aggregator executes Algorithm 1 to modify the controllable load demand of its users.
Fig. 1 shows the load profile of the load aggregators in buses
17, 30, 31 (the load in bus 2), 33 (the load in bus 5) and
the benchmark system without renewable generators and DR
6
j
4) Conventional generators: The generators with conventional units can also benefit from the DR program by reducing
the PAR in the aggregate demand. For example, Fig. 6 shows
that using the DR program makes the generation profile of the
conventional units in bus 11 and 13 smoother. In other words,
as the demand profiles in different buses change toward a profile with a lower peak demand, the required power generation
follows the same trend. In order to quantify the impact of the
DR program on the generation profile, we provide the value of
the PAR with and without DR program in Fig. 7. Results show
the reduction in PAR for the generators by 15% on average.
Load demand (MW)
40
125
20
100
0
7 am
12 pm
6 pm
12 am
6 am
75
7 am
15
15
10
10
5
7 am
12 pm
6 pm
12 am
6 am
5
7 am
12 pm
6 pm
12 am
6 am
12 pm
6 pm
12 am
6 am
Time (hour)
Load shift percentage (%)
Figure 1. Load demand profiles over 24 hours in buses 31 (up-left), 32
(up-right), 17 (down-left), and 30 (down-right) with and without DR and
renewable generators.
20
10
0
3
4
7 10 12 14 15 16 17 18 19 20 21 23 24 26 29 30 31 32 33
Figure 2. The values of shift load percentage for different load aggregators
avg
with ωi = 500, 50, 30, and 15 cents/(kWh)2 .
Total cost ($)
3) Renewable generators: Generator 11 has a PV panel and
generator 13 has a wind turbine in addition to their conventional units. Recall that these generators sell the presumed
output power of the renewable units in the day-ahead market,
and pay a penalty for generation shortage. In Algorithm 1, the
generators use the historical data of the renewable resources
and respond to the control signals (14b) and (14c) to determine
the amount of power that they plan to sell in the market. Fig.
4 shows the historical data and the presumed output power of
the PV panel in bus 11 and the wind turbine in bus 13 for the
confidence levels β11 = β13 = 0.9. The obtained presumed
output power will have the lowest risk in the day-ahead market
for these generators. Two parameters affect the presumed
output power of the renewable generator: the value of weight
coefficient ϑc in the ISO’s objective function (12) and the
values of confidence levels βj (set by generator j) and βjISO
(set by the ISO). The penalty θj,h in (14c) for the generation
shortage increases when ϑc increases. That is, when the riskaverse ISO puts a higher weight on the risk of generation
shortage, it assigns a larger penalty. Fig. 5 (a) (for bus 11) and
Fig. 5 (b) (for bus 13) show that a larger coefficient ϑc enforces
the generators to offer a lower renewable generation in the
market during most of the times. In a similar manner, equation
(14c) implies that when βj > βjISO , i.e., generator j is more
risk-averse than the ISO, then the ISO reduces penalty θj,h
1−βj
< 1. If βj < βjISO , i.e., generator j is more
by factor 1−β ISO
j
risk-taker than the ISO, then the ISO increases penalty θj,h
1−βj
by 1−β ISO
> 1 to limit the likelihood of generation shortage.
With DR and renewable generation
Without DR and renewable generation
3000
2500
2000
1500
1000
3
4
7 10 12 14 15 16 17 18 19 20 21 23 24 26 29 30 31 32 33
Figure 3. The load aggregator’s cost with low and high discomfort cost’s
scaling coefficients.
Renewable generation (MW)
program. Peak shaving can be observed in the load profiles.
Results for all load aggregators verify that by executing the
decentralized Algorithm 1, the peak load demand is reduced by
16.5% on average. Since the peak load may occur in different
times for different load aggregators, we consider the index of
load shift percentage, which is the ratio of the total shifted
load demand to the aggregate demand over 24 hours. In Fig.
2, the shift load percentage index for different values of the
average scaling coefficient ωiavg at all load buses are provided.
When ωiavg increases in a bus, the load aggregator modifies a
lower amount of controllable loads due to a higher discomfort
cost of its users, and thereby the load shift percentage index
decreases. Load shifting is performed with the goal of reducing
the total cost in (2). Fig. 3 shows that the total cost in (2)
of the load aggregators is lower (by about 18%) when the
scaling coefficients ωiavg are lower, since the load aggregators
can benefit from the price fluctuations and shift more amount
of load demand to the hours with lower price values.
10
7.5
5
5
0
2.5
0
96
192
288
384
478
10
12 pm
6 pm
12 am
6 am
12 pm
6 pm
12 am
6 am
7.5
6
6
2
0
7 am
4.5
0
96
192
288
384
478
3
7 am
Time (hour)
Figure 4. (left figures) The PV panel and wind turbine historical data samples.
(right figures) The presumed output power of the PV panel and wind turbine
in buses 11 and 13.
The reduction in the PAR can reduce the peak load demand,
and thus the generation cost of the generators. However, the
revenue of the generators will decrease due to a lower price
in the market. Fig. 8 shows the profit (revenue minus cost)
for the generators with and without DR program. The profit
of the generators increase by 17.1% on average with DR. To
complete the discussion, we consider different discomfort cost
coefficients ωiavg , i ∈ N for the load aggregators and report the
average profit of generators in Fig. 9. It is interesting that the
scenario with the most flexible load demands (i.e., ωiavg = 0)
does not lead to the maximum profit for the generators, though
the generation cost of the generators are lower in this scenario.
The reason is that a higher load flexibility will flatten the load
profiles, and thus reduces the price values and the revenue of
c
c
=0.2
10
c
=0.5
c
=1
=2
c
=5
c
7.5
5
2.5
0
7 am
12 pm
6 pm
103
75
=20
Profit ($)
PV generation (MW)
7
12 am
With DR and renewable generation
Without DR and renewable generation
50
25
0
6 am
1
5
2
11
8
13
Generator bus number
c
=0.2
c
c
=0.5
=1
=2
c
=5
c
Figure 8. The profit of the generators with and without DR and renewable
generation.
=20
7
6
5
4
3
7 am
12 pm
6 pm
12 am
6 am
Time (hour)
Figure 5. The presumed output power of (a) the PV panel in bus 11, and
(b) the wind turbine in bus 13 for different values of coefficient ϑc .
70
Generation (MW)
60
Average profit ($)
Wind turbine
generation (MW)
c
8
43
33
100
150
1.5
40
1,5
1
30
7 am
12 pm
6 pm
12 am
200
250
300
350
400
450
500
7,10
2,5
28,15 6,15
5,10
20,20 10,20
0.5
6 am
70
0
60
-0.5
-1
40
-1.5
30
7 am
12 pm
6 pm
12 am
0
Time (hour)
1.4
1.2
5
8
11
60
90
75
1250
1000
750
13
Generator bus number
Figure 7. The PAR in the generation of the generators with and without DR
and renewable generation.
the generators. Here, the reduction in the revenue is larger than
the reduction in the costs. The maximum profit is achieved for
ωiavg = 28.2 cents/(kWh)2 .
5) Algorithm convergence: We study the required number
of iterations for convergence, which can be interpreted as
an indicator of the number of message exchanges among
the load aggregators, generators and the ISO. The value of
voltage angles depends on all generators’ and load aggregators’ decision variables. Thus, the convergence of the voltage
angles is a viable indicator of the convergence of the decision
variables in all buses of the system. Since the values of
the voltage angles of all buses can be added by a constant,
we illustrate the convergence of the phase angle difference
between the voltages of the buses at the end nodes of the lines.
As an example, we provide the convergence of δ1,5 − δ2,5 ,
δ7,10 − δ5,10 , δ28,15 − δ6,15 , and δ20,20 − δ10,20 in Figures 10.
We can observe that 45 iterations are enough for convergence.
The average running time of the algorithm for different initial
conditions is 19 seconds for 200 random initial conditions.
Centeralized algorithm
Decentralized algorithm
500
250
0
2
45
30
Figure 10. The convergence of phase difference over transmission lines.
CPU time (s)
Without DR and renewable generation
With DR and renewable generation
1
15
6 am
Figure 6. The generation profile of the generators (a) 11 and (b) 13.
1
50
Figure 9. The average profit of the generators in terms of the average weight
avg
coefficients ωi , i ∈ N .
Without DR and renewable generation
With DR and renewable generation
50
PAR
63
53
0
50
1.6
103
73
IEE
E1
4-b
us
IEE
E3
0-b
us
IEE
E1
18
IEE
Po
lish
-bu
E3
s
00
-bu
s
Po
23
lish
83
wp
30
12
w
p
Figure 11. The CPU time of the centralized and decentralized algorithms.
We use MOSEK solver to solve the ISO’s centralized
problem (13). The solution is the same as the decentralized
approach, but the running time is 35 seconds. To further
elaborate the comparison, we provide the average running
time of Algorithm 1 and the centralized approach for six test
systems [32] in Fig. 11, we also use the approach in [24] that
applies semidefinite programming (SDP) to obtain the global
optimal point in a grid using AC OPF analysis and the CVaR
as the risk measure. In addition to comparing the algorithm
running time, the global optimality of the solution to the AC
OPF enables us to quantify the approximation in using the
DC OPF in our decentralized algorithm. The main difference
is that the AC OPF includes the network losses, while the DC
OPF doesn not consider the network losses. Furthermore, the
SDP approach in [24] returns the global optimal solution to
the AC OPF. The calculation results show that the values of
the ISO’s objective using Algorithm 1 with DC OPF are lower
by 3% (in Polish 2383wp) to 8% (in IEEE 300-bus system)
than the centralized method with AC OPF and SDP approach
due to the inclusion of losses and optimality of SDP method.
Whereas, the computation time is much lower in Algorithm 1.
8
Table I
T HE OPTIMAL VALUE AND AVERAGE CPU T IME FOR THE D ETERMINISTIC
M ULTI - STAGE A LGORITHM AND O UR P ROPOSED A LGORITHM .
Centralized algorithm
with AC OPF
Algorithm 1
f ISO ($)
CPU time (s)
207,780.8
10
217,991.1
24
IEEE 30-bus
330,760.2
16
347,980.8
63
IEEE 118-bus
3,290,813.4
25
3,490,815.8
153
IEEE 300-bus 18,152,944.3
42
19,845,152.8
218
Polish 2383wp 43,263,027.9
136
44,164,101.3
1,312
Polish 3012wp 64,121,518.3
182
71,426,241.2
1,636
Test system
IEEE 14-bus
f ISO ($)
CPU time (s)
V. C ONCLUSION
In this paper, we proposed a decentralized algorithm for
energy trading among load aggregators and generators in a
power grid. In our model, the ISO sends control signals to the
entities to motivate them towards optimizing their objectives
independently, while meeting the physical constraints of the
power network. We introduced the concept of CVaR to limit
the likelihood of renewable generation shortage in the dayahead market. To evaluate the performance of the proposed
decentralized algorithm, we used an IEEE 30-bus test system
connected to some renewable generators. Results confirmed
that the algorithm converges in 45 iterations. We also evaluated
the price responsive load profiles and generation values for
various discomfort costs of the load aggregators, and showed
that the proposed decentralized algorithm can benefit the load
aggregators by reducing their cost by 18%, and the generators
by reducing the PAR by 15% and increasing their profit by
17.1%. Our algorithm benefit the ISO by maintaining the
privacy issues and a lower computational time compared to
the centralized approach. When compared with a centralized
method with AC power flow equations, our approach has a
lower running time at the cost of 3% to 8% approximation
error due to using the DC power flow equations.
R EFERENCES
[1] S. Kar, G. Hug, J. Mohammadi, and J. M. Moura, “Distributed state
estimation and energy management in smart grids: A consensus innovations approach,” IEEE Journal of Selected Topics in Signal Processing,
vol. 8, no. 6, pp. 1022–1038, 2014.
[2] B. Chai, J. Chen, Z. Yang, and Y. Zhang, “Demand response management with multiple utility companies: A two-level game approach,” IEEE
Trans. on Smart Grid, vol. 5, no. 2, pp. 722–731, Mar. 2014.
[3] S. Maharjan, Q. Zhu, Y. Zhang, S. Gjessing, and T. Basar, “Dependable
demand response management in the smart grid: A stackelberg game
approach,” IEEE Trans. on Smart Grid, vol. 4, no. 1, pp. 120–132, Mar.
2013.
[4] R. Deng, Z. Yang, F. Hou, M.-Y. Chow, and J. Chen, “Distributed realtime demand response in multiseller-multibuyer smart distribution grid,”
IEEE Trans. on Power Systems, vol. PP, no. 99, pp. 1–11, Oct. 2014.
[5] F. Kamyab et al., “Demand response program in smart grid using supply
function bidding mechanism,” IEEE Trans. on Smart Grid, vol. 7, no. 3,
pp. 1277 – 1284, 2016.
[6] M. Parvania, M. Fotuhi-Firuzabad, and M. Shahidehpour, “ISO’s optimal
strategies for scheduling the hourly demand response in day-ahead
markets,” IEEE Trans. on Power Systems, vol. 29, no. 6, pp. 2636–2645,
2014.
[7] V. R. Disfani, L. Fan, and Z. Miao, “Distributed dc optimal power flow
for radial networks through partial primal dual algorithm,” in 2015 IEEE
Power & Energy Society General Meeting. IEEE, 2015, pp. 1–5.
[8] W. Shi, N. Li, X. Xie, C. Chu, and R. Gadh, “Optimal residential demand
response in distribution networks,” IEEE Journal on Selected Areas in
Comm., vol. 32, no. 7, pp. 1441–1450, Jun. 2014.
[9] N. Li, L. Gan, L. Chen, and S. Low, “An optimization-based demand
response in radial distribution networks,” in Proc. of IEEE Globecom,
Anaheim, CA, Anaheim, CA 2012.
[10] E. Dall’Anese, H. Zhu, and G. B. Giannakis, “Distributed optimal power
flow for smart microgrids,” IEEE Trans. on Smart Grid, vol. 4, no. 3,
pp. 1464–1475, Sept. 2013.
[11] A. G. Bakirtzis and P. N. Biskas, “A decentralized solution to the DCOPF of interconnected power systems,” IEEE Trans. on Power Systems,
vol. 18, no. 3, pp. 1007–1013, Aug. 2003.
[12] T. Erseghe, “Distributed optimal power flow using ADMM,” IEEE
Trans. on Power Systems, vol. 29, no. 5, pp. 2370–2380, Sept. 2014.
[13] Q. Peng and S. H. Low, “Distributed algorithm for optimal power flow
on a radial network,” in Proc. of IEEE Conf. on Decision and Control,
Dec. 2014, pp. 167–172.
[14] S. Magnússon, P. C. Weeraddana, and C. Fischione, “A distributed
approach for the optimal power-flow problem based on ADMM and
sequential convex approximations,” IEEE Trans. on Control of Network
Systems, vol. 2, no. 3, pp. 238–253, Sept. 2015.
[15] J. M. Arroyo and F. D. Galiana, “Energy and reserve pricing in security
and network-constrained electricity markets,” IEEE Trans. on Power
Systems, vol. 20, no. 2, pp. 634–643, May 2005.
[16] G. Hug-Glanzmann and G. Andersson, “Decentralized optimal power
flow control for overlapping areas in power systems,” IEEE Trans. on
Power Systems, vol. 24, no. 1, pp. 327–336, Feb. 2009.
[17] H. Zhang and P. Li, “Chance constrained programming for optimal
power flow under uncertainty,” IEEE Transactions on Power Systems,
vol. 26, no. 4, pp. 2417–2424, 2011.
[18] D. Bienstock, M. Chertkov, and S. Harnett, “Chance-constrained optimal
power flow: Risk-aware network control under uncertainty,” SIAM
Review, vol. 56, no. 3, pp. 461–495, 2014.
[19] Y. Cao, Y. Tan, C. Li, and C. Rehtanz, “Chance-constrained
optimization-based unbalanced optimal power flow for radial distribution
networks,” IEEE transactions on power delivery, vol. 28, no. 3, pp.
1855–1864, 2013.
[20] R. A. Jabr, “Adjustable robust opf with renewable energy sources,” IEEE
Transactions on Power Systems, vol. 28, no. 4, pp. 4742–4751, 2013.
[21] V. Miranda and J. Saraiva, “Fuzzy modelling of power system optimal
load flow,” in Power Industry Computer Application Conference, 1991.
Conference Proceedings. IEEE, 1991, pp. 386–392.
[22] A. Nasri, S. J. Kazempour, A. J. Conejo, and M. Ghandhari, “Networkconstrained ac unit commitment under uncertainty: A benders’ decomposition approach,” IEEE transactions on power systems, vol. 31, no. 1,
pp. 412–422, 2016.
[23] Y. Zhang and G. B. Giannakis, “Robust optimal power flow with wind
integration using conditional value-at-risk,” in Smart Grid Communications (SmartGridComm), 2013 IEEE International Conference on.
IEEE, 2013, pp. 654–659.
[24] P. Samadi, S. Bahrami, V. Wong, and R. Schober, “Power dispatch
and load control with generation uncertainty,” in Proc. of IEEE Global
Conf. on Signal and Information Processing (GlobalSIP), Dec. 2015,
pp. 1126–1130.
[25] L. Song, Y. Xiao, and M. van der Schaar, “Demand side management
in smart grids using a repeated game framework,” IEEE Selected Areas
in Comm., vol. 32, no. 7, pp. 1412–1424, July 2014.
[26] T. Li and M. Shahidehpour, “Price-based unit commitment: a case of
lagrangian relaxation versus mixed integer programming,” IEEE Trans.
on Power Systems, vol. 20, no. 4, pp. 2015–2025, 2005.
[27] J. Stewart, Calculus, 4th ed. CA: Brooks/Cole Pub Co., 1999.
[28] R. T. Rockafellar and S. Uryasev, “Optimization of conditional valueat-risk,” J. of Risk, vol. 2, pp. 21–42, Apr. 2000.
[29] T. J. Overbye, X. Cheng, and Y. Sun, “A comparison of the AC and DC
power flow models for LMP calculations,” in Proceedings of the 37th
Annual Hawaii International Conference on System Sciences. IEEE,
2004, pp. 9–pp.
[30] B. Stott, J. Jardim, and O. Alsaç, “DC power flow revisited,” IEEE
Trans. on Power Systems, vol. 24, no. 3, pp. 1290–1300, 2009.
[31] S. Boyd and L. Vandenberghe, Convex Optimization.
Cambridge
University Press, 2004.
[32] University of Washington, power systems test case archive. [Online].
Available: http://www.ee.washington.edu/research/pstca.
9
[33] Independent Electricty System Operator (IESO). [Online]. Available:
http://www.ieso.ca
| 3cs.SY
|
arXiv:1406.7392v3 [math.AT] 9 Oct 2016
ON HOCHSTER’S FORMULA FOR A CLASS OF QUOTIENT
SPACES OF MOMENT-ANGLE COMPLEXES
LI YU
Abstract. Any finite simplicial complex K and a partition of the vertex set
of K determines a canonical quotient space of the moment-angle complex of
K. We prove that the cohomology groups of such a space can be computed via
some Hochster’s type formula, which generalizes the usual Hochster’s formula
for the cohomology groups of moment-angle complexes. In addition, we show
that the stable decomposition of moment-angle complexes can also be extended
to such spaces. This type of spaces include all the quasitoric manifolds that
are pullback from the linear models. And we prove that the moment-angle
complex associated to a finite simplicial poset is always homotopy equivalent
to one of such spaces.
1. Introduction
An abstract simplicial complex on a set rms “ tv1 , ¨ ¨ ¨ , vm u is a collection K of
subsets σ Ď rms such that if σ P K, then any subset of σ also belongs to K. We
always assume that the empty set belongs to K and refer to σ P K as an abstract
simplex of K. The simplex corresponding to the empty set is denoted by 0̂. In
particular, any element of rms is called a vertex of K. We call the number of
vertices of a simplex σ the rank of σ, denoted by rankpσq. Let dimpσq denote the
dimension of a simplex σ. So rankpσq “ dimpσq ` 1.
Any finite abstract simplicial complex K admits a geometric realization in some
Euclidean space. But sometimes we also use K to denote its geometric realization
when the meaning is clear in the context.
Given a finite abstract simplicial complex K on a set rms and a pair of spaces
pX, Aq with A Ă X, we can construct of a topological space pX, AqK by:
ď
ź
ź
A.
(1)
pX, AqK “
pX, Aqσ , where pX, Aqσ “
Xˆ
vj Pσ
σPK
vj Rσ
Key words and phrases. Moment-angle complex, Hochster’s formula, Tor algebra, balanced
simple polytope.
2010 Mathematics Subject Classification. 57S15, 14M25,
The author is partially supported by Natural Science Foundation of China (grant no.11371188)
and the PAPD (priority academic program development) of Jiangsu higher education
institutions.
1
2
LI YU
ś
The symbol
here and in the rest of this paper means Cartesian product. So
pX, AqK is a subspace of the Cartesian product of m copies of X. It is called the
polyhedral product or the generalized moment-angle complex of K and pX, Aq. In
particular, ZK “ pD 2 , S 1qK and RZK “ pD 1 , S 0 qK are called the moment-angle
complex and real moment-angle complex of K, respectively (see [5]). Moreover,
we can define the polyhedral product pX, AqK of K with m pairs of spaces pX, Aq “
tpX1 , A1 q, ¨ ¨ ¨ , pXm , Am qu (see [2] or [6, Sec 4.2]).
Originally, ZK and RZK were constructed by Davis and Januszkiewicz [8] in
a different way. We will only explain the construction of ZK below (the RZK
case is completely parallel). Let K1 denote the barycentric subdivision of K.
We can consider K1 as the set of chains of simplices in K ordered by inclusions.
For each simplex σ P K, let Fσ denote the geometric realization of the poset
Kěσ “ tτ P K | σ Ď τ u. Thus, Fσ is the subcomplex of K1 consisting of all
simplices of the form σ “ σ0 Ĺ σ1 Ĺ ¨ ¨ ¨ Ĺ σl . Let PK denote the cone on K1 . If
σ is a pk ´ 1q-simplex, then we say that Fσ Ă PK is a face of codimension k in
PK . The polyhedron PK together with its decomposition into “faces” tFσ uσPK is
called a simple polyhedral complex (see [8, p.428]).
Let V pKq denote the vertex set of K. Any map λ : V pKq Ñ Zr is called a
r
Z -coloring of K, and any element of Zr is called a color. For any simplex σ P K,
‚ let V pσq denote the vertex set of σ and,
‚ let Gλ pσq denote the toral subgroup of T r “ pS 1 qr corresponding to the
subgroup of Zr generated by tλpvq | v P V pσqu.
Given a Zr -coloring λ of K, we obtain a space XpK, λq defined by
XpK, λq :“ PK ˆ T r { „
(2)
where pp, gq „ pp1 , g 1 q whenever p1 “ p P Fσ and g 1g ´1 P Gλ pσq for some σ P K.
In particular, if r “ |V pKq| “ m and tλpvj q ; 1 ď i ď mu is a basis of Zm ,
XpK, λq is homeomorphic to ZK . Let πK : PK ˆ T m Ñ ZK be the corresponding
quotient map in (2). There is a canonical action of T m on ZK defined by:
g 1 ¨ πK pp, gq “ πK pp, gg 1q, p P PK , g, g 1 P T m .
(3)
Then any subgroup of T m acts canonically on ZK through this action.
The following is another way to view the canonical T m -action on ZK . Recall
ZK “ pD 2 , S 1 qK “
ď`ź
σPK
vj Pσ
2
Dpjq
ˆ
ź
vj Rσ
ź
˘
1
2
Spjq
Ă
Dpjq
vj Prms
(4)
3
2
1
where Dpjq
and Spjq
are the copy of D 2 and S 1 associated to vj . Notice that
1
1
1
Dpjq
“ Spjq
˚ vj (the join of Spjq
with vj ). So we can write
ź
ď`ź
˘
1
1
Spjq
˚ vj ˆ
Spjq
(5)
ZK “ pD 2 , S 1 qK “
σPK
vj Pσ
vj Rσ
1
We can identify Spjq
with the j-th S 1 -factor in T m “ pS 1 qm . Then for any
1
pg1 , ¨ ¨ ¨ , gm q P T m , let gj act on Spjq
through left translations. This is equivalent
m
to the canonical T -action on ZK defined by (3).
For any map λ : V pKq Ñ Zr whose image spans the whole Zr , we can view the
space XpK, λq in (2) as a quotient space of ZK by a toral sbugroup of T m . Indeed,
let te1 , ¨ ¨ ¨ , em u be a unimodular basis of Zm and define a group homomorphism
ρλ : Zm ÝÑ Zr , ρλ pej q “ λpvj q, 1 ď i ď m.
The kernel of ρλ is a subgroup of Zm which determines an pm ´ rq-dimensional
toral subgroup Hλ Ă T m . It is easy to see that XpK, λq is homeomorphic to the
quotient space ZK {Hλ , where Hλ acts on ZK via the canonical T m -action. Note
that the action of Hλ on ZK is not necessarily free.
The cohomology groups of ZK can be computed via a Hochster’s type formula as
follows (see [4] or [5]). For any subset J Ă rms, let KJ denote the full subcomplex
of K obtained by restricting to J. Let k denote a field or Z below. We have
à q´|J|´1
r
H q pZK ; kq “
H
pKJ ; kq, q ě 0.
(6)
JĂrms
r ˚ pKJ ; kq is the reduced cohomology groups of KJ and |J| denotes the
where H
number of elements in J. Here we adopt the convention that
r ´1 pK∅ ; kq “ k.
H
Moreover, it is shown in [4] and [5] that there is a natural bigrading on H ˚ pZK ; kq
so that it is isomorphic to Torkrv1 ,¨¨¨ ,vm s pkrKs; kq as bigraded algebras, where krKs
is the face ring (or Stanley-Reisner ring) of K over k (also see [9]).
The cohomology rings of free quotient spaces of ZK can be reprsented in a
similar way. Suppose a subtorus H Ă T m acts freely on ZK through the canonical
action. It is shown in [14] (also see [5, Theorem 7.37]) that there is a graded
algebra isomorphism
H ˚ pZK {H; kq – TorH ˚ pBpT m {Hq;k pkrKs; kq
(7)
where BpT m {Hq is the classifying space for the principal T m {H-bundle. However,
TorH ˚ pBpT m {Hq;k pkrKs; kq is not so easy to compute in practice and it is not clear
whether there exists a Hochster’s type formula for H ˚pZP {H; kq as we have for
H ˚ pZP ; kq in (6).
4
LI YU
e~1
e~2
e~2
e~2
e~2
e~4
e~ 3
e~1
e~1
e~1
e~1
e~1
e~1
e~3
e~ 2
e~3
e~ 1
e~1
Figure 1. Examples of XpK, λα q
Remark 1.1. For the calculation of the cohomology ring structure of general
polyhedral products, the reader is referred to [2, 3, 16].
An important class of quotient spaces of moment-angle complexes are quasitoric
manifolds. Let KP be the simplicial sphere that is dual to an n-dimensional simple
convex polytope P with m facets. Then ZP “ ZKP is an pm ` nq-dimensional
closed connected manifolds, called the moment-angle manifold of P . Suppose
H – T m´n is a subgroup of T m that acts freely on ZP through the canonical
action, the quotient space ZP {H is called a quasitoric manifold over P . Quasitoric
manifolds are introduced by Davis and Januszkiewicz in [8].
In this paper, we study a special class of quotient spaces of ZK and show
that their cohomology groups can indeed be computed via some Hochster’s type
formula. These spaces are defined as follows.
‚ Let α “ tα1 , ¨ ¨ ¨ , αk u be a partition of the vertex set V pKq of a simplicial
complex K, i.e. αi ’s are disjoint subsets of V pKq with α1 Y¨ ¨ ¨Yαk “ V pKq.
‚ Let tẽ1 , ¨ ¨ ¨ , ẽk u be a basis of Zk . We define a Zk -coloring of K, denoted
by λα , which assigns ẽi to all the vertices in αi (1 ď i ď k).
Then we obtain a space XpK, λα q via construction (2), which can be thought
of as a quotient space of ZK by a rank m ´ k subtorus Hλα of T m . Note that it
is possible that the two vertices of a 1-simplex in K are assigned the same color
(see Figure 1 for example).
Let α˚ denote the trivial partition of V pKq, i.e. α˚ “ pα1 , ¨ ¨ ¨ , αm q where each
αj “ tvj u consists of only one vertex of K. Then according to our definition,
XpK, λα˚ q “ ZK .
For a non-trivial partition α of V pKq, the space XpK, λα q is a priori not a
moment-angle complex of any kind. But we will see that some topological properties of XpK, λα q are very similar to moment-angle complexes. In particular,
the cohomology groups of these spaces can also be computed by some Hochster’s
type formula as we do for moment-angle complexes.
5
Let rks “ t1, ¨ ¨ ¨ , ku. One should keep in mind the difference between rks and
the vertex set rms of K. For any simplex σ P K, let
Iα pσq :“ ti P rks ; V pσq X αi ‰ ∅u Ă rks,
which just tells us the set of colors on the vertices of σ defined by λα . Obviously
we have 0 ď |Iα pσq| ď rankpσq. For any subset L Ă rks, define
Kα,L :“ the subcomplex of K consisting of tσ P K ; Iα pσq Ă Lu.
(8)
The main results of this paper are the following two theorems.
Theorem 1.2. Let α “ tα1 , ¨ ¨ ¨ , αk u be a partition of the vertex set of a finite
simplicial complex K. Then we have group isomorphisms:
à q´|L|´1
r
H q pXpK, λα q; kq –
H
pKα,L; kq, @q ě 0.
LĂrks
According to the isomorphism in (7), the above formula gives us a way to
compute TorH ˚ pBpT m {Hq;k pkrKs; kq in the setting of Theorem 1.2. Note that the
above formula for α˚ gives (6).
In addition, it was shown in [2, Corollary 2.23] that the Hochster’s formula for
the cohomology groups of ZK follows from a stable decomposition of ZK . We
have parallel results for XpK, λα q as well.
Theorem 1.3. Let α “ tα1 , ¨ ¨ ¨ , αk u be a partition of the vertex set of a finite
simplicial complex K. There are homotopy equivalences:
ł
ΣpXpK, λα qq »
Σ|L|`2 pKα,Lq
LĂrks
where the bold Σ denotes the reduced suspension.
The paper is organized as follows. In section 2, we construct some natural cell
decomposition of XpK, λα q and use it to compute the the cohomology groups of
XpK, λα q, which leads to a proof of Theorem 1.2. In section 3, we use the same
strategy in [2] to study the stable decompositions of XpK, λα q and give a proof of
Theorem 1.3. In section 4, we show that the moment-angle complex of any finite
simplicial poset S is homotopy equivalent to XpK, λα q for some finite simplicial
complex K and a partition α of V pKq. In section 5, we generalize our results on
XpK, λα q to a wider range of spaces.
2. Cohomology groups of XpK, λα q
Suppose the vertex set of K is rms “ tv1 , ¨ ¨ ¨ , vm u. Let ∆rms be the simplex
with vertex set rms. For a partition α “ tα1 , ¨ ¨ ¨ , αk u of rms, let ∆αi denote the
6
LI YU
face of ∆rms whose vertex set is αi . Then K can be thought of as a simplicial
subcomplex of ∆rms . Next, we construct a cell decomposition of XpK, λα q.
2.1. The cell decomposition of XpK, λα q.
According to the construction of XpK, λα q in (2), it is easy to see that XpK, λα q
is homeomorphic to the quotient space of ZK by the canonical action of the
toral subgroup Hλα of T m corresponding to the subgroup of Zm “ xe1 , ¨ ¨ ¨ , em y
generated by the set
tej ´ ej 1 | vj , vj 1 P αi for some 1 ď i ď ku Ă Zm .
1
and Spj1 1 q
In other words, the action Hλα on ZK “ PK ˆ T m { „ identifies the Spjq
1
1
in T m “ Sp1q
ˆ ¨ ¨ ¨ ˆ Spmq
whenever vj and vj 1 belong to the same αi . Considering
the partition α of the vertex set of K, we can rewrite the decomposition of ZK
in (5) as:
ź
ď ´` ź
˘ ź 1 ¯
1
Spjq ˚ vj ˆ
Spjq .
(9)
ZK “
σPK
vj Rσ
iPIα pσq vj PV pσqXαi
Then with respect to this decomposition of ZK , we obtain a decomposition of
XpK, λα q by Lemma 2.1 below.
¯
ź
ď´ ź `
˘
1
1
Spiq
˚
˚ v ˆ
Spiq
XpK, λα q “
σPK
“
σPK
vPV pσqXαi
iPIα pσq
ď´ ź
iPIα pσq
1
Spiq
αi
˚ pσ X ∆ q ˆ
iPrkszIα pσq
ź
iPrkszIα pσq
1
Spiq
¯
Ă
ź
1
Spiq
˚ ∆αi
(10)
iPrks
1
is a copy of S 1 corresponding to i P rks, which can be considered as
where Spiq
the join of S 1 with the empty face of ∆αi .
Lemma 2.1. If we identify all the S 1 factors in a product pS 1 ˚v1 qˆ¨ ¨ ¨ˆpS 1 ˚vs q,
the quotient space pS 1 ˚v1 qˆ¨ ¨ ¨ˆpS 1 ˚vs q{ „ is homeomorphic to S 1 ˚pv1 ˚¨ ¨ ¨˚vs q.
Note that v1 ˚¨ ¨ ¨˚vs can be identified with a simplex whose vertex set is tv1 , ¨ ¨ ¨ , vs u.
Proof. The points in pS 1 ˚ v1 q ˆ ¨ ¨ ¨ ˆ pS 1 ˚ vs q can be written as
`
˘
pt1 v1 ` p1 ´ t1 qx1 q, ¨ ¨ ¨ , pts vs ` p1 ´ ts qxs q , xi P S 1 , 0 ď ti ď 1, 1 ď i ď s.
In we identify all the S 1 factors in the above product, the points in the quotient
space can be written as
`
˘
Pt1 ,¨¨¨ ,ts ,x “ pt1 v1 ` p1 ´ t1qxq, ¨ ¨ ¨ , pts vs ` p1 ´ tsqxq , x P S 1 , 0 ď ti ď 1, 1 ď i ď s.
Then mapping any Pt1 ,¨¨¨ ,ts ,x to the point P̂t1 ,¨¨¨ ,ts ,x in S 1 ˚ pv1 ˚ ¨ ¨ ¨ ˚ vs q below
t1
ts
1 ´ t1 ´ ¨ ¨ ¨ ´ ts
P̂t1 ,¨¨¨ ,ts ,x “ v1 ` ¨ ¨ ¨ ` vs `
x
s
s
s
defines a homeomorphism from pS 1 ˚v1 qˆ¨ ¨ ¨ˆpS 1 ˚vs q{ „ to S 1 ˚pv1 ˚¨ ¨ ¨˚vs q.
7
Remark 2.2. We see from (10) that the building blocks of XpK, λα q are spaces
obtained by mixtures of Cartesian products and joins of some simple spaces (i.e.
points and S 1 ). The building blocks of polyhedral products pX, AqK , however,
only involve Cartesian products of spaces. In addition, we have polyhedral join
(see [1]) and polyhedral smash product (see [2]) whose building blocks only involve
joins and smash products, respectively. It should be interesting to study spaces
whose building blocks involve mixtures of Cartesian products, joins and smash
products.
To obtain a cell decomposition of XpK, λα q from (10), we need to choose a cell
decomposition of the torus T k . First of all, a circle S 1 “ tz P C ; |z| “ 1u has
a natural cell decomposition te0 , e1 u where e0 “ t1u P S 1 and e1 “ S 1 ze0 . We
ś
1
consider T k as the product ki“1 Spiq
and equip T k with the product cell structure
(see [10, 3.B]). Then the cells in T k can be indexed by subsets of rks “ t1, ¨ ¨ ¨ , ku.
More specifically, any subset L Ă rks determines a unique cell UL in T k where
ź
ź
UL “
e1piq ˆ
e0piq , dimpUL q “ |L|.
iPL
iPrkszL
1
for each i P rks.
Here e0piq , e1piq denote the cells in Spiq
ś
1
Observe that for any σ P K, iPIα pσq Spiq
˚ pσ X ∆αi q is homeomorphic to a
closed ball of dimension rankpσq ` |Iα pσq|. Now for each L Ă rkszIα pσq, define
´ ź
¯
1
αi
Bpσ,Lq :“ the relative interior of
Spiq ˚ pσ X ∆ q ˆ UL .
iPIα pσq
Then from (10), a cell decomposition of XpK, λα q is given by:
Bα pKq :“ tBpσ,Lq | σ P K, L Ă rkszIα pσqu
(11)
Note that Bpσ,Lq is an open cell of dimension rankpσq ` |Iα pσq| ` |L|.
2.2. The cochain complex of XpK, λα q.
For any coefficient ring k, let C ˚ pXpK, λα q; kq be the cellular cochain complex
corresponding to the cell decomposition Bα pKq. If we want to write the boundary
maps of the cochains in C ˚ pXpK, λα q; kq, we need to put orientations on the base
elements. To do this, we need to first assign orientations to all the simplicies of
K. For convenience, we put a total ordering ă on the vertex set tv1 , ¨ ¨ ¨ , vm u of
∆rms so that they appear in the increasing order in α1 until αk . In other words,
for any 1 ď i ă k all the vertices of ∆αi have less order than the vertices in ∆αi`1 .
Moreover, the vertex-ordering of ∆rms induces a vertex-ordering of any simplex
8
LI YU
ω P ∆rms , which determines an orientation of ω. Then the boundary of ω is
ÿ
Bω “
εpσ, ωq σ.
(12)
σĂω
dimpσq“dimpωq´1
Here if V pωq “ V pσq Y tvu, then εpσ, ωq is equal to p´1qlpv,ωq where lpv, ωq is the
number vertices of ω that are less than v with respect to the vertex-ordering ă.
The following definition is very useful for us to simplify our argument later.
Definition 2.3 (Simplex with a ghost face). For any m ě 1, let 0̂ denote the
empty face of ∆rms (distinguished from the empty simplex 0̂ in K). In addition,
we attach a ghost face ´1̂ to ∆rms with the following conventions.
‚ dimp0̂q “ dimp´1̂q “ ´1, rankp0̂q “ rankp´1̂q “ 0.
‚ The interiors of 0̂ and ´1̂ are themselves.
‚ The boundary of any vertex of ∆rms is 0̂.
‚ The boundaries of 0̂ and ´1̂ are empty,
p rms to denote ∆rms with the ghost face ´1̂.
In the rest of the paper we use ∆
Let te0 , e1 u be the cell decomposition of S 1 where dimpe0 q “ 0, dimpe1 q “ 1,
and e0 , e1 are both oriented. Then given an orientation of each face of ∆rms , we
obtain an oriented cell decomposition of S 1 ˚ ∆rms by
te0 u, te1 u, tS 1 ˚ σ ˝ | σ is a nonempty simplex in ∆rms u.
If we formally define S 1 ˚ 0̂ “ e1 and S 1 ˚ ´1̂ “ e0 , we can write a basis of the
p rms u where the orientation
cellular chain complex C˚ pS 1 ˚∆rms ; kq as tS 1 ˚σ ˝ | σ P ∆
1
˝
of S ˚ σ is determined canonically by the orientations of S 1 and σ.
p rms u be a basis for the cellular cochain complex C ˚ pS 1 ˚ ∆rms ; kq,
Let ty σ | σ P ∆
σ
where y is the dual of S 1 ˚ σ ˝ . For any nonempty simplex σ in ∆rms ,
ÿ
dpy σ q “
εpσ, τ q ¨ y τ .
(13)
σĂτ
dimpτ q“dimpσq`1
In addition, we have
dpy ´1̂q “ 0, dpy 0̂q “
ÿ
yv.
(14)
vPrms
Note that y ´1̂ and y 0̂ are cochains in dimension 0 and 1, respectively.
By setting the σ in (10) to the big simplex ∆rms , we obtain a homeomorphism
ź
Xp∆rms , λα q –
S 1 ˚ ∆αi .
iPrks
9
Let Xp∆rms , λα q be equipped with the product cell structure of each S 1 ˚ ∆αi .
The corresponding cellular cochain complex C ˚ pXp∆rms , λα q; kq has a basis
p αi , 1 ď i ď ku.
tyΦ “ y σ1 ˆ ¨ ¨ ¨ ˆ y σk ; Φ “ pσ1 , ¨ ¨ ¨ , σk q, where σi P ∆
We need to introduce two more notations for our argument below.
‚ For any vertex v of K and a partition α “ tα1 , ¨ ¨ ¨ , αk u of V pKq, let
iα pvq P rks denote the index so that v belongs to the subset αiα pvq .
‚ For any i P L Ă rks, define κpi, Lq “ p´1qrpi,Lq , where rpi, Lq is the number
of elements in L less than i. Moreover, for any simplex σ P Kα,L we define
ź
κpσ, Lq :“
κpiα pvq, Lq.
(15)
vPV pσq
So if V pωq “ V pσq Y tvu, we have κpω, Lq “ κpσ, Lq ¨ κpiα pvq, Lq.
The differential of yΦ in C ˚ pXp∆rms , λα q; kq is given by:
ÿ
dpyΦ q :“
ιpΦ, σi q y σ1 ˆ ¨ ¨ ¨ ˆ dy σi ˆ ¨ ¨ ¨ ˆ y σk ,
where ιpΦ, σi q “ p´1q
1ďiďk
ři´1
σl
l“1 dimpy q
(16)
(17)
. For a simplex σ P ∆rms and J Ă rkszIα pσq, let
ΦJσ “ pσ1J , ¨ ¨ ¨ , σkJ q
$
α
’
&σ X ∆ i ,
p αi ,
where σiJ “ 0̂ P ∆
’
%´1̂ P ∆
p αi ,
i P Iα pσq;
i P J;
i P rkszpIα pσq Y Jq;
So by our definition when Φ “ ΦJσ “ pσ1J , ¨ ¨ ¨
$
J
’
&rankpσi q ` 1,
J
dimpy σi q “ 1,
’
% 0,
(18)
, σkJ q, we have
i P Iα pσq;
i P J;
i P rkszpIα pσq Y Jq.
For Φ “ ΦJσ , the formula (17) reads:
ÿ
J
J
J
J
dpyΦσ q “
ιpΦJσ , σiJ q y σ1 ˆ ¨ ¨ ¨ ˆ dy σi ˆ ¨ ¨ ¨ ˆ y σk
1ďiďk
(13)
“
ÿ
1ďiďk
J
ιpΦJσ , σiJ q y σ1 ˆ ¨ ¨ ¨ ˆ
´
ÿ
σiJ Ăτ Ă∆αi
dimpτ q“dimpσq`1
¯
J
εpσiJ , τ q ¨ y τ ˆ ¨ ¨ ¨ ˆ y σk .
Note that if V pωq “ V pσq Y tvu where v P αi , then ωiJ “ σiJ Y tvu and we have
κpi, Iα pσq Y Jq ¨ εpσ, ωq “ ιpΦJσ , σiJ q ¨ εpσiJ , τ q.
10
LI YU
J
This is because for each i P rkszpIα pσq Y Jq, dimpy σi q “ 0. So we obtain
J
dpyΦσ q “
ÿ
σĂω, Iα pωqĂIα pσqYJ
dimpωq“dimpσq`1
`
˘
Iα pσqYJzIα pωq
κ iα pωzσq, Iα pσq Y J ¨ εpσ, ωq yΦω
(19)
where ωzσ denote the only vertex of ω that is not in σ.
J
For the simplicial complex K Ă ∆rms and any simplex σ P K, yΦσ is a cochain
of dimension rankpσq ` |Iα pσq| ` |J| in C ˚ pXpK, λα q; kq that is dual to the cell
Bpσ,Jq (see (11)). So C ˚ pXpK, λα q; Zq has a basis
J
tyΦσ ; σ P K, J Ă rkszIα pσqu.
J
Note that for any σ P K, the differential dpyΦσ q in C ˚ pXpK, λα q; kq is only the
Iα pσqYJzIα pωq
sum of those terms yΦω
on the right hand side of (19) with ω P K.
Proof of Theorem 1.2.
For any subset L Ă rks, let C ˚,L pXpK, λα q; kq denote the k-submodule of
˚
C pXpK, λα q; kq generated by the following set
tBpσ,Jq | Iα pσq Y J “ L, σ P K, J Ă rkszIα pσqu.
From the differential of C ˚ pXpK, λα q; kq in (19), we see that C ˚,L pXpK, λα q; kq
is actually a cochain subcomplex of C ˚ pXpK, λα q; kq. We denote its cohomology
groups by H ˚,L pXpK, λα q; kq. Then we have the following decompositions:
H ˚ pXpK, λα q; kq “
à
H ˚,L pXpK, λα q; kq.
(20)
LĂrks
For any subset L Ă rks, let C ˚ pKα,L ; kq denote the simplicial cochain complex
of Kα,L. For any simplex σ P K, let σ ˚ be the cochain dual to σ in C ˚ pKα,L; kq.
Then we have the following isomorphism of k-modules:
ϕLα : C ˚ pKα,L ; kq ÝÑ C ˚,L pXpK, λα q; kq
LzIα pσq
σ ˚ ÞÝÑ κpσ, Lq yΦσ
11
LzIα pσq
Note dimpyΦσ
q “ rankpσq ` |L|. Moreover, ϕLα is actually a chain complex
isomorphism. Indeed by the differential of C ˚ pXpK, λα q; kq shown in (19),
`
LzIα pσq ˘
dpϕLα pσ ˚ qq “ d κpσ, Lq yΦσ
ÿ
LzIα pωq
“ κpσ, Lq
κpiα pωzσq, Lq ¨ εpσ, ωq yΦω
σĂωPK, Iα pωqĂL
dimpωq“dimpσq`1
ÿ
“
LzIα pωq
κpω, Lq ¨ εpσ, ωq yΦω
“ ϕLα pdσ ˚ q.
σĂωPK, Iα pωqĂL
dimpωq“dimpσq`1
The third ““” uses the relation κpω, Lq “ κpσ, Lq ¨ κpiα pωzσq, Lq (see (16)). So we
have an additive isomorphism of cohomology groups
r q pKα,L; kq – H q`|L|`1,LpXpK, λα q; kq.
H
Let ϕα “
à
ϕLα :
LĂrks
à
r q´|L|´1pKα,L ; kq ÝÑ H q pXpK, λα q; kq.
H
LĂrks
Then ϕα is an isomorphism that satisfies our requirement.
Example 1 (Balanced Simplicial Complex and Pullbacks from the Linear Model).
An pn ´ 1q-dimensional simplicial complex K is called balanced if there exists
a map φ : V pKq Ñ rns “ t1, ¨ ¨ ¨ , nu such that if tv, v 1 u is an edge of K, then
φpxq ‰ φpyq (see [15]). We call φ an n-coloring on K. It is easy to see that K is
balanced if and only if K admits a non-degenerate simplicial map onto ∆n´1 . In
fact, if we identify the vertex set of ∆n´1 with rns, any n-coloring φ on K induces
a non-degenerate simplicial map from K to ∆n´1 which sends a simplex σ P K
with V pσq “ tvi1 , ¨ ¨ ¨ , vis u to the face of ∆n´1 spanned by tφpvi1 q, ¨ ¨ ¨ , φpvis qu.
Suppose φ : V pKq Ñ rns is an n-coloring of an pn ´ 1q-dimensional simplicial
complex K. Let te1 , ¨ ¨ ¨ , en u be a basis of Zn . Then φ uniquely determines a
Zn -coloring λφ : V pKq Ñ Zn where λφ pvq “ eφpvq . The space XpK, λφ q is called
a pullback from the linear model in [8, Example 1.15]. On the other hand, we
have a partition of V pKq defined by αφ :“ tφ´1 p1q, ¨ ¨ ¨ φ´1 pnqu. By our notation in
section 1, we have XpK, λφ q “ XpK, λαφ q. Then by Theorem 1.2, the cohomology
groups of XpK, λφ q can be computed by
à q´|L|´1
r
H
pKαφ ,L; kq, @q ě 0.
H q pXpK, λφ q; kq –
LĂrns
12
LI YU
3. Stable decompositions of XpK, λα q
It is shown in [2] that the stable homotopy type of a polyhedral product pX, AqK
is a wedge of spaces, which implies corresponding homological decompositions of
pX, AqK . In this section, we prove a parallel result for XpK, λα q. Our argument
proceeds along the same line as [2].
Let α “ tα1 , ¨ ¨ ¨ , αk u be a partition of the vertex set V pKq of K. For any
1
“ e0piq Y e1piq to be e0piq . So
i P rks, choose the base-point of Spiq
1
‚ e0piq is a base-point of Spiq
˚ τ for any simplex τ P K, and
‚ pe0piq , ¨ ¨ ¨ , e0pis q q is a base-point of Spi1 1 q ˆ ¨ ¨ ¨ ˆ Spi1 s q for any i1 , ¨ ¨ ¨ , is P rks.
Then for any subset L Ă rks and any simplex σ P K, it is meaningful to define
ľ
ľ
1
1
S1
Wα,L
pσq :“
Spiq
˚ pσ X ∆αi q ^
Spiq
iPIα pσqXL
iPLzpIα pσqXLq
Ź
where ^ and
denote the smash product with respect to the based spaces. So
S1
for L “ ti1 , ¨ ¨ ¨ , is u, the base-point of Wα,L
pσq is pe0piq , ¨ ¨ ¨ , e0pis q q. We adopt the
convention that the smash product of a space with the empty space is empty.
S1
Then the following lemma is immediate from the definition of Wα,L
pσq.
Lemma 3.1. For any subset L Ă rks and any simplex σ P K, we have
1
1
S
S
(i) Wα,L
pσq “ Wα,L
pσ X Kα,L q,
1
S
pσq is contractible whenever Iα pσq X L ‰ ∅,
(ii) Wα,L
Ź
1
S
1
p0̂q “ iPL Spiq
(iii) Wα,L
– S |L| .
For any simplex σ P K, define
ź
1
Dα pσq :“
Spiq
˚ pσ X ∆αi q ˆ
iPIα pσq
ź
1
Spiq
.
(21)
iPrkszIα pσq
Note that all the Dα pσq have the same base-point pe0p1q , ¨ ¨ ¨ , e0pkq q, which is the
base-point of XpK, λα q.
By [2, Theorem 2.8], there are natural homotopy equivalences
´ ł
¯
S1
ΣpDα pσqq » Σ
Wα,L
pσq
LĂrks
Ž
where Σ denotes the reduced suspension and
denotes the wedge sum with
S1
respect to the base-point of Wα,Lpσq. Now let
ł
S1
Eα pσq :“
Wα,L
pσq.
LĂrks
13
Let CatpKq denote the face category of K whose objects are simplices σ P K
and there is a morphism from σ to τ whenever σ Ď τ . Then we can consider
Dα and Eα as functors from CatpKq to the category CW˚ of connected, based
CW-complexes and based continuous maps. It is clear that
ď
XpK, λα q “
Dα pσq “ colimpDα pσqq.
σPK
For any subset L “ tl1 , ¨ ¨ ¨ , ls u Ă rks, define
ď
ď
S1
S1
p α,L, λα q :“
Wα,L
pσq.
XpK
Wα,L
pσq “
σPK
σPKα,L
1
S
For a fixed L Ă rks, all the spaces tWα,L
pσq, σ P Kα,L u share a base-point,
p α,L , λα q. So XpK
p α,L, λα q is the colimit
which then defines the base-point of XpK
1
S
of Wα,L
pσq over the face category CatpKα,Lq of Kα,L. In addition, we clearly have
ł
p α,L , λα q.
colimpEα pσqq “
XpK
LĂrks
Since the suspension commutes with the colimits up to homotopy equivalence
(see [2, Theorem 4.3]), we obtain the following homotopy equivalences.
ΣpXpK, λα qq » colimpΣpDα pσqqq » colimpΣpEα pσqqq
` ł
˘
p α,L, λα q .
»Σ
XpK
(22)
LĂrks
Definition 3.2 (Order Complex). Given a poset (partially ordered set) pP, ăq,
the order complex ∆pPq is the simplicial complex with vertices given by the set
of points of P and k-simplices given by the ordered pk ` 1q-tuples pp1 , p2 , ..., pk`1q
in P with p1 ă p2 ă . . . ă pk`1 .
Lemma 3.3. For any L Ă rks, there is a homotopy equivalence:
ł
S1
p α,L, λα q »
|∆ppKα,Lqăσ q| ˚ Wα,L
pσq.
XpK
σPKα,L
Here ∆ppKα,Lqăσ q is the order complex of the poset tτ P Kα,L | τ Ľ σu whose
order is the reverse inclusion, and |∆ppKα,L qăσ q| is the geometric realization of
∆ppKα,L qăσ q.
Proof. Note that the natural inclusion S 1 ãÑ S 1 ˚ ∆rms is null-homotopic for any
m ě 1. Then the same argument as in the proof of [2, Theorem 2.12] shows that
S1
S1
there is a homotopy equivalence HL pσq : Wα,L
pσq Ñ Wα,L
pσq for each simplex
14
LI YU
σ P Kα,L so that the following diagram commutes for any τ Ď σ P Kα,L ,
1
S
Wα,L
pτ q
HL pτ q
/
1
S
Wα,L
pτ q
(23)
cσ,τ
ζσ,τ
1
S
Wα,L
pσq
HL pσq
/
1
S
Wα,L
pσq
where ζσ,τ is the natural inclusion and cσ,τ is the constant map to the base-point.
Then by [2, Theorem 4.1] and [2, Theorem 4.2], there is a homotopy equivalence
ł
S1
S1
p α,L , λα q “ colimpWα,L
|∆ppKα,Lqăσ q| ˚ Wα,L
pσq.
XpK
pσqq »
σPKα,L
So the lemma is proved.
Proof of Theorem 1.3. Putting the homotopy equivalences in the equation (22)
and Lemma 3.3 together gives us a homotopy equivalence:
¯
´ ł ł
S1
|∆ppKα,L qăσ q| ˚ Wα,L
pσq .
ΣpXpK, λα qq » Σ
LĂrks σPKα,L
Moreover, the space in the left hand side can be simplified by the following facts.
1
S
‚ Wα,L
pσq is contractible whenever σ ‰ 0̂ P Kα,L (see Lemma 3.1(ii)).
1
‚ ∆ppKα,Lqă0̂ q is isomorphic to the barycentric subdivision Kα,L
of Kα,L
as simplicial complexes. So the geometric realization |∆ppKα,L qă0̂ q| is
homeomorphic to that of Kα,L.
Then by Lemma 3.1, we have the following homotopy equivalences:
´ ł
¯
ł
ΣpXpK, λα qq » Σ
Σ|L|`2p|Kα,L |q.
|Kα,L | ˚ S |L| »
LĂrks
LĂrks
So Theorem 1.3 is proved.
4. A description of moment-angle complexes of simplicial posets
A poset (partially ordered set) S with the order relation ď is called simplicial
if it has an initial element 0̂ and for each σ P S the lower segment
r0̂, σs “ tτ P S : 0̂ ď τ ď σu
is the face poset of a simplex.
For each σ P S we assign a geometric simplex ∆σ whose face poset is r0̂, σs,
and glue these geometric simplices together according to the order relation in S.
We get a cell complex ∆S in which the closure of each cell is identified with a
simplex preserving the face structure, and all attaching maps are inclusions. We
15
call ∆S the geometric realization of S. For convenience, we still use ∆σ to denote
the image of each geometric simplex ∆σ in ∆S . Then ∆σ is a maximal simplex
of ∆S if and only if σ is a maximal element of S.
The notion of moment-angle complex ZS associated to a simplicial poset S is
introduced in [11] where ZS is defined via the categorical language. Note that the
barycentric subdivision also makes sense for ∆S . Let PS denote the cone of the
barycentric subdivision of ∆S . Let the vertex set of ∆S be V p∆S q “ tu1 , ¨ ¨ ¨ , uk u.
Let λS : V p∆S q Ñ Zk be a map so that tλS pui q, 1 ď i ď ku is a unimodular basis
of Zk . Then we can construct ZS from PS and λS via the same rule in (2). So
we also denote ZS by XpS, λS q.
Define a map λ : rms “ tv1 , ¨ ¨ ¨ , vm u Ñ Zm by λpvi q “ ei , 1 ď i ď m, where
te1 , ¨ ¨ ¨ , em u is a unimodular basis of Zm . We identify ∆rms with ∆rms ˆ t0u
in ∆rms ˆ r0, 1s (considered as a product of simplices). Let the vertex set of
1
∆rms ˆ r0, 1s be tv1 , ¨ ¨ ¨ , vm , v11 , ¨ ¨ ¨ , vm
u where vi1 “ vi ˆ t1u, 1 ď i ď m. Define
1
1
a map λ̃ : tv1 , ¨ ¨ ¨ , vm , v1 , ¨ ¨ ¨ , vm u Ñ Zm by
λ̃pvi q “ λ̃pvi1 q “ ei , 1 ď i ď m.
It is clear that Xp∆rms , λq can be considered as a subspace Xp∆rms ˆ r0, 1s, λ̃q.
Lemma 4.1. There is a canonical deformation retraction from Xp∆rms ˆr0, 1s, λ̃q
to Xp∆rms , λq “ Z∆rms .
Proof. Any m-simplex in ∆rms ˆ r0, 1s can be written as
1
σjm “ rv1 , ¨ ¨ ¨ , vj , vj1 , ¨ ¨ ¨ , vm
s, 1 ď j ď m.
m
1
1
m
Note that σjm X σj`1
“ rv1 , ¨ ¨ ¨ , vj , vj`1
, ¨ ¨ ¨ , vm
s. So σ1m , σ2m , ¨ ¨ ¨ , σm
is a shelling
rms
of ∆ ˆ r0, 1s. By the cell decomposition (10), we have
ź
1
1
1
Xp∆rms , λq “ pSp1q
˚ v1 q ˆ ¨ ¨ ¨ ˆ pSpmq
˚ vm q Ă
Spjq
˚ rvj , vj1 s.
1ďjďm
Xp∆rms ˆ r0, 1s, λ̃q “
ď
1ďjďm
Bj “
Bj Ă
ź
1
Spjq
˚ rvj , vj1 s, where
1ďjďm
1
1
1
1
1
1
1
˚vj`1
qˆ¨ ¨ ¨ˆpSpmq
˚vm
q.
˚rvj , vj1 sqˆpSpj`1q
pSp1q
˚v1 qˆ¨ ¨ ¨ˆpSpj´1q
˚vj´1 qˆpSpjq
There is a canonical deformation retraction from ∆rms ˆ r0, 1s to ∆rms along
the above shelling of ∆rms ˆ r0, 1s as follows. We first compress the edge rv1 , v11 s
to v1 , which induces a deformation retraction
1
1
σ1m “ rv1 , v11 , ¨ ¨ ¨ , vm
s ÝÑ rv1 , v21 ¨ ¨ ¨ , vm
s “ σ1m X σ2m .
Ť
Ť
It compresses ∆rms ˆ r0, 1s “ 1ďjďm σjm to 2ďjďm σjm . Next, we compress the
edge rv2 , v21 s to v2 which induces a deformation retraction from σ2m to σ2m X σ3m
16
LI YU
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
Figure 2. Canonical deformation retraction from ∆rms ˆ r0, 1s to ∆rms
Ť
Ť
and hence compresses 2ďjďm σjm to 3ďjďm σjm , and so on. After m steps of
retractions, ∆rms ˆ r0, 1s is deformed to ∆rms ˆ t0u “ ∆rms (see Figure 2).
Using the above
of ∆rms ˆ r0, 1s, we obtain parallel deformation
ś retractions
1
retractions from 1ďjďm Spjq ˚ rvj , vj1 s to Xp∆rms , λq in m steps. The j-th step is
1
1
to compress Spjq
˚rvj , vj1 s to Spjq
˚vj along the edge rvj , vj1 s, which compresses Bj to
Bj X Bj`1. Then starting from the first step, we obtain a sequence of retractions
ď
ď
ď
step 1
step 2
step m-1
Xp∆rms ˆ r0, 1s, λ̃q “
Bj ÝÑ
Bj ÝÑ ¨ ¨ ¨ ÝÑ
Bj ÝÑ Bm
1ďjďm
2ďjďm
m´1ďjďm
step m
1
1
˚ vm q “ Xp∆rms , λq.
ÝÑ pSp1q
˚ v1 q ˆ ¨ ¨ ¨ ˆ pSpmq
The above deformation process is canonical since it only depends on the ordering
of the vertices of ∆rms .
The canonical deformation retraction from Xp∆rms ˆ r0, 1s, λ̃q to Xp∆rms , λq in
the above lemma will serve as a model of homotopies in our argument below.
Theorem 4.2. For any finite simplicial poset S, there always exists a finite
simplicial complex K and a partition α of V pKq so that ZS is homotopy equivalent
to XpK, λα q.
Proof. We can construct K and α in the following way. Let p : ∆S ˆ r0, ns Ñ ∆S
be the projection where ∆S is identified with ∆S ˆ t0u, and n is a large enough
rσ Ă
integer. For each maximal simplex ∆σ Ă ∆S , we can choose a simplex ∆
∆S ˆ tlσ u for some 0 ď lσ ď n so that
r σ simplicially isomorphically onto ∆σ .
‚ p maps ∆
rσ X ∆
r τ “ ∅ for any maximal elements σ and τ in S.
‚ ∆
r σ a horizontal lifting of ∆σ . We consider ∆σ ˆ r0, ns as the Cartesian
We call ∆
product of ∆σ and r0, ns as simplicial complexes (see [5, Construction 2.11]),
where r0, ns is considered as a 1-dimensional simplicial complex with vertices
17
.
.
.
Figure 3. A stretch of a simplicial poset
t0, ¨ ¨ ¨ , nu and the set of 1-simplices tri, i ` 1s, 0 ď i ď n ´ 1u. If σ and τ are both
maximal, ∆σ X ∆τ is the geometric realization of σ ^ τ (the greatest common
lower bound of σ and τ ). Now define
¯
¯ď´ ď
´ ď
σ
τ
σ
r
p∆ X ∆ q ˆ rlσ , lτ s ,
∆
K“
σPS
maximal
σ,τ PS
maximal
lσ ălτ
where rlσ , lτ s is considered as a simplicial subcomplex of r0, ns. Then by our
construction, K is clearly a finite simplicial complex, called a stretch of ∆S (see
Figure 3 for example). Let V p∆S q “ tu1 , ¨ ¨ ¨ , uk u be the vertex set of ∆S . Define
a partition α “ tα1 , ¨ ¨ ¨ , αk u of V pKq by
αi “ tv P V pKq | ppvq “ ui u, 1 ď i ď k.
Then we get a Zk -coloring λα on K. In the following we show that the space
XpK, λα q is homotopy equivalent to XpS, λS q “ ZS .
For a pair of maximal elements σ, τ P S, we have a decomposition
ď ď
ď
∆ω ˆ rs, s ` 1s.
p∆σ X ∆τ q ˆ rs, s ` 1s “
p∆σ X ∆τ q ˆ rlσ , lτ s “
lσ ďsălτ
lσ ďsălτ ωPσ^τ
Given the shelling of each ∆ω ˆ rs, s ` 1s as we do for ∆rms ˆ r0, 1s in the proof
of Lemma 4.1, we obtain a canonical shelling of p∆σ X ∆τ q ˆ rlσ , lτ s. If for
all pairs of maximal elements σ, τ P S we do the deformation retractions from
p∆σ X ∆τ q ˆ rlσ , lτ s to p∆σ X ∆τ q ˆ tlσ u in K along their canonical shellings, we
obtain a space that can be identified with ∆S in the end. Note that all these
retractions are caused by compressing tvu ˆ rs, s ` 1s to tvu ˆ tsu step by step
for each vertex v P ∆σ X ∆τ . So for different pairs of maximal elements of S,
the retractions we constructed are compatible with each other and hence can be
done simultaneously at each tvu ˆ rs, s ` 1s. In addition, the Zm -coloring λα on
K naturally induces a Zm -coloring on the space in each step of the deformation
18
LI YU
and eventually recovers λS on ∆S in the end. So by applying Lemma 4.1 to every
∆ω ˆrs, s`1s Ă p∆σ X∆τ qˆrlσ , lτ s, we see that the above deformation retractions
on K induce a homotopy equivalence from XpK, λα q to XpS, λS q “ ZS . This
proves the theorem.
For any subset L Ă V pSq, let SL denote the full subposet of S with vertex set
L. By our construction of the stretch K of S in the proof of Theorem 1.2, the
geometric realization ∆SL of SL is a deformation retraction of the subcomplex
Kα,L of K. So by Theorem 4.2 and Theorem 1.2, we derive that for any q ě 0,
à
à
r q´|L|´1pKα,L; kq –
r q´|L|´1p∆SL ; kq.
H q pZS ; kq – H q pXpK, λα q; kq –
H
H
LĂV pSq
LĂV pSq
The above equality can also be derived from [11, Theorem 3.5] easily.
5. Some generalizations
We can generalize our results on XpK, λα q to a wider range of spaces as follows.
1
For a partition α “ tα1 , ¨ ¨ ¨ , αk u of V pKq, we can replace the Spiq
in (10) by a
d1
dk
sequence of spheres S “ pS , ¨ ¨ ¨ , S q and define
¯ ź
ź
ď´ ź
S di ˚ ∆αi .
S di ˚ pσ X ∆αi q ˆ
S di Ă
XpK, α, Sq “
σPK
iPIα pσq
0
iPrkszIα pσq
iPrks
0
In particular when S “ pS , ¨ ¨ ¨ , S q, the XpK, α, Sq is a quotient space of the real
moment-angle complex of K by the action of some Z2 -torus. We have the following
two theorems which are parallel to Theorem 1.2 and Theorem 1.3, respectively.
Theorem 5.1. For any coefficients k, there is a k-module isomorphism:
à q´1´ř d
r
iPL i pK
H q pXpK, α, Sq; kq –
H
α,L ; kq, @q ě 0.
LĂrks
Theorem 5.2. There is a homotopy equivalence
ł
ř
Σp iPL di q`2 pKα,Lq.
ΣpXpK, α, Sqq »
LĂrks
The proofs of the above two theorems are completely parallel to the proofs of
Theorem 1.2 and Theorem 1.3. So we leave them to the reader. There is only one
technical point in the proof here. The definition of κpi, Lq and κpσ, Lq (see (15))
in the proof of Theorem 1.2. need to be modified to be adapted to Theorem 5.1.
For XpK, α, Sq we should redefine κpi, Lq as follows and adjust the definition of
κpσ, Lq accordingly.
ÿ
κpi, Lq “ p´1qrS pi,Lq , where rS pi, Lq “
dj , @i P L Ă rks.
jPL, jăi
19
Remark 5.3. When α˚ is the trivial partition of V pKq, XpK, α˚ , Sq is nothing but
the polyhedral product KpD,Sq where pD, Sq “ tpD d1 `1 , S d1 q, ¨ ¨ ¨ , pD dk `1 , S dk qu. In
this special case, Theorem 5.1 coincides with [7, Theorem 4.2].
References
[1] A. Ayzenberg, Substitutions of polytopes and of simplicial complexes, and multigraded betti
numbers, Trans. Moscow Math. Soc. 74 (2013), no. 2, 175–202.
[2] A. Bahri, M. Bendersky, F. Cohen and S. Gitler, The Polyhedral Product Functor: a
method of computation for moment-angle complexes, arrangements and related spaces, Adv.
Math. 225 (2010) no. 3, 1634–1668.
[3] A. Bahri, M. Bendersky, F. Cohen and S. Gitler, Cup-products for the polyhedral product
functor, Math. Proc. Cambridge Philos. Soc. 153 (2012), no. 3, 457–469.
[4] I. V. Baskakov, Cohomology of K-powers of spaces and the combinatorics of simplicial
divisions, (Russian) Uspekhi Mat. Nauk 57 (2002), no. 5 (347), 147–148; translation in
Russian Math. Surveys 57 (2002), no. 5, 989–990.
[5] V. M. Buchstaber and T. E. Panov, Torus actions and their applications in topology and
combinatorics, University Lecture Series, 24. American Mathematical Society, Providence,
RI, 2002.
[6] V. M. Buchstaber and T.E. Panov, Toric Topology, Mathematical Surveys and Monographs, Vol. 204, American Mathematical Society, Providence, RI, 2015,
[7] X. Cao and Z. Lü, Möbius transform, moment-angle complexes and Halperin-Carlsson
conjecture, J. Algebraic Combin. 35 (2012), no. 1, 121–140.
[8] M. W. Davis and T. Januszkiewicz, Convex polytopes, Coxeter orbifolds and torus actions,
Duke Math. J. 62 (1991), no.2, 417–451.
[9] M. Franz, On the integral cohomology of smooth toric varieties, Proc. Steklov Inst. Math.
252 (2006), 53–62.
[10] A. Hatcher, Algebraic topology, Cambridge Univeristy Press, 2002.
[11] Z. Lü and T. Panov, Moment-angle complexes from simplicial posets, Cent. Eur. J. Math.
9 (2011), no. 4, 715–730.
[12] M. Masuda and T. Panov, On the cohomology of torus manifolds, Osaka J. Math. 43
(2006), 711–746.
[13] T. Panov, Cohomology of face rings, and torus actions, in “Surveys in Contemporary
Mathematics”, London Math. Soc. Lecture Notes Series, Vol. 347, Cambridge, U.K. 2008,
165–201.
[14] T. Panov, On the cohomology of quotients of moment-angle complexes, Uspekhi Mat. Nauk
70 (2015), no.4, 209–210 (Russian); Russian Math. Surveys 70 (2015), no.4, 779–781 (English translation).
[15] R. P. Stanley, Combinatorics and commutative algebra, Second edition, Progress in Mathematics, 41. Birkhäuser Boston, 1996.
[16] X. J. Wang and Q. B. Zheng, The homology of simplicial complements and the cohomology
of polyhedral products, Forum Math. 27 (2015), no. 4, 2267–2299.
Department of Mathematics and IMS, Nanjing University, Nanjing, 210093,
P.R.China
E-mail address: [email protected]
| 0math.AC
|
Elementary matrix factorizations over Bézout domains
arXiv:1801.02369v1 [math.AC] 8 Jan 2018
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
Center for Geometry and Physics, Institute for Basic Science, Pohang, Republic of Korea 37673
Max-Planck Institut für Mathematik, Vivatsgasse 7, 53111 Bonn, Germany
E-mail: [email protected], [email protected], [email protected]
Abstract: We study the homotopy category hef(R, W ) (and its Z2 -graded version HEF(R, W ))
of elementary factorizations, where R is a Bézout domain which has prime elements and W =
W0 Wc , where W0 ∈ R× is a square-free element of R and Wc ∈ R× is a finite product of
primes with order at least two. In this situation, we give criteria for detecting isomorphisms in
hef(R, W ) and HEF(R, W ) and formulas for the number of isomorphism classes of objects. We
also study the full subcategory hef (R, W ) of the homotopy category hmf(R, W ) of finite rank
matrix factorizations of W which is additively generated by elementary factorizations. We show
that hef (R, W ) is Krull-Schmidt and we conjecture that it coincides with hmf(R, W ). Finally,
we discuss a few classes of examples.
Contents
1.
2.
3.
4.
5.
A.
B.
Matrix factorizations over an integral domain . . . . . . . . . . . . . . . .
Elementary matrix factorizations over a Bézout domain . . . . . . . . . .
The additive category hef (R, W ) for a Bézout domain and critically-finite
Counting elementary factorizations . . . . . . . . . . . . . . . . . . . . . .
Some examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
GCD domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bézout domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. .
. .
W
. .
. .
. .
. .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
9
22
24
35
42
43
Introduction
The study of topological Landau-Ginzburg models [1, 2, 3, 4] often leads to the problem of understanding the triangulated category hmf(R, W ) of finite rank matrix factorizations of an element
W ∈ R, where R is a non-Noetherian commutative ring. For example, the category of B-type
topological D-branes associated to a holomorphic Landau-Ginzburg pair (Σ, W ) with Σ a noncompact Riemann surface and W : Σ → C a non-constant holomorphic function has this form
with R = O(Σ), the non-Noetherian ring of holomorphic functions defined on Σ. When Σ is
connected, the ring O(Σ) is a Bézout domain (in fact, an elementary divisor domain). In this
situation, this problem can be reduced [5] to the study of the full subcategory hef(R, W ) whose
objects are the elementary factorizations, defined as those matrix factorizations of W for which
2
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
the even and odd components of the underlying supermodule have rank one. In this paper, we
study the category hef(R, W ) and the full category hef (R, W ) of hmf(R, W ) which is additively
generated by elementary matrix factorizations, for the case when R is a Bézout domain. We say
that W is critically-finite if it is a product of a square-free element W0 of R with an element
Wc ∈ R which can be written as a finite product of primes of multiplicities strictly greater than
one. When W is critically-finite, the results of this paper provide a detailed description of the
categories hef(R, W ) and hef (R, W ), reducing questions about them to the divisibility theory
of R.
The paper is organized as follows. In Section 1, we recall some basic facts about finite rank
matrix factorizations over unital commutative rings and introduce notation and terminology
which will be used later on. In Section 2, we study the category hef(R, W ) and its Z2 -graded
completion HEF(R, W ) when W is any non-zero element of R, describing these categories in
terms of the lattice of divisors of W and giving criteria for deciding when two objects are
isomorphic. We also study the behavior of these categories under localization at a multiplicative
set as well their subcategories of primary matrix factorizations. In Section 3, we show that
the additive category hef (R, W ) is Krull-Schmidt when R is a Bézout domain and W is a
critically-finite element of R and propose a few conjectures about hmf(R, W ). In Section 4,
we give a formula for the number of isomorphism classes in the categories HEF(R, W ) and
hef(R, W ). Finally, Section 5 discusses a few classes of examples. Appendices A and B collect
some information on greatest common denominator (GCD) domains and Bézout domains.
Notations and conventions. The symbols 0̂ and 1̂ denote the two elements of the field Z2 = Z/2Z,
where 0̂ is the zero element. Unless otherwise specified, all rings considered are unital and
commutative. Given a cancellative Abelian monoid (M, ·), we say that an element x ∈ M
divides y ∈ M if there exists q ∈ M such that y = qx. In this case, q is uniquely determined by
x and y and we denote it by q = x/y or xy .
def.
Let R be a unital commutative ring. The set of non-zero elements of R is denoted by R× =
R \ {0}, while the multiplicative group of units of R is denoted by U (R). The Abelian categories
of all R-modules is denoted ModR , while the Abelian category of finitely-generated R-modules
is denoted modR . Let ModZR2 denote the category of Z2 -graded modules and outer (i.e. even)
morphisms of such and ModZR2 denote the category of Z2 -graded modules and inner morphisms of
such. By definition, an R-linear category is a category enriched in the monoidal category ModR
while a Z2 -graded R-linear category is a category enriched in the monoidal category ModZR2 .
With this definition, a linear category is pre-additive, but it need not admit finite bi-products
(direct sums). For any Z2 -graded R-linear category C, the even subcategory C 0̂ is the R-linear
category obtained from C by keeping only the even morphisms.
For any unital integral domain R, let ∼ denote the equivalence relation defined on R× by
association in divisibility:
x ∼ y iff ∃γ ∈ U (R) : y = γx .
The set of equivalence classes of this relation coincides with the set R× /U (R) of orbits for
the obvious multiplicative action of U (R). Since R is a commutative domain, the quotient
R× /U (R) inherits a multiplicative structure of cancellative Abelian monoid. For any x ∈ R× ,
let (x) ∈ R× /U (R) denote the equivalence class of x under ∼. Then for any x, y ∈ R× , we
have (xy) = (x)(y). The monoid R× /U (R) can also be described as follows. Let G+ (R) be
the set of non-zero principal ideals of R. If x, y are elements of R× , we have hxihyi = hxyi,
so the product of principal ideals corresponds to the product of the multiplicative group R×
and makes G+ (R) into a cancellative Abelian monoid with unit h1i = R. Notice that G+ (R)
coincides with the positive cone of the group of divisibility (see Subsection 5.2) G(R) of R, when
Elementary matrix factorizations over Bézout domains
3
the latter is viewed as an Abelian group ordered by reverse inclusion. The monoids R× /U (R)
and G+ (R) can be identified as follows. For any x ∈ R× , let hxi ∈ G+ (R) denote the principal
ideal generated by x. Then hxi depends only on (x) and will also be denoted by h(x)i. This gives
a group morphism h i : R× /U (R) → G+ (R). For any non-zero principal ideal I ∈ G+ (R), the
set of all generators x of I is a class in R× /U (R) which we denote by (I); this gives a group
morphism ( ) : G+ (R) → R× /U (R). For all x ∈ R× , we have (hxi) = (x) and h(x)i = hxi, which
implies that h i and ( ) are mutually inverse group isomorphisms.
If R is a GCD domain (see Appendix A) and x1 , . . . , xn are elements of R such that x1 . . . xn 6=
0, let d be any greatest common divisor (gcd) of x1 , . . . , xn . Then d is determined by x1 , . . . , xn
up to association in divisibility and we denote its equivalence class by (x1 , . . . , xn ) ∈ R× /U (R).
The principal ideal hdi = h(x1 , . . . , xn )i ∈ G+ (R) does not depend on the choice of d. The
elements x1 , . . . , xn also have a least common multiple (lcm) m, which is determined up to
association in divisibility and whose equivalence class we denote by [x1 , . . . , xn ] ∈ R× /U (R).
For n = 2, we have:
(x1 )(x2 )
[x1 , x2 ] =
.
(x1 , x2 )
def.
If R is a Bézout domain (see Appendix B), then we have hx1 , . . . , xn i = h(x1 , . . . , xn )i =
hx1 i + . . . + hxn i, so the gcd operation transfers the operation given by taking the finite sum of
principal ideals from G+ (R) to R× /U (R) through the isomorphism of groups described above.
In this case, we have (x1 , . . . , xn ) = (hx1 , . . . , xn i). We also have h[x1 , . . . , xn ]i = ∩ni=1 hxi i and
hence [x1 , . . . , xn ] = (∩ni=1 hxi i). Thus the lcm corresponds to the finite intersection of principal
ideals.
1. Matrix factorizations over an integral domain
Let R be an integral domain and W ∈ R× be a non-zero element of R.
1.1. Categories of matrix factorizations. We shall use the following notations:
1. MF(R, W ) denotes the R-linear and Z2 -graded differential category of R-valued matrix factorizations of W of finite rank. The objects of this category are pairs a = (M, D), where M
is a free Z2 -graded R-module of finite rank and D is an odd endomorphism of M such that
D2 = W idM . For any objects a1 = (M1 , D1 ) and a2 = (M2 , D2 ) of MF(R, W ), the Z2 -graded
R-module of morphisms from a1 to a2 is given by the inner Hom:
HomMF(R,W ) (a1 , a2 ) = HomR (M1 , M2 ) = Hom0̂R (M1 , M2 ) ⊕ Hom1̂R (M1 , M2 ) ,
endowed with the differential da1 ,a2 determined uniquely by the condition:
da1 ,a2 (f ) = D2 ◦ f − (−1)κ f ◦ D1 , ∀f ∈ HomκR (M1 , M2 ) ,
where κ ∈ Z2 .
2. ZMF(R, W ) denotes the R-linear and Z2 -graded cocycle category of MF(R, W ). This has the
same objects as MF(R, W ) but morphism spaces given by:
def.
HomZMF(R,W ) (a1 , a2 ) = {f ∈ HomMF(R,W ) (a1 , a2 )|da1 ,a2 (f ) = 0} .
4
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
3. BMF(R, W ) denotes the R-linear and Z2 -graded coboundary category of MF(R, W ), which is
an ideal in ZMF(R, W ). This has the same objects as MF(R, W ) but morphism spaces given
by:
def.
HomBMF(R,W ) (a1 , a2 ) = {da1 ,a2 (f )|f ∈ HomMF(R,W ) (a1 , a2 )} .
4. HMF(R, W ) denotes the R-linear and Z2 -graded total cohomology category of MF(R, W ).
This has the same objects as MF(R, W ) but morphism spaces given by:
def.
HomHMF(R,W ) (a1 , a2 ) = HomZMF(R,W ) (a1 , a2 )/HomBMF(R,W ) (a1 , a2 ) .
5. The subcategories of MF(R, W ), ZMF(R, W ), BMF(R, W ) and HMF(R, W ) obtained by
def.
restricting to morphisms of even degree are denoted respectively by mf(R, W ) = MF0̂ (R, W ),
def.
def.
def.
zmf(R, W ) = ZMF0̂ (R, W ), bmf(R, W ) = BMF0̂ (R, W ) and hmf(R, W ) = HMF0̂ (R, W ).
The categories MF(R, W ), BMF(R, W ) and ZMF(R, W ) admit double direct sums (and hence all
finite direct sums of at least two elements) but do not have zero objects.
On the other hand, the
0 1
category HMF(R, W ) is additive, the matrix factorization
being a zero object. Finally,
W 0
it is well-known that the category hmf(R, W ) is triangulated (see [6] for a detailed treatment).
For later reference, recall that the biproduct (direct sum) of MF(R, W ) is defined as follows:
Definition 1.1 Given two matrix factorizations ai = (Mi , Di ), (i = 1, 2) of W ∈ R, their
def.
direct sum a1 ⊕ a2 is the matrix factorization a = (M, D) of W , where M = M 0̂ ⊕ M 1̂ and
def. 0 v
, with:
D =
u0
κ
M =
M1κ
⊕
M2κ
∀κ ∈ Z2
u1 0
and u =
0 u2
v 0
, v= 1
0 v2
.
Given a third matrix factorization a3 = (M3 , D3 ) of W and two morphisms fi ∈ HomM F (R,W ) (ai , a3 ) =
HomR (ai , a3 ) (i = 1, 2) in MF(R, W ), their direct sum of f1 ⊕ f2 ∈ HomMF(R,W ) (a1 ⊕ a2 , a3 ) =
HomR (a1 ⊕ a2 , a3 ) is the ordinary direct sum of the R-module morphisms f1 and f2 .
As a consequence, MF(R, W ) admits all finite but non-empty direct sums. The following result
is elementary:
Lemma 1.2 The following statements hold:
1. The subcategories ZMF(R, W ) and BMF(R, W ) of MF(R, W ) are closed under finite direct
sums (but need not have zero objects).
2. The direct sum induces a well-defined biproduct (which is again denoted by ⊕) on the R-linear
categories HMF(R, W ) and hmf(R, W ).
3. (HMF(R, W ), ⊕) and (hmf(R, W ), ⊕) are additive categories, a zero object in each being
given by any of the elementary factorizations e1 and eW , which are isomorphic to each other
in hmf(R, W ). In particular, any finite direct sum of the elementary factorizations e1 and eW
is a zero object in HMF(R, W ) and in hmf(R, W ).
Elementary matrix factorizations over Bézout domains
5
1.2. Reduced rank and matrix description. Let a = (M, D) be an object of MF(R, W ), where
M = M 0̂ ⊕ M 1̂ . Taking the supertrace in the equation D2 = W idM and using the fact that
W 6= 0 shows that rkM 0̂ = rkM 1̂ . We call this natural number the reduced rank of a and denote
it by ρ(a); we have rkM = 2ρ(a). Choosing a homogeneous basis of M (i.e. a basis of M 0̂
and a basis of M 1̂ ) gives an isomorphism of R-supermodules M ' Rρ|ρ , where ρ = ρ(a) and
Rρ|ρ denotes the R-supermodule with Z2 -homogeneous components (Rρ|ρ )0̂ = (Rρ|ρ )1̂ = R⊕ρ .
This isomorphism allows us to identify D with a square matrix of size 2ρ(a) which has block
off-diagonal form:
0v
D=
,
u0
where u and v are square matrices of size ρ(a) with entries in R. The condition D2 = W idM
amounts to the relations:
uv = vu = W Iρ ,
(1.1)
where Iρ denotes the identity matrix of size ρ. Since W 6= 0, these conditions imply that the
matrices u and v have maximal rank1 :
rku = rkv = ρ .
Matrix factorizations for which M = Rρ|ρ form a dg subcategory of MF(R, W ) which is essential
in the sense that it is dg-equivalent with MF(R, W ). Below, we often tacitly identify MF(R, W )
with this essential subcategory and use similar identifications for ZMF(R, W ), BMF(R, W ) and
HMF(R, W ).
ρ1 |ρ1 , D ) and a = (Rρ2 |ρ2 , D ) of W , write D =
1
2
2
i
Given
two matrix factorizations a1 = (R
0 vi
, with ui , vi ∈ Mat(ρi , ρi , R). Then:
ui 0
• An even morphism f ∈ Hom0̂MF(R,W ) (a1 , a2 ) has the matrix form:
0
f
f = 0̂0̂
0 f1̂1̂
with f0̂0̂ , f1̂1̂ ∈ Mat(ρ1 , ρ2 , R) and we have:
0
v2 ◦ f1̂1̂ − f0̂0̂ ◦ v1
da1 ,a2 (f ) = D2 ◦ f − f ◦ D1 =
u2 ◦ f0̂0̂ − f1̂1̂ ◦ u1
0
;
• An odd morphism g ∈ Hom1̂MF(R,W ) (a1 , a2 ) has the matrix form:
0 g1̂0̂
g=
g0̂1̂ 0
with g1̂0̂ , g0̂1̂ ∈ Mat(ρ1 , ρ2 , R) and we have:
v ◦ g + g1̂0̂ ◦ u1
0
da1 ,a2 (g) = D2 ◦ g + g ◦ D1 = 2 0̂1̂
0
u2 ◦ g1̂0̂ + g0̂1̂ ◦ v1
1
To see this, it suffices to consider equations (1.1) in the field of fractions of R.
.
6
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
Remark 1.1. The cocycle condition da1 ,a2 (f ) = 0 satisfied by an even morphism f ∈ Hom0̂ZMF(R,W ) (a1 , a2 )
amounts to the system:
(
v2 ◦ f1̂1̂ = f0̂0̂ ◦ v1
u2 ◦ f0̂0̂ = f1̂1̂ ◦ u1 ,
which in turn amounts to any of the following equivalent conditions:
f1̂1̂ =
u2 ◦ f0̂0̂ ◦ v1
v2 ◦ f1̂1̂ ◦ u1
⇐⇒ f0̂0̂ =
.
W
W
Similarly, the cocycle condition da1 ,a2 (g) = 0 defining an odd morphism g ∈ Hom1̂ZMF(R,W ) (a1 , a2 )
amounts to the system:
(
v2 ◦ g0̂1̂ + g1̂0̂ ◦ u1 = 0
u2 ◦ g1̂0̂ + g0̂1̂ ◦ v1 = 0 ,
which in turn amounts to any of the following equivalent conditions:
u2 ◦ g1̂0̂ ◦ u1
v2 ◦ g0̂1̂ ◦ v1
⇐⇒ g0̂1̂ = −
.
g1̂0̂ = −
W
W
1.3. Strong isomorphism. Recall that zmf(R, W ) denotes the even subcategory of ZMF(R, W ).
This category admits non-empty finite direct sums but does not have a zero object.
Definition 1.3 Two matrix factorizations a1 and a2 of W over R are called strongly isomorphic
if they are isomorphic in the category zmf(R, W ).
It is clear that two strongly isomorphic factorizations are also isomorphic in hmf(R, W ), but the
converse need not hold.
Proposition 1.4 Let a1 = (Rρ1 |ρ1 , D1 ) and a2 = (Rρ2 |ρ2 , D2 ) be two matrix factorizations of
0 vi
W over R, where Di =
. Then the following statements are equivalent:
ui 0
(a) a1 and a2 are strongly isomorphic.
(b) ρ1 = ρ2 (a quantity which we denote by ρ) and there exist invertible matrices A, B ∈ GL(ρ, R)
such that one (and hence both) of the following equivalent conditions is satisfied:
1. v2 = Av1 B −1 ,
2. u2 = Bu1 A−1 .
Proof. a1 and a2 are strongly isomorphic iff there exists U ∈ Homzmf(R,W ) (a1 , a2 ) which is an
isomorphism in zmf(R, W ). Since U is an even morphism in the cocycle category, we have:
U D 1 = D2 U .
(1.2)
A 0
The condition that U be even allows us to identify it with a matrix of the form U =
,
0B
while invertibility of U in zmf(R, W ) amounts to invertibility of the matrix U , which in turn
means that A and B are square matrices (thus ρ1 = ρ2 = ρ) belonging to GL(ρ, R). Thus relation
(1.2) reduces to either of conditions 1. or 2., which are equivalent since v1 u1 = u1 v1 = W Iρ and
u2 v2 = v2 u2 = W Iρ . t
u
Elementary matrix factorizations over Bézout domains
7
1.4. Critical divisors and the critical locus of W .
Definition 1.5 A divisor d of W which is not a unit is called critical if d2 |W .
Let:
def.
C(W ) =
d∈R
d2 |W
be the set of all critical divisors of W . The ideal:
def.
IW = ∩d∈C(W ) hdi
(1.3)
is called the critical ideal of W . Notice that IW consists of those elements of R which are divisible
by all critical divisors of W . In particular, we have (W ) ⊂ IW and hence there exists a unital
ring epimorphism R/(W ) → R/IW .
Definition 1.6 A critical prime divisor of W is a prime element p ∈ R such that p2 |W . The
critical locus of W is the subset of Spec(R) consisting of the principal prime ideals of R generated
by the critical prime divisors of W :
def.
Crit(W ) =
hpi ∈ Spec(R)
p2 |W
.
1.5. Critically-finite elements. Let R be a Bézout domain. Then R is a GCD domain, hence
irreducible elements of R are prime. This implies that any factorizable element2 of R has a
unique prime factorization up to association in divisibility.
Definition 1.7 A non-zero non-unit W of R is called:
• non-critical, if W has no critical divisors;
• critically-finite if it has a factorization of the form:
W = W0 Wc with Wc = pn1 1 . . . pnNN ,
(1.4)
where nj ≥ 2, p1 , . . . , pN are critical prime divisors of W (with pi 6∼ pj for i 6= j) and W0 is
non-critical and coprime with Wc .
Notice that the elements W0 , Wc and pi in the factorization (1.4) are determined by W up to
association, while the integers ni are uniquely determined by W . The factors W0 and Wc are
called respectively the non-critical and critical parts of W . The integers ni ≥ 2 are called the
orders of the critical prime divisors pi .
For a critically-finite element W with decomposition (1.4), we have:
Crit(W ) = {hp1 i, . . . , hpN i} and IW = hWred i ,
where3 :
def.
b
Wred = p1
n1
c
2
b
nN
. . . pN 2
c
is called the reduction of W . Notice that Wred is determined up to association in divisibility.
2
3
I.e. an element of R which has a finite factorization into irreducibles.
The notation bxc ∈ Z indicates the integral part of a real number x ∈ R.
8
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
1.6. Two-step factorizations of W . Recall that a two-step factorization (or two-step multiplicative partition) of W is an ordered pair (u, v) ∈ R×R such that W = uv. In this case, the divisors
u and v are called W -conjugate. The transpose of (u, v) is the ordered pair (v, u) (which is again a
two-step factorization of W ), while the opposite transpose is the ordered pair σ(u, v) = (−v, −u).
This defines an involution σ of the set MP2 (W ) of two-step factorizations of W . The two-step
factorizations (u, v) and (u0 , v 0 ) are called similar (and we write (u, v) ∼ (u0 , v 0 )) if there exists
γ ∈ U (R) such that u0 = γu and v 0 = γ −1 v. We have σ(u, v) ∼ (v, u).
Definition 1.8 The support of a two-step factorization (u, v) of W is the principal ideal hu, vi ∈
G+ (R).
def.
def.
Let d be a gcd of u and v. Since W = uv = d2 u1 v1 (where u1 = u/d, v1 = v/d), it is clear
that d is a critical divisor of W . Notice that the opposite transpose of the two step factorization
(u, v) has the same support as (u, v).
1.7. Elementary matrix factorizations.
Definition 1.9 A matrix factorization a = (M, D) of W over R is called elementary if it has
unit reduced rank, i.e. if ρ(a) = 1.
def.
Any elementary factorizationis strongly
isomorphic to one of the form ev = (R1|1 , Dv ), where v
def.
def. 0 v
is a divisor of W and Dv =
, with u = W/v ∈ R. Let EF(R, W ) denote the full subcateu0
gory of MF(R, W ) whose objects are the elementary factorizations of W over R. Let ZEF(R, W )
and HEF(R, W ) denote respectively the cocycle and total cohomology categories of EF(R, W ).
def.
def.
We also use the notations zef(R, W ) = ZEF0̂ (R, W ) and hef(R, W ) = HEF0̂ (R, W ). Notice
that an elementary factorization is indecomposable in zmf(R, W ), but it need not be indecomposable in the triangulated category hmf(R, W ).
The map Φ : ObEF(M, W ) → MP2 (W ) which sends ev to the ordered pair (u, v) is a
bijection. The suspension of ev is given by Σev = e−u = (R1|1 , D−u ), since:
D−u =
0 −u
−v 0
.
In particular, Σev corresponds to the opposite transpose σ(u, v) and we have:
Φ◦Σ =σ◦Φ .
Hence Σ preserves the subcategory EF(M, W ) of MF(R, W ) and the subcategories HEF(R, W )
and hef(R, W ) of HMF(R, W ) and hmf(R, W ). This implies that HEF(R, W ) is equivalent with
the graded completion grΣ hef(R, W ). We thus have natural isomorphisms:
Hom1̂HEF(R,W ) (ev1 , ev2 ) 'R Homhef(R,W ) (ev1 , Σev2 ) = Homhef(R,W ) (ev1 , e−u2 ) ,
Hom1̂HEF(R,W ) (ev1 , ev2 ) 'R Homhef(R,W ) (Σev1 , ev2 ) = Homhef(R,W ) (e−u1 , ev2 ) ,
for any divisors v1 , v2 of W , where u1 = W/v1 and u2 = W/v2 .
(1.5)
Elementary matrix factorizations over Bézout domains
9
Definition 1.10 The support of an elementary matrix factorization ev is the ideal of R defined
through:
def.
supp (ev ) = supp (Φ(ev )) = hv, W/vi .
Notice that this ideal is generated by any gcd d of v and W/v and that d is a critical divisor of
W.
We will see later that an elementary factorization is trivial iff its support equals R.
Definition 1.11 Two elementary matrix factorizations ev1 and ev2 of W are called similar if
v1 ∼ v2 or equivalently u1 ∼ u2 . This amounts to existence of a unit γ ∈ U (R) such that v2 = γv1
and u2 = γ −1 u1 .
Proposition 1.12 Two elementary factorizations ev1 and ev2 are strongly isomorphic iff they
are similar. In particular, strong isomorphism classes of elementary factorization are in bijection
with the set of those principal ideals of R which contain W .
Proof. Suppose that ev1 and ev2 are strongly isomorphic. By Proposition 1.4, there exist units
def.
def.
x, y ∈ U (R) such that v2 = xv1 y −1 and u2 = yu1 x−1 , where ui = W/vi . Setting γ = xy −1
gives v1 = γv1 and u2 = γ −1 u1 , hence ev1 and ev2 are similar. Conversely, suppose that ev1 ∼ ev2 .
Then there exists a unit γ ∈ U (R) such that v2 = γv1 and u2 = γ −1 u1 . Setting x = γ and y = 1
gives v2 = xv1 y −1 and u2 = yu1 x−1 , which shows that ev1 and ev2 are strongly isomorphic upon
using Proposition 1.4. The map which sends the strong isomorphism class of ev to the principal
ideal (v) gives the bijection stated. t
u
It is clear that ev1 and ev2 are similar iff the corresponding two-step factorizations (v1 , u1 ) and
(v2 , u2 ) of W are similar. Since any strong isomorphism induces an isomorphism in hef(R, W ),
it follows that similar elementary factorizations are isomorphic in hef(R, W ).
1.8. The categories HEF(R, W ) and hef (R, W ). Let EF(R, W ) denote the smallest full R-linear
subcategory of MF(R, W ) which contains all objects of EF(R, W ) and is closed under finite direct sums. It is clear that EF(R, W ) is a full dg subcategory of MF(R, W ). Let HEF(R, W )
def.
denote the total cohomology category of EF(R, W ). Let hef (R, W ) = HEF0̂ (R, W ) denote
the subcategory obtained from HEF(R, W ) by keeping only the even morphisms. Notice that
hef (R, W ) coincides with the smallest full subcategory of hmf(R, W ) which contains all elementary factorizations of W .
2. Elementary matrix factorizations over a Bézout domain
Throughout this section, let R be a Bézout domain and W be a non-zero element of R.
2.1. The subcategory of elementary factorizations. Let v1 , v2 be divisors of W and e1 := ev1 ,
def.
e2 := ev2 be the corresponding elementary matrix factorizations of W . Let u1 = W/v1 , u2 =
W/v2 . Let a be a gcd of v1 and v2 . Define:
def.
def.
def.
b = v1 /a , c = v2 /a , d =
W
def.
def.
, a0 = a/s , d0 = d/s ,
abc
(2.1)
10
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
where s is a gcd of a and d. Then a = a0 s and d = d0 s with (a0 , d0 ) = (1) = (b, c) and
W = abcd = s2 a0 bcd0 . In particular, s is a critical divisor of W . Moreover:
v1 = ab = sa0 b , v2 = ac = sa0 c , u1 = cd = scd0 , u2 = bd = sbd0
(2.2)
and we have:
(d) = (u1 , u2 ) , (s) = (v1 , v2 , u1 , u2 ) .
(2.3)
Notice the following relations in the cancellative monoid R× /U (R):
(v1 , v2 ) = (sa0 ) , (u1 , u2 ) = (s)(d0 ) , (u1 , v1 ) = (s)(a0 , c)(b, d0 ) ,
(u1 , v2 ) = (s)(c) , (u2 , v1 ) = (s)(b) , (u2 , v2 ) = (s)(a0 , b)(c, d0 ) .
(2.4)
In this notation:
0 v2
0 ac
0 a0 c
0 v1
0 ab
0 a0 b
=
=s
.
=
=s
and Dv2 =
Dv1 =
u2 0
bd 0
bd0 0
u1 0
cd 0
cd0 0
For f ∈ Hom0̂MF(R,W ) (e1 , e2 ) = Hom0̂R (R1|1 , R1|1 ) and g ∈ Hom1̂MF(R,W ) (e1 , e2 ) = Hom1̂R (R1|1 , R1|1 ),
we have:
c0
0 a
.
(2.5)
and de1 ,e2 (g) = (ag0̂1̂ + g1̂0̂ d)
de1 ,e2 (f ) = (cf1̂1̂ − f0̂0̂ b)
0b
−d 0
Remark 2.1. Relations (2.1) and (2.4) imply the following equalities in the cancellative monoid
R× /U (R):
(u1 , v2 )
(u2 , v1 )
=
.
(s) =
(2.6)
v2
(u1 , v2 ), (v1 ,v2 )
(u2 , v1 ), (v1v,v1 2 )
2.1.1. Morphisms in HEF(R, W ). Let Mat(n, R× /U (R)) denote the set of square matrices of size
n with entries from the multiplicative semigroup R/U (R). Any matrix S ∈ Mat(n, R× /U (R))
can be viewed as an equivalence class of matrices A ∈ Mat(n, R× ) under the equivalence relation:
A ∼n B iff ∀i, j ∈ {1, . . . , n} : ∃qij ∈ U (R) such that Bij = qij Aij .
(2.7)
Proposition 2.1 With the notations above, we have:
1. Hom0̂ZMF(R,W ) (e1 , e2 ) is the free R-module of rank one generated by the matrix:
def.
0̂ (v1 , v2 ) =
" (v2 )
c0
∈ (v1 ,v2 )
0b
0
#
0
def.
(v1 )
(v1 ,v2 )
= 0̂ (v1 , v2 ) ,
where the matrix 0̂ (v1 , v2 ) ∈ Mat(2, R/U (R)) in the right hand side is viewed as an equivalence class under the relation (2.7).
2. Hom1̂ZMF(R,W ) (e1 , e2 ) is the free R-module of rank one generated by the matrix:
def.
1̂ (v1 , v2 ; W ) =
"
0
0 a0
∈
0
(u
−d 0
− (u1 ,v1 )2 )
(v2 )
(u1 ,v2 )
0
#
def.
= 1̂ (v1 , v2 ; W )
and we have 1̂ (v1 , v2 ; W ) = 1̂ (v2 , v1 ; W ) in Mat(2, R/U (R)).
Elementary matrix factorizations over Bézout domains
11
Proof. Relations (2.4) imply:
(v2 )
(v1 )
(v2 )
(v1 )
(u1 )
(u2 )
= (c) ,
= (b) ,
=
= (a0 ) ,
=
= (d0 ) .
(v1 , v2 )
(v1 , v2 )
(u1 , v2 )
(u2 , v1 )
(u1 , v2 )
(u2 , v1 )
(2.8)
These relations show that 0̂ (v1 , v2 ) and 1̂ (v1 , v2 ; W ) belong to the equivalence classes 0̂ (v1 , v2 )
and 1̂ (v1 , v2 ; W ) and that we have 1̂ (v1 , v2 ; W ) = 1̂ (v2 , v1 ; W ).
For an even morphism f : e1 → e2 in MF(R, W ), the first equation in (2.5) shows that the
condition de1 ,e2 (f ) = 0 amounts to:
f1̂1̂ c − f0̂0̂ b = 0 .
Since b and c are coprime, this condition is equivalent with the existence of an element γ ∈ R
such that f0̂0̂ = γc and f1̂1̂ = γb. Thus:
c0
f =γ
= γ0̂ (v1 , v2 ) .
(2.9)
0b
On the other hand, the second equation in (2.5) shows that an odd morphism g : e1 → e2 in
MF(R, W ) satisfies de1 ,e2 (g) = 0 iff:
ag0̂1̂ + dg1̂0̂ = 0 .
Since a0 and d0 are coprime, this condition is equivalent with the existence of an element γ ∈ R
such that g1̂0̂ = γa0 and g1̂1̂ = −γd0 . Thus:
0 a0
(2.10)
= γ1̂ (v1 , v2 ; W ) .
g=γ
−d0 0
t
u
Proposition 2.2 Let vi be as in Proposition 2.1. Then Hom0̂HMF(R,W ) (e1 , e2 ) and Hom1̂HMF(R,W ) (e1 , e2 )
are cyclically presented cyclic R-modules generated respectively by the matrices 0̂ (v1 , v2 ) and
1̂ (v1 , v2 ; W ), whose annihilators are equal to each other and coincide with the following principal ideal of R:
def.
αW (v1 , v2 ) = hv1 , u1 , v2 , u2 i = hsi .
Proof. Let f ∈ Hom0̂ZMF(R,W ) (e1 , e2 ). Then f is exact iff there exists an odd morphism g ∈
Hom1̂MF(R,W ) (e1 , e2 ) such that:
c0
f = de1 ,e2 (g) = (ag0̂1̂ + g1̂0̂ d)
0b
.
Comparing this with (2.9), we find that f is exact if and only if s ∈ (a, d) divides γ. This implies
that the principal ideal generated by the element:
s ∈ ((v1 , v2 ), (u1 , u2 )) = (v1 , u1 , v2 , u2 )
is the annihilator of Hom0̂ZMF(R,W ) (e1 , e2 ).
On the other hand, an odd morphism g ∈ Hom1̂ZMF(R,W ) (e1 , e2 ) is exact iff there exists an
even morphism f ∈ Hom0̂MF(R,W ) (e1 , e2 ) such that:
0 a
g = de1 ,e2 (f ) = (f1̂1̂ c − f0̂0̂ b)
−d 0
.
Comparing with (2.10) and recalling that (b, c) = (1), we find that g is exact iff (a, d)|γ. Hence
the annihilator of Hom1̂HMF(R,W ) (e1 , e2 ) coincides with that of Hom0̂HMF(R,W ) (e1 , e2 ). t
u
12
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
Remark 2.2. Since s is a critical divisor of W , we have IW HomHEF(R,W ) (e1 , e2 ) = 0, where IW
denotes the critical ideal of W defined in (1.3). In particular, HEF(R, W ) can be viewed as an
R/IW -linear category.
def.
Let Div(W ) = {d ∈ R d|W } and consider the function αW : Div(W ) × Div(W ) → G+ (R)
defined in Proposition 2.2. This function is symmetric since αW (v1 , v2 ) = αW (v2 , v1 ). Let 1G(R) =
h1i = R denote the neutral element of the group of divisibility G(R), whose group operation we
write multiplicatively.
Proposition 2.3 The symmetric function αW (v1 , v2 ) is multiplicative with respect to each of
its arguments in the following sense:
• For any two relatively prime elements v2 and ve2 of R such that v2 ve2 is a divisor of W , we
have:
αW (v1 , v2 ve2 ) = αW (v1 , v2 )αW (v1 , ve2 )
(2.11)
and αW (v1 , v2 ) + αW (v1 , ve2 ) = 1G(R) , where + denotes the sum of ideals of R.
• For any two relatively prime elements v1 and ve1 of R such that v1 ve1 is a divisor of W , we
have:
αW (v1 ve1 , v2 ) = αW (v1 , v2 )αW (e
v1 , v2 )
(2.12)
and αW (v1 , v2 ) + αW (e
v1 , v2 ) = 1G(R) , where + denotes the sum of ideals of R.
Proof. To prove the first statement, we start from relation (2.6), which allows us to write:
(u1 , v2 ve2 )
D
αW (v1 , v2 ve2 ) =
(u1 , v2 ve2 ), (v1v,v2 ve22ve2 )
E
,
(2.13)
where u1 = W/v1 . Recall that the function (−, r) is multiplicative on relatively prime elements
for any r ∈ R× , i.e. (xy, r) = (x, r)(y, r). Thus:
(u1 , v2 ve2 ) = (u1 , v2 )(u1 , ve2 ) ,
(v1 , v2 ve2 ) = (v1 , v2 )(v1 , ve2 ) .
(2.14)
v2 )
(v2 )
(e
v2 )
The second of these relations gives (v(v1 ,v2 ve22ve)2 ) = (v(v1 ,v2 )2 ) (v(e
v2 ) . Notice that ( (v1 ,v2 ) , (v1 ,e
v2 ) ) = (1)
1 ,e
since v2 and ve2 are coprime. Hence:
(v2 )
(e
v2 )
(v2 )(e
v2 )
(u1 , v2 ve2 ),
= (u1 , v2 ve2 ),
(u1 , v2 ve2 ),
=
(v1 , v2 ve2 )
(v1 , v2 )
(v1 , ve2 )
(v2 )
(v2 )
(e
v2 )
(e
v2 )
= (u1 , v2 ),
(u1 , ve2 ),
(u1 , v2 ),
(u1 , ve2 ),
,
(v1 , v2 )
(v1 , v2 )
(v1 , ve2 )
(v1 , ve2 )
where in the last equality we used the first relation in (2.14) and noticed that (u1 , v2 ) and (u1 , ve2 )
are coprime (since (v2 , ve2 ) = (1)), which allows us to use similar-multiplicativity
of the function
(e
v2 )
(v2 )
(−, r) for (r) = (v1 ,v2 ) and for (r) = (v1 ,ev2 ) . Since (v2 , ve2 ) = (1), we have (u1 , ve2 ), (v(v1 ,v2 )2 ) =
v2 )
(u1 , v2 ), (v(e
v2 ) = (1). Thus:
1 ,e
(v2 )(e
v2 )
(v2 )
(e
v2 )
(u1 , v2 ve2 ),
= (u1 , v2 ),
(u1 , ve2 ),
.
(v1 , v2 ve2 )
(v1 , v2 )
(v1 , ve2 )
Using this and the first equation of (2.14) in the expression (2.13) gives relation (2.11). The
second statement now follows from the first by symmetry of αW . t
u
Elementary matrix factorizations over Bézout domains
13
2.1.2. Isomorphisms in HEF(R, W ).
We start with a few lemmas.
Lemma 2.1. Let s, x, y, z be four elements of R. Then the equation:
s(g1 x + g2 y) + g3 z = 1
(2.15)
has a solution (g1 , g2 , g3 ) ∈ R3 iff (s(x, y), z) = (1).
Proof. Let t be a gcd of x and y. We treat each implication in turn:
1. Assume that (g1 , g2 , g3 ) ∈ R3 is a solution. Then t divides g1 x+g2 y, so there exists g4 ∈ R such
that g1 x + g2 y = g4 t. Multiplying both sides with s and using (2.15), this gives stg4 + g3 z = 1,
which implies (st, z) = (1).
2. Assume that (st, z) = (1). Then there exist g3 , g4 ∈ R such that:
stg4 + g3 z = 1 .
(2.16)
Since (t) = (x, y), the Bézout identity shows that there exist ge1 , ge2 ∈ R such that ge1 x+e
g2 y = t.
def.
Substituting this into (2.16) shows that (g1 , g2 , g3 ) satisfies (2.15), where g1 = ge1 g4 and
def.
g2 = ge2 g4 . t
u
Lemma 2.2. Let s, a0 , b, c, d0 be five elements of R such that (a0 , d0 ) = (1). Then the system of
equations:
(
bcg − s(a0 bg1 + cd0 g2 ) = 1
(2.17)
bcg − s(a0 ch1 + bd0 h2 ) = 1
has a solution (g, g1 , g2 , h1 , h2 ) ∈ R5 iff a0 , b, c, d0 are pairwise coprime and (bc, s) = (1).
Proof. Consider the two implications in turn.
1. Assume that (2.17) has a solution (g, g1 , g2 , h1 , h2 ) ∈ R5 . By Lemma 2.1, we must have
(bc, s(a0 b, cd0 )) = (1) and (bc, s(a0 c, bd0 )) = (1). This implies (bc, s) = (1) and (b, c) = (1). If a
prime element p ∈ R divides (a0 b, cd0 ), then it divides both a0 b and cd0 , hence p|c or p|b since
(a0 , d0 ) = (1). Thus p|bc, which contradicts the fact that that bc and s(a0 b, cd0 ) are coprime.
It follows that we must have (a0 b, cd0 ) = (1). Similarly, the second equation implies that we
must have (a0 c, bd0 ) = (1). Since (a0 , d0 ) = (1) and (b, c) = (1), the last two conditions imply
that a0 , b, c, d0 must be pairwise coprime.
2. Conversely, assume that a0 , b, c, d0 are pairwise coprime and (bc, s) = (1). Following the strategy and notations of the previous lemma, we first solve the equation bcg − sg4 = 1 for g and
g4 using the Bézout identity. Using the same identity, we solve the system:
(
a0 be
g1 + cd0 ge2 = 1
,
(2.18)
a0 ce
h1 + bd0e
h2 = 1
obtaining the solution (g, g4 ge1 , g4 ge2 , g4e
h1 , g4e
h2 ) of (2.17).
t
u
14
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
Proposition 2.4 With the notations (2.1), we have:
1. e1 and e2 are isomorphic in hef(R, W ) iff a0 , b, c, d0 are pairwise coprime and (bc, s) = (1).
2. An odd isomorphism between e1 and e2 in HEF(R, W ) exists iff a0 , b, c, d0 are pairwise coprime
and (a0 d0 , s) = (1).
Proof. 1. Proposition 2.1 gives:
c0
b0
Homzef(R,W ) (e1 , e2 ) = R
and Homzef(R,W ) (e2 , e1 ) = R
.
0c
0b
c0
b0
Two non-zero morphisms f12 = α
∈ Homzef(R,W ) (e1 , e2 ) and f21 = β
∈ Homzef(R,W ) (e2 , e1 )
0b
0c
(where α, β ∈ R× ) induce mutually inverse isomorphisms in hef(R, W ) iff:
f21 f12 = 1 + de1 ,e1 (g) , f12 f21 = 1 + de2 ,e2 (h)
for some g, h ∈ End1̂R (R1|1 ). These conditions read:
bc 0
10
10
10
10
αβ
=
+ (abg0̂1̂ + g1̂0̂ cd)
=
+ s(a0 bg0̂1̂ + g1̂0̂ cd0 )
0 bc
01
01
01
01
10
10
bc 0
10
10
0
0
+ s(a ch0̂1̂ + h1̂0̂ bd )
=
αβ
=
+ (ach0̂1̂ + h1̂0̂ bd)
01
01
01
0 bc
01
and hence amount to the following system of equations for αβ and g, h:
(
αβbc − s(a0 bg0̂1̂ + g1̂0̂ cd0 ) = 1
αβbc − s(a0 ch0̂1̂ + h1̂0̂ bd0 ) = 1 .
(2.19)
Since this system has the form (2.17), Lemma 2.2 shows that it has solutions iff a0 , b, c, d0 are
pairwise coprime and (bc, s) = (1).
0 a0
1̂
1̂
. Two non2. Proposition 2.1 gives HomZEF(R,W ) (e1 , e2 ) = HomZEF(R,W ) (e2 , e1 ) = R
−d0 0
0 a0
0 a0
1̂
∈
∈ HomZEF(R,W ) (e1 , e2 ) and g21 = β
zero odd morphisms g12 = α
−d0 0
−d0 0
Hom1̂ZEF(R,W ) (e2 , e1 ) (with α, β ∈ R× ) induce mutually inverse isomorphisms in HEF(R, W )
iff:
g21 g12 = 1 + de1 ,e1 (f ) , g12 g21 = 1 + de2 ,e2 (q)
for some f, q ∈ End0̂R (R1|1 ). This gives the equations:
0 0
ad 0
10
10
αβ
=
+ (acf0̂0̂ + f1̂1̂ bd)
,
0 a0 d0
01
01
0 0
ad 0
10
10
αβ
=
+ (abq0̂0̂ + q1̂1̂ cd)
0 a0 d0
01
01
which amount to the system:
(
αβa0 d0 − s(a0 cf0̂0̂ + f1̂1̂ bd0 ) = 1
αβa0 d0 − s(a0 bq0̂0̂ + q1̂1̂ cd0 ) = 1 .
(2.20)
This system again has the form (2.17), as can be seen by the substitution of the quadruples
(b, c, a0 , d0 ) := (a0 , d0 , b, c). As a consequence, it has a solution iff a0 , b, c, d0 are pairwise coprime
and (a0 d0 , s) = (1). t
u
Elementary matrix factorizations over Bézout domains
15
Corollary 2.5 Similar elementary matrix factorizations of W are isomorphic in hef(R, W ).
Proof. The statement follows immediately from Proposition 2.4 by taking a = vγ, b = γ −1 ,
c = 1, d = u (where γ ∈ U (R)), since the gcd in R is defined modulo U (R). t
u
Proposition 2.6 Any elementary matrix factorization of W is odd-isomorphic in HEF(R, W )
to its suspension:
ev 'HEF(R,W ) Σev = e−u ,
where u = W/v.
Proof. Let s ∈ (u, v). The isomorphism follows from Proposition 2.4 for a0 = 1 = d0 , b = −v/s,
c = −u/s:
0v
0 bs
0 cs
0u
'
'
'
,
u0
cs 0
bs 0
v0
since (ad, s) = (1).
t
u
Remark 2.3. An odd isomorphism in HEF(R, W ) between ev and Σev = e−u can also be obtained
more abstractly by transporting the identity endomorphism of ev through the isomorphism of Rmodules Hom1̂ (ev , e−u ) = Hom1̂ (ev , Σev ) ' Homhmf(R,W ) (ev , ev ) which results by taking v1 = v
and v2 = −u in the first line of (1.5). Since e−u is similar to eu , the Proposition implies that ev
and eu are oddly isomorphic. When v = 1, both ev = e1 and eu = eW are zero objects and we
have Hom0̂HMF(R,W ) (e1 , eW ) = Hom1̂HMF(R,W ) (e1 , eW ) = {0}, so the odd isomorphism is the zero
morphism.
Proposition 2.7 Let W = W1 W2 with (W1 , W2 ) = (1) and let v be a divisor of W1 . Then
eW2 v 'hmf(R,W ) ev .
def. W1
v .
Proof. Let u0 =
compute:
Setting v1 = v, u1 =
W
v
= W2 u0 , v2 = W2 v and u2 =
W
v2
= u0 , we
v1
v2
W
W1
a ∈ (v1 , v2 ) = (v) , b =
∈ (1) , c =
∈ (W2 ) , d ∈
=
a
a
[v1 , v2 ]
v
v
(W1 )
(u0 )
s ∈ (u1 , v1 , u2 , v2 ) = (u0 , v) , a0 = a/s ∈
, d0 = d/s ∈
=
.
(u0 , v)
(v)(u0 , v)
(u0 , v)
It is clear that a0 , b, c, d0 are mutually coprime and that (s, bc) = (1).
t
u
2.1.3. The composition of morphisms in HEF(R, W ).
Proposition 2.8 Given three divisors v1 , v2 and v3 of W , we have the following relations:
(v2 )(v1 , v3 )
(v1 , v3 )
(v1 , v2 )(v2 , v3 ) 0̂
(v2 )(u1 , v3 )
0̂ (v2 , v3 )1̂ (v1 , v2 ; W ) =
(v1 , v3 ; W )
(u1 , v2 )(v2 , v3 ) 1̂
(v3 )(v1 , u3 )
1̂ (v2 , v3 ; W )0̂ (v1 , v2 ) =
(v3 , v1 ; W )
(v1 , v2 )(u2 , v3 ) 1̂
(u1 )(v1 , v3 )
1̂ (v2 , v3 ; W )1̂ (v1 , v2 ; W ) = −
(v1 , v3 ) .
(u2 , v3 )(u1 , v2 ) 0̂
0̂ (v2 , v3 )0̂ (v1 , v2 ) =
(2.21)
16
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
Proof. Given three divisors v1 , v2 and v3 of W , we have:
#
"
v2 v3
0
v2 (v1 , v3 )
[v1 , v2 , v3 ](v1 , v3 )
=
0̂ (v2 , v3 )0̂ (v1 , v2 ) = (v1 ,v2 )(v2 ,v3 )
0̂ (v1 , v3 ) =
(v1 , v3 )
v1 v2
0
(v1 , v2 )(v2 , v3 )
(v1 , v2 , v3 )[v1 , v3 ] 0̂
(v1 ,v2 )(v2 ,v3 )
where we used the identity:
[a, b, c](a, b)(b, c)(c, a) = (a)(b)(c) (a, b, c) .
(2.22)
This establishes the first of equations (2.21). The remaining equations follow similarly t
u.
Corollary 2.9 Let v be a divisor of W and u = W/v. Then:
1. The R-algebra Endzmf(R,W ) (ev ) is isomorphic with R.
2. We have an isomorphism of Z2 -graded R-algebras:
EndZMF(R,W ) (ev ) '
where ω is an odd generator and t ∈
Z2 -graded ring.
[u,v]
(u,v) .
R[ω]
,
hu2 + ti
In particular, EndZMF(R,W ) (ev ) is a commutative
Proof. For v1 = v2 = v, we have αW (v, v) = hu, vi. Proposition 2.1 gives:
"
#
(v)
0 (u,v)
def.
def. 1 0
, 1̂ (v, v) =
0̂ (v, v) =
01
− (u) 0
(u,v)
and we have:
0̂ (v, v)2 = 0̂ (v, v)
0̂ (v, v)1̂ (v, v; W ) = 1̂ (v, v; W )0̂ (v, v) = 1̂ (v, v)
[u, v]
(v, v) ,
1̂ (v, v; W )2 = −
(u, v) 0̂
which also follows from Proposition 2.8. Setting ω = 1̂ (v, v; W ), these relations imply the desired
statements upon using Proposition 2.1. t
u
Corollary 2.10 Let v be a divisor of W and u = W/v. Then:
1. The R-algebra Endhmf(R,W ) (ev ) is isomorphic with R/hdi = R/hu, vi, where d ∈ (u, v).
2. We have an isomorphism of Z2 -graded R-algebras:
EndHMF(R,W ) (ev ) '
where ω is an odd generator, d ∈ (u, v) and t ∈
supercommutative Z2 -graded ring.
(R/hdi) [ω]
,
hu2 + ti
[u,v]
(u,v) .
In particular, EndZMF(R,W ) (ev ) is a
Proof. The same relations as in the previous Corollary imply the conclusion upon using Proposition 2.2. t
u
Corollary 2.11 An elementary matrix factorization ev is a zero object of hmf(R, W ) iff (u, v) =
(1), where u = W/v.
Proof. The R-algebra End0̂HMF(R,W ) (ev ) ' R/t (where u = W/v) vanishes iff (u, v) = (1). t
u
Elementary matrix factorizations over Bézout domains
17
2.2. Localizations. Let S ⊂ R be a multiplicative subset of R containing the identity 1 ∈ R and
λS : R → RS denote the natural ring morphism from R to the localization RS = S −1 R of R
def.
at S. For any r ∈ R, let rS = λS (r) = 1r ∈ RS denote its extension. For any R-module N , let
NS = S −1 N = N ⊗R RS denote the localization of N at S. For any morphism of R-modules
def.
f : N → N 0 , let fS = f ⊗R idRS : NS → NS0 denote the localization of f at S. For any Z2 -graded
R-module M = M 0̂ ⊕ M 1̂ , we have MS = MS0̂ ⊕ MS1̂ , since the localization functor is exact. In
particular, localization at S induces a functor from the category of Z2 -graded R-modules to the
category of Z2 -graded RS -modules.
Let a = (M, D) be a matrix factorization of W . The localization of a at S is the following
matrix factorization of WS over the ring RS :
def.
aS = (MS , DS ) ∈ MF(RS , WS ) .
It is clear that this extends to an even dg functor locS : MF(R, W ) → MF(RS , WS ), which is Rlinear and preserves direct sums. In turn, this induces dg functors ZMF(R, W ) → ZMF(RS , WS ),
BMF(R, W ) → BMF(RS , WS ), HMF(R, W ) → HMF(RS , WS ) and hmf(R, W ) → hmf(RS , WS ),
which we again denote by locS . We have locS (a) = aS for any matrix factorization a of W over
R.
Proposition 2.12 The functor locS : hmf(R, W ) → hmf(RS , WS ) is a triangulated functor.
Moreover, the strictly full subcategory of hmf(R, W ) defined through:
def.
KS =
a ∈ Ob[hmf(R, W )]
aS 'hmf(RS ,WS ) 0
is a triangulated subcategory of hmf(R, W ).
Proof. It is clear that locS commutes with the cone construction (see [6] for a detailed account of
the latter). It is also clear that the subcategory KS is closed under shifts. Since any distinguished
triangle in which two objects vanish has the property that its third object also vanishes, KS is
also closed under forming triangles. t
u
Proposition 2.13 For any matrix factorizations a, b of W , there exists a natural isomorphism
of Z2 -graded RS -modules:
HomHMF(RS ,WS ) (aS , bS ) 'RS HomHMF(R,W ) (a, b)S .
Proof. Follows immediately from the fact that localization at S is an exact functor from ModR
to ModRS . t
u
2.3. Behavior of hef(R, W ) under localization.
Lemma 2.14 The following statements are equivalent for any elements s, r of R:
1. (s, r) = (1)
2. The class of s modulo the ideal hri is a unit of the ring R/hri.
Proof. We have (s, r) = (1) iff there exist elements a, b ∈ R such that as + br = 1. In turn,
this is equivalent with the condition ās̄ = 1̄ in the ring R/hri, where x̄ = x + hri denotes the
equivalence class of an element x ∈ R modulo the ideal hri. t
u
18
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
Consider the multiplicative set:
def.
SW = {s ∈ R
(s, W ) = (1)} .
Since 0 6∈ SW , the localization RS = S −1 R of R at any multiplicative set S ⊂ SW is a sub-ring
of the field of fractions K of R:
r
RS = {
r ∈ R, s ∈ S} ⊂ K .
s
In particular, RS is an integral domain.
Proposition 2.15 Let S be any multiplicative subset of R such that S ⊂ SW . Then the localization functor locS : hmf(R, W ) → hmf(RS , WS ) restricts to an R-linear equivalence of categories
between hef(R, W ) and hef(RS , WS ).
Proof. Since locS preserves the reduced rank of matrix factorizations, it is clear that it restricts
to a functor from hef(R, W ) to hef(RS , WS ). Given two elementary factorizations ev1 , ev2 ∈
Ob[hef(R, W )], let r ∈ (v1 , v2 , W/v1 , W/v2 ). By Proposition 2.13, we have:
Homhef(RS ,WS ) ((ev1 )S , (ev2 )S ) 'R Homhef(R,W ) (ev1 , ev2 )S .
(2.23)
Let s be any element of S. Since S is a subset of SW , we have (s, W ) = (1) and hence (s, r) = (1)
since r is a divisor of W . By Lemma 2.14, the image s̄ = s + hri is a unit of the quotient ring
R/hri, hence the operator of multiplication with s is an isomorphism of the cyclic R-module
HomHMF(R,W ) (ev1 , ev2 ) ' R/hri. Thus every element of S acts as an automorphism of this module, which implies that the localization map Homhmf(R,W ) (ev1 , ev2 ) → Homhmf(R,W ) (ev1 , ev2 )S
is an isomorphism of R-modules (where Homhmf(R,W ) (ev1 , ev2 )S is viewed as an R-module
by the extension of scalars R → RS ). Combining this with (2.23) shows that the restriction
locS : hef(R, W ) → hef(RS , WS ) is a full and faithful functor.
Now let ex be an elementary factorization of WS corresponding to the divisor x of WS = W/1
in the ring RS . Let y = WS /x ∈ RS . Write x = v/s and y = u/t with x, y ∈ R and s, t ∈ S
chosen such that (v, s) = (u, t) = (1). Then the relation xy = WS amounts to uv = stW . Since
S is a subset of SW , we have (s, W ) = (t, W ) = (1). Thus st|uv, which implies s|v and t|u since
(v, s) = (u, t) = (1). Thus v = v1 t and u = u1 s with u1 , v1 ∈ R and we have u1 v1 = W . This
def.
gives x = γv1 and y = γ −1 u1 , where γ = t/s is a unit of RS . It follows that ex is similar to the
elementary matrix factorization ev1 of WS over RS , and hence isomorphic to the latter in the
category hef(RS , WS ) by Proposition 2.5. Since u1 and v1 are divisors of W satisfying u1 v1 = W ,
we can view ev1 as an elementary factorization of W over R (it lies in the image of the functor
locS ). This shows that any objects of hef(RS , WS ) is even-isomorphic with an object lying in
the image of the restricted localization functor, hence the latter is essentially surjective. t
u
2.4. Behavior of HEF(R, W ) under multiplicative partition of W . For any divisor W1 of W , let
HEFW1 (R, W ) denote the full subcategory of HEF(R, W ) whose objects are those elementary
factorization ev of W for which v is a divisor of W1 .
Proposition 2.16 Let e1 and e2 be as above. Consider elements of R chosen as follows:
s1 ∈ (u1 , v1 ) = (s)(a0 , c)(b, d0 ) , s2 ∈ (u2 , v2 ) = (s)(a0 , b)(c, d0 ) ,
def.
def.
def.
def.
u01 = u1 /s = cd0 , u02 = u2 /s = bd0 , v10 = v1 /s = a0 b , v20 = v2 /s = a0 c ,
x(e1 ) ∈ (s, v10 ) = (s, a0 b) , y(e1 ) ∈ (s, u01 ) = (s, cd0 ) ,
x(e2 ) ∈
(s, v20 )
0
= (s, a c) , y(e2 ) ∈
(s, u02 )
0
= (s, bd ) .
(2.24)
Elementary matrix factorizations over Bézout domains
19
Then:
1. e1 and e2 are isomorphic in hef(R, W ) iff:
(i) (s1 ) = (s2 ) and
(ii) (x(e1 )), (y(e1 )) = (x(e2 )), (y(e2 )) as ordered pairs of elements in R× /U (R).
(2.25)
2. e1 and e2 are isomorphic in HEF(R, W ) iff:
(i) (s1 ) = (s2 ) and
(ii) (x(e1 )), (y(e1 )) = (x(e2 )), (y(e2 ))
as unordered pairs of elements in R× /U (R).
(2.26)
Notice that (s1 ) = (s2 ) implies (s1 ) = (s) = (s2 ), with s defined in (2.3).
Proof.
1. Assume that e1 'hmf(R,W ) e2 . By Proposition 2.4, part 1, to such pair of elementary factorizations we can associate four pairwise coprime divisors a0 , b, c, d0 of W such that v1 = a0 bs,
u1 = d0 cs, v2 = a0 cs, u2 = d0 bs and together with the equality (bc, s) = (1). Thus (s1 ) =
(v1 , u1 ) = (a0 bs, d0 cs) = (s), since a0 b and d0 c are coprime. Similarly, (s2 ) = (s). The equality
(bc, s) = (1) is equivalent to (b, s) = (1) and (c, s) = (1). Using this, we compute:
(x(e1 )) = (s, v10 ) = (s, a0 b) = (s, a0 ) = (s, a0 c) = (s, v20 ) = (x(e2 ))
(2.27)
Acting similarly, we also find (y(e1 )) = (s, d0 ) = (y(e2 )). Thus (2.25) holds.
Now assume that (2.25) is satisfied for two elementary factorizations e1 and e2 . Let s ∈ (s1 ) =
(s2 ) and define a0 , b, c, d0 as before, following (2.1). By the very construction, (b, c) = (1) and
(a0 , d0 ) = (1). We first show that s1 ∼ s2 all a0 , b, c, d0 are pairwise coprime. Indeed, if we
assume that p|(a0 , b) then s1 ∼ s2 implies:
(s)(a0 , c)(b, d0 ) = (s1 ) = (s2 ) = (s)(a0 , b)(c, d0 ) .
Since p divides the right hand side, it should divide (a0 , c)(b, d0 ) and (c, b) = (1) and (a0 , d0 ) =
(1). Thus p ∈ U (R). It much the same way we prove that other pairs from a0 , b, c, d0 are
coprime.
Condition (ii) in (2.25) reads:
(s, a0 b) = (x(e1 )) = (x(e2 )) = (s, a0 c) .
If p|b and p|s then p|(s, b) and thus p|(s, a0 b). By the equality above, we also have p|(s, a0 c)
and hence p|a0 c. But b is coprime with both a0 and c, thus p ∈ U (R). Similarly, p|c and
p|s implies p ∈ U (R). Thus (bc, s) = (1). Note that (y(e1 )) = (y(e2 )) is now automatically
satisfied. Proposition 2.4, part 1 implies that e1 'hef(R,W ) e2 .
2. Assume e1 'HEF(R,W ) e2 . If the isomorphism is even, then it comes from the isomorphism
in hef(R, W ) and part 1 above already proves that (2.25) and thus also (2.26). Thus we can
assume that the isomorphism is odd. We will prove that (s1 ) = (s2 ) and (x(e1 )) = (y(e2 )),
(x(e2 )) = (y(e1 )). Applying Proposition 2.4, part 2, we obtain a0 , b, c, d0 pairwise coprime
and s such that (s, a0 d0 ) = (1). Then (s1 ) = (s) = (s2 ) similarly to part 1 above. Using
(s, a0 d0 ) = (1), we also compute:
(x(e1 )) = (s, v10 ) = (s, a0 b) = (s, b) = (s, d0 b) = (s, u02 ) = (y(e2 ))
20
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
and also (x(e2 )) = (y(e1 )). Thus (2.26).
Assume now that (2.26) is satisfied. Since the statement for even morphisms is covered by
(2.25), we only need to consider the situation (x(e1 )) = (y(e2 )) and (x(e2 )) = (y(e1 )). As in
part 1, (s1 ) = (s2 ) implies that a0 , b, c, d0 are pairwise coprime. Condition (ii) reads:
(s, a0 b) = (x(e1 )) = (y(e2 )) = (s, d0 b) .
If we assume that p|a0 and p|s then p|(s, a0 ) and p|(s, a0 b). The equality implies (p|d0 b). Since
a0 is coprime with both d0 and b, we obtain p ∈ U (R). Similarly (d0 , s) = (1) and thus
(a0 d0 , s) = (1). Proposition 2.4, part 2 implies that e1 'HEF(E,W ) e2 by an odd isomorphism.
t
u
Proposition 2.17 Let W1 and W2 be divisors of W such that W = W1 W2 and (W1 , W2 ) = (1).
Then there exist equivalences of R-linear Z2 -graded categories:
HEF(R, W1 ) ' HEFW1 (R, W ) , HEF(R, W2 ) ' HEFW2 (R, W ) .
which are bijective on objects.
Proof. For any divisor v of W , let e0v = (R1|1 , Dv0 ) and ev = (R1|1 , Dv ) be the corresponding
elementary factorizations of W1 and W , where:
Dv0
0 v
=
W1 /v 0
0 v
, Dv =
W/v 0
.
1
For any two divisors v1 , v2 of W1 and any κ ∈ Z2 , we have W/vi = W2 W
vi and (vi , W2 ) = (1).
Thus (v1 , v2 , W1 /v1 , W1 /v2 ) = (v1 , v2 , W/v1 , W/v2 ). By Proposition 2.2, this gives:
Ann(HomκHEF(R,W1 ) (e0v1 , e0v2 )) = Ann(HomκHEF(R,W ) (ev1 , ev2 )) , ∀κ ∈ Z2 .
On the other hand, the modules Hom0̂HEF(R,W1 ) (e0v1 , e0v2 ) and Hom0̂HEF(R,W1 ) (ev1 , ev2 ) are generated by the same element 0̂ (v1 , v2 ) while Hom1̂HEF(R,W1 ) (e0v1 , e0v2 ) and Hom1̂HEF(R,W1 ) (ev1 , ev2 )
are generated by the elements 1̂ (v1 , v2 ; W1 ) and 1̂ (v1 , v2 ; W ), respectively. Hence the functor
which maps e0v to ev for any divisor v of W1 and takes 0̂ (v1 , v2 ) to 0̂ (v1 , v2 ) and 0̂ (v1 , v2 ; W1 )
to 1̂ (v1 , v2 ; W ) for any two divisors v1 , v2 of W is an R-linear equivalence from HEF(R, W1 ) to
HEFW1 (R, W ). A similar argument establishes the equivalence HEF(R, W2 ) ' HEFW2 (R, W ).
t
u
2.5. Primary matrix factorizations. Recall that an element of R is called primary if it is a power
of a prime element.
Definition 2.18 An elementary factorization ev of W is called primary if v is a primary divisor
of W .
Let HEF0 (R, W ) denote the full subcategory of HEF(R, W ) whose objects are the primary
factorizations of W .
Elementary matrix factorizations over Bézout domains
21
Proposition 2.19 Let W = W1 W2 be a factorization of W , where W1 and W2 are coprime
elements of R. Then there exists an equivalence of R-linear Z2 -graded categories:
HEF0 (R, W ) ' HEF0 (R, W1 ) ∨ HEF0 (R, W2 ) ,
where ∨ denotes the coproduct of ModR -enriched categories.
Proof. Let HEF0,Wi (R, W ) denote the full subcategory of HEF0 (R, W ) whose objects are the
primary factorizations ev of W for which v is a (primary) divisor of Wi . Since W = W1 W2
and (W1 , W2 ) = (1), a primary element v ∈ R is a divisor of W iff it is either a divisor of W1
or a divisor of W2 . Hence ObHEF0 (R, W ) = ObHEF0,W1 (R, W ) t ObHEF0,W2 (R, W ). For any
primary divisors v1 and v2 of W and any κ ∈ Z2 , we have:
κ
HomHEF(R,W1 ) (ev1 , ev2 ) if v1 |W1 & v2 |W1
HomκHEF(R,W ) (ev1 , ev2 ) ' R/hdi ' HomκHEF(R,W2 ) (ev1 , ev2 ) if v1 |W2 & v2 |W2 ,
0
if v1 |W2 & v2 |W1
where d ∈ (v1 , v2 , W/v1 , W/v2 ) and in the third case we used the fact that v1 |W2 and v2 |W1 implies (v1 , v2 ) = (1) since W1 and W2 are coprime. This shows that HEF0 (R, W ) = HEF0,W1 (R, W )∨
HEF0,W2 (R, W ). By Proposition 2.17, we have R-linear equivalences HEF0,Wi (R, W ) ' HEF0 (R, Wi )
which are bijective on objects. This implies the conclusion. t
u
Definition 2.20 A reduced multiplicative partition of W is a factorization:
W = W1 W2 . . . Wn
where W1 , . . . , Wn are mutually coprime elements of R.
Corollary 2.21 Let W = W1 . . . Wn be a reduced multiplicative partition of W . Then there
exists a natural equivalence of R-linear categories:
HEF0 (R, W ) ' ∨ni=1 HEF0 (R, Wi ) .
Proof. Follows immediately from Proposition 2.19. t
u
Let ev be a primary matrix factorization of W . Then v = pi for some prime divisor p of W and
some integer i ∈ {0, . . . , n}, where n is the order of p as a divisor of W . We have W = pn W1
for some element W1 ∈ R such that p does not divide W1 and u = pn−i W1 . Thus (u, v) =
(pmin(i,n−i) ).
Definition 2.22 The prime divisor p of W is called the prime locus of ev . The order n of p is
called the order of ev while the integer i ∈ {0, . . . , n} is called the size of ev .
Let R be a Bézout domain and p ∈ R be a prime element. Fix an integer n ≥ 2 and consider
the quotient ring:
def.
An (p) = R/hpn i .
Let mn (p) = pAn (p) = hpi/hpn i and kp = R/hpi.
Lemma 2.23 The following statements hold:
1. The principal ideal hpi generated by p is maximal.
22
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
2. The primary ideal hpn i is contained in a unique maximal ideal of R.
3. The quotient An (p) is a quasi-local ring with maximal ideal mn (p) and residue field kp .
4. An (p) is a generalized valuation ring.
Proof.
1. Let I be any ideal containing hpi. If hpi =
6 I, then take any element x ∈ I \ hpi. Then we have
the proper inclusion hpi ( hp, xi. Since R is a Bézout domain, the ideal hp, xi is generated by
a single element y. We have y|p, so y is a unit of R since p is prime. Since y belongs to I, this
gives I = R. Thus hpi is a maximal ideal.
2. Let m be a maximal ideal of R containing hpn i. Then pn ∈ m, which implies p ∈ m since m
is prime. Thus hpi ⊂ m, which implies m = hpi since hpi is maximal by point 1. This shows
that R/hpn i has a unique maximal ideal, namely hpi/hpn i.
3. Since R is Bézout and hpn i is finitely-generated, the quotient R/hpn i is a Bézout ring (which
has divisors of zero when n ≥ 2). By point 2. above, R/hpn i is also a quasi-local ring.
4. Follows from [7, Lemma 1.3 (b)] since R is a valuation ring. t
u
Recall that an object of an additive category is called indecomposable if it is not isomorphic with
a direct sum of two non-zero objects.
Proposition 2.24 Let ev be a primary factorization of W with prime locus p, order n and size
i. Then ev is an indecomposable object of hmf(R, W ) whose endomorphism ring Endhmf(R,W ) (ev )
is a quasi-local ring isomorphic with Amin(i,n−i) (p).
Proof. We have Endhmf(R,W ) (ev ) = R/hu, vi = R/hpmin(i,n−i) i by Corollary 2.10. This ring is
quasi-local by Lemma 2.23. Since quasi-local rings have no nontrivial idempotents, it follows
that ev is an indecomposable object of hmf(R, W ). t
u
Lemma 2.25 Let v1 and v2 be two divisors of W which are mutually coprime. Then Homhmf(R,W ) (ev1 , ev2 ) =
0.
Proof. Let ui := W/vi . Then (v1 , v2 , u1 , u2 ) = (1) since (v1 , v2 ) = (1). Thus hv1 , v2 , u1 , u2 i = R
and the statement follows from Proposition 2.2. t
u
Proposition 2.26 Let p be a prime divisor of W of order n and i ∈ {1, . . . , n}. Then:
Σepi 'hmf(R,W ) epn−i .
def.
def.
def.
def.
Proof. Let W1 = pn , W2 = W/pn and v = pi , u = W/v = pn−i W2 . We have Σepi =
Σev = e−u 'hmf(R,W ) eu . Since pn−i is a divisor of W1 and (W1 , W2 ) = 1, Proposition 2.7 gives
eu = epn−i W2 'hmf(R,W ) epn−i . t
u
3. The additive category hef (R, W ) for a Bézout domain and critically-finite W
Let R be a Bézout domain and W be a critically-finite element of R.
Elementary matrix factorizations over Bézout domains
23
Q
Proposition 3.1 Let ev be an elementary factorization of W over R such that v = ni=1 vi ,
where vi ∈ R are mutually coprime divisors of W . Then there exists a natural isomorphism in
hmf(R, W ):
n
M
ev 'hmf(R,W )
evi
i=1
In particular, an elementary factorization ev for which v is finitely-factorizable divisor of W is
isomorphic in hmf(R, W ) with a direct sum of primary factorizations.
Proof. Let d be any divisor of W . By Proposition 2.2, we have isomorphisms of R-modules:
Homhmf(R,W ) (ed , evi ) 'R R/αW (d, vi ) and Homhmf(R,W ) (ed , ev ) 'R R/αW (d, v) .
Q
Since vi are mutually coprime, Proposition 2.3 gives αW (d, v) = ni=1 αW (d, vi ), where αW (d, vi )
are principal ideals generated by mutually coprime elements. The Chinese reminder theorem
gives an isomorphism of R-modules:
R/αW (d, v) 'R
n
M
[R/αW (d, vi )]
i=1
Combining the above, we conclude that there exist natural isomorphisms of R-modules:
∼
ϕd : Homhmf(R,W ) (ed , ev ) −→ Homhmf(R,W ) (ed ,
n
M
evi ) ,
(3.1)
i=1
Ln
Ln
where we used the fact that Homhmf(R,W ) (ed , i=1 evi ) 'R
i=1 Homhmf(R,W ) (ed , evi ). This
implies that the functors Homhef (R,W ) (−, ev ) and Homhef (R,W ) (−, ⊕ni=1 evi ) are isomorphic.
By
L
the Yoneda lemma, we conclude that there exists a natural ismorphism ev 'hef (R,W ) ni=1 evi .
t
u
Recall that a Krull-Schmidt category is an additive category for which every object decomposes
into a finite direct sum of objects having quasi-local endomorphism rings.
Theorem 3.2 The additive category hef (R, W ) is Krull-Schmidt and its non-zero indecomposable objects are the non-trivial primary matrix factorizations of W . In particular, hef (R, W ) is
additively generated by hef 0 (R, W ).
Proof. Suppose that W has the decomposition (1.4). Any elementary factorization ev of W
corresponds to a divisor v of W , which must have the form v = v0 pls11 . . . plsmm , where 1 ≤ s1 <
. . . < sm ≤ N and 1 ≤ li ≤ nsi , while v0 is a divisor of W0 . Applying Proposition 3.1 with vi = plsii
nN
W0 n1
for i ∈ {1, . . . , m}, we find ev 'hmf(R,W ) ⊕m
i=0 evi , where we defined u0 = W/v0 = v0 p1 . . . pN
ns −l
and ui = W/vi = W0 pn1 1 . . . psi i i . . . pnNN for i ∈ {1, . . . , m}. We have (u0 , v0 ) = (u0 , W0 /u0 ).
Since W0 = u0 v0 , it follows that (u0 , v0 )2 |(W0 ). Since W0 has no critical divisors, we must have
(u0 , v0 ) = (1) and hence ev0 'hmf(R,W ) 0. For i ∈ {1, . . . , m}, we have (ui , vi ) = pµsii , where
def.
µi = min(li , nsi − li ). Thus evi is primary of order µsi when µsi ≥ 1 and trivial when µi = 0.
This gives a direct sum decomposition:
ev 'hmf(R,W ) ev0 ⊕i∈{1,...,m|li ≤nsi −1} epli ' ⊕i∈{1,...,m|li <nsi } epli ,
si
si
24
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
where all matrix factorizations in the direct sum are primary except for ev0 . If li = nsi for all
i ∈ {1, . . . , m}, then the sum in the right hand side is the zero object of hmf(R, W ). We conclude
that any elementary matrix factorization decomposes into a finite direct sum of primary matrix
factorizations. On the other hand, any matrix factorization of W decomposes as a finite direct
sum of elementary factorizations and hence also as a finite direct sum of primary factorizations
whose prime supports are the prime divisors of W . By Proposition 2.24, every primary matrix
factorization has a quasi-local endomorphism ring. t
u
Theorem 3.3 Suppose that R is a Bézout domain and W has the decomposition (1.4). Then
there exists an equivalence of categories:
ni
hef (R, W ) ' ∨N
i=1 hef (R, pi ) ,
where ∨ denotes the coproduct of additive categories.
Proof. Theorem 3.2 and Proposition 3.1 imply that hef (R, W ) is additively generated by the
additive subcategories hef pni (R, W ) ' hef (R, pni i ), where we used Proposition 2.17. These
i
categories are mutually orthogonal by Lemma 2.25. t
u
3.1. A conjecture. Consider the inclusion functor:
ι : hef (R, W ) → hmf(R, W )
Conjecture 3.4 The inclusion functor ι is an equivalence of R-linear categories.
Conjecture 3.4 and Theorem 3.2 imply:
Conjecture 3.5 Let R be a Bézout domain and W be a critically-finite element of R. Then
hmf(R, W ) is a Krull-Schmidt category.
In [5], we establish Conjecture 3.4 for the case when R is an elementary divisor domain. This
shows that Conjecture 3.4 is implied by the still unsolved conjecture [8] that any Bézout domain
is an elementary divisor domain. Some recent work on that conjecture can be found in [9].
4. Counting elementary factorizations
In this section, we give formulas for the number of isomorphism classes of objects in the categories
HEF(R, W ) and hef(R, W ) when W is critically-finite.
4.1. Counting isomorphism classes in HEF(R, W ). Let W = W0 Wc be a critically-finite element
of R, where W0 ∈ R is non-critical and Wc = pn1 1 . . . pnr r with prime pj ∈ R and nj ≥ 2 (see
Definition 1.7). Let Hef (R, W ) denote the set of isomorphism classes of objects in the category
HEF(R, W ). We are interested in the cardinality:
def.
N (R, W ) = |Hef (R, W )|
Elementary matrix factorizations over Bézout domains
25
of this set. In this subsection, we derive a formula for N (R, W ) as a function of the orders
ni of the prime elements pi arising in the prime decomposition of Wc . The main result of this
subsection is Theorem 4.12 below.
Lemma 4.1 The cardinality N (R, W ) depends only on the critical part Wc of W .
Proof. Let W = puv with a divisor p coprime with both u and v. Taking b = p and c = 1 in
Proposition 2.4 gives:
0 pv
0 v
'
.
(4.1)
u 0
pu 0
Together with Corollary 2.5, this implies N (R, W ) = N (R, Wc ). t
u
From now on, we will assume that W ∈ R is fixed and is of the form:
W = Wc = pn1 1 pn2 2 . . . pnr r .
(4.2)
To simplify notations, we will omit to indicate the dependence of some quantities on W .
Definition 4.2 Let T be a non-empty set. A map f : ObEF(R, W ) → T is called an elementary
invariant if f (e1 ) = f (e2 ) for any e1 , e2 ∈ ObEF(R, W ) such that e1 'HEF(R,W ) e2 . An elementary invariant f is called complete if the map f : Hef (R, W ) → T induced by f is injective.
To determine N (R, W ), we will construct a complete elementary invariant. Let:
def.
I = {1, . . . , r} ,
where r is the number of non-associated prime factors of W , up to association in divisibility.
Similarity classes of elementary factorizations and normalized divisors of W . Let HEFsim (R, W )
be the groupoid having the same objects as HEF(R, W ) and morphisms given by similarity
transformations of elementary factorizations and let Hef sim (R, W ) be its set of isomorphism
classes. Since the similarity class of an elementary factorization ev is uniquely determined by
the principal ideal hvi generated by the divisor v of W , the map ev → hvi induces a bijection:
Hef sim (R, W ) ' Div(W ) ,
where:
def.
Div(W ) =
hvi, v|W
= hvi v ∈ R : W ∈ hvi
is the set of principal ideals of R containing W . Let:
Y k
def.
Div1 (W ) = {
pi i ∀i ∈ I : ki ∈ {0, . . . , ni } } ,
(4.3)
i∈I
be the set of normalized divisors of W . The map v → hvi induces a bijection between Div1 (W )
and Div(W ). Indeed, any principal ideal of R which contains W has a unique generator which
belongs to Div1 (W ), called its normalized generator. Given any divisor v of W , its normalization
v0 is the unique normalized divisor v0 ∈ Div1 (W ) such that hvi = hv0 i. Given two divisors t, s
of W , their normalized greatest common divisor is the unique normalized divisor (t, s)1 of W
which generates the ideal Rt + Rs. The set of exponent vectors of W is defined through:
def.
AW =
r
Y
{0, . . . , ni } .
i=1
26
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
The map AW 3 k = (k1 , . . . , kr ) 7→
AW given by:
Q
i∈I
pki i ∈ Div1 (W ) is bijective, with inverse µ : Div1 (W ) →
µ(v) = (ordp1 (v), . . . , ordpn (v)) = (k1 , . . . , kn ) for v =
Y
pki i ∈ Div1 (W ) .
i∈I
Combining everything, we have natural bijections:
Hef sim (R, W ) ' Div(R, W ) ' Div1 (R, W ) ' AW .
Remark 4.1. In Proposition 2.16, the quantity s1 was an arbitrary element of the class (u1 , v1 )
for ev1 . For a fixed critically-finite W , we have a canonical choice for this quantity, namely the
normalized gcd of u1 and v1 . Thus we define s(e1 ) = (u1 , v1 )1 . The two definitions are connected
by the relation (s1 ) = (s(e1 )). Below we introduce “normalized” quantities x(e), y(e) which
belong to the same classes in R× /U (R) as the quantities x and y defined in Section 2. The
results of Section 2 hold automatically for these normalized choices.
Given t ∈ Div(W ), its index set is the subset of I given by:
def.
I(t) = supp µ(t0 ) = i ∈ I
pi |t
.
Notice that I(t) depends only on the principal ideal hti, that in turn depends only on the class
(t) ∈ R× /U (R). This gives a map from Div(W ) to the power set P(I) of I. Note that (t) = (1)
iff I(t) = ∅.
The essence and divisorial invariant of an elementary factorization. Consider an elementary
factorization e of W and let v(e) ∈ Div1 (W ) be the unique normalized divisor of W for which
def.
def.
e is similar to ev . Let u = u(e) = W/v ∈ Div1 (W ) and let s(e) = (v, u)1 ∈ Div1 (W ) be
def.
def.
the normalized greatest common denominator of v and u. Let Is (e) = I(s(e)). Let mi (e) =
ordpi (s(e)) and m(e) = µ(s(e)) = (m1 (e), . . . , mn (e)). Then Is (e) = supp m(e) and:
s(e) =
Y
mi (e)
pi
.
(4.4)
i∈Is (e)
def.
def.
Let v 0 (e) = v/s(e) and u0 (e) = u/s(e). Then (v 0 (e), u0 (e)) = (1) and W = v(e)u(e) =
v 0 (e)u0 (e)s(e)2 . Define:
def.
def.
x(e) = (s(e), v 0 (e))1 , y(e) = (s(e), u0 (e))1
(4.5)
and:
def.
def.
Ix (e) = I(x(e)) , Iy (e) = I(y(e)) .
def.
def.
Notice that (x(e), y(e))1 = 1, thus Ix (e) ∩ Iy (e) = ∅. Defining v 00 (e) = v 0 (e)/x(e) and u00 (e) =
u0 (e)/y(e), we have:
W = x(e)y(e)v 00 (e)u00 (e)s(e)2 ,
where v 00 (e), u00 (e) and s(e) are mutually coprime. Moreover, we have:
ordpi (v 0 (e)) = ni − 2mi (e) ∀i ∈ Ix (e) and ordpi (u0 (e)) = ni − 2mi (e) ∀i ∈ Iy (e) ,
Elementary matrix factorizations over Bézout domains
27
which implies:
ordpi x(e) = max(mi (e), ni − 2mi (e)) for i ∈ Ix (e) ,
ordpi y(e) = max(mi (e), ni − 2mi (e)) for i ∈ Iy (e) .
(4.6)
Notice that ordpi x(e)y(e) ≡ ni mod 2 for i ∈ Ix ∪ Iy if 3mi < ni .
Definition 4.3 The essence z := z(e) of an elementary factorization e of W is the normalized
divisor of W defined through:
Y m (e)
def.
def.
pi i , where Iz (e) = I(s)\ Ix (e) ∪ Iy (e) .
(4.7)
z(e) =
Iz (e)
An elementary factorization e is called essential if z(e) = 1, i.e. if Iz (e) = ∅.
The divisor s defines x, y and z uniquely by (4.6) and (4.7). These 3 divisors in turn also define
s uniquely that can be seen by the inverting the max functions above:
ordpi x(e)
if i ∈ Ix (e) and 3 ordpi x(e) ≥ ni
(ni − ordpi x(e))/2 if i ∈ Ix (e) and 3 ordpi x(e) < ni
ordpi s = mi = ordpi y(e)
(4.8)
if i ∈ Iy (e) and 3 ordpi y(e) ≥ ni
(ni − ordp y(e))/2 if i ∈ Iy (e) and 3 ordp y(e) < ni
i
i
ordpi z(e),
if i ∈ Iz (e) .
The fundamental property of an essential factorization of e is the equality of sets Is (e) = Ix (e) t
Iy (e), which will allow us to compute the number N∅ (R, W ) of isomorphism classes of such
factorizations (see Proposition 4.11 below). Then N (R, W ) will be determined by relating it to
N∅ for various reductions of the potential W .
Notice that the essence z(e) is a critical divisor of W and that we have (z(e), v 0 (e))1 =
(z(e), u0 (e))1 = 1. Since W = v(e)u(e) = v 0 (e)u0 (e)s(e)2 , this gives:
ordpi W = ni = 2mi (e) = 2ordpi z(e) for any i ∈ Iz (e) .
(4.9)
Definition 4.4 The divisorial invariant of an elementary factorization e of W is the element
h(e) of the set Div1 (W ) × Sym2 (P(I)) defined through:
h(e) = (s(e), {Ix (e), Iy (e)}) .
This gives a map h : EF(R, W ) → Div1 (W ) × Sym2 (P(I)).
We have already given a criterion for two elementary factorizations of W to be isomorphic in
Proposition 2.16. There exists another way to characterize when two objects of HEF(R, W ) (and
also of hef(R, W )) are isomorphic, which will be convenient for our purpose.
Proposition 4.5 Consider two elementary factorizations of W . The following statements are
equivalent:
1. The two factorizations are isomorphic in HEF(R, W ) (respectively in hef(R, W )).
28
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
2. The two factorizations have the same (s, {x, y}) (respectively same (s, x, y)).
3. The two factorizations have the same divisorial invariant (s, {Ix , Iy }) (respectively same
(s, Ix , Iy )).
In particular, the divisorial invariant h : ObEF(R, W ) → Div1 (W ) × Sym2 (P(I)) is a complete
elementary invariant.
Proof. The equivalence between 1. and 2. follows from Proposition 2.16. Indeed, the proposition
shows that for two isomorphic factorizations e1 and e2 the corresponding s1 and s2 are similar: (s1 ) = (s2 ) in the notations of Section 2. We compute (s2 ) = (u2 , v2 ) = ((u2 )1 , (v2 )1 ) =
((u(e2 ), v(e2 ))1 ) = (s(e2 )) with the last s(e2 ) defined in Div1 by (4.4). Similarly (s1 ) = (s(e1 )).
By the very definition of Div1 we have (s(e1 )) = (s(e2 )) implies s(e1 ) = s(e2 ).
The implication 2. ⇒ 3. is obvious. Thus it suffices to prove that 3. implies 2. For this,
let e1 and e2 be the two elementary factorizations of W . Assume
that s(e1 ) = s(e2 ) and
Q
i
{Ix (e1 ), Iy (e1 )} = {Ix (e2 ), Iy (e2 )} and let s := s(e1 ) = s(e2 ) = i∈I(s) pm
i . Consider the case
Ix (e1 ) = Ix (e2 ) and Iy (e1 ) = Iy (e2 ). Applying (4.6) to v = v(e1 ) and v = v(e2 ) and using the
relations mi (e1 ) = ordpi s = mi (e2 ) gives:
x(e1 ) = x(e2 ) and y(e1 ) = y(e2 ) .
When Ix (e1 ) = Iy (e2 ) and Ix (e2 ) = Iy (e1 ), a similar argument gives x(e1 ) = y(e2 ) and y(e1 ) =
x(e2 ). t
u
Proposition 4.6 The map z : ObEF(R, W ) → Div1 (W ) which gives the essence of an elementary factorization is an elementary invariant.
Proof. Let e1 and e2 be two factorizations of W which are isomorphic in HEF(R, W ). By Proposition 4.5, we have s(e1 ) = s(e2 ) and {Ix (e1 ), Iy (e1 )} = {Ix (e2 ), Iy (e2 )}. Hence:
I(z(e1 )) = I(s(e1 ))\ Ix (e1 ) ∪ Iy (e1 ) = I(s(e2 ))\ Ix (e2 ) ∪ Iy (e2 ) = I(z(e2 )) .
(4.10)
Applying (4.9) for e = e1 and e = e2 gives ordpi z(e1 ) = ordpi z(e2 ) for any i ∈ I(z(e1 )) = I(z(e2 )).
Thus z(e1 ) = z(e2 ). t
u
The essential reduction of an elementary factorization. For any normalized critical divisor z of
W , let HEFz (R, W ) denote the full subcategory of HEF(R, W ) consisting of those elementary
factorizations whose essence equals z and let Hef z (R, W ) be its set of isomorphism classes.
Then Hef 1 (R, W ) consists of the isomorphism classes of essential factorizations.
Definition 4.7 The essential reduction of an elementary factorization e := ev of W is the
essential elementary factorization of W/z(e)2 defined through:
def.
essred(e) = ev/z(e) .
This gives a map essred : ObEF(R, W ) → ObHEF1 (R, W/z(e)2 ).
To see that essred is well-defined, consider the elementary factorization ẽ = ev/z(e) :
f def.
W
= W/z(e)2 = u(ẽ)v(ẽ) ,
(4.11)
Elementary matrix factorizations over Bézout domains
29
where v(ẽ) = v(e)/z(e) and u(ẽ) = u(e)/z(e). We compute:
def.
s(ẽ) = (v(ẽ), u(ẽ))1 = (v(e)/z(e), u(e)/z(e))1 = s(e)/z(e)
def.
def.
def.
and v 0 (ẽ) = v(ẽ)/s(ẽ) = v 0 (e), u0 (ẽ) = u(ẽ)/s(ẽ) = u0 (e). Thus x(ẽ) = (s(ẽ), v 0 (ẽ))1 = x(e)
def.
and y(ẽ)
= (s(ẽ), u0 (ẽ))1 = y(e). By
(4.7) applied to ẽ and e, we derive Iz (ẽ) = Is (ẽ)\ Ix (ẽ) ∪
Iy (ẽ) = Iz (e) ∪ Is (e)\ Ix (e) ∪ Iy (e) = ∅, which implies z(ẽ) = 1. Hence essred(e) is an essential
f . Also notice the relation:
elementary matrix factorization of W
z(e), W/z(e)2 = (1) ,
which follows from the fact that z(e) is coprime with v 0 (e) and u0 (e).
Lemma 4.8 For any critical divisor z of W such that (z, W/z 2 ) ∼ 1, the map essred induces a
∼
well-defined bijection essredz : Hef z (R, W ) → Hef 1 (R, W/z 2 ) .
Proof. We perform the proof in two steps:
1. Let e1 := ev1 and e2 :=ev2 be two elementary factorizations of W such that z(e1 ) = z(e2 ) = z
def.
def.
and let v1 = v(e1 ), v2 = v(e2 ). Define also v3 = v1 /z and v4 = v2 /z. To show that essredz is
well-defined, we have to show that e1 'HEF(R,W ) e2 implies that the two essential elementary
f def.
f ). For this,
factorizations e3 := ev3 and e4 := ev4 of W
= W/z 2 are isomorphic in HEF(R, W
we compute:
def.
s(e1 ) = (v1 , u1 )1 = (z(e1 ) · v3 , z(e1 ) · u3 )1 = z(e1 ) · (v3 , u3 )1 = z(e1 ) · s(e3 ) .
Thus:
def.
x(e1 ) = (s(e1 ), v10 )1 = (z(e1 ) · s(e3 ), z(e1 )v3 /z(e1 )s(e3 ))1 = (s(e3 ), v30 )1 = x(e3 ).
The third equality above holds since (z(e1 ), v10 )1 = 1 and thus (z(e1 ), v3 )1 = 1. Similarly, we
have s(e2 ) = z(e2 ) · s(e4 ) and we find y(e1 ) = y(e3 ) as well as x(e2 ) = x(e4 ) and y(e2 ) =
y(e4 ). By Proposition 4.5, the condition e1 'HEF(R,W ) e2 implies
s(e1 ) = s(e2 ) = s and
I(s(e1 )) = I(s(e2 )), thus z(e
)
=
z(e
)
=
z.
If
s(e
),
{x(e
),
y(e
)}
=
s(e
),
{x(e
),
y(e
)}
,
1
2
1
1
1
2
2
2
then s(e3 ), {x(e3 ), y(e3 )} = s(e4 ), {x(e4 ), y(e4 )} . Thus e3 ' e4 .
2. Let z be a critical divisor of W such that (z, W/z 2 ) = (1). For any essential elementary
f , the elementary factorization ezv of W is an object of HEFz (R, W )
factorization ev of W
and we have essred(evz ) = ev . This shows that essredz is surjective. Now let e3 and e4 be
f def.
f ).
two essential elementary factorizations of W
= W/z 2 which are isomorphic in HEF(R, W
def.
def.
Let v1 = zv3 and v2 = zv4 . To show that essredz is injective, we have to show that the
two elementary factorizations e1 := ev1 and
in HEF(R, W ).
e2 := ev2 of W are isomorphic
For this, notice that s(e3 ), {x(e
3 ), y(e3 )} = s(e4 ), {x(e
4 ), y(e4 )} by Proposition 4.5. This
implies s(e1 ), {x(e1 ), y(e1 )} = s(e2 ), {x(e2 ), y(e2 )} , with z(e1 ) = z(e2 ) = z. Hence e1 and
e2 are isomorphic in HEF(R, W ) by the same proposition. t
u
A formula for N (R, W ) in terms of essential reductions. Let:
def.
S = im h ⊂ Div1 (W ) × Sym2 (P(I)) .
(4.12)
30
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
The degrees of the prime factors pi in the decomposition (4.2) of W define on I = {1, . . . , r} a
Z2 -grading given by:
def.
def.
I 0̂ = {i ∈ I ni is even} , I 1̂ = {i ∈ I ni is odd} .
(4.13)
Let:
def.
def.
r0̂ = |I 0̂ | , r1̂ = |I 1̂ | .
Since I = I 0̂ t I 1̂ , we have r = r0̂ + r1̂ . Any non-empty subset K ⊂ I is endowed with the Z2 grading induced from I. For any critical divisor z of W , we have z 2 |W , which implies I(z) ⊂ I 0̂ .
For any subset J ⊂ I 0̂ , define:
Y n /2
def.
pi i
,
(4.14)
zJ =
i∈J
which is a normalized critical divisor of W satisfying (zJ , W/zJ2 )1 = 1. Also define:
def.
SJ = h(Hef zJ (R, W )) ⊂ S .
and:
def.
NJ (R, W ) = |SJ | .
(4.15)
Since h is a complete elementary invariant, we have N∅ (R, W ) = |h(Hef 1 (R, W ))| = |Hef 1 (R, W )|.
Moreover, Lemma 4.8 gives:
NJ (R, W ) = N∅ (R, W/zJ2 ) .
(4.16)
Proposition 4.9 We have:
N (R, W ) =
X
N∅ (R, W/zJ2 ) .
(4.17)
J⊂I 0̂
Proof. Follows immediately from Lemma 4.8 and the remarks above. t
u.
Computation of N∅ (R, W ). Notice that z∅ = 1. Since h is a complete elementary invariant, we
have N∅ (R, W ) = |S∅ |, where:
S∅ = {h(e) e ∈ ObHEF(R, W ) : z(e) = 1} .
We will first determine the cardinality of the set:
def.
S∅,k = {h(e) e ∈ ObHEF(R, W ) : z(e) = 1 and |Is (e)| = k} .
We have:
S∅ = trk=1 S∅,k .
Lemma 4.10 For k ≥ 1, we have:
|S∅,k | = 2k−1 ·
X Y ni − 1
.
2
K⊂I, i∈K
|K|=k
(4.18)
Elementary matrix factorizations over Bézout domains
31
Proof. Consider a subset K ⊂ I(s) of cardinality |K| = k. Since s2 |W , we have:
1 ≤ mi (ev ) ≤ b(ni − 1)/2c ∀i ∈ I(s) .
Q
There are i∈K ni2−1 different possibilities for s such that I(s) = K. We can have several
elements (s, {x, y}) of S1,k with the same s since x and y can vary. This is where the coefficient
2k−1 in front in (4.18) comes from, as we now explain. Fixing the set I(s) with |I(s)| = k, we
have a set P(I(s)) of 2k partitions I(s) = Ix t Iy as disjoint union of 2 sets. These can be
parameterized by the single subset Iv ⊂ I(s) since Iu = I(s)\Iv . Define:
S∅,k,s = {h(e) e ∈ ObHEF(R, W ) : z(e) = 1 and |Is (e)| = k and s(e) = s} .
Consider the surjective map
Ψ : P(I(s)) → S∅,k,s
which sends a partition β = (I1 , I2 ) of I(s) to the element α = (s, {I1 , I2 }). The preimage
Ψ −1 (h(e)) of an element h(e) ∈ S∅,k,s consist of two elements : (I1 , I2 ) and (I2 , I1 ). Thus the
map is 2:1. This holds for every K with |K| = k. Comparing the cardinalities of P(I(s)) and
S∅,k,s , we find:
|S∅,k,s | = |P(I(s))|/2 = 2k /2 = 2k−1 .
This holds for any s with I(s) = K, where K ⊂ I has cardinality k. Since S∅,k = ts S∅,k,s and
since the cardinality |S∅,k,s | does not depend on s, we find:
|S∅,k | =
X
s
Y ni − 1
X
|S∅,k,s | = 2k−1
K⊂I:|K|=k i∈K
2
.
t
u
Proposition 4.11 With the definitions above, we have:
N∅ (R, W ) = |S∅ | = 1 +
r
X
2k−1
X Y ni − 1
.
2
(4.19)
K⊂I i∈K
|K|=k
k=1
Proof. Since S∅ = trk=1 S∅,k , the the previous lemma gives:
|S∅ | = 1 +
r
X
|S∅,k | = 1 +
r
X
2k−1
k=1
k=1
X Y ni − 1
.
2
(4.20)
K⊂I i∈K
|K|=k
The term 1 in front corresponds to the unique element (1, {∅, ∅}) of S. t
u
Computation of N (R, W ). The main result of this subsection is the following:
Theorem 4.12 The number of isomorphism classes of HEF(R, W ) for a critically-finite W as
in (4.2) is given by:
1̂
N (R, W ) = 2
r0̂
+
r
X
k=0
2r
0̂ +k−1
X Y ni − 1
.
2
K(I i∈K
|K 1̂ |=k
(4.21)
32
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
Proof. Combining Proposition 4.9 and Proposition 4.11, we have:
N (R, W ) =
X
1+
r−j
X
2k−1
K⊂I\J i∈K
|K|=k
k=1
J⊂I 0̂
X Y ni − 1
2
,
(4.22)
def.
where j = |J|. We will simplify this expression by changing the summation signs and applying
the binomial formula.
Since r0 = |I0 | and J ⊂ I 0̂ , we have j ≤ r0 . For fixed j, we have rj0 different subsets J ⊂ I 0̂
f ) of any such J has the free coefficient 1. Then
of this cardinality. The contribution to N∅ (R, W
the free coefficient of N (R, W ) is:
r0̂ 0̂
X
r
0̂
= 2r .
(4.23)
j
j=0
For the other coefficients of (4.22), we consider a subset K ⊂ I as in (4.19). Such an index set
f ) if K 0̂ ⊂ I 0̂ \J. The coefficient
K = K 1̂ t K 0̂ ⊂ I of cardinality k = k 1̂ + k 0̂ appears in N∅ (R, W
Q
f ) is 2k−1 for any of r0̂ −k0̂ choices of J. It follows that this coefficient
of i∈K ni2−1 in N (R, W
j
in N (R, W ) is:
0̂ −k 0̂
0̂
rX
0̂
0̂
1̂
0̂
0̂
0̂
1̂
k−1 r − k
2
= 2k +k −1 2r −k = 2r +k −1 .
(4.24)
j
j=0
Together with (4.23) and (4.24), relation (4.22) gives:
0̂
N (R, W ) = 2r +
X
∅⊂K(I
|K 1̂ |=k1̂
2r
0̂ +k 1̂ −1
Y ni − 1
i∈K
2
,
which is equivalent to (4.21). t
u
The two examples below illustrate how the coefficients behave for r = 2 and r = 3.
Example 4.1. Let W = pn1 pm
2 for prime elements p1 , p2 ∈ R such that (p1 ) 6= (p2 ) and n, m ≥ 2
with odd n and m. Then:
N (R, W ) = 1 +
n − 1 m − 1
+
.
2
2
k
Example 4.2. Consider W = pn1 pm
2 p3 for primes p1 , p2 , p3 ∈ R which are mutually non-associated
in divisibility and orders n, m, k ≥ 2 subject to the condition that n and m are even while k is
odd. Then:
n − 1
m − 1 k − 1
N (R, W ) = 2 + 2
+2
+
+
2
2
2
n − 1 m − 1
n − 1 k − 1
n − 1 m − 1
4
+2
+2
. (4.25)
2
2
2
2
2
2
Elementary matrix factorizations over Bézout domains
33
4.2. Counting isomorphism classes in hef(R, W ). We next derive a formula for the number of
isomorphism classes in the category hef(R, W ) for a critically-finite W (see Theorem 4.21 below).
Since the morphisms of hef(R, W ) coincide with the even morphisms of HEF(R, W ), the number
of isomorphism classes of hef(R, W ) is larger than N (R, W ). The simplest difference between
the two cases arises from the fact that suspension does not preserve the ismorphism class of an
elementary factorization in the category hef(R, W ). Let Ȟef (R, W ) be the set of isomorphism
classes of objects in hef(R, W ) and:
Ň (R, W ) = |Ȟef (R, W )| .
Lemma 4.13 The cardinality Ň (R, W ) depends only on the critical part Wc of W .
Proof. The proof is identical to that of Lemma 4.1.
Definition 4.14 Let T be a non-empty set. A map f : ObEF(R, W ) → T is called an even
elementary invariant if f (e1 ) = f (e2 ) for any e1 , e2 ∈ ObEF(R, W ) such that e1 'hef(R,W ) e2 .
An even elementary invariant f is called complete if the map f : Ȟef (R, W ) → T induced by f
is injective.
As in the previous subsection, we will compute Ň (R, W ) by constructing an even complete
elementary invariant.
Definition 4.15 The even divisorial invariant of an elementary factorization e of W is the
element ȟ(e) of the set Div1 (W ) × P(I)2 defined through:
ȟ(e) = (s(e), Ix (e), Iy (e)) .
This gives a map ȟ : EF(R, W ) → Div1 (W ) × P(I)2 .
Lemma 4.16 The even divisorial invariant ȟ : ObEF(R, W ) → Div1 (W ) × P(I)2 is a complete
even elementary invariant.
Proof. By Proposition 4.5, two elementary factorizations of W are isomorphic in hef(R, W ) iff
they have the same (s, x, y), which in turn is equivalent with coincidence of their even elementary
invariants. t
u
Using the essence z(e) defined in (4.7), each elementary factorization ev of W determines an
def.
f def.
elementary factorization essred(e) = ev/z(e) of W
= W/z(e)2 (see Definition 4.7). For any
normalized critical divisor z of W , let hef z (R, W ) denote the full subcategory of hef(R, W )
consisting of those elementary factorizations whose essence equals z and let Ȟef z (R, W ) be its
set of isomorphism classes.
Lemma 4.17 For any critical divisor z of W such that (z, W/z 2 ) = (1), the map essred induces
∼
a well-defined bijection ěssredz : Ȟef z (R, W ) → Ȟef 1 (R, W/z 2 ).
Proof. The proof is almost identical to that of Lemma 4.8, but taking into account that in
hef(R, W ) we deal only with the even morphisms of HEF(R, W ). t
u
34
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
Let:
def.
Š = im h ⊂ Div1 (W ) × P(I)2 .
(4.26)
For a subset J ⊂ I 0̂ , let:
def.
ŠJ = ȟ(Ȟef zJ (R, W )) ⊂ Š ,
def.
where zJ was defined in (4.14). Define ŇJ (R, W ) = |ŠJ | and Ň∅ (R, W ) = |ȟ(Ȟef 1 (R, W ))| =
|Ȟef 1 (R, W )|, where the last equality holds since ȟ is a complete even elementary invariant. We
can again compute Ň (R, W ) in terms of Ň∅ (R, W ):
Proposition 4.18 We have:
X
Ň (R, W ) =
Ň∅ (R, W/zJ2 ) .
(4.27)
J⊂I 0̂
Proof. Follows from Lemma 4.17. t
u
Define:
Š∅ = ȟ(e) e ∈ ObHEF(R, W ) and z(e) = 1
and:
def.
Š∅,k =
ȟ(e) e ∈ ObHEF(R, W ) , z(e) = 1 and |Is (e)| = k
.
Lemma 4.19 For k ≥ 1, we have:
|Š∅,k | = 2k ·
X Y ni − 1
.
2
(4.28)
K⊂I, i∈K
|K|=k
Proof. The proof is similar toQthat of Lemma
4.10. Consider a subset K ⊂ I(s) of cardinality k.
ni −1
As in Lemma 4.10, there are i∈K
different possibilities for s such that I(s) = K. Fixing
2
the set I(s) with |I(s)| = k, we have a set P(I(s)) of 2k partitions I(s) = Ix t Iy . Define:
Š∅,k,s = ȟ(e) e ∈ Ob hef(R, W ) , z(e) = 1 , |Is (e)| = k and s(e) = s .
The map which sends a partition β = (I1 , I2 ) of I(s) to the element α = (s, I1 , I2 ) is a bijection.
We compute:
|Š∅,k,s | = |P(I(s))| = 2k .
This holds for any s with I(s) = K, where K ⊂ I has cardinality k. Since Š∅,k = ts Š∅,k,s and
since the cardinality |Š∅,k,s | does not depend on s, we find:
|Š∅,k | =
X
X
|Š∅,k,s | = 2k ·
s
Y ni − 1
K⊂I:|K|=k i∈K
2
.
t
u
An immediate consequence is the following:
Proposition 4.20 With the definitions above, we have:
Ň∅ (R, W ) = |Š∅ | =
r
X
k=0
2k
X Y ni − 1
.
2
K⊂I i∈K
|K|=k
(4.29)
Elementary matrix factorizations over Bézout domains
35
We are now ready to compute Ň (R, W ).
Theorem 4.21 The number of isomorphism classes of the category hef(R, W ) for a criticallyfinite W as in (4.2) is given by:
1̂
Ň (R, W ) =
r
X
X
2r
0̂ +k
k=0 K(I,
|K 1̂ |=k
Y ni − 1
i∈K
2
,
(4.30)
Proof. Using Proposition 4.18 and Proposition 4.20, we write:
Ň (R, W ) =
X
r−j
X
J⊂I 0̂
k=0
X Y ni − 1
2k
K⊂I\J i∈K
|K|=k
2
,
(4.31)
def.
where j = |J|. Consider a subset K ⊂ I as in (4.29). Such an index set K = K 1̂ t K 0̂ ⊂ I of
ni −1
f ) if K 0̂ ⊂ I 0̂ \J. The coefficient of Q
cardinality k = k 1̂ + k 0̂ appears in Ň∅ (R, W
in
i∈K
2
r0̂ −k0̂
k
f
Ň (R, W ) is 2 for any of
choices of J. It follows that this coefficient in Ň (R, W ) is:
j
0̂ −k 0̂
rX
j=0
0̂
r − k 0̂
0̂
1̂
0̂
0̂
0̂
1̂
2
= 2k +k 2r −k = 2r +k .
j
k
Together with (4.32), relation (4.31) yields (4.30).
(4.32)
t
u
5. Some examples
In this section, we discuss a few classes of examples to which the results of the previous sections
apply. Subsection 5.1 considers the ring of complex-valued holomorphic functions defined on a
smooth, non-compact and connected Riemann surface, which will be discussed in more detail in
a separate paper. Subsection 5.2 considers rings arising through the Krull-Kaplansky-JaffardOhm construction, which associates to any lattice-ordered Abelian group a Bézout domain
having that ordered group as its group of divisibility. Subsection 5.3 discusses Bézout domains
with a specified spectral poset, examples of which can be produced by a construction due to
Lewis.
5.1. Elementary holomorphic factorizations over a non-compact Riemann surface. Let Σ be any
non-compact connected Riemann surface (notice that such a surface need not be algebraic and
that it may have infinite genus and an infinite number of ends). It is known that the cardinal
Krull dimension of O(Σ) is independent of Σ and is greater than or equal to 2ℵ1 (see [10, 11]).
The following classical result (see [12, 13]) shows that the C-algebra of holomorphic functions
entirely determines the complex geometry of Σ.
Theorem 5.1 (Bers) Let Σ1 and Σ2 be two connected non-compact Riemann surfaces. Then
Σ1 and Σ2 are biholomorphic iff O(Σ1 ) and O(Σ2 ) are isomorphic as C-algebras.
36
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
A Bézout domain R is called adequate if for any a ∈ R× and any b ∈ R, there exist r, s ∈ R such
that a = rs, (r, b) = (1) and any non-unit divisor s0 of s satisfies (s0 , b) 6= (1). It is known that
any adequate Bézout domain R is an elementary divisor domain, i.e. any matrix with elements
from R admits a Hermite normal form. The following result provides a large class of examples
of non-Noetherian adequate Bézout domains:
Theorem 5.2 For any smooth and connected non-compact Riemann surface Σ, the ring O(Σ)
is an adequate Bézout domain and hence an elementary divisor domain.
Proof. The case Σ = C was established in [8, 14]. This generalizes to any Riemann surface using
[11,12]. Since O(Σ) is an adequate Bézout domain, it is also a P M ∗ ring4 [8] and hence [15]
an elementary divisor domain. The fact that O(Σ) is an elementary divisor domain can also be
seen as follows. Guralnick [16] proved that O(Σ) is a Bézout domain of stable range one. By
[17,18], this implies that O(Σ) is an elementary divisor domain. t
u
The prime elements of O(Σ) are those holomorphic functions f : Σ → C which have a single
simple zero on Σ. This follows, for example, from the Weierstrass factorization theorem on noncompact Riemann surfaces (see [19, Theorem 26.7]). A critically-finite element W ∈ O(Σ) has
the form W = W0 Wc , where W0 : Σ → C is a holomorphic function with (possibly infinite)
number of simple zeroes and no multiple zeroes while Wc is a holomorphic function which has
only a finite number of zeroes, all of which have multiplicity at least two. All results of this
paper apply to this situation, allowing one to determine the homotopy category hef(R, W )
of elementary D-branes (and to count the isomorphism classes of such) in the corresponding
holomorphic Landau-Ginzburg model [3, 4] defined by (Σ, W ).
5.2. Constructions through the group of divisibility. Recall that the group of divisibility G(R)
of an integral domain R is the quotient group K × /U (R), where K is the quotient field of R
and U (R) is the group of units of R. This is a partially-ordered Abelian group when endowed
with the order induced by the R-divisibility relation, whose positive cone equals R× /U (R).
Equivalently, G(R) is the group of principal non-zero fractional ideals of R, ordered by reverse
inclusion. Since the positive cone generates G(R), a theorem due to Clifford implies that G(R)
is a directed group (see [20, par. 4.3]). It is an open question to characterize those directed
Abelian groups which arise as groups of divisibility of integral domains. It is known that G(R)
is totally-ordered iff R is a valuation domain, in which case G(R) is order-isomorphic with the
value group of R and the natural surjection of K × to G(R) gives the corresponding valuation.
Moreover, a theorem due to Krull [21] states that any totally-ordered Abelian group arises as
the group of divisibility of some valuation domain. It is also known5 that R is a UFD iff G(R) is
order-isomorphic with a (generally infinite) direct sum of copies of Z endowed with the product
order (see [20, Theorem 4.2.2]).
An ordered group (G, ≤) is called lattice-ordered if the partially ordered set (G, ≤) is a lattice,
i.e. any two element subset {x, y} ⊂ G has an infimum inf(x, y) and a supremum sup(x, y) (these
two conditions are in fact equivalent for a group order); in particular, any totally-ordered Abelian
group is lattice-ordered. Any lattice-ordered Abelian group is torsion-free (see [22, p. 10] or [23,
A P M ∗ -ring is a unital commutative ring R which has the property that any non-zero prime ideal of R is
contained in a unique maximal ideal of R.
5
Notice that a UFD is a Bézout domain iff it is Noetherian iff it is a PID (see Appendix B).
4
Elementary matrix factorizations over Bézout domains
37
15.7]). The divisibility group G(R) of an integral domain R is lattice-ordered iff R is a GCD
domain [20]. In particular, the group of divisibility of a Bézout domain is a lattice-ordered group.
When R is a Bézout domain, the prime elements of R are detected by the lattice-order of
def.
G(R) as follows. Given any Abelian lattice-ordered group (G, ≤) and any x ∈ G, let ↑ x = {y ∈
def.
G|x ≤ y} and ↓ x = {y ∈ G|y ≤ x} denote the up and down sets determined by x. A positive
filter of (G, ≤) is a filter of the lattice (G+ , ≤), i.e. a proper subset F ⊂ G+ having the following
two properties:
1. F is upward-closed, i.e. x ∈ F implies ↑ x ⊂ F
2. F is closed under finite meets, i.e. x, y ∈ F implies inf(x, y) ∈ F .
Notice that ↑ x is a positive filter for any x ∈ G+ . A positive filter F of (G, ≤) is called:
(a) prime, if G+ \ F is a semigroup, i.e. if x, y ∈ G+ \ F implies x + y ∈ G+ \ F .
(b) principal, if there exists x ∈ F such that F =↑ x.
If R is a Bézout domain with field of fractions K and group of divisibility G = K × /U (R), then
the natural projection π : K × → G induces a bijection between the set of proper ideals of R
and the set of positive filters of G, taking a proper ideal I to the positive filter π(I \ {0}) and
a positive filter F to the proper ideal {0} ∪ π −1 (F ) (see [24, 25]). This correspondence maps
prime ideals to prime positive filters and non-zero principal ideals to principal positive filters.
In particular, the prime elements of R correspond to the principal prime positive filters of G.
The following result shows (see [7, Theorem 5.3, p. 113]) that any lattice-ordered Abelian
group is the group of divisibility of some Bézout domain, thus allowing one to construct a very
large class of examples of such domains using the theory of lattice-ordered groups:
Theorem 5.3 (Krull-Kaplansky-Jaffard-Ohm) If (G, ≤) is a lattice-ordered Abelian group, then
there exists a Bézout domain R whose group of divisibility is order-isomorphic to (G, ≤).
For any totally ordered group G0 , the result of Krull mentioned above gives a valuation ring
whose divisibility group is order-isomorphic to G0 . This valuation ring can be taken to be the
group ring k[G0 ], where k is a field together with the following valuation on the field of fractions
k(G0 ):
m
n
X
X
v
ai Xgi /
bj Xhj = inf(g1 , . . . , gm ) − inf(h1 , . . . , hn ) ,
i=1
j=1
where it is assumed that all coefficients appearing in the expression are non-zero. Lorenzen
[26] proved that every lattice-ordered group can be embedded into a direct product of totally
ordered groups with the product ordering. This embedding is used by Kaplansky and Jaffard
to construct the valuation domain R of Theorem 5.3. By the result of Lorenzen, there exists
def. Q
a lattice embedding f : G → H =
γ∈Γ Gγ , where Gγ is a totally ordered group for all
γ ∈ Γ and H has the product ordering. Let Q = k({Xg : g ∈ G}) be the group field with
coefficients in a field k with the set of formal variables Xg indexed by elements of G. There
is a valuation ϕ : Q× → H. The integral domain R is the domain defined by this valuation,
def.
i.e. R = {0} ∪ {x ∈ Q× : ϕ(x) ≥ 0}. It is proved by Ohm that the divisibility group of R is
order-isomorphic to G. Combining this with the results of the previous sections, we have:
38
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
Proposition 5.4 Let G be any lattice-ordered group and consider the ring R constructed by
Theorem 5.3. Assume further that W is a critically-finite element of R. Then statements of
Theorems 3.2, 3.3 and 4.12 hold.
The simplest situation is when the lattice-ordered group G is totally ordered, in which case R
is a valuation domain. Then a proper subset F ⊂ G+ is a positive filter iff it is upward-closed,
in which case the complement G \ F is non-empty and downward-closed. If G \ F has a greatest
element m, then G \ F =↓ m and F = (↑ m) \ {m}. If G \ m does not have a greatest element,
then (G \ F, F ) is a Dedekind cut of the totally-ordered set (G, ≤). For example we can take
G ∈ {Z, Q, R} with the natural total order:
• When G = Z, the Bézout domain R is a discrete valuation domain and thus a PID with a
unique non-zero prime ideal and hence with a prime element p ∈ R which is unique up to
association in divisibility. In this case, any positive filter of Z is principal and there is only
one prime filter, namely ↑ 1 = Z+ \ {0}. A critically-finite potential has the form W = W0 pk ,
where k ≥ 2 and W0 is a unit of R.
• When G = Q, there are two types of positive filters. The first have the form F = (↑ q)\{q} =
(q, +∞) ∩ Q with q ∈ Q≥0 , while the second correspond to Dedekind cuts and have the form
F = [a, +∞) ∩ Q with a ∈ R>0 . In particular, a positive filter is principal iff it has the
form F = [q, +∞) ∩ Q with q ∈ Q>0 . A principal positive filter can never be prime, since
Q+ \ F = [0, q) ∩ Q is not closed under addition for q > 0. Hence the Bézout domain R has
no prime elements when G = Q.
• When G = R, any proper subset of R+ has an infimum and hence positive filters have the
form F = (a, +∞) or F = [a, +∞) with a ∈ R>0 , the latter being the principal positive
filters. No principal positive filters can be prime, so the corresponding Bézout domain has no
prime elements.
We can construct more interesting examples as follows. Let (Gi , ≤i )i∈I be any family of latticeordered Abelian groups, where the non-empty index set I is arbitrary. Then the direct product
def. Q
group G =
i∈I Gi is a lattice-ordered Abelian group when endowed with the product order
≤:
(gi )i∈I ≤ (gi0 )i∈I iff ∀i ∈ I : gi ≤i gi0 .
def.
def.
Let supp (g) = {i ∈ I|gi 6= 0}. The direct sum G0 = ⊕i∈I Gi = {g = (gi ) ∈ G||supp (g) < ∞}
is a subgroup of G which becomes a lattice-ordered AbelianQgroup when endowed with the order
induced by ≤. For any x = (xi )i∈I ∈ G, we have ↑G x = i∈I ↑ xi while for any x0 ∈ G0 , we
have ↑G0 x0 = ⊕i∈I ↑ x0i , where ↑G and ↑G0 denote respectively the upper sets computed in G
and G0 . Hence:
Q
I. The principal positive filters of G have the form F = i∈I Fi , where:
1. each Fi is a non-empty subset of Gi+ which either coincides with ↑ 0i or is a principal
positive filter of (Gi , ≤i )
2. at least one of Fi is a principal positive filter of (Gi , ≤i ).
Such a principal positive filter F of G is prime iff the set Gi+ \ Fi is empty or a semigroup
for all i ∈ I. In particular, the principal prime ideals of the Bézout domain R associated
to G by the construction of Theorem 5.3 are in bijection with families (indexed by I) of
principal prime ideals of the Bézout domains Ri associated to Gi by the same construction.
The non-zero principal prime ideals of R are in bijection with families (Ji )i∈S , where S is
a non-empty subset of I and Ji is a non-zero principal prime ideal of Ri for each i ∈ S.
Elementary matrix factorizations over Bézout domains
39
II. The principal positive filters of G0 have the form F 0 = ⊕i∈I Fi , where (Fi )i∈I is a family
def.
of subsets of Fi ⊆ Gi+ such that the set supp F = {i ∈ I|Fi 6=↑Gi 0i } is finite and nonempty and such that Fi is a principal positive filter of (Gi , ≤i ) for any i ∈ supp F . Such a
principal positive filter F 0 of G0 is prime iff Fi is prime in (Gi , ≤i ) for any i ∈ supp F . In
particular, the non-zero principal prime ideals of the Bézout domain R0 associated to G0 by
the construction of Theorem 5.3 are in bijection with finite families of the form Ji1 , . . . , Jin
(n ≥ 1), where i1 , . . . , in are distinct elements of I and Jik is a non-zero principal prime
ideal of the Bézout domain Rik associated to Gik by the same construction.
It is clear that this construction produces a very large class of Bézout domains which have prime
elements and hence to which Proposition 5.4 applies. For example, consider the direct power
G = ZI and the direct sum G = Z(I) , endowed with the product order. Then the Bézout domain
R0 associated to Z(I) is a UFD whose non-zero principal prime ideals are indexed by the finite
non-empty subsets of I. On the other hand, the non-zero principal prime ideals of the Bézout
domain R associated to ZI are indexed by all non-empty subsets of I. Notice that R and R0
coincide when I is a finite set.
5.3. Constructions through the spectral poset. Given a unital commutative ring R, its spectral
poset is the prime spectrum Spec(R) of R viewed as a partially-ordered set with respect to the
order relation ≤ given by inclusion between prime ideals. Given a poset (X, ≤) and two elements
x, y ∈ X, we write x y if x < y and x is an immediate neighbor of y, i.e. if there does not
exist any element z ∈ X such that x < z < y. It was shown in [27] that the spectral poset of any
unital commutative ring satisfies the following two conditions, known as Kaplansky’s conditions:
I. Every non-empty totally-ordered subset of (Spec(R), ≤) has a supremum and an infimum
(in particular, ≤ is a lattice order).
II. Given any elements x, y ∈ Spec(R) such that x < y, there exist distinct elements x1 , y1 of
Spec(R) such that x ≤ x1 < y1 ≤ y and such that x1 y1 .
It is known [28, 29] that these conditions are not sufficient to characterizes spectral posets. It
was shown in [30] that a poset (X, ≤) is order-isomorphic with the spectral poset of a unital
commutative ring iff (X, ≤) is profinite, i.e. iff (X, ≤) is an inverse limit of finite posets; in
particular, any finite poset is order-isomorphic with a spectral poset [29].
A partially ordered set (X, ≤) is called a tree if for every x ∈ X, the lower set ↓ x = {y ∈
X|y ≤ x} is totally ordered. The following result was proved by Lewis:
Theorem 5.5 [29] Let (X, ≤) be a partially-ordered set. Then the following statements are
equivalent:
(a) (X, ≤) is a tree which has a unique minimal element θ ∈ X and satisfies Kaplansky’s
conditions I. and II.
(b) (X, ≤) is isomorphic with the spectral poset of a Bézout domain.
Moreover, R is a valuation domain iff (X, ≤) is a totally-ordered set.
An explicit Bézout domain R whose spectral poset is order-isomorphic with a tree (X, ≤) satisfying condition (a) of Theorem 5.5 is found by first constructing a lattice-ordered Abelian group
G associated to (X, ≤) and then constructing R from G is in Theorem 5.3. The lattice-ordered
group G is given by [29]:
G = {f : X ∗ → Z | |supp (f )| < ∞} ,
40
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
def.
def.
where X ∗ = {x ∈ X | ∃y ∈ X : y x} and supp f = {x ∈ X ∗ | f (x) 6= 0} is a tree when
endowed with the order induced from X. The group operation is given by pointwise addition.
The lattice order on G is defined by the positive cone:
def.
G+ = {f ∈ G | f (x) > 0 ∀x ∈ minsupp (f )} = {f ∈ G | f (x) ≥ 1 ∀x ∈ minsupp (f )} ,
(5.1)
where the order on Z is the natural order and the minimal support of f ∈ G is defined through:
def.
minsupp (f ) = {x ∈ supp (f ) | ∀y ∈ X ∗ such that y < x : f (y) = 0} .
(5.2)
Notice that f ∈ G+ if minsupp (f ) = ∅ (in particular, we have 0 ∈ G+ ). The lattice-ordered
Abelian group G has the property that the set of its prime positive filters6 (ordered by inclusion)
is order-isomorphic with the tree obtained from (X, ≤) by removing the minimal element θ
(which corresponds to the zero ideal of R). Explicitly, the positive prime filter Fx associated to
an element x ∈ X \ {θ} is defined through [29, p. 432]:
def.
Fx = {f ∈ G+ | ∃y ∈ minsupp (f ) : y ≤ x} = {f ∈ G+ | minsupp (f ) ∩ (↓ x) 6= ∅} .
(5.3)
By Lemma 2.23, a principal prime ideal of a Bézout domain is necessarily maximal. This implies
that the prime elements of R (considered up to association in divisibility) correspond to certain
maximal elements of the tree (X, ≤). Notice, however, that a Bézout domain can have maximal
ideals which are not principal (for example, the so-called “free maximal ideals” of the ring of
complex-valued holomorphic functions defined on a non-compact Riemann surface Σ [11]). For
any maximal element x of X which belongs to X ∗ , let 1x ∈ G be the element defined by the
characteristic function of the set {x} in X ∗ :
def.
1x (y) =
1 if y = x
0 if y ∈ X ∗ \ {x}
.
Then supp (1x ) = minsupp (1x ) = {x} and 1x ∈ G+ \ {0}. Notice that 1x ∈ Fx .
Proposition 5.6 Let (X, ≤) be a tree which has a unique minimal element and satisfies Kaplansky’s conditions I. and II. and let R be the Bézout domain determined by (X, ≤) as explained
above.
(a) For each maximal element x of X which belongs to X ∗ , the principal positive filter ↑ 1x is
prime and hence corresponds to a principal prime ideal of R. Moreover, we have:
↑ 1x = {f ∈ G+ | supp (f )∩ ↓ x 6= ∅}
(5.4)
Fx = {f ∈ ↑ 1x | inf Sf (x) ∈ Sf (x)} = {f ∈ ↑ 1x | ∃ min Sf (x)} ,
(5.5)
and:
where:
def.
Sf (x) = supp (f )∩ ↓ x .
(b) Let W be a critically-finite element of R. Then the statements of Theorem 3.2, 3.3 and 4.12
hold.
6
Called “prime V-segments” in [29].
Elementary matrix factorizations over Bézout domains
41
Proof. The second statement follows from the results of the previous sections. To prove the first
statement, let x be a maximal element of X which belongs to X ∗ . We have:
↑ 1x = {f ∈ G+ | f − 1x ∈ G+ } = {f ∈ G+ | f (y) > 1x (y) ∀y ∈ minsupp (f − 1x )} .
(5.6)
On the other hand, we have minsupp (f − 1x ) = {y ∈ X ∗ | f (y) 6= 1x (y) & ∀z ∈ X ∗ such that z <
y : f (z) = 1x (z)}. Since x is maximal, any element z ∈ X ∗ for which there exists y ∈ X ∗ such
that z < y satisfies z 6= x and hence 1x (z) = 0. This gives:
minsupp (f − 1x ) = {y ∈ X ∗ | f (y) 6= 1x (y) & ∀z ∈ X ∗ such that z < y : f (z) = 0} =
minsupp (f ) ∪ {x} if f ∈ Qx
=
,
minsupp (f ) \ {x} if f ∈ G+ \ Qx
where:
def.
Qx = {f ∈ G+ | f (x) 6= 1 & ∀z ∈ X ∗ such that z < x : f (z) = 0} = Ax t Bx ,
with:
def.
Ax = {f ∈ G+ | ∀z ∈ X ∗ such that z ≤ x : f (z) = 0} = {f ∈ G+ | supp (f )∩ ↓ x = ∅}
def.
Bx = {f ∈ G+ | x ∈ minsupp (f ) & f (x) > 1} ⊂ Fx ⊂ G+ \ Ax .
This gives:
↑ 1x = (G+ \ Qx ) ∪ Bx = (G+ \ Ax ) ∪ Bx = G+ \ Ax = {f ∈ G+ | supp (f )∩ ↓ x 6= ∅} , (5.7)
which establishes (5.4). Notice that G+ \ (↑ 1x ) = Ax is a semigroup, so ↑ 1x is a prime principal
positive filter and hence it corresponds to a principal prime ideal of R. Also notice that Fx ⊂ ↑ 1x .
def.
Consider an element f ∈ ↑ 1x . Then the non-empty set Sf (x) = supp (f )∩ ↓ x is totally
ordered (since X is a tree and hence ↓ x is totally ordered). By Kaplansky’s condition I., this
set has an infimum which we denote by xf = inf Sf (x); notice that xf ∈ ↓ x. For any y ∈ X ∗
with y < xf , we have y 6∈ Sf (x) and hence f (y) = 0. Hence if xf belongs to Sf (x) (i.e. if Sf (x)
has a minimum), then xf = min Sf (x) is an element of minsupp (f )∩ ↓ x and in this case we
have f ∈ Fx . Conversely, given any element f ∈ Fx , it is easy to see that the totally-ordered
set minsupp (f )∩ ↓ x must be a singleton, hence minsupp (f )∩ ↓ x = {xf } for a unique element
xf ∈ Sf (x). This element must be a minimum (and hence an infimum) of the totally-ordered
set Sf (x), since xf belongs to minsupp (f ). We conclude that (5.5) holds. t
u
Remark 5.1. Statement (a) of Proposition 5.6 allows us to construct particular critically-finite
elements of R as follows. For each maximal element of X which belongs to X ∗ , let px be prime
element of R which generates the principal prime ideal corresponding to the principal prime
positive filter ↑ 1x (notice that px is determined up to association in divisibility). For any finite
collection x1 , . . . , xN (N ≥ 1) of maximal elements of X which belong to X ∗ Q
and any integers
nj
n1 , . . . , nN such that nj ≥ 2 for each j ∈ {1, . . . , N }, the element W = N
j=1 pxj ∈ R is
critically-finite.
The following statement will be used in the construction of some examples below:
Proposition 5.7 Let (S, ≤) be a well-ordered set. Then (S, ≤) is a tree with a unique minimal
element. Moreover, (S, ≤) satisfies Kaplansky’s conditions I. and II. iff S has a maximum.
42
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
Proof. Since S is well-ordered, it is totally ordered and has a minimum, therefore it is a tree with
a unique minimal element. Given x, y ∈ S such that x < y, we have x ≤ x1 y1 ≤ y, where
def.
def.
x1 = min{x < s ≤ y} and y1 = min{x1 < s ≤ y}. Thus S satisfies Kaplansky’s condition I.
Any non-empty totally-ordered subset A ⊂ S has a minimum since S is well-ordered. Moreover,
A has a supremum (namely min{s ∈ S | ∀x ∈ A : x ≤ s}) iff it has an upper bound. Hence S
satisfies Kaplansky’s condition II. iff it every non-empty subset of S has an upper bound, which
amounts to the condition that S has a greatest element. t
u
Remark 5.2. Every element of S (except a possible greatest element) has an immediate successor
(upper neighbor). In particular, S has a maximal element iff it has a maximum M , which in
turn happens iff the order type α of S is a successor ordinal. In this case, M has a predecessor
iff α is a double successor ordinal, i.e. iff there exists an ordinal β such that α = β + 2.
Example 5.8 Consider the tree T whose underlying set is the set N = Z≥0 of non-negative
integers together with the following partial order: 0 < n for every n ∈ N and there is no further
strict inequality; notice that any maximal vertex n ∈ N∗ = Z>0 has an immediate lower neighbor,
namely 0. This corresponds to a countable corolla, i.e. a tree rooted at 0 and with an edge
connecting the root to n for every n ∈ N∗ (and no other edges). By Proposition 5.6, each
maximal vertex n ∈ N∗ corresponds to a principal prime ideal of the associated Bézout domain.
Example 5.9 We can make the previous example more interesting by replacing the edges of T
with a tree. For each x ∈ N∗ , consider a tree Tx with a unique root (minimal element) rx ∈ Tx
and which satisfies Kaplansky’s conditions I. and II. Consider the tree T obtained by connecting
0 to rx for x ∈ N∗ . Then T has a unique minimal element (namely 0) and satisfies Kaplansky’s
conditions I. and II. By Proposition 5.6, those maximal elements of each of the trees Tx which
have an immediate lower neighbor correspond to prime elements of the associated Bézout domain
R. We obtain many examples of Bézout domains by varying the trees Tx :
1. Assume that for every x ∈ N∗ , the tree Tx is reduced to the single point rx = x. Then we
recover Example 5.8.
2. For any element x ∈ N∗ , consider a finite tree Tx and let Σx be the set of maximal elements
of Tx . Then T ∗ = T \ {0} and any maximal element of T different from 0 has an immediate
lower neighbor. The corresponding Bézout domain R has a principal prime ideal for every
element of the set ∪x∈N∗ Σx .
3. For each x ∈ N∗ , consider a well-ordered set Sx which has a maximum mx and denote
the minimum element of Sx by rx . By Proposition 5.7, we can take Tx = Sx in the general
construction above, thus obtaining a tree T and a corresponding Bézout domain R. Let U ⊂ N∗
be the set of those x ∈ N∗ for which Sx is a double successor ordinal. Then each element of
U corresponds to a principal prime ideal of R.
A. GCD domains
Let R be an integral domain and U (R) its multiplicative group of units. For any finite sequence
of elements f1 , . . . , fn ∈ R, let hf1 , . . . , fn i denote the ideal generated by the set {f1 , . . . , fn }. An
element u ∈ R is a unit iff hui = R. Two elements f, g ∈ R are called associated in divisibility
(we write f ∼ g) if there exists u ∈ U (R) such that g = uf . This is equivalent with the condition
hf i = hgi. The association relation is an equivalence relation on R.
Elementary matrix factorizations over Bézout domains
43
Definition A.1 An integral domain R is called a GCD domain if any two elements f, g admit
a greatest common divisor (gcd).
Let R be a GCD domain. In this case, the gcd of two elements f, g is determined up to association and the corresponding equivalence class is denoted by (f, g). Any two elements f, g of R
also admit a least common multiple (l.c.m.), which is determined up to association and whose
equivalence class is denoted by [f, g]. By induction, any finite collection of elements f1 , . . . , fn
admits a gcd and and lcm, both of which are determined up to association and whose equivalence
classes are denoted by:
(f1 , . . . , fn ) and [f1 , . . . , fn ] .
Remark A.1. Any irreducible element of a GCD domain is prime, hence primes and irreducibles
coincide in a GCD domain. In particular, any element of a GCD domain which can be factored
into primes has unique prime factorization, up to permutation and association of the prime
factors.
B. Bézout domains
Let R be a GCD domain. We say that the Bézout identity holds for two elements f and g of R
if for one (equivalently, for any) gcd d of f and g, there exist a, b ∈ R such that d = af + bg.
This amounts to the condition that the ideal hf, gi is principal, namely we have hf, gi = hdi.
B.1. Definition and basic properties.
Definition B.1 An integral domain R is called a Bézout domain if any (and hence all) of the
following equivalent conditions hold:
• R is a GCD domain and the Bézout identity holds for any two non-zero elements f, g ∈ R.
• The ideal generated by any two elements of R is principal.
• Any finitely-generated ideal of R is principal.
More generally, a Bézout ring is a unital commutative ring R which has the property that
its finitely-generated ideals are principal. Hence a Bézout domain is a Bézout ring which is
an integral domain. The following well-known statement shows that the Bézout property is
preserved under quotienting by principal ideals:
Proposition B.2 Let R be a Bézout ring and I be a finitely-generated (hence principal) ideal
of R. Then R/I is a Bézout ring.
If R is a Bézout domain and f1 , . . . , fn ∈ R, then we have hf1 , . . . , fn i = hdi for any d ∈
(f1 , . . . , fn ) and there exist a1 , . . . , an ∈ R such that d = a1 f1 +. . .+an fn . The elements f1 , . . . , fn
are called coprime if (f1 , . . . , fn ) = (1), which amounts to the condition hf1 , . . . , fn i = R. This
happens iff there exist elements a1 , . . . , an ∈ R such that a1 f1 + . . . + an fn = 1. Notice that
every Bézout domain is integrally closed [31].
Remark B.1. Bézout domains coincide with those Prüfer domains which are GCD domains. Since
any Prüfer domain is coherent, it follows that any Bézout domain is a coherent ring.
The following result characterizes finitely-generated projective modules over Bézout domains:
44
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
Proposition B.3 [7] Every finitely-generated projective module over a Bézout domain is free.
In particular, finitely-generated projective factorizations over a Bézout domain coincide with
finite-rank matrix factorizations.
B.2. Examples of Bézout domains. The following rings are Bézout domains:
• Principal ideal domains (PIDs) coincide with the Noetherian Bézout domains. Other characterizations of PIDs among Bézout domains are given below.
• Any generalized valuation domain is a Bézout domain.
• The ring O(Σ) of holomorphic complex-valued functions defined on any7 smooth connected
non-compact Riemann surface Σ is a non-Noetherian Bézout domain. In particular, the ring
O(C) of entire functions is a non-Noetherian Bézout domain.
• The ring A of all algebraic integers (the integral closure of Z inside C) is a non-Noetherian
Bézout domain which has no prime elements.
B.3. The Noetherian case. The following is well-known:
Proposition B.4 Let R be a Bézout domain. Then the following statements are equivalent:
• R is Noetherian
• R is a principal ideal domain (PID)
• R is a unique factorization domain (UFD)
• R satisfies the ascending chain condition on principal ideals (ACCP)
• R is an atomic domain.
B.4. Characterizations of Bézout domains.
Definition B.5 Let R be a commutative ring. The Bass stable rank bsr(R) of R is the smallest
integer n, such that for any collection {a0 , a1 , . . . , an } of generators of the unit ideal, there exists
a collection {λ1 , . . . , λn } in R such that the collection {ai − λi a0 : 1 ≤ i ≤ n} also generate the
def.
unit ideal. If no such n exists, then bsr(R) = ∞.
Definition B.6 A unital commutative ring R is called a Hermite ring (in the sense of Kaplansky) if every matrix A over R is equivalent with an upper or a lower triangular matrix.
The following result is proved in [32, Theorem 8.1]
Theorem B.7 [32] Let R be a Bézout domain. Then bsr(R) ≤ 2. Moreover, R is a Hermite
ring.
Acknowledgements. This work was supported by the research grant IBS-R003-S1.
7
Notice that Σ need not be algebraic. In particular, Σ can have infinite genus and an infinite number of ends.
Elementary matrix factorizations over Bézout domains
45
References
1. C. I. Lazaroiu, On the boundary coupling of topological Landau-Ginzburg models, JHEP 05 (2005)
037.
2. M. Herbst, C. I. Lazaroiu, Localization and traces in open-closed topological LandauGinzburg models,
JHEP 05 (2005) 0449.
3. M. Babalic, D. Doryn, C. I. Lazaroiu, M. Tavakol, Differential models for B-type open-closed topological Landau-Ginzburg theories, arXiv:1610.09103 [math.DG].
4. M. Babalic, D. Doryn, C. I. Lazaroiu, M. Tavakol, On B-type open-closed Landau-Ginzburg theories
defined on Calabi-Yau Stein manifolds, arXiv:1610.09813 [math.DG].
5. D. Doryn, C. I. Lazaroiu, M. Tavakol, Matrix factorizations over elementary divisor domains, to
appear.
6. M. B. Langfeldt, Triangulated Categories and Matrix Factorizations, M. Sc. thesis, Norwegian University of Science and Technology, 2016.
7. L. Fuchs, L. Salce, Modules over non-Noetherian domains, Mathematical Surveys and Monographs
84, AMS, 2001.
8. O. Helmer, The elementary divisor theorem for certain rings without chain conditions, Bull. Amer.
Math. Soc. 49 (1943), 225–236.
9. D. Lorenzini, Elementary Divisor domains and Bézout domains, J. Algebra 371 (2012) 609–619.
10. M. Henriksen, On the prime ideals of the ring of entire functions, Pacific J. Math. 3 (1953), 711–720.
11. N. Alling, The valuation theory of meromorphic function fields over open Riemann surfaces, Acta
Math. 110 (1963), 79–96.
12. N. Alling, The valuation theory of meromorphic function fields, Proc. Sympos. Pure Math. vol. 11
Amer. Math. Soc., 1968, pp. 8–29.
13. R. Remmert, Classical Topics in Complex Function Theory, Graduate Texts in Mathematics,
Springer, 1998.
14. M. Henriksen, Some remarks on elementary divisor rings, II , Michigan Math. J. 3 (1955/56),
159–163.
15. B. Zabavsky, A. Gatalevych, A commutative Bézout P M ∗ domain is an elementary divisor ring,
Algebra and Discrete Mathematics 19 (2015) 2, 295–301.
16. R. Guralnick, Similarity of holomorphic matrices, Linear Algebra Appl. 99 (1988), 85–96.
17. D. Estes, J. Ohm, Stable range in commutative rings, J. Algebra 7 (1967), 343–362.
18. M. Roitman, The Kaplansky condition and rings of almost stable range 1, Trans. Amer. Math. Soc.
191 (2013), 3013–3019.
19. O. Forster, Lectures on Riemann surfaces, Springer, 1991.
20. J. L. Mott, Groups of divisibility: A unifying concept for integral domains and partially ordered
groups, in Lattice-Ordered Groups: Advances and Techniques, eds. A. M. W. Glass, W. Charles Holland, Kluver, 1989, pp. 80–104.
21. W. Krull, Algemeine Bewertungstheorie, J. Reine Angew. Math. 117 (1931), 160–196.
22. P. Jaffard, Contribution á la théorie des groupes ordonnés, J. Math. Pures Appl. 9 (1953) 32, 203–
280.
23. R. Gilmer, Multiplicative Ideal Theory, M. Dekker, N.Y., 1972.
24. P. Sheldon, Two counterexamples involving complete integral closure in finite dimensional Prüfer
domains, J. Algebra 27 (1973), 462–474.
46
Dmitry Doryn, Calin Iuliu Lazaroiu, Mehdi Tavakol
25. W. Brandal, Constructing Bézout domains, Rocky Mountain J. Math 6 (1976) 3, 383–399.
26. P. Lorenzen, Abstrakte Begründung der multiplicativen Idealtheorie, Math. Z. 45 (1939), 533–553.
27. I. Kaplansky, Commutative Rings, Allyn and Bacon, Boston, 1970.
28. M. Hochster, Prime ideal structure in commutative rings, Trans. Amer. Math. Soc. 142 (1969),
43–60.
29. W. J. Lewis, The spectrum of a ring as a partially-ordered set, J. Algebra 25 (1973), 419–434.
30. T. P. Speed, On the order of prime ideals, Algebra Universalis 2 (1972), 85–87.
31. N. Bourbaki, Commutative algebra, Chap. 1–7, Springer, 1989.
32. R. Mortini, R. Rupp, The Bézout properties for some classical function algebras, Indagationes Mathematicae 24 (2013), 229–253.
| 0math.AC
|
1
Wireless Information and Power Transfer:
Nonlinearity, Waveform Design and Rate-Energy
Tradeoff
arXiv:1607.05602v3 [cs.IT] 20 Sep 2017
Bruno Clerckx
Abstract—The design of Wireless Information and Power
Transfer (WIPT) has so far relied on an oversimplified and
inaccurate linear model of the energy harvester. In this paper, we
depart from this linear model and design WIPT considering the
rectifier nonlinearity. We develop a tractable model of the rectifier
nonlinearity that is flexible enough to cope with general multicarrier modulated input waveforms. Leveraging that model, we
motivate and introduce a novel WIPT architecture relying on
the superposition of multi-carrier unmodulated and modulated
waveforms at the transmitter. The superposed WIPT waveforms
are optimized as a function of the channel state information so
as to characterize the rate-energy region of the whole system.
Analysis and numerical results illustrate the performance of the
derived waveforms and WIPT architecture and highlight that
nonlinearity radically changes the design of WIPT. We make key
and refreshing observations. First, analysis (confirmed by circuit
simulations) shows that modulated and unmodulated waveforms
are not equally suitable for wireless power delivery, namely
modulation being beneficial in single-carrier transmissions but
detrimental in multi-carrier transmissions. Second, a multicarrier unmodulated waveform (superposed to a multi-carrier
modulated waveform) is useful to enlarge the rate-energy region
of WIPT. Third, a combination of power splitting and time
sharing is in general the best strategy. Fourth, a non-zero
mean Gaussian input distribution outperforms the conventional
capacity-achieving zero-mean Gaussian input distribution in
multi-carrier transmissions. Fifth, the rectifier nonlinearity is
beneficial to system performance and is essential to efficient
WIPT design.
Index Terms—Nonlinearity, optimization, waveform, wireless
power, wireless information and power transfer
I. I NTRODUCTION
Wireless Information and Power Transfer/Transmission
(WIPT) is an emerging research area that makes use of
radiowaves for the joint purpose of wireless communications
or Wireless Information Transfer (WIT) and Wireless Power
Transfer (WPT). WIPT has recently attracted significant attention in academia. It was first considered in [2], where the rateenergy tradeoff was characterized for some discrete channels,
and a Gaussian channel with an amplitude constraint on the
input. WIPT was then studied in a frequency-selective AWGN
channel under an average power constraint [3]. Since then,
WIPT has attracted significant interests in the communication literature with among others MIMO broadcasting [4]–
[6], architecture [7], interference channel [8]–[10], broadband
Bruno Clerckx is with the EEE department at Imperial College London,
London SW7 2AZ, UK (email: [email protected]). This work has
been partially supported by the EPSRC of UK, under grant EP/P003885/1.
The material in this paper was presented in part at the ITG WSA 2016 [1].
system [11]–[13], relaying [14]–[16], wireless powered communication [17], [18]. Overviews of potential applications and
promising future research avenues can be found in [19], [20].
Wireless Power Transfer (WPT) is a fundamental building
block of WIPT and the design of an efficient WIPT architecture fundamentally relies on the ability to design efficient
WPT. The major challenge with WPT, and therefore WIPT,
is to find ways to increase the end-to-end power transfer
efficiency, or equivalently the DC power level at the output
of the rectenna for a given transmit power. To that end,
the traditional line of research (and the vast majority of the
research efforts) in the RF literature has been devoted to the
design of efficient rectennas [21], [22] but a new line of
research on communications and signal design for WPT has
emerged recently in the communication literature [23].
A rectenna is made of a nonlinear device followed by a lowpass filter to extract a DC power out of an RF input signal. The
amount of DC power collected is a function of the input power
level and the RF-to-DC conversion efficiency. Interestingly,
the RF-to-DC conversion efficiency is not only a function of
the rectenna design but also of its input waveform (power and
shape) [24]–[30]. This has for consequence that the conversion
efficiency is not a constant but a nonlinear function of the
input waveform (power and shape).
This observation has triggered recent interests on systematic
wireless power waveform design [29]. The objective is to
understand how to make the best use of a given RF spectrum
in order to deliver a maximum amount of DC power at the
output of a rectenna. This problem can be formulated as a
link optimization where transmit waveforms (across space and
frequency) are adaptively designed as a function of the channel
state information (CSI) so as to maximize the DC power
at the output of the rectifier. In [29], the waveform design
problem for WPT has been tackled by introducing a simple and
tractable analytical model of the diode nonlinearity through
the second and higher order terms in the Taylor expansion of
the diode characteristics. Comparisons were also made with a
linear model of the rectifier, that only accounts for the second
order term, which has for consequence that the harvested DC
power is modeled as a conversion efficiency constant (i.e. that
does not reflect the dependence w.r.t. the input waveform)
multiplied by the average power of the input signal. Assuming
perfect Channel State Information at the Transmitter (CSIT)
can be attained, relying on both the linear and nonlinear
models, an optimization problem was formulated to adaptively
change on each transmit antenna a multisine waveform as a
2
function of the CSI so as to maximize the output DC current
at the energy harvester. Important conclusions of [29] are that
1) multisine waveforms designed accounting for nonlinearity
are spectrally more efficient than those designed based on
a linear model of the rectifier, 2) the derived waveforms
optimally exploit the combined effect of a beamforming gain,
the rectifier nonlinearity and the channel frequency diversity
gain, 3) the linear model does not characterize correctly the
rectenna behavior and leads to inefficient multisine waveform
design, 4) rectifier nonlinearity is key to design efficient
wireless powered systems. Following [29], various works have
further investigated WPT signal and system design accounting
for the diode nonlinearity, including among others waveform
design complexity reduction [31]–[34], large-scale system
design with many sinewaves and transmit antennas [32],
[33], multi-user setup [32], [33], imperfect/limited feedback
setup [35], information transmission [36] and prototyping and
experimentation [37]. Another type of nonlinearity leading to
an output DC power saturation due to the rectifier operating in
the diode breakdown region, and its impact on system design,
has also appeared in the literature [30], [38].
Interestingly, the WIPT literature has so far entirely relied
on the linear model of the rectifier, e.g. see [2]–[20]. Given
the inaccuracy and inefficiency of this model and the potential
of a systematic design of wireless power waveform as in
[29], it is expected that accounting for the diode nonlinearity
significantly changes the design of WIPT and is key to efficient
WIPT design, as confirmed by initial results in [1].
In this paper, we depart from this linear model and revisit
the design of WIPT in light of the rectifier nonlinearity. We
address the important problem of waveform and transceiver
design for WIPT and characterize the rate-energy tradeoff,
accounting for the rectifier nonlinearity. In contrast to the
existing WIPT signal design literature, our methodology in
this paper is based on a bottom-up approach where WIPT
signal design relies on a sound science-driven design of the
underlying WPT signals initiated in [29].
First, we extend the analytical model of the rectenna nonlinearity introduced in [29], originally designed for multi-carrier
unmodulated (deterministic multisine) waveform, to multicarrier modulated signals. We investigate how a multi-carrier
modulated waveform (e.g. OFDM) and a multi-carrier unmodulated (deterministic multisine) waveform compare with each
other in terms of harvested energy. Comparison is also made
with the linear model commonly used in the WIPT literature.
Scaling laws of the harvested energy with single-carrier and
multi-carrier modulated and unmodulated waveforms are analytically derived as a function of the number of carriers and
the propagation conditions. Those results extend the scaling
laws of [29], originally derived for unmodulated waveforms,
to modulated waveforms. We show that by relying on the
classical linear model, an unmodulated waveform and a modulated waveform are equally suitable for WPT. This explains
why the entire WIPT literature has used modulated signals.
On the other hand, the nonlinear model clearly highlights that
they are not equally suitable for wireless power delivery, with
modulation being beneficial in single-carrier transmission but
detrimental in multi-carrier transmissions. The behavior is fur-
thermore validated through circuit simulations. This is the first
paper where the performance of unmodulated and modulated
waveforms are derived based on an tractable analytical model
of the rectifier nonlinearity and the observations made from
the analysis are validated through circuit simulations.
Second, we introduce a novel WIPT transceiver architecture
relying on the superposition of multi-carrier unmodulated
and modulated waveforms at the transmitter and a powersplitter receiver equipped with an energy harvester and an
information decoder. The WIPT superposed waveform and the
power splitter are jointly optimized so as to maximize and
characterize the rate-energy region of the whole system. The
design is adaptive to the channel state information and results
from a posynomial maximization problem that originates from
the nonlinearity of the energy harvester. This is the first paper
that studies WIPT and the characterization of the rate-energy
tradeoff considering the diode nonlinearity.
Third, we provide numerical results to illustrate the performance of the derived waveforms and WIPT architecture.
Key observations are made. First, a multi-carrier unmodulated
waveform (superposed to a multi-carrier modulated waveform)
is useful to enlarge the rate-energy region of WIPT if the
number of subbands is sufficiently large (typically larger
than 4). Second, a combination of power splitting and time
sharing is in general the best strategy. Third, a non-zero
mean Gaussian input distribution outperforms the conventional
capacity-achieving zero-mean Gaussian input distribution in
multi-carrier transmissions. Fourth, the rectifier nonlinearity is
beneficial to system performance and is essential to efficient
WIPT design. This is the first paper to make those observations
because they are direct consequences of the nonlinearity.
Organization: Section II introduces and models the WIPT
architecture. Section III optimizes WIPT waveforms and characterizes the rate-energy region. Section IV derives the scaling
laws of modulated and unmodulated waveforms. Section V
evaluates the performance and section VI concludes the work.
Notations: Bold lower case and upper case letters stand for
vectors and matrices respectively whereas a symbol not in bold
font represents a scalar. k.k2F refers to the Frobenius norm a
matrix. A {.} refers to the DC component of a signal. EX {.}
refers to the expectation operator taken over the distribution of
the random variable X (X may be omitted for readability if
T
the context is clear). .∗ refers to the conjugate of a scalar. (.)
H
and (.) represent the transpose and conjugate transpose of a
matrix or vector respectively. The distribution of a circularly
symmetric complex Gaussian (CSCG) random vector with
mean µ and covariance matrix Σ is denoted by CN (µ, Σ)
and ∼ stands for “distributed as”.
II. A N OVEL WIPT T RANSCEIVER A RCHITECTURE
In this section, we introduce a novel WIPT transceiver
architecture and detail the functioning of the various building
blocks. The motivation behind the use of such an architecture
will appear clearer as we progress through the paper.
A. Transmitter and Receiver
We consider a single-user point-to-point MISO WIPT system in a general multipath environment. The transmitter is
3
equipped with M antennas that transmit information and
power simultaneously to a receiver equipped with a single
receive antenna. We consider the general setup of a multicarrier/band transmission (with single-carrier being a special
case) consisting of N orthogonal subbands where the nth
subband has carrier frequency fn and equal bandwidth Bs ,
n = 0, ..., N − 1. The carrier frequencies are evenly spaced
such that fn = f0 + n∆f with ∆f the inter-carrier frequency
spacing (with Bs ≤ ∆f ).
Uniquely, the WIPT signal transmitted on antenna m,
xm (t), consists in the superposition of one multi-carrier unmodulated (deterministic multisine) power waveform xP,m (t)
at frequencies fn , n = 0, ..., N − 1 for WPT and one multicarrier modulated communication waveform xI,m (t) at the
same frequencies for WIT1 , as per Fig 1(a). The modulated
waveform carries N independent information symbols x̃n (t)
on subband n = 0, ..., N − 1. Hence, the transmit WIPT signal
at time t on antenna m = 1, ..., M writes as
xm (t) = xP,m (t) + xI,m (t),
=
N
−1
X
(wP,n,m + wI,n,m x̃n (t)) e
j2πfn t
)
(1)
where we denote the complex-valued baseband signal transmitted by antenna m at subband n for the unmodulated (deterministic multisine) waveform as wP,n,m = sP,n,m ejφP,n,m
and for the modulated waveform as xn,m (t) = wI,n,m x̃n (t) =
s̃I,n,m (t)ej φ̃I,n,m (t) . wP,n,m is constant across time (for a
given channel state) and xP,m (t) is therefore the weighted
summation of N sinewaves inter-separated by ∆f Hz, and
hence occupies zero bandwidth. On the other hand, xn,m (t)
has a signal bandwidth no greater than Bs with symbols x̃n (t)
assumed i.i.d. CSCG2 random variable with zero-mean and
unit variance (power), denoted as x̃n ∼ CN (0, 1). Denoting
the input symbol x̃n = |x̃n | ejφx̃n , we further express the
magnitude and phase of xn,m as follows s̃I,n,m = sI,n,m |x̃n |
with
sI,n,m = |wI,n,m | and φ̃I,n,m = φI,n,m + φx̃n . Hence
E |xn,m |2 = s2I,n,m and xn,m ∼ CN (0, s2I,n,m ).
The transmit WIPT signal propagates through a multipath
channel, characterized by L paths. Let τl and αl be the
delay and amplitude gain of the lth path, respectively. Further,
denote by ζn,m,l the phase shift of the lth path between
transmit antenna m and the receive antenna at subband n.
Denoting vn,m (t) = wP,n,m +wI,n,m x̃n (t), the signal received
at the single-antenna receiver due to transmit antenna m
can be expressed as the sum of two contributions, namely
one originating from WPT yP,m (t) and the other from WIT
1x
I,m (t) can be implemented using e.g. OFDM.
2 following the capacity achieving input distribution
with average power constraint.
≈ℜ
hn,m (wP,n,m +wI,n,m x̃n (t)) e
j2πfn t
n=0
)
(2)
where we have assumed maxl6=l′ |τl − τl′ | << 1/Bs so that
vn,m (t) and x̃n (t) for each subband are narrowband signals,
thus vn,m (t − τl ) = vn,m (t) and x̃n (t − τl ) = x̃n (t), ∀l. The
PL−1
quantity hn,m = An,m ej ψ̄n,m = l=0 αl ej(−2πfn τl +ζn,m,l ) is
the channel frequency response between antenna m and the
receive antenna at frequency fn .
Stacking up all transmit signals across all antennas, we can
write the transmit WPT and WIT signal vectors as
)
(N −1
X
j2πfn t
,
(3)
xP (t) = ℜ
wP,n e
(N −1
X
n=0
wI,n x̃n (t)e
j2πfn t
)
(4)
T
where wP/I,n = wP/I,n,1 . . . wP/I,n,M
. Similarly,
we define the vector channel as hn = hn,1 ... hn,M .
The total received signal comprises the sum of (2) over all
transmit antennas, namely
n=0
n=0
l=0 n=0
(N −1
X
xI (t) = ℜ
sP,n,m cos(2πfn t + φP,n,m )
+ s̃I,n,m (t) cos(2πfn t + φ̃I,n,m (t)),
)
(N −1
X
=ℜ
(wP,n,m + xn,m (t)) ej2πfn t ,
=ℜ
ym (t) = yP,m (t) + yI,m (t),
(L−1 N −1
)
XX
j2πfn (t−τl )+ζn,m,l
=ℜ
,
αl vn,m (t − τl )e
n=0
n=0
(N −1
X
yI,m (t), namely
in a Gaussian channel
y(t) = yP (t) + yI (t),
(N −1
)
X
j2πfn t
=ℜ
hn (wP,n + wI,n x̃n ) e
.
(5)
n=0
The magnitudes and phases of the sinewaves can be collected into N × M matrices SP and ΦP . The (n, m) entry
of SP and ΦP write as sP,n,m and φP,n,m , respectively.
Similarly, we define N × M matrices such that the (n, m)
entry of matrix SI and ΦI write as sI,n,m and φI,n,m ,
respectively. We define the average power of the WPT and
2
2
WIT waveforms as PP = 12 kSP kF and PI = 12 kSI kF . Due
to the superposition of the two waveforms, the total average
transmit power constraint writes as PP + PI ≤ P .
Following Fig 1(b), using a power splitter with a power
splitting ratio ρ and assuming perfect matching
(as in
√
Section
II-C1),
the
input
voltage
signals
ρR
y(t)
and
ant
p
(1 − ρ)Rant y(t) are respectively conveyed to the energy
harvester (EH) and the information decoder (ID).
Remark 1: As it will appear clearer throughout the paper,
the benefit of choosing a deterministic multisine power waveform over other types of power waveform (e.g. modulated,
pseudo-random) is twofold: 1) energy benefit: multisine will be
shown to be superior to a modulated waveform, 2) rate benefit:
multisine is deterministic and therefore does not induce any
rate loss at the communication receiver.
Remark 2: It is worth noting the effect of the deterministic multisine waveform on the input distribution in (1).
Recall that xn,m = s̃I,n,m ej φ̃I,n,m ∼ CN (0, s2I,n,m ). Hence
wP,n,m + xn,m ∼ CN (wP,n,m , s2I,n,m ) and the effective input
distribution on a given frequency and antenna is not zero
4
xP,1(t)
Power WF
+
xI,1(t)
vd
Ant. 1
vout
Rant
Comm. WF
vs
xP,M(t)
+
Ant. M
~
vin Rin
id
vin
~
C
iout
RL
xI,M(t)
non-linear
device
(a) Transmitter
rectifier
Fig. 2.
low-pass
filter and load
Antenna equivalent circuit (left) and a single diode rectifier (right).
power
splitter
RF-BB conversion
ADC
Power WF
cancellation
BB
receiver
(b) Receiver with waveform cancellation
Fig. 1. Transceiver (Tx and Rx) architecture for WIPT with superposed
communication and power waveform (WF).
mean3 . The magnitude |wP,n,m + xn,m | is Ricean distributed
with a K-factor on frequency n and antenna m given by
Kn,m = s2P,n,m /s2I,n,m.
Remark 3: The superposition of information and power signals has appeared in other works, but for completely different
purposes; namely for multiuser WIPT in [6], collaborative
WIPT in interference channel in [43], [44], and for secrecy
reasons in [45], [46]. Since those works relied on the linear
model, the superposition was not motivated by the rectifier
nonlinearity. Moreover, the properties of the power signals are
completely different. While the power signal is a deterministic
multisine waveform leading to non-zero mean Gaussian input
and the twofold benefit (Remark 1) in this work, it is complex
(pseudo-random) Gaussian ∼ CN (0, Σ) in those works.
B. Information Decoder
Since xP,m (t) does not contain any information, it is
deterministic. This has for consequence that the differential
entropy of vn,m and xn,m are identical (because translation
does not change the differential entropy) and the achievable
rate is always equal to
!
N
−1
X
(1 − ρ) |hn wI,n |2
, (6)
I(SI , ΦI , ρ) =
log2 1 +
σn2
n=0
where σn2 is the variance of the AWGN from the antenna and
the RF-to-baseband down-conversion on tone n.
Naturally, I(SI , ΦI , ρ) is larger than the maximum rate
achievable when ρ = 0, i.e. I(S⋆I , Φ⋆I , 0), which is obtained
by performing Maximum Ratio Transmission (MRT) on each
subband and water-filling power allocation across subbands.
The rate (6) is achievable irrespectively of the receiver
architecture, e.g. with and without waveform cancellation. In
the former case, after down-conversion from RF-to-baseband
(BB) and ADC, the contribution of the power waveform is
subtracted from the received signal (as illustrated in Fig 1(b))4.
In the latter case, the “Power WF cancellation” box of Fig 1(b)
is removed and the BB receiver decodes the translated version
of the codewords.
3 If using OFDM, x
I,m (t) and xm (t) are OFDM waveforms with CSCG
inputs and non-zero mean Gaussian inputs, respectively.
4 If using OFDM, conventional OFDM processing (removing the cyclic
prefix and performing FFT) is then conducted in the BB receiver.
C. Energy Harvester
In [29], a tractable model of the rectifier nonlinearity in
the presence of multi-carrier unmodulated (deterministic multisine) excitation was derived and its validity verified through
circuit simulations. In this paper, we reuse the same model and
further expand it to modulated excitation. The randomness due
to information symbols x̃n impacts the amount of harvested
energy and needs to be captured in the model.
1) Antenna and Rectifier: The signal impinging
an on the
2
tenna is y(t) and has an average power Pav = E |y(t)| . A
lossless antenna is modelled as a voltage source vs (t) followed
by a series resistance5 Rant (Fig 2 left). Let Zin = Rin +jXin
denote the input impedance of the rectifier with the matching
network. Assuming perfect matching (Rin = Rant , Xin = 0),
due to the power splitter, a fraction ρ of the available RF power
Pav is transferred to the rectifier and absorbed by Rin , so that
2
the
power to the rectifier is Pin = ρE |y(t)| =
actual input
2
E |vin (t)| /Rin and√vin (t) = vs (t)/2.
√ Hence, vin (t) can be
formed as vin (t) = y(t) ρRin = y(t) ρRant . We also assume
that the antenna noise is too small to be harvested.
Let us now look at Fig 2(right) and consider a rectifier
composed of a single series diode6 followed by a low-pass
filter with load. Denoting the voltage drop across the diode as
vd (t) = vin (t) − vout (t) where vin (t) is the input voltage to
the diode and vout (t) is the output voltage across the load
resistor, a tractable behavioral diode model is obtained by
Taylor series expansion of the diode characteristic equation
vd (t)
id (t) = is e nvt − 1 (with is the reverse bias saturation
current, vt the thermal voltage, n the ideality factor assumed
equal to 1.05) around
a quiescent operating point vda = a,
P
i
′
′
nvt −1
namely id (t) = ∞
k
i=0 i (vd (t) − a) where k0 = is e
a
nv
e t
and ki′ = is i!(nv
i , i = 1, . . . , ∞. Assume a steady-state
t)
response and an ideal low pass filter such that vout (t) is at
constant DC level. Choosing a = E {vd (t)} = −vout , we can
P∞
P∞
i/2
write id (t) = i=0 ki′ vin (t)i = i=0 ki′ ρi/2 Rant y(t)i .
Under the ideal rectifier assumption and a deterministic
incoming waveform y(t), the current delivered to the load
in a steady-state response is constant and given by iout =
A {id (t)}. In order to make the optimization tractable, we
truncate the Taylor expansion to the nth
o order. A nonlinear
model truncates the Taylor expansion to the nth
o order but
retains the fundamental nonlinear behavior of the diode while
a linear model truncates to the second order term.
5 Assumed
6 The
real for simplicity. A more general model can be found in [31].
model holds also for more general rectifiers as shown in [31].
5
2) Linear and Nonlinear Models: After truncation, the
output DC current approximates as
iout = A {id (t)} ≈
no
X
i=0
i/2
ki′ ρi/2 Rant A y(t)i .
(7)
Let us first consider a multi-carrier unmodulated (multisine)
waveform, i.e. y(t) = yP (t). Following [29], we get an
approximation of the DC component of the current at the
output of the rectifier (and the low-pass filter) with a multisine
excitation over a multipath channel as
iout ≈ k0′ +
no
X
i even,i≥2
i/2
ki′ ρi/2 Rant A yP (t)i
(8)
where A yP (t)2 and A yP (t)4 are detailed in (9) and
(11), respectively (at the top of next page). The linear model
is a special case of the nonlinear model and is obtained by
truncating the Taylor expansion to order 2 (no = 2).
Let us then consider the multi-carrier modulated waveform,
i.e. y(t) = yI (t). It can be viewed as a multisine waveform for
a fixed set of input symbols {x̃n }. Hence, we can also write the
DC component of the current at the output of the rectifier (and
the low-pass filter) with a multi-carrier modulated excitation
and fixed set of input symbols
over a multipath channel as
P
i/2
k0′ + ni oeven,i≥2 ki′ ρi/2 Rant A yI(t)i . Similar expressions
as
2
4
(9) and (11) can be written for A yI (t) and A yI (t) for
a fixed set of input symbols {x̃n }. However, contrary to the
multisine waveform, the input symbols {x̃n } of the modulated
waveform change randomly at symbol rate 1/Bs . For a given
channel impulse response, the proposed model for the DC
current with a modulated waveform is obtained as
no
X
i/2
iout ≈ k0′ +
ki′ ρi/2 Rant E{x̃n } A yI (t)i , (12)
i even,i≥2
by taking the expectation
the distribution of the input
over
symbols {x̃n }. For E A yI (t)i
with i even, the DC
component is first extracted
for
a
given
set of amplitudes
{s̃I,n,m } and phases φ̃I,n,m and then expectation is taken
over the randomness of the input symbols x̃n . Due to the i.i.d.
2
CSCG distribution of
the 2input symbols, |x̃n | is exponentially
distributed with E x̃n
= 1 and φx̃n is uniformly distributed. From the moments of an exponential distribution, we
4
also have that E x̃n
= 2. We can then express (13) and
(14) as a function of sI,n,m and ψI,n,m = φI,n,m + ψ̄n,m . Note
4
that this factor of E x̃n
= 2 does not appear in (11) due
to the absence of modulation, which explains why (11) and
(14) enjoy a multiplicative factor of 38 and 86 , respectively. Here
again, the linear model is obtained by truncating to no = 2.
Let us finally consider the superposed waveform, i.e. y(t) =
yP (t)+yI (t). Both yP (t) and yI (t) waveforms now contribute
to the DC component
iout ≈ k0′ +
no
X
i even,i≥2
i/2
ki′ ρi/2 Rant E{x̃n } A y(t)i
.
(16)
Taking for
instance no = 4 and further expanding the term
E{x̃n } A y(t)i
using the fact that
E {A {yP (t)yI (t)}} =
0, E A yP (t)3 yI (t) = 0, E A yP (t)yI (t)3
= 0 and
E A yP (t)2 yI (t)2
can be written as
= A yP (t)2 E A yI (t)2
, iout
2
iout ≈ k0′ + k2′ ρRant A yP (t)2 + k4′ ρ2 Rant
A yP (t)4
2
E A yI (t)4
+ k2′ ρRant E A yI (t)2 + k4′ ρ2 Rant
2
+ 6k4′ ρ2 Rant
A yP (t)2 E A yI (t)2 . (17)
Observation 1: The linear model highlights that there is
no difference in using a multi-carrier unmodulated (multisine)
waveform and a multi-carrier modulated (e.g. OFDM) waveform for WPT, since according to this model the harvested
PN −1
2
energy is a function of n=0 hn wP/I,n , as seen from (9)
and (13). Hence modulated and unmodulated waveforms are
equally suitable. On the other hand, the nonlinear model highlights that there is a clear difference between using a multicarrier unmodulated over a multi-carrier modulated waveform
in WPT. Indeed, from (13) and (15) of the modulated waveform, both the second and fourth order terms exhibit the
same behavior and samePdependencies, namely they are both
N −1
2
exclusively function of n=0 |hn wI,n | . That suggests that
for a multi-carrier modulated waveform with CSCG inputs, the
linear and nonlinear models are equivalent, i.e. there is no need
in modeling the fourth and higher order term. On the other
hand, for the unmodulated waveform, the second and fourth
order terms, namely (9) and (11), exhibit clearly different
behaviors with the second order term being linear and the
fourth order being nonlinear and function of terms expressed
as the product of contributions from different frequencies.
Remark 4: The linear model is motivated by its simplicity rather than its accuracy and is the popular model
used throughout the WIPT literature, e.g. [4]. Indeed, it is
always assumed that
the harvested DC power is modeled
as ηPin (y(t)) = ηE A y(t)2
where η is the RF-toDC conversion efficiency assumed constant. By assuming η
constant, those works effectively only care about maximizing
the input power Pin (y(t)) (function of y(t))to
the rectifier,
i.e. the second order term (or linear term) E A y(t)2
in
the Taylor expansion. Unfortunately this is inaccurate as η is
not a constant and is itself a function of the input waveform
(power and shape) to the rectifier, as recently highlighted in the
communication literature [23], [28]–[30] but well recognized
in the RF literature [21], [22]. This linear model was shown
through circuit simulations in [29] to be inefficient to design
multisine waveform but also inaccurate to predict the behavior
of such waveforms in the practical low-power regime (-30dBm
to 0dBm). On the other hand, the nonlinear model, rather than
explicitly expressing the DC output power as η(y(t))Pin (y(t))
with η(y(t)) a function of the input signal power and shape,
it directly expresses the output DC current as a function of
y(t) (and therefore as a function of the transmit signal and
wireless channel) and leads to a more tractable formulation.
Such a nonlinear model with no = 4 has been validated for
the design of multisine waveform in [28], [29], [31] using
circuit simulators with various rectifier topologies and input
power and in [37] through prototyping and experimentation.
Nevertheless, the use of a linear vs a nonlinear model for the
design of WPT based on other types of waveforms and the
design of WIPT has never been addressed so far.
6
A yP (t)2
A yP (t)4
1
=
2
=
"N−1
X
3
ℜ
8
3
=
8
E A yI (t)2
E A yI (t)4
hn wP,n
n=0
X
2
n0 ,n1 ,n2 ,n3
n0 +n1 =n2 +n3
X
N−1
1 X X
sP,n,m0 sP,n,m1 An,m0 An,m1 cos ψP,n,m0 − ψP,n,m1 ,
=
2 n=0 m ,m
0
1
∗
∗
hn3 wP,n3
,
hn0 wP,n0 hn1 wP,n1 hn2 wP,n2
#
X
m0 ,m1 ,
n0 ,n1 ,n2 ,n3
n0 +n1 =n2 +n3 m2 ,m3
"
3
Y
j=0
#
(9)
(10)
sP,nj ,mj Anj ,mj cos(ψP,n0 ,m0 + ψP,n1 ,m1 − ψP,n2 ,m2 − ψP,n3 ,m3 ) .
(11)
"N−1
#
N−1
1 X
1 X X
2
sI,n,m0 sI,n,m1 An,m0 An,m1 cos ψI,n,m0 − ψI,n,m1 =
=
hn wI,n
(13)
2 n=0 m ,m
2 n=0
0
1
"
"
#"
#
#
X
Y
Y
6 X
=
sI,n0 ,mj An0 ,mj
sI,n1 ,mj An1 ,mj cos(ψI,n0 ,m0 +ψI,n1 ,m1 −ψI,n0 ,m2 −ψI,n1 ,m3 ) (14)
8 n ,n m0 ,m1 , j=0,2
j=1,3
0
1
m2 ,m3
"N−1
#2
6 X
2
hn wI,n
=
8 n=0
Remark 5: The above model deals with the diode nonlinearity under ideal low pass filter and perfect impedance
matching. However there exist other sources of nonlinearities
in a rectifier, e.g. impedance mismatch, breakdown voltage and
harmonics. Recently, another nonlinear model has emerged
in [30]. This model accounts for the fact that for a given
rectifier design, the RF-to-DC conversion efficiency η is a
function of the input power and sharply decreases once the
input power has reached the diode breakdown region. This
leads to a saturation nonlinearity where the output DC power
saturates beyond a certain input power level. There are multiple differences between those two models.
First, our diode nonlinearity model assumes the rectifier is
not operating in the diode breakdown region. Circuit evaluations in [29], [31] and in Section V-B also confirm that
the rectifier never reached the diode breakdown voltage under
all investigated scenarios. We therefore do not model the
saturation effect. On the other hand, [30] assumes the rectifier
can operate in the breakdown region and therefore models the
saturation. However, it is to be reminded that operating diodes
in the breakdown region is not the purpose of a rectifier and
should be avoided. A rectifier is designed in such a way that
current flows in only one direction, not in both directions as it
would occur in the breakdown region. Hence, [30] models a
saturation nonlinearity effect that occurs in an operating region
where one does not wish to operate in. In other words, the
rectifier is pushed in an input power range quite off from the
one it has originally been designed for. This may only occur
in applications where there is little guarantee to operate the
designed rectifier below that breakdown edge.
Second, the diode nonlinearity is a fundamental, unavoidable and intrinsic property of any rectifier, i.e. any rectifier,
irrespectively of its design, topology or implementation, is
always made of a nonlinear device (most commonly Schottky
diode) followed by a low pass filter with load. This has for
consequence that the diode nonlinearity model is general and
(15)
valid for a wide range of rectifier design and topology (with
one and multiple diodes) as shown in [31]. Moreover, since
it is driven by the physics of the rectenna, it analytically
links the output DC metric to the input signal through the
diode I-V characteristics. On the other hand, the saturation
nonlinearity in [30] is circuit-specific and modeled via curve
fitting based on measured data. Hence changing the diode
or the rectifier topology would lead to a different behavior.
More importantly, the saturation effect, and therefore the
corresponding nonlinearity, is actually avoidable by properly
designing the rectifier for the input power range of interest.
A common strategy is to use an adaptive rectifier whose
configuration changes as a function of the input power level,
e.g. using a single-diode rectifier at low input power and
multiple diodes rectifier at higher power, so as to generate
consistent and non-vanishing η over a significantly extended
operating input power range [39], [40].
Third, the diode nonlinearity model accomodates a wide range
of multi-carrier modulated and unmodulated input signals
and is therefore a function of the input signal power, shape
and modulation. The saturation nonlinearity model in [30] is
restricted to a continuous wave input signal and is a function
of its power. Hence it does not reflect the dependence of the
output DC power to modulation and waveform designs.
Fourth, the diode nonlinearity is a beneficial feature that is
to be exploited as part of the waveform design to boost the
output DC power, as shown in [29]. The saturation nonlinearity
is detrimental to performance and should therefore be avoided
by operating in the non-breakdown region and using properly
designed rectifier for the input power range of interest.
Fifth, the diode nonlinearity is more meaningful in the lowpower regime (-30dBm to 0dBm with state-of-the-art rectifiers7 ) while the saturation nonlinearity is relevant in the high
power regime (beyond 0dBm input power).
7 At
lower power levels, the diode may not turn on.
7
III. WIPT WAVEFORM O PTIMIZATION AND
R ATE -E NERGY R EGION C HARACTERIZATION
Leveraging the energy harvester model, we now aim at
characterizing the rate-energy region of the proposed WIPT
architecture. We define the achievable rate-energy region as
n
CR−IDC (P ) , (R, IDC ) : R ≤ I,
o
1
2
2
IDC ≤ iout , kSI kF + kSP kF ≤ P . (18)
2
Assuming the CSI (in the form of frequency response hn,m ) is
known to the transmitter, we aim at finding the optimal values
of amplitudes, phases and power splitting ratio, denoted as
S⋆P ,S⋆I ,Φ⋆P ,Φ⋆I , ρ⋆ , so as to enlarge as much as possible the
rate-energy region. We derive a methodology that is general
to cope with any truncation order no 8 .
Characterizing such a region involves solving the problem
max
SP ,SI ,ΦP ,ΦI ,ρ
iout (SP , SI , ΦP , ΦI , ρ)
(19)
1
2
2
(20)
kSI kF + kSP kF ≤ P.
2
Following [29], (19)-(20) can equivalently be written as
subject to
max
SP ,SI ,ΦP ,ΦI ,ρ
zDC (SP , SI , ΦP , ΦI , ρ)
(21)
1
2
2
kSI kF + kSP kF ≤ P,
(22)
2
Pno
i/2 i/2
Rant E{x̃n } A y(t)i
where we
=
i even, ki ρ
subject to
with zDC
i≥2
is
define ki = i!(nv
i . Assuming is = 5µA, a diode ideality
t)
factor n = 1.05 and vt = 25.86mV , we get k2 = 0.0034 and
k4 = 0.3829. For no = 4, similarly to (17), we can compute
zDC as in (23).
This enables to re-define the achievable rate-energy region
in terms of zDC rather than iout as follows
n
CR−IDC (P ) , (R, IDC ) : R ≤ I,
o
1
IDC ≤ zDC , kSI k2F + kSP k2F ≤ P , (24)
2
This definition of rate-energy region will be used in the sequel.
A. WPT-only: Energy Maximization
In this section, we first look at energy maximization-only
(with no consideration for rate) and therefore assume ρ = 1.
We study and compare the design of multi-carrier unmodulated
(multisine) waveform (y(t) = yP (t)) and modulated waveforms (y(t) = yI (t)) under the linear and nonlinear models.
Since a single waveform is transmitted (either unmodulated or
modulated), the problem simply boils down to the following
for i ∈ {P, I}
1
2
kSi kF ≤ P,
(25)
2
Pno
i/2
i
where zDC (SP ,ΦP ) =
for
i even,i≥2 ki Rant A yP (t)
multi-carrier unmodulated
(multisine)
waveform,
and
P
i/2
zDC (SI ,ΦI ) = ni oeven,i≥2 ki Rant E{x̃n } A yI (t)i
for the
multi-carrier modulated waveform.
max zDC (Si , Φi ) subject to
Si ,Φi
8 We
display terms for no ≤ 4 but the derived algorithm works for any no .
The problem of multisine waveform design with a linear and nonlinear rectenna model has been addressed in
[29]. The Plinear model leads to the P
equivalent problem
N −1
2
N −1
2
1
maxwP,n
n=0 kwP,n k ≤ P
n=0 |hn wP,n | subject to 2
whose solution is the adaptive single-sinewave (ASS) strategy
√
2P hH
⋆
n / khn k , n = n̄,
wP,n
=
(26)
0, n 6= n̄.
The ASS performs a matched (also called MRT) beamformer
on a single sinewave, namely the one corresponding to the
2
strongest channel n̄ = arg maxn khn k . On the other hand,
the nonlinear model leads to a posynomial maximization
problem that can be formulated as a Reversed Geometric
Program and solved iteratively. Interestingly, for multisine
waveforms, the linear and nonlinear models lead to radically
different strategies. The former favours transmission on a
single frequency while the latter favours transmission over
multiple frequencies. Design based on the linear model was
shown to be inefficient and lead to significant loss over the
nonlinear model-based design.
The design of multi-carrier modulated waveform is rather
different. Recall that from (13) and (15), both the second
and fourth order terms are exclusively function of
PN −1
2
n=0 |hn wI,n | . This shows that both the linear and nonlinear model-based designs of multi-carrier modulated waveforms for WPT lead to the ASS strategy and the optimum
⋆
wI,n
should be designed according to (26). This is in sharp
contrast with the multisine waveform design and originates
from the fact that the modulated waveform is subject to CSCG
randomness due to the presence of input symbols x̃n . Note that
this ASS strategy has already appeared in the WIPT literature,
e.g. in [11], [41] with OFDM transmission.
B. WIPT: A General Approach
We now aim at characterizing the rate-energy region of the
proposed WIPT architecture. Looking at (6) and (23), it is
easy to conclude that matched filtering w.r.t. the phases of the
channel is optimal from both rate and harvested energy maximization perspective. This leads to the same phase decisions
as for WPT in [28], [29], namely
φ⋆P,n,m = φ⋆I,n,m = −ψ̄n,m
(27)
and
all arguments of the
guarantees
cosine
functions in
A yP (t)i i=2,4 ((9), (11)) and in E A yI (t)i
i=2,4
((13), (14)) to be equal to 0. Φ⋆P and Φ⋆I are obtained by
collecting φ⋆P,n,m and φ⋆I,n,m ∀n, m into a matrix, respectively.
With such phases Φ⋆P and Φ⋆I , zDC (SP , SI , Φ⋆P , Φ⋆I , ρ) can
be finally written as (28). Similarly we can write
!
N
−1
Y
(1 − ρ)
⋆
(29)
Cn
1+
I(SI , ΦI , ρ) = log2
σn2
n=0
P
Q1
where Cn = m0 ,m1 j=0 sI,n,mj An,mj .
Recall from [47] that a monomial is defined as the function
a1 a2
aN
g : RN
++ → R : g(x) = cx1 x2 . . . xN where c > 0 and
ai ∈ R. A sum of K monomials
PKis called a posynomial
and can be written as f (x) =
k=1 gk (x) with gk (x) =
ck xa1 1k xa2 2k . . . xaNN k where ck > 0. As we can see from (28),
zDC (SP , SI , Φ⋆P , Φ⋆I , ρ) is a posynomial.
8
zDC (SP , SI , ΦP , ΦI , ρ) = k2 ρRant A yP (t)2 + k4 ρ2 R2ant A yP (t)4 + k2 ρRant E A yI (t)2
+ k4 ρ2 R2ant E A yI (t)4
+ 6k4 ρ2 R2ant A yP (t)2 E A yI (t)2
zDC (SP , SI , Φ⋆P , Φ⋆I , ρ) =
N−1
X
k2 ρ
Rant
2
n=0 m
X
0 ,m1
"
#
2
3k4 ρ 2
Rant
sP,n,mj An,mj +
8
j=0
1
Y
X
X
m0 ,m1 ,
n0 ,n1 ,n2 ,n3
n0 +n1 =n2 +n3 m2 ,m3
"
3
Y
j=0
0
1
0
In order to identify the achievable rate-energy region, we
formulate the optimization problem as an energy maximization
problem subject to transmit power and rate constraints
max
subject to
zDC (SP , SI , Φ⋆P , Φ⋆I , ρ)
1
2
2
kSI kF + kSP kF ≤ P,
2
I(SI , Φ⋆I , ρ) ≥ R̄.
min
SP ,SI ,ρ,ρ̄,t0
(30)
(31)
(32)
min
subject to
1/t0
1
2
2
kSI kF + kSP kF ≤ P,
2
t0 /zDC (SP , SI , Φ⋆P , Φ⋆I , ρ) ≤ 1,
"N −1
#
Y
(1 − ρ)
R̄
1+
2 /
≤ 1.
Cn
σn2
n=0
1/t0
(37)
1
2
2
kSI kF + kSP kF ≤ P,
(38)
2
K
−γk
Y
gk (SP , SI , Φ⋆P , Φ⋆I , ρ)
≤ 1, (39)
t0
γk
k=1
−γnk
Kn
N
−1 Y
Y
gnk (SI , ρ̄)
R̄
2
≤ 1,
(40)
γnk
n=0
k=1
It therefore consists in maximizing a posynomial subject
to constraints. Unfortunately this problem is not a standard
Geometric Program (GP) but it can be transformed to an
equivalent problem by introducing an auxiliary variable t0
SP ,SI ,ρ,t0
(28)
1
and write the standard GP
subject to
SP ,SI ,ρ
(23)
#
sP,nj ,mj Anj ,mj
" 1
" 1
#2
#
N−1
N−1
2
X X
Y
X X
Y
k2 ρ
3k
ρ
4
2
+
Rant
Rant
sI,n,mj An,mj
sI,n,mj An,mj +
2
4
n=0 m0 ,m1 j=0
n=0 m0 ,m1 j=0
" 1
# N−1
" 1
#
N−1
Y
X X
Y
3k4 ρ2 2 X X
Rant
sP,n,mj An,mj
sI,n,mj An,mj
+
2
n=0 m ,m
n=0 m ,m
j=0
j=0
.
(33)
ρ + ρ̄ ≤ 1,
(41)
that can be solved using existing software, e.g. CVX [48].
It is important to note that the choice of {γk , γnk } plays a
great role in the tightness of the AM-GM inequality. An iterative procedure can be used where at each iteration the standard
GP (37)-(41) is solved for an updated set of {γk , γnk }. Assum(i−1)
(i−1)
ing a feasible set of magnitude SP
and SI
and power
splitting ratio ρ(i−1) at iteration i − 1, compute at iteration i
(i−1)
(34)
(35)
(36)
This is known as a Reversed Geometric Program [47]. A
similar problem also appeared in the WPT waveform op⋆
⋆
timization
Note that
1/zDC (SP , SI , ΦP , ΦI , ρ) and
QN −1 [28]. (1−ρ)
are not posynomials, therefore
Cn
1/
2
n=0 1 +
σn
preventing the use of standard GP tools. The idea is to replace
the last two inequalities (in a conservative way) by making use
of the arithmetic mean-geometric mean (AM-GM) inequality.
Let
{gk (SP , SI , Φ⋆P , Φ⋆I , ρ)}
be
the
monomial
terms in the posynomial zDC (SP , SI , Φ⋆P , Φ⋆I , ρ) =
PK
⋆
⋆
Similarly
we
define
k=1 gk (SP , SI , ΦP , ΦI , ρ).
{gnk (SI , ρ̄)} asPthe set of monomials of the posynomial
Kn
1 + σρ̄2 Cn =
k=1 gnk (SI , ρ̄) with ρ̄ = 1 − ρ. For a
n
given
choice
of
{γ
PK
PKn k } and {γnk } with γk , γnk ≥ 0 and
γ
=
k
k=1
k=1 γnk = 1, we perform single condensations
(i−1)
⋆ (i−1)
,SI
,Φ⋆
)
P ,ΦI ,ρ
(i−1)
(i−1)
⋆
(i−1) )
zDC (SP
,SI
,ΦP ,Φ⋆
I ,ρ
(i−1)
(i−1)
gk (SP
k = 1, . . . , K and γnk =
(i−1)
) , n = 0, . . . , N −1,
gnk (SI
, ρ̄(i−1) )/ 1+ ρ̄ σ2 Cn (SI
n
k = 1, . . . , Kn and then solve problem (37)-(41) to obtain
(i)
(i)
SP , SI and ρ(i) . Repeat the iterations till convergence. The
whole optimization procedure is summarized in Algorithm 1.
γk =
Algorithm 1 WIPT Waveform and R-E Region
⋆
1: Initialize: i ← 0, R̄, Φ⋆
P and ΦI , SP , SI , ρ, ρ̄ = 1 − ρ,
(0)
zDC = 0
2: repeat
3:
i ← i + 1, S̈P ← SP , S̈I ← SI , ρ̈ ← ρ, ρ̄¨ ← ρ̄
4:
γk ← gk (S̈P , S̈I , Φ⋆P , Φ⋆I , ρ̈)/zDC (S̈P , S̈I , Φ⋆P , Φ⋆I , ρ̈),
k = 1, . . . , K
¨
5:
γnk ← gnk (S̈I , ρ̄¨)/ 1+ σρ̄2 Cn (S̈I ) , n = 0, . . . , N −1,
n
k = 1, . . . , Kn
6:
SP , SI , ρ, ρ̄ ← arg min (37) − (41)
(i)
7:
zDC ← zDC (SP , SI , Φ⋆P , Φ⋆I , ρ)
(i)
(i−1)
8: until zDC − zDC
< ǫ or i = imax
Note that the successive approximation method used in the
9
Algorithm 1 is also known as a sequential convex optimization
or inner approximation method [49]. It cannot guarantee to
converge to the global solution of the original problem, but
only to yield a point fulfilling the KKT conditions [49], [50].
C. WIPT: Decoupling Space and Frequency
When M > 1, previous section derives a general methodology to optimize the superposed waveform weights jointly
across space and frequency. It is worth wondering whether we
can decouple the design of the spatial and frequency domain
weights without impacting performance. The optimal phases
in (27) are those of a MRT beamformer. Looking at (6), (9),
(10), (13) and (15), the optimum weight vectors wP,n and
wI,n that maximize the 2nd and 4th order terms and the rate,
respectively, are MRT beamformers of the form
wP,n = sP,n hH
n / khn k ,
wI,n = sI,n hH
n / khn k ,
(42)
that,
from
(5), P yP (t)
=
N −1
jwn t
kh
k
s
cos
(w
t)
=
ℜ
kh
k
s
e
n
P,n
n
n
P,n
n=0
n=0
PN −1
kh
k
s
x̃n cos (wn t)
=
and
y
(t)
=
n
I,n
I
n=0
PN −1
jwn t
ℜ
. Hence, with (42), the
n=0 khn k sI,n x̃n e
multi-antenna multisine WIPT weight optimization is
converted into an effective single antenna multi-carrier WIPT
optimization with the effective channel gain on frequency n
given by khn k and the power allocated to the nth subband
given by s2P,n and s2I,n for the multi-carrier unmodulated
(multisine)
waveform, respectively (subject
PN −1 2and modulated
2
to
n=0 sP,n + sI,n = 2P ). The optimum magnitude
sP,n and sI,n in (42) can now be obtained by using the
posynomial maximization methodology of Section III-B.
Namely, focusing on no = 4 for simplicity, plugging (42)
into (9), (10), (13) and (15), we get (43) and (44). The
DC component zDC as defined in (23) simply writes as
zDC (sP , sI , ρ), expressing that it is
now only a function of
the N -dimensional vectors sP/I = sP/I,0 , . . . , sP/I,N −1 .
QN −1
2
(1−ρ) 2
Similarly, I(sI , ρ) = log2
sI,n khn k
.
2
n=0 1 + σn
Following the posynomial maximization
methodology,
we
PK
can write zDC (sP , sI , ρ) =
k=1 gk (sP , sI , ρ) and 1 +
P
Kn
2
ρ̄
2
g
(s
,
ρ̄)
with
C
C
=
nk
I
n = sI,n khn k , apply the
2
n
k=1
σn
AM-GM inequality and write the standard GP problem
such
P
N −1
1/t0
sP ,sI ,ρ,ρ̄,t0
i
1h
subject to
ksP k2 + ksI k2 ≤ P,
2
−γk
K
Y
gk (sP , sI , ρ)
≤ 1,
t0
γk
k=1
−γnk
Kn
N
−1 Y
Y
gnk (sI , ρ̄)
2R̄
≤ 1,
γnk
n=0
min
(45)
D. WIPT: Characterizing the Twofold Benefit of using Deterministic Power Waveforms
Superposing a power waveform to a communication waveform may boost the energy performance but may lead to the
drawback that the power waveform interferes with the communication waveform at the information receiver. Uniquely, with
the proposed architecture, this issue does not occur since the
power waveform is deterministic, which leads to this twofold
energy and rate benefit highlighted in Remark 1. Nevertheless,
in order to get some insights into the contributions of the deterministic power waveform to the twofold benefit, we compare
in Section V the R-E region of Section III-B to that of two
baselines. The first baseline is the simplified architecture where
the deterministic multisine waveform is absent and for which
the R-E region can still be computed using Algorithm 1 by
forcing SP = 0. In this setup, the twofold benefit disappears.
The second baseline is the hypothetical system where the
power waveform is a deterministic multisine from an energy
perspective but CSCG distributed from a rate perspective.
The energy benefit is retained but the rate benefit disappears
since
the power waveform now creates an interference term
√
1 − ρhn wP,n (with power level given by (1−ρ) |hn wP,n |2 )
at the information receiver and therefore a rate loss. The
achievable rate of that system is given by
ILB (SP , SI , ΦP , ΦI , ρ)
=
(46)
(47)
(48)
k=1
ρ + ρ̄ ≤ 1.
Algorithm 2 WIPT Waveform with Decoupling
1: Initialize: i ← 0, R̄, wP/I,n and in (42), sP , sI , ρ, ρ̄ =
(0)
1 − ρ, zDC = 0
2: repeat
3:
i ← i + 1, s̈P ← sP , s̈I ← sI , ρ̈ ← ρ, ρ̄¨ ← ρ̄
4:
γk ← gk (s̈P , s̈I , ρ̈)/zDC (s̈P , s̈I , ρ̈),
k = 1, . . . , K
¨
5:
γnk ← gnk (s̈I , ρ̄¨)/ 1 + σρ̄2 Cn (s̈I ) , n = 0, . . . , N − 1,
n
k = 1, . . . , Kn
6:
sP , sI , ρ, ρ̄ ← arg min (45) − (49)
(i)
7:
zDC ← zDC (sP , sI , ρ)
(i)
(i−1)
8: until zDC − zDC
< ǫ or i = imax
(49)
Algorithm 2 summarizes the design methodology with
spatial and frequency domain decoupling. Such an approach
would lead to the same performance as the joint spacefrequency design of Algorithm 1 but would significantly reduce the computational complexity since only N -dimensional
vectors sP and sI are to be optimized numerically, compared
to the N × M matrices SP and SI of Algorithm 2.
N
−1
X
n=0
log2
1+
(1 − ρ) |hn wI,n |2
σn2 + (1 − ρ) |hn wP,n |
2
!
(50)
and the corresponding R-E region is a lower bound on that
achieved in Section III-B. Comparing those two R-E regions,
we get a sense of the rate benefit of using a deterministic
power waveform over a modulated one.
The lower bound on the R-E region is obtained by replacing
I by ILB in (24). Because of the interference term in the
SINR expression, decoupling the space frequency design by
choosing the weights vectors as in (42) is not guaranteed to
be optimal. To characterize the R-E region of the hypothetical
system, we therefore resort to the more general approach
where space and frequency domain weights are jointly designed, similarly to the one used in Section III-B.
Let us assume the same phases Φ⋆P and Φ⋆I as in (27).
Such a choice is optimal for M = 1 (even though there is no
10
A yP (t)2
E A yI (t)2
1
=
2
"N−1
X
khn k2 s2P,n
1
2
"N−1
X
#
khn k2 s2I,n , E A yI (t)4
=
n=0
n=0
#
, A yP (t)4
claim of optimality for M > 1). With such a choice of phases,
zDC (SP , SI , Φ⋆P , Φ⋆I , ρ) also writes as (28), while the lower
bound on the achievable rate is written as
!
N
−1
Y
(1−ρ)Cn
⋆
1+ 2
ILB (SI , ΦI , ρ) = log2
(51)
σn +(1−ρ)Dn
n=0
where Cn =
P
Q1
P
m0 ,m1
Q1
j=0 sI,n,mj An,mj
and Dn
=
m0 ,m1
j=0 sP,n,mj An,mj .
The optimization problem can now be written as (33)-(36)
with (36) replaced by
QN −1
ρ̄
D
1
+
2
n
n=0
σn
≤ 1.
2R̄ Q
(52)
N −1
ρ̄
1
+
(D
+
C
)
n
n
n=0
σ2
3
=
8
=
X
n0 ,n1 ,n2 ,n3
n0 +n1 =n2 +n3
6
8
"N−1
X
n=0
"
3
Y
sP,nj hnj
j=0
khn k2 s2I,n
#2
.
#
,
(43)
(44)
Algorithm 3 Lower-Bound on R-E Region
⋆
1: Initialize: i ← 0, R̄, Φ⋆
P and ΦI , SP , SI , ρ, ρ̄ = 1 − ρ,
(0)
zDC = 0
2: repeat
3:
i ← i + 1, S̈P ← SP , S̈I ← SI , ρ̈ ← ρ, ρ̄¨ ← ρ̄
4:
γk ← gk (S̈P , S̈I , Φ⋆P , Φ⋆I , ρ̈)/zDC (S̈P , S̈I , Φ⋆P , Φ⋆I , ρ̈),
k = 1, . . . , K
¨
5:
γnj ← fnj (S̈P , S̈I , ρ̈)/ 1 + σρ̄2 (Dn (S̈P ) + Cn (S̈I )) ,
n
n = 0, . . . , N − 1, j = 1, . . . , Jn
6:
SP , SI , ρ, ρ̄ ← arg min (53) − (57)
(i)
7:
zDC ← zDC (SP , SI , Φ⋆P , Φ⋆I , ρ)
(i)
(i−1)
8: until zDC − zDC
< ǫ or i = imax
n
Define the set of monomials {fnj (SP , SI , ρ)} of the posynP Jn
fnj (SP , SI , ρ). For a given
omial 1 + σρ̄2 (Dn +Cn ) = j=1
n
P Jn
choice of {γnj } with γnj ≥ 0 and j=1
γnj = 1, we perform
single condensations and write the standard GP as
min
SP ,SI ,ρ,ρ̄,t0
subject to
1/t0
(53)
1
2
2
kSI kF + kSP kF ≤ P,
(54)
2
K
Y gk (SP , SI , Φ⋆ , Φ⋆ , ρ) −γk
P
I
≤ 1, (55)
t0
γk
k=1
N
−1
Y
ρ̄
1 + 2 Dn (SP )
2R̄
σn
n=0
−γnj
J
n
Y
fnj (SP , SI , ρ)
≤ 1, (56)
γnj
j=1
ρ + ρ̄ ≤ 1.
(57)
The whole optimization procedure is summarized in Algorithm
3. Algorithm 3 boils down to Algorithm 1 for Dn = 0 ∀n.
IV. S CALING L AWS
In order to get some insights and assess the performance
gain/loss of an unmodulated waveform over a modulated
waveform for WPT, we quantify how zDC scales as a function
of N for M = 1. For simplicity we truncate the Taylor expansion to the fourth order (no = 4). We consider
frequency-flat (FF) and frequency-selective (FS) channels. We
assume that the complex channel gains αl ejξl are modeled
as independent CSCG random variables. αl are therefore
independent Rayleigh distributed
such that α2l ∼ EXPO(λl )
2
with 1/λl = βl = E αl . The impulse responses have a
constant
PL−1 average received power normalized to 1 such that
l=0 βl = 1. In the frequency flat channel, ψ̄n = ψ̄ and
An = A ∀n. This is met when (N −1)∆f is much smaller than
the channel coherence bandwidth. In the frequency selective
channel, we assume that L >> 1 and frequencies fn are far
apart from each other such that the frequency domain CSCG
random channel gains hn,m fade independently (phase and
amplitude-wise) across frequencies and antennas. Taking the
expectation over the distribution of the channel, we denote
z̄DC = E {zDC }. Leveraging the derivation of the scaling laws
for multisine waveforms in [29], we can compute the scaling
laws of multi-carrier modulated waveforms designed using the
ASS strategy (relying on CSIT) and the uniform power (UP)
allocation√strategy
√ (not relying on CSIT). UP is characterized
by Si = 2P / N 1N and Φi = 0N , i ∈ {P, I} [29].
Table I summarizes the scaling laws for both modulated
and unmodulated waveforms for N = 1 (single-carrier SC)
and for N > 1 (multi-carrier) with ASS and UP strategies.
It also compares with the UPMF strategy for multi-carrier
unmodulated (multisine) waveforms (introduced in [29]) that
consists in uniformly allocating power to all sinewaves and
matching the waveform phase to the channel phase (hence
it requires CSIT). Such a UPMF strategy is suboptimal for
multisine excitation with the nonlinear model-based design
and its scaling law is therefore a lower bound on what can
be achieved with the optimal multisine strategy of (25) [29].
Let us first discuss multi-carrier transmission (N >> 1)
with CSIT. Recall that, in the presence of CSIT, the ASS
strategy for multi-carrier modulated waveforms is optimal
for the maximization of zDC with the linear and nonlinear
model-based designs. The ASS for unmodulated (multisine)
waveform is only optimal for the linear model-based design.
As it can be seen from the scaling laws9 , unmodulated
(multisine) waveform with UPMF strategy leads to a linear
9 Taking N to infinity does not imply that the harvested energy reaches
infinity as explained in Remark 4 of [29].
11
TABLE I
S CALING L AWS OF M ULTI -C ARRIER U NMODULATED VS M ODULATED WAVEFORMS .
Strategy
N, M
Frequency-Flat (FF)
z̄DC,SC
z̄DC,U P
z̄DC,SC
z̄DC,U P
N = 1, M = 1
N > 1, M = 1
N = 1, M = 1
N >> 1, M = 1
k2 Rant P
k2 Rant P
k2 Rant P
k2 Rant P
z̄DC,ASS
z̄DC,ASS
z̄DC,U P M F
N >> 1, M = 1
N >> 1, M = 1
N >> 1, M = 1
k2 Rant P + 6k4 R2ant P 2
k2 Rant P + 3k4 R2ant P 2
k2 Rant P + 2k4 R2ant P 2 N
Frequency-Selective (FS)
+ 6k4 R2ant P 2
+ 6k4 R2ant P 2
+ 3k4 R2ant P 2
+ 2k4 R2ant P 2 N
increase of z̄DC with N in FF and FS channels while the
ASS strategy for both modulated and unmodulated only lead
to at most a logarithmic increase with N (achieved in FS
channels). Hence, despite being suboptimal for the nonlinear
model-based design, an unmodulated (multisine) waveform
with UPMF provides a better scaling law than those achieved
by modulated/unmodulated waveforms with ASS. In other
words, even a suboptimal multi-carrier unmodulated waveform
outperforms the optimal design of the multi-carrier modulated
(with CSCG inputs) waveform.
Let us now look at multi-carrier transmission in the absence
of CSIT. A modulated waveform does not enable a linear
increase of z̄DC with N in FF channels, contrary to the unmodulated waveform. This is due to the CSCG distributions of
the information symbols that create random fluctuations of the
amplitudes and phases across frequencies. This contrasts with
the periodic behavior of the unmodulated multisine waveform
that has the ability to excite the rectifier in a periodic manner
by sending higher energy pulses every 1/∆f .
Finally, for the single-carrier transmission in the absence
of CSIT case (N = 1), an opposite behavior is observed
with z̄DC of the modulated waveform outperforming that of
the unmodulated waveform thanks to the fourth order term
being twice as large. This gain originates from
the presence
of the fourth order moment of x̃n , namely E |x̃n |4 = 2, in
the fourth order term of (12). Hence modulation through the
CSCG distribution of the input symbols is actually beneficial
to WPT in single-carrier transmissions. Actually, this also
suggests that a modulation with an input distribution leading
to a large fourth order moment is beneficial to WPT 10 .
The scaling laws also highlight that the effect of channel
frequency selectivity on the performance depends on the type
of waveform. Indeed, without CSIT, frequency selectivity is
detrimental to unmodulated waveform performance (as already
confirmed in [29]) but has no impact on modulated waveform
performance. On the other hand, with CSIT, frequency selectivity leads to a frequency diversity gain that is helpful to the
performance of both types of waveforms11.
Observation 2: The scaling laws highlight that, due to the
diode nonlinearity of the fourth order term, there is a clear
difference between using an unmodulated (multisine) waveform and a modulated waveform in WPT. For single-carrier
10 Asymmetric Gaussian signaling (asymmetric power allocations to the real
and imaginary dimensions) may be a better option than CSCG for WPT [51].
11 This behavior was already confirmed in [29] for multisine waveforms.
k2 Rant P + 6k4 R2ant P 2
k2 Rant P + 3k4 R2ant P 2
k2 Rant P log N + 3k4 R2ant P 2 log2 N
k2 Rant P log N + 32 k4 R2ant P 2 log2 N
≥
k2 Rant P
+
π 2 /16k4 R2ant P 2 N
≤ k2 Rant P + 2k4 R2ant P 2 N
3
2.5 MHz
1 MHz
2.8
2.6
Frequency response
Waveform
No CSIT
Modulated
Modulated
Unmodulated
Unmodulated
CSIT
Modulated
Unmodulated
Unmodulated
2.4
2.2
2
1.8
1.6
1.4
1.2
1
−1.25
Fig. 3.
−1
−0.75
−0.5
−0.25
0
0.25
Frequency [MHz]
0.5
0.75
1
1.25
Frequency response of the wireless channel.
transmissions (N = 1), the scaling law of the modulated
waveform outperforms that of the unmodulated waveform due
to the beneficial effect of the fourth order moment of the
CSCG distribution to boost the fourth order order term. On
the other hand, for multi-carrier transmissions (N >> 1), the
scaling law of multisine significantly outperforms that of the
modulated waveform. While z̄DC scales linearly with N with
a multisine waveform thanks to all carriers being in-phase,
it scales at most logarithmically with N with the modulated
waveform due to the independent CSCG randomness of the
information symbols across subbands. This also shows that
the diode nonlinearity can be beneficial to WPT performance
in two different ways: first, by involving higher order moments
of the input distribution, and second by enabling constructive
contributions from various frequencies.
Observation 2 further motivates the WIPT architecture of
Section II that is based on a superposition of multi-carrier
unmodulated (deterministic multisine) waveform for efficient
WPT and multi-carrier modulated waveform for efficient WIT.
V. P ERFORMANCE E VALUATIONS
We consider two types of performance evaluations, the first
one is based on the simplified nonlinear model of Section
II-C with no = 4, while the second one relies on an actual
and accurate design of the rectenna in PSpice.
A. Nonlinear Model-Based Performance Evaluations
We now illustrate the performance of the optimized WIPT
architecture. Parameters are taken as k2 = 0.0034, k4 =
0.3829 and Rant = 50Ω. We assume a WiFi-like environment
at a center frequency of 5.18GHz with a 36dBm EIRP, 2dBi
receive antenna gain and 58dB path loss. This leads to an
12
30
14
N=2
N=4
N=8
N=16
WPT, =1
25
10
Time sharing
I DC [ A]
zDC [ A]
20
N=1
N=2
N=4
N=8
N=16
12
15
A, 0< <1
8
6
10
4
Time sharing
5
2
WIT with Water-Filling, =0
0
0
1
2
3
4
5
6
7
8
0
9
0
Rate [bits/s/Hz]
(a) Superposed waveforms.
Fig. 4.
1
2
3
4
5
6
Rate [bits/s/Hz]
7
8
9
10
(b) No power waveform.
CR−IDC as a function of N for M = 1, B = 1MHz and SNR=20dB.
average received power of about -20dBm. We assume a
large open space environment with a NLOS channel power
delay profile with 18 taps obtained from model B [52]. Taps
are modeled as i.i.d. CSCG random variables, each with an
average
P18power βl . The multipath response is normalized such
that
l=1 βl = 1. Equivalently, this system model can be
viewed as a transmission over the aforementioned normalized
multipath channel with an average transmit power fixed to
-20dBm. Fig. 3 illustrates the frequency response of one
realization of the multipath channel.
The superposed multi-carrier modulated and unmodulated
waveforms are designed assuming a frequency gap ∆f = B/N
with B = 1M Hz, and centered around 5.18GHz.
Fig 4(a),(b) illustrate the rate-energy (R-E) region obtained
with superposed waveforms (with Algorithm 1) and without
power (multisine) waveform (with Algorithm 1 by allocating
zero power to the unmodulated multisine waveform). M = 1
and N = 1, 2, 4, 8, 16 are considered and transmission is
performed over the frequency response of Fig 3. The noise
power σn2 is fixed at -40dBm in each subband, corresponding
to a SNR (defined as P/σn2 ) of 20dB. For the two strategies
(superposed waveforms and no power waveform), the extreme
right point on the x-axis refers to the rate achieved by the
water-filling solution (with all transmit power allocated to the
multi-carrier communication waveform) with ρ = 0. Note that
in all the figures, the rate has been normalized w.r.t. the bandwidth N Bs . Hence, the x-axis refers to a per-subband rate.
It decreases as N increases because each subband receives a
fraction of the total power. In Fig 4(a), the points on the y-axis
refer to the zDC achieved with the multisine waveform (WPT
only) of [29] with ρ = 1 for N sufficiently large (e.g. 8,16).
A first observation is that with superposed waveforms,
increasing N boosts the harvested energy. On the other hand,
in the absence of multisine waveform, the harvested energy
does not change much. This was predicted from the scaling
laws in Table I that showed that the average zDC with a
multi-carrier modulated waveform is independent of N in
a frequency-flat channel and only increases logarithmically
with N in a frequency-selective channel (as a consequence
of a frequency diversity gain), and is lower than that achieved
with the multisine waveform (which scaled linearly with N ).
We also note that the benefit of using superposed waveforms
really kicks in for N typically larger than 4. For N smaller,
there is no incentive to use a superposed waveform. This is
also well aligned with the observations made from the scaling
laws where a modulated waveform was outperforming the
unmodulated waveform for small N and inversely. Because of
the frequency selectivity of the channel, a frequency diversity
gain is exploited in Fig 4. Its effect on IDC is for instance
noticeable in Fig 4(b) at very low rate. At such rate, the
power is allocated to a single subcarrier/subband, namely the
one corresponding to the strongest channel. For N = 1, the
subcarrier is located at the center frequency while for N = 2,
the two subcarriers (in baseband) are located at -0.25 and 0.25
MHz. At -0.25 MHz the channel gain is very close to the
maximum experienced within the 1MHz band, which explains
why the harvested energy for N = 2 at low rate is only very
slightly lower than that achieved for N > 2. On the other hand,
for N = 1, the loss in harvested energy is more clearly visible
due to the lower channel gain at the center frequency.
A second observation is that without power/WPT waveform,
the R-E region appears convex such that power-splitting (PS)
outperforms time sharing (TS), irrespectively of N . On the
other hand, with the superposed waveform, the R-E region
with PS is convex for low N (N = 2, 4) but concave-convex
for larger N (N = 8, 16). This has for consequences that PSonly is suboptimal. Looking at Fig. 4(b) for N = 8, the convex
hull is obtained by TS between WPT (multisine waveform only
and ρ = 1) and WIPT with 0 < ρ < 1 (illustrated by point
A). Going for N = 16, the convex hull is obtained by TS
between the two extreme points: WPT (multisine waveform
only and ρ = 1) and WIT with water-filling (with ρ = 0).
This shows that the convex hull can in general be obtained
by a combination of TS and PS. This behavior is observed
because of the concacity-convexity of the R-E region, and
fundamentally originates from the rectifier nonlinearity.
Note that those observations are in sharp contrast with the
existing WIPT literature (relying on the linear model). It is
indeed shown in [7], [12] that PS always outperforms TS.
Fig 5(a),(b),(c),(d) further studies the performance of the
WIPT architecture with superposed waveform for N = 16 and
four different SNRs (10,20,30,40 dB) under the same channel
response as in Fig 3. Comparisons are made with two baselines, namely the simplified system with “no power waveform”
and the hypothetical system denoted by “LB” as described in
Section III-D. Comparing the R-E with superposed waveform
13
30
30
Superposed waveform
No power waveform
LB
25
20
zDC [ A]
zDC [ A]
20
Superposed waveform
No power waveform
LB
WPT, =1
25
15
Time sharing
10
15
Time sharing
10
WIT with ASS, =1
5
5
WIT with Water-Filling, =0
0
0
0
0.5
1
1.5
Rate [bits/s/Hz]
2
2.5
0
1
2
(a) SNR=10dB.
5
6
30
Superposed waveform
No power waveform
LB
25
Superposed waveform
No power waveform
LB
25
20
20
zDC [ A]
zDC [ A]
4
(b) SNR=20dB.
30
15
15
10
10
5
5
0
0
0
1
2
3
4
5
Rate [bits/s/Hz]
6
7
8
9
(c) SNR=30dB.
Fig. 5.
3
Rate [bits/s/Hz]
0
2
4
6
8
Rate [bits/s/Hz]
10
12
14
(d) SNR=40dB.
CR−IDC for N = 16 and M = 1 over the multipath channel of Fig. 3 with B = 1MHz and SNR=10,20,30,40 dB.
to the “no power waveform” case, we gain insight into the
joint twofold rate and energy benefit of using a deterministic
power waveform. Comparing it to the the “LB” R-E region,
we get insight into the rate benefit of using a power waveform
that does not incur a rate at the information receiver.
A first observation is that the R-E region with superposed
waveform and its lower bound (LB) is augmented thanks to
the presence of the deterministic multisine power waveform.
This can be seen by noting that the corresponding R-E regions
are larger than that obtained with no multisine waveform (No
power waveform). This key observation fundamentally comes
from the nonlinear behavior of the wireless power channel.
The usefulness of the multisine originates from the fact that
the rectifier is a nonlinear device and that the multisine
is deterministic. On the contrary, the modulated waveform
with CSCG inputs exhibits independent randomness of the
information symbols across sub-carriers that leads to some
uncontrollable fluctuation of the waveform at the input of the
rectifier and therefore some loss in terms of harvested energy.
Interestingly, this observation is in sharp contrast with the
WIPT literature relying on the conventional linear model. It
was indeed shown in [6] that, in the event where the power
waveform is not eliminated (and therefore treated as noise) by
the communication receiver, the power waveform is useless
and does not help enlarging the R-E region. As we can see
from Fig. 5 and Observation 2, even in the worst-case “LB”
where the power waveform is assumed to create interference
to the information receiver, the multisine power waveform is
always useful when the rectifier nonlinearity is properly taken
into account in the design of WIPT.
A second observation is that the gain over “LB” increases
as the SNR increases. At low SNR, the interference from
the power waveform in “LB” is drawn within the noise and
therefore does not impact much the R-E region. Both regions
are therefore quite similar at very low SNR. On the other
hand, as the SNR increases, in order for “LB” not to become
interference limited, the hypothetical system of Section III-D
needs to allocate a very small power to the multisine waveform
over a wide range of rates. For a given zDC , this leads to a rate
loss. This shows that the rate benefit of using a deterministic
waveform increases as the SNR increases.
A third observation is that at lower SNRs (10 and 20dB), the
R-E region achieved superposed waveforms with PS is actually
outperformed by a TS between WPT (multisine waveform
only and ρ = 1) and WIT (communication waveform only
and ρ = 0). At higher SNR (30 and 40dB SNR), PS
performs better than TS between those two extreme points.
This behavior is observed because of the concacity-convexity
of the R-E region. On the other hand, in the absence of the
WPT waveform (e.g. WIT-only transmission), the R-E region
appears convex because of the inefficiency of the modulated
waveform to boost zDC . At low SNR (as on Fig 5(a)), the
water-filling (WF) strategy for rate maximization coincides
with the ASS strategy for energy maximization. Hence the
ASS strategy is used across the boundary and only changing ρ
enables to draw the region boundary. As the SNR increases and
WF allocates power over an increasing number of subbands,
the region boundary with “No power waveform” is getting
rounded because of the log behavior of the rate expression.
In summary, we draw the following conclusions from Fig
4 and 5: 1) A multi-carrier power (multisine) waveform
(superposed to a multi-carrier communication waveform with
14
It is also noteworthy to recall that if we had used the
linear model for the design and evaluation of the R-E region,
following Observation 1, there would not be any benefit of
using a multisine waveform on top of the modulated waveform
since both are equally suitable from an energy harvesting perspective. Moreover, according to Section III-A, the optimum
design of a multi-carrier modulated waveform (with CSCG
inputs) for WPT is the same for both the linear and nonlinear model, namely based on the ASS strategy. These facts
imply that the R-E region achieved with a design of WIPT
based on the linear model is the same as the one achieved
by “No power waveform” in Fig 4 and 512 . Moreover, since
the “No power waveform” strategy is always significantly
outperformed by the superposed waveform in Fig 4 and 5, this
also implies that the design of WIPT based on the nonlinear
rectifier model is more spectrally efficient than a design based
on the linear model. The importance of accounting for the
nonlinearity of the rectifier in the design and evaluations of
WPT was highlighted in [29]. Evaluations in this paper shows
that this is also true for WIPT.
As a final but very interesting observation, we note that the
input distribution in every subband with “No power waveform”
or with a linear model-based design is CSCG. The mean is
always zero and the variance is frequency-dependent. This
is the classical capacity achieving input distribution over a
Gaussian channel with an average transmit power constraint.
Following Remark 2, we note however that with the superposed waveform, the input distribution is Gaussian but
can have a non-zero mean. It is zero-mean as long as only
the multi-carrier modulated communication waveform (with
CSCG inputs) is used, but as we reduce the threshold R̄
and aim at a higher harvested energy, the waveform design
algorithm starts allocating power over the multisine waveform.
This has the effect of changing the input distribution from
zero mean to non-zero mean. Concurrently the variance of the
distribution decreases since the transmit power allocated to the
modulated waveform is decreased, which leads to increasing
K-factors (defined in Remark 2). This also suggests an alternative way of interpreting the transmit waveform (1). Rather than
viewing it as the superposition of a deterministic multisine
and a modulated waveform with CSCG inputs, we can view
it as a multi-carrier modulated waveform with a non-zero
mean Gaussian input distribution. Fig 4 and 5 effectively show
that designing a multi-carrier modulated waveform for WIPT
with non-zero mean Gaussian inputs leads to an enlarged R-E
region compared to the CSCG inputs. An interpretation of Fig
5(b) in terms of K-factor is provided in Fig 6.
12 Validated
by simulations but are not reported here for brevity.
30
Superposed waveform
No power waveform
LB
K-factor=
25
20
zDC [ A]
CSCG inputs) is useful to enlarge the R-E region of WIPT,
for N sufficiently large (N > 4). 2) Without multisine power
waveform, PS is preferred over TS. 3) PS is favoured for low
N , a combination of PS and TS for medium N and TS for
large N . 4) For sufficiently large N , TS is favoured at low
SNR and PS at high SNR. It is important to note that 1), 3)
and 4) are consequences of the non-linearity of the rectifier.
K-factor increasing
15
10
K-factor=0
K-factor=0
5
K-factor=0
0
0
1
2
3
Rate [bits/s/Hz]
4
5
6
Fig. 6. CR−IDC for N = 16 and M = 1 over the channel of Fig. 3 with
B = 1MHz and SNR=20dB. Evolution of the input distribution and K-factor.
B. Validation of the Model and the Scaling Laws
The rectifier model and the scaling laws for multisine were
validated through circuit simulations using PSpice in [29]. We
here conduct a similar evaluation to validate the rectifier model
and the scaling laws for modulated waveforms with CSCG
inputs. To that end, PSpice simulations have been conducted
using the realistic rectenna of Fig 7 (same circuit as the one
used in Fig 10 of [29]) designed for an average input power of
-20dBm (10µW). Details on the circuit design can be obtained
in [29]. Applying the UP strategy to both a multisine waveform
and an OFDM waveform with CSCG inputs and assuming no
wireless channel, i.e. A = 1 and ψ̄ = 0, Fig 8(left) displays the
harvested DC power Pdc measured at the rectifier output load
versus N , for M = 1, ∆f = B/N with B = 10MHz and an
OFDM symbol duration T = 1/∆f .
The DC power of the OFDM waveform has a flat behavior
as a function of N while that of the multisine increases rapidly
with N (up to N = 64)13 . OFDM outperforms multisine for
small N and is outperformed for larger N . This validates the
scaling laws of No CSIT in FF of Table I and Observation 1. It
also highlights the inaccuracy of the linear model that would
have predicted that multisine and OFDM are equally suitable
for WPT. The loss of OFDM compared to multisine for most
N > 2 comes from the random and independent fluctuations of
the input symbols (magnitude and phase) across frequencies
that lead to a random fluctuation of the input waveform, in
contrast with the periodic behavior of the multisine waveform
(which is more suitable to turn on and off the rectifier
periodically). The gain for N ≤ 2 comes from the fourth order
moment of the CSCG distribution that boosts the fourth order
term in z̄DC by a factor 2 compared to the unmodulated case.
It is also worth contrasting with RF experiment results.
In [42], it was shown that (single-carrier) modulations such
QPSK and 16QAM lead to amplitude and phases variations
that are detrimental to the RF-to-DC conversion efficiency
compared to a CW (i.e. N=1) in the input power range 0-10
dBm. This may appear contradicting our results since modulation hurts performance for N = 1. Recall that our results
are obtained with CSCG inputs, not with finite constellations.
The flat scaling law of the OFDM waveform may appear
quite surprising. Recall that both OFDM and multisine waveforms exhibit a larger PAPR as N increases, as illustrated
13 For N > 64, a decrease is observed because the rectenna has been
optimized for N = 4. Further explanations are provided in Fig 11 of [29].
Fig. 7.
Rectenna with a single diode and a L-matching network.
10
-3
0.5
10
-4
0.4
10
-5
10
-6
0.2
10
-7
0.1
10
-8
10
Diode breakdown region
0.3
0
-6
10
-5
10
-4
x 10
10
-3
10
-2
10
-1
P in [W]
−6
3.5
RF-to-DC conversion efficiency
P dc [W]
15
0
10
Fig. 9.
Pdc vs Pin and RF-to-DC conversion efficiency η.
was designed for 10µW input power but as we can see it can
operate typically between 1µW and 1mW. Beyond 1mW, a
rectifier with multiple diodes will perform better and avoid
the saturation problem [21], [22], [39], [40].
2.5
CCDF
PDC [W]
3
−1
10
2
max PAPR, N=1
max PAPR, N=2
max PAPR, N=4
max PAPR, N=8
max PAPR, N=16
PAPR, ∀ N
1.5
Multisine
OFDM with Gaussian inputs
1
0
Fig. 8.
2
4
log2(N)
6
−2
8
10
−2
0
2
4
6
PAPR [dB]
VI. C ONCLUSIONS
8
10
AND
F UTURE W ORKS
12
Pdc vs N (left) and CCDF of PAPR with OFDM vs N (right).
in Fig 8(right) for OFDM. We distinguish the CCDFs of
“max PAPR” for N = 1, 2, 4, 8, 16 and CCDF of “PAPR
∀N ”. The former ones are obtained by drawing the CCDF
of max0≤t≤T |x(t)|2 /E |x(t)|2 (i.e. defined based on the
maximum instantaneous power of the signal over a symbol
2
duration). It is sensitive to N because max0≤t≤T |x(t)|
is likely to increase as N increases. On the
the
contrary,
2
2
latter one displays the CCDF of |x(t)| /E |x(t)|
(i.e.
defined based on the instantaneous power of the signal). It
is insensitive to N because the modulated symbols are CSCG
distributed and so is also the time domain signal irrespectively
of N . For comparison,
of a multisine is given by
the PAPR
2
2
max0≤t≤T |x(t)| /E |x(t)|
= 10 log(2N )[dB], e.g. 6dB
for N = 2 and 9dB for N = 4. Large PAPR has been
advertised in the RF literature as a way to enhance the RF-toDC conversion efficiency [24]–[27]. With OFDM, in contrast
with the multisine, Pdc is insensitive to N despite the increase
of the (max) PAPR with N . Such a behavior of the OFDM
waveform therefore cannot be explained by looking at PAPR.
This shows that just PAPR is not an accurate enough metric
to judge the suitability of a waveform for WPT. It should
also be stressed that in [29] (discussions along Fig 7 and 8 in
that paper), the authors have already stressed and shown that
maximizing PAPR is not a right approach to design efficient
WPT multisine signals in frequency selective channels.
In Fig 9, we illustrate the discussion in Remark 5. We plot
the DC power Pdc harvested at the load of the circuit in Fig
7 as a function of the input power Pin to the rectifier when
a CW (i.e. a single sinewave) signal is used for excitation.
We also display the RF-to-DC conversion efficiency defined
as η = Pdc /Pin . The diode SMS-7630 becomes reverse
biased at 2Volts, corresponding to an input power of about
1mW. We note that beyond 1mW input power, the output
DC power saturates and η suddenly significantly drops, i.e.
the rectifier enters the diode breakdown region. This circuit
The paper derived a methodology to design waveforms
and characterize the rate-energy region of a point-to-point
MISO WIPT. Contrary to the existing WIPT literature, the
nonlinearity of the rectifier is modeled and taken into account
in the WIPT waveform and transceiver optimization. Motivated by the fact that a multisine waveform outperforms a
multi-carrier modulated waveform from a WPT perspective, a
WIPT waveform is introduced as the superposition of a WPT
waveform (multisine) and a WIT (multi-carrier modulated)
waveform. The waveforms are adaptive to the CSI and result
from a non-convex posynomial maximization problem.
Results highlight that accounting for the rectifier nonlinearity radically changes the design of WIPT. It favours a different waveform, modulation, input distribution and transceiver
architecture as well as a different use of the RF spectrum.
Exploiting the rectifier nonlinearity in the WIPT design also
makes a more efficient use of the resources by enabling
enlarged rate-energy regions compared to those obtained by
ignoring the nonlinearity in the system design.
A lot of interesting questions arise from this work. To name
a few, a fundamental question is, how to make the best use
of the RF spectrum for WIPT? Results here highlight that
the conventional capacity-achieving CSCG input distribution
is suboptimal for WIPT. What is the optimal input distribution and how to build corresponding WIPT architecture?
Results also highlight that the design of WIPT differs from
conventional communication due to the inherent nonlinearity
of the wireless power channel. Interesting works consist in
re-thinking WIPT architecture in light of nonlinearity for
broadcast, multiple access, interference and relay channels.
The problem of nonlinearity, waveform design and rate-energy
tradeoff also occurs in other types of communication systems,
such as in backscatter communications [53].
VII. ACKNOWLEDGMENTS
We thank E. Bayguzina for providing Fig 8(left) and Fig 9.
R EFERENCES
[1] B. Clerckx, “Waveform Optimization for SWIPT with Nonlinear Energy
Harvester Modeling,” 20th International ITG WSA, March 2016, Munich.
16
[2] L.R. Varshney, “Transporting information and energy simultaneously,” in
Proc. Int. Symp. Inf. Theory, Jul. 2008, pp. 1612-1616.
[3] P. Grover and A. Sahai, “Shannon meets Tesla: Wireless information and
power transfer,” in Proc. Int. Symp. Inf. Theory, Jun. 2010.
[4] R. Zhang and C. K. Ho, “MIMO broadcasting for simultaneous wireless
information and power transfer,” IEEE Trans. Wireless Commun., vol.
12, no. 5, pp. 1989-2001, May 2013.
[5] H. Son and B. Clerckx, “Joint beamforming design for multi-user wireless
information and power transfer,” IEEE Trans. Wireless Commun., vol. 13,
no. 11, pp. 6397-6409, Nov. 2014.
[6] J. Xu, L. Liu, R. Zhang, “Multiuser MISO Beamforming for Simultaneous
Wireless Information and Power Transfer,” IEEE Transactions on Signal
Processing, vol.62, no.18, pp.4798-4810, Sept.15, 2014.
[7] X. Zhou, R. Zhang and C. K. Ho, “Wireless Information and Power
Transfer: Architecture Design and Rate-Energy Tradeoff,” IEEE Trans.
on Commun., vol. 61, no. 11, pp. 4754-4767, Nov 2013.
[8] J. Park and B. Clerckx, “Joint Wireless Information and Energy Transfer
in a Two-User MIMO Interference Channel,” IEEE Trans. Wireless
Commun., vol. 12, no. 8, pp. 4210-4221, Aug. 2013.
[9] J. Park and B. Clerckx, “Joint Wireless Information and Energy Transfer
in a K-User MIMO Interference Channel,” IEEE Trans. Wireless Commun., vol. 13, no. 10, pp. 5781-5796, Oct. 2014.
[10] J. Park and B. Clerckx, “Joint Wireless Information and Energy Transfer
with Reduced Feedback in MIMO Interference Channels,” IEEE JSAC,
vol. 33, no. 8, pp. 1563-1577, Aug 2015.
[11] K. Huang and E, G. Larsson, “Simultaneous information and power
transfer for broadband wireless systems,” IEEE Trans. Sig. Process., vol.
61, no. 23, pp. 5972-5986, Dec. 2013.
[12] X. Zhou, R. Zhang and C. K. Ho, “Wireless information and power
transfer in multiuser OFDM systems,” IEEE Trans. Wireless Commun.,
vol. 13, no. 4, pp. 2282-2294, Apr. 2014.
[13] D. W. K. Ng, E. S. Lo and R. Schober, “Wireless information and power
transfer: energy efficiency optimization in OFDMA systems,” IEEE Trans.
Wireless Commun., vol. 12, no. 12, pp. 6352-6370, Dec. 2013.
[14] A. A. Nasir, X. Zhou, S. Durrani, and R. A. Kennedy, “Relaying
protocols for wireless energy harvesting and information processing,”
IEEE Trans. Wireless Commun., vol. 12, no. 7, pp. 3622-3636, Jul. 2013.
[15] Y. Huang and B. Clerckx, “Joint Wireless Information and Power
Transfer for an Autonomous Multiple-Antenna Relay System,” IEEE
Comm. Letters, vol. 19, no. 7, pp. 1113-1116, July 2015.
[16] Y. Huang and B. Clerckx, “Relaying Strategies for Wireless-Powered
MIMO Relay Networks,” IEEE Trans. on Wireless Comm, vol 15, no 9,
pp 6033-6047, Sept 2016.
[17] H. Ju and R. Zhang, “Throughput maximization in wireless powered
communication networks,” IEEE Trans. Wireless Commun., vol. 13, no.
1, pp. 418-428, Jan. 2014.
[18] H. Lee, et al “Resource allocation techniques for wireless powered
communication networks with energy storage constraint,” IEEE Trans.
Wireless Commun., vol. 15, no. 4, pp. 2619-2628, Apr. 2016.
[19] K. Huang, C. Zhong, and G. Zhu, “Some new research trends in
wirelessly powered communications,” IEEE Wireless Commun., vol. 23,
no. 2, pp. 19-27, Apr. 2016.
[20] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, “Wireless networks
with RF energy harvesting: a contemporary survey,” IEEE Commun.
Surveys Tuts., pp. 757-789, 2015.
[21] A. Boaventura, A. Collado, N. B. Carvalho, A. Georgiadis, “Optimum
behavior: wireless power transmission system design through behavioral
models and efficient synthesis techniques,” IEEE Microwave Magazine,
vol. 14, no. 2, pp. 26-35, March/Apr. 2013.
[22] A. Costanzo and D. Masotti, “Smart solutions in smart spaces: Getting
the most from far-field wireless power transfer,” IEEE Microw. Mag., vol.
17, no. 5, pp. 30-45, May 2016.
[23] Y. Zeng, B. Clerckx and R. Zhang, “Communications and Signals Design
for Wireless Power Transmission,” IEEE Trans. on Comm, 2017.
[24] M.S. Trotter, J.D. Griffin and G.D. Durgin, “Power-Optimized Waveforms for Improving the Range and Reliability of RFID Systems,” 2009
IEEE International Conference on RFID.
[25] A. S. Boaventura and N. B. Carvalho, “Maximizing DC Power in Energy
Harvesting Circuits Using Multisine Excitation,” 2011 IEEE MTT-S.
[26] A. Collado and A. Georgiadis, “Optimal Waveforms for Efficient Wireless Power Transmission,” IEEE Microw. Wirel. Compon. Lett., May 2014.
[27] C.R. Valenta, M.M. Morys and G.D. Durgin, “Theoretical EnergyConversion Efficiency for Energy-Harvesting Circuits Under PowerOptimized Waveform Excitation,” IEEE Trans. on MTT, May 2015.
[28] B. Clerckx, E. Bayguzina, D. Yates, and P.D. Mitcheson, “Waveform
Optimization for Wireless Power Transfer with Nonlinear Energy Harvester Modeling,” IEEE ISWCS 2015, August 2015, Brussels.
[29] B. Clerckx and E. Bayguzina, “Waveform Design for Wireless Power
Transfer,” IEEE Trans on Sig Proc, vol. 64, no. 23, Dec 2016.
[30] E. Boshkovska, DWK Ng, N. Zlatanov, R. Schober, “Practical non-linear
energy harvesting model and resource allocation for SWIPT systems,”
IEEE Comm. Letters, Vol. 19, No. 12, Dec 2015.
[31] B. Clerckx and E. Bayguzina, “A Low-Complexity Adaptive Multisine
Waveform Design for Wireless Power Transfer,” IEEE Antennas and
Wireless Propagation Letters, Vol. 16, 2017.
[32] Y. Huang and B. Clerckx, “Waveform Optimization for Large-Scale
Multi-Antenna Multi-Sine Wireless Power Transfer,” IEEE SPAWC 2016.
[33] Y. Huang and B. Clerckx, “Large-Scale Multi-Antenna Multi-Sine
Wireless Power Transfer,” IEEE Trans on Sig Proc, in press.
[34] M. R. V. Moghadam, Y. Zeng, and R. Zhang, “Waveform optimization
for radio-frequency wireless power transfer,” arXiv:1703.04006.
[35] Y. Huang and B. Clerckx, “Waveform Design for Wireless Power
Transfer with Limited Feedback,” arXiv:1704.05400.
[36] D. I. Kim, J. H. Moon, and J. J. Park, “New SWIPT using PAPR: How
it works,” IEEE Wireless Commun. Lett., vol. 5, no. 6, Dec. 2016.
[37] J. Kim, B. Clerckx, and P.D. Mitcheson, “Prototyping and Experimentation of a Closed-Loop Wireless Power Transmission with Channel
Acquisition and Waveform Optimization,” IEEE WPTC 2017.
[38] X. Xu, A. Ozcelikkale, T. McKelvey, and M. Viberg, “Simultaneous
information and power transfer under a non-linear RF energy harvesting
model,” in IEEE ICC2017 Workshops, Paris, May 2017.
[39] H. Sun, Z. Zhong, and Y.-X. Guo, “An Adaptive Reconfigurable Rectifier for Wireless Power Transmission,” IEEE Microwave and Wireless
Components Letters,Vol. 23, No. 9, Sept 2013.
[40] C.-J. Li and T.-C. Lee, “2.4-GHz High-Efficiency Adaptive Power
Harvester,” IEEE Trans. on Very Large Scale Integration (VLSI) Systems,
Vol. 22, No. 2, Feb 2014.
[41] Z.B. Zawawi, J. Park and B. Clerckx, “Simultaneous Wireless Information and Power Transfer in a Two-User OFDM Interference Channel,”
IEEE ISWCS 2015, August 2015, Brussels.
[42] H. Sakaki et al., “Analysis of rectifier RF-DC power conversion behavior
with QPSK and 16 QAM input signals for WiCoPT system,” in Proc. 2014
Asia-Pacific Microwave Conf. (APMC), Nov. 2014, pp. 603-605.
[43] S. Lee, L. Liu, and R. Zhang, “Collaborative wireless energy and information transfer in interference channel,” IEEE Trans. Wireless Commun.,
vol. 14, no. 1, pp. 545-557, Jan. 2015.
[44] H. Lee, S.-R. Lee, K.-J. Lee, and I. Lee, “Optimal beamforming
designs for wireless information and power transfer in MISO interference
channels,” IEEE Trans. Wireless Commun., vol. 14, no. 9, Sep. 2015.
[45] L. Liu, R. Zhang, and K.-C. Chua, “Secrecy wireless information and
power transfer with MISO beamforming,” IEEE Trans. Signal Process.,
vol. 62, no. 7, pp. 1850-1863, Apr. 2014.
[46] D. W. K. Ng, E. S. Lo, and R. Schober, “Robust beamforming for secure
communication in systems with wireless information and power transfer,”
IEEE Trans. Wireless Commun., vol. 13, no. 8, pp. 4599-4615, Aug. 2014.
[47] M. Chiang, “Geometric Programming for Communication Systems,”
Foundations and Trends in Communications and Info. Theory, 2005.
[48] M. Grant, S. Boyd, and Y. Ye, “CVX: MATLAB software for disciplined
convex programming [Online],” Available: http://cvxr.com/cvx/, 2015.
[49] B.R. Marks and G.P. Wright, “A general inner approximation algorithm
for nonconvex mathematical programs,” Oper Res., vol. 26, no. 4, 1978.
[50] M. Chiang, C.W. Tan, D.P. Palomar, D. ONeill, and D. Julian, “Power
Control By Geometric Programming,” IEEE Trans. on Wireless Comm.,
Vol. 6, No. 7, pp. 2640-2651, July 2007.
[51] M. Varasteh, B. Rassouli and B. Clerckx, “Wireless Information and
Power Transfer over an AWGN channel: Nonlinearity and Asymmetric
Gaussian Signaling,” IEEE ITW 2017, arXiv:1705.06350.
[52] J. Medbo, P. Schramm, “Channel Models for HIPERLAN/2 in Different
Indoor Scenarios,” 3ERI085B, ETSI EP BRAN, March 1998.
[53] B. Clerckx, Z. Bayani Zawawi and K. Huang, “Wirelessly Powered
Backscatter Communications: Waveform Design and SNR-Energy Tradeoff,” IEEE Commun. Letters, in press.
| 7cs.IT
|
On the number of isomorphism classes of quasigroups
arXiv:1701.06847v1 [math.GR] 24 Jan 2017
P. Jiménez-Serala , I. Lizasoainb , G.Ochoab
a
Departamento de Matemáticas,
Facultad de Ciencias, Universidad de Zaragoza,
Pedro Cerbuna 12, 50009 Zaragoza, Spain
[email protected]
b
Departamento de Matemáticas,
Universidad Pública de Navarra,
Campus de Arrosadı́a, 31006 Pamplona, Spain
[email protected], [email protected]
Abstract
In this paper the number of isomorphism classes of the right transversals of
the symmetric group Σn−1 in the symmetric group Σn is calculated. This
result together with the techniques employed along the paper may be used
to go forward in the study and classification of the subgroups with a certain
number of isomorphism classes of transversals in the group.
Keywords: Right quasigroups, Transversals, Permutation groups.
1. Introduction
Throughout this paper, all groups considered are finite.
The classical theory of extensions of groups has focused on aspects which
are independent of the choice of transversal of the smaller group in the largest
one. For instance, either the representation of the group as a permutation
group of its right cosets, or the induced representation of a group from one
of its subgroups, does not depend, up to equivalence, on the choice of a right
transversal of the subgroup in the whole group (see for instance [5] or [6]).
However, this point of view leads to some lack of accuracy in certain
definitions. For this reason some recent papers have analysed the set of all
right transversals of a subgroup in a group as an abstract object ([2], [3]).
1
The research of the first and the second authors is supported by Proyecto MTM201019938-C03-01/03 of the Ministerio de Ciencia e Innovación de España.
Preprint submitted to Journal of Algebra
March 16, 2018
Any right transversal of a subgroup in a group has the structure of a right
quasigroup with identity. But what really makes this structure important is
that any right quasigroup with identity of order n is isomorphic to a right
transversal of the symmetric group Σn−1 in Σn , as R. Lal showed in the key
paper [4].
The main aim of this paper is to establish a formula to compute the
number of all isomorphism classes of right quasigroups with identity of a
given order. This result is presented in Theorem 3.2.
2. Preliminaries
Definition 2.1. A non-empty set S endowed with an inner operation ◦ is said
to be a right quasigroup if, for any pair a, b ∈ S, there exists a unique x ∈ S
such that x ◦ a = b. If there exists an element e such that e ◦ a = a = a ◦ e,
for any a ∈ S, S will be called a right quasigroup with identity.
Given two right quasigroups with identity S and T , a bijective map f :
S → T is said to be an isomorphism if (x ◦ y)f = xf ◦ y f for any x, y ∈ S.
(Here we use the same symbol ◦ to denote the inner operation of both S and
T ). In this case we will say that S and T are isomorphic.
Remarks 2.2. Let S be a right quasigroup with identity.
(i) The element e which appears in the definition is unique and is called
the identity element of S.
(ii) If S and T are right quasigroups with identity, f : S → T is an isomorphism and e is the identity element of S, then ef is the identity
element of T .
(iii) Each element a ∈ S induces a permutation of the elements of S given
by xa = b if and only if x ◦ a = b, for each x ∈ S.
(iv) Assume that the operation ◦ is associative. For any a ∈ S there exists
a unique element b ∈ S such that b ◦ a = e. Note that (a ◦ b) ◦ a =
a ◦ (b ◦ a) = a ◦ e = a. Then a ◦ b = e. Hence, every element of S
possesses an inverse in S.
Thus, if the operation ◦ is associative, S is a group.
(v) We say that S has order n if the cardinal of the set S is n.
The introduction of right quasigroups with identity, as an algebraic structure, enables us the analysis of the right transversals of a subgroup in a group.
If H is a subgroup of a group G, e denotes the identity element of G and S
2
is a right transversal of H in G such that e ∈ S, then for any x, y ∈ S, there
exists a unique z ∈ S with Hxy = Hz. Therefore the inner operation ◦ in S
defined by
x ◦ y = z ⇐⇒ Hxy = Hz for any x, y ∈ S,
gives an structure of right quasigroup with identity to the set S.
In fact, this example is prototypical.
As usual, we denote the group of all permutations of the set {1, 2, . . . , n}
by Σn . The subgroup of Σn consisting of all permutations fixing the figure
n is isomorphic to Σn−1 . This fact determines an embedding of Σn−1 in Σn
and, in what follows, we consider Σn−1 as a subgroup of Σn . Note that for
any x ∈ Σn , the coset Σn−1 x consists of all permutations moving n to nx .
Next we prove that any finite right quasigroup with identity is isomorphic to
a right transversal of Σn−1 in Σn . This fact appears in [8] (Theorem 3.7) but
we include here a direct, and simpler, proof for the sake of completeness.
Proposition 2.3 ([8] Theorem 3.7). Any right quasigroup with identity of
order n is isomorphic to some right transversal of Σn−1 in Σn .
Proof. Let S = {x1 , x2 , . . . , xn = e} be a right quasigroup with identity
element e. For each 1 ≤ i ≤ n, consider the permutation σi ∈ Σn given by
j σi = k whenever xj ◦ xi = xk . Notice that, for all 1 ≤ i ≤ n, σi is a bijective
map with nσi = i. In addition, σn agrees with the identity permutation of
{1, 2, . . . , n}.
In order to see that {σ1 , . . . , σn } is a right transversal of Σn−1 in Σn ,
suppose that σi = ρσj for some ρ ∈ Σn−1 . Then,
i = nσi = nρσj = nσj = j.
We conclude that σ1 , . . . , σn belong to different right cosets of Σn−1 in Σn and
hence {σ1 , . . . , σn } is a complete set of representatives of the n right cosets
of Σn−1 in Σn .
′
Notice that, for any ρ, ρ′ ∈ Σn , Σn−1 ρ = Σn−1 ρ′ ⇐⇒ nρ = nρ . Hence,
Σn−1 σi σj = Σn−1 σk ⇐⇒ nσi σj = nσk ⇐⇒ xi ◦ xj = xk .
Consequently, the map from S to the right transversal {σ1 , . . . , σn } which
maps xi to σi for each 1 ≤ i ≤ n, provides an isomorphism between both
right quasigroups with identity.
3
For any subgroup H of G, T (G, H) will denote the set of all right transversals of H in G containing e considered as right quasigroups with identity. The
set of all isomorphism classes of these right quasigroups with identity will be
denoted by I(G, H).
Remarks 2.4. Let G be a group and H a subgroup of G. Put |G : H| = n.
(i) Assume that N is a normal subgroup of G with N ≤ H. For any
element g ∈ G, let g = Ng denote the image of g under the natural
epimorphism of G onto G/N. Analogously, for any subset X of G we
write X = {x; x ∈ X}. For any right transversal S ∈ T (G, H), then
S is a right transversal of H in G. In addition, for any x, y ∈ S, if
Hxy = Hz, with z ∈ S, then xy = hz for some h ∈ H and Hx y =
Hxy = H hz = Hz. Hence S and S are isomorphic quasigroups with
identity. In other words, the natural epimorphism of G onto G induces
a bijection between I(G, H) and I(G, H).
(ii) If H and K be subgroups of G with HK = G, then any right transversal
of H ∩ K in K is also a right transversal of H in G. In addition,
I(K, H ∩ K) ⊆ I(G, H).
(iii) If f : G1 → G2 is a group isomorphism and H1 ≤ G1 , then the restriction of f to any right transversal S of H1 in G1 provides a quasigroup
isomorphism between S and the right transversal S f of H1f in G2 .
(iv) The natural action of G on the set of right cosets of H in G provides a
group homomorphism ϕ from G to Σn . In fact, for any right transversal
{x1 , . . . , xn = e} of H in G, iϕ(g) = k if and only if Hxi g = Hxk . In
particular,
iϕ(xj ) = k ⇐⇒ Hxi xj = Hxk ⇐⇒ xi ◦ xj = xk .
Notice that nϕ(xi ) = i for any 1 ≤ i ≤ n and nϕ(h) = n for every h ∈ H.
Hence, ϕ(H) ⊆ Σn−1 . Moreover, {ϕ(x1 ), . . . , ϕ(xn ) = id} is a right
transversal of Σn−1 in Σn .
In addition, if {x1 , . . . , xn = e} and {y1 , . . . , yn = e} are right transversals of H in G, then they are isomorphic as quasigroups if and only if
{ϕ(x1 ), . . . , ϕ(xn ) = id} and {ϕ(y1 ), . . . , ϕ(yn ) = id} are isomorphic.
(v) If CoreG (H), the largest normal subgroup of G contained in H, is the
identity subgroup, then ϕ is one-to-one and we can consider G ≤ Σn
and H ≤ Σn−1 . Therefore any right transversal of H in G is a right
transversal of Σn−1 in Σn in this case.
4
3. Counting the isomorphism classes of right quasigroups with
identity
For any positive integer n, let QG(n) denote the number of isomorphism
classes of right quasigroups with identity of order n. In this section we
present a formula to reckon QG(n). By Proposition 2.3 we have to calculate
the cardinal of the set I(Σn , Σn−1 ). Thus it is necessary to give some notation
of permutation groups as it will appear in this section.
Let t denote any conjugacy class of Σn−1 . The cardinal of t will be denoted
by at and we write ct = |CΣn−1 (σ)|, where σ is a representative of the elements
in t and CΣn−1 (σ) denotes the centralizer of σ in Σn−1 . Recall that at ct =
(n − 1)!. It is well-known that each t is characterized by an (n − 1)-tuple of
non-negative integers (rt,1 , . . . , rt,n−1 ) with rt,1 +2rt,2 · · ·+(n−1)rt,n−1 = n−1,
where rt,k is the number of cycles of length k for each k = 1, 2, . . . , n − 1,
occurring in any element of t written as a product of disjoint cycles (see [7]
Theorems 3.1, 3.2 and 3.10).
Note that next Lemma 3.1 is a particular case of a result proved in the
paper of R. Lal, [4] Theorem 2.6. We include here a proof written in terms
of permutation groups.
Lemma 3.1. Two right transversals of Σn−1 en Σn are isomorphic (as right
quasigroups with identity) if and only if they are conjugate by some σ ∈ Σn−1 .
Proof. Suppose that S = {x1 , · · · , xn = e} and T = {y1 , · · · , yn = e} are
right transversals of Σn−1 in Σn such that, for any 1 ≤ i ≤ n, nxi = nyi = i.
Note that if τi denotes the transposition (i, n), then Σn−1 xi = Σn−1 τi =
Σn−1 yi .
Let f : S → T be a bijective map providing the isomorphism between S
and T . Define σ ∈ Σn−1 by means of iσ = m whenever xif = ym . This means
xif = yiσ . For any 1 ≤ i, j ≤ n, take k such that xi ◦ xj = xk . Then,
σ
(iσ )xj = (iσ )σ
f
−1 x
jσ
f
= ixj σ = n(xi ◦xj )σ = nxk σ = k σ = nykσ = nxk
f
= n(xi ◦xj ) = n(yiσ ◦yjσ ) = (iσ )yjσ ,
which shows that xjσ = yj σ .
For the converse, conjugation by any σ ∈ Σn−1 is an automorphism of Σn
which fixes Σn−1 . Then it provides a quasigroup isomorphism between any
right transversal of Σn−1 in Σn and its σ-conjugate.
5
Theorem 3.2. The number of isomorphism classes of right quasigroups with
identity of order n is
QG(n) =
1
|Σn−1 |
X
r
r
r
t,n−1
at ct1t,1 ct2t,2 · · · ctn−1
t
where t runs over the set of all conjugacy classes of Σn−1 and, for each t, if
σ ∈ t, we write tk to denote the conjugacy class of σ k , for k = 1, . . . , n − 1.
Proof. As commented above, by Proposition 2.3, QG(n) = |I(Σn , Σn−1 )|. By
Lemma 3.1, QG(n) is the number of orbits of the action by conjugation of
the group Σn−1 on the set Tn = T (Σn , Σn−1 ). By a formula due to Burnside
(see [1], Theorem 14.4), we have
QG(n) =
1
X
|Σn−1 | σ∈Σ
|fix(σ)|
(1)
n−1
where fix(σ) = {S ∈ Tn ; S σ = S}.
Fix σ ∈ Σn−1 and consider that σ = ρ1 . . . ρm is the decomposition of σ as
a product of disjoint cycles including the cycles of length one. Consider any
right transversal S = {x1 , . . . , xn = e} of Σn−1 in Σn . Assume that, for each
σ
−1
i = 1, . . . , n, xi ∈ Σn−1 τi where τi = (i, n). Since nxi = nσ xi σ = nxi σ = iσ ,
we have xiσ ∈ Σn−1 xiσ for any i = 1, . . . , n − 1. Therefore S σ = S if and only
if xiσ = xiσ , for all i = 1, . . . , n − 1. Note that i and iσ belong to the same
cycle in the decomposition of σ. This is to say that if j and i appear in the
same cycle of the decomposition of σ and S σ = S, then the element xj ∈ S
is determined by xi and σ. Thus, for each l ∈ {1, . . . , m} choose a figure
i(l) belonging to the cycle ρl . From the above argument we deduce that in
order to get a right transversal S ∈ fix(σ), we only have to give elements
xi(1) , . . . , xi(m) since the rest are determined by these elements and σ. If for
each l ∈ {1, . . . , m}, we write Pσ (l) to denote the number of possible xi(l) ,
then we have
m
Y
|fix(σ)| =
Pσ (l).
l=1
Let l ∈ {1, . . . , m}. Write i for i(l) and kl for the length of the cycle ρl .
Note that σ kl fixes both figures i and n. Now, any representative of the coset
Σn−1 τi is of the form xi = hτi for some h ∈ Σn−1 . If, in addition, xi ∈ S
6
k
for some S ∈ fix(σ), then xiσ l = xiσkl = xi and consequently xi ∈ CΣn (σ kl ).
But τi also belongs to CΣn (σ kl ) because σ kl fixes both i and n. Hence,
h = xi τi−1 ∈ CΣn (σ kl ) ∩ Σn−1 = CΣn−1 (σ kl ).
Conversely, if h ∈ CΣn−1 (σ kl ), then
k
k
{xiσ = hσ τiσk ; k = 0, 1, . . . , kl − 1}
is a set of representatives of the right cosets Σn−1 τiσk , k = 0, 1, . . . , kl − 1,
which is fixed by σ.
Hence, writing t to denote the conjugacy class of σ, we have
Pσ (l) = |CΣn−1 (σ kl )| = ctkl
and then
|fix(σ)| =
m
Y
ctkl .
l=1
Note that |fix(σ)| depends only on the structure of σ as a product of disjoint
cycles and two conjugate permutations have the same cycle structure. Thus
in formula (1) we can write
QG(n) =
1
X
|Σn−1 | σ∈Σ
|fix(σ)| =
n−1
1
|Σn−1 |
X
at |fix(σt )|
t
where t runs over all conjugacy classes of Σn−1 and σt is a representative
of the class t. Now, if the members of the class t are characterized by the
(n − 1)-tuple (rt,1 , . . . , rt,n−1 ), then
r
r
r
t,n−1
.
|fix(σt )| = ct1t,1 ct2t,2 · · · ctn−1
Now, the formula becomes clear.
Example 3.3. Let us calculate QG(6), the number of isomorphism classes of
right quasigroups with identity of order 6. Equivalently we want to calculate
the cardinal |I(Σ6 , Σ5 )| of the isomorphism classes of right transversals of Σ5
in Σ6 .
Since there are seven ways to decompose 5 as a sum of positive integers:
1 + 1 + 1 + 1 + 1 = 1 + 1 + 1 + 2 = 1 + 2 + 2 = 1 + 1 + 3 = 1 + 4 = 5 = 2 + 3,
7
there are seven conjugacy classes in Σ5 which can be denoted as
t11111 , t1112 , t122 , t113 , t14 , t5 , t23 ,
A direct calculation gives the following table:
class represent. at ct (rt,1 , rt,2 , rt,3 , rt,4 , rt,5 )
(ct1 , ct2 , ct3 , ct4 , ct5 )
t11111
id
1 120
(5, 0, 0, 0, 0)
(120, 120, 120, 120, 120)
t1112
(12)
10 12
(3, 1, 0, 0, 0)
(12, 120, 12, 120, 12)
t122
(12)(34) 15 8
(1, 2, 0, 0, 0)
(8, 120, 8, 120, 8)
t113
(123)
20 6
(2, 0, 1, 0, 0)
(6, 6, 120, 6, 6)
t14
(1234)
30 4
(1, 0, 0, 1, 0)
(4, 8, 4, 120, 4)
t5
(12345) 24 5
(0, 0, 0, 0, 1)
(5, 5, 5, 5, 120)
t23
(12)(345) 20 6
(0, 1, 1, 0, 0)
(6, 6, 12, 6, 6)
Hence, the number of isomorphism classes of right quasigroups with identity of order 6 is
1 5
120 + 10(123 · 120) + 15(8 · 1202)+
120
20(62 · 120) + 30(4 · 120) + 24 · 120 + 20(6 · 12) = 207392556.
QG(6) =
Remarks 3.4. (i) Note that the above formula to calculate QG(n) looks
similar to an evaluation of the poynomial associated to the symmetric
group Σn−1 (see [1]). However, appearances deceive!
(ii) Any group of order n is a right quasigroup with identity and therefore it
is isomorphic, as a right quasigroup with identity, to a right transversal
of Σn−1 in Σn . Moreover, two groups are isomorphic if and only if they
are isomorphic as right quasigroups with identity.
Acknowledgement: The authors would like to thank Professor L. M.
Ezquerro (Dpto. de Matemáticas, Universidad Pública de Navarra) for his
very helpful comments and suggestions in improving this paper.
References
[1] N. L. Biggs. Discrete mathematics. Oxford University Press, New York,
2002.
8
[2] V.K. Jain and R. P. Shukla. On the isomorphism classes of transversals.
Com. in Algebra 36 (2008) 1717-1725.
[3] V.K. Jain an R. P. Shukla. On the isomorphism classes of transversals
II. Com. in Algebra 39 (2011) 2024-2036.
[4] R. Lal. Transversals in Groups. J. Algebra 181 (1996) 70-81.
[5] A. Lucchini, F. Menegazzo, M. Morigi. Asymptotic results for transitive
permutation groups. Bull. London Math. Soc. 32 (2000) 191-195.
[6] A. Lucchini, M. C. Tamburini. Tensor products of primitive modules.
Archiv der Mathematik 77 (2001) 149-154.
[7] J. Rotman. An Introduction to the Theory of Groups. Springer, New
York, 1995.
[8] R. P. Shukla. Congruences in right quasigroups with identity and general
extensions. Com. in Algebra 23 (1995) 2679-2695.
9
| 4math.GR
|
1
Joint Content Delivery and Caching Placement
via Dynamic Programming
arXiv:1801.00924v1 [cs.IT] 3 Jan 2018
Bojie Lv, Lexiang Huang and Rui Wang
Abstract
In this paper, downlink delivery of popular content is optimized with the assistance of wireless cache
nodes. Specifically, the requests of one file is modeled as a Poisson point process with finite lifetime,
and two downlink transmission modes are considered: (1) the base station multicasts file segments to the
requesting users and selected cache nodes; (2) the base station proactively multicasts file segments to the
selected cache nodes without requests from users. Hence the cache nodes with decoded files can help to
offload the traffic upon the next file request via other air interfaces, e.g. WiFi. Without proactive caching
placement, we formulate the downlink traffic offloading as a Markov decision process with random
number of stages, and propose a revised Bellman’s equation to obtain the optimal control policy. In order
to address the prohibitively huge state space, we also introduce a low-complexity sub-optimal solution
based on linear approximation of the value functions, where the gap between the approximated value
functions and the real ones is bounded analytically. The approximated value functions can be calculated
from analytical expressions given the spatial distribution of requesting users. Moreover, we propose
a learning-based algorithm to evaluate the approximated value functions for unknown distribution of
requesting users. Finally, a proactive caching placement algorithm is introduced to exploit the temporal
diversity of shadowing effect. It is shown by simulation that the proposed low-complexity algorithm
based on approximated value functions can significantly reduce the resource consumption at the base
station, and the proactive caching placement can further improve the performance.
I.
INTRODUCTION
Caching is a promising technology to improve the network spectral efficiency [2] or cut down
energy consumption [3], [4] by predicting the potential content requests of users. In this paper,
we consider a flexible deployment scenario where there is no wired connection or dedicated
Bojie Lv, Lexiang Huang and Rui Wang are with Department of Electrical and Electronic Engineering, The Southern University
of Scienece and Technology, China, Email: {[email protected], [email protected], [email protected]}
Part of this work has been submitted to IEEE ICC 2018 [1]. We have extended the conference paper by including the learningbased algorithm in Section IV-C and proactive scheduling algorithm in Section V, and adding more illustrative simulation results.
2
spectrum between the base station (BS) and cache nodes. The cache nodes have to update
their buffers via downlink, sharing the same transmission resources with ordinary users. Hence
the joint content delivery and caching placement problem shall be addressed to minimize the
downlink resource consumption.
A. Related Works
With the wired connection between cache nodes and BSs, the buffer capacity limitation at
the cache nodes becomes the major performance bottleneck, and there have been a number of
research efforts spent on the file placement of the cache nodes. For example in [5], it was shown
that prediction of future demand will improve the performance of file placement at the cache
nodes. The authors in [6] evaluated the performance of wireless heterogeneous networks, where
the small BSs (cache nodes) cache the most popular files under the buffer capacity limitation
at the cache nodes. Alternatively, the authors in [7] showed that caching the files randomly
with optimized probabilities is better than saving the most popular files when each user can
be served by multiple cache nodes. In order to offload more traffics to cache nodes, a file
placement algorithm was derived in [8] with the constrains of cache nodes’ capacity and backhaul
rates between the cache nodes and macro BS. In cooperative interference networks, the authors
proposed a short-term MIMO precoding design and a long-term file placement policy to minimize
the transmission power [4]. In [9], [10], the authors proposed multicast algorithms with coded
caching scheme. All the above works assumed that there are wired links between the cache
nodes and BSs. However, it might be costly to deploy cache nodes with wired connection in
some areas, and hence the caching placement via wireless links (e.g., cellular downlink) should
also be studied [11]. In [12], the transmissions from BS to cache nodes and from cache nodes to
users shared the same spectrum. However, it neglected the possibility that both cache nodes and
users could listen to the BS simultaneously via a multicast mode. In fact, if file caching has to
be made via downlink, the phase of caching placement can be coupled with the phase of content
delivery. For example, at the first few transmissions of one file, both cache nodes and requesting
users could listen to the BS simultaneously; and the cache nodes can help to forward the file
as long as they have been able to decode it. Thus the downlink strategy should be optimized
spanning the whole lifetime of a file (including the transmission to both cache nodes and users).
However, this has not been addressed by the existing literature.
3
When the caching placement shares the same transmission resource as the downlink of ordinary
users, there will be a multi-stage optimization issue. For example, if more transmission resource
is spent on caching placement in the previous stages, more traffic may be offloaded in the
following stages, and vice versa. Thus the transmission dynamics in time domain should be
captured to minimize the overall downlink resource consumption, and the method of dynamic
programming can be utilized. In fact, the dynamic programming via Markov decision process
(MDP) has been considered in the delay-aware resource allocation of wireless systems. For
example, the infinite-horizon MDP has been used to optimize the cellular uplink [13], [14] and
downlink transmissions [15], and relay networks [16], where the average transmission delay
is either minimized or constrained. Moreover, the low-complexity algorithm design is usually
considered in the above works to avoid the curse of dimensionality [17]. However, the popular
files to be buffered at the cache nodes should usually have a finite lifetime, and the infinitehorizon MDP may not be suitable in modeling anymore. Nevertheless, the MDP with finite
stages is usually more complicated [18], and it is still an open issue on the low-complexity
algorithm design with finite-horizon MDP.
B. Our Contributions
In this paper, we consider the downlink file transmission with the assistance of cache nodes,
where the caching placement shares the downlink resource with ordinary users. Specifically,
two downlink transmission modes are considered: (1) the BS multicasts file segments to the
requesting users and selected cache nodes; (2) the BS proactively multicasts file segments to
the selected cache nodes without requests from users (namely proactive caching placement).
The communication links between cache nodes and users are via different spectrum from the
downlink (e.g., Wi-Fi) as [19], [20], and the BS tries to minimize the average downlink resource
consumption by offloading traffic to cache nodes. The main contributions of this work are
summarized below:
•
Without proactive caching placement, we model the downlink transmission of one file within
his lifetime as a Markov decision process (MDP) with random number of stages. Note that
this is not conventional MDP problem with finite and fixed number of stages, we propose
a revised version of Bellman’s equation, where the optimal control policy can be obtained
given the value functions. Then we introduce a linear approximation on the value functions
so that the exponential complexity can be reduced into linear. With the knowledge of spatial
4
distribution of requesting users, the approximated value functions can be calculated via
analytically expression; whereas, a learning algorithm is also introduced to evaluate the
approximated value functions with unknown distribution of requesting users.
•
The bounds on the approximation error of value functions are also obtained. Based on it,
we further derive the bounds on the overall average transmission cost of the BS.
•
A per-stage optimization approach is proposed to include the proactive caching placement
into the above optimization framework.
It is shown by simulation that the proposed low-complexity algorithm based on approximated
value functions can significantly reduce the resource consumption at the base station, and the
proactive caching placement can further improve the performance.
The remainder of this paper is organized as follows. In Section II, the system model is
introduced. In Section III, we formulate the downlink resource allocation without proactive
caching placement as a MDP with random number of stages, and introduce the optimal solution.
In Section IV, a linear approximation is proposed to the value functions, and the bound on
approximation error is derived. In Section V, a per-stage optimization approach is proposed to
include the proactive caching placement. The numerical simulation is provided in Section VI,
and the conclusion is drawn in Section VII.
II. S YSTEM M ODEL
A. File Request Model
As illustrated in Fig. 1, the downlink transmission in a cell with one multi-antenna BS, NC
single-antenna cache nodes is considered. Let NT be the number of antennas at the BS. Without
loss of generality, it is assumed that the BS locates in the origin, and the locations of cache nodes
are denoted as c1 , c2 , ..., and cNC respectively, which can be arbitrary in the cell coverage. Let
Ci (∀i = 1, 2, ..., NC ) be the coverage region of the i-th cache node, and C = Ci ∪ Ci ∪ ... ∪ CNC .
It is reported in the existing literature that a significant portion of the downlink data may
be requested by multiple users of one cell, e.g. popular videos, web pages and etc, which can
be buffered at the cache nodes for future transmissions. This type of data is usually organized
as files. We shall focus on the offloading of downlink resource for the these files. However,
we do not exclude the possibility that BS may simultaneously transmit data dedicated to one
certain user, which is not necessarily buffered at the cache nodes. Without loss of generality,
5
Cache
node
User
Coverage Region of Cache
node
BS with multiantenna
Cache Node
Fig. 1. Illustration of network model with one BS and multiple wireless cache nodes.
it is assumed that each file consists of RF information bits, which are further divided into NS
segment equally.
In most of the existing literature on wireless caching networks, the popularity of downlink
files is modeled by an access probability, which may follow the Zipf distribution [21]. This
model is applicable when the caching placement and content delivery to users are scheduled
with separated resource (i.e. different time or spectrum). However, these two phases are coupled
in this paper. Specifically, the scenario we consider is described below:
•
There is no wired or dedicated wireless connection between cache nodes and BS. Hence
the cache nodes can only receive files via downlink.
•
The downlink files can be requested by users since it is generated, and the downlink transmission to users and cache nodes can happen simultaneously. Hence there is no dedicated
phase of caching placement as assumed in the existing literature.
For instance, users may not find the desired file from the nearby cache nodes in the very
beginning of the file’s lifetime (the cache nodes have not decoded the file). After certain times
of downlink file transmission, the situation can be different. This is because some cache nodes
may have successfully decoded the above file. Thus the following file transmission is strongly
related to the previous transmissions. In order to capture this temporal dynamics of the coupled
user requesting and caching placement, we introduce the following file request model based on
6
Poisson point process (PPP).
Model 1 (File Request Model). For the elaboration convenience, it is assumed that each file
has the same lifetime T . Suppose the f -th file (f = 1, 2, ...) is available for access since time
instance tf , we consider the requests on this file happening during the time period [tf , tf + T ].
The Poisson point process is adopted to model the event of file request within a file’s life time,
and λf is the process intensity of the f -th file. Hence the probability mass function of request
number on the f -th file (∀f ), when the remaining lifetime is Trem , is given by
(λf Trem )n −λf Trem
e
.
Pr(Request Number = n) =
n!
(1)
The locations of new requesting users are independently and identically (i.i.d.) distributed
following certain spatial distribution in the cell coverage. It is assumed that users’ locations
do not change during the file transmission, and the users become inactive after receiving their
desired files.
Note that Poisson point process has been widely used to model the phone call at an exchange.
Moreover, given a time duration, the above model could provide an probability that one file is
requested by at least one user, which degenerates to the access model in the existing literature.
For example, the probability that the f -th file is requested by at least one user within a time
duration τ is given by
Pf = 1 −
∞
X
(λf τ )n
n=1
n!
e−λf τ .
In this paper, we shall first assume that the spatial distribution of new requesting users is
known to the BS, and derive a low-complexity resource scheduling algorithm. Then we shall
continue to consider the scenario that the BS does not know the distribution of new requesting
users, and introduce a learning-based algorithm to adapt the proposed low-complexity resource
scheduling algorithm.
B. File Transmission with Wireless Caching
Since the cache nodes rely on downlink to obtain the popular files and each file can be
requested at the very beginning of its lifetime, the phases of caching placement and user
requesting are coupled, and it is necessary to consider them jointly (Note that these two phases are
usually considered in a separated way in most of the existing literature [22], [23]). Specifically,
7
there are three types of communication links in the cellular networks with wireless cache nodes,
which are introduced below.
•
Requested downlink multicast: the BS delivers the requested file segments to user and the
chosen cache nodes simultaneously. This will happen when the requested segments cannot
be found in the nearby cache nodes. For example, if one user is within the coverage of Ci ,
it will receive the file segments from the BS when these file segments cannot be found in
the i-th cache node. Since both requesting user and cache nodes can listen when the BS is
transmitting, the transmission mode is multicast.
•
Device-to-device (D2D) communications: the cache nodes forward the requested file
segments to users. This will happen when the requested segments hit the buffer of nearby
cache nodes. For example, if one user is within the coverage of Ci , it will receive the
file segments from the i-th cache node when the file segments can be found there. This
communication is made directly from cache nodes to users. It can use Wi-Fi, bluetooth,
or other air interfaces, which is not in the same spectrum as downlink. For example, the
offloading from cellular to Wi-Fi has attracted a number of attenuations [19] [20].
•
Proactive caching placement: the BS delivers some file segments to cache nodes without
the request from users. It is assumed that the BS will determine whether to deliver one file
segment, which is still in its lifetime, to cache nodes proactively in every Tp seconds.
Without proactive caching placement, the first transmission of each file is made directly from
the BS, and some cache nodes may also be able to decode the whole file or some segments. For
the successive requests of the same file, user will receive one segment from the cache node if
the following two conditions are satisfied: (1) the user is in the coverage of certain cache node;
(2) the aforementioned file segment has been successfully decoded by the aforementioned cache
node. If more transmission resource (transmission power or symbols) is spent in the previous
rounds of file transmission, more cache nodes would be able to buffer the file, which may save the
transmission resource of the BS in the future. Clearly, it is necessary to exploit the correlation
among different transmissions of the same file in downlink resource allocation. Furthermore,
if proactive caching placement is allowed, the BS may multicast certain file segments to the
chosen cache nodes according to their downlink channels even without the requests from users.
This decision might be made if the BS believes the proactive caching placement can save more
transmission resource in the future transmission.
8
Example 1 (Downlink File Delivery via Wireless Caching). Consider the following example:
the downlink file is divided into two segments; there are two users A and B, which will raise
the requests on this file in sequence; and User B is in the coverage of cache node (say Cache
Node C). Suppose when User A requests the file, the Cache Node C has not buffered it. The BS
chooses User A and Cache Node C as the downlink receivers. Hence when the User B requests
the same file, it can be served by Cache Node C.
Moreover, if the proactive caching placement is allowed, the BS may find the channel to the
Cache Node C is very good, and deliver the first segment of the file to the Cache Node C before
the request from User A. When the User A requests the first file, it can obtain the first segment
from the Cache Node C and another segment from the BS.
Since the data rate of wireless communications is usually much smaller than wired, and the
received files can be removed from the buffer of cache nodes when it is expired, we ignore the
limitation on capacity of cache buffer in this paper 1 . Moreover, since the D2D communications
can be done distributively and parallelly with relatively low transmission power, we shall focus
on the transmission resource consumption at the BS only, which is the bottleneck of the overall
system.
C. Downlink Physical Layer Model
In downlink, the receivers include the requesting user and cache nodes, and the space-time
block code (STBC) with full diversity is used at the BS to facilitate the multicast communications.
There are two main reasons that STBC is chosen for downlink multicast: (1) the BS need not to
collect channel state information (CSI); (2) the full diversity can be achieved at all the receivers.
It has been well studied in the existing literature that the user selection strategy in downlink
frames (the typical duration of one frame is 10 milliseconds) will affect the transmission delay.
For example, the delay-aware downlink control has been investigated in [24], [25]. In this paper,
however, we shall consider the resource allocation in the time scale of file segment transmission, which consists of thousands of frames. In order to avoid the prohibitively complicated
optimization on multiple time scales, we make the following simple and practical assumption:
the transmission of one file segment is within the coherent time of shadowing attenuation, and
1
For example, suppose that one BS is transmitting downlink files with overall data rate of 1 Gbps, and the lifetime of each file
is 24 hours. Then the maximum required storage capacity of one cache node is around 10T bytes, which is a mild requirement
for ignoring the buffer capacity limitation.
9
the shadowing attenuation for the transmission of different segments of one file is i.i.d. Based
on this assumption, we shall consider the optimization of cache node selection, transmission
power level and transmission time of each file segment for both requested downlink multicast
and proactive caching placement. Note that when the above three parameters are determined for
each segment transmission, the user selection for each frame can be further considered in the
finer time scale to optimize the average transmission delay, which is outside the scope of this
paper.
In the requested downlink multicast, we shall refer to the user, which raises the n-th request
on the f -th file, as the (f, n)-th user, and refer to the s-th segment of the f -th file as the (f, s)th segment. Since the transmission time of one file segment is much larger than the channel
coherent time of small-scale fading, it is assumed that the ergodic channel capacity span all
possible small-scale channel fading can be achieved during one segment transmission. Let ρf,n
and ρc be the pathloss from the BS to the (f, n)-th user and the c-th cache node respectively,
c
ηf,n,s and ηf,n,s
be the corresponding shadowing attenuation in n-th transmission of the (f, s)-th
segment. In the downlink transmission of one file, the location of requesting user is assumed to
be static. Let Pf,n,s be the downlink transmission power of the s-th file segment in response to
request of the (f, n)-th user, Nf,n,s be the number of downlink transmission symbols scheduled to
deliver the s-th segment to the (f, n)-th user, the throughput achieved by the (f, n)-th downlink
user in the transmission of the s-th segment is given by
||hf,n,s ||2 Pf,n,s
,
Rf,n,s = Nf,n,s Ehf,n,s log2 1 +
NT σz2
(2)
where σz2 is the power of noise as well as inter-cell interference, hf,n,s is the i.i.d. channel vector
from the BS to the requesting user. Each element of hf,n,s is complex Gaussian distributed with
zero mean and variance ρf,n ηf,n,s . As a remark note that the transmission of one segment may
consume a large number of frames, and the channel vector hf,n,s can be different from frame
to frame. However, since we consider the ergodic channel capacity, the randomness in smallscale fading is averaged. Hence, the (f, n)-th user can decode the s-th segment only when
Rf,n,s ≥ RF /NS . Similarly, the achievable data rate of the c-th cache node is given by
||hcf,n,s ||2 Pf,n,s
c
Rf,n,s = Nf,n,s Ehcf,n,s log2 1 +
,
NT σz2
(3)
where hcf,n,s is the i.i.d. channel vector from the BS to c-th cache node. Each element of hcf,n,s
10
c
is complex Gaussian distributed with zero mean and variance ρc ηf,n,s
. The c-th cache node can
c
decode the s-th segment only when Rf,n,s
≥ RF /NS .
Moreover, let ηkc be the shadowing attenuation in the k-th proactive transmission opportunities
of the BS to the c-th cache node, Pk and Nk be the corresponding downlink transmission power
and transmission symbols. The throughput achieved by the c-th cache node is given by
||hck ||2 Pk
c
,
Rk = Nk Ehck log2 1 +
NT σz2
(4)
where hck is the i.i.d. channel vector from the BS to c-th cache node. Each element of hck is
complex Gaussian distributed with zero mean and variance ρc ηkc . The file segment can be decoded
when Rkc ≥ RF /NS .
There is an scheduling problem in the downlink transmission: if more power or time resource
is spent in downlink file transmission, more cache nodes are able to buffer the file (or some
segments of the file), which may save the resource of the BS in the successive transmissions of
the same file. Thus it is necessary to optimize the resource allocation during the whole lifetime
of one file, instead of single transmission. This is a multi-stage optimization problem, which
shall be addressed by the following two steps in this paper:
•
Without proactive caching placement of the BS, we shall first formulate the downlink transmission as a random-stage MDP, and propose both optimal and low-complexity solutions,
which is elaborated in Section III and IV.
•
In Section V, proactive caching placement is considered to further suppress the transmission
cost of BS.
III. O PTIMAL C ONTROL WITHOUT P ROACTIVE C ACHING P LACEMENT
A. Problem Formulation
Without proactive caching placement, the system state and scheduling policy are defined below.
Definition 1 (System State). Before the transmission to the (f, n)-th user, the system status is
c
c
c
uniquely specified by Sf,n = Bf,s
, ρf,n , ηf,n,s , ηf,n,s
|∀c = 1, ..., NC ; s = 1, ..., NS , where Bf,s
=
1 means that the s-th segment of the f -th file has been successfully decoded by the c-th cache
c
node and Bf,s
= 0 means otherwise.
Definition 2 (Downlink Multicast Policy). Suppose the s-th segment is transmitted to the (f, n)th user via downlink. Given the system state Sf,n , the scheduling policy Ωf,n (∀f, n) is a mapping
11
from system state and the remaining lifetime Tf,n to the scheduling parameters Pf,n,s and Nf,n,s
(∀s). Thus
Ωf,n (Sf,n , Tf,n ) = {(Pf,n,s , Nf,n,s )|∀s}.
Moreover, to guarantee that the requesting user can successfully decode the downlink data, the
constraint Rf,n,s ≥ RF /NS (∀s) should be satisfied.
s
i =1 Ci be the area where the reqeusting users is able to receive the (f, s)-th
= ∪∀i,Bf,s
Let Cf,n
file segment from cache nodes, and lf,n be the location of the (f, n)-th user. We use the following
cost function to measure the weighted sum of energy and transmission time of the BS, which
is spent on the (f, n)-th user for the s-th segment.
s
gf,n,s (Pf,n,s , Nf,n,s ) = I(lf,n ∈
/ Cf,n
) × (we Pf,n,s Nf,n,s + wt Nf,n,s ),
where we and wt are the weights on transmission energy and transmission opportunities respectively, and I(·) is the indicator function. Since transmission from cache nodes to users is made
via other air interfaces, the transmission resource used by the cache nodes is not counted in this
cost function. Hence the average cost spent on the overall lifetime of the f -th file is given by
"
#
N NS
X
(λf T )N −λf T X X
Eη,ρ
e
gf,n,s ,
g f ({Ωf,n |∀n}) =
N!
n=1 s=1
N
where the expectation is taken over all possible large-scale channel fading (including the shadowing effect η and requesting users’ pathloss ρ) in the system. The summation on N is due to
the random number of requests as elaborated in (1). As a result, the overall system cost function
is given by
F
1 X
G({Ωf,n |∀f, n}) = lim
g f ({Ωf,n |∀n}) ,
F →+∞ F
f =1
and the system optimization problem can be written as
Problem 1 (Overall System Optimization).
min
G({Ωf,n |∀f, n})
s.t.
Rf,n,s ≥ RF /NS , ∀f, n, s.
{Ωf,n |∀f,n}
Since there is no constraint on the cache nodes’ buffer capacity and transmission resources,
12
the above optimization problem can be further decoupled into the following sub-problems with
respect to each file.
Problem 2 (Optimization on the f -th File).
min
g f ({Ωf,n |∀n})
s.t.
Rf,n,s ≥ RF /NS , ∀n, s.
{Ωf,n |∀n}
(5)
B. Optimal Control Policy
Note that Problem 2 is a dynamic programming problem with random number of stages, which
cannot be solved by the standard approach as [26]. We shall shown in this section that the optimal
solution of Problem 2 (i.e. policy iteration) can be obtained by applying value iteration on another
MDP problem (Problem 3 as follows) with finite and fixed number of stages first, then solving
a revised version of Bellman’s equation. First of all, we introduce the following MDP problem
with fixed number of stages.
Problem 3 (Optimization with Fixed Stage Number).
min
{Ωf,n |∀n}
s.t.
NR X
NS
X
Eη,ρ [
gf,n,s ]
n=1 s=1
Rf,n,s ≥ RF /NS , ∀n, s.
where NR is the fixed number of requests on the f -th file.
As introduced in [27], there is standard solution for the MDP problem with finite and fixed
number of stages. The optimal solution of Problem 3 can be deduced via the following Bellman’s
equation
VNR −n+1 (Sf,n ) = min
X
Ωf,n (Sf,n )
s
gf,n,s (Ωf,n )+
X
VNR −n (Sf,n+1 )P r(Sf,n+1 |Sf,n , Ωf,n ) , ∀Sf,n , (6)
Sf,n+1
where VNR −n+1 (Sf,n ) is usually named as value function of the n-th stage, and Sf,n+1 denotes
the next state of the f -th file given the current state Sf,n . According to the definition of value
function in finite-stage dynamic programming, VNR −n+1 (Sf,n ) is the average remaining cost of
the f -th file from the n-th transmission to the NR -th transmission, given the system state of the
n-th stage Sf,n .
13
Note that the large-scale fading is i.i.d. in each file transmission, the expectation on large-scale
fading can be taken on both side of the above equation. Hence we have the following conclusion.
Lemma 1 (Bellman’s Equation with Reduced Space). The optimal control policy of Problem 3
is the solution of the Bellman’s equation with reduced state space as follows
X
X
e
e
e
e
e
VNR −n+1 (Sf,n ) = min Eη,ρ
gf,n,s (Ωf,n )+
VNR −n (Sf,n+1 )P r(Sf,n+1 |Sf,n , Ωf,n ) ,
ef,n )
Ωf,n (S
s
(7)
ef,n+1
S
c
where Sef,n = {Bf,s
∈ Sf,n |∀c, s}, VeNR −n (Sef,n ) = Eη,ρ [VNR −n (Sf,n+1 )], and Ωf,n (Sef,n ) =
c
{Ωf,n (Sf,n )|∀ρf,n , ηf,n,s , ηf,n,s
}.
Proof. Please refer to Appendix A.
The standard iterative value iteration can be used to solve the Bellman’s equation (7), and
obtain the value functions VeNR −n+1 (Sef,n ) (∀n, Sef,n ). Since this is not the contribution of this
paper, the detail procedure of value iteration is omitted here due to limited space. In the following
lemma, we show that the optimization of Problem 2 can also be solved given the value functions
VeN −n+1 (Sef,n ) (∀n, Sef,n ).
R
Lemma 2 (Optimal Control Policy of Problem 2). With the value function VeNR −n+1 (Sef,n ) (∀n),
the optimal control policy for Problem 2, denoted as Ω∗ (Sef,n , Tf,n ), can be calculated as
f,n
Ω∗f,n (Sef,n , Tf,n ) = arg
min
ef,n )
Ωf,n (S
X
Eη,ρ
X
gf,n,s (Ωf,n ) +
s
(λf Tf,n )N −λf Tf,n e e
e
e
VN (Sf,n+1 )P r(Sf,n+1 |Sf,n , Ωf,n ) ,
N!
(8)
ef,n+1
N,S
where Sef,n+1 is the next state of the f -th file given the current state Sf,n .
Proof. Note that (1) VeN (Sef,n+1 ) denotes the averaged cost for N times of requests on the f -th
(λ T )N −λf Tf,n
file, give the system state Sef,n+1 ; (2) f Nf,n
e
is the probability that there are N times
!
of file requests within the duration Tf,n . The conclusion is straightforward since the average cost
of the BS, given the current state Sef,n and remaining lifetime Tf,n , is
Eη,ρ
X
s
gf,n,s (Ωf,n ) +
X
ef,n+1
N,S
(λf Tf,n )N −λf Tf,n e e
e
VN (Sf,n+1 )P r(Sef,n+1 |Sf,n , Ωf,n ) .
N!
14
As a summary, the optimal control policy of the Problem 2 can be solved via the following
two steps.
•
Value Iteration: Calculate the value function VeNR −n+1 (Sef,n )(∀n, Sef,n ) via Bellman’s equation with reduced space (7).
•
Policy Iteration: Obtain optimal control policy from (8).
Note that the state space of the system is actually huge. For example, suppose there are 20 cache
nodes in the system and 10 segments per file, the dimensional of system state is 2200 , which is
referred to as the curse of dimensionality in dynamic programming problem. Thus although the
optimal solution algorithm can be provided, the computation complexity is intractable. Hence,
we continue to propose a low-complexity solution based on the technique of approximated MDP
and reinforcement learning in the following section.
IV. L OW-C OMPLEXITY S OLUTION VIA A PPROXIMATED MDP
In this section, we shall introduce a novel linear approximation approach on the value function
VeNR −n+1 (Sef,n ), and elaborate the on-line control policy to determine the control actions given the
current system state and approximated value function. In order to evaluate the approximated value
function, we first provide analytical expressions given the distribution of requesting users (e.g.
uniform distribution); then an online learning algorithm is introduced for unknown distribution
statistics of requesting users.
A. Approximation on Value Function
We first define the notations for following reference system states.
•
•
Sef∗ is the state of f -th file where all the cache nodes have successfully decoded the whole
∗
c
file. Thus Sef,n
= {Bf,s
= 1|∀c, s}.
Sei,s is the state of f -th file transmission where only the s-th segment at the i-th cache node
f
j
i
is not successfully decoded. Sefi,s = {Bf,s
= 0, Bf,t
= 1|∀(j, t) 6= (i, s)}.
Hence, we approximate the value function VeNR −n+1 (Sef,n ) linearly as
X
∗
VeN −n+1 (Sef,n ) ≈ VeN −n+1 (Sef ) +
VeN −n+1 (Sei,s ) − VeN
R
R
R
f
(Sef∗ )
R −n+1
,
(9)
i (S
ef,n )=0}
{(i,s)|∀Bf,s
i
i
where Bf,s
(Sef,n ) means the parameter of Bf,s
in the system state Sef,n . An example of approxi-
mated value function is elaborated below.
15
T
1
2
≈
Node 1
1
2
Node 2
S*
S 1,2
1
1
2
Node 1
1
2
Node 1
2
1
Node 2
2
Node 2
: Decoded
S*
1
2
Node 1
1
S 2,2
S*
1
1
2
Node 1
2
1
Node 2
2
Node 1
1
Node 2
2
2
Node 2
: Not Decoded
Fig. 2. One example of approximated value function.
Example 2 (Approximated Value Function). An example on the approximated value function is
illustrated in Fig. 2, where there are two cache nodes and downlink file (say the f -th file) is
1
1
2
2
divided into two segments. For the system state Te = [Bf,1
, Bf,2
, Bf,1
, Bf,2
] = [1, 0, 1, 0], the value
function on the n-th stage can be approximated as
∗
VeN −n+1 (Te) ≈ VeN −n+1 (Se ) + VeN
(Sef1,2 )
R −n+1
VeNR −n+1 (Sef∗ )
−
2,2
∗
e
e
e
e
+ VNR −n+1 (Sf ) − VNR −n+1 (Sf ) .
R
R
f
In the right hand side of the above approximation, the first term counts the transmission cost
for the users outside the coverage of cache nodes; the second term approximates the cost on
the second segment transmission to the users within the coverage of the first cache node; and
the third term approximates the cost on the second segment transmission to the users within the
coverage of the second cache node. Note that there is no transmission cost on the first segment
for the users within the coverage of both cache nodes.
In order to apply this approximation on all value function, it is necessary to obtain the value
of VeN −n+1 (Se∗ ) and VeN −n+1 (Sei,s ) for all n, i, and s via (7). In the following, we provide
R
f
f
R
the analytically expressions for them with the knowledge of the requesting users’ distribution.
Moreover, in the next section, an online learning algorithm is proposed for the evaluation of
VeNR −n+1 (Sef∗ ) and VeNR −n+1 (Sefi,s ) with unknown spatial distribution of requesting users.
1) Evaluation of VeN −n+1 (Se∗ ): Note that system state Se∗ represents the situation that all the
R
f
f
cache nodes have already decoded the f -th file, the purpose of downlink transmission is only
to make sure that the requesting users, which are outside of the coverage of cache nodes, can
16
decode the downlink file. Hence it is clear that
VeNR −n+1 (Sef∗ )
s
)
(NR − n + 1) Pr(lf,n ∈
/ Cf,n
X
s
Eρ,η
min we Pf,n,s Nf,n,s + wt Nf,n,s lf,n ∈
/ Cf,n
=
s
Pf,n,s
Nf,n,s
s.t. Rf,n,s ≥ RF /NS , ∀s.
The above value function can be calculated with analytical expression, which is elaborated below.
Lemma 3. The value function VeNR −n+1 (Sef∗ ) is given by
X
∗
s
∗
∗
∗
s
VeNR −n+1 (Sef ) ≈ (NR − n + 1) Pr(lf,n ∈
/ Cf,n )Eρ,η
(we Pf,n,s Nr,f,s + wt Nf,n,s ) lf,n ∈
/ Cf,n ,
s
where
∗
Pf,n,s
=
wt
we W(
2θ wt
)
ewe
∗
, Nf,n,s
=
RF
,
∗
NS [θ+log2 (Pf,n,s
)]
h
θ = Ehf,n,s log2
||hf,n,s ||2
NT σz2
i
, and W(x) is
the Lambert-W function [28].
Proof. Please refer to Appendix B.
2) Evaluation of VeNR −n+1 (Sefi,s ): Given system state Sefi,s for arbitrary stage, there are only
two possible next system states Sefi,s and Sef∗ , which are discussed below.
•
i
i
When ρf,n ηf,n,s ≤ ρi ηf,n,s
, thus Rf,n,s ≤ Rf,n,s
, the i-th cache node is alway able to decode
the s-th file segment give that the transmission constraint (5) should be satisfied. Thus the
next state must be Se∗ . In this case, the optimized RHS of (7) is given by
f
Qc (Sefi,s )
=
mini,s
e )
Ωf,n (S
f
s.t. Rf,n,t =
•
X
gf,n,t (Sefi,s , Ωf,n ) + VeNR −n (Sef∗ ),
(10)
t
RF
, ∀t.
NS
i
i
When ρf,n ηf,n,s > ρi ηf,n,s
, thus Rf,n,s > Rf,n,s
, the BS can choose to secure the transmission
of the s-th segment to the (f, n)-th user or the i-th cache node. Hence the optimized RHS
of (7) is given by
i,s
i,s
e
e
E min Qu (Sf ), Qi,s (Sf ) ,
where Qu and Qi,s are defined below.
Qu (Sefi,s ) =
mini,s
e )
Ωf (S
f
X
t
RF
gf,n,t (Sefi,s , Ωf,n ) + VeNR −n (Sefi,s ), s.t. Rf,n,t =
, ∀t (11)
NS
17
Qi,s (Sefi,s )
=
mini,s
X
e )
Ωf (S
f
i
s.t. Rf,n,s
=
gf,n,t (Sefi,s , Ωf ) + VeNR −n (Sef∗ )
(12)
t
RF
RF
and Rf,n,t =
,∀t 6= s.
NS
NS
As a result, the expression of VeNR −n+1 (Sefi,s ) is summarized by the following lemma.
Lemma 4. The value function VeNR −n+1 (Sefi,s ) is given below
i
i
)+
]P r(Rf,n,s ≤ Rf,n,s
VeNR −n+1 (Sefi,s ) = Eη,ρ [Qc (Sefi,s )|Rf,n,s ≤ Rf,n,s
i
i
]P r(Rf,n,s > Rf,n,s
). (13)
Eη,ρ [min{Qu (Sefi,s ), Qi,s (Sefi,s )}|Rf,n,s > Rf,n,s
Moreover, the optimal control actions for Qc and Qu are the same as Lemma 3. The optimal
control action for Qi,s is given by
∗
Pf,n,s
=
wt
θi
we W( 2ewwe t )
∗
, Nf,n,s
"
RF
=
, θi = Ehf,n,s log2
NS [θi + log2 (Pf,n,s )]
||hif,n,s ||2
NT σz2
!#
and ∀t 6= s
∗
Pf,n,t
wt
RF
||hf,n,t ||2
∗
=
, Nf,n,t =
, θ = Ehf,n,t log2
.
θ
∗
NS [θ + log2 (Pf,n,s
)]
NT σz2
we W( 2ewwet )
Proof. The proof is similar to that of Lemma 3, and it is omitted here.
With the distribution knowledge of large-scale fading, the value functions VeNR −n+1 (Sef∗ ) and
VeNR −n+1 (Sefi,s ) can be calculated according to above analytical expressions. Note that these value
functions are the same for different files, they can be applied on all the files’ transmission once
calculated. Thus the complexity is low.
B. Low-Complexity Downlink Multicast Policy
With the value function VeNR −n+1 (Sef∗ ) and VeNR −n+1 (Sefi,s ), the value function for arbitrary
system state in arbitrary transmission stage can be approximated via (9). Hence the online
control action for arbitrary system state Sf,n , denoted as Ω∗f,n (Sf,n ), can be obtained by the
following optimization problem.
18
Problem 4 (Online Optimization).
Ω∗f,n (Sf,n ) = arg min
X
gf,n,s (Sf,n , Ωf,n ) +
s
s.t.
X (λf Tf,n )N
N!
N
e−λf Tf,n VeN (Sef,n+1 )
Rf,n,s ≥ RF /NS , ∀s.
Since the value function VeN (Sef,n+1 ) is approximated by (9), the optimization in Problem 4
can be further decoupled for each segment. For the s-th segment (∀s), the solution of Problem 4
can be obtained by the following problem, given that the requesting user cannot find the segment
from nearby cache nodes.
Problem 5 (Online Optimization for the s-th Segment).
∗
∗
{Pf,n,s
, Nf,n,s
}
= arg min gf,n,s (Sf,n , Ωf,n ) +
X (λf Tf,n )N
N
N!
X
e−λf Tf,n
VeN (Sefi,s ) − VeN (Sef∗ )
i (S
ef,n+1 )=0}
{i|∀Bf,s
Rf,n,s ≥ RF /NS ,
s.t.
i
where Bf,s
(Sef,n+1 ) represents the buffer status for the (f, s)-th segment in the i-th cache node.
Due to the second term of objective in Problem 5, the BS should first choose the cache
nodes for downlink receiving, in addition to the requesting user. Based on the selection, the
optimal power and transmission time can be derived. This is an integrated continuous and discrete
optimization, its solution algorithm is summarized below.
Algorithm 1 (Online Scheduling with Approximated Value Function). Given the system state
Sf,n , let d1 , d2 , .. be the indexes of cache nodes, whose large-scale attenuation to the BS in the
s-th segment is worse than the (f, n)-th user. Moreover, without loss of generality, it is assumed
d1
d2
that ρd1 ηf,n,s
≤ ρd2 ηf,n,s
≤ ... ≤ ρf,n,s ηf,n,s . The optimal control action for the s-th segment (∀s)
can be obtained below.
•
For each i, suppose the di -th cache node are involved for downlink receiving, the optimal
power and transmission time control is given by
Q∗di ,s (Sf,n )
= min gf,n,s (Sf,n , Ωf,n ) +
s.t.
Pf,n,s
Nf,n,s
di
Rf,n,s
= RF /NS .
X (λf Tf,n )N
N
N!
−λf Tf,n
e
X
j,s
∗
e
e
e
e
VN (Sf ) − VN (Sf )
j={d1 ,...,di−1 }
19
di
di
The optimal solution, denoted as [Pf,n,s
, Nf,n,s
], can be derived similar to Lemma 3.
•
∗
∗
Let d∗ = arg min Q∗di ,s , the solution of Problem 5 is then given by [Pf,n,s
, Nf,n,s
] =
∗
∗
di
d
d
[Pf,n,s
, Nf,n,s
].
C. Learning Algorithm for Approximated Value Function
In Section IV-A, the values of VeNR −n+1 (Sef∗ ) and VeNR −n+1 (Sefi,s ) are evaluated analytically
by assuming the available distribution knowledge of requesting users. However in practice, the
statistics of the distribution may be unknown to the BS. In order to address this issue, we
introduce in the following a learning-based online algorithm to evaluate the value functions
VeN −n+1 (Se∗ ) and VeN −n+1 (Sei,s ).
R
f
R
f
Algorithm 2 (Reinforcement Learning for Value Function). Suppose requesting users for all
files are distributed i.i.d. according to certain spatial distribution with unknown statistics. When
one user becomes active to request one file, its location and channel information can be used to
update all the value functions. Note that VeN −n+1 (Se∗ ) and VeN −n+1 (Sei,s ) (∀n, i, s) are identical
R
f
R
f
for all files, we only elaborate the update on the f -th file’s value functions. Specifically, the
learning procedure is described below.
•
Step 1: Let t = 0. Initialize the value of VeNR −n+1 (Sef∗ ) and VeNR −n+1 (Sefi,s ) (∀n, i, s), and
denote them as VeNt R −n+1 (Sef∗ ) and VeNt R −n+1 (Sefi,s ). This initialization can be done by assuming
all the users appear uniformly in the cell coverage, hence the approach in Section IV-A can
be applied.
•
Step 2: t = t + 1.
•
Step 3: When one new user appears to request any downlink file, update the value functions
with its location and channel information. Suppose it is the m-th request on the g-th file,
its location is defined as lm,g , we have
VeNt R −n+1 (Sef∗ ) =
t e t−1
(Se∗ ) +
V
t + 1 NR −n+1 f
X
1
∗
∗
∗
),
(NR − n + 1)I(lg,m ∈
/ C)
(we Pg,m,s
Ng,m,s
+ wt Ng,m,s
t+1
s
20
where
∗
Pg,m,s
=
wt
we W(
2θt wt
)
ewe
VeNt R −n+1 (Sefi,s ) =
•
∗
, Ng,m,s
=
RF
,
∗
NS [θt +log2 (Pf,n,s
)]
h
θt = Ehg,m,s log2
||hg,m,s ||2
NT σz2
i
.
t e t−1
1
i,s
i
e
)
Qc (Segi,s )I(Rg,m,s ≤ Rg,m,s
V
(S ) +
t + 1 NR −n+1 f
t+1
i,s
i,s
i
+ min{Qu (Seg ), Qi,s (Seg )}I(Rg,m,s > Rg,m,s ) ,
where Qc , Qu and Qi,s are defined in (10), (11) and (12) respectively.
(Sef∗ )| ∀n, i, s}
Step 4: If max{|VeNt R −n+1 (Sefi,s ) − VeNt−1
(Sefi,s )|, |VeNt R −n+1 (Sef∗ ) − VeNt−1
R −n+1
R −n+1
is greater than one threshold τ , the algorithm goes to Step 2; otherwise, the algorithm
terminates.
Moreover, we have the following lemma on the convergence of above iterative learning
algorithm.
Lemma 5. The Algorithm 2 will converge to the true value of VeNR −n+1 (Sef∗ ) and VeNR −n+1 (Sefi,s )
(∀f, n, i, s). Thus
lim VeNt R −n+1 (Sef∗ ) = VeNR −n+1 (Sef∗ ),
t→+∞
lim VeNt R −n+1 (Sefi,s ) = VeNR −n+1 (Sefi,s ).
t→+∞
Proof. Please refer to Appendix C.
D. Bounds on Approximated Value function
Since the control policy is derived according to approximated value function in Section IV-B,
it is necessary to know the bound on the approximation error. Hence we introduce the following
conclusion.
Lemma 6 (Bounds of Value Functions). With arbitrary spatial distribution of requesting users,
the upper-bound in (14) holds for value function VeN −n+1 (Sef,n ) (∀n). Moreover, if there is no
R
overlap in the service region of cache nodes, the lower-bound in (15) also holds for VeNR −n+1 (Sef,n ).
∗
VeNR −n+1 (Sef,n ) ≤ VeNR −n+1 (Sef,n
)+
X
ef,n )=0}
{(i,s)|∀Bif,s (S
i,s
∗
e
e
e
e
VNR −n+1 (Sf ) − VNR −n+1 (Sf,n ) (14)
21
VeNR −n+1 (Sef,n ) ≥
∗
VeNR −n+1 (Sef,n
)
X
+
i,s
∗
Ve1 (Sef ) − Ve1 (Sef,n )
(15)
ef,n )=0}
{(i,s)|∀Bif,s (S
Proof. Please refer to Appendix D.
Notice that the proposed linear approximation on value function is actually the upper-bound in
(14), the gap between the approximated value function and the actual value function VeN −n+1 (Sef,n ),
R
denoted as ENR −n+1 (Sef,n ), is given by
X
i,s
i,s
∗
∗
e
e
e
e
e
e
e
e
e
ENR −n+1 (Sf,n ) ≤
VNR −n+1 (Sf ) − VNR −n+1 (Sf ) − V1 (Sf ) + V1 (Sf ) .
i (S
ef,n )=0}
{(i,s)|∀Bf,s
According to the definition of value function, the average system cost on the f -th file with
optimal control can be written as
g ∗f =
X (λf T )NR
NR
NR !
e−λf T VeNR (Sef0 ),
(16)
where Sef0 denotes the system state with empty buffer in all cache nodes. Hence the upper and
lower bounds on g ∗ can be obtained by applying (14) and (15) on VeN (Se0 ) in (16).
f
R
f
V. P ROACTIVE C ACHING P LACEMENT
In the above sections, the downlink transmission happens only when users raise requests on
file transmission. This cannot exploit the channel temporal diversity on the shadowing effect.
For example, when the channels from the BS to some cache nodes are good, the downlink
resource consumption on caching placement can be low, and the proactive file transmission to
these cache nodes may further save the overall downlink cost. Hence, in this section, we consider
the scenario that the BS can deliver one file segment to cache nodes per every Tp seconds. It
is assumed that the transmission of one file segment can be completed within the coherent time
of shadowing attenuation. Thus we define the following control policy.
Definition 3 (Proactive Caching Placement Policy). In the k-th proactive transmission opportuc
nity, given the buffer status of each cache node {Bf,s
|∀c, f, s}, the shadowing from the BS to each
cache nodes {ηkc |∀c}, and the remaining lifetime of each file {Tfk |∀f }, the BS should determine
the selected file segment (Fk , Sk ) and the downlink transmission parameters Pk and Nk for
22
c
the selected (Fk , Sk )-th file segment. Thus denote S k = {Bf,s
|∀c, f, s}, {ηkc |∀c}, {Tfk |∀f } , the
placement policy can be written as
Ωk (S k ) = [Fk , Sk , Pk , Nk ].
Suppose before the k-th proactive transmission opportunity, the buffer status of cache nodes
c
is Sefk = {Bf,s
|∀c, f, s}. The average cost of the BS for the f -th file without proactive caching
placement is given by
g kf (Sefk , Tfk ) =
X (λf Tfk )N
N
N!
k
e−λf Tf VeN (Sefk ),
where VeN (Sefk ) can be approximated linearly via (9). In each proactive caching placement opportunity, the BS can choose one file segment which is expected to mostly suppress the expected
cost of the BS. As a result, we have the following algorithm on the proactive caching placement.
Algorithm 3 (Proactive Caching Placement). On each proactive transmission opportunity (say
the k-th opportunity), the algorithm to determine the proactive caching placement policy Ωk is
elaborated below.
•
Step 1: For each file segment (say the (f, s)-th one), evaluate the expected cost saving if
the BS choose deliver this segment proactively by solving the following Problem 6.
•
Step 2: The (Fk , Sk )-th segment is chosen when the following two conditions are satisfied:
k
k
– (Fk , Sk ) = arg max ∆gf,s
, where ∆gf,s
is defined in Problem 6.
(f,s)
–
∆gFk k ,Sk
0
0
≥ τ , where τ > 1 is certain threshold.
Problem 6 (Proactive placement for (f, s)-th Segment).
k
∆gf,s
=
max
k ,N k ,Ak
Pf,s
f,s
f
g kf (Sefk , Tfk )/
h
k
k
we Pf,s
Nf,s
+
k
wt Nf,s
+
i
g kf (S̆fk (Akf ), Tfk )
s.t. Rkc ≥ RF /NS , ∀c ∈ Akf ,
k
k
where Pf,s
, Nf,s
and Akf represents the transmission power, transmission time and the set of
multicast cache nodes, S̆fk (Akf ) denotes the cache buffer status where the cache nodes in Akf
have successfully decoded the (f, s)-th segment given the previous status Sek .
f
Algorithm 4 (Solution Algorithm for Problem 6). Given the system state Sefk and the remaining
lifetime T k , let d1 , d2 , ...(∀di , B di (Sek ) = 0) be the indexes of cache nodes satisfying ρd1 η d1 ≤
f
f,s
f
k
23
200
220
200
Baseline1
Baseline2
Proposed(MDP)
180
160
160
140
Averge Cost
Average Cost
Baseline1
Baseline2
Proposed(MDP)
180
140
120
100
120
100
80
80
60
60
40
40
20
20
1
2
3
4
5
6
7
8
1
2
3
Expectation of Request Number
4
5
6
7
8
Expectation of Request Number
(a) NC = 20
(b) NC = 25
Fig. 3. The average total cost versus the average times of request in file lifetime, where the number of cache nodes is 20 and
25 respectively.
k
ρd2 ηkd2 ≤ · · · . Then ∆gf,s
can be obtained below.
•
Step 1: For each i, suppose the di -th cache node are involved for downlink receiving, the
optimal power and transmission time control is given by
Q∗di ,s
= min
k,di
f,s
k,di
N
f,s
k,di k,di
k,di
we Pf,s
Nf,s +wt Nf,s
+
P
X (λf Tfk )N
N
N!
e
−λf Tfk
X
j,s
∗
e
e
e
e
VN (Sf )−VN (Sf ) ,
j={d1 ,...,di−1 }
subject to Rkdi = RF /NS . The optimal solution can be derived similar to Lemma 3.
•
k,di∗
k
= Pf,s
,
Step 2: Let di∗ = arg min Q∗di ,s , the solution of Problem 6 is given by Pf,s
di
k
Nf,s
=
k,di∗
Nf,s
,
and
Akf
= {di |∀i ≥ i∗ }.
VI. S IMULATION
In the simulation, the cell radiusis 500 meters, cache nodes are randomly deployed on the
cell-edge region with a service radius of 90 meters. The number of antennas at the BS is 8. The
downlink path loss exponent is 3.5. The standard derivation of shadow fading in 6dB. The file
size is 140Mb, and is further divided into 10 segments. The transmission bandwidth is 20MHz.
The weights on transmission energy and time are we = 1 and wt = 100 respectively. The
performance of the proposed algorithm will be compared with the following two baselines.
Baseline 1. The BS only ensures the segment delivery to the requesting users in each transmission. The cache nodes with better channel condition to the BS can also decode the file segments.
24
300
220
250
Approximated Value Function
Accurate Value Function
Lower Bound
200
Approximated Value Function
Accurate Value Function
Lower Bound
180
160
200
Value
Value
140
150
120
100
100
80
60
50
40
0
20
1
2
3
4
5
6
7
8
1
Total Transmission Times NR
i
(a) Se = {Bf,s
= 0, ∀f, s, i}
2
3
4
5
6
7
8
Total Transmission Times NR
i
i
(b) Se = {Bf,s
= 0, ∀f, s, i = 1, 2, ..., 10} ∪ {Bf,s
=
1, ∀f, s, i = 11, 12, ..., 20}
Fig. 4. Illustration of value function and its bounds, where NC = 20.
Baseline 2. The BS ensures that all the cache nodes can decode the downlink file in the first
transmission. Hence, all the cache nodes can help to forward the file since the second file request.
The performance of the proposed low-complexity algorithm (Algorithm 1) is compared with
the above two baselines in Fig.3. In the simulation, the number of cache nodes is 20 and
25 respectively, and the requesting users are uniformly distributed in the cell coverage, and
distribution statistics is known to the BS. Hence, the analytical expressions derived in Section
IV-A can be used to calculate the approximated value functions. It can be observed that the
proposed Algorithm 1 is superior to the two baselines for any expected number of requests
per file lifetime. Moreover, the Baseline 1 has better performance than Baseline 2 when the
popularity of the file is high (larger expected number of file requests).
The approximation error of the value function versus different stages is illustrated in Fig.4,
where the true value function and the bounds derived in Lemma 6 are plotted. Fig. 4(a) shows
the state where all the cache nodes are empty, and Fig. 4(b) shows the state where half of cache
nodes have decoded the whole file. It is shown that for both states, both upper and lower bounds
are tight, and therefore approximation error is small.
In Fig. 5, there are 3 and 4 hot zones in the cell coverage, each with radius 90 m. The
statistics of users’ distribution is unknown to the BS. The probability that the user appears on
the one hot zone is 12.5% (larger than the other regions). The performance of two baselines, the
proposed Algorithm 1 assuming users are uniformly distributed, and the proposed Algorithm 1
25
160
160
Baseline1
Baseline2
MDP without Learning
Reinforcement Learning
140
120
Average Cost
120
Average Cost
Baseline1
Baseline2
MDP without Learning
Reinforcement Learning
140
100
80
100
80
60
60
40
40
20
20
1
2
3
4
5
6
7
8
1
2
3
Expectation of Request Number
4
5
6
7
8
Expectation of Request Number
(a) 3 hot zones
(b) 4 hot zones
Fig. 5. The average total cost versus the expectation of request times, where exits 3 hot zones in the cell.
1600
Baseline1
Baseline2
MDP without Learning
Reinforcement Learning
Proactive Caching Placement
1400
Average Cost
1200
1000
800
600
400
200
1
2
3
4
5
6
7
8
Expectation of Request Number
Fig. 6. The Performance of Proactive Caching Placement
with learning-based evaluation of value functions (Algorithm 2) are compared. It can be observed
that the proposed learning algorithm has the best performance. Moreover, the performance gain
of the learning-based algorithm is larger for more hot zones.
Finally, the performance of proactive caching placement is demonstrated in Fig. 6, where
there are 10 files, 3 hot zones in the cell, 50000 times of proactive transmission opportunities in
the file’s lifetime. The performance of the proactive caching placement algorithm is compared
with the above two baselines, Algorithm 1 assuming uniform user distribution, and Algorithm 1
with learning-based evaluation of value functions. It can be observed that the proposed proactive
26
caching placement algorithm can further improve the offloading performance, compared with
the algorithms without it.
VII. C ONCLUSION
We consider the downlink file transmission with the assistance of cache nodes in this paper. The
number of requests of one file within its lifetime is modeled as a Poisson point process, and the
downlink resource minimization problem without proactive caching placement can be formulated
as a MDP with random number of stages. We first propose a revised Bellman’s equation, where
the optimal control policy can be derived. In order to address the curse of dimensionality, we
also introduce a low-complexity sub-optimal solution based on linear approximation of value
functions. The approximated value function can be calculated analytically with the knowledge of
distribution statistics of users. In practice, the statistics of the distribution may be unknown to the
BS. We continue to propose a learning-based online algorithm to evaluate the value functions.
Furthermore, we derive a bound on the gap between the approximated value functions and the
real value functions. Finally, we propose a proactive caching placement algorithm, which can
exploit the channel temporal diversity of shadowing effect. It is shown by numerical simulation
that the proposed algorithms can significantly reduce the resource consumption at the BS, and
the proactive caching placement algorithm can further improve the performance.
A PPENDIX A: P ROOF O F L EMMA 1
Let VeNR −n (Sef,n ) = Eη,ρ [VNR −n (Sf,n+1 )], where the expectation is taken over the randomness
of shadowing and requesting users’ pathloss, we have
X
X
VNR −n (Sf,n+1 )P r(Sf,n+1 |Sf,n , Ωf,n ) =
Sf,n+1
VeNR −n (Sef,n+1 )P r(Sef,n+1 |Sf,n , Ωf )
ef,n+1
S
Taking expectation with respect to the shadowing and pathloss on (6), we have
X
X
e
e
VNR −n+1 (Sf,n ) = Eη,ρ min
gf,n,s (Sf,n , Ωf,n ) +
VNR −n (Sf,n+1 )P r(Sf,n+1 |Sf,n , Ωf,n )
Ω(Sf,n )
= min Eη,ρ
ef,n )
Ω(S
s
X
s
Sf,n+1
gf,n,s (Sf,n , Ωf,n ) +
X
ef,n+1
S
e
e
e
VNR −n (Sf,n+1 )P r(Sf,n+1 |Sf,n , Ωf ) .
27
A PPENDIX B: P ROOF O F L EMMA 3
First of all, we have the following high SNR (signal-to-noise ratio) approximation on the
throughput Rf,n,s .
Rf,n,s ≈ Nf,n,s Ehf,n,s log2
With Rf,n,s =
RF
,
NS
we have Nf,n,s =
||hf,n,s ||2 Pf,n,s
NT σz2
RF
.
NS [θ+log2 (Pf,n,s )]
min
Pf,n,s
= Nf,n,s [θ + log2 (Pf,n,s )].
Hence the original optimization becomes
RF (we Pf,n,s + wt )
NS [θ + log2 (Pf,n,s )]
∗
Taking first-order derivative on f (Pf,n,s ), the optimal transmission power Pf,n,s
can be obtained.
A PPENDIX D: P ROOF O F L EMMA 5
We only prove the convergence of VeNt R −n+1 (Sef∗ ), and the convergence of VeNt R −n+1 (Sefi,s ) can
be applied similarly. Let
εt = VeNR −n+1 (Sef∗ ) − (NR − n + 1)I(lg,m ∈
/ C)
X
∗
∗
∗
(we Pg,m,s
Ng,m,s
+ wt Ng,m,s
)
s
denotes the estimate error in t-th iteration. It is clear that the estimation errors are i.i.d. with
respect to t, E[εt ] = 0 and V ar[εt ] < +∞. Note that VeNt R −n+1 (Sef∗ ) can be written as
VeNt R −n+1 (Sef∗ ) =
t e
X
VNR −n+1 (Sef∗ ) − εi
t+1
i=0
where the total estimate error is
Pt
εi
i=0 t+1 .
= VeNR −n+1 (Sef∗ ) −
t
X
εi
,
t
+
1
i=0
The mean and variance of total estimate error are
summarized below:
X
t
εi
= 0,
E
t
+
1
i=0
X
t
εi
V ar[εi ]
V ar
=
.
t
+
1
t
+
1
i=0
When t → +∞, the variance of estimation error tends to zero, and VeNt R −n+1 (Sef∗ ) converges to
VeN −n+1 (Se∗ ).
R
f
28
A PPENDIX E: P ROOF O F L EMMA 6
1) Proof of Upper Bound: The approach of mathematical induction will be used in the proof.
Thus without loss of generality, we shall assume that the upper bound holds when the first l-th
cache nodes have not decoded the (f, s)-th segment, and prove that the upper bound also holds
when the first (l + 1)-th cache nodes have not decoded the (f, s)-th segment. Define the system
ff c,s = [B i = 1, ∀j 6= s, ∀i] ∪ [B i = 0, ∀i = 1, 2, · · · , c] ∪ [B i = 1, ∀i > c]
state T
f,j
•
f,s
Step 1: When c = 1, the upper bound holds as follows
c,s
∗
ff ) = VeN −n+1 (Se ) + VeN −n+1 (Sec,s ) − VeN
VeN −n+1 (T
R
•
f,s
f
R
f
R
(Sef∗ )
R −n+1
Step 2: Suppose the following bound holds for c = l
X
l,s
c,j
∗
∗
e
f
e
e
e
e
e
e
VNR −n+1 (Tf ) ≤ VNR −n+1 (Sf ) +
VNR −n+1 (Sf ) − VNR −n+1 (Sf )
(17)
j=1,2,··· ,l
•
Step 3: When c = l + 1, we can apply the following sub-optimal control policy: (1) if the
requesting users appear in the coverage of C1 ∪ C2 ∪ ... ∪ Cl , the optimal control policy for
ff l,s is applied; (2) if the requesting users appear in the coverage of Cl+1 , the
system state T
optimal control policy for system state Sefl+1,s is applied; (3) if the requesting users appear
outside the coverage of any cache nodes, choose the one from the above two policies with
ff l+1,s ) be the average cost of the
larger transmission resource consumption. Let V̆N −n+1 (T
R
above sub-optimal control policy, we have
ff
VeNR −n+1 (T
l+1,s
ff
) ≤ V̆NR −n+1 (T
l+1,s
c,l+1
∗
ff )+ VeN −n+1 (Se
) ≤ VeNR −n+1 (T
)−VeNR −n+1 (Sef ) .
R
f
l,s
Although the above proof is for the (f, s)-th file segment, it can be trivially extended to
arbitrary file segments. Thus the upper bound is proved.
2) Proof of Lower Bound: Let Ω∗f,n be the optimal control policy, we have
VeNR −n+1 (Sef,n ) − VeNR −n+1 (Sef∗ )
X
X
∗
∗
e
e
e
e
= Eη,ρ
gf,n,s (Sf,n , Ωf,n ) + Eη,ρ
VNR −n (Sf,n+1 )P r(Sf,n+1 |Sf,n , Ωf,n )
s
−Eη,ρ
X
s
ef,n+1
S
gf,n,s (Sef∗ , Ω∗f,n ) − VeNR −n (Sef∗ ).
29
∗
e
s gf,n,s (Sf,n , Ωf,n )
P
∗
e
s gf,NR ,s (Sf,n , Ωf,NR )
P
Because Eη,ρ
≥ Eη,ρ
P
P
e∗ ∗
e∗ ∗
and Eη,ρ
= Eη,ρ
s gf,n,s (Sf , Ωf,n )
s gf,NR ,s (Sf , Ωf,NR ) , We have
Eη,ρ
X
X
∗
∗
∗
e
e
gf,n,s (Sf,n , Ωf,n ) − Eη,ρ
gf,n,s (Sf , Ωf )
s
≥ Eη,ρ
s
X
X
∗
∗
∗
e
e
gf,NR ,s (Sf,n , Ωf,NR ) − Eη,ρ
gf,NR ,s (Sf , Ωf,NR )
s
X
=
s
i,s
∗
Ve1 (Sef ) − Ve1 (Sef ) .
ef,n )=0}
{(i,s)|∀Bif,s (S
We also have
Eη,ρ
X
∗
e
e
e
VNR −n (Sf,n+1 )P r(Sf,n+1 |Sf,n , Ωf,n ) − VeNR −n (Sef∗ ) ≥ 0.
ef,n+1
S
As a result, the lower bound is straightforward.
R EFERENCES
[1] B. Lv, L. Huang, and R. Wang, “Cellular offloading via downlink cache placement,” in submitted to 2018 IEEE International
Conference on Communications (ICC), 2018.
[2] X. Wang, M. Chen, T. Taleb, A. Ksentini, and V. C. M. Leung, “Cache in the air: exploiting content caching and delivery
techniques for 5G systems,” IEEE Communications Magazine, vol. 52, no. 2, pp. 131–139, February 2014.
[3] E. Batu, M. Bennis, and M. Debbah, “Cache-enabled small cell networks: Modeling and tradeoffs,” in 2014 11th
International Symposium on Wireless Communications Systems (ISWCS), Aug 2014, pp. 649–653.
[4] A. Liu and V. K. N. Lau, “Mixed-timescale precoding and cache control in cached MIMO interference network,” IEEE
Transactions on Signal Processing, vol. 61, no. 24, pp. 6320–6332, Dec 2013.
[5] E. Bastug, M. Bennis, and M. Debbah, “Social and spatial proactive caching for mobile data offloading,” in 2014 IEEE
International Conference on Communications Workshops (ICC), June 2014, pp. 581–586.
[6] C. Yang, Y. Yao, Z. Chen, and B. Xia, “Analysis on cache-enabled wireless heterogeneous networks,” IEEE Transactions
on Wireless Communications, vol. 15, no. 1, pp. 131–145, Jan 2016.
[7] S. H. Chae and W. Choi, “Caching placement in stochastic wireless caching helper networks: Channel selection diversity
via caching,” IEEE Transactions on Wireless Communications, vol. 15, no. 10, pp. 6626–6637, Oct 2016.
[8] X. Li, X. Wang, and V. C. M. Leung, “Weighted network traffic offloading in cache-enabled heterogeneous networks,” in
2016 IEEE International Conference on Communications (ICC), May 2016, pp. 1–6.
[9] M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,” IEEE Transactions on Information Theory, vol. 60,
no. 5, pp. 2856–2867, May 2014.
[10] X. Xu and M. Tao, “Modeling, analysis, and optimization of coded caching in small-cell networks,” IEEE Transactions
on Communications, vol. 65, no. 8, pp. 3415–3428, Aug 2017.
[11] A. Checko, H. L. Christiansen, Y. Yan, L. Scolari, G. Kardaras, M. S. Berger, and L. Dittmann, “Cloud RAN for mobile
networks a technology overview,” IEEE Communications Surveys Tutorials, vol. 17, no. 1, pp. 405–426, Jan. 2015.
30
[12] J. Koh, O. Simeone, R. Tandon, and J. Kang, “Cloud-aided edge caching with wireless multicast fronthauling in fog radio
access networks,” in 2017 IEEE Wireless Communications and Networking Conference (WCNC), March 2017, pp. 1–6.
[13] M. Moghadari, E. Hossain, and L. B. Le, “Delay-optimal distributed scheduling in multi-user multi-relay cellular wireless
networks,” IEEE Transactions on Communications, vol. 61, no. 4, pp. 1349–1360, April 2013.
[14] Y. Cui and V. K. N. Lau, “Distributive stochastic learning for delay-optimal OFDMA power and subband allocation,” IEEE
Transactions on Signal Processing, vol. 58, no. 9, pp. 4848–4858, Sept 2010.
[15] Y. Cui and D. Jiang, “Analysis and optimization of caching and multicasting in large-scale cache-enabled heterogeneous
wireless networks,” IEEE Transactions on Wireless Communications, vol. 16, no. 1, pp. 250–264, Jan 2017.
[16] R. Wang and V. K. N. Lau, “Delay-aware two-hop cooperative relay communications via approximate MDP and stochastic
learning,” IEEE Transactions on Information Theory, vol. 59, no. 11, pp. 7645–7670, Nov 2013.
[17] W. A. Shewhart and S. S. Wilks, Approximate Dynamic Programming: Solving the Curses of Dimensionality, Second
Edition, 2011.
[18] Q.-S. Jia, “A potential-based method for finite-stage markov decision process,” in 2008 American Control Conference,
June 2008, pp. 5029–5034.
[19] K. Poularakis, G. Iosifidis, I. Pefkianakis, L. Tassiulas, and M. May, “Mobile data offloading through caching in residential
802.11 wireless networks,” IEEE Transactions on Network and Service Management, vol. 13, no. 1, pp. 71–84, March
2016.
[20] M. Ji, G. Caire, and A. F. Molisch, “Wireless device-to-device caching networks: Basic principles and system performance,”
IEEE Journal on Selected Areas in Communications, vol. 34, no. 1, pp. 176–189, Jan 2016.
[21] B. Zhou, Y. Cui, and M. Tao, “Optimal dynamic multicast scheduling for cache-enabled content-centric wireless networks,”
IEEE Transactions on Communications, vol. 65, no. 7, pp. 2956–2970, July 2017.
[22] K. Shanmugam, N. Golrezaei, A. G. Dimakis, A. F. Molisch, and G. Caire, “Femtocaching: Wireless content delivery
through distributed caching helpers,” IEEE Transactions on Information Theory, vol. 59, no. 12, pp. 8402–8413, Dec
2013.
[23] T. Liu, J. Li, F. Shu, M. Tao, W. Chen, and Z. Han, “Design of contract-based trading mechanism for a small-cell caching
system,” IEEE Transactions on Wireless Communications, vol. 16, no. 10, pp. 6602–6617, Oct 2017.
[24] Y. Cui, V. K. N. Lau, and Y. Wu, “Delay-aware BS discontinuous transmission control and user scheduling for energy
harvesting downlink coordinated MIMO systems,” IEEE Transactions on Signal Processing, vol. 60, no. 7, pp. 3786–3795,
July 2012.
[25] V. Chamola, B. Sikdar, and B. Krishnamachari, “Delay aware resource management for grid energy savings in green
cellular base stations with hybrid power supplies,” IEEE Transactions on Communications, vol. 65, no. 3, pp. 1092–1104,
March 2017.
[26] D. P. Bertsekas, Dynamic Programming and Optimal Control.
[27] S. K. Jayaweera, Markov Decision Processes.
Athena Scientific,, 2000.
Wiley Telecom, 2015, pp. 1–768. [Online]. Available: http:
//ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=8045654
[28] U. Tamm, “Some refelections about the Lambert W function as inverse of x*log(x),” in 2014 Information Theory and
Applications Workshop (ITA), Feb 2014, pp. 1–4.
| 7cs.IT
|
Simple Methods for Scanner Drift Normalization Validated for
Automatic Segmentation of Knee Magnetic Resonance Imaging –
with data from the Osteoarthritis Initiative
Erik B. Dam
- Biomediq A/S, Copenhagen, Denmark
- Department of Computer Science, University of Copenhagen
- Data Science Lab, Faculty of Natural Sciences, University of Copenhagen
Abstract. Scanner drift is a well-known magnetic resonance imaging (MRI) artifact characterized by gradual signal
degradation and scan intensity changes over time. In addition, hardware and software updates may imply abrupt
changes in signal. The combined effects are particularly challenging for automatic image analysis methods used in
longitudinal studies. The implication is increased measurement variation and a risk of bias in the estimations (e.g. in
the volume change for a structure). We proposed two quite different approaches for scanner drift normalization and
demonstrated the performance for segmentation of knee MRI using the fully automatic KneeIQ framework. The
validation included a total of 1975 scans from both high-field and low-field MRI. The results demonstrated that the
pre-processing method denoted Atlas Affine Normalization significantly removed scanner drift effects and ensured
that the cartilage volume change quantifications became consistent with manual expert scores.
Address all correspondence to: Erik B Dam, [email protected]
1
Introduction
Variation in signal intensity distribution over time is a well-known effect for magnetic resonance
imaging scans, and is typically termed scanner drift. For automatic methods, including
segmentation and biomarker quantification, scanner drift may impact the measurements.
Ideally, analysis methods could be developed to be invariant to scanner drift. A classical
example is registration using the normalized mutual information (NMI) measure to steer the
optimization1. NMI is invariant to bijective intensity transformations. Another example is use of
the scale-invariant feature transformation2 (SIFT) to generate features used in voxel
classification methods for segmentation. However, frameworks that implicitly or explicitly are
invariant to intensity scaling will by definition ignore aspects of the intensity information.
Therefore, it seems sensible to normalize the intensities by a simple transformation that preserves
most of the information captured by the intensity levels.
Ideally, an intensity transformation could be based on a physical model of the actual sources
of the intensity drift. However, these sources are numerous, including the scanner (equipment
drift, updates of software, replacement of spare parts), environment (temperature, pressure,
humidity), operator (patient positioning), intra-patient (liquid status, movement), and interpatient variation. To directly model the interplay between these sources of intensity variation is a
daunting task – particularly in the light of the large inter-subject biological variation that may
mask the more subtle drift effects.
Another appealing approach could be to include a calibration object in the scan. This has
traditionally been used to normalize intensity and geometry in traditional x-rays. It would also
allow normalization with respect to geometric drift. However, for MRI, the construction of a
phantom calibration object is much more challenging since it may be needed to include multiple
“tissues” designed for each combination of MRI sequence and anatomical target. Even if it may
1
be feasible to design such phantoms3, they would need to fit within the coil together with the
anatomy in order to be used for drift correction.
The focus of this paper is normalization of intensity variation. We will limit ourselves to two
processes we denote drift and shift. The drift is a consequence of wear and tear and results in
gradual loss of signal over time. Contrarily, shift is the result of hardware/software/environment
changes that affect the overall signal abruptly, often improving the signal to compensate for a
period of drift. The challenge is to compensate for drift and shift due to the scanner without
normalizing away intensity variation related to biological variation including pathology.
Specifically, for the application of knee MRI segmentation used in this work, the articular
cartilage should not be normalized to a uniform intensity within a scan or across a collection of
scans. We consider the following potential approaches that are illustrated in Fig 1:
a) Normalize individual scan
The simplest normalization where each scan is, for instance, normalized to have an
intensity distribution with zero mean and standard deviation one.
The challenge here is that knee MRI have large variation in the joint tissues included
in the scan due to differences in knee size and anatomy, as well as large differences in
muscle and fat tissue present. Therefore, such normalization will likely introduce large
variation in small tissues such as the central cartilage structures.
b) Normalize individual scan using reference template or atlas
A common approach is to normalize the scan against a single template or a multi-atlas
and several methods for global intensity transformations have been published. These
methods typically perform histogram matching to a template and transform the intensities
to normalize the histograms4–6 (e.g. using a polynomial intensity transformation4). For
brain MRI, a more sophisticated patch-based method for focal intensity correction was
demonstrated7. This used a library of image patches from scans with homogeneous
intensity and matched focal patches to replace the patch center intensity with the intensity
from the best matching library patch. This method potentially handles noise removal,
intensity inhomogeneity, and scanner drift. However, there is a risk of normalizing away
pathologies and the authors suggest limiting the use to normal subjects. These methods
were validated for brain MRI, except one validated for prostate analys5.
Straightforward histogram-driven normalization is problematic for knee MRI, where
the leg is partially outside the scan and tissue distributions vary greatly due to biological
variation, as mentioned above.
c) Normalize baseline and follow-up scans relatively to each other
Particularly in brain MRI, where focal atrophy is a key pathology indicator, registrationbased analysis is very common. This approach is also applicable to scanner drift
normalization. When only longitudinal change analysis is required, this is an appealing
approach since the large biological variation between subjects is not a challenge when
two scans from the subject are analyzed together. Therefore, intensity drift as well as
inhomogeneity field difference can be robustly corrected8.
d) Normalize quantified biomarkers across all scans in completed study
Instead of directly addressing the scanner drift in a pre-processing step, the drift effects
could also be normalized in post-processing using the entire collection of scans. The
inclusion of an entire study may allow robust estimation of the subtle drift and abrupt
shift effects as a function of time for each site.
2
Fig. 1: Each of the scanner drift normalization options a)-d) are illustrated in the workflow of the KneeIQ
framework that includes a multi-atlas registration followed by voxel classification to quantify cartilage volume
scores for each scan in the study. The framework is trained using a collection of scans with manual expert
annotations for the cartilage compartments. The solid lines show the KneeIQ data flow. The dotted lines show the
pre-processing steps of normalization options a)-c) and the post-processing step of option d).
Among the four approaches above, we will not consider option c) since we aim to allow crosssectional analysis of quantified biomarkers and not limit ourselves to longitudinal analysis. We
previously implemented the trivial zero-mean, std-one normalization mentioned in option a) as
an early experiment and experienced slightly worse segmentation accuracy compared to no
normalization (these results are not reported). This experiment confirmed the challenge of tissue
distribution variation in knee scans. As a result of the above, we propose two simple and very
different methods for scanner drift normalization:
• A pre-processing method for normalization of each scan using a collection of training
scans. To avoid the problem with tissue distribution variation, we apply a multi-atlas
registration to ensure the intensities are analyzed robustly in a well-defined structural
window. The method is termed Atlas Affine Normalization (AAN).
• A post-processing method that models the scan intensity drift and shift across the entire
study using the scanner maintenance logs and normalizes the impact on the estimated
biomarker scores (e.g. cartilage volumes) directly from this. This is termed Piecewise
Linear Drift Shift (PLDS) correction.
3
We implemented these methods and demonstrated them in combination with a recently
published, fully automatic method for segmentation of knee MRI9. The validation included two
longitudinal studies of the progression of knee osteoarthritis (OA). Knee OA is the most
dominant from of arthritis and among the highest contributors to disability in the world10. A
central imaging biomarker for knee OA is to measure the cartilage degradation as the volume
change over time. Typically, only small yearly cartilage loss rates less than 1% are observed11
meaning that measurement accuracy and precision are essential. Given a typical re-measurement
precision of around 5% for cartilage volume scores12, disease effects risk drowning in
measurement variation. Furthermore, if scanner drift introduces a measurement bias, it may be
impossible to observe cartilage-protective treatments effects. Therefore, control of scanner drift
is essential for the task of cartilage volume quantification.
2
Methods
The methods section includes descriptions of the cohorts, the existing segmentation framework,
the two drift normalization methods, and finally the experiments. The descriptions of the cohorts
and the segmentation framework are brief, for more detail we refer to the recent KneeIQ
segmentation paper9.
2.1
Knee MRI Collections
We used cohorts of knee MRI from the Center for Clinical and Basic Research (CCBR) and the
Osteoarthritis Initiative (OAI) study. The collections included scans with manual segmentations
that we used for training and validation. The relevant ethical review boards approved the studies.
The collections differed substantially in scanner type, sequence design, and disease severity
as described below. The scanner and sequence specifications are given in Table 1.
Table 1: Study descriptions in terms of scanners, sequences, and available manual annotations.
CCBR
Esaote 0.18T C-span
Sagittal Turbo 3D T1
40⁰ flip angle,
50 ms RT, 16 ms ET,
0.7 x 0.7 mm pixels,
0.8 mm slices,
scan time 10 min
Compartments with
Tibia Bone
manual segmentations Medial Tibial Cartilage (MTC)
Medial Femoral Cartilage (MFC)
Scanner
Sequence
OAI
Siemens 3T Trio
Sagittal 3D DESS WE
25⁰ flip angle,
16 ms RT, 4.7 ms ET,
0.36 x 0.36 mm pixels,
0.7 mm slices,
scan time 10 min
Medial Tibial Cartilage (MTC)
Lateral Tibial Cartilage (LTC)
Femoral Cartilage (FC)
Medial Femoral Cartilage (MFC)
Lateral Femoral Cartilage (LFC)
Patellar Cartilage (PC)
Medial Meniscus (MM)
Lateral Meniscus (LM)
The CCBR cohort included 159 subjects in a longitudinal study of 21 month. Approximately
half the subjects had no baseline radiographic signs of OA, defined as Kellgren and Lawrence13
4
(KL) score 0, and the rest were evenly distributed from KL 1 to KL 3. From this cohort, 30 and
100 knee MRIs were used for training and validation, respectively. These MRI also had manual
segmentations performed by an experienced radiologist. Among the validation knees, 38 also had
rescans a week after baseline to evaluate scan-rescan variation.
The OAI cohort included 4796 subjects and six main, yearly visits, from which we included
two sub-cohorts of 88 and 565 subjects and analyzed visits V00 and V01.
The small OAI sub-cohort mainly included KL 2 and KL 3 knees, similar to a typical phase 3
trial population. The large sub-cohort also had around one third knees that are KL 0 or KL 1. The
88 subjects were semi-manually segmentation by iMorphics and are available via the OAI
website14. We divided the baseline scans for this set into 44 scans for training and 44 for
validation. Further, we split the femoral cartilage segmentations into medial and lateral subcompartments for the training set.
The large OAI sub-cohort was selected as OAI project 9B14, where cartilage volume scores
for the double-echo steady-state (DESS) scans made by Chondrometrics are publicly available14.
This set includes 565 knees with volumes for both V00 and V01.
Both cohorts were evenly distributed among men and women and between left and right
knees. For details on the cohorts, see the segmentation paper9.
2.2
Existing KneeIQ Segmentation Framework
The segmentation framework used for the experiments combines rigid multi-atlas registration
with voxel classification in a multi-structure setting9. The framework is trained from a collection
of scans with expert segmentations of a set of anatomical structures. The framework name
“KneeIQ” comes from Knee Image Quantification.
The multi-atlas registration step allows geometric transformation of each scan to a common
space where structures are located consistently. This allows focused training of a voxel classifier
for each structure. The classifier is based on features selected for each structure among a large
set of potential features computed for each voxel including position, intensity, the 3-Jet at 3
scales15, and non-linear combinations of derivatives such as the Hessian eigenvalues. This
defines a k-nearest neighbors (kNN) classifier for each anatomical structure. The segmentation
of the full scan is defined by choosing the structure with strongest classifier response for each
voxel.
Since neither intensity nor the 3-Jet features are invariant to scan intensity changes, the
framework may be sensitive to scanner drift. Particularly, an overall scanner drift between
baseline and follow-up visits may create a general over- or under-segmentation bias for the
follow-up scans.
2.2.1 Multi-Atlas Rigid Registration
The aim of the registration step in KneeIQ is to produce a geometric transformation from a given
scan to a common atlas space. Since we augmented this step to perform scanner drift
normalization in one of the proposed methods, we describe the registration step briefly.
In the multi-atlas registration, a study scan is registered to each atlas scan using multi-scale
rigid registration where the optimization criterion is NMI. To avoid boundary extrapolation
artifacts, an outer margin of 15% of the scan is excluded in the evaluation of NMI. Each
registration results in a transformation consisting of translation, rotation and scaling parameters.
The final transformation for the scan is defined as the element-wise median. Since the
registration is driven by NMI, this step is invariant to bijective intensity transformations.
5
This registration is summarized as follows. During training, each atlas scan Imt is registered
to all other atlas scans and the median transformation Rt gives the transform to the atlas center.
When segmenting a new scan S, the scan is registered to all atlas scans resulting in
transformations St. The compositions StxRt provide transformations to the atlas center via each
atlas scan. Then element-wise median of these compositions defines the transformation from S to
the atlas center.
2.3
Scanner Drift Normalization: Affine Atlas Normalization
The Atlas Affine Normalization (AAN) augmented the multi-atlas registration step in KneeIQ.
Given the rigid registration of a study scan to an atlas scan, the study scan was transformed using
the geometric parameters.
We defined an affine intensity transformation as the least squares solution mapping the
intensities from the transformed scan to the atlas scan – where the 15% periphery of the scans
were ignored to avoid boundary extrapolation artifacts (illustrated in Fig 2). The affine intensity
transformation was defined as the element-wise median of the transformations to each atlas scan.
As done for the geometric transformation outlined above, this normalization step was first
performed for each atlas scan. This produced an affine intensity transformation for each atlas
scan At, scaletAt + offsett, and an affine transformation for the study scan S to the atlas At,
scaleSS + offsetS. The transformation for the study scan over the atlas scan to the atlas space
center was the combination of the two transformations: scalet(scaleSS + offsetS) + offsett.
We defined the final affine intensity transformation for the study scan as the median of the
combination transformation over each atlas scan. This affine transformation was applied to the
study scan as a pre-processing step to the voxel classification step in the KneeIQ framework.
The AAN is not strictly a pre-processing stop to KneeIQ since it is applied between the two
main operations of KneeIQ. However, since the multi-atlas registration is invariant to a global,
affine intensity transformation due to the NMI criterion, we conceptually think of AAN as a preprocessing step.
As illustrated in Fig. 2, by ignoring a 15% margin, we excluded boundary artifacts as well as
part of the muscle and fat tissues.
Fig. 2: The Atlas Affine Normalization. A Study Scan from the OAI V01 visit is transformed via the first atlas scan
(Atlas 1). The geometric transformation is applied to the scan (Scan Geom Trans) and the least squares affine
mapping of the scan intensities inside a 15% margin from scan to atlas is computed. This intensity mapping is
combined with the mapping for the atlas scan to give the intensity mapping to the atlas space center. This final
mapping is applied to the study scan (Scan Geom+Int Trans) as a preprocessing step before voxel classification.
6
2.4
Scanner Drift Normalization: Piecewise Linear Drift-Shift
It may not be possible to estimate scanner drift effects robustly from the individual scan due to
the large biological variation between subjects. In this case, the drift may be modeled by
assessing the overall change in intensities over time including all scans in the collection.
We designed the Piece-wise Linear Drift-Shift (PLDS) method under the assumption that
scanner drift results in a slow, linear change in overall intensity over time, interrupted by abrupt
shifts in overall intensity. Further, we assumed that the intensity changes are linearly related to
changes in the measured imaging markers. Finally, the method assumed the availability of a
scanner maintenance log to identify the dates where shifts can occur. Alternatively, header
information in the study scans may be used to detect shift events such as scanner software
updates.
The method is illustrated by the example in Fig. 3. The first step was to optimize a piecewise linear model for each site of the scan mean intensity as a function of the scan day. We
optimized using the Matlab “fmincon” function with the sum of absolute difference to the model
as penalty function. This model described the population-averaged scanner drift and shift. The
second step was to model the impact of intensity changes on biomarkers. This was done by
modeling the linear relationship between scan mean intensity changes and biomarker changes.
The normalization using the PLDS method was then done for each scan by using the scan
day and the biomarker values as the input. The scan day defined the change in intensity from
study start using the piece-wise linear intensity model. The biomarker values were then corrected
by the linear biomarker changes as function of the given intensity change.
For the CCBR study, the scanner maintenance log were available, defining dates for software
updates and replacement of spare parts.
For the OAI scans, we used the attributes StationName and SoftwareVersion from the scan
DICOM headers to search for shift events. For all four OAI sites, the software was updated at
least once. For one OAI site, the station was updated. In addition to these automatically
discovered shift events, we defined two speculative shift events by inspection of the intensity
drift visualizations (see Fig. 3).
2.5
Experiments and Statistical Analysis
The AAN was used during pre-processing and affected the segmentations directly. Therefore it
could be evaluated in terms of segmentation accuracy and precision. The PLDS was used as
post-processing and could therefore only be validated on the resulting volume measurements.
We evaluated segmentation accuracy by the Dice volume overlap and as the linear
correlation between biomarker scores, depending on availability of validation segmentations or
just volume scores. Precision was evaluated by the root mean squared coefficient of variation
(RMS CV) of the cartilage volume pairs from segmentations of repeated scans from the same
subject acquired within a short period (scan-rescan precision). Such scan-rescan pairs were only
available for the CCBR study.
The central validation of the drift normalization methods was to analyze the longitudinal
changes for the biomarkers. For the OAI cohort, validation cartilage volumes were available, and
the distributions of biomarker change scores were compared using a paired t-test. For the CCBR
cohort, no validation volume scores were available at follow-up, and the volume changes were
evaluated by comparing to scores typically reported in the literature.
7
Fig. 3: The Piece-Wise Linear Drift Shift method. For each site, the relationship between scan day and scan mean
intensity is modeled as piece-wise linear function. The shift events are marked by vertical lines. The top row shows
the intensity models for two OAI sites. For each marker, the linear relationship between changes in intensity and
changes in marker values are modeled. The bottom row shows this for two cartilage volume markers.
3
Results
The numbers of analyzed scans from each of the cohorts are listed in Table 2. The baseline (BL)
scans were used for training and evaluation of segmentation accuracy and precision. The
exception was the iMorphics set where the validation consisted of both BL and follow-up (FU)
scans to investigate the effect of drift on the segmentation accuracy. For all collections, the pairs
of BL and FU scans were used to quantify longitudinal changes and thereby investigate drift
effects.
Table 2: Number of knee MRIs from each cohort used for training of the framework, for validation of segmentation
accuracy and precision, and the number of BL/FU pairs for assessment of longitudinal changes. The iMorphics
validation set included 44 knees at both V00 and V01 to evaluate drift in segmentation accuracy.
CCBR
OAI iMorphics
OAI Chondrometrics
8
Training
30
44
Validation
110
44+44
Rescans
38
Longitudinal
266
44
565
3.1
Segmentation Accuracy
The validation segmentations accuracies are shown in Table 3 for the CCBR and OAI iMorphics
collections. The Dice volume overlaps were typically around 0.82, with patellar cartilage being
the most challenging with scores around 0.75.
For comparison, for the CCBR study the radiologist repeated the manual segmentations for
the 38 scans also used for the scan-rescan precision analysis. The intra-operator, intra-scan mean
Dice volume overlaps for both medial cartilages were 0.86.
There was no clear difference in accuracy between the original KneeIQ framework and the
version augmented with the Atlas Affine normalization. There was no clear different between
accuracy for the V00 and V01 visits for the OAI sub-cohort.
Table 3: Segmentation accuracy for training and validation in the CCBR and OAI iMorphics collections. Accuracy
is given as the Dice volume overlap showing mean±std values. The number of validation knees (n) is given for each
set. The empty entries are for compartments with no available manual segmentations in each sub-cohort. The
compartment abbreviations are defined in Table 1.
Compartment
MTC
LTC
MFC
LFC
FC
PC
MM
LM
Cartilages,
Mean
3.2
CCBR (n=110)
KneeIQ
KneeIQ+AAN
0.839±0.048 0.839±0.046
0.804±0.059
0.820±0.049
0.798±0.064
0.816±0.052
OAI V00 (n=44)
KneeIQ
KneeIQ+AAN
0.811±0.056 0.816±0.050
0.865±0.034 0.868±0.037
0.814±0.043 0.819±0.037
0.841±0.043 0.847±0.045
0.838±0.037 0.843±0.036
0.748±0.101 0.765±0.102
0.758±0.085 0.754±0.081
0.830±0.055 0.822±0.064
0.816±0.076 0.823±0.072
OAI V01 (n=44)
KneeIQ
KneeIQ+AAN
0.809±0.052 0.814±0.052
0.861±0.037 0.863±0.037
0.836±0.032
0.750±0.099
0.754±0.088
0.823±0.057
0.814±0.073
0.839±0.030
0.756±0.101
0.751±0.089
0.818±0.058
0.818±0.072
Cartilage Volume Precision
The scan-rescan precision for the CCBR collection is given in Table 4. For comparison, the
radiologist had intra-scan precision 6-7% (repeated manual segmentation of same scan). For
OAI, no scan-rescan examples are publicly available.
All variants of the automatic KneeIQ segmentation framework had better scan-rescan
precision than the radiologist. The Atlas Affine drift normalization had better precision than the
other variants for the medial tibial cartilage compartment, however this was not the case for the
total tibial and femoral medial cartilage volume.
Table 4: Precision of cartilage volumes for the 38 scan-rescan knees from the CCBR collection. The scan-rescan
results are for segmentations of repeated scans acquired approximately 1 week apart. Precision is given as the root
mean squared coefficient of variation (RMS CV).
Compartment
MTC
MFC
9
CCBR, Manual, Radiologist (n=38)
Radiologist KneeIQ KneeIQ+PLDS
8.9%
4.9%
4.9%
9.2%
4.9%
4.8%
KneeIQ+AAN
3.8%
5.1%
Total Cartilage
3.3
6.1%
4.1%
4.1%
4.1%
Cartilage Volumes, Cross-sectional Agreements
The cross-sectional agreement between volumes for the baseline visit is given for each cohort in
Table 5. The volumes from the variants of the automatic method were compared to the reference
volumes from manual or semi-automated segmentations. There were slight variations in both
linear correlation and bias offset across the automated KneeIQ framework and the two drift
normalization methods, but no clear trends were apparent.
Table 5: Cross-sectional agreement between the automatic KneeIQ volume scores and the validation segmentation
volumes on the CCBR collection and OAI sub-cohorts. The table includes number of knees available (n), Pearson
linear correlation coefficient (r), and mean signed relative difference (%). Abbreviations are given in Table 1.
Cohort
Compartment
KneeIQ
r
%
KneeIQ+PLDS
r
KneeIQ+AAN
%
r
%
CCBR (Validation: Radiologist, manual, n=110)
MTC
0.91
-7
0.90
-7
0.90
-9
MFC
0.92
-5
0.92
-5
0.90
-4
Total Cartilage
0.94
-5
0.93
-5
0.92
-5
OAI (Validation: iMorphics, semi-automatic, n=44)
MTC
0.93
11
0.93
12
0.92
12
LTC
0.94
6
0.94
6
0.94
4
MFC
0.89
6
0.89
4
0.91
0
LFC
0.91
5
0.91
5
0.90
4
FC
0.92
5
0.92
4
0.93
2
PC
0.88
15
0.88
17
0.90
8
MM
0.85
12
0.85
14
0.87
11
LM
0.87
4
0.87
7
0.79
6
Total Cartilage
0.92
9
0.92
10
0.92
7
OAI (Validation: Chrondrometrics, manual, n=565)
3.4
MTC
0.92
11
0.92
11
0.92
14
LTC
0.87
17
0.87
17
0.88
17
Total Cartilage
0.90
14
0.90
14
0.90
15
Cartilage Volumes, Longitudinal Agreements
The longitudinal volume changes are listed in Table 6 for the three cohorts. The each cohort, the
manual or semi-automatic validation segmentations defined a reasonable level for the mean
volume change for each structure. For each variant of the automatic framework, the
corresponding volume changes were compared to validation changes using a Student’s t-test.
Here, low p-values indicate that it was unlikely that the two methods captured the same
10
distribution. Therefore, the automatic method and the drift normalizations should ideally result in
high p-values, indicating that they captured the same volume changes as the validation scores.
For the CCBR cohort, no validation segmentations were available for the follow-up scans.
These results were therefore qualitatively compared to the expected cartilage volume changes
over 21 months, as suggested by previous studies. For the small OAI cohort, there was no clear
pattern in agreement, neither across the structures nor across the method variants.
For the large OAI cohort, the validation segmentations indicated a moderate cartilage loss
given by changes around -1%. The original KneeIQ method and the PLDS normalization both
resulted in cartilage growths around 1%. The combination of KneeIQ and Atlas Affine
normalization resulted in cartilage loss levels similar to the validation segmentations.
Table 6: Longitudinal cartilage volume changes for the automatic KneeIQ volume scores and the corresponding
validation scores on the CCBR and the OAI sub-cohorts. The table includes number of knees available (n), mean
relatively changes ± standard deviation (%). For each of the KneeIQ variants, the p-value from a t-test comparison
to the validation scores is given, where p>0.05 indicates that the scores are not statistically different (large p value
indicate the method agrees with the validation segmentations). The compartment abbreviations are in Table 1.
Cohort
Validation
Compartment
%
KneeIQ
KneeIQ+PLDS
%
p
%
KneeIQ+AAN
p
%
p
CCBR (Validation: Radiologist, manual, n=266)
MTC
4.3% ± 10.0%
5.3% ± 10.1%
3.1% ± 8.9%
MFC
-0.2% ± 13.7%
1.5% ± 13.9%
-1.8% ± 12.7%
Total Cartilage
1.0% ± 11.6%
2.4% ± 11.7%
-0.5% ± 10.4%
OAI (Validation: iMorphics, semi-automatic, n=44)
MTC
0.2% ± 7.4%
1.7% ± 6.4%
0.2555
2.5% ± 6.3%
0.0868
2.1% ± 6.3% 0.1560
LTC
-0.5% ± 5.4%
1.2% ± 5.2%
0.0955
0.2% ± 5.2%
0.4634
0.8% ± 5.0% 0.1221
FC
0.9% ± 4.2%
1.8% ± 5.8%
0.3922
0.4% ± 5.8%
0.5671
2.0% ± 4.1% 0.1732
PC
-2.0% ± 8.3%
0.2% ± 10.5% 0.3131
4.2% ± 10.5% 0.0070
2.6% ± 15.0% 0.0932
MM
2.9% ± 10.0% -3.0% ± 10.4% 0.0143 -0.4% ± 10.1% 0.1540 -2.6% ± 11.1% 0.0199
LM
1.3% ± 6.8%
1.3% ± 11.7% 0.9630
6.2% ± 11.9% 0.0045
0.3% ± 9.5% 0.4533
Total Cartilage
0.3% ± 3.2%
1.5% ± 4.9%
1.2% ± 4.9%
1.7% ± 3.8% 0.0248
0.1104
0.2445
OAI (Validation: Chrondrometrics, manual, n=565)
4
MTC
-1.0% ± 4.9%
1.0% ± 6.5% <0.0001
1.3% ± 6.2% <0.0001 -0.4% ± 6.0% 0.0152
LTC
-1.4% ± 4.7%
1.2% ± 6.7% <0.0001
0.5% ± 6.6% <0.0001 -1.3% ± 6.2% 0.7453
Total Cartilage -1.2% ± 3.5%
1.0% ± 4.9% <0.0001
0.8% ± 4.9% <0.0001 -0.9% ± 4.7% 0.1999
Discussion
Two fundamentally different scanner drift and shift normalization methods were implemented
and validated for the task of cartilage segmentation from knee MRI.
11
Fig. 4: Visualization of the intensity drift for an OAI site (station name 20576). Each BL scan was transformed to a
common space using the registration transform and the mean intensity was computed at each position (left). The
same was done for each FU scan, and the relative, focal intensity change was computed. To ignore boundary
extrapolation effect, the 15% outer margin was cleared (right). For this site, there was a mean intensity loss of 3.3%
from BL to FU, but the intensity changes were not uniformly distributed across the scan. Some (yellow) regions had
intensity increase, whereas most (green-blue) regions had some loss.
4.1
Limitations
The simple, affine, global normalization methods that we investigated have strong limitations.
The first is the assumption that the drift is global for a scan. For PLDS, there was further the
assumption that the drift impact on the measured biomarkers is linear. Whether or not these
assumptions are sufficiently close approximations of the actual drift depends on the scanner, the
sequence, and the biomarker quantification method. As a consequence, they may be appropriate
for one study, but not for others.
Specifically, the drift may result in changes in noise rather than changes in intensity levels.
Or changes in scanner inhomogeneity. These interplay between these factors was previously
addressed for the task of analysis of prostate cancer from MRI5. Their results demonstrated that
scanner drift normalization alone worsened the classification results, these results improved
when performing combination of bias field correction and noise filtering prior to intensity
standardization.
Further, the drift may have different effects on different tissue, again depending on the
sequence. The illustration in Fig. 4 hints that this is to some degree the case for the OAI scans.
In addition, the observed changes in intensity may not be the consequence of scanner drift,
but rather of changes in the subject – either due to pathology or changes in body composition. In
these cases, drift normalization may introduce added measurement variation. Particularly, PLDS
may respond to changes in body composition since added weight will “fill” the air around the
knee in the scan field of view. Since AAN excluded a boundary margin from the drift estimate,
the risk of weight dependency was smaller.
In general, due to the non-physical modeling of the scanner drift, our proposed normalization
schemes will at best capture part of the scanner drift effects. The intention of the evaluation is to
determine whether the methods improve the resulting biomarkers in terms of accuracy, precision,
and particularly longitudinal change consistency. However, the evaluation is itself possibly
slightly confounded by scanner drift effects since the validation segmentations may also be
12
affected by scanner drift. Since manual segmentation is performed with the ability to change the
intensity windowing interactively, the effect of simple changes in intensity levels should be
minimal. However, the iMorphics segmentations that were performed using a semi-automated
live-wire method may to some degree suffer from scanner drift. Due to this risk and the small
size of the iMorphics sub-cohort, we have chosen to put less focus on the change measurement in
this group. However, they are included in Table 6 for completeness.
4.2
Visualizing the Scanner Drift
The assumptions regarding the nature of the observed that be visually inspected. Since all scans
were registered to the atlas space, we can compute the mean BL and the mean FU scans, and
visualize the focal changes in intensity (see Fig. 4). The figure demonstrates that intensity loss
occurred throughout the scan, but also that the distribution was not uniform.
This means that our drift normalization methods are likely to simplistic to fully capture these
drift patterns since the methods assumed a global, affine drift. As a consequence, more
sophisticated methods with spatially varying correction could in principle provide better drift
normalization. However, the development of such methods is challenging since they could
normalize away focal pathology. The evaluation demonstrated to what degree our simplistic
assumptions approximated the observed drift and shift sufficiently to improve the cartilage
volume quantifications.
4.3
Results
Both PLDS and AAN scanner drift normalizations could potentially improve consistency in
segmentations when the collection is acquired over longer time periods. The AAN scanner drift
may also improve robustness against fluctuations in signal occurring over the course of a day and
thereby potentially improve scan-rescan precision. Here, we summarize the results for each
method.
The original KneeIQ methods had good accuracy with Dice volume overlaps around 0.82 and
volume linear correlations around 0.9 compared to validation segmentations. The scan-rescan
precision was 4.9% for both evaluated compartments. However, the volume change
measurements differed significantly from the manual validation segmentation scores, with mean
tibial cartilage volume changes of +1.0% compared to -1.2% for KneeIQ and validation,
respectively.
The PLDS normalization provided similar cross-sectional volume correlations and scanrescan precision scores as the original KneeIQ. The volume change measurements were also
significantly different from the manual validation scores.
The AAN normalization provided similar accuracy, but improved scan-rescan precision in
the medial tibial compartment to 3.8%, albeit with a slightly worse precision in the femoral
compartment at 5.1%. Regarding longitudinal volume changes, the AAN normalization resulted
in scores similar to the manual validation set with mean total cartilage volume change of -0.9%
compared to -1.2% from the manual segmentations.
4.4
Segmentation of Meniscus
The focus in this study was articular cartilage and robust estimation of volume changes.
However, for the OAI study, we also quantified medial and lateral meniscus volumes. We have
not discussed these results as the only longitudinal validation data was the small OAI sub-cohort
13
with semi-automated segmentations that may also be confounded by scanner drift effects. Future
research will show the validity of the meniscus segmentation and the meniscal volume changes.
4.5
Why AAN worked and PLDS did not?
Both AAN and PLDS relied on simple affine modeling of the drift effects. However, there were
two fundamental differences. First, the intensity modeling for AAN was performed in a window
where anatomical alignment was ensured by multi-atlas registration. This limited the impact of
the differences in body composition and scan background that confounded the PLDS model.
Secondly, the post-processing modeling by PLDS assumed a linear relationship between
intensity changes and biomarker changes. This assumption may be too strong for many
biomarkers.
4.6
Conclusions
There are complex, highly non-linear processes leading to scanner drift and subsequent changes
in biomarkers quantified from the scans using automated analysis. In this study, we investigated
whether two simple, linear methods could approximate the drift effects sufficiently to improve
biomarkers in terms of accuracy, precision, and most importantly change measurements.
The post-processing PLDS method retained accuracy and precision, but was unable to
improve the change measurements. The pre-processing AAN method retained accuracy,
marginally improved precision, and improved longitudinal cartilage volume measurements. The
resulting KneeIQ+AAN total cartilage volume changes of -0.5% and -0.9%, for the CCBR and
OAI studies respectively, are in line with the literature and were very similar to the manual
segmentation scores. Furthermore, the standard deviations on the change scores in Table 6 show
that the measurement variation was also improved due to the AAN correction.
Thereby, the evaluation showed that the AAN method was apparently able to model scanner
drift effects sufficiently well to improve the results significantly. And the improvement was
sufficient to ensure that the resulting biomarker scores became very similar to manual expert
scores in both cross-sectional and longitudinal comparisons. Future research will focus on the
relationship between the volume changes and clinical observations with the hypothesis that the
AAN scanner drift normalization will allow stronger clinical conclusions.
5
Acknowledgements
We gratefully acknowledge financial support from the Danish Research Foundation (“Den
Danske Forskningsfond”).
The research leading to these results has received funding from the D-BOARD consortium, a
European Union Seventh Framework Programme (FP7/2007-2013) under grant agreement n°
305815.
The CCBR collection was provided by the Center for Clinical and Basic Research, Ballerup,
Denmark, with manual cartilage segmentations by radiologist Paola C. Pettersen.
The OAI collection was provided by the Osteoarthritis Initiative with the cartilage and
menisci segmentations performed by iMorphics. The OAI is a public-private partnership
comprised of five contracts (N01-AR-2-2258; N01-AR-2-2259; N01-AR-2-2260; N01-AR-22261; N01-AR-2-2262) funded by the National Institutes of Health, a branch of the Department
of Health and Human Services, and conducted by the OAI Study Investigators. Private funding
partners include Merck Research Laboratories; Novartis Pharmaceuticals Corporation,
GlaxoSmithKline; and Pfizer, Inc. Private sector funding for the OAI is managed by the
14
Foundation for the National Institutes of Health. This manuscript was prepared using an OAI
public use data set and does not necessarily reflect the opinions or views of the OAI
investigators, the NIH, or the private funding partners.
6
References
1. D. Rueckert, A. F. Frangi, and J. A. Schnabel, “Automatic construction of 3-D statistical
deformation models of the brain using nonrigid registration,” IEEE Tranon Med. Imaging
22(8), 1014–1025 (2003).
2. D. G. Lowe, “Object recognition from local scale-invariant features,” in The Proceedings of
the Seventh IEEE International Conference on Computer Vision, 1999 2, pp. 1150–1157
vol.2 (1999) [doi:10.1109/ICCV.1999.790410].
3. K. Yoshimura et al., “Development of a tissue-equivalent MRI phantom using carrageenan
gel,” Magn. Reson. Med. 50(5), 1011–1017 (2003) [doi:10.1002/mrm.10619].
4. S. Prima et al., “Statistical Analysis of Longitudinal MRI Data: Applications for Detection of
Disease Activity in MS,” in Medical Image Computing and Computer-Assisted Intervention
— MICCAI 2002, T. Dohi and R. Kikinis, Eds., pp. 363–371, Springer Berlin Heidelberg (2002).
5. D. Palumbo et al., “Interplay between bias field correction, intensity standardization, and
noise filtering for T2-weighted MRI,” Conf. Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.
IEEE Eng. Med. Biol. Soc. Annu. Conf. 2011, 5080–5083 (2011)
[doi:10.1109/IEMBS.2011.6091258].
6. L. G. Nyúl, J. K. Udupa, and X. Zhang, “New variants of a method of MRI scale
standardization,” IEEE Trans. Med. Imaging 19(2), 143–150 (2000) [doi:10.1109/42.836373].
7. S. Roy et al., “Intensity inhomogeneity correction of magnetic resonance images using
patches,” 2011, 79621F – 79621F – 6 [doi:10.1117/12.877466].
8. B. Zou et al., “Simultaneous Registration and Bilateral Differential Bias Correction in Brain
MRI,” in Intelligent Imaging: Linking MR Acquisition and Processing (2015).
9. E. B. Dam et al., “Automatic segmentation of high- and low-field knee MRIs using knee
image quantification with data from the osteoarthritis initiative,” J. Med. Imaging 2(2),
024001–024001 (2015) [doi:10.1117/1.JMI.2.2.024001].
10. M. Cross et al., “The global burden of hip and knee osteoarthritis: estimates from the Global
Burden of Disease 2010 study,” Ann. Rheum. Dis., annrheumdis – 2013–204763 (2014)
[doi:10.1136/annrheumdis-2013-204763].
11. W. Wirth et al., “Comparison of 1-year vs 2-year change in regional cartilage thickness in
osteoarthritis results from 346 participants from the Osteoarthritis Initiative,” Osteoarthr.
Cartil. OARS Osteoarthr. Res. Soc. 19(1), 74–83 (2011) [doi:10.1016/j.joca.2010.10.022].
12. E. Schneider et al., “Equivalence and precision of knee cartilage morphometry between
different segmentation teams, cartilage regions, and MR acquisitions,” Osteoarthritis
Cartilage 20(8), 869–879 (2012) [doi:10.1016/j.joca.2012.04.005].
13. J. H. Kellgren and J. S. Lawrence, “Radiological assessment of osteo-arthrosis,” Ann Rheum
Dis 16(4), 494–501 (1957).
14. “The Osteoarthritis Initiative (OAI),” Publicly available at www.oai.ucsf.edu, 2012,
<www.oai.ucsf.edu>.
15. L. Florack, The Syntactical Structure of Scalar Images, PhD dissertation (1993).
15
| 1cs.CV
|
A New Optimal Algorithm for Computing the Visibility
Area of a simple Polygon from a Viewpoint
Hamid Hoorfar
∗
Alireza Bagheri∗
arXiv:1803.10184v2 [cs.CG] 29 Mar 2018
Abstract
Given a simple polygon P of n vertices in the Plane. We study the problem of computing the
visibility polygon from a given viewpoint q inside P where only few working variables can be used.
The problem was previously solved in O(nr) time using O(1) variables for working space, where r
denotes the cardinality of reflex vertices of P in the output. In this paper, we present an optimal
in-place algorithm of O(n) time, and O(1) extra variables (in addition to the being in-place) for
computing visibility area. If reporting the parts of P that are not in the visibility area is acceptable
as result, our algorithm can be a constant-memory for computing visibility area. We improve timecomplexity such that it just depends on n. Actually, we provide a linear-time and in-place algorithm
for computing visibility area of a viewpoint inside a simple polygon.
Keywords: Visibility Area, Constant-Memory Mode, Memory-constrained Algorithm, Two-pass
Algorithm, In-place Algorithm, Streaming Algorithm, Guarding Polygon.
1
Introduction
The visibility polygon of a simple polygon P from an internal point q is the set of all points of P that
can be visible from q, where two points p and q are visible from each other whenever the line-segment
pq does not intersects the boundary of P. The visibility area of viewpoint q inside P is the set of all
points in P that are visible from q. One of the important problem studies in computational geometry is
visibility [8]. Joe and Simpson found the first optimal algorithm for computing visibility area (region)
from a point in linear-time and space [9, 11]. In computing, when the input is presented as a sequence
of items, it is called data streams, also, if an algorithm is presented for processing data streams is called
streaming algorithm. In streaming algorithms, the input is presented as a sequence and is computed in a
constant p passes and typically in one pass, so, they are called p-pass algorithm. In most models, this type
of constrained-memory algorithms has restricted memory (generally less than logarithmic in the size of
input and typically constant). So, a p-pass algorithm is a streaming algorithm which sees its input exactly
p times, in order, without unbounded buffering. A p-pass algorithm generally requires less than O(n)
and typically constant space. As mentioned, these constant-memory algorithms do not have a long-term
memory structure. Deciding and processing of current data must be at the moment. Using constant
space, there is not enough memory for storing previously seen data or even for the index of this data. An
in-place algorithm is an algorithm which transforms input using no auxiliary data structure. Only a small
amount of extra workspace is allowed for auxiliary variables. The input is overwritten by the output as
the algorithm executes. In-place algorithm updates input sequence only through replacement or swapping
of elements. If extra O(1) bits for workspace is allowed, It is called strict in-place. The strict definition of
in-place algorithms includes all algorithms with O(1)-bits space complexity, the class DSPACE(1). This
class is very limited; it equals the regular languages [10]. However, in constraint-memory mode, there
is O(s) bits workspace during running algorithm, typically s ∈ O(log n). Several researchers worked on
constrained memory algorithm in visibility [1, 3, 5, 7]. Some related works are reviewed here. Computing
visibility area of a viewpoint q in the simple polygon P with extra O(1) variables has time-complexity
of O(nr) where r is the cardinality of the reflex vertices of P in the output [6] while there exists a
linear-time algorithm for computing the visibility area without constrained memory [2]. If there is O(s)
variables for the computing, the time complexity of computing visibility area decreases to O( nr
2s + n log r)
where s ∈ O(log r). It was presented a time-space trade-off algorithm which reports the k-visibility area
kn
for q ∈ P in O( cn
s + n log s + min{ s , n log logs n}) time, the algorithm uses O(s) words with O(log n)
bits each of workspace and c ≤ n is the number of vertices of P where the visibility area changes for the
viewpoint q [4]. If k = 1 and s = 1, their algorithm time complexity is O(cn), in fact, using O(log n) bits
∗ Department of Computer Engineering and Information Technology, Amirkabir University of Technology (Tehran Polytechnic), {hoorfar,ar bagheri}@aut.ac.ir
A New Optimal Algorithm for Computing the Visibility Area of a simple Polygon from a Viewpoint
V12
V11
V10
V8
V7
θ20
q
V1
positive X-axis
V20
Figure 1: Illustration of the angle of pi , the angle, θ20 is the angle of qp20 , the ∠V1 V7 V8 is a clockwise
turn and the angle ∠V10 V11 V12 is a counterclockwise turn.
space.
Our General Result:
We provide an optimal algorithm of O(n) time, and O(1) extra variables for computing visibility area of
a view point q inside P in the in-place model and we improve time-complexity such that it just depends
on n. With the difference that if we change our algorithm to be not-in-place(to be constant-memory),
we can just report parts of P that are not in the visibility area of q.
2
Preliminaries
Assume that a simple polygon P is given in a doubly linked list L of n vertices (every vertices in one word
with log(ω) bits each) in counterclockwise order along the boundary and there is a read-only variable
that is contained a query point q ∈ P (one word with log(ω) bits), where ω is the maximum storage
size of vertices, it differs from n. The storage size of variables that are used as working space, also, is
log(ω) bits. In this model, the input is provided in a sequential-access with two directions forward and
backward. If 2-direction sequential-access is acceptable, our algorithm is streaming. Sequential access is
the only way of accessing a random element in L (in our model). Let l0 , l1 , l2 , . . . , ln−1 be the elements
of L, in forward traversal, the input is streamed from l0 to ln−1 , and in backward traversal, the input is
streamed from ln−1 to l0 , sequentially. There is a few additional variables of memory as working space or
workspace. The workspace is both writable and readable during the running of algorithm. The output is
presented with modifying input L as the algorithm executes. At the end of processing algorithm, Let ti is
the number of reading element li in L where 0 ≤ i < n, and p = maxn−1
i=0 (li ), then the algorithm is called
p-pass. The aim is to report a suitable representation of the visibility area of P in the doubly linked list
L, using O(1) variables, so, there is no extra space for writing the result. For every vertex pi of P, we
→ and the positive X-axis as denoted by θ (radian). For every
compute the angle between vector (ray) −
qp
i
i
θi , where i between 0 and n − 1, 0 ≤ i ≤ 2π. For every 3 vertices v1 ,v2 and v3 , the angle ∠v1 v2 v3 is called
→
counterclockwise turn, if and only if v3 is placed on the left side of the vector −
v−
1 v2 , otherwise, if and only
if v3 is placed on the right side, it is called clockwise turn, see Figure 1. Sometimes, for simplicity we
denote angle ∠v1 v2 v3 as ∠v2 . For two vertices ν and ν 0 , let Γν is the line which is contained ν and q, let
e1 and e2 are two edges which are contained ν 0 , so, the intersection of e1 and Γν , or e2 and Γν , that one
touches the polygon from internal of P(from direction of q), is called shadow of ν, we denote it by S(ν)
or ν, see Figure 2(A) and 2(B). In boundary traversal of P, if θc−1 < θc > θc+1 or θc−1 > θc < θc+1 , pc
is called critical vertex. If θc−1 < θc > θc+1 , then pc is called critical-max vertex and if θc−1 > θc < θc+1 ,
pc is called critical-min vertex. In the Table 2, There is a list of functions that are useful for our data
structure, doubly linked list.
Hamid Hoorfar and Alireza Bagheri, 2018
q
e1
e2
ν
ν
q
ν0
e2
ν0
ν
e1
ν
(A)
(B)
Figure 2: (A)the red box is shadow of ν, in this case, there are two intersections between e1 , e2 and qν,
so, the red circle is not the shadow. (B)the red box is shadow of ν, in this case, there is one intersection
between e1 , e2 and qν.
Function Name
Insert(x)
Delete()
Update(x)
Forward()
Backward()
Return()
3
Table 1: Functions
Time Complexity
θ(1)
θ(1)
θ(1)
θ(1)
θ(1)
θ(1)
List for doubly linked list structure
Description
Insert a new element before current element(data=x).
Delete current element.
Replace data of current element with value of x.
Go to next element.
Go to previous element.
Print current element in the output.
A Constant-Memory Algorithm for Computing Visibility area of Simple Polygons
In this section, the aim is presenting an algorithm for computing the visibility area of q inside P and
only O(1) extra variables is available(allowed). We assume that vertices of simple polygon P is given in
doubly linked list L. Let l0 , l1 , l2 , . . . , ln−1 be the elements of L and p0 , p1 , p2 , . . . , pn−1 be the vertices
of P in counterclockwise order along the boundary and assume that θi is the angle of qpi . We use 3
variables as workspace. Our algorithm contains two traverse in L, forward and backward. Note that the
aim is to report a suitable representation of the visibility area of P from q.
Forward Traversal:
Using forward traversal in L, actually, we traverse the vertices of P along the boundary in counterclockwise
order. Considering angles, we continue to traverse until finding the first element that is not in the
ascending order which is named critical-max element(vertex). When we encounter such element pi , we
check whether ∠pi−2 pi−1 pi is clockwise turn or not. If yes, we delete (pi ) and continue deleting until
we occur first element pk such that θk > θi , we update and replace value of pk−1 with the shadow of
(pi ) (shadow pi on pk ) and continue. Otherwise, if no, we ignore current element and next descending
elements until we occur next critical-min vertex and continue. See Figure 3(A).
Backward Traversal:
After forward traversal, we traverse the vertices of P along the boundary in clockwise order. Considering
angles, we continue to traverse until finding the first element that is not in the descending order which
is named critical-min element(vertex). When we encounter such element pj , we delete (pj ) and continue
deleting until we occur first element pl such that θl < θj , we update and replace value of pl−1 with the
shadow of (pj ) (on pl ) and continue. See Figure 3(B).
Our algorithm do the above two procedures, respectively. By doing these, no critical vertex will remained
in L. Others vertices that are remained in L are the boundary of visibility area of q in P, see Algorithm 1.
Since our algorithm is 2-pass (according to our definition), and since processing each vertex takes O(1)
time, the running time of the algorithm is O(n), using O(1) workspace(variables). Now, we have the
following theorem:
Theorem 1 Given a simple polygon P with n vertices in a doubly linked list and a view point q inside
P. Using O(1) variables, there is an algorithm which computes visibility area of q in P in O(n) time.
See Algorithm 1 and Video 1 to illumination.
In our strategy, we modify and delete unused elements from input to obtain the output. Of course, we
can save the output in a new write-only linked list or array by a little change in Algorithm 1. If we put
the Return function instead of the Insert and Delete functions, then we can report the parts of polygon
A New Optimal Algorithm for Computing the Visibility Area of a simple Polygon from a Viewpoint
Algorithm 1: Computing the visibility area using constant workspace.
Input: Vertices of simple polygon P in doubly linked list L, a view point q in P
Output: Vertices of visibility area of P from q in L, ordered counterclockwise.
1 set V0 , V1 , V2 three vertices with smallest angle in L, respectively;
2 if V2 = N ull then
3
End of Algorithm.
4 end
5 while V1 is not a critical-max vertex do
6
V0 = V1 ; V1 = V2 ; V2 =Forward();
7 end
8 if ∠V1 is clockwise turn then
9
while V1 > V2 do
10
Delete(); V2 =Forward();
11
end
12
Insert(S(V1 )); Goto 2;
13 else
14
while V0 > V1 do
15
V0 = V1 ; V1 = V2 ; V2 =Forward();
16
end
17
Goto 2;
18 end
19 set V0 , V1 , V2 three vertices with largest angle in L, respectively;
20 while V1 is not critical-min vertex do
21
V0 = V1 ; V1 = V2 ; V2 =Backward();
22 end
23 while V1 < V2 do
24
Delete(); V2 =Backward();
25 end
26 Insert(S(V1 )); Goto 2;
(A)
(B)
Figure 3: (A)gray areas are the deleted parts of P using the forward traversal. (B)gray areas are the
deleted parts of P using the backward traversal.
that are not in the visibility area. It leads to reporting the visibility area, uniquely. With the difference
that if we change our algorithm to be not-in-place, we can just report parts of P that are not in the
visibility area. Also, our current algorithm can be used for polygons with holes, polygonal chains, sets of
non-crossed line segments and sets of disjointed polygons(obstacles) in the plane.
In statically typed languages, a variable also has a type, meaning that only certain kinds of values can
be stored in it, so, we are allowed to assume ω is constant without no assumption about n(the number of
vertices). Also we can say our algorithm is used O(1) bits space. But, in dynamically typed languages, it
is values, not variables, which carry type, so, our algorithm is used extra O(log ω) bits for its workspace,
as mentioned ω and n are not same (maybe ω n).
Hamid Hoorfar and Alireza Bagheri, 2018
4
Conclusion
We studied the problem of computing visibility of a viewpoint inside simple polygon. We used a different
data structure for finding visibility area. we presented an in-place algorithm with using O(log n) additional
space to compute visibility area of a viewpoint in simple polygon using a different data structure.
References
[1] Mikkel Abrahamsen. Constant-workspace algorithms for visibility problems in the plane. Master’s
thesis, University of Copenhagen, 2013.
[2] Takao Asano, Tetsuo Asano, Leonidas Guibas, John Hershberger, and Hiroshi Imai. Visibility of
disjoint polygons. Algorithmica, 1(1-4):49–63, 1986.
[3] Tetsuo Asano, Kevin Buchin, Maike Buchin, Matias Korman, Wolfgang Mulzer, Günter Rote, and
André Schulz. Memory-constrained algorithms for simple polygons. Computational Geometry,
46(8):959–969, 2013.
[4] Yeganeh Bahoo, Bahareh Banyassady, Prosenjit Bose, Stephane Durocher, and Wolfgang Mulzer.
Time-space trade-off for finding the k-visibility region of a point in a polygon. In International
Workshop on Algorithms and Computation, pages 308–319. Springer, 2017.
[5] Luis Barba, Matias Korman, Stefan Langerman, and Rodrigo I Silveira. Computing the visibility
polygon using few variables. In International Symposium on Algorithms and Computation, pages
70–79. Springer, 2011.
[6] Luis Barba, Matias Korman, Stefan Langerman, and Rodrigo I Silveira. Computing a visibility
polygon using few variables. Computational geometry, 47(9):918–926, 2014.
[7] Minati De, Anil Maheshwari, and Subhas C Nandy. Space-efficient algorithms for visibility problems
in simple polygon. arXiv preprint arXiv:1204.2634, 2012.
[8] Subir Kumar Ghosh, Thomas Caton Shermer, Binay Kumar Bhattacharya, and Partha Pratim
Goswami. Computing the maximum clique in the visibility graph of a simple polygon. Journal of
Discrete Algorithms, 5(3):524–532, 2007.
[9] Barry Joe and Richard B Simpson. Corrections to lee’s visibility polygon algorithm. BIT Numerical
Mathematics, 27(4):458–473, 1987.
[10] Maciej Liskiewicz and Rüdiger Reischuk. The complexity world below logarithmic space. In Structure
in Complexity Theory Conference, 1994., Proceedings of the Ninth Annual, pages 64–78. IEEE, 1994.
[11] Csaba D Toth, Joseph O’Rourke, and Jacob E Goodman. Handbook of discrete and computational
geometry. CRC press, 2017.
A
A.1
Appendices
List of Videos
Video 1 The video is available at YouTube and DropBox.
| 8cs.DS
|
Assumption-Free and Near-Optimal Sparse Inverse Covariance Selection
Sidhant Misra 1 Marc Vuffray 2 Andrey Y. Lokhov 1 Michael Chertkov 2 3
arXiv:1703.04886v2 [cs.LG] 26 Feb 2018
Abstract
We study the problem of reconstructing the graph
of a sparse Gaussian Graphical Model from independent observations, which is equivalent to
finding non-zero elements of an inverse covariance matrix. For a model of size p and maximum
degree d, the information theoretic lower bound
requires that the number of samples needed for recovering the graph perfectly is at least d log p/κ2 ,
where κ is the minimum normalized non-zero
entry of the inverse covariance matrix. Existing algorithms require additional assumptions to
guarantee perfect graph reconstruction, and consequently, their sample complexity is dependent
on parameters that are not present in the lower
bound. We show that even for very simple examples where the precision matrix is close to singularity, there is a huge gap between the samples
required by existing methods and the lower bound.
We propose an estimator, called S LICE, that consists of a cardinality constrained least-squares regression followed by a thresholding procedure.
We show that S LICE attains a sample complexity
of κ644 d log p, which differs from the lower bound
by only a factor proportional to 1/κ2 . To the best
of our knowledge, this is the first result in the
literature that provides a completely assumptionfree reconstruction guarantee and shows the existence of a polynomial-time algorithm whose sample complexity depends only on the parameters
present in the information theoretic lower bound.
1
Theoretical Division T-5, Los Alamos National Laboratory, Los Alamos, NM 87545, USA 2 Theoretical Division
T-4, Los Alamos National Laboratory, Los Alamos, NM
87545, USA 3 Skolkovo Institute of Science and Technology, 143026 Moscow, Russia. Correspondence to: Sidhant
Misra <[email protected]>, Marc Vuffray <[email protected]>,
Andrey Y. Lokhov <[email protected]>, Michael Chertkov
<[email protected]>.
1. Introduction
Graphical Models (GMs) are powerful and intuitive modelling tools to represent the dependency structure of a set
of random variables. In undirected graphical models, also
known as Markov Random Fields (MRFs), the sparsity pattern of the graph succinctly captures the conditional independence between the variables through the separation property.
In this paper, we study Gaussian Graphical Models (GGMs)
that are described on a graph G = (V, E) with |V| = p and
with maximum degree d. The probability density function
of the multivariate Gaussian distribution1 is given by
p
X
X
det(Θ)
1
P(x) =
exp −
θii x2i −
θij xi xj ,
p
2
(2π) 2
i∈V
(i,j)∈E
(1)
where Θ is the inverse covariance matrix, also known as
the precision matrix. For the special case of GGMs, the
sparsity pattern of the precision matrix Θ and the graph G
is identical; (i, j) ∈ E iff θij 6= 0.
We consider the problem of reconstructing the underlying
graph G of an unknown sparse GGM from n i.i.d. samples
drawn from the distribution described by Eq. (1), where
samples {xki }i∈V are indexed by k ∈ {1, . . . , n}. We are
interested in finding tractable algorithms that provide an
estimate Ĝ of the graph G using as few samples as possible.
Moreover we require that with high-probability the graph
should be perfectly reconstructed with a given confidence,
meaning that for a prescribed δ > 0, we require that P(Ĝ =
G) > 1 − δ.
This problem has attracted great interest due to its application in numerous fields including protein interactions (Friedman, 2004; Jones et al., 2012), gene regulatory networks
(Basso et al., 2005; Menéndez et al., 2010) and neuroscience
(Huang et al., 2010; Varoquaux et al., 2010).
Numerous methods have been proposed in the literature
to reconstruct sparse GGMs. Meinhausen and Buhlman
in (Meinshausen & Bühlmann, 2006) adopted a regression
based approach and used the L ASSO (Tibshirani, 1996) to estimate the neighborhood of each vertex. Yuan and Lin (Yuan
1
Our analysis only relies on properties of the unbiased empirical covariance estimator and can then be easily extended to the
case of a distribution with non-zero mean.
Assumption-Free and Near-Optimal Sparse Inverse Covariance Selection
& Lin, 2007) introduced the `1 regularized log-likelihood
estimator, commonly known as G RAPH L ASSO, which was
further explored in (d’Aspremont et al., 2008) and (Ravikumar et al., 2009). Anandkumar et al. in (Anandkumar et al.,
2012) used empirical conditional covariance thresholding
and proved structural consistency for walk-summable models. In (Cai et al., 2011), Cai, Liu and Luo introduced an
`1 regularized estimator called C LIME and later an adaptive version called ACLIME (Cai et al., 2016), that can be
implemented by linear programming. With suitable assumptions, all the above methods have been shown to successfully
exploit sparsity and reconstruct the underlying graph G perfectly with a sample complexity of O(log p).
Among other quantities, all the above algorithms exhibit
dependence on the condition number of Θ. The most prevalent assumption in the literature is the so-called Restricted
Eigenvalues (RE), which directly translates to the bound on
the condition number of Θ. As a result, it appears to be a
common belief that the RE condition is necessary to enable
model reconstruction. However, the condition number has
no impact on the IT lower bound in (2). To illustrate this,
consider the following very simple precision matrix
1
κ0
κ0
1
1 − ,
Θ = κ0
(4)
κ0 1 −
1
An information theoretic lower bound for the minimum
number of samples required for perfect graph reconstruction
was derived in (Wang et al., 2010). The lower bound reads
p
log p−d − 1
log
−
1
2
d
2
,
,
n∗ > max
dκ
dκ
4κ2
log 1 + 1−κ
− 1+(d−1)κ
where 1 − > κ0 > 0. The minimum normalized edge
strength for the above matrix is given by κ = κ0 . On the
(Θ)
other hand the condition number is given by λλmax
≥ −1 .
min (Θ)
This means that if we keep κ0 fixed and let → 0, the IT
bound in (2) remains fixed whereas the condition number,
and hence the sample complexity of existing algorithms
diverges to ∞.
(2)
where the parameter κ denotes the minimum normalized
edge strength and is given by
|θij |
κ = min p
.
(i,j)∈E
θii θjj
(3)
The inverse square dependence on κ in (2) appears due to
the difficulty in detecting edges of low strength from their
absence. Noticeably, the IT lower bound only depends on
the dimension p, sparsity d, and minimum edge strength κ.
In contrast, existing estimators in the literature always require additional assumptions to prove perfect graph reconstruction. The analysis of L ASSO in (Meinshausen &
Bühlmann, 2006) assumes certain incoherence properties
of the matrix Θ, reminiscent of the compressed sensing
problem. In (Ravikumar et al., 2009), a variant of the incoherence condition is assumed, which by itself can be hard to
verify. The proof of C LIME (Cai et al., 2011) and ACLIME
(Cai et al., 2016) require that the eigenvalues of the precision
matrix are bounded. The analysis of the conditional covariance thresholding algorithm (Anandkumar et al., 2012) was
conducted only for the class of so-called walk-summable
models. More recently, non-convex optimization based
methods (Johnson et al., 2012; Wang et al., 2016) have been
proposed to improve sample complexity but they require
that the eigenvalues of Θ are bounded. The very popular PC
algorithm for GM selection (Kalisch & Bühlmann, 2007;
Spirtes et al., 2000) requires the so-called strong faithfulness
assumption, which has been shown to eliminate a non-trivial
volume in the space of all precision matrices (Uhler et al.,
2013). Due to these additional assumptions, the sample
complexity of the algorithms depends on extra parameters
that are not present in the IT lower bound (2).
With the above illustration, we see that there is a gap between the information theoretic lower bound and the sample
complexity of known algorithms. It remains an open question whether the IT lower bound in (2) is loose and fails
to capture dependence on additional parameters, or if the
existing algorithms fail to achieve the same parametric dependence as the lower bound in their sample complexity.
In this manuscript, we propose a two stage algorithm consisting of a sparse least-squares regression followed by a
thresholding procedure. We call this algorithm Sparse Leastsquares Inverse Covariance Estimator (S LICE). In a completely assumption-free way we prove that S LICE successfully reconstructs the graph G perfectly
d+1with
probability
greater than 1 − δ using d + κ324 log 4p δ
samples. The
sample complexity of our algorithm is off from the lower
bound only by a O(1/κ2 ) factor. More importantly, it depends only on the same set of parameters as the lower bound,
thus closing the information theoretic gap with respect to
parametric dependence. To the best of our knowledge, this
is the first result in the literature that shows the existence of
an algorithm that achieves the same parametric dependence
as the IT lower bound with a completely assumption-free
proof.
The first step in the algorithm adopts a regression perspective and similar to the L ASSO, minimizes an empirical least
squares objective for each vertex. However, instead of using
an `1 penalty we use a cardinality constraint, often referred
to as an `0 constraint, which enforces that the number of
non-zero elements of the regression vector is at most d. The
second step, based on a product and threshold procedure is
unique to our algorithm and is critical to keep the sample
complexity low.
Assumption-Free and Near-Optimal Sparse Inverse Covariance Selection
The computational complexity of algorithm is dictated primarily by the heavier non-convex regression step. In its
bare form, this step is equivalent to an exhaustive search
over all possible size d neighborhoods of a vertex, resulting
in a computational complexity of O(pd+1 ). However, as
we show in our numerical experiments, the problem can be
solved much faster in practice by recasting it as a Mixed Integer Quadratic Program (MIQP), an approach that has been
explored previously in the compressed sensing literature
(Bertsimas et al., 2016). The MIQP formulation leverages
significant recent advances in mixed integer programming
technology and availability of powerful commercial solvers.
The paper is organized as follows: In Section 2 we introduce
our algorithm S LICE and its sub-routines. In Section 3 we
provide rigorous mathematical guarantees on the sample
complexity required to achieve perfect graph reconstruction
with our algorithm. Proofs and analysis of the optimization step and the thresholding procedure can be found in
Section 4 and in Section 5 respectively. Section 6 contains
numerical simulations on synthetic and real data. Section 7
contains conclusions and discussions on future work.
2. Algorithm
s.t.
kβi k0 ≤ d,
(6)
where the `0 -norm counts the number of non-zero components and Σ̂ denotes the empirical covariance matrix whose
components are given by
n
1X k k
xi xj .
n
Σ̂ij =
(7)
k=1
As the notation suggests, the objective in (6) can be rewritten
only in terms of Σ̂ as
Li (βi , Σ̂) = βiT Σ̂īī βi + 2Σ̂iī βi + Σ̂ii ,
(8)
where ī = [p] \ i.
Step 2: Product and Threshold
Once the estimates β̂i have been obtained for all i ∈ V, we
estimate the edge-set Ê through the following thresholding
procedure
q
Ê = (i, j) ∈ V × V : |β̂ij × β̂ji | > κ/2 .
(9)
The estimated graph is then declared as Ĝ = (V, Ê).
In this section we describe the details of the S LICE algorithm, and the interpretation and implementation of its subroutines.
2.1. Details of the S LICE Algorithm
In this section we specify the details of the algorithm,
which consists of a minimization routine followed by a
post-processing routine.
Step 1: Least Squares with `0 -constraint
Our first step consists of minimizing the well-known
quadratic least squares objective with an added `0 constraint.
This step is based on the interpretation of a multivariate
Gaussian distribution where each variable can be expressed
as a noisy linear combination of its neighbors,
X θij
Xi =
− Xj + ηi ,
(5)
θii
j6=i
where ηi is independent from X[p]\i , where [p] =
{1, 2, . . . , p}. The above representation suggests the use
of sparse linear regression, e.g., see (Meinshausen &
Bühlmann, 2006), where the regression coefficients corθ
respond to βij = θij
, and the non-zero values of the coeffiii
cients determine the graph structure.
Define βi = (βij )j6=i . For each i ∈ V, let
2
n
1 X k X
βij xj ,
β̂i = argmin Li (βi , Σ̂) =
xi +
n
βi ∈Rp−1
k=1
j6=i
2.2. Interpretation as Minimum Conditional Variance
For a given i ∈ V, the `0 -constraint in (6) is tantamount
to an exhaustive search over all candidate neighborhoods
of vertex i, denoted by B̂i , of size exactly d. Note that
neighborhoods of size strictly less than d can be eliminated
from the search as they are strictly contained in bigger candidate neighborhoods. Thus, the optimization in (6) can be
rewritten as
min
B̂i ⊆[p]\i : |B̂i |=d
βi ∈
min
R
p−1 :Supp(β )⊂B̂
i
i
Li (βi , Σ̂).
(10)
Since we will be restricting ourselves to the case when
d < n, each d × d submatrix of Σ̂ has full rank, and the
inner minimization in (10) can be explicitly resolved to get
β̂i (B̂i ) = −Σ̂−1
Σ̂B̂i i .
B̂ B̂
i
(11)
i
The corresponding optimal value is given by
L∗i (B̂i , Σ̂) = Li (β̂i (B̂i ), Σ̂) = Σ̂ii − Σ̂iB̂i Σ̂−1
Σ̂
B̂i B̂i B̂i i
h
i
= Σ̂−1
,
(12)
(iB̂ )(iB̂ )
i
i
11
where in the last line we use the matrix inversion lemma.
The expression in (12) is the empirical conditional variance
of Xi conditioned on XB̂i ,
d i |X ).
L∗i (B̂i , Σ̂) = Var(X
B̂i
(13)
Assumption-Free and Near-Optimal Sparse Inverse Covariance Selection
Thus, the first step of the S LICE algorithm is equivalent to
finding a candidate neighborhood B̂i ∈ [p] for vertex i that
minimizes the empirical conditional variance. To see why
this is a suitable objective, it is helpful to consider the limit
of large number of samples, when the empirical conditional
variance is replaced by the true one. In this case, any set B̂i
that minimizes the conditional variance must be a superset
of the neighborhood of i. This follows from the observation
that conditioning reduces variance in the case of multivariate
Gaussians, and the separation property of GMs,
Var (Xi | XA ) ≥ Var Xi | X[p]\{i} , ∀A ⊆ [p] \ {i},
(14)
Var Xi | X[p]\{i} = Var (Xi | XBi ) ,
(15)
where Bi ⊂ [p] denots the set of neighbors of i.
We remark here that (14) is a property specific to Gaussians, but can also be thought of as a special case of
the more general information theoretic inequality which
states that conditioning reduces entropy. Indeed, (14)
and (15) holds for any graphical model if we replace
Var(.) by the Shannon entropy function H(.), which are
equivalent in the case of Gaussians since H (Xi | XA ) =
1
2 log [2πeVar (Xi | XA )]. Incidentally, minimizing conditional entropy, more popularly referred to as the pseudolikelihood loss function has been proved to reconstruct graph
structure in (Ravikumar et al., 2010) under incoherence assumption, and recently in (Lokhov et al., 2017) without any
assumptions.
3. Sufficient Conditions for Structure
Recovery
In this section we present our main result which states that
the S LICE estimator successfully recovers the graph structure G with a number of samples that depends only on κ, d
and p.
Theorem 1. Given δ > 0, the probability of perfect graph
reconstruction using S LICE is lower bounded as P(Ĝ =
G) > 1 − δ, provided that the number of samples satisfies
d+1
32
4p
n − d > 4 log
.
(16)
κ
δ
We prove Theorem 1 through the results below that provide
guarantees for each step of the S LICE estimator.
Proposition 1 (Optimal support contains the true support).
For each i ∈ V, let B̂i ⊂ [p] be the support of the optimal
solution in (6) and let Bi ⊂ [p] be the neighbors of i. Then
for any δ > 0, the support B̂i satisfies Bi ⊆ B̂i with
probability greater than 1 − δ/2, provided that the number
of samples satisfies
d+1
32
4p
n − d > 4 log
.
(17)
κ
δ
Proposition 2 (Post-processing Proposition). Assume that
8dp
64
n − d > 2 log
.
(18)
κ
δ
Then with probability greater than 1 − δ/2, the post processing procedure consisting of Product and Threshold terminates with exactly the correct support.
Theorem 1 follows by combining Proposition 1 and Proposition 2 and applying the union bound.
4. Proof of Proposition 1
To prove Proposition 1, it is sufficient to show that if Bi is
not a subset of B̂i then the corresponding objective value
L∗i (B̂i , Σ̂) in (12) is too large. We establish this in two steps.
We first show that when Σ̂ is replaced by the true underlying covariance matrix Σ in (12), then the corresponding
objective value L∗i (Bi , Σ) is smaller than L∗i (B̂, Σ) by a
factor that depends on κ, for any |B̂| = d that does not
contain Bi (see Lemma 1). We then show that the ordering
is preserved even with the estimation error that results when
we replace Σ by Σ̂ provided that we have sufficient number
of samples as specified by Proposition 1. These statements
are formalized in the following lemmas.
Lemma 1 (Multiplicative gap in noiseless optimal solutions). Fix i ∈ V and let Bi ⊂ [p] be the neighbors of i.
Let B̂ ⊂ [p] be any subset such that |B̂| = d and Bi 6⊆ B̂.
Then
L∗i (B̂, Σ) ≥ L∗i (Bi , Σ)(1 − κ2 )−1 .
(19)
Lemma 2 (Large deviations on L∗i (., Σ̂)). Fix i ∈ V and
let 0 < < 1. Then for every subset A ⊂ [p] \ {i} with
|A| = d, we have
(1 − )L∗i (A, Σ) ≤ L∗i (A, Σ̂) ≤ (1 + )L∗i (A, Σ), (20)
−(n−d)2 /8
with probability at least 1 − 2 p−1
.
d e
We show that Proposition 1 follows from the above two
lemmas.
Proof of Proposition 1. Combining Lemma 1 and Lemma 2
and using = κ2 /2 we have for any i ∈ V that the sequence
of inequalities
L∗i (Bi , Σ̂) < (1 + )L∗i (Bi , Σ) < (1 + )(1 − κ2 )L∗i (B̂, Σ)
<
(1 − κ2 )(1 + ) ∗
Li (B̂i , Σ̂) < L∗i (B̂, Σ̂),
1−
is satisfied for all {B̂ ⊂ [p] \ {i} : |B̂| = d, Bi 6⊆ B̂} with
−(n−d)κ4 /32
probability at least 1 − 2 p−1
. To complete
d e
the proof we take the union bound for all i ∈ [p] to get
p − 1 −(n−d)κ4 /32
P ∃i ∈ [p] : Bi 6⊆ B̂i < 2p
e
< δ/2,
d
Assumption-Free and Near-Optimal Sparse Inverse Covariance Selection
≥ 1 − 4dpe−(n−d+1)κ
where the last
inequality follows from n − d
4pd+1
32
log
.
κ4
δ
>
5. Proof of Proposition 2
In this section, we prove the post-processing proposition.
From Proposition 1 we have that for all i ∈ V the estimated
neighborhoods B̂i in Step 1 of S LICE satisfies Bi ⊂ B̂i . The
post processing aims at eliminating the vertices u ∈ B̂i \ Bi
so that we obtain the exact neighborhood. For the rest of
the section we assume that the statement in Proposition 1
holds. Our proof of Proposition 2 is based on showing
that the estimates β̂ij are sufficiently accurate. First we
show that the random quantities
h i β̂ij estimated in step one
θ
have the correct mean E β̂ij = θij
, and are conditionally
ii
distributed as normal random variables. Then we show that
although the variance of the normal random variable β̂ij
is dependent on the conditioning, it can be bounded from
above with high probability. These steps are contained in
the following lemmas.
Lemma 3. Fix i ∈ V. The conditional distribution of β̂ij
for any j ∈ B̂i is given by
Θij −1 −1
,
(21)
,Θ
Σ̂B̂ B̂
β̂ij | Σ̂B̂i B̂i ∼ N
i i jj
Θii ii
where N (., .) denotes the normal distribution.
Lemma 4. For any > 0 the random variable Σ̂−1
B̂ B̂
i
i
jj
satisfies the following inequality
h
i
(n−d+1)2
−1
−1
8
P Σ̂B̂ B̂
> (1 − ) Θjj ≤ e−
. (22)
i
i
jj
We combine the above two lemmas to prove that the estiθ
mated quantities β̂ij are close to θij
.
ii
Lemma 5. Fix any i ∈ V and j ∈ B̂i . The estimates β̂ij
satisfy the following inequality
!
r
2
θij
κ θjj
P |β̂ij −
|≥
≤ 4e−(n−d+1)κ /64 , (23)
θii
4 θii
where κ denotes the minimum normalized edge strength
defined in (3).
We show that Proposition 2 follows from Lemma 5.
Proof of Proposition 2. Using Lemma 5 and using the
union bound for all i ∈ V and j ∈ B̂i , we get
!
r
θij
κ θjj
P |β̂ij −
|≤
∀j ∈ B̂i , i ∈ V
(24)
θii
4 θii
2
/64
(a)
δ
≥ 1− ,
2
(25)
where
the implication
(a) is obtained by using n − d >
8dp
64
in the premise of Proposition 2.
κ2 log
δ
Using (25) for both i and j we get with probability greater
than least 1 − 2δ ,
s
!
!
r
|θij | κ θjj
|θij | κ θii
|β̂ij β̂ji | ≥
(26)
−
−
θii
4 θii
θjj
4 θjj
!2
|θij |
κ
= p
−
.
(27)
4
θii θjj
From
q (27), we get that for (i, j) ∈ E the estimates satisfy
|β̂ij ||β̂ji | ≥ 3κ/4 > κ/2. An identical argument can be
q
used to show that |β̂ij ||β̂ji | ≤ κ/4 < κ/2. This proves
that the post-processing step recovers the exact support.
6. Numerical Simulations
6.1. Illustration of the condition number problem
In this section, we construct a very simple counterexample,
consisting of a sequence of matrices with growing condition
number λλmax
but fixed minimum normalized edge strength
min
κ. The primary purpose of this experiment is to demonstrate that the sample complexity of existing reconstruction
algorithms are indeed sensitive to the condition number of
the precision matrix Θ, whereas the sample complexity dictated by the IT lower bound in (2) as well as our proposed
algorithm S LICE shows no such dependence.
The counter example sequence inspired by (4) consists of
a triangle with two weak links and one stronger link and a
collection of independent nodes. This family of GGMs is
parametrized by the following inverse covariance matrix,
1
κ
κ
0
κ
1
1−
0
,
Θκ,,σ =
κ 1−
1
0
1
0
0
0 σ2 I(p−3)×(p−3)
(28)
where 1 − is the strength of the strong link, κ < 1 −
is the strength of the weak links and σ 2 is the variance of
the independent nodes. This family of graphs are chosen
such that κ in (28) corresponds to the minimum normalized
edge strength in (3). Note that the maximum degree is
d = 2. This problem can be interpreted as detecting a
triangle within a cloud of independent nodes, a situation
that is very plausible in practice.
The simulations are performed for matrix dimension p =
200 and n = 175 samples which satisfies n < p. We repeat
Assumption-Free and Near-Optimal Sparse Inverse Covariance Selection
the reconstruction procedure 50 times with independent
samples for different values of σ 2 ∈ {1, . . . , 104 } while
κ = 0.4 and = 0.01 are fixed. The regularizer parameters in ACLIME, L ASSO and G RAPH L ASSO have been
optimized to yield the best possible results for each value of
σ 2 , an advantage that cannot be availed in practice. S LICE
inherently does not have this issue.
For each algorithm we compute its estimate of the normalized link values (1,2) and (1,4),
r
r
θ12 θ21
θ14 θ41
, γ14 =
.
(29)
γ12 =
θ11 θ22
θ11 θ44
succeed with high-probability and are able to correctly identify that there is a link (1, 2) and no link between (1, 4).
However for larger value of σ 2 , the probability of failure
of ACLIME, L ASSO and G RAPH L ASSO is close to one
while S LICE remains insensitive to changes in σ 2 . This
simple example highlights that when the sample complexity of algorithms depends on parameters not present in the
information theoretic bound, the graph reconstruction can
be adversarially affected even by the presence of additional
independent nodes.
1
Pfailure
0.8
We declare that an algorithm fails to reconstruct the graph
whenever γ12 ≤ γ14 : if this condition is satisfied, then links
(1,2) or/and (1,4) are incorrectly reconstructed regardless
of the thresholding procedure. Note that this choice of reconstruction failure criterion is quite generous. It is very
unlikely that one can devise a successful thresholding procedure solely based on the criterion γ12 ≤ γ14 when γ12
and γ14 are close to each other. This is particularly true
for several reconstructions provided by G RAPH L ASSO and
ACLIME as illustrated in Figure 1, whereas the procedure
appears to provide no advantage to SLICE. Note that we
compare normalized link strengths γij , which are invariant
to rescaling of the Θ matrix, instead of matrix element ratios
βij or matrix elements θij . Thus reconstruction based on
the latter quantities would fail for some rescaling of Θ.
γ12
0.5
SLICE
0.4
0.3
0.3
0.2
0.2
0.1
γ14
0
0
0.1
γ12
1
0.2
0.3
γ12
0.5
0.4
0.4
0.1
γ14
0
0.5
0
0.1
γ12
GRAPH LASSO
0.4
0.8
LASSO
0.2
0.3
0.4
0.5
ACLIME
0.3
0.6
0.2
0.4
0.2
γ14
0
0
0.2
0.4
0.6
0.8
1
0.1
γ14
0
0
0.1
0.2
0.3
2
0.4
√
Figure 1. Illustration of reconstruction failure for σ = 1000
We show the scatter plot of reconstructed values γ12 and γ14 obtained through 50 trial reconstructions. We declare that reconstruction fails when γ12 < γ14 which corresponds to points in the lower
right part of the graphs (highlighted in red). S LICE demonstrates an
almost ideal behavior with γ14 = 0 and γ12 > κ/2 = 0.2. Note
that G RAPH L ASSO and ACLIME systematically yield γ14 > γ12
although (1,4) is not an existing edge.
Simulation results are summarized in Figure 2 where the
probability of failure is plotted against the variance σ 2 of the
independent nodes. For σ 2 close to one, all four algorithms
ACLIME
0.6
GRAPH LASSO
0.4
LASSO
SLICE
0.2
σ2
0
100
101
102
103
104
Figure 2. Detecting a triangle in a cloud of independent nodes.
Empirical probability of failure averaged over 50 trial reconstructions with a fixed number of samples n = 175. The values for all
algorithms except S LICE have been optimized over regularization
coefficients. All algorithms except SLICE fail for large σ 2 values.
6.2. Tests on synthetic and real data
In this section, we present several tests on synthetic and real
data. Our goal is merely to illustrate that the use of modern Mixed-Integer Quadratic Programming (MIQP) solvers
such as Gurobi (Gurobi Optimization, 2016) allows one to
run SLICE in a reasonable time even on relatively large
realistic problems.
As a first test, we run SLICE on synthetic random graph
instances of different degrees (d = 3 and d = 4) and sizes
(p = 10, p = 100 and p = 1000). The link strengths γij
have been randomly generated in the ranges [0.2, 0.4] for
d = 3 and [0.2, 0.3] for d = 4 instances. The family of regular random graphs has been chosen to eliminate potential
dependencies on the heterogeneity in the degree distributions. For implementation, we used one possible MIQP
formulation presented in the Supplementary Material, and
the JuMP framework (Dunning et al., 2017) in julia for running the Gurobi solver. The running times for SLICE with
n = 104 samples for each problem instance are presented
in the Table 1. Notice that the practical scaling of running
times is significantly better than what one would expect
from the worst-case complexity O(pd+1 ) for the full graph
reconstruction.
For the illustration on real data, we use the biological data
set related to the Riboflavin production with B. subtilis.
This data set contains the logarithm of the Riboflavin production rate alongside the logarithms of normalized expres-
Assumption-Free and Near-Optimal Sparse Inverse Covariance Selection
Table 1. Comparison of running times for SLICE on various
regular random graphs with n = 104 samples: MIQP Gurobi
solver time and total running time for reconstruction of the neighborhood of one node, and total time for learning the entire graph.
G RAPH
(p, d)
1 NODE
(G UROBI )
1 NODE
( TOTAL )
F ULL PROBLEM
( TOTAL )
(10, 3)
(10, 4)
(100, 3)
(100, 4)
(1000, 3)
(1000, 4)
0.01 SEC
0.03 SEC
0.03 SEC
0.04 SEC
15.7 SEC
92.3 SEC
2.7 SEC
2.8 SEC
2.7 SEC
2.8 SEC
19.3 SEC
96 SEC
7.2 SEC
7.6 SEC
19.8 SEC
21.7 SEC
18 HOURS
29.3 HOURS
tically impact the reconstruction even for simple models as
illustrated in this paper.
In future work, we would like to perform a more exhaustive
numerical study of S LICE with comparison against state
of art methods including the recently developed “Interaction Screening” algorithm (Vuffray et al., 2016) that was
shown to be optimal for reconstructing discrete graphical
models. On the theoretical side, it would be interesting
to know if the information theoretic lower bound on the
sample complexity is perfectly tight with respect to scaling
in all involved parameters, and if this scaling is efficiently
achievable. This requires either improving the IT bound or
to devise a polynomial time reconstruction algorithm that
closes the gap.
A. Proofs
In this Appendix we prove Lemma 3, 4 and 5. In the process we make heavy use of the properties of the inverse
Wishart distribution based on the properties of the Schur
complement stated in the following lemma.
Lemma 6 ((Ouellette, 1981)). Let X ∈ k×k ∼ W (V, l)
be a random matrix distributed according to the Wishart
distribution with parameter V 0 and order l > k − 1. Let
Y = X −1 with Y ∼ W −1 (U, l) where U = V −1 . Let
X11 X12
V11 V12
X=
, V =
X21 X22
V21 V22
R
Figure 3. Graph learned with SLICE from Riboflavin data set.
This real-world data set (Bühlmann et al., 2014) contains p = 101
variables and n = 71 samples. In the reconstruction procedure,
the maximum degree has been set to d = 6.
sion levels of 100 genes that are most responsive to the
Riboflavin production. Hybridization under different fermentation conditions lead to the acquisition of n = 71
samples, see (Bühlmann et al., 2014) for more details and
raw data. The graph reconstructed with SLICE and constraint d = 6 is depicted in the Figure 3. It took about 2.5
days for the algorithm to learn this graph (with the proof of
optimality of the obtained solution) in this high-dimensional
regime. Notice that again the practical running time for
SLICE using MIQP technology is much lower than the one
required to search over the 1014 candidate neighborhoods
of size d = 6. This example is a perfect illustration of a
trade-off between sample and algorithmic complexity in
real-world problems where the collection of samples might
be very costly.
7. Conclusions and Path Forward
In this paper we propose a polynomial time algorithm, called
S LICE, that perfectly recovers the graph of arbitrary Gaussian graphical models. Contrary to existing algorithms, we
prove that the number of samples needed by S LICE neither
depends on the incoherence properties of the precision matrix nor on its conditioning number, but only on parameters
present in the information theoretic lower-bound. This is
critical when samples are scarce or expensive and can dras-
be any compatible block matrix representation of X and
V . Consider block representations for Y, U with the same
dimensions k1 , k2 that satisfy k1 + k2 = k. Then,
(a) The Schur complement of Y11 is distributed as
−1
−1
−1
Y11 − Y12 Y22
Y21 = X11
∼ W −1 (V11
, l)
−1
= W −1 (U11 − U12 U22
U21 , l),
−1
−1
(b) The random matrix Y22
Y21 conditioned on X11
is
distributed as a matrix normal distribution
−1
−1
−1
−1
−1
Y22
Y21 | X11
∼ N U22
U21 , X11
⊗ U22
.
Proof of Lemma 2. Fix i ∈ V and A ⊂ [p] \ {i} with |A| =
d. Using properties of the Wishart distribution we get that,
−1
2
Σ̂ii − Σ̂iA Σ̂−1
AA Σ̂Ai ∼ (Σii − Σi,A ΣAA ΣAi )χn−d (30)
= L∗i (A, Σ)χ2n−d ,
(31)
where χ2t denotes the standard Chi-squared distribution with
t degrees of freedom. Using the Chernoff bound,
2
P(χ2n−d ≶ 1 ∓ ) < e−(n−d)( 2 − 2 log(1+)) < e−(n−d) /8 .
(32)
1
The proof is completed by using the union bound for all
A ⊂ [p] \ i with |A| = d.
Assumption-Free and Near-Optimal Sparse Inverse Covariance Selection
Proof of Lemma 1. From (13) we have that
(a)
(b)
−1
L∗i (Bi , Σ) = Var(Xi |XBi ) = Var(Xi |X[p]\i ) = θii
,
(33)
Proof of Lemma 4. From Lemma 6, the random matrix
Σ̂B̂i B̂i is distributed according to the Wishart distribution
Σ̂B̂i B̂i ∼ W ΣB̂i B̂i , n . Hence
where (a) follows from the separation property of graphical
models, and (b) follows from (12). Similarly,
h
L∗i (B̂, Σ) = Var(Xi |XB̂ ).
(a)
Using the law of total variance we get that
h
i
Var(Xi |XB̂ ) = E Var(Xi |XBi ∪B̂ )|XB̂
i
h
+ Var E Xi |XBi ∪B̂ |XB̂
X
1
1
θij Xj | XB̂ .
=
+ 2 Var
θii
θii
(34)
Σ̂−1
B̂ B̂
i
i −1
jj
i
= Σ̂jj − Σ̂j(B̂i \j) Σ̂−1
(B̂ \j)(B̂
∼ (Σjj − Σj(B̂i \j) Σ−1
(B̂ \j)(B̂
i
(42)
(35)
αj−1 = Σjj − Σj(B̂i \j) Σ−1
(B̂ \j)(B̂
i
The probability P
h
i
Σ̂−1
B̂ B̂
(a)
−1
= θii
i
≤ e−(n−d+1)
jj
2
/8
(44)
,
(45)
where (a) follows from (43) and (32).
θij Xj | X[p]\{i,u}
j∈Bi ∪B̂
−1
(43)
> (1 − )−1 Θjj
= P χ2n−d+1 < (1 − )αj−1 Θ−1
jj
X
Σ(B̂i \j)j
(36)
j∈Bi ∪B̂
−2
≥ θii
Var
i \j)
= Ψjj − Ψ2ji Ψ−1
ii ≤ Ψjj ≤ Θjj .
i
(a)
Σ(B̂i \j)j )χ2n−d+1
where (a) follows from Lemma 6 and the constant αj can
be bounded from above for every j ∈ B̂i as
Let u ∈ Bi \ B̂. From above, we get
i \j)
Σ̂(B̂i \j)j
, αj χ2n−d+1 ,
j∈Bi ∪B̂
Var(Xi |XB̂ ) − Var(Xi |XBi )
X
−2
θij Xj | XB̂
= θii
Var
i \j)
i
−1 (b)
θii θuu
κ2
. (37)
−
1
≥
Var(X
|X
)
i
B
i
2
θiu
1 − κ2
The inequality (a) follows from the fact that conditioning
reduces variance in Gaussian and observing that B̂ ⊆ [p] \
{i, u}. The inequality (b) follows from (3).
Proof of Lemma 3. For any i ∈ V let Ψ̂(iB̂i )(iB̂i ) =
−1
−1
Σ̂(iB̂i )(iB̂i )
and Ψ(iB̂i )(iB̂i ) = Σ(iB̂i )(iB̂i )
. Using the block matrix decomposition for matrix inverse
Ψ(iB̂i )(iB̂i ) = Θ(iB̂i )(iB̂i ) − Θ(iB̂i )D Θ−1
DD ΘD(iB̂i ) , (38)
where D = [p] \ {i ∪ B̂}. Since Bi ⊆ B̂i , we must have
ΘiD = 0. Hence the matrix Ψ(iB̂i )(iB̂i ) satisfies the following properties for all i and j ∈ Bi :
Ψ(i)(iBi ) = Θ(i)(iBi ) ,
Ψjj ≤ Θjj .
(39)
i
h
Proof of Lemma 5. Define the event E = Σ̂−1
B̂ B̂
≤
(1 − 1 )−1 Θjj . We bound the deviation of β̂ij from
as
i
where (a) follows from (39).
(41)
r
√
√
2
≤ 2Φc 1 − 1 n + e−(n−d+1)1 /8 ,
(a)
(46)
where (a) follows by bounding the first term using Lemma 3
and the definition of the event E, and bounding the second
term by the probability of the event E using Lemma 4.
Setting = κ/4 and 1 = κ/2, and a tight upper bound on
the Gaussian CDF, we get
√
√
2
Φc 1 − 1 n + e−(n−d+1)1 /8
≤ e−κ
(40)
jj
θij
θii
!
!
r
θjj
θjj
θij
= P |β̂ij −
|≥
|E
θii
θii
θii
!
r
θjj
θij
c
|≥
|E
+ P |β̂ij −
θii
θii
θij
P |β̂ij −
|≥
θii
2
From Lemma 6, part (b) we get that
Ψij −1 −1
β̂ij | Σ̂B̂i B̂i ∼ N
,Ψ
Σ̂B̂ B̂
i i jj
Ψii ii
Θij −1
(a)
−1
= N
,Θ
Σ̂B̂ B̂
,
i i jj
Θii ii
i
n/64
2
+ e−(n−d+1)κ
/32
2
≤ 2e−(n−d+1)κ
Combining (46) and (47) we complete the proof.
/64
.
(47)
Assumption-Free and Near-Optimal Sparse Inverse Covariance Selection
References
Anandkumar, Animashree, Tan, Vincent YF, Huang, Furong,
and Willsky, Alan S. High-dimensional gaussian graphical model selection: Walk summability and local separation criterion. Journal of Machine Learning Research, 13
(Aug):2293–2337, 2012.
Basso, Katia, Margolin, Adam A, Stolovitzky, Gustavo,
Klein, Ulf, Dalla-Favera, Riccardo, and Califano, Andrea.
Reverse engineering of regulatory networks in human b
cells. Nature genetics, 37(4):382–390, 2005.
Bertsimas, Dimitris, King, Angela, Mazumder, Rahul, et al.
Best subset selection via a modern optimization lens. The
Annals of Statistics, 44(2):813–852, 2016.
Bühlmann, Peter, Kalisch, Markus, and Meier, Lukas. Highdimensional statistics with a view toward applications in
biology. Annual Review of Statistics and Its Application,
1:255–278, 2014.
Cai, T Tony, Liu, Weidong, and Zhou, Harrison H. Estimating sparse precision matrix: Optimal rates of convergence
and adaptive estimation. The Annals of Statistics, 44(2):
455–488, 2016.
Cai, Tony, Liu, Weidong, and Luo, Xi. A constrained `1
minimization approach to sparse precision matrix estimation. Journal of the American Statistical Association, 106
(494):594–607, 2011.
d’Aspremont, Alexandre, Banerjee, Onureena, and
El Ghaoui, Laurent. First-order methods for sparse covariance selection. SIAM Journal on Matrix Analysis and
Applications, 30(1):56–66, 2008.
Dunning, Iain, Huchette, Joey, and Lubin, Miles. Jump: A
modeling language for mathematical optimization. SIAM
Review, 59(2):295–320, 2017.
Friedman, Nir. Inferring cellular networks using probabilistic graphical models. Science, 303(5659):799–805,
2004.
Gurobi Optimization, Inc. Gurobi optimizer reference manual, 2016. URL http://www.gurobi.com.
Huang, Shuai et al. Learning brain connectivity of
Alzheimer’s disease by sparse inverse covariance estimation. NeuroImage, 50(3):935–949, 2010.
Johnson, Christopher, Jalali, Ali, and Ravikumar, Pradeep.
High-dimensional sparse inverse covariance estimation
using greedy methods. In Artificial Intelligence and Statistics, pp. 574–582, 2012.
Jones, David T, Buchan, Daniel WA, Cozzetto, Domenico,
and Pontil, Massimiliano. Psicov: precise structural contact prediction using sparse inverse covariance estimation
on large multiple sequence alignments. Bioinformatics,
28(2):184–190, 2012.
Kalisch, Markus and Bühlmann, Peter.
Estimating
high-dimensional directed acyclic graphs with the pcalgorithm. Journal of Machine Learning Research, 8
(Mar):613–636, 2007.
Lokhov, Andrey Y, Vuffray, Marc, Misra, Sidhant,
and Chertkov, Michael. Optimal structure and parameter learning of Ising models. arXiv preprint
arXiv:1612.05024, 2017.
Meinshausen, Nicolai and Bühlmann, Peter.
Highdimensional graphs and variable selection with the lasso.
The annals of statistics, pp. 1436–1462, 2006.
Menéndez, Patricia, Kourmpetis, Yiannis AI, ter Braak,
Cajo JF, and van Eeuwijk, Fred A. Gene regulatory networks from multifactorial perturbations using graphical
lasso: application to the dream4 challenge. PloS one, 5
(12):e14147, 2010.
Ouellette, Diane Valerie. Schur complements and statistics.
Linear Algebra and its Applications, 36:187–295, 1981.
Ravikumar, Pradeep, Raskutti, Garvesh, Wainwright, Martin J, and Yu, Bin. Model selection in gaussian graphical
models: High-dimensional consistency of `1 -regularized
MLE. In Advances in Neural Information Processing
Systems 21, pp. 1329–1336, 2009.
Ravikumar, Pradeep, Wainwright, Martin J, Lafferty,
John D, et al. High-dimensional Ising model selection
using `1 -regularized logistic regression. The Annals of
Statistics, 38(3):1287–1319, 2010.
Spirtes, Peter, Glymour, Clark N, and Scheines, Richard.
Causation, prediction, and search. MIT press, 2000.
Tibshirani, Robert. Regression shrinkage and selection via
the lasso. Journal of the Royal Statistical Society. Series
B (Methodological), pp. 267–288, 1996.
Uhler, Caroline, Raskutti, Garvesh, Bühlmann, Peter, Yu,
Bin, et al. Geometry of the faithfulness assumption in
causal inference. The Annals of Statistics, 41(2):436–463,
2013.
Varoquaux, Gael, Gramfort, Alexandre, Poline, JeanBaptiste, and Thirion, Bertrand. Brain covariance selection: better individual functional connectivity models
using population prior. In Advances in Neural Information Processing Systems 23, pp. 2334–2342, 2010.
Assumption-Free and Near-Optimal Sparse Inverse Covariance Selection
Vuffray, Marc, Misra, Sidhant, Lokhov, Andrey, and
Chertkov, Michael. Interaction screening: Efficient and
sample-optimal learning of Ising models. In Advances
in Neural Information Processing Systems 29, pp. 2595–
2603, 2016.
Wang, Lingxiao, Ren, Xiang, and Gu, Quanquan. Precision
matrix estimation in high dimensional gaussian graphical
models with faster rates. In Proceedings of the 19th
International Conference on Artificial Intelligence and
Statistics, pp. 177–185, 2016.
Wang, Wei, Wainwright, Martin J, and Ramchandran, Kannan. Information-theoretic bounds on model selection for
gaussian markov random fields. In IEEE International
Symposium on Information Theory Proceedings (ISIT),
pp. 1373–1377, 2010.
Yuan, Ming and Lin, Yi. Model selection and estimation
in the gaussian graphical model. Biometrika, pp. 19–35,
2007.
| 10math.ST
|
arXiv:1504.05289v2 [math.PR] 20 Jul 2017
Distance-based species tree estimation under
the coalescent: information-theoretic trade-off
between number of loci and sequence length∗
Elchanan Mossel
MIT
[email protected]
Sebastien Roch
UW Madison
[email protected]
July 24, 2017
Abstract
We consider the reconstruction of a phylogeny from multiple genes under the multispecies coalescent. We establish a connection with the sparse
signal detection problem, where one seeks to distinguish between a distribution and a mixture of the distribution and a sparse signal. Using this connection, we derive an information-theoretic trade-off between the number of
genes, m, needed for an accurate reconstruction and the sequence length, k,
of the genes. Specifically,
we show that to detect a branch of length f , one
√
needs m = Θ(1/[f 2 k]) genes.
∗
Keywords: phylogenetic reconstruction, multispecies coalescent, sequence length requirement. S.R. is supported by NSF grants DMS-1007144, DMS-1149312 (CAREER), and an Alfred
P. Sloan Research Fellowship. E.M. is supported by NSF grants DMS-1106999 and CCF 1320105
and DOD ONR grants N000141110140 and N00014-14-1-0823 and grant 328025 from the Simons Foundation. E.M. and S.R. thank the Simons Institute for the Theory of Computing at U.C.
Berkeley where this work was done.
1
Introduction
In the sparse signal detection problem, one is given m i.i.d. samples X1 , . . . , Xm
(m)
and the goal is to distinguish between a distribution P0
(m)
H0
(m)
: Xi ∼ P0 ,
(m)
and the same distribution corrupted by a sparse signal P1
(m)
H1
(m)
: Xi ∼ Q(m) := (1 − σm ) P0
(m)
+ σ m P1 .
Typically one takes σm = m−β , where β ∈ (0, 1). This problem arises in a
number of applications [Dob58, JCL10, CJT05, KHH+ 05]. The Gaussian case
in particular is well-studied [Ing97, DJ04, CJJ11]. For instance it is established
(m)
(m)
in [Ing97,
√ DJ04] that, in the case P0 ∼ N (0, 1) and P1 ∼ N (λm , 1) with
λm = 2r log m, a test with vanishing error probability exists if and only if r
exceeds an explicitly known detection boundary r∗ (β).
In this paper, we establish a connection between sparse signal detection and
the reconstruction of phylogenies from multiple genes or loci under the multispecies coalescent, a standard population-genetic model [RY03]. The latter problem is of great practical interest in computational evolutionary biology and is currently the subject of intense study. See e.g. [LYK+ 09, DR09, ALPE12, Nak13] for
surveys. The problem is also related to the reconstruction of demographic history
in population genetics [MFP08, BS14, KMRR15].
By taking advantage of the connection to sparse signal detection, we derive a
“detection boundary” for the multilocus phylogeny problem and use it to characterize the trade-off between the number of genes needed to accurately reconstruct
a phylogeny and the quality of the signal that can be extracted from each separate
gene. Our results apply to distance-based methods, an important class of reconstruction methods. Before stating our results more formally, we begin with some
background. See e.g. [SS03] for a more general introduction to mathematical
phylogenetics.
Species tree estimation An evolutionary tree, or phylogeny, is a graphical representation of the evolutionary relationships between a group of species. Each
leaf in the tree corresponds to a current species while internal vertices indicate
past speciation events. In the classical phylogeny estimation problem, one sequences a single common gene (or other locus such as pseudogenes, introns, etc.)
2
from a representative individual of each species of interest. One then seeks to
reconstruct the phylogeny by comparing the genes across species. The basic principle is simple: because mutations accumulate over time during evolution, more
distantly related species tend to exhibit more differences between their genes.
Formally, phylogeny estimation boils down to learning the structure of a latent tree graphical model from i.i.d. samples at the leaves. Let T = (V, E, L, r)
be a rooted leaf-labelled binary tree, with n leaves denoted by L = {1, . . . , n}
and a root denoted by r. In the Jukes-Cantor model [JC69], one of the simplest
Markovian models of molecular evolution, we associate to each edge e ∈ E a
mutation probability
(1)
pe = 1 − e−νe te ,
where νe is the mutation rate and te is the time elapsed along the edge e. (The
analytical form of (1) derives from a continuous-time Markov process of mutation
along the edge. See e.g. [SS03].) The Jukes-Cantor process is defined as follows:
• Associate to the root a sequence sr = (sr,1 , . . . , sr,k ) ∈ {A, C, G, T}k of
length k where each site sr,i is uniform in {A, C, G, T}.
• Let U denote the set of children of the root.
• Repeat until U = ∅:
– Pick a u ∈ U .
– Let u− be the parent of u.
– Associate a sequence su ∈ {A, C, G, T}k to u as follows: su is obtained
from su− by mutating each site in su− independently with probability
p(u− ,u) ; when a mutation occurs at a site i, replace su,i with a uniformly
chosen state in {A, C, G, T}.
– Remove u from U and add the children (if any) of u to U .
Let T −r be the tree T where the root is suppressed, i.e., where the two edges
adjacent to the root are combined into a single edge. We let L[T, (pe )e , k] be
the distribution of the sequences at the leaves s1 , . . . , sn under the Jukes-Cantor
process. We define the single-locus phylogeny estimation problem as follows:
Given sequences at the leaves (s1 , . . . , sn ) ∼ L[T, (pe )e , k], recover
the (leaf-labelled) unrooted tree T −r .
3
(One may also be interested in estimating the pe s, but we focus on the tree.
The root is in general not identifiable.) This problem has a long history in evolutionary biology. A large number of estimation techniques have been developed. See e.g. [Fel04]. For a survey of the learning perspective on this problem, see e.g. [MSZ+ 13]. On the theoretical side, much is known about the sequence length—or, in other words, the number of samples—required for a perfect reconstruction with high probability, including both information-theoretic
lower bounds [SS02, Mos03, Mos04, MRS11] and matching algorithmic upper
bounds [ESSW99a, DMR11a, DMR11b, Roc10]. More general models of molecular evolution have also been considered in this context; see e.g. [ESSW99b,
CGG02, MR06, DR13, ADHR12].
Nowadays, it is common for biologists to have access to multiple genes—or
even full genomes. This abundance of data, which on the surface may seem like a
blessing, in fact comes with significant challenges. See e.g. [DBP05, Nak13] for
surveys. One important issue is that different genes may have incompatible evolutionary histories—represented by incongruent gene trees. In other words, if one
were to solve the phylogeny estimation problem separately for several genes, one
may in fact obtain different trees. Such incongruence can be explained in some
cases by estimation error, but it can also result from deeper biological processes
such as horizontal gene transfer, gene duplications and losses, and incomplete
lineage sorting [Mad97]. The latter phenomenon, which will be explained in Section 2, is the focus of this paper.
Accounting for this type of complication necessitates a two-level hierarchical
model for the input data. Let S = (V, E, L, r) be a rooted leaf-labelled binary
species tree, i.e., a tree representing the actual succession of past divergences for
a group of organisms. To each gene j shared by all species under consideration,
we associate a gene tree Tj = (Vj , Ej , L), mutation probabilities (pje )e∈Ej , and
sequence length kj . The triple (Tj , (pje )e∈Ej , kj ) is picked at random according
to a given distribution G[S, (νe , te )e∈E ] which depends on the species tree, mutation parameters νe and inter-speciation times te . It is standard to assume that
the gene trees are conditionally independent given the species tree. In the context
of incomplete lineage sorting, the distribution of the gene trees, G, is given by
the so-called multispecies coalescent, which is a canonical model for combining
speciation history and population genetic effects [RY03]. (Readers familiar with
the multispecies coalescent may observe that our model is a bit richer than the
standard model, as it includes mutational parameters in addition to branch length
information. Note that we also incorporate sequence length in the model.) The
detailed description of the model is deferred to Section 2, as it is not needed for
4
a high-level overview of our results. For the readers not familiar with population
genetics, it is useful to think of Tj as a noisy version of S (which, in particular,
may result in Tj having a different (leaf-labelled) topology than S).
Our two-level model of sequence data is then as follows. Given a species tree
S, parameters (νe , te )e∈E and a number of genes m:
1. [First level: gene trees] Pick m independent gene trees and parameters
(Tj , (pje )e∈Ej , kj ) ∼ G[S, (νe , te )e∈E ],
j = 1, . . . , m.
2. [Second level: leaf sequences] For each gene j = 1, . . . , m, generate sequence data at the leaves L according to the (single-locus) Jukes-Cantor
process, as described above,
(sj1 , . . . , sjn ) ∼ L[Tj , (pje )e , kj ],
j = 1, . . . , m,
independently of the other genes.
We define the multi-locus phylogeny estimation problem as follows:
Given sequences at the leaves (sj1 , . . . , sjn ), j = 1, . . . , m, generated
by the process above, recover the (leaf-labelled) unrooted species tree
S −r .
In the context of incomplete lineage sorting, this problem is the focus of very active research in statistical phylogenetics [LYK+ 09, DR09, ALPE12, Nak13]. In
particular, there is a number of theoretical results, including [DR06, DDBR09,
DD10, MR10, LYP10, ADR11, Roc13, DNR15, RS15, RW15]. However, many
of these results concern the statistical properties (identifiability, consistency, convergence rate) of species tree estimators that (unrealistically) assume perfect knowledge of the Tj s. A very incomplete picture is available concerning the properties
of estimators based on sequence data, i.e., that do not require the knowledge of
the Tj s. (See below for an overview of prior results.)
Here we consider the data requirement of such estimators based on sequences.
To simplify, we assume that all genes have the same length, i.e., that kj = k for
all j = 1, . . . , m for some k. (Because our goal is to derive a lower bound, such
simplification is largely immaterial.) Our results apply to an important class of
methods known as distance-based methods, which we briefly describe now. In
the single-locus phylogeny estimation problem, a natural way to infer T −r is to
5
use the fraction of substitutions between each pair, i.e., letting k · k1 denote the
`1 -distance,
θ(sa , sb ) := ksa − sb k1 ,
∀a, b ∈ [n].
(2)
We refer to reconstruction methods relying solely on the θ(sa , sb )s as distancebased methods. Assume for instance that νe = ν for all e, i.e., the so-called
molecular clock hypothesis. Then it is easily seen that single-linkage clustering
(e.g., [HTF09]) applied to the distance matrix (θ(sa , sb ))a,b∈[n] converges to T −r
as k → +∞. (In this special case, the root can be recovered as well.) In fact, T
can be reconstructed perfectly as long as, for each a, b, k1 θ(sa , sb ) is close enough
to its expectation (e.g. [SS03])
3
θa,b := k −1 E[θ(sa , sb )] = (1 − e−dab )
4
with dab :=
X
νe te ,
e∈P (a,b)
where P (a, b) is the edge set on the unique path between a and b in T . Here “close
enough” means O(f ) where f := mine νe te . This observation can been extended
to general νe s. See e.g. [ESSW99a] for explicit bounds on the sequence length
required for perfect reconstruction with high probability.
Finally, to study distance-based methods in the multi-locus case, we restrict
ourselves to the following multi-locus distance estimation problem:
Given an accuracy ε > 0 and distance matrices θ(sja , sjb )a,b∈[n] , j =
1, . . . , m, estimate dab as defined above within ε for all a, b.
Observe that, once the dab s are estimated within sufficient accuracy, i.e., within
O(f ), the species tree can be reconstructed using the techniques referred to in the
single-locus case.
Our results How is this related to the sparse signal detection problem? Our
main goal is to provide a lower bound on the amount of data required for perfect
reconstruction, in terms of m (the number of genes) and k (the sequence length).
Consider the three possible (rooted, leaf-labelled) species trees with three leaves,
as depicted in Figure 1, where we let the time to the most recent divergence be
1 − f (from today) and the time to the earlier divergence be 1. Thus f is the
time between the two divergence events. In order for a distance-based method
to distinguish between these three possibilities, i,e., to determine which pair is
closest, we need to estimate the dab s within O(f ) accuracy. Put differently, within
the multi-locus distance estimation problem, it suffices to establish a lower bound
6
1
2
1
3
3
2
f
1
2
3
1
Figure 1: Three species trees.
on the data required to distinguish between a two-leaf species tree S with d12 = 2
and a two-leaf species tree S + with d12 = 2 − 2f , where in both cases νe = 1 for
all e. We are interested in the limit f → 0.
Let P0 and Q be the distributions of θ(s11 , s12 ) for a single gene under S and S +
respectively, where for ease of notation the dependence on k is implicit. For m
genes, we denote the corresponding distributions by P⊗m
and Q⊗m . To connect
0
the problem to sparse signal detection we observe below that, under the multispecies coalescent, Q is in fact a mixture of P0 and a sparse signal P1 , i.e.,
Q = (1 − σf ) P0 + σf P1 ,
(3)
where σf = O(f ) as f → 0.
When testing between P⊗m
and Q⊗m , the optimal sum of Type-I (false posi0
tive) and Type-II (false negative) errors is given by (see, e.g., [CT91])
⊗m
inf {P⊗m
(Ac )} = 1 − kP⊗m
− Q⊗m kTV ,
0 (A) + Q
0
A
(4)
where k · kTV denotes the total variation distance. Because σf = O(f ), for any
k, in order to distinguish between P0 and Q one requires that, at the very least,
m = Ω(f −1 ). Otherwise the probability of observing a sample originating from
P1 under Q is bounded away from 1. In [MR10] it was shown that, provided that
k = Ω(f −2 log f −1 ), m = Ω(f −1 ) suffices. At the other end of the spectrum,
when k = O(1), a lower bound for the single-locus problem obtained by [SS02]
7
implies that m = Ω(f −2 ) is needed. An algorithm achieving this bound under the
multispecies coalescent was recently given in [DNR15].
We settle the full spectrum between these two regimes. Our results apply when
k = f −2+2κ and m = f −1−µ where 0 < κ, µ < 1 as f → 0.
Theorem 1 (Lower bound). For any δ > 0, there is a c > 0 such that
− Q⊗m kTV ≤ δ,
kP⊗m
0
whenever
m≤c
1
√ .
f2 k
Notice that the lower bound on m interpolates between the two extreme regimes discussed above. As k increases, a more accurate estimate of the gene trees
can be obtained and one expects that the number of genes required for perfect
reconstruction should indeed decrease. The form of that dependence is far from
clear however. We in fact prove that our analysis is tight.
Theorem 2 (Matching upper bound). For any δ > 0, there is a c0 > 0 such that
kP⊗m
− Q⊗m kTV ≥ 1 − δ,
0
whenever
m ≥ c0
1
√ .
k
f2
Moreover, there is an efficient test to distinguish between P⊗m
and Q⊗m in that
0
case.
Our proof of the upper bound actually gives an efficient reconstruction algorithm under the molecular clock hypothesis. We expect that the insights obtained
from proving Theorem 1 and 2 will lead to more accurate practical methods as
well in the general case.
Our results were announced without proof in abstract form in [MR15].
Proof sketch Let Z be an exponential random variable with mean 1. We first
show that, under P0 (respectively
Q), θ(s11 , s12 ) is binomial with k trials and success
3
−2(ζ+Z)
, where ζ = 1 (respectively ζ = 1 − f ). Equation (3)
probability 4 1 − e
then follows from the memoryless property of the exponential, where σf is the
probability that Z ≤ f .
8
A recent result of [CW14] gives a formula for the detection boundary of the
sparse signal detection problem for general P0 , P1 . However, applying this formula here is non-trivial. Instead we bound directly the total variation distance between P⊗m
and Q⊗m . Similarly to the approach used in [CW14], we work with the
0
⊗m
Hellinger distance H 2 (P⊗m
) which tensorizes as follows (see e.g. [CT91])
0 ,Q
m
1 2
1 2 ⊗m ⊗m
H (P0 , Q ) = 1 − 1 − H (P0 , Q)
,
(5)
2
2
and further satisfies
kP⊗m
0
−
Q⊗m k2TV
≤H
2
⊗m
(P⊗m
)
0 ,Q
1 2 ⊗m ⊗m
1 − H (P0 , Q ) .
4
(6)
All the work is in proving that, as f → 0,
√
H 2 (P0 , Q) = O f 2 k .
The details are in Section 3.
The proof of Theorem 2 on the other hand involves the construction of a statistical test that distinguishes between P⊗m
and Q⊗m . In the regime k = O(1), an
0
optimal test (up to constants) compares the means of the samples [DNR15]. See
also [LYPE09] for a related method (without sample complexity). In the regime
k = ω(f −2 ), an optimal test (up to constants) compares the minima of the samples [MR10]. A natural way to interpolate between these two
√ tests is to consider
an appropriate quantile. We show that a quantile of order 1/ k leads to the optimal choice.
Organization. The gene tree generating model is defined in Section 2. The
proof of Theorem 1 can be found in Section 3. The proof of Theorem 2 can be
found in Section 4.
2
Further definitions
In this section, we give more details on the model.
9
1
2
3
Figure 2: An incomplete lineage sorting event. Although 1 and 2 are more closely
related in the species tree (fat tree), 2 and 3 are more closely related in the gene tree
(thin tree). This incongruence is caused by the failure of the lineages originating
from 1 and 2 to coalesce within the shaded branch.
Some coalescent theory As we mentioned in the previous section, our gene
tree distribution model G[S, (νe , te )e∈E ] is the multispecies coalescent [RY03].
We first explain the model in the two-species case. Let 1 and 2 be two species and
consider a common gene j. One can trace back in time the lineages of gene j from
an individual in 1 and from an individual in 2 until the first common ancestor. The
latter event is called a coalescence. Here, because the two lineages originate from
different species, coalescence occurs in an ancestral population. Let τ be the time
of the divergence between 1 and 2 (back in time). Then, under the multispecies
coalescent, the coalescence time is τ + Z where Z is an exponential random
variable whose mean depends on the effective population size of the ancestral
population. Here we scale time so that the mean is 1. (See, e.g., [Dur08] for an
introduction to coalescent theory.)
We get for the two-level model of sequence data:
Lemma 1 (Distance distribution). Let S be a two-leaf species tree with d12 = 2τ
and νe = 1 for all e and let θ(s11 , s12 ) be as in (2) for some k. Then the (ran1 1
dom) distribution
of θ(s1 , s2 ) is binomial with k trials and success probability
3
−2(τ +Z)
1−e
.
4
The memoryless property of the exponential gives:
Lemma 2 (Mixture). Let S be a two-leaf species tree with d12 = 2 and let S + be
a two-leaf species tree with d12 = 2 − 2f , where in both cases νe = 1 for all e.
10
Let P0 and Q be the distributions of θ(s11 , s12 ) for a single gene under S and S +
respectively. Then, there is P1 such that,
Q = (1 − σf ) P0 + σf P1 ,
where σf = O(f ), as f → 0.
Proof. The proof of the lemma is straightforward: We couple perfectly the coalescence time for Q conditioned on Z ≥ f and the unconditional coalescence
time for P0 and this extends to a coupling of the distances between the sequences.
Thus P1 is obtained by conditioning Q on the event that Z is ≤ f and σf is the
probability of that event.
More generally (this paragraph may be skipped as it will not play a role below), consider a species tree S = (V, E; L, r) with n leaves. Each gene j =
1, . . . , m has a genealogical history represented by its gene tree Tj distributed according to the following process: looking backwards in time, on each branch of
the species tree, the coalescence of any two lineages is exponentially distributed
with rate 1, independently from all other pairs; whenever two branches merge in
the species tree, we also merge the lineages of the corresponding populations, that
is, the coalescence proceeds on the union of the lineages. More specifically, the
probability density of a realization of this model for m independent genes is
e −O e
h
e h
j
i
i IjY
`
Oj
e,Oje
e,Oje +1
e,`
e,`−1
σj − σj
exp −
,
− σj
σj
exp −
2
2
j=1 e∈E
`=1
m Y
Y
where, for gene j and branch e, Ije is the number of lineages entering e, Oje is
the number of lineages exiting e, and σje,` is the `th coalescence time in e; for
e,I e −Oe +1
convenience, we let σje,0 and σj j j be respectively the divergence times of
e and of its parent population. The resulting trees Tj s may have topologies that
differ from that of the species tree S. This may occur as a result of an incomplete
lineage sorting event, i.e., the failure of two lineages to coalesce in a population.
See Figure 2 for an illustration.
A more abstract setting Before proving Theorem 1, we re-set the problem in
a more generic setting that will make the computations more transparent. Let P0
and P1 denote two different distributions for a random variable X supported on
[0, 1]. Given these distributions, we define two distributions, which we will also
11
denote by P0 and P1 , for a random variable θ taking values in {0, . . . , k} for some
k. These are defined by
k
Pi [θ = `] =
Ei [X ` (1 − X)k−` ],
(7)
`
where Ei is the expectation operator corresponding to Pi for the random variable
X defined on [0, 1]. As before, we let
Q = (1 − σf ) P0 + σf P1 ,
for some σf = O(f ). We make the following assumptions which are satisfied in
the setting of the previous section.
A1. Disjoint supports: X admits a density whose support is (p0 , p0 ) under P0
and (p0 − φf , p0 ) under P1 , where 0 < p0 < p0 < 1 (independent of f )
and φf = O(f ). (In the setting of Lemma 2, p0 = 43 (1 − e−2 ), p0 − φf =
3
(1 − e−(2−2f ) ), and p0 = 3/4.)
4
A2. Bounded density around p0 : There exist ρ ∈ (0, 1) and p̄ ∈ (p0 , p0 ),
not depending on f , such that the following holds. Under P0 , the density
of X on (p0 , p̄) lies in the interval [ρ, ρ−1 ], i.e., for any measurable subset
X ⊆ (p0 , p̄) we have
P0 [X ∈ X ] ∈ ρ|X |, ρ−1 |X | ,
where |X | is the Lebesgue measure of X . (In the setting of Lemma 2, under
1/2
P0 the density of X on (p0 , p0 ) is 4e3 (1 − 4x/3)−3/4 . Notice that this
density is not bounded from below over the entire interval (p0 , p0 ).)
The first assumption asserts that the supports of X under P0 and P1 are disjoint,
while also being highly concentrated under P1 (as f → 0). The key point being
that, under P1 , X lies near the lower end of the support under P0 , which partly
explains the effectiveness of a quantile-based test to distinguish between P0 and
Q. The second, more technical, assumption asserts that, under P0 , the density of
X is bounded from above and below in a neighborhood of the lower end of its
support. As we will see in Section 3, the dominant contribution to the difference
between P0 and Q comes from the regime where X lies close to p0 and we will
need to control the probability of observing X there.
12
3
Lower bound
The proof of the lower bound is based on establishing an upper bound on the
Hellinger distance between P0 and Q. The tensoring property of the Hellinger
distance then allows to directly obtain an upper bound on the Hellinger distance
between P⊗m
and Q⊗m . Using a standard inequality, this finally gives the desired
0
bound on the total variation distance between P⊗m
and Q⊗m .
0
We first rewrite the Hellinger distance in a form that is convenient for asymptotic expansion. In the abstract setting of Section 2, the Hellinger distance can be
written as
k h
i2
X
p
p
Q[θ = j] − P0 [θ = j]
H (P0 , Q) =
2
j=0
#2
P1 [θ = j]
=
1 + σf
− 1 − 1 P0 [θ = j]
P0 [θ = j]
j=0
"s
#2
k
X
E1 [X j (1 − X)k−j ]
=
1 + σf
− 1 − 1 P0 [θ = j]
j (1 − X)k−j ]
E
[X
0
j=0
k
X
E1 [X j (1 − X)k−j ]
=
hσf
P0 [θ = j],
(8)
E0 [X j (1 − X)k−j ]
j=0
k
X
"s
where we define
p
hb (s) := ( 1 + b(s − 1) − 1)2 .
(9)
We will refer to
E1 [X j (1 − X)k−j ]
=
E0 [X j (1 − X)k−j ]
k
E1 [X j (1
j
k
E0 [X j (1
j
− X)k−j ]
− X)k−j ]
=
P1 [θ = j]
,
P0 [θ = j]
as the likelihood ratio and to
P0 [θ = j],
as the null probability.
We prove the following proposition, which implies Theorem 1.
Proposition 1. Assume that k = f −2+2κ where 0 < κ < 1 and that Assumptions
A1 and A2 hold. As f → 0,
√
H 2 (P0 , Q) = O f 2 k .
13
The proof of Proposition 1 follows in the next section.
Finally:
Proof of Theorem 1. The tensorization property of the Hellinger distance, as stated
in (5), together with Proposition 1 imply that
m
√ m
1 2
1 2 ⊗m ⊗m
< δ,
H (P0 , Q ) = 1 − 1 − H (P0 , Q)
= 1 − 1 − O f2 k
2
2
if m ≤ cf −2 k −1/2 for a small enough constant c. Thus, by (6), we have
1 2 ⊗m ⊗m
⊗m
⊗m 2
2
⊗m
⊗m
kP0 − Q kTV ≤ H (P0 , Q ) 1 − H (P0 , Q ) < δ,
4
as needed.
3.1
Proof of Proposition 1
From (8), in order to bound the Hellinger distance from above, we need upper
j (1−X)k−j ]
and on the null probability P0 [θ = j]
bounds on the likelihood ratio EE10 [X
[X j (1−X)k−j ]
for each term in the sum. The basic intuition is that the contributions of those
terms where θ is far from its mean under P1 (which is ≈ p0 ) are negligible. Indeed:
• When θ is much smaller than p0 , the null probability is negligible because,
under P0 , X is almost surely greater than p0 . We establish
√ that this leads to
2
an overall contribution to the Hellinger distance of o(f k). See (21).
• When θ is much larger than p0 , the likelihood ratio is negligible because X
has a much broader support under P0 than it does under P1 . In that case, we
show that the overall contribution to the Hellinger distance is O(f 2 ). To get
a sense of why that is, note that as f → 0
p
hσf (0) = [ 1 − σf − 1]2 = O(f 2 ).
See Claims 1, 2 and 3.
On the other hand, by (7), under both P0 and P1 , the random variable θ √
conditioned
on X is binomial with mean kX and standard deviation of order k. In the
regime considered under Theorem 1, i.e., k = f −2+2κ , we have further that f =
14
√
o(1/ k). Hence by Assumption A1, under P1 , X has support of size O(f
√) and
the unconditional random variable θ also has standard deviation of order k. In
this bulk regime, our analysis relies on the following insight:
• How big is each term in the Hellinger sum? In order√for E0 [X j (1 − X)k−j ]
to be non-negligible, X must lie within
√ roughly k of p0 , which under
Assumption A2 has probability Θ(1/ k). On the other hand, under P1√
,X
k.
is almost surely close to p√
0 . That produces a likelihood ratio of order
Therefore, recalling that f k = o(1), the term
hσf
j
k−j
E1 [X (1 − X) ]
E0 [X j (1 − X)k−j ]
"s
=
1 + σf
#2
E1 [X j (1 − X)k−j ]
−1 −1 ,
E0 [X j (1 − X)k−j ]
is of order f 2 k. Moreover, by the
√argument above, the overall null probability of the bulk is of order 1/√ k. Thus, we expect that the Hellinger
distance in this regime is of order kf 2 as stated in Proposition 1. It will be
convenient to divide the analysis into θ-values below p0 (see Claims 4, 5, 6
and 7) and above p0 (see Claims 8, 9, 10 and 11).
The full details are somewhat delicate, as we need to carefully consider various
intervals of summands j according to the behavior of the null probability P0 [θ =
j (1−X)k−j ]
.
j] and the likelihood ratio EE10 [X
[X j (1−X)k−j ]
In the next subsection we introduce some notation and prove some simple
estimates that will be used in the proofs.
3.2
Some useful lemmas
The following is Lemma 4 in [CW14]:
p
Lemma 3. For b > 0, let hb (s) = ( 1 + b(s − 1) − 1)2
1. For any b > 0, the function hb (s) is strictly decreasing on [0, 1] and strictly
increasing on [1, +∞).
2. For any b > 0 and s ≥ 1,
hb (s) ≤ [b(s − 1)] ∧ [b(s − 1)]2 ≤ [bs] ∧ [bs]2 .
The following lemmas follow from straightforward calculus.
15
Lemma 4. For j ∈ {0, . . . , k} and x ∈ (0, 1), let
Φj (x) =
Then
Φ0j (x)
k−j
j
log x +
log(1 − x).
k
k
j 1 k−j 1
1
=
−
=
kx
k 1−x
x(1 − x)
j
−x .
k
As a result Φj is increasing on [0, kj ] and decreasing on [ kj , 1], and Φ0j ( kj ) = 0.
Lemma 5. For j ∈ {0, . . . , k}, p ∈ (0, 1), and x ∈ [0, p), let
Ψj,p (x) =
p
k−j
1−p
j
log
+
log
.
k
p−x
k
1−p+x
Then:
1. The first two derivatives are:
Ψ0j,p (x)
j 1
k−j
1
1
=
−
=
kp−x
k 1−p+x
(p − x)(1 − p + x)
j
− (p − x) ,
k
and
Ψ00j,p (x)
j
=
k
1
(p − x)2
k−j
+
k
1
(1 − p + x)2
1
≥ ,
2
(since the terms in curly brackets are at least 1 and one of kj or k−j
is greater
k
or equal than 1/2).
2. By a Taylor expansion around x = 0, we have for x ∈ [0, p) and some
x∗ ∈ [0, x]
1
j
x2 00 ∗
1
j
1
Ψj,p (x) =
− p x+ Ψj,p (x ) ≥
− p x+ x2 .
p(1 − p) k
2
p(1 − p) k
4
3.3
Proof
Let C be a large constant (not depending on f ) to be determined later. We divide
up the sum in (8) into intervals with distinct behaviors. We consider the following
intervals for kj :
#
"
#
"
r
r
log k
log k
J0 = p0 , p0 + C
,
J1 = p0 + C
,1 ,
k
k
16
and
J00 = [p0 − φf , p0 ] ,
"
#
r
log k
J10 = p0 − C
, p0 − φf ,
k
"
#
r
log k
J20 = 0, p0 − C
.
k
In words J10 ∪ J00 ∪ J0 is the bulk of P1 , i.e., where j/k sampled from P1 takes
its typical values, with J00 being the support of X under P1 . (This bulk interval is
further sub-divided into three intervals whose analyses are slightly different.) The
intervals J20 and J1 are where j/k takes atypically small and large values under
P1 respectively. For a subset of kj -values J, we write the contribution of J to the
Hellinger distance as
X
E1 [X j (1 − X)k−j ]
2
H (P0 , Q)|J =
hσf
P0 [θ = j].
E0 [X j (1 − X)k−j ]
j:j/k∈J
Below, it will be convenient to break up the analysis into three regimes: J1 , which
we refer to as the high-substitution regime; J20 ∪ J10 ∪ J00 , the low-substitution
regime; and J0 , the border regime. (Refer back to Section 3.1 for an overview
of the proof in these different regimes. Note in particular that we combine the
analyses of the atypically low values, J20 , and the typical values below p0 , J10 ∪ J00 ,
because they follow from related derivations.)
High substitution regime We consider J1 first. As we previewed in Section 3.1,
the argument in this case involves proving that the likelihood ratio is small. Let
E1 [X j (1 − X)k−j ]
J≤1 = 0 ≤ j ≤ k :
≤1 .
(10)
E0 [X j (1 − X)k−j ]
I.e., J≤1 is where the likelihood ratio is bounded by 1. Note that Lemma 3
in Section 3.2 says that hσf is monotone decreasing in the interval [0, 1] and
we can therefore bound the sum of terms in J≤1 assuming the likelihood ratio
17
E1 [X j (1−X)k−j ]
E0 [X j (1−X)k−j ]
in fact equals 0, as follows,
#2
E1 [X j (1 − X)k−j ]
− 1 − 1 P0 [J = j]
1 + σf
j
k−j ]
E
0 [X (1 − X)
j∈J≤1
X p
2
≤
1 − σf − 1 P0 [J = j]
"s
X
j∈J≤1
= O(σf2 )
= O(f 2 ).
We have thus proved the following claim.
Claim 1 (Ratio less than 1).
H 2 (P0 , Q)|J≤1 = O(f 2 ).
Hence, to bound the sum in J1 , it suffices to show that J1 ⊆ J≤1 , which we prove
in the next claim.
Claim 2 (High substitution implies ratio less than 1). It holds that J1 ⊆ J≤1 .
Since the support of X under P1 is below p0 while it is above p0 under P0 , we
might expect that the likelihood ratio will be bounded by 1 on J1 , which is what
we prove next.
Proof. By Assumption A1, under P1 , X is a.s. less than p0 . Since Lemma 4
implies that Φj (x) is monotone increasing on [0, j/k], which includes [0, p0 ] since
j
∈ J1 , it follows that
k
E1 [X j (1 − X)k−j ] = E1 [exp(kΦj (X))]
≤ exp(kΦj (p0 ))
= pj0 (1 − p0 )k−j .
Let E be the event that
(
E=
"
X ∈ p0 + C
r
log k 1
− , p0 + C
k
k
18
r
log k
k
(11)
#)
.
By Assumption A2, P0 [E] ≥ ρ/k. Hence, using Lemma 4 again, for
j
k
∈ J1
E0 [X j (1 − X)k−j ] = E0 [X j (1 − X)k−j | E]P0 [E] + E0 [X j (1 − X)k−j | E c ]P0 [E c ]
ρ j
≥
p (1 − p)k−j ,
(12)
k
q
where p = p0 + C logk k − k1 .
Combining (11) and (12), and using Lemma 5 with x = p − p0 ≥ 0 (for k > 1
and C large enough), we have
E0 [X j (1 − X)k−j ]
ρpj (1 − p)k−j
≥
E1 [X j (1 − X)k−j ]
kpj0 (1 − p0 )k−j
ρ
exp (kΨj,p (p − p0 ))
=
k
1
ρ
j
1 2
≥
exp k
−p x+ x
k
p(1 − p) k
4
!2
r
ρ
k
log k 1
exp
C
−
≥
k
4
k
k
2
ρ
C
≥
exp
log k
k
5
≥ 1,
for C large enough (assuming k is large), where on the fourth line we used that
q
j/k − p ≥ 0 for j/k ∈ J1 = p0 + C logk k , 1 . We have thus established
J1 ⊆ J≤1
Combining Claims 1 and 2, we thus obtain:
Claim 3 (High substitution: Hellinger distance).
H 2 (P0 , Q)|J1 = O(f 2 ).
Low substitution regime In order to estimate the sum in J00 ∪ J10 ∪ J20 we need
to further subdivide it into intervals of doubling length. The basic intuition is
that for far enough intervals the null probabilities P0 [θ = j] are small enough so
[X j (1−X)k−j ]
we can estimate the likelihood ratio term EE01 [X
j (1−X)k−j ] by its worst value in the
19
interval.
However, when the intervals are close to the mean, the fluctuations in
E1 [X j (1−X)k−j ]
are too big so we need to work with shorter intervals. The partition
E0 [X j (1−X)k−j ]
is defined as follows:
I00
I`0
= p0 −
= p0 −
1
√ , p0
k
`
2
2`−1
√ , p0 − √
,
k
k
` ≥ 1.
√
Define L by 2L = C log k (where we may choose C so that it is integer-valued).
We first upper bound E1 [X j (1 − X)k−j ] using Lemma 4 and Assumption A1:
• On J00 ,
E1 [X j (1 − X)k−j ] = E1 [exp(kΦj (X))]
≤ E1 [exp(kΦj (j/k))]
= (j/k)j (1 − j/k)k−j .
(13)
• On J10 ∪ J20 we have that X ≥ p0 − φf a.s. and therefore
E1 [X j (1 − X)k−j ] ≤ (p0 − φf )j (1 − p0 + φf )k−j .
(14)
To lower bound E0 [X j (1 − X)k−j ], we consider the event
(
"
r #)
1
.
E = X ∈ p0 , p0 +
k
By Assumption A2 and Lemma 4, on J00 ∪ J10 ∪ J20 , arguing as in (12),
ρ
E0 [X j (1 − X)k−j ] ≥ √ pj (1 − p)k−j ,
(15)
k
q
where p = p0 + k1 (assuming k is large). Combining (13), (14), and (15), and
using Lemma 5:
• On J00 ,
E0 [X j (1 − X)k−j ]
E1 [X j (1 − X)k−j ]
≥
ρ
√ exp(kΨj,p (p − j/k))
k
ρ
1
≥ √ exp k −
p(1 − p)
k
ρ
≥ C10 √ ,
k
20
2
2 !!
1 j
j
−p +
−p
k
4 k
for some constant C10 (not depending on f ), where we used that φf
2
so that kj − p = O(1/k) and, further, p(1 − p) ∈ (0, 1/4).
p
1/k
• On J10 ∪ J20 ,
E0 [X j (1 − X)k−j ]
E1 [X j (1 − X)k−j ]
ρ
≥ √ exp(kΨj,p (p − p0 + φf ))
k
ρ
j
1
1
2
≥ √ exp k
− p (p − p0 + φf ) + (p − p0 + φf )
p(1 − p) k
4
k
√
ρ
j
≥ √ exp −C1 k p −
k
k
!!
r
√
ρ
1
j
= √ exp −C1 k p0 +
−
k k
k
√
ρ
j
= C2 √ exp −C1 k p0 −
,
k
k
for some
on f ), where again we used that
p constants C1 , C2 (not depending
2
φf 1/k so that (p − p0 + φf ) = O(1/k 2 ).
By decreasing C2 appropriately we combine the two bounds into:
Claim 4 (Low substitution: Likelihood ratio). For all j/k ∈ J00 ∪ J10 ∪ J20 ,
√
√
E1 [X j (1 − X)k−j ]
k
j
≤
exp C1 k p0 −
.
E0 [X j (1 − X)k−j ]
C2
k
(16)
We now bound the integrand in H 2 (P0 , Q) over J00 ∪ J10 ∪ J20 . As noted after
the definition of J≤1 in equation (10), Lemma 3 implies that on J≤1
E1 [X j (1 − X)k−j ]
hσf
≤ C0 f 2 ,
(17)
E0 [X j (1 − X)k−j ]
for some constant C0 > 0.
• On J00 ∪ J10 , we will further use Lemma 3 (Part 2) which, recall, says that
for s ≥ 1 and b > 0
hb (s) ≤ [b(s − 1)] ∧ [b(s − 1)]2 ≤ [bs] ∧ [bs]2 .
21
In particular observe that, if s ≥ 1, b > 0 and bs < 1, then we have simply
hb (s) ≤ [bs]2 . Here b = σf and s is bounded above by the expression
in (16). We show first that bs is therefore small. Indeed,
√
p
√
k
j
exp C1 k p0 −
= O(f κ ) exp O( log f −1 ) = o(1).
σf
C2
k
Hence, for those j/k-values where the likelihood ratio is bounded below by
1, we have by Lemma 3 (Part 2) that
√
σf2 k
j
E1 [X j (1 − X)k−j ]
≤ 2 exp 2C1 k p0 −
.
hσf
E0 [X j (1 − X)k−j ]
C2
k
For those j/k-values where the likelihood ratio is not bounded below by 1,
we instead use (17). Changing the constants we obtain finally the following
bound valid on all of on J00 ∪ J10 :
√
E1 [X j (1 − X)k−j ]
j
2
hσf
≤ C2 f k exp C1 k p0 −
. (18)
E0 [X j (1 − X)k−j ]
k
• On J20 , arguing as in the previous case, we note that the likelihood ratio
multiplied by σf may be larger than 1 this time. Therefore by Lemma 3
(Part 2) and (17) we have
E1 [X j (1 − X)k−j ]
hσf
E0 [X j (1 − X)k−j ]
"
√
# " √
#2
√
√
k
k
j
j
≤ C0 f 2 ∨
σ
exp C1 k p0 −
exp C1 k p0 −
∧ σf
,
f C2
k
C2
k
Changing the constants we re-write this expression as
√
√
E1 [X j (1 − X)k−j ]
j
hσf
≤ C2 f k exp C1 k p0 −
,
E0 [X j (1 − X)k−j ]
k
where, to upper bound the minimum in square brackets above, we√only
squared the exponential (which is larger
than 1) and used the fact that f k =
√
k
o(1) (which implies that the term σf C2 is on the other hand asymptotically
√
smaller than 1). We also used that f 2 < f k to deal with the maximum
above.
We combine the two bounds into:
22
Claim 5 (Low substitution: Integrand). For all j/k ∈ J00 ∪ J10 ∪ J20 ,
hσf
E1 [X j (1 − X)k−j ]
E0 [X j (1 − X)k−j ]
√
≤ C2 (f k 1
2
j/k∈J00 ∪J10
+f k 1
j/k∈J20
) exp C1
√
j
k p0 −
.
k
(19)
It remains to bound the integrator, for which we rely on Chernoff’s bound. We
let
1
I0 = p0 , p0 + √
k
`−1
2
2`
I` = p0 + √ , p0 + √ ,
` ≥ 1.
k
k
√
Let Λ > 0 be such that 2Λ = (p̄ − p0 ) k. Then by Assumption A2
X
P0 [θ/k ∈ I`0 ] =
P0 [θ/k ∈ I`0 | X ∈ Iλ ] P0 [X ∈ Iλ ]
λ≥0
≤
X
2λ−1
P0 [θ/k ∈ I`0 | X ∈ Iλ ].
P0 [θ/k ∈ I`0 | X ∈ Iλ ] √ ρ−1 +
k
λ>Λ
λ≥0
X
By Chernoff’s bound
P0 [θ/k ∈ I`0 | X ∈ Iλ ] ≤ exp −2(2`−1 + 2λ−1 )2 ≤ exp −22`−1 − 22λ−1 .
In particular
X
P0 [θ/k ∈ I`0 | X ∈ Iλ ] ≤ exp −22`−1
X
λ>Λ
exp −22λ−1
λ>Λ
2`−1
≤ exp −2
exp (−C30 k) ,
for some constant C30 > 0 (not depending on f ). On the other hand,
X
exp −22`−1 X λ−1
2λ−1 −1
0
√
≤
2
exp −22λ−1
P0 [θ/k ∈ I` | X ∈ Iλ ] √ ρ
k
ρ k
λ≥0
λ≥0
2`−1
C3 exp −2
√
≤
,
k
for a constant C3 > 0 (not depending on f ). Combining the bounds and increasing
C3 appropriately, we get
23
Claim 6 (Low substitution: Integrator). For all ` ≥ 0,
P0 [θ/k ∈
I`0 ]
C3 exp −22`−1
√
≤
.
k
(20)
We can now compute the contribution
of J00 ∪ J10 ∪ J20 to the Hellinger distance.
√
L
Recall that L is defined by 2 = C log k. From (19) and (20), we get:
• For 0 ≤ ` ≤ L,
H 2 (P0 , Q)|I`0
√
2`−1
`
C
exp
−2
2
3
√
≤ C2 f 2 k exp C1 k √
k
k
√
2
2`−1
`
≤ C2 C3 f k exp −2
+ C1 2
√
2`
2
≤ C5 f k exp −C4 2 ,
for some constants C4 , C5 > 0. Summing over ` we get
L
X
√
H 2 (P0 , Q)|I`0 ≤ C6 f 2 k,
`=0
for some constant C6 > 0.
• Similarly, for ` > L,
H 2 (P0 , Q)|I`0 ≤ C5 f exp −C4 22` ,
adapting constants C4 , C5 > 0. Summing over ` we get
X
√
H 2 (P0 , Q)|I`0 ≤ C8 f exp −C7 C 2 log k = o(f 1+κ ) = o(f 2 k),
`>L
(21)
by choosing C large enough.
Combining these bounds we get finally:
Claim 7 (Low substitution: Hellinger distance).
√
H 2 (P0 , Q)|J00 ∪J10 ∪J20 = O(f 2 k).
24
Border regime. We now consider J0 , i.e., the bulk regime above p0 . The highlevel argument is similar to the case of J00 ∪ J10 above, although some details differ.
We first bound E1 [X j (1 − X)k−j ] using Lemma 4 and Assumption A1
E1 [X j (1 − X)k−j ] ≤ pj0 (1 − p0 )k−j .
(22)
To bound E0 [X j (1 − X)k−j ], we consider the event
(
"
r #)
1
j j
, +
.
E= X∈
k k
k
By Assumption A2 and Lemma 4, on J0 , arguing as in (12),
ρ
E0 [X j (1 − X)k−j ] ≥ √ pj (1 − p)k−j ,
k
where p =
j
k
+
q
E0 [X j (1 − X)k−j ]
E1 [X j (1 − X)k−j ]
1
.
k
(23)
Combining (22) and (23), and using Lemma 5, on J0 ,
≥
≥
ρ
√ exp(kΨj,p (p − p0 ))
k
1
ρ
j
1
2
√ exp k
− p (p − p0 ) + (p − p0 )
.
p(1 − p) k
4
k
For j/k ∈ I` , ` ≤ L,
E0 [X j (1 − X)k−j ]
E1 [X j (1 − X)k−j ]
1
ρ
≥ √ exp k −
p(1
− p)
k
1
≥ C2 √ exp C1 22` ,
k
r
2 !!
1 2` + 1
1 2`−1 + 1
√
√
+
k
4
k
k
for constants C1 , C2 > 0.
Claim 8 (Border regime: Likelihood ratio). For all j/k ∈ I` , ` ≤ L,
√
E1 [X j (1 − X)k−j ]
k
2`
≤
exp
−C
2
.
1
E0 [X j (1 − X)k−j ]
C2
(24)
We now bound the√
integrand in H 2 (P0 , Q). We follow the argument leading
up to (18). Because σf k = o(1), by Lemma 3 (Part 2) and (17) we have on I`
σf2 k
E1 [X j (1 − X)k−j ]
2
hσf
≤
C
f
∨
exp −2C1 22` .
0
2
j
k−j
E0 [X (1 − X) ]
C2
25
Changing the constants we re-write this expression as
E1 [X j (1 − X)k−j ]
2
2
2`
.
≤
C
f
∨
C
f
k
exp
−C
2
hσf
0
2
1
E0 [X j (1 − X)k−j ]
Claim 9 (Border regime: Integrand). For all j/k ∈ I` , 0 ≤ ` ≤ L,
E1 [X j (1 − X)k−j ]
2
2
2`
hσf
≤
C
f
∨
C
f
k
exp
−C
2
.
0
2
1
E0 [X j (1 − X)k−j ]
(25)
It remains √
to bound the integrator. We have by Assumption A2 (recall that
2Λ = (p̄ − p0 ) k)
X
P0 [θ/k ∈ I` ] =
P0 [θ/k ∈ I` | X ∈ Iλ ]P0 [X ∈ Iλ ]
λ≥0
X
≤
P0 [X ∈ Iλ ] +
P0 [θ/k ∈ I` | X ∈ Iλ ]P0 [X ∈ Iλ ]
`<λ≤Λ
0≤λ≤`
+
X
X
P0 [θ/k ∈ I` | X ∈ Iλ ]
λ>Λ
`
X
2
2λ−1
≤ √ ρ−1 +
P0 [θ/k ∈ I` | X ∈ Iλ ] √ ρ−1
k
k
`<λ≤Λ
X
+
P0 [θ/k ∈ I` | X ∈ Iλ ].
λ>Λ
By Chernoff’s bound, for λ > `,
P0 [θ/k ∈ I` | X ∈ Iλ ] ≤ exp −2(−2` + 2λ−1 )2 ≤ exp −22`+1 (2λ−`−1 − 1)2 .
In particular
X
P0 [θ/k ∈ I` | X ∈ Iλ ] ≤ exp (−C30 k) ,
λ>Λ
for some constant C30 > 0 (not depending on f ). On the other hand,
C3 2`
2λ−1
P0 [θ/k ∈ I` | X ∈ Iλ ] √ ρ−1 ≤ √ ,
k
k
`<λ≤Λ
X
for a constant C3 > 0 (not depending on f ). Combining the bounds and increasing
C3 appropriately, we get
26
Claim 10 (Border substitution: Integrator). For all 0 ≤ ` < L,
C3 2`
P0 [θ/k ∈ I` ] ≤ √ .
k
(26)
We can now compute the contribution of J0 to the Hellinger distance. From (25)
and (26), we get for 0 ≤ ` ≤ L
H 2 (P0 , Q)|I` ≤
C0 f 2 ∨ C2 f 2 k exp −C1 22`
C3 2`
√ .
k
Summing over ` we get
L
X
√
H 2 (P0 , Q)|I` ≤ C4 f 2 k,
`=0
for some constant C4 > 0.
Claim 11 (Border regime: Hellinger distance).
√
H 2 (P0 , Q)|J0 = O(f 2 k).
Wrapping up We now prove Proposition 1.
Proof of Proposition 1.
√
H 2 (P0 , Q) ≤ HJ21 (P0 , Q) + HJ200 ∪J10 ∪J20 (P0 , Q) + HJ20 (P0 , Q) ≤ O(f 2 k),
by Claims 3, 7 and 11. That implies Proposition 1.
4
Matching upper bound
We give two proofs of the upper bound.
4.1
Proof of Theorem 2
Proof. We use (4) and construct an explicit test A as follows:
27
• Let W be the number of genes such that θ/k ≤ p0 . Let w = P0 [θ/k ≤ p0 ],
w0 = Q[θ/k ≤ p0 ] and
w∗ = m
w + w0
m
m
= mw + (w0 − w) = mw0 − (w0 − w).
2
2
2
We consider the following event
A = {W ≥ w∗ }.
It remains to show that the event A is highly unlikely under P⊗m
while being
0
highly likely under Q⊗m . We do this by bounding the difference w0 − w and
applying Chebyshev’s inequality to W .
Note that W ∼ Bin(m, w) under P⊗m
and W ∼ Bin(m, w0 ) under Q⊗m .
0
By Assumption A1, X ∈ [p0 − φf , p0 ] under P1 . By the Berry-Esseen theorem
(e.g. [Dur96]),
1
1
1
(27)
P1 [θ/k ≤ p0 ] ≥ E1 [P1 [θ ≤ kX | X]] = − O √
≥ ,
2
3
k
for k large enough. Hence,
w0 = σf P1 [θ/k ≤ p0 ] + (1 − σf )w
1
≥
σf + (1 − σf )w,
3
(28)
whereas by the computations in the previous section (more specifically, by summing over ` in (20))
1
w=O √
,
(29)
k
√
and, similarly, since f k = o(1)
1
0
w =O √
,
(30)
k
from (28) and (29). Consequently
0
w − w ≥ σf
1
−w
3
28
= Ω(f ).
(31)
By Chebyshev’s inequality,
P⊗m
0
4mw(1 − w)
[A] ≤ 2 0
=O
m (w − w)2
1
√
mf 2 k
δ
≤ ,
2
for c0 large enough, where we used (29) and (31). Similarly,
Q⊗m [Ac ] ≤
4.2
4mw0 (1 − w0 )
δ
.
≤
m2 (w0 − w)2
2
Agnostic version
Although Theorem 2 shows that our bound in Theorem 1 is tight, it relies on a test
(i.e., the set A) that assumes knowledge of the null and alternative hypotheses.
Here we relax this assumption.
Pairwise distance comparisons We assume that we have two (independent)
collections of genes, T1 and T2 , one from each model, P0 and Q as in the previous
section. We split the genes into two equal-sized disjoint sub-collections, (T11 , T12 )
and (T21 , T22 ). Assume for convenience that the total number of genes is in fact 2m
for each dataset. Let C > 0 be a constant, to be determined later (in equation (33)).
We proceed in two steps.
1. We first compute p̂1 and p̂2 , the
tively. Let p̂ = max{p̂1 , p̂2 }.
C
√
-quantiles
k
based on T11 and T21 respec-
2. Compute the fraction of genes, ŵ1 and ŵ2 , with θ/k ≤ p̂ in T12 and T22
respectively.
We infer that the first dataset comes from P⊗2m
if ŵ1 < ŵ2 , and vice versa.
0
Remark 1. Simply comparing the √Ck -quantiles breaks down when f k1 , as
it is quite possible that the quantiles will be identical since they can only take k
possible values. However, even if the quantiles are identical, the probability of
a gene being lower than the quantile is bigger if the distance is smaller. This
explains the need for the second phase in our algorithm. We remark further that
the partition of the data into two sets is used for analysis purposes as it allows for
better control of dependencies.
29
We show that this approach succeeds with probability at least 1 − δ whenever
m ≥ c0 f 21√k , for c0 large enough. This proceeds from a series of claims.
Claim 12 (p̂ is close to p0 ). For c0 large enough, there is C1 > 0 such that
C1
p̂ ∈ p0 , p0 + √
k
(32)
with probability 1 − δ/2.
Proof. The argument is similar to that in the proof of Theorem 2.
By summing over ` in (20),
C0
P0 [θ/k ≤ p0 ] ≤ √1 ,
k
for some C10 > 0. For any C100 > C10 , there is C1 > 0 such that
C1
C1
C1
P0 θ/k ≤ p0 + √
≥ P0 θ/k ≤ p0 + √ X ∈ p0 , p0 + √
k
k
k
C1
× P0 X ∈ p0 , p0 + √
k
1 ρ C1
C100
√ ≥√
≥
3 k
k
by the Berry-Esseen theorem (as in (27)), for C1 large enough.
Let
C 0 + C100
.
(33)
C= 1
2
Let W be the number of genes (among m) such that θ/k ≤ p0 and w = P0 [θ/k ≤
p0 ]. Repeating the calculations in the proof of Theorem 2,
C
4mw(1 − w)k
1 √
1
δ
⊗m
P0
W ≥ m√ ≤ 2
=
O
k ≤ 0 O(f 2 k) ≤ ,
0 2
m (C − C1 )
m
c
8
k
for c0 large
the number of genes such that θ/k ≤
√ enough. Similarly, let W̃ be √
p0 + C1 / k and w̃ = P0 [θ/k ≤ p0 + C1 / k]. Then
C
δ
⊗m
P0
W̃ ≤ m √ ≤ .
8
k
30
√
That implies that with probability 1 − δ/4 the C/ k-quantile under
P⊗m
lies in
0
√
C
1
√
the interval [p0 , p0 + k ]. By monotonicity, P1 [θ/k ≤ p0 + C1 / k] ≥ w̃, and we
also have
δ
C
⊗m
Q
W̃ ≤ m √ ≤ ,
8
k
which implies the claim.
Claim 13 (Test). For c0 large enough, if T1 comes from P0⊗2m , T2 comes from
Q⊗2m and (32) holds, then
ŵ1 < ŵ2
with probability 1 − δ/2, and vice versa.
Proof. The proof is identical to that of Theorem 2 with W now being the number
of genes such that θ/k ≤ p̂, w = P0 [θ/k ≤ p̂], w0 = Q[θ/k ≤ p̂], and (29)
and (30) now following from Claim 12 together with (20) and (26).
Triplet reconstruction Consider again the three possible species trees depicted
in Figure 2. By comparing the pairs two by two as described in the agnostic algorithm, we can determine which is the correct species tree topology. Such “triplet”
information is in general enough (assuming the molecular clock hypothesis) to
reconstruct a species tree on any number of species (e.g. [SS03]). We leave out
the details.
Acknowledgments
We thank Gautam Dasarathy and Rob Nowak for helpful discussions.
References
[ADHR12] Alexandr Andoni, Constantinos Daskalakis, Avinatan Hassidim, and
Sebastien Roch. Global alignment of molecular sequences via ancestral state reconstruction. Stochastic Processes and their Applications, 122(12):3852 – 3874, 2012.
[ADR11]
Elizabeth S. Allman, James H. Degnan, and John A. Rhodes. Identifying the rooted species tree from the distribution of unrooted
gene trees under the coalescent. Journal of Mathematical Biology,
62(6):833–862, 2011.
31
[ALPE12]
Christian N.K. Anderson, Liang Liu, Dennis Pearl, and Scott V. Edwards. Tangled trees: The challenge of inferring species trees from
coalescent and noncoalescent genes. In Maria Anisimova, editor,
Evolutionary Genomics, volume 856 of Methods in Molecular Biology, pages 3–28. Humana Press, 2012.
[BS14]
Anand Bhaskar and Yun S. Song. Descartes’ rule of signs and
the identifiability of population demographic models from genomic
variation data. Ann. Statist., 42(6):2469–2493, 2014.
[CGG02]
M. Cryan, L. A. Goldberg, and P. W. Goldberg. Evolutionary trees
can be learned in polynomial time. SIAM J. Comput., 31(2):375–
397, 2002.
[CJJ11]
T. Tony Cai, X. Jessie Jeng, and Jiashun Jin. Optimal detection of
heterogeneous and heteroscedastic mixtures. J. R. Stat. Soc. Ser. B
Stat. Methodol., 73(5):629–662, 2011.
[CJT05]
L. Cayon, J. Jin, and A. Treaster. Higher criticism statistic: detecting
and identifying non-gaussianity in the wmap first-year data. Monthly
Notices of the Royal Astronomical Society, 362(3):826–832, 2005.
[CT91]
T. M. Cover and J. A. Thomas. Elements of information theory.
Wiley Series in Telecommunications. John Wiley & Sons Inc., New
York, 1991. A Wiley-Interscience Publication.
[CW14]
T.T. Cai and Yihong Wu. Optimal detection of sparse mixtures
against a given null distribution. Information Theory, IEEE Transactions on, 60(4):2217–2232, April 2014.
[DBP05]
Frederic Delsuc, Henner Brinkmann, and Herve Philippe. Phylogenomics and the reconstruction of the tree of life. Nat Rev Genet,
6(5):361–375, 05 2005.
[DD10]
Michael DeGiorgio and James H Degnan. Fast and consistent estimation of species trees using supermatrix rooted triples. Molecular
Biology and Evolution, 27(3):552–69, March 2010.
[DDBR09] James H. Degnan, Michael DeGiorgio, David Bryant, and Noah A.
Rosenberg. Properties of consensus methods for inferring species
trees from gene trees. Systematic Biology, 58(1):35–54, 2009.
32
[DJ04]
David Donoho and Jiashun Jin. Higher criticism for detecting sparse
heterogeneous mixtures. Ann. Statist., 32(3):962–994, 06 2004.
[DMR11a] Constantinos Daskalakis, Elchanan Mossel, and Sébastien Roch.
Evolutionary trees and the ising model on the bethe lattice: a proof of
steel’s conjecture. Probability Theory and Related Fields, 149:149–
189, 2011. 10.1007/s00440-009-0246-2.
[DMR11b] Constantinos Daskalakis, Elchanan Mossel, and Sébastien Roch.
Phylogenies without branch bounds: Contracting the short, pruning
the deep. SIAM J. Discrete Math., 25(2):872–893, 2011.
[DNR15]
Gautam Dasarathy, Robert D. Nowak, and Sébastien Roch. Data
requirement for phylogenetic inference from multiple loci: A new
distance method. IEEE/ACM Trans. Comput. Biology Bioinform.,
12(2):422–432, 2015.
[Dob58]
R. L. Dobrusin. A statistical problem arising in the theory of detection of signals in the presence of noise in a multi-channel system
and leading to stable distribution laws. Theory of Probability & Its
Applications, 3(2):161–173, 1958.
[DR06]
J. H. Degnan and N. A. Rosenberg. Discordance of species trees
with their most likely gene trees. PLoS Genetics, 2(5), May 2006.
[DR09]
James H. Degnan and Noah A. Rosenberg. Gene tree discordance,
phylogenetic inference and the multispecies coalescent. Trends in
Ecology and Evolution, 24(6):332 – 340, 2009.
[DR13]
Constantinos Daskalakis and Sebastien Roch. Alignment-free phylogenetic reconstruction: sample complexity via a branching process
analysis. Ann. Appl. Probab., 23(2):693–721, 2013.
[Dur96]
Richard Durrett. Probability: theory and examples. Duxbury Press,
Belmont, CA, second edition, 1996.
[Dur08]
Richard Durrett. Probability models for DNA sequence evolution.
Probability and its Applications (New York). Springer, New York,
second edition, 2008.
33
[ESSW99a] P. L. Erdös, M. A. Steel;, L. A. Székely, and T. A. Warnow. A few
logs suffice to build (almost) all trees (part 1). Random Struct. Algor.,
14(2):153–184, 1999.
[ESSW99b] P. L. Erdös, M. A. Steel;, L. A. Székely, and T. A. Warnow. A few
logs suffice to build (almost) all trees (part 2). Theor. Comput. Sci.,
221:77–118, 1999.
[Fel04]
J. Felsenstein. Inferring Phylogenies. Sinauer, New York, New York,
2004.
[HTF09]
Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The elements of statistical learning. Springer Series in Statistics. Springer,
New York, second edition, 2009. Data mining, inference, and prediction.
[Ing97]
Yu. I. Ingster. Some problems of hypothesis testing leading to infinitely divisible distributions. Math. Methods Statist., 6(1):47–69,
1997.
[JC69]
T. H. Jukes and C. Cantor. Mammalian protein metabolism. In H. N.
Munro, editor, Evolution of protein molecules, pages 21–132. Academic Press, 1969.
[JCL10]
X. Jessie Jeng, T. Tony Cai, and Hongzhe Li. Optimal sparse segment identification with application in copy number variation analysis. J. Amer. Statist. Assoc., 105(491):1156–1166, 2010.
[KHH+ 05] Martin Kulldorff, Richard Heffernan, Jessica Hartman, Renato Assuno, and Farzad Mostashari. A space?time permutation scan statistic for disease outbreak detection. PLoS Med, 2(3):e59, 02 2005.
[KMRR15] Junhyong Kim, Elchanan Mossel, Mikls Z. Rcz, and Nathan Ross.
Can one hear the shape of a population history? Theoretical Population Biology, 100(0):26 – 38, 2015.
[LYK+ 09]
Liang Liu, Lili Yu, Laura Kubatko, Dennis K. Pearl, and Scott V.
Edwards. Coalescent methods for estimating phylogenetic trees.
Molecular Phylogenetics and Evolution, 53(1):320 – 328, 2009.
34
[LYP10]
Liang Liu, Lili Yu, and DennisK. Pearl. Maximum tree: a consistent estimator of the species tree. Journal of Mathematical Biology,
60(1):95–106, 2010.
[LYPE09]
Liang Liu, Lili Yu, Dennis K. Pearl, and Scott V. Edwards. Estimating species phylogenies using coalescence times among sequences.
Systematic Biology, 58(5):468–477, 2009.
[Mad97]
Wayne P. Maddison. Gene trees in species trees. Systematic Biology,
46(3):523–536, 1997.
[MFP08]
Simon Myers, Charles Fefferman, and Nick Patterson. Can one learn
history from the allelic spectrum? Theoretical Population Biology,
73(3):342 – 348, 2008.
[Mos03]
E. Mossel. On the impossibility of reconstructing ancestral data and
phylogenies. J. Comput. Biol., 10(5):669–678, 2003.
[Mos04]
E. Mossel. Phase transitions in phylogeny. Trans. Amer. Math. Soc.,
356(6):2379–2404, 2004.
[MR06]
Elchanan Mossel and Sébastien Roch. Learning nonsingular phylogenies and hidden Markov models. Ann. Appl. Probab., 16(2):583–
614, 2006.
[MR10]
Elchanan Mossel and Sébastien Roch. Incomplete lineage sorting:
Consistent phylogeny estimation from multiple loci. IEEE/ACM
Trans. Comput. Biology Bioinform., 7(1):166–171, 2010.
[MR15]
Elchanan Mossel and Sébastien Roch. Distance-based species tree
estimation: Information-theoretic trade-off between number of loci
and sequence length under the coalescent. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, APPROX/RANDOM 2015, August 24-26, 2015, Princeton,
NJ, USA, pages 931–942, 2015.
[MRS11]
Elchanan Mossel, Sébastien Roch, and Allan Sly. On the inference of large phylogenies with long branches: How long is too
long? Bulletin of Mathematical Biology, 73:1627–1644, 2011.
10.1007/s11538-010-9584-6.
35
[MSZ+ 13]
Raphaël Mourad, Christine Sinoquet, Nevin Lianwen Zhang,
Tengfei Liu, and Philippe Leray. A survey on latent tree models
and applications. J. Artif. Intell. Res. (JAIR), 47:157–203, 2013.
[Nak13]
Luay Nakhleh. Computational approaches to species phylogeny inference and gene tree reconciliation. Trends in ecology & evolution,
28(12):10.1016/j.tree.2013.09.004, 12 2013.
[Roc10]
Sebastien Roch. Toward extracting all phylogenetic information
from matrices of evolutionary distances. Science, 327(5971):1376–
1379, 2010.
[Roc13]
Sebastien Roch. An analytical comparison of multilocus methods
under the multispecies coalescent: The three-taxon case. In Pacific
Symposium in Biocomputing 2013, pages 297–306, 2013.
[RS15]
Sebastien Roch and Mike Steel. Likelihood-based tree reconstruction on a concatenation of aligned sequence data sets can be statistically inconsistent. Theoretical Population Biology, 100:56 – 62,
2015.
[RW15]
Sebastien Roch and Tandy Warnow. On the robustness to gene tree
estimation error (or lack thereof) of coalescent-based species tree
methods. Systematic Biology, 64(4):663–676, 2015.
[RY03]
Bruce Rannala and Ziheng Yang. Bayes estimation of species divergence times and ancestral population sizes using DNA sequences
from multiple loci. Genetics, 164(4):1645–1656, 2003.
[SS02]
M. A. Steel and L. A. Székely. Inverting random functions. II.
Explicit bounds for discrete maximum likelihood estimation, with
applications. SIAM J. Discrete Math., 15(4):562–575 (electronic),
2002.
[SS03]
C. Semple and M. Steel. Phylogenetics, volume 22 of Mathematics
and its Applications series. Oxford University Press, 2003.
36
| 10math.ST
|
arXiv:quant-ph/0508116v1 16 Aug 2005
A Probabilistic Branching Bisimulation for
Quantum Processes
Marie L ALIRE∗
Leibniz Laboratory
46, avenue Félix Viallet - 38000 Grenoble, France
12th August 2017
Abstract
Full formal descriptions of algorithms making use of quantum principles
must take into account both quantum and classical computing components
and assemble them so that they communicate and cooperate. Moreover, to
model concurrent and distributed quantum computations, as well as quantum communication protocols, quantum to quantum communications which
move qubits physically from one place to another must also be taken into
account.
Inspired by classical process algebras, which provide a framework for
modeling cooperating computations, a process algebraic notation is defined,
which provides a homogeneous style to formal descriptions of concurrent
and distributed computations comprising both quantum and classical parts.
Based upon an operational semantics which makes sure that quantum objects, operations and communications operate according to the postulates of
quantum mechanics, a probabilistic branching bisimulation is defined among
processes considered as having the same behavior.
1
Introduction
Quantum algorithms are frequently described by means of quantum gate networks.
This has several drawbacks, for instance, gate networks do not allow descriptions
of loops nor conditional executions of parts of networks. So as to overcome these
∗ [email protected]
1
difficulties, a few quantum programming languages have been developed, such
as: QCL [12], an imperative language designed by Bernhard Ömer which aims at
simulating quantum programs, qGCL [15] by Paolo Zuliani which allows the construction of proved correct quantum programs through a refinement method, QPL
[13], a functional language designed by Peter Selinger with a denotational semantics, and a few others. Several quantum λ-calculus have also been developed:
for example [14] by André Van Tonder, which is based on a simplified linear λcalculus and [2] by Pablo Arrighi and Gilles Dowek, which is a ”linear-algebraic
λ-calculus”.
Cooperation between quantum and classical computations is inherent in quantum algorithmics. Teleportation of a qubit state from Alice to Bob [4] is a good
example of this cooperation. Indeed, Alice carries out a measurement, the result
of which (two bits) is sent to Bob, and Bob uses this classical result to determine which quantum transformation he must apply. Moreover, initial preparation
of quantum states and measurement of quantum results are two essential forms
of interactions between the classical and quantum parts of computations which a
language must be able to express. Process algebras are a good candidate for such
a language since they provide a framework for modeling cooperating computations. In addition, they have well defined semantics and permit the transformation
of programs as well as the formal study and analysis of their properties. A key
point in their semantics is the definition of an equivalence relation on processes.
Bisimulation is an adequate equivalence relation to deal with communicating processes since it relates processes that can execute the same flows of actions while
having the same branching structure.
Simon Gay and Rajagopal Nagarajan have also developed CQP, a language
to describe communicating quantum processes [11]. This language is based on
π-calculus. An important point in their work is the definition of a type system,
and the proof that the operational semantics preserves typing.
This paper presents first the main points of the definition of a Quantum Process
Algebra (QPAlg). Then, examples of short quantum programs are given in section
3. Finally, a bisimulation among processes is defined in section 4.
2
Quantum communicating processes
The process algebra developed here is based upon process algebras such as CCS
[10] and Lotos [5]. In this process algebra, basic actions are communications
among processes (emission, denoted g !x and reception, denoted g ?x , where g
is a communication gate) and quantum actions (unitary transformations and mea-
2
surements). To create a process from basic actions, the prefix operator ”.” is used:
if α is an action and P , a process, α.P is a new process which performs α first,
then behaves as P .
There are two predefined processes: nil, the process that cannot perform any
transition, and end, which performs a ”δ-transition” for signaling successful termination, and becomes nil (”δ-transitions” are necessary in the semantics of sequential composition of processes).
The operators of the process algebra are: sequential composition (P ; Q), parallel composition (P k Q), conditional choice ([[ c1 → P1 , . . . , cn → Pn ] ) and
restriction (P \L ). As for sequential composition, process Q is executed if process
P terminates successfully, that is to say if P performs a δ-transition. The process
[ c1 → P1 , . . . , cn → Pn ] , where ci is a condition and Pi a process, evolves as a
process chosen nondeterministically among the processes Pj such that cj is true.
Restriction is useful for disallowing the use of some gates (the gates listed in L),
thus forcing internal communication within process P . Communication can occur
between two parallel processes whenever a value emission in one of them and a
value reception in the other one use the same gate name.
The main points of QPAlg involving the quantum world are developed in the
rest of this section. The precise syntax and the main inference rules of the semantics are given in appendix A. For more details, see [9].
2.1
Quantum variables
For the purpose of this paper, we consider that there are two types of variables,
one classical: Nat, for variables taking integer values, and one quantum: Qubit for
variables standing for qubits. An extended version of the process algebra would
of course also include quantum registers and other types of variables.
In classical process algebras, variables are instantiated when communications
between processes occur and cannot be modified after their instantiation. As a
consequence, it is not necessary to store their values. In fact, when a variable is
instantiated, all its occurrences are replaced by the value received.
Here, quantum variables stand for physical qubits. Applying a unitary transformation to a variable which represents a qubit modifies the state of that qubit.
This means that values of variables are modified. For that reason, it is necessary
to keep track of both variable names and variable states.
Variables are declared, the syntax is: [ x1 : t1 , . . . , xn : tn . P ] where
x1 , . . . , xn is a list of variables, t1 , . . . , tn are their types, and P is a process
which can make use of these classical and quantum variables. To simplify the rest
of this paper, the names of variables will always be considered distinct.
3
In the inference rules which describe the semantics of processes, the states of
processes are process terms P together with contexts C, of the form P/C. The
main purpose of a context is to maintain the quantum state, stored as q = ρ where
q is a sequence of quantum variable names and ρ a density matrix representing
their current quantum state. Moreover, in order to treat classical variables in a
similar way, modifications of classical variables are also allowed. So, for the
same reason as in the case of quantum variables, classical values are stored in
the context. Storing and retrieving classical values is represented by functions
f : names → values. The context must also keep track of the embedding of
variable scopes. To keep track of parallel composition, this is done via a ”cactus
stack” structure of sets of variables, called the environment stack (s), which stores
variable scopes and types. The set of all the variables in s is denoted Var(s), ”.”
adds an element on top of a stack, and ”|” concatenates two stacks.
In summary, the context has three components < s, q = ρ, f >, where:
• s is the environment stack;
• q is a sequence of quantum variable names;
• ρ is a density matrix representing the quantum state of the variables in q;
• f is the function which associates values with classical variables.
The rules for declaration and liberation of variables are the following:
Declaration:
[ x1 : t1 , . . . , xn : tn . P ] /C
′
−−τ−→ [ P ] /C
with C =< s, q = ρ, f >, C ′ =< s′ , q = ρ, f >
and s′ = {(x1 , t1 ), . . . , (xn , tn )}.s
This rule adds the new variable names and types on top of the stack s. Because
the variables do not have values yet, the quantum state and the classical function
do not have to be modified at this point.
Evolution of a process within the scope of declared variables:
P/C
[ P ] /C
′
′
−−α
−→ P /C
′
′
−−α
−→ [ P ] /C
4
α 6= δ
In short: if the process P can perform a transition, then the process [ P ] can
perform the same transition, provided that the action of the transition is not δ.
Termination of a process with exit from a scope and liberation of the variables:
′
−−δ−→ P / < e.s, q = ρ, f >
nil / < s, q\Var(e) = T rVar(e)/q (ρ), f \Var(e) >
P/C
[ P ] /C
δ
−−−→
If the action is δ, this means that P has successfully terminated, so the context
must be cleaned up by eliminating the variables having their scope limited to that
process. These variables have their names listed in the head e of the stack. So,
cleaning up the context means eliminating the head of the stack, removing the
variables in e from the sequence q and from the domain of the function f . The
new quantum state is obtained by performing a partial trace on ρ over the qubits
in Var(e), which is denoted T rVar(e)/q (ρ).
2.2
Basic actions
The classical basic actions are classical to classical communications and will not
be further defined here. Classical to quantum communications and quantum to
quantum communications are introduced for respectively initializing qubits and
allowing the description of quantum communication protocols. Quantum to classical communications are part of measurement and are dealt with in the next paragraph.
The semantics of communications involving the quantum world is based upon
the following rules concerning the quantum side of such communications:
g !x .P/C
g !x
−−−−−−→
P/C ′
where
• C =< s, q = ρ, f > and C ′ =< s\{x}, q\{x} = T r{x}/q (ρ), f >
• x ∈ Var(s) and x ∈ q
and:
g ?x .P/C
g ?x
P/C ′
−−−−−−→
where
5
• C =< s, q = ρ, f >, C ′ =< s, x.q = ν ⊗ ρ, f >
• x ∈ Var(s), x of type Qubit, x 6∈ q
• ν density matrix of dimension 2
The first rule deals with qubit sending, and the other one, with reception of a
qubit. For qubit sending, because of the no-cloning theorem, the sent qubit must
be removed from the context. The two rules concerning classical value sending
and classical value reception are given in appendix A.2.
In the operational semantics of parallel composition, the combination of the
rules for emission and reception defines communication. In a classical to quantum communication, the qubit is initialized in the basis state |vihv|, where v is
the classical value sent (in this case, v must be 0 or 1). In a quantum to quantum
communication, the name of the sent qubit is replaced in q by the name of the
receiving qubit.
Other basic actions are unitary transformations which perform the unitary evolution of qubit states. Given a set U of predefined unitary transformations, the
action corresponding to the application of U ∈ U to a list of quantum variables is
denoted by U [x1 , . . . , xn ].
The inference rule for unitary transformations is:
U [x1 , . . . , xn ].P/ < s, q = ρ, f >
′
−−τ−→ P/ < s, q = ρ , f >
where
• U ∈ U, x1 , . . . , xn ∈ Var(s), and x1 , . . . , xn ∈ q
• x1 , . . . , xn are pairwise distinct
• ρ′ = TU (ρ)
The condition x1 , . . . , xn ∈ q prevents from applying a unitary transformation
to qubits which have not been initialized.
TU is the super-operator which must be applied to ρ, to describe the evolution
of the quantum state due to the application of the unitary transformation U to the
qubits x1 , . . . , xn . In general, with A a 2n × 2n matrix:
TA : ρ 7→ Π† .(A ⊗ I ⊗k ).Π.ρ.Π† .(A† ⊗ I ⊗k ).Π
where
6
• Π is the permutation matrix which places the xi ’s at the head of q
• k = size(q) − n
• I ⊗k = I ⊗ · · · ⊗ I , where I is the identity matrix on C2
| {z }
k
Since the unitary transformation U may be applied to qubits which are anywhere within the list q, a permutation Π must be applied first. This permutation
moves the xi ’s so that they are placed at the head of q in the order specified by
[x1 , . . . , xn ]. Then U can be applied to the first n elements and I to the remainder.
Finally, the last operation is the inverse of the permutation Π so that at the end,
the arrangement of the elements in ρ is consistent with the order of the elements
in q.
2.3
Measurement and probabilistic processes
Last but not least, an essential basic action has to be introduced into the process
algebra: quantum measurement. Let M be an observable in a set O of predefined
observables, x1 , . . . , xn a list of distinct quantum variables and g a gate. The
syntax for measurement is the following:
• M [x1 , . . . , xn ] is a measurement of the n qubits of the list with respect to
observable M , but the classical result is neither stored nor transmitted.
• g !M [x1 , . . . , xn ] is a measurement of the n qubits of the list with respect
to observable M , followed by sending the classical result through gate g.
Measurement is probabilistic: more precisely, the classical result and the quantum state after measurement are probabilistic. In the case of measurement without
communication of the classical result, only the quantum state is probabilistic after
measurement, so the probabilities can be reflected in the density matrix:
M [x1 , . . . , xn ].P/ < s, q = ρ, f >
′
−−τ−→ P/ < s, q = ρ , f >
with
• x1 , . . . , xn ∈ Var(s), x1 , . . . , xn ∈ q and x1 , . . . , xn are pairwise distinct
P
• M ∈ O with i λi Pi as spectral decomposition
P
• ρ′ = i TPi (ρ)
7
As in the case of unitary transformations, TPi is the super-operator corresponding
to the application of the projector Pi to measured qubits. The computation of ρ′
stems from the projective measurement postulate of quantum mechanics.
When the value coming out of the measurement is sent out, the classical result
is probabilistic. This requires the introduction of a probabilistic composition operator for contexts. This operator is denoted ⊞p : the state P/C1 ⊞p C2 is P/C1
with probability p and P/C2 with probability 1 − p. In general, a context is either
of the form < s, q = ρ, f >, or of the form ⊞pi < si , qi = ρi , fi > where the pi ’s
are probabilities adding to 1. Then, the rule for measurement followed by sending
the classical result is:
g !M [x1 , . . . , xn ] .P/C
τ
[ g !y .end ] ; P/ ⊞pi Ci
−−−→
where
• C =< s, q = ρ, f >
• Ci =< {(y, Nat)}.s, q = ρi , f ∪ {y 7→ λi } >
• x1 , . . . , xn ∈ Var(s), x1 , . . . , xn ∈ q and x1 , . . . , xn are pairwise distinct
• y is a new variable (introduced as y : Nat by this rule)
P
• M ∈ O with i λi Pi as spectral decomposition
• pi = T r(TPi (ρ)),
ρi =
1
pi TPi (ρ)
As explained in [6] and [7], if a process contains both a probabilistic and a
nondeterministic choice, the probabilistic choice must always be solved first. In
the process algebra presented here, nondeterminism appears with parallel composition and conditional choice. So as to guarantee that probabilistic choice is
always solved first, the notion of probabilistic stability for contexts is introduced:
a context C is probabilistically stable, which is denoted C ↓, if it is of the form
< s, q = ρ, f >. If the context of a process state is not stable, a probabilistic
transition must be performed first:
X
pi = 1
where
P/ ⊞pi Ci −→pj P/Cj
i
where S1 −→p S2 means that state S1 becomes S2 with probability p.
8
3
Examples
In the following examples, the set U of unitary transformations is:
U = {H, CN ot, I, X, Y, Z}
where H is Hadamard transformation, CN ot is ”controlled not”, I is the identity,
and X, Y, Z are Pauli operators. The set O of observables contains the observables corresponding to measurement of one and two qubits in the standard basis, denoted respectively Mstd,1 and Mstd,2 , and the observable corresponding to
measurement of a qubit in the basis {|+i, |−i}, denoted M+− .
3.1 Teleportation
Once upon a time, there were two friends, Alice and Bob who had to separate
and live away from each other. Before leaving, each one took a qubit of the same
EPR pair. Then Bob went very far away, to a place that Alice did not know. Later
on, someone gave Alice a mysterious qubit in a state |ψi = α|0i + β|1i, with a
mission to forward this state to Bob. Alice could neither meet Bob and give him
the qubit, nor clone it and broadcast copies everywhere, nor obtain information
about α and β. Nevertheless, Alice succeeded thanks to the EPR pair and the
teleportation protocol [4]:
BuildEPR
Alice
def
=
def
=
[ x : Qubit, y : Qubit .
((g1 ?x .g2 ?y .H[x].CN ot[x, y].end)
k (g1 !0 .g2 !0 .end))\{g1 , g2 }
]
[ x : Qubit, y : Qubit .
CN ot[x, y].H[x].meas !Mstd,2 [x, y] .end
]
9
Bob
def
=
Teleport
[ z : Qubit .
[ k : Nat .
meas ?k .
[ k = 0 → I[z].end,
k = 1 → X[z].end,
k = 2 → Z[z].end,
k = 3 → Y [z].end ]
]
]
def
=
[ ψ : Qubit .
[ a : Qubit, b : Qubit .
BuildEPR[a, b] ;
(Alice[ψ, a] k Bob[b])\{meas}
]
]
The inference rules can be used to show that this protocol results in Bob’s z
qubit having the state initially possessed by the x qubit of Alice, with only two
classical bits sent from Alice to Bob.
3.2
Communication protocols
Alice sends qubits to Bob through a non secure channel and Eve eavesdrops this
channel to get information on the qubits sent by Alice. In the following example
A, B, and E are processes modeling whatever Alice, Bob, and Eve may respectively apply to their qubits. The actions of these processes, which are not made
explicit here, will be specified in the next example of the BB84 protocol.
The communication protocols which are described here could be used to model cryptographic protocols so as to check if they are secure.
Eve intercepts all qubits
Eve intercepts qubits because of a flaw in the channel that Alice and Bob are using
to communicate.
Alice
def
=
[ a : Qubit . A[a] ; fill !a .end ] ; Alice
10
Bob
Eve
Flaw
def
=
[ b : Qubit . empty ?b .B[b] ] ; Bob
def
[ e : Qubit, f : Qubit .
emptyFlaw ?e .E[e, f ] ; fillFlaw !f .end
] ; Eve
=
def
=
Channel
Protocol
[ u : Qubit, v : Qubit . emptyFlaw !u .fillFlaw ?v .end ]
def
=
[ x : Qubit, y : Qubit . fill ?x .Flaw[x, y] ; empty !y .end ] ;
Channel
def
(Alice k Bob k Eve k Channel)
\{ fill, empty, fillFlaw, emptyFlaw}
=
Eve intercepts some of the qubits
This part assumes that a nondeterministic process composition P +Q is introduced
in the process algebra. This operator is not presented in the operational semantics
in appendix A.2 but it can be simulated by [ true → P, true → Q ] .
To model the fact that Eve does not succeed in intercepting all qubits, the flaw
in the channel is made nondeterministic:
Channel
def
=
[ x : Qubit .
fill ?x .
(
[ y : Qubit . Flaw[x, y] ; empty !y .end ] +
(empty !x .end)
)
] ; Channel
3.3 The BB84 protocol
The BB84 protocol [3] is a protocol for secure quantum key distribution: Alice
and Bob must agree on a private key, i.e. a list of bits that should remain secret.
To communicate, they send qubits through a non secure channel. In fact, the
processes A and B left unspecified in the previous paragraph can be used to model
this protocol. The process Alice is redefined and the process B used by Bob is
made explicit. In addition, another process is defined: the process Random which
11
initializes a bit randomly at 0 or 1. The gates keepDataA and keepDataB are used
by Alice and Bob respectively to send the bits that they want to keep.
def
=
Alice
Random
A1
A2
[ a : Qubit, dataA : Nat, baseA : Nat .
A1 [a, dataA, baseA] ; fill !a .A2 [dataA, baseA]
] ; Alice
def
=
[ r : Nat .
[ x : Qubit .
(g !0 .end k g ?x .end)\{g} ;
H[x].
(h !Mstd,1 [x] .end k h ?r .end)\{h}
]
]
def
=
[ a : Qubit, dataA : Nat, baseA : Nat .
Random[dataA][a] ;
Random[baseA] ;
[ baseA = 1 → H[a].end ]
]
def
[ dataA : Nat, baseA : Nat .
[ bool : Nat, ok : Nat .
received ?ok .
base !baseA .
keep ?bool .
[ bool = 1 → keepDataA !dataA .end ]
]
]
=
12
B
4
def
=
[ b : Qubit .
[ baseA : Nat, baseB : Nat, dataB : Nat .
Random[baseB] ;
(
[ baseB = 0 → g !Mstd,1 [b] .end,
baseB = 1 → g !M+− [b] .end ]
k g ?dataB .end
)\{g} ;
received !1 .
base ?baseA .
[ baseA = baseB → keep !1 .keepDataB !dataB .end,
baseA 6= baseB → keep !0 .end ]
]
]
Probabilistic branching bisimulation
The operational semantics associates a process graph with a process state. A process graph is a graph where vertices are process states and edges are transitions
labeled with actions or probabilities. Each process graph has an initial state.
A bisimulation is an equivalence relation on process states. It identifies states
when they are associated with process graphs having the same branching structure.
The bisimulation defined here is probabilistic because of probabilities introduced by quantum measurement and branching because some transitions are considered as silent. It is inspired from the definitions in [8] and [1].
4.1
Preliminary definitions and notations
Process states
The set of all possible process states is denoted S. Let S, T ∈ S, then S can be
written P/CP and T , Q/CQ where P , Q are process terms and CP , CQ contexts
(possibly probabilistic).
Assuming that S = P/CP and CP =< s, q = ρ, f >, if x is a qubit in S
and x ∈ q, then ρSx is the state of x and this state can be obtained with a trace out
operation on ρ:
ρSx = T r{x}/q (ρ)
13
Silent transitions
τ
The transitions considered as silent are of course internal transitions ( −→ ) but
also probabilistic transitions. The reason is that we want, for example, the following states S1 and S2 to be equivalent.
S2
S1
0.2
✠
a
❄
T
a
❄
T
Silent transitions will be denoted
empty) of silent transitions.
.
∗
❅
0.8
❅
❘
❅
a
❄
T
stands for a sequence (possibly
Function µ
Probabilistic transitions are considered silent, nonetheless, in two equivalent states, the corresponding actions that can be performed on both sides must occur with
the same probability.
Let ≡ be an equivalence on process states, S be a process state and S̄, its
equivalence class with respect to ≡. If M is a set of process states and S a state,
then S ⊲ M means that there exists a sequence of transitions remaining in M ∪ S̄,
from S to a state of M .
A function µ≡ : S × P(S) → [0, 1] is defined for computing the probability to
reach a state in the set M from a state S without leaving S̄ ∪M . It should be noted
that, for this function to yield a probability, nondeterminism must be eliminated
in a way which allows the computation of µ. Here, nondeterminism is treated as
equiprobability, but this is just a convention for the definition of µ≡ . For example,
it does not imply the equivalence of the following two process states:
S2
S1
a ❅ a
❅
✠
❘
❅
T
T
0.5
✠
a
❄
T
14
❅
0.5
❅❅
❘
a
❄
T
A bisimulation is an equivalence relation ≡ which must verify some properties, among which: if S ≡ T then µ≡ (S, M ) = µ≡ (T, M ) for all M equivalence
class of ≡. In this case, µ≡ (S, M ) is the probability to perform an action.
The function µ≡ is defined by:
• if S ∈ M then µ≡ (S, M ) = 1
• else if ∃ T ∈ M ∪ S̄ such that S −→p T ⊲ M
then let ES = {R ∈ M ∪ S̄ | S −→pR R ⊲ M }
in
X
µ≡ (S, M ) =
pR µ≡ (R, M )
S
p3
T1
T2
a
• else if ∃ T ∈ M ∪ S̄ such thataS −→ T ⊲ M
R
then let ES = {R ∈ M ∪ S̄ | S −→
R ⊲ M } in
1
|ES |
T3
p2
R∈ES
µ≡ (S, M ) =
S
E S p1
X
S
ES
µ≡ (R, M )
a
T1
M
S
c
b
T3
R∈ES
T2
M
• else µ≡ (S, M ) = 0
Equivalence of contexts
Let σ be a renaming on variables. The extensions of σ to environment stacks and
sequences of quantum variables are also called σ.
Let C =< s, q = ρ, f > and C ′ =< s′ , q ′ = ρ′ , f ′ > be two contexts, C and
′
C are equivalent, if and only if there exists a renaming σ such that:
• σ(s) = s′
• ∃ a permutation π such that π(σ(q)) = q ′ and π(ρ) = ρ′
• ∀x ∈ dom(f ), σ(x) ∈ dom(f ′) and f (x) = f ′ (σ(x))
• ∀y ∈ dom(f ′), σ −1 (y) ∈ dom(f ) and f (σ −1 (y)) = f ′ (y)
This equivalence relation can easily be extended to probabilistic contexts.
15
4.2
Probabilistic branching bisimulation: definition
An equivalence relation is a probabilistic branching bisimulation if and only if:
•
if S and T are equivalent and if an action a can be performed from S, then the
same action can performed from T , possibly after several silent transitions;
′
T
S
a
′′
•
the reached states (S and T ) are equivalent;
•
the action a must occur with the same
probability in both cases.
S’
T’
*
a
T’’
Let ≡ be an equivalence relation on process states. ≡ is a probabilistic branching bisimulation if and only if it satisfies:
• Termination
δ
if S ≡ T and S −→ S ′ then
δ
∗
∃ T ′ , T ′′ such that (T
T ′ −→ T ′′ ∧ S ≡ T ′ ∧ S ′ ≡ T ′′ ) and the
′
′′
contexts in S and T are equivalent
• Value sending
g !v
if S ≡ T and S −→ S ′ (v classical value) then
g !v
∃ T ′ , T ′′ such that (T ∗ T ′ −→ T ′′ ∧ S ≡ T ′ ∧ S ′ ≡ T ′′ )
• Qubit sending
g !x
if S ≡ T and S −→ S ′ (x variable) then
g !y
∃ T ′ , T ′′ such that (T ∗ T ′ −→ T ′′ ∧ ρSx = ρTy ∧ S ≡ T ′ ∧ S ′ ≡ T ′′ )
• Value reception
g ?v
if S ≡ T and S −→ S ′ (v classical value) then
g ?v
∃ T ′ , T ′′ such that (T ∗ T ′ −→ T ′′ ∧ S ≡ T ′ ∧ S ′ ≡ T ′′ )
• Qubit reception
g ?x
if S ≡ T and S −→ S ′ (x variable) then
′
′′
g ?y
∃ T ′ , T ′′ such that (T ∗ T ′ −→ T ′′ ∧ ρSx = ρTy ∧ S ≡ T ′ ∧ S ′ ≡ T ′′ )
16
• Silent transition
if S ≡ T and S
S ′ then S ′ ≡ T
• Probabilities
if S ≡ T then µ≡ (S, M ) = µ≡ (T, M ), ∀M ∈ (S/ ≡)\{S̄}
4.3
Bisimulation and recursion
Recursion in a process definition introduces circuits in the associated process
graph. As a consequence, it must be proved that in this case, the notion of probabilistic branching bisimulation is well-defined. In fact, the only point that could
be a problem is the definition of µ≡ .
Let S be a process state and M be a set of process states, the computation of
µ≡ (T, M ) for all T in S̄ leads to a linear system of equations where the µ≡ (T, M )
are the unknowns:
X = AX + B
The ith row in this system can be written: xi = ai1 x1 + · · · + ain xn + bi .
The coefficients aij and bi are either probabilities or average coefficients in
the case of nondeterminism. As a consequence: 0 ≤ aij ≤ 1 and 0 ≤ bi ≤ 1.
Moreover, in the definition of µ≡ , every state in the set ES is such that there exists
a path from that state to the set M . Therefore, the system of equations obtained
can be transformed into a system such that bi > 0, ∀i ∈ J1, nK. From now on, we
consider that the system verifies this property.
P
P
Another property of the system is: 0 ≤ j aij +bi ≤ 1, thus 0 ≤ j aij < 1.
To prove that the system has a unique solution, it is sufficient to prove kAk < 1
and use the fixpoint theorem. The norms for matrices and vectors are:
kAk = sup kAXk
kXk = max |xi |
i
kXk=1
We obtain:
kAXk = max |
i
X
aij xi | ≤ max
i
j
and then kAk ≤ max
i
X
X
(|aij ||xi |)
j
|aij | < 1
j
kAk < 1 implies that the function f : X 7→ AX + B is strictly contracting, so
from the fixpoint theorem, we infer that the equation X = AX + B has a unique
solution. Moreover, as f ([0, 1]) ⊆ [0, 1], this solution belongs to [0, 1].
As a consequence, the function µ≡ is well-defined even in case of recursive
processes.
17
5
Conclusion
This paper has presented a process algebra for quantum programming which can
describe both classical and quantum programming, and their cooperation. Without this cooperation, the implementation of protocols like BB84 is not possible.
Another feature of this language is that measurement and initialization of quantum
registers occur through communications between quantum and classical parts of
the language, which happens to be a faithful model of physical reality.
Moreover, a thorough semantics has been defined, thus allowing the study
and analysis of programs. One peculiarity of this semantics is the introduction
of probabilistic processes, due to quantum measurement. Probabilistic processes
perform probabilistic transitions. As a consequence, the execution tree obtained
from a process presents action branches and probabilistic branches.
Finally a semantical equivalence relation on processes has been defined. This
equivalence is a bisimulation which identifies processes associated with process
graphs having the same branching structure. This is the first step toward the verification of quantum cryptographic protocols.
Several extensions are possible. As already mentioned, a nondeterministic
process composition operator can be introduced. A probabilistic composition of
processes could be added. This would allow, for example, the description of communication protocols in which Eve intercepts qubits with a given probability.
References
[1] S. Andova. Process algebra with probabilistic choice. Lecture Notes in
Computer Science, 1601:111–129, 1999.
[2] P. Arrighi and G. Dowek. Operational semantics for formal tensorial calculus. Proceedings of The 2nd International Workshop on Quantum Programming Languages, pages 21 – 38, July 2004.
[3] C. H. Bennett and G. Brassard. Quantum cryptography: Public-key distribution and coin tossing. In Proceedings of the IEEE International Conference
on Computer, Systems and Signal Processing, pages 175–179, Bangalore,
India, December 1984.
[4] C. H. Bennett, G. Brassard, C. Crépeau, R. Jozsa, A. Peres, and W. Wootters.
Teleporting an unknown quantum state via dual classical and EPR channels.
Physical Review Letters, 70:1895–1899, 1993.
18
[5] T. Bolognesi and E. Brinksma. Introduction to the ISO specification language LOTOS. Computer Networks and ISDN Systems, 14(1):25–59, 1987.
[6] D. Cazorla, F. Cuartero, V. Valero, and F. L. Pelayo. A process algebra
for probabilistic and nondeterministic processes. Information Processing
Letters, 80(1):15–23, 2001.
[7] D. Cazorla, F. Cuartero, V. Valero, F. L. Pelayo, and J. Pardo. Algebraic
theory of probabilistic and nondeterministic processes. The Journal of Logic
and Algebraic Programming, 55(1–2):57–103, 2003.
[8] W. Fokkink. Introduction to Process Algebra. Springer, 2000.
[9] M. Lalire and Ph. Jorrand. A process algebraic approach to concurrent
and distributed quantum computation: operational semantics. Proceedings
of The 2nd International Workshop on Quantum Programming Languages,
pages 109 – 126, July 2004.
[10] R. Milner. Communication and Concurrency. Prentice-Hall, London, 1999.
[11] R. Nagarajan and S. Gay. Communicating quantum processes. Proceedings
of The 2nd International Workshop on Quantum Programming Languages,
pages 91 – 107, July 2004.
[12] B. Omer. Quantum programming in QCL. Master’s thesis, Institute Information System, Technical University of Vienna, 2000.
[13] P. Selinger. Towards a quantum programming language. To appear in Mathematical Structures in Computer Science, 2003.
[14] A. Van Tonder. A lambda calculus for quantum computation. Los Alamos
arXive e-print quant-ph/0307150, 2003.
[15] P. Zuliani. Quantum Programming. PhD thesis, St Cross College, University
of Oxford, 2001.
19
A
The quantum process algebra
A.1 Syntax
process
::=
|
|
|
|
|
|
|
|
nil
end
action . process
process ; process
process k process
process \{ gate list }
[ cond list ]
[ var decl list . process ]
process name [[[ var list ] ]
action
::=
|
|
communication
unit transf
measurement
communication
::=
|
|
gate ! exp
gate ! measurement
gate ? variable
unit transf
::=
unitary operator [ var list ]
measurement
::=
observable [ var list ]
var decl
::=
variable : var type
proc def
::=
process name = process
def
A.2 Main inference rules of the semantics
The semantics is specified with inference rules which give the evolution of the
states of processes. There are four kinds of transitions:
α
• action transition: −→ where α is g !x or g ?x ;
τ
• silent transition: −→ , for internal transition;
δ
• delta transition: −→ , for successful termination;
20
• probabilistic transition: −→p , where p is a probability.
In the following, P, Q, P ′ , Q′ , Pi and Pi′ are processes, C, C ′ and Ci are
contexts, α is an action, g is a communication gate, v is a value, x is a variable,
and cj is a condition.
Successful termination
end/C
−−δ−→ nil/C
C↓
Action Prefix
g !v .P/C
g !v
P/C
−−−−−−→
v ∈ IN, C ↓
g !x .P/C
g !f (x)
P/C
−−−−−−−−→
where C =< s, q = ρ, f >, x ∈ Var(s) and x ∈ dom(f )
g !x .P/C
g !x
−−−−−−→
P/C ′
where
• C =< s, q = ρ, f >, C ′ =< s\{x}, q\{x} = T r{x}/q (ρ), f >
• x ∈ Var(s) and x ∈ q
g ?x .P/C
g ?v
P/C ′
−−−−−−→
where
• C =< s, q = ρ, f >, C ′ =< s, q = ρ, f ∪ {x 7→ v} >
• x ∈ Var(s), x of type Nat, v ∈ IN
g ?x .P/C
g ?x
P/C ′
−−−−−−→
where
• C =< s, q = ρ, f >, C ′ =< s, x.q = ν ⊗ ρ, f >
• x ∈ Var(s), x of type Qubit, x 6∈ q
21
• ν density matrix of dimension 2
U [x1 , . . . , xn ].P/C
′
−−τ−→ P/C
where
• C =< s, q = ρ, f >, C ′ =< s, q = ρ′ , f >
• U ∈ U, x1 , . . . , xn ∈ Var(s), and x1 , . . . , xn ∈ q
• ∀ i, j ∈ {0, . . . , n} such that i 6= j : xi 6= xj
• ρ′ = TU (ρ)
T is defined in the following way: if A is a 2n × 2n matrix, then
TA : ρ 7→ Π† .(A ⊗ I ⊗k ).Π.ρ.Π† .(A† ⊗ I ⊗k ).Π
where Π is the permutation matrix which places the xi ’s at the head of q, and
k = size(q) − n .
M [x1 , . . . , xn ].P/ < s, q = ρ, f >
−−τ−→ P/ < s, q = ρ, f >
with
• x1 , . . . , xn ∈ Var(s) and x1 , . . . , xn ∈ q
• ∀ i, j ∈ {0, . . . , n} such that i 6= j : xi 6= xj
P
• M ∈ O with i λi Pi as spectral decomposition
P
• ρ′ = i TPi (ρ)
g !M [x1 , . . . , xn ] .P/C
−−τ−→ [ g !y .end ] ; P/ ⊞pi Ci
where
• C =< s, q = ρ, f > (which implies C ↓)
• Ci =< {(y, Nat)}.s, q = ρi , f ∪ {y 7→ λi } >
• x1 , . . . , xn ∈ Var(s) and x1 , . . . , xn ∈ q
22
• ∀ i, j ∈ {0, . . . , n} such that i 6= j : xi 6= xj
• y is a new variable
P
• M ∈ O with i λi Pi as spectral decomposition
• pi = T r(TPi (ρ)),
ρi =
1
pi TPi (ρ)
Probabilistic contexts
P/ ⊞pi Ci
−→pj
P/Cj
where
X
pi = 1
i
Sequential composition
P/C
P ; Q/C
α
P ′ /C ′
−−−→
′
′
−−α
−→ P ; Q/C
α 6= δ
′
′
−−δ−→ P /C
P ; Q/C −−τ−→ Q/C ′
P/C
Parallel composition
In the rules for parallel composition, C, CP and CQ are defined as:
• C =< (sP k sQ ).s, q = ρ, f >
• CP =< sP |s, q = ρ, f >
• CQ =< sQ |s, q = ρ, f >
In the definition of C, the operator k permits to build a cactus stack (see paragraph 2.1). In the cactus stack (sP k sQ ).s of the process P k Q, the names in s
correspond to variables shared by P and Q whereas the names in sP (resp. sQ )
correspond to variables declared in P (resp. Q).
P/CP
P k Q/C
′
′
−−α
−→ P /CP
α
P ′ k Q/C ′
−−−→
α 6= δ
where
• If CP′ =< s′ , q ′ = ρ′ , f ′ > then C ′ =< (s′P k sQ ).s, q ′ = ρ′ , f ′ > with
s′P such that s′ = s′P |s (P can neither add to nor remove variables from s)
23
• If CP′ = ⊞pi < s′i , qi′ = ρ′i , fi′ > then C ′ = ⊞pi < (sP ′i k sQ ).s, qi′ =
ρ′i , fi′ > with sP ′i such that s′i = sP ′i |s
′
g !v
P ′ /CP′
Q/CQ
g ?v
Q′ /CQ
−−−−−−→
−−−−−−→
P k Q/C
τ
P ′ k Q′ /C ′
−−−→
′
where v ∈ IN , CQ
=< s′ , q ′ = ρ′ , f ′ >, and C ′ =< (sP k sQ ).s, q = ρ, f ′ >
P/CP
P/CP
′
g !v
P ′ /CP′
Q/CQ
g ?x
Q′ /CQ
−−−−−−→
−−−−−−→
P k Q/C −−τ−→ P ′ k Q′ /C ′
where
• x ∈ Var(s) ∪ Var(sQ ), x of type Qubit, x 6∈ q, v ∈ {0, 1}
• C ′ =< (sP k sQ ).s, x.q = |vihv| ⊗ ρ, f >
P/CP
′
g !x
P ′ /CP′
Q/CQ
g ?y
Q′ /CQ
−−−−−−→
−−−−−−→
P k Q/C −−τ−→ P ′ k Q′ /C ′
where
• x ∈ Var(s) ∪ Var(sP ), x ∈ q
• y ∈ Var(s) ∪ Var(sQ ), y 6∈ q, y of type Qubit
• C ′ =< ((sP k sQ ).s)\{x}, q[x ← y] = ρ, f >
′
′
′
Q/CQ −−δ−→ Q′ /CQ
−−δ−→ P /CP
P k Q/C −−δ−→ nil/C ′
with C ′ =< s, q\e = T re/q (ρ), f \e > and e = (Var(sP ) ∪ Var(sQ ))
P/CP
Variable declaration
′
−−τ−→ [ P ] /C
with C =< s, q = ρ, f >, C ′ =< s′ , q = ρ, f >
and s′ = {(x1 , t1 ), . . . , (xn , tn )}.s
[ x1 : t1 , . . . , xn : tn . P ] /C
24
End of scope of variables
′
′
−−α
−→ P /C
α 6= δ
[ P ] /C
α
[ P ′ ] /C ′
−−−→
P/C −−δ−→ P ′ / < e.s, q = ρ, f >
P/C
[ P ] /C
−−δ−→ nil / < s, q\Var(e) = T rVar(e)/q (ρ), f \Var(e) >
25
| 2cs.AI
|
ON THE LENGTH AND DEPTH OF FINITE GROUPS
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
arXiv:1802.02194v1 [math.GR] 6 Feb 2018
With an appendix by D.R. Heath-Brown
Abstract. An unrefinable chain of a finite group G is a chain of subgroups G = G0 >
G1 > · · · > Gt = 1, where each Gi is a maximal subgroup of Gi−1 . The length (respectively, depth) of G is the maximal (respectively, minimal) length of such a chain. We
studied the depth of finite simple groups in a previous paper, which included a classification of the simple groups of depth 3. Here we go much further by determining the
finite groups of depth 3 and 4. We also obtain several new results on the lengths of finite
groups. For example, we classify the simple groups of length at most 9, which extends
earlier work of Janko and Harada from the 1960s, and we use this to describe the structure of arbitrary finite groups of small length. We also present a number-theoretic result
of Heath-Brown, which implies that there are infinitely many non-abelian simple groups
of length at most 9.
Finally we study the chain difference of G (namely the length minus the depth). We
obtain results on groups with chain difference 1 and 2, including a complete classification
of the simple groups with chain difference 2, extending earlier work of Brewster et al.
We also derive a best possible lower bound on the chain ratio (the length divided by the
depth) of simple groups, which yields an explicit linear bound on the length of G/R(G)
in terms of the chain difference of G, where R(G) is the soluble radical of G.
1. Introduction
An unrefinable chain of length t of a finite group G is a chain of subgroups
G = G0 > G1 > · · · > Gt−1 > Gt = 1,
(1)
where each Gi is a maximal subgroup of Gi−1 . The length of G, denoted by l(G), is the
maximal length of an unrefinable chain. This notion arises naturally in several different
contexts, finding a wide range of applications. For example, Babai [3] investigated the
length of symmetric groups in relation to the computational complexity of algorithms for
finite permutation groups. In a different direction, Seitz, Solomon and Turull studied the
length of finite groups of Lie type in a series of papers in the early 1990s [30, 32, 33],
motivated by applications to fixed-point-free automorphisms of finite soluble groups. In
fact, the notion predates both the work of Babai and Seitz et al. Indeed, Janko and
Harada studied the simple groups of small length in the 1960s, culminating in Harada’s
description of the finite simple groups of length at most 7 in [16].
Given the definition of l(G), it is also natural to consider the minimal length of an
unrefinable chain for G. Following [8], we call this number the depth of G, denoted by
λ(G). For example, if G is a cyclic group of order n > 2, then λ(G) = Ω(n), the number
of prime divisors of n (counting multiplicities). In particular, λ(G) = 1 if and only if G
has prime order. This notion appears in the earlier work of several authors. For example,
in [31] Shareshian and Woodroofe investigate the length of various chains of subgroups
The third author acknowledges the hospitality of Imperial College, London, while part of this work was
carried out. He also acknowledges the support of ISF grant 686/17 and the Vinik chair of mathematics
which he holds.
Date: February 8, 2018.
2010 Mathematics Subject Classification. Primary 20E32, 20E15; Secondary 20E28.
1
2
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
of finite groups G in the context of lattice theory (in their paper, the depth of G is
denoted by minmaxl(G)). There are also several papers on the so-called chain difference
cd(G) = l(G) − λ(G) of a finite group G. For example, a well known theorem of Iwasawa
[18] states that cd(G) = 0 if and only if G is supersoluble. The simple groups G with
cd(G) = 1 have been determined by Brewster et al. [7] (also see [17] and [28] for related
results).
In [8], we focus on the depth of finite simple groups. One of the main results is [8,
Theorem 1], which determines the simple groups of depth 3 (it is easy to see that λ(G) > 3
for every non-abelian simple group G); the groups that arise are recorded in Table 1. We
also show that alternating groups have bounded depth (indeed, λ(An ) 6 23 for all n,
whereas l(An ) tends to infinity with n) and we obtain upper bounds on the depth of each
simple group of Lie type. We refer the reader to [9] for results on analogous notions of
length and depth for connected algebraic groups over algebraically closed fields.
G
Ap
L2 (q)
Lǫn (q)
2B
2 (q)
M23 , B
Conditions
p and (p − 1)/2 prime, p 6∈ {7, 11, 23}
(q + 1)/(2, q − 1) or (q − 1)/(2, q − 1) prime, q 6= 9; or
q prime and q ≡ ±3, ±13 (mod 40); or
q = 3k with k > 3 prime
n −ǫ
both prime, n > 3 and
n and (q−ǫ)q (n,q−ǫ)
(n, q, ǫ) 6= (3, 4, +), (3, 3, −), (3, 5, −), (5, 2, −)
q − 1 prime
Table 1. The simple groups G with λ(G) = 3
Our goal in this paper is to extend the depth results in [8] in several different directions,
both for simple groups, as well as arbitrary finite groups. We also revisit some of the
aforementioned results of Janko and Harada from the 1960s, providing a precise description
of the simple groups of small length. In turn, this allows us to describe the structure of
arbitrary finite groups of small length and we can use this to classify the simple groups G
with cd(G) = 2, which extends one of the main results in [7].
By a theorem of Shareshian and Woodroofe [31, Theorem 1.4], it follows that λ(G) > 3
for every insoluble finite group G. Our first main result determines all finite groups of
depth 3. In particular, notice that an obvious consequence of the theorem is that almost
simple groups of depth 3 are simple.
Theorem 1. A finite group G has depth 3 if and only if either G is soluble of chief length
3, or G is a simple group as in Table 1.
The next result classifies the finite groups of depth 4. In part (iv) of the statement, a
twisted wreath product T twrφ S for non-abelian simple groups S, T is as defined and studied
in [4]. The ingredients are a transitive action of S on k points with point stabiliser S1 , and
a homomorphism φ : S1 → Aut(T ) with image containing Inn(T ). Thus T is isomorphic
to a proper section of S; indeed, the subgroups C = φ−1 (Inn(T )) and D = ker(φ) ∩ C
satisfy C/D ∼
= T , and (C, D) forms an S-maximal section of S, as defined in [4, Definition
4.1]. Moreover T twrφ S is a semidirect product T k .S having S as a maximal subgroup.
Theorem 2. Suppose G is a finite group of depth 4. Then one of the following holds,
where p is a prime:
(i) G is soluble of chief length 4.
ON THE LENGTH AND DEPTH OF FINITE GROUPS
3
(ii) G = T × T , T × Cp or T p .Cp , where T is simple of depth 3 (as in Table 1).
(iii) G = (Cp )k .T (a possibly nonsplit extension), where T is simple of depth 3, and
acts irreducibly on (Cp )k .
(iv) G = T twrφ S, a twisted wreath product, where S, T are simple, T is a proper
section of S, and S has depth 3.
(v) G is quasisimple, and Z(G) is either 1 or Cp .
(vi) G is almost simple with socle T , and G/T is either 1 or Cp .
Remark 1. Let us make some comments on the statement of Theorem 2.
(a) Note that in part (iv), and also (ii) (apart from T p .Cp ), the group G has a maximal
subgroup that is simple of depth 3. In (iii), for split extensions this is also the case;
for nonsplit, G must have a maximal subgroup M = pk .M0 with M0 maximal in
T acting irreducibly on pk (so λ(M0 ) = 2 and λ(M ) = 3). See Remark 3.1 for
further comments on the nonsplit groups G = pk .T arising in part (iii).
(b) In (iv), being of depth 3, the possibilities for S can be found in Table 1. Every
proper non-abelian simple section T of S can occur. The simple sections of the
groups Ap and Lǫn (q) in Table 1 cannot be listed. However, the proper simple
sections of the other groups in the table can be listed: for L2 (q) they are A5 ,
L2 (q0 ) (with q = q0k and q0 > 4); for 2B2 (q) there are none (since q − 1 is prime);
and those for M23 and B can be listed using [12].
(c) Consider case (v), where G is quasisimple with Z(G) = Cp (p prime) and let
T = G/Z(G), a simple group. Here λ(G) = λ(T ) + 1, as explained in Remark 3.2,
so λ(G) = 4 if and only if T is one of the simple groups in Table 1. In particular, by
considering the Schur multipliers of the simple groups in Table 1 (see [21, Theorem
5.1.4], for example), we deduce that a quasisimple group with nontrivial centre has
depth 4 if and only if it appears in Table 2.
G
2.Ap
SLǫn (q)
SL2 (q)
2.2 B2 (8), 2.B
Conditions
p and (p − 1)/2 prime, p 6∈ {7, 11, 23}
q n −ǫ
both prime, n > 3 and
n = (n, q − ǫ) and n(q−ǫ)
(n, q, ǫ) 6= (3, 4, +), (3, 5, −)
(q + 1)/2 or (q − 1)/2 prime, q 6= 9; or
q prime and q ≡ ±3, ±13 (mod 40); or
q = 3k with k > 3 prime
Table 2. The quasisimple groups G with Z(G) 6= 1 and λ(G) = 4
The next result sheds further light on the almost simple groups of depth 4 (case (vi) of
Theorem 2).
Theorem 3. Let G be an almost simple group with socle T and depth 4. Then one of the
following holds:
(i) G/T = Cp , p prime and λ(T ) = 3;
(ii) (G, T ) is one of the cases in Table 3 (in each case, λ(T ) = 4);
(iii) G = T has a soluble maximal subgroup M of chief length 3; the possibilities for
(G, M ) are listed in Table 4;
(iv) G = T has a simple maximal subgroup of depth 3.
4
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
In Table 4, the required conditions when G = L2 (q) and q is odd are rather complicated
to state (mainly due to the fact that we need λ(G) 6= 3). To simplify the presentation of
the table, we refer to the following conditions on q:
Ω(q ± 1) > 3
q 6≡ ±3, ±13 (mod 40) if q is prime
(2)
q 6= 3k with k > 3 prime.
We refer the reader to Remark 3.3 for further details on the simple groups that arise in
part (iv) of Theorem 3.
T
A6
A7 , A11 , A23
L2 (q)
L3 (4)
U3 (5)
G
Conditions
PGL2 (9), M10
S7 , S11 , S23
PGL2 (q)
q prime, q ≡ ±11, ±19 (mod 40), Ω(q ± 1) > 3
PGL3 (4)
PGU3 (5)
Table 3. The almost simple groups G = T.p with λ(G) = λ(T ) = 4
G
Ap
A6
L2 (q)
Lǫ3 (q)
Lǫn (q)
2B
2 (q)
M
p:((p − 1)/2)
S4 , 32 :4
Fq :((q − 1)/2), Dq−1
Dq+1
S4
Fq :(q − 1), D2(q−1)
D2(q+1)
(Cq−ǫ )2 :S3
q n −ǫ
(q−ǫ)(n,q−ǫ)
:n
D2(q−1)
√
(q ± 2q + 1):4
√
2 G (q) (q ± 3q + 1):6
2
3 D (q) (q 4 − q 2 + 1):4
4
J1
7:6, 11:10, 19:6, 23 :7:3
J4
43:14
Ly
67:22
Fi′24
29:14
Th
31:15
Conditions
p prime, Ω(p − 1) = 3
q
q
q
q
q
q
odd, Ω(q − 1) = 3, (2) holds
odd, Ω(q + 1) = 3, (2) holds
prime, q ≡ ±1 (mod 8), (2) holds
even, Ω(q − 1) = 2, Ω(q + 1) > 2
even, Ω(q + 1) = 2, Ω(q − 1) > 2
> 8 even, q − ǫ prime, Ω(q 2 + ǫq + 1) > 2
n
q −ǫ
n > 3 prime, Ω( (q−ǫ)(n,q−ǫ)
)=2
Ω(q − 1) = 2
√
q ± 2q + 1 prime, Ω(q − 1) > 2
√
q ± 3q + 1 prime, q > 3
q 4 − q 2 + 1 prime
Table 4. The simple groups G of depth 4 with a soluble maximal subgroup
M of depth 3
Next we turn to our main results on the lengths of finite groups. Recall that the finite
simple groups of small length were studied by Janko and Harada in the 1960s, beginning
with [19] which classifies the simple groups of length 4 (since λ(G) > 3, Iwasawa’s theorem
implies that l(G) > 4 for every non-abelian simple group G). In a second paper [20], Janko
describes the simple groups of length 5 and this was extended by Harada [16] to length at
ON THE LENGTH AND DEPTH OF FINITE GROUPS
5
most 7. In both papers, the main results state that either G = L2 (q) for some unspecified
prime powers q, or G belongs to a short list of specific groups.
The following result extends this earlier work by giving a precise classification of the
simple groups of length at most 9.
Theorem 4. Let G be a non-abelian finite simple group of length at most 9. Then G and
l(G) are recorded in Table 5, where p is a prime number.
l(G)
4
5
6
7
8
9
G
A5 , L2 (q)
L2 (q)
A7 , J1 , L2 (q)
M11 , U3 (3), U3 (5)
L2 (q)
Conditions
q = p > 5, max{Ω(q ± 1)} = 3 and q ≡ ±3, ±13 (mod 40)
q ∈ {7, 8, 9, 11, 19, 27, 29}, or q = p and max{Ω(q ± 1)} = 4
q ∈ {25, 125}, or q = p and max{Ω(q ± 1)} = 5
q ∈ {16, 32, 49, 121, 169}, or q = p and max{Ω(q ± 1)} = 6, or
q = p3 , Ω(q − 1) = 4 and Ω(q + 1) 6 6
M12 , 2 B2 (8), L3 (3)
L2 (q)
q
q
q
q
q
A8 , U4 (2), L3 (4)
U3 (q)
q
q
L2 (q)
q
q
q
q
q
q
q
= p and max{Ω(q ± 1)} = 7, or
= p2 , Ω(q − 1) = 6 and Ω(q + 1) 6 7, or
= p3 , Ω(q − 1) = 5 and Ω(q + 1) 6 7, or
= p3 , Ω(q − 1) 6 4 and Ω(q + 1) = 7, or
= p5 , Ω(q − 1) = 3 and Ω(q + 1) 6 7
∈ {4, 11, 13, 29}, or q = p, Ω(q ± 1) = 3, Ω(q 2 − q + 1) 6 8,
≡ 2 (mod 3) and q ≡ ±3, ±13 (mod 40)
∈ {81, 128, 2187}, or q = p and max{Ω(q ± 1)} = 8, or
= p2 , Ω(q − 1) = 7 and Ω(q + 1) 6 8, or
= p2 , Ω(q − 1) = 6 and Ω(q + 1) = 8, or
= p3 , Ω(q − 1) = 6 and Ω(q + 1) 6 8, or
= p3 , Ω(q − 1) 6 5 and Ω(q + 1) = 8, or
= p5 , Ω(q − 1) = 4 and Ω(q + 1) 6 8, or
= p5 , Ω(q − 1) = 3 and Ω(q + 1) = 8
Table 5. The simple groups G of length at most 9
The proof of Theorem 4 is given in Section 4.1, together with the proof of the following
corollary, which describes the structure of finite groups of small length. Recall that soluble
groups G have length Ω(|G|), so we focus on insoluble groups.
Corollary 5. Let G be a finite insoluble group, in which case l(G) > 4.
(i) l(G) = 4 if and only if G is simple as in line 1 of Table 5.
(ii) l(G) = 5 if and only if one of the following holds:
(a) G is simple as in line 2 of Table 5; or
(b) G = T × Cp with T simple of length 4 (as in Table 5) and p a prime; or
(c) G = SL2 (q) or PGL2 (q), and either q = 5, or q > 5 is a prime such that
max{Ω(q ± 1)} = 3 and q ≡ ±3, ±13 (mod 40).
(iii) l(G) = 6 if and only if one of the following holds:
(a) G is simple as in line 3 of Table 5; or
6
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
(b) G = T ×Cp , or a quasisimple group p.T , or an almost simple group T.p, where
T is simple of length 5 (as in Table 5) and p a prime; the quasisimple groups
occurring are SL2 (q), 3.L2 (9), and the almost simple groups are PGL2 (q),
M10 , S6 , L2 (8).3 and L2 (27).3; or
(c) G = L2 (q) × (p.r), (L2 (q) × p).2, SL2 (q) × p or 2.L2 (q).2, where p, r are primes
and L2 (q) has length 4, as in Table 5.
Let G = L2 (q), where q is a prime, and consider the conditions on q in the first row
of Table 5. One checks that the first ten primes that satisfy the given conditions are as
follows:
q ∈ {13, 43, 67, 173, 283, 317, 653, 787, 907, 1867},
but it is not known if there are infinitely many such primes. The following more general
problem is addressed in [1]: Does there exist an infinite set S of non-abelian finite simple
groups and a positive integer N such that l(G) 6 N for all G ∈ S? The main result of [1]
gives a positive answer to this question. The key ingredient is a purely number theoretic
result [1, Theorem C], which states that for each positive integer n, there is an infinite
set of primes P and a positive integer N such that Ω(pn − 1) 6 N for all p ∈ P. More
precisely, for n = 2 they show that the conclusion holds with N = 21, which immediately
implies that there are infinitely many primes p with l(L2 (p)) 6 20. The same problem
arises in work of Gamburd and Pak (see [14, p.416]), who state that l(L2 (p)) 6 13 for
infinitely many primes p (giving [15] as a reference).
We establish the following strengthening of the results in [1, 14].
Theorem 6. There are infinitely many finite non-abelian simple groups G with l(G) 6 9.
In fact we show that l(L2 (p)) 6 9 for infinitely many primes p. As explained in Section
4.2, this is easily deduced from the following number-theoretic result of Heath-Brown,
which is of independent interest.
Theorem 7 (Heath-Brown). There are infinitely many primes p ≡ 5 (mod 72) for which
Ω((p2 − 1)/24)) 6 7.
See Appendix A for the proof of this theorem, which implies that there are infinitely
many primes p for which max{Ω(p ± 1)} 6 8.
Next we consider the chain difference of simple groups. The following result determines
the simple groups of chain difference two (see Section 5.1 for the proof). This extends
earlier work of Brewster et al. [7, Theorem 3.3] (also see Theorem 5.1), who described the
simple groups of chain difference one.
Theorem 8. Let G be a finite simple group. Then cd(G) = 2 if and only if one of the
following holds:
(i) G = A7 , J1 or U3 (5).
(ii) G = L2 (q) and either q ∈ {7, 8, 11, 27, 125}, or q is a prime and one of the following
holds:
(a) max{Ω(q ± 1)} = 4 and either min{Ω(q ± 1)} = 2, or q ≡ ±3, ±13 (mod 40).
(b) max{Ω(q ± 1)} = 5, min{Ω(q ± 1)} > 3 and q 6≡ ±3, ±13 (mod 40).
The chain ratio of a finite group G is given by cr(G) = l(G)/λ(G) and the next result establishes a best possible lower bound on the chain ratio of simple groups. In the
statement, we define
G = {G : G simple, l(G) = 5 and λ(G) = 4}.
(3)
ON THE LENGTH AND DEPTH OF FINITE GROUPS
7
By combining Theorem 4 with [8, Theorem 1], it follows that the groups in G are of the form
L2 (q) and either q ∈ {9, 19, 29}, or q is a prime with max{Ω(q ±1)} = 4, min{Ω(q ±1)} > 3
and q 6≡ ±3, ±13 (mod 40).
Theorem 9. Let G be a non-abelian finite simple group. Then
5
cr(G) > ,
4
with equality if and only if G ∈ G.
It follows from [8, Corollary 9] that there exists an absolute constant a such that l(G) 6
a cd(G) for every non-abelian simple group G. As an immediate corollary of Theorem 9,
we deduce that a = 5 is the best possible constant.
Corollary 10. Let G be a non-abelian finite simple group. Then l(G) 6 5 cd(G), with
equality if and only if G ∈ G.
Our final result, which applies Theorem 9, relates the structure of an arbitrary finite
group G with its chain difference. We let R(G) denote the soluble radical of G.
Theorem 11. Let G be a finite group. Then
l(G/R(G)) 6 10 cd(G).
Combining this theorem with [1, Proposition 2.2], it follows that
Ω(|G/R(G)|) 6 100 cd(G)2 .
Note that the length of G itself need not be bounded in terms of cd(G); indeed, if G
is supersoluble then cd(G) = 0 while l(G) may be arbitrarily large. However, we show
in Proposition 5.10 below, that, if ss(G) denotes the direct product of the non-abelian
composition factors of G (with multiplicities), then
l(ss(G)) 6 5 cd(G).
This extends Corollary 10 dealing with simple groups, and serves as a useful tool in the
proof of Theorem 11 above.
2. Preliminaries
We begin by recording some preliminary results, which will be needed in the proofs of
our main theorems. Given a finite group G, we write chiefl(G) for the length of a chief
series of G. Recall that l(G) and λ(G) denote the length and depth of G, respectively, as
defined in the Introduction. Let cd(G) = l(G) − λ(G) be the chain difference of G.
Lemma 2.1. Let G be a finite group and let N be a normal subgroup of G.
(i) l(G) = l(N ) + l(G/N ).
(ii) λ(G/N ) 6 λ(G) 6 λ(N ) + λ(G/N ).
Proof. This is straightforward. For example, see [10, Lemma 2.1] for part (i).
Notice that Lemma 2.1(i) implies that the length of a finite group is equal to the sum
of the lengths of its composition factors. In particular, if G is soluble then l(G) = Ω(|G|),
which is the number of prime divisors of |G| (counting multiplicities).
The next result is [7, Lemma 1.3].
Lemma 2.2. If G is a finite group, B 6 G and A is a normal subgroup of B, then
cd(G) > cd(B/A) + cd(A).
In particular, cd(G) > cd(L) for every section L of G.
8
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
Lemma 2.3. Let G be a finite group.
(i) If G is soluble, then λ(G) = chiefl(G).
(ii) If G is insoluble, then λ(G) > chiefl(G) + 2.
Proof. Part (i) is [22, Theorem 2] and part (ii) is [31, Theorem 1.4].
Lemma 2.4. Let H be soluble, p a prime, and suppose G = H p hαi where αp ∈ H p and α
permutes the p factors transitively. Then λ(G) > λ(H) + 1.
Proof. We proceed by induction on |H|. The base case H = 1 is trivial. Now assume
H 6= 1, and let N be a minimal normal subgroup of H. By Lemma 2.3(i) we have
λ(H/N ) = λ(H) − 1.
Q
αi
Let M = p−1
i=0 N . Then 1 6= M ⊳ G, and λ(G) > λ(G/M ) + 1, again by Lemma 2.3(i).
Applying the induction hypothesis to G/M ∼
= (H/N )p .p, we have
λ(G/M ) > λ(H/N ) + 1.
It follows that λ(G) > λ(H/N ) + 2 = λ(H) + 1, as required.
The next lemma on the length of L2 (q) will be useful later.
Lemma 2.5. Let G = L2 (q), where q = pf > 5 and p is a prime.
(i) If q is even, then l(G) = Ω(q − 1) + f + 1.
(ii) If q is odd, then either
l(G) = max{Ω(q − 1) + f, Ω(q + 1) + 1},
(4)
or q ∈ {7, 11, 19, 29} and l(G) = 5, or q = 5 and l(G) = 4.
Proof. Part (i) is a special case of [32, Theorem 1], noting that 2f :(2f − 1) is a Borel
subgroup of G. Now assume q is odd. The case f = 1 follows from [10, Proposition 5.2],
so let us assume f > 2. We proceed by induction on Ω(f ).
First assume Ω(f ) = 1, so f is a prime, and let M be a maximal subgroup of G. By
inspecting [6, Tables 8.1, 8.2], either M = PGL2 (p) or A5 (for f = 2 only), or M =
pf :((pf − 1)/2), Dpf ±1 or L2 (p), which gives
l(G) = max{Ω(q − 1) + f, Ω(q + 1) + 1, l(L2 (p)) + 1 + δ2,f },
where δi,j is the familiar Kronecker delta. It is easy to check that (4) holds if p ∈
{3, 5, 7, 11, 19, 29}. For example, if p = 29 and f = 2, then Ω(q − 1) = 6 and l(L2 (p)) = 5.
For any other prime p,
l(L2 (p)) + 1 + δ2,f = max{Ω(p ± 1)} + 2 + δ2,f 6 max{Ω(q − 1) + f, Ω(q + 1) + 1}
and the result follows.
Similarly, if Ω(f ) > 2 then
l(G) = max{Ω(q − 1) + f, Ω(q + 1) + 1, l(L2 (q 1/r )) + 1 + δ2,r : r ∈ π(f )},
where π(f ) is the set of prime divisors of f , and induction gives
Therefore
l(L2 (q 1/r )) = max{Ω(q 1/r − 1) + f /r, Ω(q 1/r + 1) + 1}.
l(L2 (q 1/r )) + 1 + δ2,r 6 max{Ω(q − 1) + f, Ω(q + 1) + 1}
and we conclude that (4) holds.
3. Depth
In this section we prove our results on depths, namely Theorems 1, 2 and 3.
ON THE LENGTH AND DEPTH OF FINITE GROUPS
9
3.1. Proof of Theorem 1. Let G be a finite group. For soluble groups, the theorem is
an immediate corollary of Lemma 2.3(i), so let us assume G is insoluble. Here Lemma
2.3(ii) implies that λ(G) > chiefl(G) + 2. Hence if λ(G) = 3, then chiefl(G) = 1 and thus
G is simple, and is as in Table 1 by [8, Theorem 1]. Conversely, the groups in the table
indeed have depth 3.
3.2. Proof of Theorem 2. Suppose G is a finite group and λ(G) = 4. If G is soluble
then it has chief length 4 by Lemma 2.3(i), as in part (i) of Theorem 2. Now assume G is
insoluble. Since λ(G) > chiefl(G) + 2 by Lemma 2.3(ii), it follows that chiefl(G) 6 2. If
chiefl(G) = 1 then G is simple, so (v) holds.
Now assume that chiefl(G) = 2. Then G has a minimal normal subgroup N ∼
= Tk
for some simple (possibly abelian) group T , and G/N ∼
= S is simple (also possibly nonabelian).
Suppose first that k = 1 and S, T are both non-abelian. Then G ∼
= S ×T by the Schreier
hypothesis. When S ∼
6 T , any maximal subgroup of G is of the form S0 ×T or S ×T0 (with
=
S0 , T0 maximal in S, T respectively), and neither of these can have depth 3, by Theorem
1. Hence S ∼
= T . Now G has a maximal subgroup M of depth 3, and M cannot be of the
above form S0 × T or S × T0 . It follows that M is a diagonal subgroup isomorphic to T ,
and hence λ(T ) = 3 and G is as in conclusion (ii) of Theorem 2.
Next suppose k = 1 and S or T is an abelian simple group Cp . Then G is one of T × Cp ,
S ×Cp, a quasisimple group p.S or an almost simple group T.p. The latter two possibilities
are conclusions (v) and (vi). Now assume G = T × Cp , and let M be a maximal subgroup
of G of depth 3. Then M is either T or T0 × Cp , where T0 is maximal in T . In the
latter case Theorem 1 shows that T0 × Cp is soluble of chief length 3, hence chiefl(T0 ) = 2.
Therefore in both cases T has depth 3, and so G is as in conclusion (ii).
We may now assume that k > 1. Suppose T is non-abelian and S = Cp . Then k = p
and G = N hαi = T p hαi, where αp ∈ N and α permutes the p factors transitively. Let M
be a maximal subgroup of G of depth 3. Then M 6= N by Theorem 1, so, replacing α by
Q
αi
∼ p
another element in the coset N α if necessary, M is of the form p−1
i=0 H hαi = H .p for
some maximal subgroup H of T . Also M is soluble, again by Theorem 1. It now follows
from Lemma 2.4 that λ(M ) > λ(H) + 1, and hence λ(H) 6 2. Therefore λ(T ) 6 3, so
λ(T ) = 3 and conclusion (ii) of Theorem 2 holds.
Next consider the case where T = Cp and S is non-abelian. Here G = (Cp )k .S, where
S acts irreducibly on V := (Cp )k . Let M be a maximal subgroup of G of depth 3. If
V 66 M then M maps onto S, hence M ∼
= S by Theorem 1, and conclusion (iii) holds.
Now assume V 6 M , so that M is soluble, by Theorem 1. Then M/V = S0 , a maximal
subgroup of S, and by Lemma 2.3(i), λ(S0 ) < λ(M ) = 3. Hence λ(S) = 3 and again (iii)
holds.
It remains to handle the case where both T and S are non-abelian. Here G = T k .S
and S acts transitively on the k factors. Let M be a maximal subgroup of G of depth 3.
Then T k 66 M by Theorem 1. Hence M maps onto S, so M ∼
= S, again by Theorem 1. In
particular, λ(S) = 3. Write Ω = (G : M ), the coset space of M in G. As M is a core-free
maximal subgroup of G, it follows that G acts primitively on Ω and T k is a regular normal
subgroup. At this point the O’Nan-Scott theorem (see [24], for example) implies that G
is a twisted wreath product T twrφ S, as in conclusion (iv) of Theorem 2.
This completes the proof of Theorem 2.
Remark 3.1. Let us consider the nonsplit groups G = pk .T arising in part (iii) of Theorem
2. Here T is a simple group of depth 3 (so the possibilities for T are given in Table 1) and
V = (Cp )k is a nontrivial irreducible module for T over Fq with q = pf for a prime p. In
10
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
particular, dim V > 2, p divides |T | and the second cohomology H 2 (T, V ) is nontrivial.
As noted in Remark 1(a), G has a maximal subgroup M = pk .S with λ(M ) = 3, where
S < T is maximal and acts irreducibly on V . Note that S is soluble and has chief length
2. It will be difficult to give a complete classification of the depth 4 groups of this form,
but we can identify some genuine examples:
Example. Let T = M23 , so S = 23:11. Now T has an 11-dimensional irreducible module
V over F2 . Moreover, one checks that S acts irreducibly on V and H 2 (T, V ) 6= 0, hence
there is a nonsplit group G = 211 .M23 of depth 4.
Example. Take T = A5 and S = A4 . Set q = 5f with f > 1 and let V be a 3-dimensional
irreducible module for T over Fq . Then S acts irreducibly on V and H 2 (T, V ) 6= 0, so for
each f > 1 there is a nonsplit group 53f .A5 of depth 4.
n
−1
:n.
Example. Suppose T = Ln (r), where n > 3 is a prime and (n, r −1) = 1, so S = rr−1
Let V be the natural module for T = SLn (r). Then S acts irreducibly on V , and a theorem
of Bell [5] implies that H 2 (T, V ) 6= 0 if and only if
(n, r) ∈ {(3, 3a > 3), (3, 2), (3, 5), (4, 2), (5, 2)}.
In particular, there is a nonsplit group 39 .L3 (27) of depth 4 (note that we need λ(L3 (r)) =
3, which in this case means that r 2 + r + 1 is a prime). Thanks to Bell’s result, there are
also nonsplit groups 23 .L3 (2), 53 .L3 (5) and 25 .L5 (2), each of which has depth 4.
Remark 3.2. Let G be a quasisimple group with Z(G) = Cp (p prime) and G/Z(G) = T .
We claim that λ(G) = λ(T ) + 1, so λ(G) = 4 if and only if T is one of the simple groups
in Table 1. To see this, first note that λ(T ) 6 λ(G) 6 λ(T ) + 1 by Lemma 2.1(ii). Also
observe that if M is a maximal subgroup of a group of the form N.K, where N = Cp is
normal, then either M = N.L with L < K maximal, or M ∼
= K. Seeking a contradiction,
suppose λ(G) = λ(T ) = t and consider an unrefinable chain of G as in (1). Set Z = Z(G).
Since λ(G) = λ(T ), it follows that G1 ∼
6 T , so G1 = Z.T1 where T1 < T is maximal.
=
Similarly, G2 = Z.T2 with T2 < T1 maximal, and so on. In particular, Gt−1 = Z.Tt−1 with
Tt−1 < Tt−2 maximal. But Tt−1 6= 1 and thus |Gt−1 | is composite. This is a contradiction
and the claim follows.
3.3. Proof of Theorem 3. Let G be an almost simple group with λ(G) = 4 and socle
T . First assume G 6= T , so Theorem 2 implies that G/T = Cp for a prime p. If λ(T ) = 3,
then we are in case (i) of Theorem 3. Now assume λ(T ) > 4. We claim that (ii) holds, so
λ(T ) = 4 and (G, T ) is one of the cases in Table 3.
To see this, let M be a maximal subgroup of G of depth 3. Then M 6= T , so G = T M
and M ∩ T ⊳ M has index p. By Theorem 1, M is soluble and Lemma 2.3(i) implies that
λ(M ∩ T ) = 2, so M ∩ T is not maximal in T . Therefore, G has a novelty soluble maximal
subgroup M of chief length 3. This property is highly restrictive and we can determine
all the possibilities for G and M .
First assume T = An is an alternating group, so G = An .2. If n = 6 then λ(T ) = 4 and
one checks that λ(PGL2 (9)) = λ(M10 ) = 4, while λ(S6 ) = 5. Now assume n 6= 6, so that
G = Sn . By part (I) of the main theorem of [23], the novelty soluble maximal subgroups
of Sn are S2 ≀ S4 < S8 and Cp :Cp−1 < Sp for p ∈ {7, 11, 17, 23}. Of these, only Cp :Cp−1
for p ∈ {7, 11, 23} have depth 3, so S7 , S11 and S23 are the only depth 4 groups arising in
this case.
If T is a sporadic group then G = T.2 and one checks (by inspection of the Atlas [12])
that G does not have a maximal subgroup M with the required properties.
Now assume T is a simple group of Lie type over Fq . If T is an exceptional group of Lie
type, then all the maximal soluble subgroups of G are known (see [11, 25]) and one checks
ON THE LENGTH AND DEPTH OF FINITE GROUPS
11
that there are no relevant examples (it is helpful to note that if T = 2 B2 (q), 2 G2 (q), 2 F4 (q)
or 3 D4 (q), then G does not have any novelty maximal subgroups). Finally, suppose T is a
classical group. For the low-rank groups, it is convenient to consult the relevant tables in
[6]; in this way, one checks that the only cases that arise are the ones listed in Table 3 (in
each case, λ(T ) = 4). For example, if G = PGL2 (q), where q is a prime and q ≡ ±11, ±19
(mod 40), then G has a maximal subgroup M = S4 and M ∩ T = A4 is non-maximal in T
(note that the additional condition Ω(q ± 1) > 3 is needed to ensure that λ(T ) > 4, which
means that λ(T ) = 4 by [8, Lemma 3.1]). By inspecting [21], it is easy to check that no
examples arise when G is one of the remaining classical groups not covered by [6]. We
conclude that part (ii) of Theorem 3 holds.
To complete the proof, we may assume G = T has depth 4. Let M be a maximal
subgroup of G with λ(M ) = 3. By Theorem 1, either M is simple (and we are in part
(iv) of Theorem 3), or M is soluble of chief length 3. It remains to show that in the latter
case, the possibilities for G and M are given in Table 4. To do this, we essentially repeat
the above argument, but now there are more cases to consider because G = T and there
is no novelty condition.
First assume G = An . It is easy to verify the result for n 6 16 (with the aid of Magma,
for example), so let us assume n > 17. By the O’Nan-Scott theorem (see [24]), the only
soluble maximal subgroups of G are of the form M = AGL1 (p) ∩ G = Cp :C(p−1)/2 , with
n = p a prime. Here Lemma 2.3(i) implies that λ(M ) = Ω(p − 1), which explains the
condition Ω(p − 1) = 3 in Table 4.
Next assume G is a sporadic group. The groups with λ(G) = 4 can be read off from
[8, Lemma 3.3] and the cases appearing in Table 4 are obtained by inspecting the lists of
maximal subgroups of G in the Atlas [12].
Finally suppose G is a simple group of Lie type over Fq . As noted above, if G is an
exceptional group then all of the soluble maximal subgroups of G are known and it is
routine to read off the cases with such a subgroup of depth 3 (for G = 2 B2 (q), note that
we need the extra condition Ω(q −1) > 2 to ensure that λ(G) = 4). Similarly, the result for
classical groups is obtained by carefully inspecting [6] (for the low-rank groups) and [21]
(in the remaining cases). Once again, extra conditions on q are needed to get λ(G) = 4.
This completes the proof of Theorem 3.
Remark 3.3. It is not feasible to give a complete description of the depth 4 examples
(G, M ) arising in part (iv) of Theorem 3, but we can give some partial information:
Sporadic groups. By inspecting the Atlas [12], it is easy to show that if G 6= M is a
sporadic group, then (G, M ) is one of the following:
(M11 , L2 (11)) (M12 , L2 (11)) (M22 , L2 (11)) (M24 , L2 (7)) (M24 , L2 (23))
(M24 , M23 )
(J1 , L2 (11))
(J2 , A5 )
(Suz, L2 (25)) (Co2 , M23 )
(Co3 , M23 )
(Fi23 , L2 (23)) (Th, L3 (3))
At present, the maximal subgroups of the Monster group have not been completely determined; among the 44 conjugacy classes of known maximal subgroups, L2 (59) is the only
simple maximal group of depth 3. According to Wilson’s recent survey article [34], any
additional maximal subgroup of the Monster is almost simple with socle L2 (8), L2 (13),
L2 (16), U3 (4) and U3 (8). Note that each of these simple groups has depth 3. The latter
possibility is eliminated in [35], and Wilson reports that he has also ruled out L2 (8) and
U3 (4) in unpublished work (see the final paragraph in Section 3 of [34]). This leaves open
the two remaining possibilities.
12
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
Alternating groups. Let G = An be an alternating group. With the aid of Magma, it is
easy to check that for n 6 100, the possibilities for (M, n) are as follows:
(A5 , 6)
(L5 (2), 31)
(L2 (59), 60)
(L2 (13), 78)
(L2 (7), 7)
(L2 (37), 38)
(A59 , 60)
(L2 (83), 84)
(L3 (3), 13)
(L2 (43), 44)
(L2 (61), 62)
(A83 , 84)
(L2 (13), 14) (M23 , 23)
(L3 (5), 31)
(L2 (47), 48) (A47 , 48)
(L2 (53), 54)
(L2 (25), 65) (L2 (67), 68) (L2 (73), 74)
(A87 , 88)
The main theorem of [24] on the maximal subgroups of symmetric and alternating groups
provides some useful information in the general case, but it is not possible to state a precise
result.
Exceptional groups. Let G be an exceptional group of Lie type over Fq . If G = 2 B2 (q),
2 G (q), 2 F (q)′ , 3 D (q) or G (q), then the maximal subgroups of G are known and we can
2
4
4
2
read off the relevant examples: either G = 2 F4 (2)′ and M = L2 (25), or
• G = 2 B2 (q) and M = 2 B2 (q0 ) with q = q0k , q0 > 2 and both k and q0 − 1 are
primes; or
• G = G2 (q) and M = L2 (13) or L2 (8), with various conditions on q needed for the
maximality of M (see [6], for example).
In the remaining cases, by combining results of Liebeck and Seitz [27] with recent work
of Craven [13], we deduce that there are no examples with M an alternating or sporadic
group. Strong restrictions on the remaining possibilities when M is a group of Lie type can
be obtained by applying [26, Theorem 8] (defining characteristic) and the main theorem
of [27] (non-defining characteristic).
Classical groups. Finally, suppose G is a simple classical group with natural n-dimensional
module V . By Aschbacher’s subgroup structure theorem [2], either M belongs to a collection C(G) of geometric subgroups, or M ∈ S(G) is almost simple and acts irreducibly on
V . By inspecting [6, 21], it is possible to determine the relevant examples with M ∈ C(G)
(the precise list of cases will depend on some delicate number-theoretic conditions). For
example, suppose G = Un (q) is a unitary group, where n = q + 1 and q > 5 is a prime. If
(q n−1 + 1)/(q + 1) is also a prime, then G has a simple maximal subgroup M = Un−1 (q)
of depth 3 (here M is the stabiliser of a non-degenerate 1-space). For instance, G = U6 (5)
has a maximal subgroup M = U5 (5) of depth 3. It is not feasible to determine all the
cases that arise with M ∈ S(G), although this can be achieved for the low-dimensional
classical groups (that is, the groups with n 6 12) by inspecting the relevant tables in [6,
Chapter 8].
4. Length
In this section we prove our main results on length, namely Theorems 4 and 6, and
Corollary 5.
4.1. Proof of Theorem 4 and Corollary 5. Let G be a non-abelian finite simple group.
First recall that λ(G) > 3 and cd(G) > 1, so l(G) > 4. The simple groups of length 4 were
classified by Janko [19, Theorem 1]. In a second paper [20], he proves that every simple
group of length 5 is of the form L2 (q) for some prime power q (but he does not give any
further information on the prime powers that arise). In later work of Harada [16], this
result was extended to simple groups of length at most 7.
Theorem 4.1 (Harada, [16]). Let G be a finite simple group with l(G) 6 7. Then either
(i) G = U3 (3), U3 (5), A7 , M11 , J1 ; or
(ii) G = L2 (q) for some prime power q.
ON THE LENGTH AND DEPTH OF FINITE GROUPS
13
For the groups in part (i) of Theorem 4.1, it is easy to check that A7 and J1 have length
6, the others have length 7. We are now ready to prove Theorem 4.
Lemma 4.2. Theorem 4 holds if G ∼
= L2 (q).
Proof. Write q = pf , where p is a prime. In view of Lemma 2.5, the result is clear if p = 2
or f = 1, so let us assume p > 3 and f > 2, in which case
l(G) = max{Ω(q − 1) + f, Ω(q + 1) + 1}.
Since l(G) 6 9, it follows that f 6 7 and it is easy to verify the result when p ∈ {3, 5}.
Now assume p > 7, in which case Ω(p2 − 1) > 5 and f ∈ {2, 3, 5}.
Suppose f = 5, so Ω(q − 1) > 3 and we have l(G) ∈ {8, 9}. For l(G) = 8 we must have
Ω(q − 1) = 3 and Ω(q + 1) 6 7; one checks that there are primes p with these properties:
p ∈ {3, 7, 23, 83, 263, 1187, . . .}.
Similarly, for l(G) = 9 we need Ω(q − 1) = 4 and Ω(q + 1) 6 8, or Ω(q − 1) = 3 and
Ω(q + 1) = 8; in both cases, there are primes satisfying these conditions.
Next consider the case f = 3, so Ω(q − 1) > 3 and l(G) ∈ {6, 7, 8, 9}. First assume
l(G) = 6, so Ω(q − 1) = 3 and thus (p − 1)/2 and p2 + p + 1 are both primes. In particular,
p ≡ −1 (mod 12) and p > 11. Therefore, Ω(p + 1) > 4 and p2 − p + 1 is divisible by
3, hence Ω(q + 1) > 6 and we have reached a contradiction. Next assume l(G) = 7, so
Ω(q − 1) = 3 or 4. If Ω(q − 1) = 3 then we need Ω(q + 1) = 6, which forces Ω(p + 1) = 4
and p2 − p + 1 = 3r for some prime r. But 7 divides p6 − 1, so 7 must divide p + 1 and
thus p = 83 is the only possibility. But then p2 + p + 1 is composite, so this case does not
arise. However, there are primes
p ∈ {7, 11, 83, 1523, 20507, 28163, . . .}
satisfying the conditions Ω(q − 1) = 4 and Ω(q + 1) 6 6, so this case is recorded in Table
5. Similarly, if l(G) = m ∈ {8, 9} then either Ω(q − 1) = m − 3 and Ω(q + 1) 6 m − 1, or
Ω(q − 1) 6 m − 4 and Ω(q + 1) = m − 1. Moreover, one can check that there are primes p
satisfying these conditions.
Finally, let us assume f = 2. Here the condition p > 7 implies that Ω(q − 1) > 5, so
l(G) ∈ {7, 8, 9}. Suppose l(G) = 7. Here Ω(q − 1) = 5 and either (p − 1)/2 or (p + 1)/2
is a prime. Suppose (p − 1)/2 is a prime, so p ≡ 3 (mod 4) and (p + 1)/2 = 2r for some
prime r. Therefore, r, 2r − 1 and 4r − 1 are all primes. If r ∈ {2, 3} then p ∈ {7, 11} and
one checks that l(G) = 7. Now assume r > 5. If r ≡ 1 (mod 3) then 4r − 1 is divisible by
3. Similarly, if r ≡ 2 (mod 3) then 3 divides 2r − 1, so there are no examples with r > 5.
A similar argument applies if we assume (p + 1)/2 is a prime: here we need a prime r such
that 2r + 1 and 4r + 1 are also primes, and one checks that r = 3 is the only possibility,
which corresponds to the case G = L2 (169) with l(G) = 7.
Next assume l(G) = 8 and f = 2, so Ω(q − 1) = 5 or 6. The case Ω(q − 1) = 5 is
ruled out by arguing as in the previous paragraph. On the other hand, if Ω(q − 1) = 6
then we need Ω(q + 1) 6 7 and there are primes p with these properties. Finally, let us
assume l(G) = 9, so Ω(q − 1) ∈ {5, 6, 7}. The case Ω(q − 1) = 5 is ruled out as above,
whereas there are primes p such that Ω(q − 1) = 7 and Ω(q + 1) 6 8, or Ω(q − 1) = 6 and
Ω(q + 1) = 8.
In view of Theorem 4.1, it remains to determine the simple groups G ∼
6 L2 (q) with
=
l(G) = 8 or 9.
Lemma 4.3. Theorem 4 holds if l(G) = 8.
14
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
Proof. Let G be a simple group with l(G) = 8. By inspection, if G is a sporadic group then
G = M12 is the only example (also see [10, Tables III and IV]). For alternating groups,
[10, Theorem 1] gives
3n − 1
− bn − 1,
(5)
l(An ) =
2
where bn is the number of ones in the base 2 expansion of n. From the formula, it is easy
to check that no alternating group has length 8.
Now assume G is a simple group of Lie type over Fq , where q = pf with p a prime.
First we handle the exceptional groups. If G = 2 B2 (q) then p = 2, f > 3 is odd and
l(G) = Ω(q − 1) + 2f + 1
by [32, Theorem 1], hence l(G) = 8 if and only if f = 3. Next assume G = 2 G2 (q), so
p = 3 and f > 3 is odd. Since a Borel subgroup of G has order q 3 (q − 1), it follows that
l(G) > Ω(q − 1) + 3f + 1 > 8.
If G = G2 (q) then q > 3 and l(G) > l(GL2 (q)) + 5f + 1 > 8, so no examples arise. All of
the other exceptional groups can be eliminated in a similar fashion.
Finally, let us assume G is a classical group. If G = PΩǫn (q) is an orthogonal group with
n > 7, then it is clear that l(G) > 8 (indeed, a Sylow p-subgroup of G has length greater
than 8). Similarly, we can eliminate symplectic groups PSpn (q) with n > 6. Now assume
G = PSp4 (q) with q = pf > 3. Here l(G) > l(SL2 (q))+ 3f + 1, so we may assume f = 1, in
which case G has a maximal subgroup 24 .A5 or 24 .S5 (according to the value of q modulo
8) and thus l(G) > 1 + 4 + l(A5 ) = 9. Similarly, it is easy to show that l(Lǫn (q)) > 8 if
n > 4.
To complete the proof, we may assume that G = U3 (q) or L3 (q). Let B be a Borel
subgroup of G and first assume G = U3 (q), so q > 3. If q is even, then [32, Theorem 1]
gives
l(G) = Ω(|B|) + 1 = Ω(q 2 − 1) + 3f + 1 − Ω((3, q + 1))
(6)
and thus l(G) > 9. For q odd we have
l(G) > Ω(q 2 − 1) + 3f + 1 − Ω((3, q + 1))
(7)
l(G) = Ω(|B|) + 2 = 2Ω(q − 1) + 3f + 2 − Ω((3, q − 1))
(8)
l(G) > l(L2 (q)) + 2f + 2 + Ω(q − 1) − Ω((3, q − 1))
(9)
and we quickly deduce that f = 1, so q > 7 (since U3 (3) and U3 (5) have length 7). Now
Ω(q 2 − 1) > 5, so we must have Ω(q 2 − 1) = 5 and q ≡ 2 (mod 3), hence (q − 1)/2 and
(q + 1)/6 are both prime. This implies that q = 6r − 1, where r and 3r − 1 are primes, so
r = 2 is the only option and one checks that l(U3 (11)) = 9.
Finally, let us assume G = L3 (q). If q is even then
and it is easy to see that l(G) 6= 8. Now assume q is odd. If q = 3 then one can check that
l(G) = 8, so let us assume q > 5. Let H be a maximal parabolic subgroup of G. Then
l(G) > l(H) + 1, so
and we deduce that l(G) > 9.
Lemma 4.4. Theorem 4 holds if l(G) = 9.
Proof. This is very similar to the proof of the previous lemma. Let G be a simple group
with l(G) = 9. By inspection, G is not a sporadic group. In view of (5), G = A8 is the
only alternating group of length 9. Now assume G is a group of Lie type over Fq , where
q = pf with p a prime. The exceptional groups are easily eliminated by arguing as in the
proof of Lemma 4.3. Similarly, if G is a classical group then it is straightforward to reduce
to the cases G = PSp4 (q)′ and Lǫ3 (q) (note that L4 (2) ∼
= A8 and U4 (2) ∼
= PSp4 (3)).
ON THE LENGTH AND DEPTH OF FINITE GROUPS
15
Suppose G = PSp4 (q)′ . If q = 2 then G ∼
= A6 and l(G) = 5. Now assume q > 3. As
noted in the proof of the previous lemma, l(G) > l(SL2 (q)) + 3f + 1 and so we may assume
q = p is odd. Now G has a maximal subgroup H of type Sp2 (q) ≀ S2 , which implies that
l(G) > l(H) + 1 = 3 + 2l(L2 (p)).
If p = 3 then this lower bound is equal to 9 and one checks that l(PSp4 (3)) = 9. For p > 3
we get l(G) > 11.
Next assume G = L3 (q). If q is even, then (8) holds and one checks that l(G) = 9 if and
only if q = 4. Now assume q is odd. We have already noted that l(L3 (3)) = 8, so we may
assume q > 5. Moreover, in view of (9), we may assume that q = p. Since l(L2 (p)) > 4
and Ω(p − 1) > 2, it follows that l(L2 (p)) = 4, Ω(p − 1) = 2 and p ≡ 1 (mod 3). Clearly,
p = 7 is the only prime satisfying the latter two conditions, but l(L2 (7)) = 5.
To complete the proof of the lemma, we may assume G = U3 (q). If q is even then (6)
holds and we deduce that l(G) = 9 if and only if q = 4. Now suppose q is odd, so (7)
holds. If f > 2 then Ω(q 2 − 1) > 5 and thus l(G) > 11. Therefore, we may assume q = p
is odd. We have already noted that l(U3 (3)) = l(U3 (5)) = 7 and l(U3 (11)) = 9, and it is
straightforward to check that l(U3 (7)) = 10. Now assume q > 13. If Ω(q 2 − 1) > 7 then
(7) implies that l(G) > 10, so we must have Ω(q 2 − 1) = 5 or 6.
If Ω(q 2 − 1) = 5 then q = 13 is the only possibility (see the proof of Lemma 4.2) and
one checks that l(U3 (13)) = 9. Now assume Ω(q 2 − 1) = 6, so q ≡ 2 (mod 3) by (7). By
considering the maximal subgroups of G (see [6, Tables 8.5 and 8.6]), we see that
l(G) = max{9, Ω(q + 1) + l(L2 (q)) + 1, Ω(q 2 − q + 1) + 1}.
Note that Ω(q + 1) > 3 and l(L2 (q)) > 4 since q > 13 and q ≡ 2 (mod 3). If Ω(q + 1) > 4
then l(L2 (q)) > 5 and thus l(G) > 10. Therefore, Ω(q +1) = Ω(q −1) = 3 and l(L2 (q)) 6 5,
so either q = 29 or q ≡ ±3, ±13 (mod 40). In addition, we need Ω(q 2 − q + 1) 6 8 and one
checks there are primes p that satisfy these conditions:
p ∈ {173, 317, 653, 2693, 3413, 3677, . . .}.
This completes the proof of Theorem 4.
Proof of Corollary 5. First observe that part (i) is clear from the additivity of length (see
Lemma 2.1(i)) and the fact that every non-abelian simple group has length at least 4.
Next, assume G is a finite insoluble group with l(G) = 5. The simple groups of length 5
are given in Theorem 4, so we may assume that G is not simple. Therefore, G must have
exactly two composition factors: a non-abelian simple group T of length 4 and depth 3,
and a cyclic group Cp of prime order. In particular, λ(G) = 4 and so G is one of the groups
in Theorem 2. By inspecting the various possibilities, we see that either G = T × Cp , or
G is quasisimple with G/Z(G) = T and Z(G) = Cp , or G is almost simple with socle
T and G/T = Cp . Since l(T ) = 4, [19, Theorem 1] implies that T = L2 (q) and q is a
prime satisfying the conditions in the first row of Table 5. In particular, the only valid
quasisimple and almost simple groups are of the form SL2 (q) and PGL2 (q), respectively.
This completes the proof of part (ii) of Corollary 5.
Finally, suppose G is insoluble of length 6. Again, the simple groups of length 6 are
given by Theorem 4, so we may assume G is not simple. Then G has a unique non-abelian
composition factor T of length 4 or 5, and 6 − l(T ) abelian composition factors. It is
readily checked that the possibilities for G when l(T ) = 5 (resp. 4) are those in (iii)(b)
(resp. (c)) of Corollary 5.
16
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
4.2. Proof of Theorem 6. Set G = L2 (p), where p > 5 is a prime. By Lemma 2.5(ii),
l(G) 6 1 + max{4, Ω(p ± 1)}.
The result follows from Theorem 7 (see Appendix A), which implies that there are infinitely
many primes p with max{Ω(p ± 1)} 6 8.
5. Chain differences and ratios
In this section we prove our main results concerning chan differences and chain ratios,
namely Theorems 8, 9 and 11.
5.1. Proof of Theorem 8. For comparison, we start by recalling [7, Theorem 3.3], which
describes the simple groups of chain difference one.
Theorem 5.1 (Brewster et al. [7]). Let G be a finite simple group. Then cd(G) = 1 if
and only if G = L2 (q) and either q ∈ {4, 5, 9}, or q is a prime and one of the following
holds:
(i) 3 6 Ω(q ± 1) 6 4 and either q ≡ ±1 (mod 10) or q ≡ ±1 (mod 8).
(ii) Ω(q ± 1) 6 3 and q ≡ ±3, ±13 (mod 40).
We begin the proof of Theorem 8 by handling the alternating and sporadic groups.
Lemma 5.2. Let G be a simple alternating or sporadic group. Then cd(G) = 2 if and
only if G = A7 or J1 .
Proof. First assume G = An is an alternating group. A formula for l(G) is given in (5) and
it is easy to compute λ(An ) directly for small values of n: we get cd(A5 ) = cd(A6 ) = 1,
cd(A7 ) = 2 and cd(A8 ) = 4. By Lemma 2.2, it follows that cd(An ) > 4 for all n > 8.
The length and depth of each sporadic group G is given in [10, Tables III and IV] and [8,
Table 2], respectively, and we immediately deduce that cd(G) = 2 if and only if G = J1 .
(Note that the “possible values” for l(G) listed in [10, Table IV] are in fact the exact
values.)
Lemma 5.3. Let G be a simple group of Lie type over Fq with G ∼
6 L2 (q). Then cd(G) = 2
=
if and only if G = U3 (5).
Proof. We will follow a similar approach to the proof of [7, Theorem 3.3]. Set q = pf ,
where p is a prime and f > 1.
First assume G = L3 (q), in which case q > 3 since L3 (2) ∼
= L2 (7). If q = 3 then
l(G) = 8 and λ(G) = 3, so we may assume q > 4 and thus l(G) > 9 by Theorem 4.
If p is odd then L3 (p) has a maximal subgroup SO3 (p) ∼
= PGL2 (p), so λ(L3 (p)) 6 6
by [8, Corollary 3.4] and thus cd(L3 (p)) > 3. In view of Lemma 2.2, this implies that
cd(G) > 3 since L3 (p) 6 G. Now assume p = 2 and let H = QL be a maximal parabolic
subgroup of G, where Q is elementary abelian of order q 2 and L 6 GL2 (q) has index
d = (3, q − 1). Note that L acts irreducibly on Q, so L is a maximal subgroup of H.
Therefore, l(H) = Ω((q − 1)/d) + 2f + l(L2 (q)) and
λ(H) 6 λ(L) + 1 = Ω((q − 1)/d) + λ(L2 (q)) + 1,
so cd(H) > 2f − 1 + cd(L2 (q)) > 2f and the result follows since f > 2.
Now assume G = Ln (q) with n > 4. Since G has a subgroup L3 (q), the desired result
follows from Lemma 2.2 if q > 3. Similarly, if q = 2 then G has a subgroup L4 (2) and it
is easy to check that cd(L4 (2)) = 4.
Next suppose G is a classical group of rank r > 3, or an untwisted exceptional group,
or one of 3 D4 (q) or 2 E6 (q). It is easy to see that G has a section isomorphic to L3 (q) and
ON THE LENGTH AND DEPTH OF FINITE GROUPS
17
thus cd(G) > 3 if q > 3. Now assume q = 2. If G = G2 (2)′ ∼
= U3 (3) then cd(G) = 3. Since
−
U3 (3) < Sp6 (2) < Ω−
(2),
it
follows
that
cd(G)
>
3
if
G
=
Sp
6 (2) or Ω8 (2). In each of the
8
remaining cases (with q = 2), G has a section isomorphic to L4 (2) and thus cd(G) > 4.
Next assume G = U3 (q), so q > 3. Let H = QL be a Borel subgroup of G, where
d = (3, q + 1). Here Q = q 1+2 , L = (q 2 − 1)/d and Q/Z(Q) is elementary abelian of
order q 2 . Moreover, L acts irreducibly on Q/Z(Q). Therefore, l(H) = 3f + Ω(L) and
λ(H) 6 f + 1 + Ω(L), so cd(H) > 2f − 1 and we may assume q = p. If p ∈ {3, 7, 11}
then it is easy to check that cd(G) > 3, whereas cd(G) = 2 if p = 5. For p > 11,
Theorem 4 implies that l(G) > 9 and we get cd(G) > 3 since G has a maximal subgroup
SO3 (p) ∼
= PGL2 (p) of depth at most 5. Now assume G = Un (q) with n > 4. If q is even
then G has a section isomorphic to U4 (2) and one checks that cd(U4 (2)) = 4. Similarly,
if q is odd and q 6= 5 then G has a section U3 (q) with cd(U3 (q)) > 3. Finally, suppose
q = 5. One checks that λ(U4 (5)) = 5 and thus cd(U4 (5)) > 5 by Theorem 4. The result
now follows since G has a section isomorphic to U4 (5).
To complete the proof, we may assume G = PSp4 (q), 2 F4 (q)′ , 2 G2 (q) or 2 B2 (q). Suppose
G = PSp4 (q) with q > 3. If q is even then G has a maximal subgroup H = L2 (q) ≀ S2 .
Now l(H) = 2l(L2 (q)) + 1 and λ(H) 6 λ(L2 (q)) + 2, so cd(H) > l(L2 (q)) > 4. Similarly,
if q is odd then H = 24 .Ω−
4 (2) < G and the result follows since cd(H) = 4. Next assume
2
′
G = F4 (q) . One checks that the Tits group 2 F4 (2)′ has depth 4, so cd(2 F4 (2)′ ) > 6 by
Theorem 4 and thus cd(G) > 6 by Lemma 2.2.
Next assume G = 2 G2 (q), so q = 3f and f > 3 is odd. Let H be a Borel subgroup of G
and let K = 2 × L2 (q) be the centralizer of an involution. Then
l(G) > l(H) + 1 = Ω(q − 1) + 3f + 1
λ(G) 6 λ(K) + 1 6 λ(L2 (q)) + 2 6 Ω(q − 1) + 3
and thus cd(G) > 3f − 2 > 7. Finally, suppose G = 2 B2 (q), where q = 2f and f > 3 is
odd. Here l(G) = Ω(q − 1) + 2f + 1 by [32, Theorem 1] and λ(G) 6 Ω(q − 1) + 2 (since G
has a maximal subgroup D2(q−1) ). Therefore, cd(G) > 2f − 1 > 5.
The next result completes the proof of Theorem 8.
Lemma 5.4. If G = L2 (q) with q > 5, then cd(G) = 2 if and only if
(i) q ∈ {7, 8, 11, 27, 125}; or
(ii) q is a prime and one of the following holds:
(a) max{Ω(q ± 1)} = 4 and either min{Ω(q ± 1)} = 2, or q ≡ ±3, ±13 (mod 40).
(b) max{Ω(q ± 1)} = 5, min{Ω(q ± 1)} > 3 and q 6≡ ±3, ±13 (mod 40).
Proof. As before, write q = pf . First assume p = 2, so q > 4. Here l(G) = Ω(q − 1) + f + 1
by Lemma 2.5(i) and λ(G) 6 Ω(q−1)+2 (since D2(q−1) is a maximal subgroup). Therefore,
cd(G) > f − 1 and thus f ∈ {2, 3}. If f = 2 then cd(G) = 1, while cd(G) = 2 if f = 3
(the case q = 8 is recorded in part (ii)(a) of Theorem 8).
Now assume p > 3. For q 6 11, one checks that cd(G) = 2 if and only if q = 7 or 11,
so we may assume q > 13. This implies that Dq−1 is a maximal subgroup of G and thus
λ(G) 6 Ω(q − 1) + 1. Now l(G) > Ω(q − 1) + f by Lemma 2.5(ii), so cd(G) > f − 1 and
thus we may assume f ∈ {1, 2, 3}.
First assume q = p > 13. By [8, Corollary 3.4] we have
3 min{Ω(p ± 1)} = 2 or p ≡ ±3, ±13 (mod 40)
λ(G) =
4 otherwise.
18
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
If λ(G) = 3 then we need l(G) = 5, in which case Theorem 4 implies that max{Ω(p±1)} =
4. There are primes p that satisfy these conditions. For example, if
p ∈ {23, 59, 83, 227, 347, 563, . . .},
then Ω(p − 1) = 2 and Ω(p + 1) = 4. Similarly, we have λ(G) = 4 and l(G) = 6 if and
only if Ω(p ± 1) > 3, p 6≡ ±3, ±13 (mod 40) and max{Ω(p ± 1)} = 5. Once again, there
are primes p with these properties.
Next assume q = p2 with p > 5. Since PGL2 (p) is a maximal subgroup of G, it
follows that λ(G) 6 6 and thus l(G) 6 8. Now, if l(G) 6 7 then Theorem 4 implies
that p ∈ {5, 7, 11, 13} and in each case one checks that cd(G) > 3. Therefore, we may
assume λ(G) = 6 and l(G) = 8. By Theorem 4, l(G) = 8 if and only if Ω(q − 1) = 6
and Ω(q + 1) 6 7. Similarly, λ(G) = 6 if and only if Ω(q ± 1) > 5, p ≡ ±1 (mod 10) and
λ(L2 (p)) = 4. The latter constraint yields the additional condition Ω(p ± 1) > 3, so we
need Ω(p − 1) = Ω(p + 1) = 3 since Ω(q − 1) = 6. We claim that there are no primes p that
satisfy these conditions. For example, suppose p ≡ 1 (mod 10). Then (p − 1)/10 must be
a prime. If p ≡ 1 (mod 3) then p = 31 is the only possibility, but this gives Ω(p + 1) = 5.
On the other hand, if p ≡ 2 (mod 3) then (p + 1)/6 is a prime. But p2 ≡ 1 (mod 8) and
thus (p − 1)/10 = 2 or (p + 1)/6 = 2, which implies that p = 11 and Ω(p − 1) = 2. A very
similar argument handles the case p ≡ −1 (mod 10).
Finally, suppose q = p3 . If p = 3 then one checks that λ(G) = 3 and l(G) = 5, so
cd(G) = 2 in this case. Now assume p > 5. Since L2 (p) is a maximal subgroup of G, it
follows that
4 min{Ω(p ± 1)} = 2 or p ≡ ±3, ±13 (mod 40)
λ(G) =
5 otherwise
and thus l(G) 6 7. By applying Theorem 4, we see that λ(G) = 4 and l(G) = 6 if and
only if p = 5. Similarly, λ(G) = 5 and l(G) = 7 if and only if Ω(q − 1) = 4, Ω(q + 1) 6 6,
Ω(p ± 1) > 3 and p 6≡ ±3, ±13 (mod 40). Note that the conditions Ω(q − 1) = 4 and
Ω(p ± 1) > 3 imply that Ω(p − 1) = 3 and p2 + p + 1 is a prime, so p ≡ 2 (mod 3) and
p2 − p + 1 is divisible by 3, whence Ω(p2 − p + 1) > 2. There are primes p such that
Ω(p3 − 1) = 4, Ω(p3 + 1) 6 6 and Ω(p ± 1) > 3: the smallest one is 433373. However, we
claim that there is no prime p that also satisfies the condition p 6≡ ±3, ±13 (mod 40).
If p ≡ 1, 9, 17, 33 (mod 40) then p − 1 is divisible by 8 and thus Ω(p − 1) > 4. Similarly,
if p ≡ 7, 23, 31, 39 (mod 40) then p + 1 is divisible by 24, and p 6= 23 since we need
Ω(p − 1) = 3, so Ω(p + 1) > 5. But we have already noted that Ω(p2 − p + 1) > 2,
whence Ω(p3 + 1) > 7. Finally, suppose p ≡ 11, 19, 29 (mod 40). These cases are similar,
so let us assume p ≡ 11 (mod 40). Here (p − 1)/10 is a prime and p + 1 is divisible by
12, so Ω(p + 1) > 4 since p 6= 11. Since Ω(p3 + 1) 6 6, it follows that (p + 1)/12 and
(p2 − p + 1)/3 are both primes. Now p6 ≡ 1 (mod 7), so one of (p − 1)/10, p2 + p + 1,
(p + 1)/12 or (p2 − p + 1)/3 must be equal to 7, but it is easy to see that this is not
possible. For example, if (p − 1)/10 = 7 then p = 71 does not satisfy the required
congruence condition.
5.2. Proof of Theorem 9. Recall that cr(G) = l(G)/λ(G) is the chain ratio of G. In this
section we prove Theorem 9, which states that cr(G) > 5/4 for every finite non-abelian
simple group G, with equality if and only if G ∈ G (see (3)). We partition the proof into
several cases.
Lemma 5.5. Theorem 9 holds if G is a sporadic or alternating group.
Proof. First assume G is a sporadic group. The length and depth of G is given in [10,
Tables III and IV] and [8, Table 2], respectively, and we immediately deduce that cr(G) >
3/2, with equality if and only if G = J1 .
ON THE LENGTH AND DEPTH OF FINITE GROUPS
19
Now assume G = An is an alternating group. The length of G is given in (5) and [8,
Theorem 2] states that λ(G) 6 23. One checks that this bound is sufficient if n > 23. For
example, if n = 23 then l(G) = 34 − 4 − 1 = 29 and thus cr(G) > 29/23 > 5/4. For n < 23
we can use Magma to show that λ(G) 6 6, with equality if and only if n = 16. In view of
the above formula for l(G), we deduce that cr(G) > 5/4 if n > 8. For the smallest values
of n, we get cr(A5 ) = 4/3, cr(A6 ) = 5/4 and cr(A7 ) = 3/2.
Lemma 5.6. Theorem 9 holds if G is an exceptional group of Lie type.
Proof. Let G be a finite simple exceptional group of Lie type over Fq , where q = pk and p
is a prime. Let B be a Borel subgroup of G and let r be the twisted Lie rank of G. Then
l(G) > Ω(|B|) + r
(10)
and [8, Theorem 4] gives
λ(G) 6 3Ω(k) + 36
2B
if G 6= 2 (q).
First assume G = E8 (q). Here |B| = q 120 (q − 1)8 so
5
5
5
l(G) > 120k + 8 > (3 log2 k + 36) > (3Ω(k) + 36) > λ(G)
4
4
4
and the result follows. The case E7 (q) is handled in exactly the same way, and similarly
E6ǫ (q) and F4 (q) with k > 2. Suppose G = F4 (p), so l(G) > 28 by (10). If p = 2 then
2 F (2) < G is maximal and λ(2 F (2)) = 5, so λ(G) 6 6 and the result follows. Similarly,
4
4
if p is odd then
λ(F4 (p)) 6 λ(2.Ω9 (p)) + 1 6 λ(Ω9 (p)) + 2
and one of A10 , S10 or A11 is a maximal subgroup of Ω9 (p) (see [6, Table 8.59]). Therefore
λ(Ω9 (p)) 6 7, so λ(G) 6 9 and once again we deduce that cr(G) > 5/4. If G = E6ǫ (p) then
F4 (p) < G is maximal, so the previous argument yields λ(G) 6 10 and the result quickly
follows.
Next assume G = G2 (q)′ . If q = 2 then G ∼
= U3 (3) and one checks that λ(G) = 4 and
6
l(G) = 7. Now assume q > 2. Since |B| = q (q − 1)2 it follows that l(G) > 6k + 4 and
by considering a chain of subfield subgroups we deduce that λ(G) 6 Ω(k) + λ(G2 (p)). If
p > 5 then G2 (2) < G2 (p) is maximal, so λ(G2 (p)) 6 6 and it is easy to check that the
same bound holds if p = 2 or 3. Therefore λ(G) 6 Ω(k) + 6 and we deduce that
5
5
5
l(G) > 6k + 4 > (log2 k + 6) > (Ω(k) + 6) > λ(G).
4
4
4
If G = 3 D4 (q) then G2 (q) < G is maximal and thus λ(G) 6 Ω(k) + 7. In addition,
|B| = q 12 (q 3 − 1)(q − 1), so l(G) > 12k + 2 and the result follows.
To complete the proof of the lemma, we may assume G = 2 F4 (q)′ , 2 G2 (q) or 2 B2 (q).
Suppose G = 2 F4 (q)′ , so q = 2k with k odd. If k = 1 then λ(G) = 4 and l(G) > 13
since G has a soluble maximal subgroup of the form 2.[28 ].5.4. Similarly, if k > 1 then
λ(G) 6 Ω(k) + 5 (see the proof of [8, Theorem 4]), l(G) > 12k + 2 and these bounds are
sufficient. The case G = 2 G2 (q)′ , where q = 3k with k odd, is very similar. If k = 1 then
G∼
= L2 (8), so λ(G) = 3 and l(G) = 5. If k > 1, then the proof of [8, Theorem 4] gives
λ(G) 6 Ω(k) + 4 and we have l(G) > 3k + 2 since |B| = q 3 (q − 1). It is easy to check that
these bounds are sufficient.
Finally, let us assume G = 2 B2 (q), where q = 2k with k > 3 odd. Since |B| = q 2 (q − 1),
it follows that l(G) > 2k + 1 + Ω(q − 1) > 2k + 2. By [8, Theorem 4], we also have
λ(G) 6 Ω(k) + 1 + Ω(q − 1) < Ω(k) + k + 1.
20
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
Therefore,
l(G) > 2k + 2 >
5
5
(log2 k + k + 1) > λ(G)
4
4
as required.
Lemma 5.7. Theorem 9 holds if G is a classical group and G ∼
6 L2 (q).
=
Proof. Let G be a finite simple classical group over Fq and r be the twisted rank of G. As
before, write q = pk with p a prime. Let B be a Borel subgroup of G and recall that (10)
holds. Our initial aim is to reduce the problem to groups of small rank. To do this, we
will consider each family of classical groups in turn.
First assume G = Lr+1 (q). We claim that cr(G) > 5/4 if r > 9. To see this, first
observe that
1
l(G) > Ω(|B|) + r > kr(r + 1) + r
2
and λ(G) 6 3Ω(k) + 36 by [8, Theorem 4]. For r > 9, it is routine to check that
1
5
5
l(G) > kr(r + 1) + r > (3 log2 k + 36) > λ(G),
2
4
4
which justifies the claim. In a similar fashion, we can reduce the problem to r 6 6 when
−
G = PSp2r (q), Ω2r+1 (q) or PΩ+
2r (q); r 6 5 when G = PΩ2r+2 (q); and r 6 4 for G = U2r (q).
Finally, suppose G = U2r+1 (q). Here l(G) > kr(2r + 1) + r and [8, Theorem 4] states
that λ(G) 6 3Ω(k) + 36 if q or k is odd. If q = 2k and k is even, then the same theorem
gives
a
λ(G) 6 3Ω(k) + 35 + 2Ω(22 + 1),
a
where k = 2a b and b is odd. Since Ω(22 + 1) 6 k, it follows that λ(G) 6 3Ω(k) + 2k + 35
for all possible values of q and k, and one checks that
5
5
5
l(G) > kr(2r + 1) + r > (3 log2 k + 2k + 35) > (3Ω(k) + 2k + 35) > λ(G)
4
4
4
if r > 5.
Therefore, in order to complete the proof of the lemma, we may assume that we are in
one of the following cases:
(a)
(b)
(c)
(d)
(e)
(f)
(g)
G = Ω2r+1 (q) with 3 6 r 6 6 and q odd;
G = PSp2r (q) with 2 6 r 6 6;
G = PΩ+
2r (q) with 4 6 r 6 6;
−
G = PΩ2r+2 (q) with 3 6 r 6 5;
G = Lr+1 (q) with 2 6 r 6 8;
G = U2r (q) with 2 6 r 6 4;
G = U2r+1 (q) with 1 6 r 6 4.
Let us start by handling case (a). By considering a chain of subfield subgroups, we see
that λ(G) 6 2Ω(k) + λ(Ω2r+1 (p)). In addition, the proof of [8, Theorem 4] implies that
λ(Ω2r+1 (p)) 6 4 + λ(Sn ) for some n 6 2r + 3 = 15. One checks that λ(Sn ) 6 6 for n 6 15,
2
hence λ(G) 6 2Ω(k) + 10. Now |B| = 12 (q − 1)r q r , so (10) yields l(G) > kr 2 + 2r − 1 and
one checks that
5
kr 2 + 2r − 1 > (2 log2 k + 10)
4
for all possible values of k and r. The result follows.
Next consider (b). First assume p = 2, in which case λ(G) 6 Ω(k) + λ(Sp2r (2)) and
λ(Sp2r (2)) 6 4 + λ(Sn ) for some n 6 14 (see the proof of [8, Theorem 4]). Therefore,
λ(G) 6 Ω(k) + 10. Since l(G) > kr 2 + r by (10), the result follows unless (r, k) = (3, 1),
ON THE LENGTH AND DEPTH OF FINITE GROUPS
21
or if r = 2 and k 6 3. If (r, k) = (2, 1) then G ∼
= A6 and cr(G) = 5/4. In each of the
remaining cases we have λ(G) 6 5 and cd(G) > 2, which implies the desired bound. Now
assume p is odd, so l(G) > kr 2 + 2r − 1 and the proof of [8, Theorem 4] yields
λ(G) 6 2Ω(k) + λ(PSp2r (p)) 6 2Ω(k) + 8 + λ(Sr ) 6 2Ω(k) + 13.
This gives cr(G) > 5/4 unless (r, k) = (3, 1), or if r = 2 and k 6 4. If G = PSp6 (p) then
G > L2 (p3 ).3 > L2 (p3 ) > L2 (p)
is unrefinable, so λ(G) 6 7 and the result follows since cd(G) > 2. Now assume r = 2
and k 6 4. Note that |B| = 12 q 4 (q − 1)2 , so Ω(|B|) = 4k + 2Ω(q − 1) − 1. If k = 4 then
λ(G) 6 4 + λ(PSp4 (p)) and we note that one of A6 , S6 or S7 is a maximal subgroup of
PSp4 (p), so λ(PSp4 (p)) 6 7 and thus λ(G) 6 11. In addition, Ω(q − 1) > 5, so l(G) > 27
and the result follows. Similarly, if k = 2 or 3 then λ(G) 6 8 and l(G) > 15. Finally, if
k = 1 then λ(G) 6 7 and the result follows since cd(G) > 2.
Now let us turn to case (c), so G = PΩ+
2r (q) and r = 4, 5 or 6. First assume p = 2, in
which case l(G) > kr(r − 1) + r and λ(G) 6 Ω(k) + λ(Ω+
2r (2)). It is easy to check that
+
λ(Ω2r (2)) 6 9. For example, if r = 6 then there is an unrefinable chain
−
−
Ω+
12 (2) > Sp10 (2) > Ω10 (2).2 > Ω10 (2) > A12
and λ(A12 ) = 5, so λ(Ω+
12 (2)) 6 9. Therefore, l(G) > 12k + 4, λ(G) 6 Ω(k) + 9 and one
checks that these bounds are sufficient. Now assume p > 2. Here l(G) > kr(r − 1) + 2r − 2
+
and λ(G) 6 3Ω(k) + λ(PΩ+
2r (p)). One checks that λ(PΩ2r (p)) 6 9. For instance, if r = 6
then there is an unrefinable chain
PΩ+
12 (p) > PSO11 (p) > Ω11 (p) > H
with H = A12 , S12 or A13 , and the claim follows since λ(H) 6 6. Therefore, l(G) > 12k+6,
λ(G) 6 3Ω(k) + 9 and we conclude that cr(G) > 5/4. A very similar argument applies in
case (d) and we omit the details.
Next consider case (e), so G = Lr+1 (q) and λ(G) 6 2Ω(k) + λ(Lr+1 (p)). Note that
|B| =
q r(r+1)/2 (q − 1)r
.
(r + 1, q − 1)
Suppose r ∈ {3, 5, 7} is odd. Now Lr+1 (p) has a maximal subgroup of the form PSpr+1 (p)
or PSpr+1 (p).2, and we noted that λ(PSpr+1 (p)) 6 13 in the analysis of case (b), whence
λ(G) 6 2Ω(k) + 15. One now checks that the bound l(G) > kr(r + 1)/2 + r from (10)
is sufficient when r = 5 or 7. Now suppose r = 3. If q = 2 then cr(G) = 9/5 so we
can assume q > 2, in which case l(G) > 6k + 5. Now λ(L4 (p)) = 5 if p = 2 or 3, and
λ(L4 (p)) 6 9 if p > 5 (this follows from the fact that PSp4 (p).2 is a maximal subgroup
of L4 (p)). Therefore, λ(G) 6 2Ω(k) + 9 and the result follows if k > 1. Finally suppose
G = L4 (p) with p > 3. If p = 3 then λ(G) = 5 and l(G) > 11. Similarly, λ(G) 6 9 and
l(G) > 13 if p > 5. The result follows.
Now let us assume G = Lr+1 (q) and r ∈ {2, 4, 6, 8}. First assume p is odd. There
is an unrefinable chain Lr+1 (p) > PSOr+1 (p) > Ωr+1 (p), so λ(Lr+1 (p)) 6 12 and thus
λ(G) 6 2Ω(k) + 12. Now l(G) > kr(r + 1)/2 + 2r − 1 and the desired bound follows if
r > 2. Now assume G = L3 (q). Since Ω3 (p) ∼
= L2 (p) we deduce that λ(L3 (p)) 6 6, so
λ(G) 6 2Ω(k) + 6 and one checks that the bound l(G) > 3k + 3 is good enough if k > 2.
If G = L3 (p2 ) then l(G) > 11 and λ(G) 6 8 since there is an unrefinable chain
L3 (p2 ) > L3 (p).2 > L3 (p) > PSO3 (p) > Ω3 (p).
Similarly, if G = L3 (p) then λ(G) 6 6 and we note that l(G) > 10 if p > 5 (this follows
from Theorem 4). Finally, if G = L3 (3) then cr(G) = 8/3.
22
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
To complete the analysis of case (e), let us assume r ∈ {2, 4, 6, 8} and p = 2. Here
l(G) > kr(r + 1)/2 + r and λ(G) 6 2Ω(k) + λ(Lr+1 (2)). As noted in the proof of [8,
Theorem 4], there is an unrefinable chain Lr+1 (2) > 2r .Lr (2) > Lr (2) and one can check
that λ(Lr (2)) 6 5, so λ(G) 6 2Ω(k) + 7. This gives the desired bound unless r = 2 and
k 6 3. We can exclude the case k = 1 since L3 (2) ∼
= L2 (7). For k ∈ {2, 3} we get λ(G) 6 4,
l(G) > 9 and the result follows.
To complete the proof of the lemma, it remains to handle the unitary groups of dimension at most 9 arising in cases (f) and (g). First consider (f), so G = U2r (q), r ∈ {2, 3, 4}
and
q r(2r−1) (q 2 − 1)r
.
|B| =
(2r, q + 1)
By arguing as in the proof of [8, Theorem 4], we see that λ(G) 6 λ(PSp2r (q)) + 2. If
p = 2, it follows that
λ(G) 6 Ω(k) + 2 + λ(Sp2r (2)) 6 Ω(k) + 12
(recall that λ(Sp2r (2)) 6 10). In view of (10) we have l(G) > kr(2r − 1) + 2r − 1 and
one checks that these bounds are sufficient unless r = 2 and k = 1, 2. Here we compute
λ(U4 (4)) = λ(U4 (2)) = 5 and the result follows. Now assume p > 2. Here
λ(G) 6 2Ω(k) + 2 + λ(PSp2r (p)) 6 2Ω(k) + 15
and (10) gives l(G) > kr(2r − 1) + 4r − 2. These estimates give the result, unless r = 2
and k = 1, 2. There is an unrefinable chain
U4 (p2 ) > PSp4 (p2 ).2 > PSp4 (p2 ) > L2 (p2 ) > L2 (p).2 > L2 (p)
and thus λ(G) 6 9 for G = U4 (p2 ). Similarly, one checks that λ(G) 6 9 if G = U4 (p). In
both cases l(G) > 12 and the result follows.
Finally, let us consider case (g), where G = U2r+1 (q), r ∈ {1, 2, 3, 4} and
|B| =
q r(2r+1) (q 2 − 1)r
.
(2r + 1, q + 1)
First assume p > 2. Here
λ(G) 6 λ(Ω2r+1 (q)) + 2 6 2Ω(k) + 2 + λ(Ω2r+1 (p)) 6 2Ω(k) + 12
and (10) gives l(G) > kr(2r + 1) + 4r − 2. One checks that these bounds are sufficient
unless r = 1 and k 6 5. Suppose G = U3 (pk ) with k 6 5. If k = 3 or 5 then λ(G) 6
2 + λ(U3 (p)) 6 8 and the result follows since l(G) > 11. If k = 4 then l(G) > 14 and there
is an unrefinable chain
U3 (p4 ) > PSO3 (p4 ) > Ω3 (p4 ) > L2 (p2 ).2 > L2 (p2 ) > L2 (p).2 > L2 (p),
so λ(G) 6 10. Similarly, if k = 2 then λ(G) 6 8 and l(G) > 11. Finally, suppose
G = U3 (p). If p > 7 then U3 (p) > PSO3 (p) > Ω3 (p) is unrefinable, so λ(G) 6 6. It is easy
to check that the same bound holds when p = 3 or 5, and the desired result now follows
since cd(G) > 2.
Now suppose p = 2. If k is odd then by considering a chain of subfield subgroups we
get λ(G) 6 2Ω(k) + λ(U2r+1 (2)) and one checks that λ(U2r+1 (2)) 6 6. For example,
J3 < U9 (2) is maximal and λ(J3 ) = 5, so λ(U9 (2)) 6 6. Therefore, λ(G) 6 2Ω(k) + 6.
Since l(G) > kr(2r + 1) + 2r − 1, the result follows unless r = 1 and k = 3 (note that
(r, k) 6= (1, 1) since U3 (2) is soluble). A routine computation gives cr(U3 (8)) = 4.
Finally, suppose p = 2 and k is even. Here l(G) > kr(2r + 1) + 3r − 1 and we recall
that λ(G) 6 3Ω(k) + 2k + 35. These bounds are sufficient unless (r, k) = (3, 2), or r = 2
and k ∈ {2, 4, 6}, or if r = 1. If (r, k) = (3, 2) then G = U7 (4) has a maximal subgroup
ON THE LENGTH AND DEPTH OF FINITE GROUPS
23
3277:7 of depth 3, so λ(G) 6 4 and the result follows. Similarly, if r = 2 and k ∈ {2, 4, 6}
then by considering an unrefinable chain through the maximal subgroup
(q 5 + 1)
:5,
(q + 1)(5, q + 1)
we deduce that λ(G) 6 5 and the result follows since l(G) > 10k + 5. Finally, let us
assume G = U3 (2k ) with k even. Now G has a reducible maximal subgroup of the form
q+1
.L2 (q)
(3, q + 1)
and thus
λ(G) 6 λ(L2 (q)) + Ω(q + 1) − Ω((3, q + 1)) + 1 6 Ω(q 2 − 1) + Ω(k) + 2 − Ω((3, q + 1))
since λ(L2 (q)) 6 Ω(q − 1) + Ω(k) + 1 by [8, Theorem 4]. We also have
l(G) = Ω(q 2 − 1) + 3k + 1 − Ω((3, q + 1))
by [32, Theorem 1], and one checks that
5
(Ω(q 2 − 1) + Ω(k) + 2 − Ω((3, q + 1)))
4
if Ω(q 2 −1)+6 < 7k. Since q = 2k we have Ω(q 2 −1)+6 < 2k +6 and the result follows.
Ω(q 2 − 1) + 3k + 1 − Ω((3, q + 1)) >
Lemma 5.8. Theorem 9 holds if G ∼
= L2 (q).
Proof. Write q = pk with p a prime. If k = 1 then λ(G) ∈ {3, 4} by [8, Corollary 3.4],
and we have l(G) > λ(G) + 1, so cr(G) > 5/4 and equality holds if and only if λ(G) = 4
and l(G) = 5 (that is, if and only if G ∈ G; see (3)). The result now follows by combining
Theorems 4 and 5.1. For the remainder, we may assume k > 2.
Suppose p = 2. Here l(G) = Ω(q − 1) + k + 1 and λ(G) 6 Ω(q − 1) + Ω(k) + 1 by [32,
Theorem 1] and [8, Theorem 4(i)]. Now
5
Ω(q − 1) + k + 1 > (Ω(q − 1) + Ω(k) + 1)
4
if and only if
Ω(q − 1) + 5Ω(k) + 1 < 4k.
(11)
Since
Ω(q − 1) + 5Ω(k) + 1 < k + 5 log 2 k + 1,
it is routine to check that (11) holds for all k > 2.
Now suppose p > 2 and observe that l(G) > Ω(q − 1) + k by Lemma 2.5(ii). First
assume k > 3 is odd. By considering a chain of subfield subgroups (as in the proof of
[8, Theorem 4]), we deduce that λ(G) 6 Ω(k) + λ(L2 (p)). Therefore, λ(G) 6 Ω(k) + 2 if
p = 3, so
5
5
5
l(G) > Ω(q − 1) + k > k + 2 > (log3 k + 2) > (Ω(k) + 2) > λ(G)
4
4
4
as required. Similarly, if p > 5 then l(G) > k + 3, λ(G) 6 Ω(k) + 4 and for k > 3 the result
follows in the same way. If k = 3 then λ(G) 6 5 and Theorem 5.1 implies that cd(G) > 2,
so cr(G) > 5/4.
Finally, let us assume p > 2 and k is even. If q = 9 then G ∼
= A6 and we have
already noted that cr(G) = 5/4 in this case. Now assume q > 9, so Ω(q − 1) > 4 and
thus l(G) > k + 4. Also observe that λ(G) 6 2Ω(k) + λ(L2 (p)). If p = 3 then k > 4,
λ(G) 6 2Ω(k) + 2 and one checks that
5
k + 4 > (2 log 2 k + 2),
4
24
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
which gives the desired result. Now assume p > 5. Here λ(G) 6 2Ω(k) + 4 and we have
5
k + 4 > (2 log2 k + 4)
4
if k > 8. If k ∈ {4, 6, 8} then Ω(q − 1) > 6, so l(G) > k + 6 and the result follows. Finally,
if k = 2 then λ(G) 6 6 (since PGL2 (p) < G is maximal) and cd(G) > 2 by Theorem 5.1,
so cr(G) > 5/4 as required.
This completes the proof of Theorem 9. Notice that Corollary 10 follows immediately.
Indeed, we have l(G) 6 a cd(G) if and only if cr(G) > a/(a − 1), so Theorem 9 implies
that a = 5 is the best possible constant.
5.3. Proof of Theorem 11. We begin by recording some immediate consequences of
Lemma 2.2.
Proposition 5.9. Let G be a finite group.
(i) If 1 = GP
m ✁ Gm−1 ✁ · · · ✁ G1 ✁ G0 = G is a chain of subgroups of G, then
cd(G) > i cd(Gi−1 /Gi ).
P
(ii) If G = G1 × · · · × Gm , then cd(G) > i cd(Gi ).
(iii) If T1 , . . . P
, Tm are the composition factors of G, listed with multiplicities, then
cd(G) > i cd(Ti ).
Note that in part (iii) above we havePcd(Ti ) = 0 if Ti is abelian, so only the nonabelian composition factors contribute to i cd(Ti ). Now let T1 , . . . , Tm be the non-abelian
composition factors of G (listed with multiplicities), and let
ss(G) =
m
Y
Ti
i=1
be their direct product.
Proposition 5.10. We have l(ss(G)) 6 5 cd(G) for every finite group G.
Proof. With the above notation we have
l(ss(G)) =
m
X
l(Ti ).
i=1
By Corollary 10 we have l(Ti ) 6 5 cd(Ti ) for each i, and by combining this with Proposition
5.9(iii) we obtain
l(ss(G)) 6
m
X
5 cd(Ti ) = 5
m
X
cd(Ti ) 6 5 cd(G)
i=1
i=1
as required.
By a semisimple group we mean a direct product of (non-abelian) finite simple groups.
Lemma 5.11. If G is a finite semisimple group, then l(Aut(G)) 6 2l(G).
Q
ki
Proof. Write G = m
i=1 Ti where the Ti are pairwise non-isomorphic finite (non-abelian)
simple groups and ki > 1. Then
Aut(G) ∼
=
m
Y
i=1
Aut(Tiki )
∼
=
m
Y
i=1
Aut(Ti ) ≀ Ski .
ON THE LENGTH AND DEPTH OF FINITE GROUPS
Hence Out(G) ∼
=
Qm
i=1 Out(Ti )
m
X
≀ Ski , so
ki l(Out(Ti )) + l(Ski ) 6
l(Out(G)) =
m
X
i=1
i=1
25
ki (log2 |Out(Ti )| + 3/2),
where the last inequality follows from the main theorem of [10] on the length of the
symmetric group. Using the well known orders of Out(T ) for the finite simple groups T
(for example, see [21], pp. 170-171), it is easy to verify that log2 |Out(Ti )| + 3/2 6 l(Ti )
for all i. We conclude that
m
X
ki l(Ti ) = l(G)
l(Out(G)) 6
i=1
and thus l(Aut(G)) 6 2l(G) as required.
We are now ready to prove Theorem 11. Let R(G) be the soluble radical of G and
consider the semisimple group Soc(G/R(G)). Applying Proposition 5.10 to the group
G/R(G) we obtain
l(Soc(G/R(G))) 6 l(ss(G/R(G))) 6 5 cd(G/R(G)).
It is well known that G/R(G) 6 Aut(Soc(G/R(G))). Applying the inequality above with
Lemma 5.11 we obtain
l(G/R(G)) 6 l(Aut(Soc(G/R(G)))) 6 2l(Soc(G/R(G))) 6 10 cd(G/R(G)) 6 10 cd(G)
and the result follows.
This completes the proof of Theorem 11.
Appendix A. On the number of prime divisors of p ± 1
by D.R. Heath-Brown
In this appendix, we prove the following result.
Theorem A.1. There are infinitely many primes p ≡ 5 (mod 72) for which
Ω((p2 − 1)/24)) 6 7.
Hence there are infinitely many primes p for which
max{Ω(p ± 1)} 6 8.
We begin by showing how the second claim follows from the first. For any prime
p ≡ 5 (mod 72) one has 24|(p2 − 1). Indeed for such primes one has (p − 1, 72) = 4
and (p + 1, 72) = 6. One necessarily has Ω((p − 1)/4) > 1 when p > 5, so that if
Ω((p2 − 1)/24)) 6 7 one must have Ω((p + 1)/6)) 6 6. It then follows that Ω(p + 1) 6 8.
The proof that Ω(p − 1) 6 8 is similar.
To handle the first statement of the theorem we use sieve methods, as described in
the book by Halberstam and Richert [15], and in particular the weighted sieve, as in [15,
Chapter 10]. To be specific, we apply [15, Theorem 10.2] to the set
A = {(p2 − 1)/24 : p ≡ 5 (mod 72), p 6 x}
and the set P of all primes. The expected value of
|Ad | := #{n ∈ A : d|n}
is Xω(d)/d, with X = Li(x)/24, and where ω(d) is a multiplicative function satisfying
0 if p = 2, 3,
ω(p) =
2 if p > 5.
26
TIMOTHY C. BURNESS, MARTIN W. LIEBECK, AND ANER SHALEV
Condition (Ω1 ), see [15, p.29], is then satisfied with A1 = 2, while condition (Ω∗2 (κ)), see
[15, p.252], holds with κ = 2 and a suitable numerical constant A2 . Moreover |Ap2 | =
O(xp−2 ), which shows that condition (Ω3 ), see [15, p.253], also holds, for an appropriate
numerical constant A3 . Finally we consider the condition (Ω(R(2, α)) given in [15, p.219].
The primes p ≡ 5 (mod 72) for which d divides (p2 − 1)/24 fall into ω(d) residue classes
modulo 72d, so that (Ω(R(2, 12 )) holds by an appropriate form of the Bombieri–Vinogradov
theorem, as in [15, Lemma 3.5]. This verifies all the necessary conditions for Theorem 10.2
of [15], and the inequality (2.2) of [15, p.278] is satisfied (with α = 21 ) for any constant
µ > 4, if x is large enough.
Theorem 10.2 of [15] then tells us that there are ≫ X(log X)−2 elements n ∈ A which
are “Pr -numbers” (that is to say, one has Ω(n) 6 r), provided that
Rv
1
1 − ut dtt
2 u σ2 (v(α−1/t))
.
r > 2u − 1 +
1 − η2 (αv)
Finally, we refer to the calculations of Porter [29], and in particular the last 3 lines of [29,
p.420], according to which it will suffice to have r > 6.7 if one takes u = 2.2 and v = 22.
Since we then have α−1 < u < v and αv = 11 > ν2 = 4.42 . . ., by Porter [29, Table 2], the
final conditions (2.3) of [15, Theorem 10.2] are satisfied, and our theorem follows.
References
[1] K. Alladi, R. Solomon and A. Turull, Finite simple groups of bounded subgroup chain length, J. Algebra
231 (2000), 374–386.
[2] M. Aschbacher, On the maximal subgroups of the finite classical groups, Invent. Math. 76 (1984),
469–514.
[3] L. Babai, On the length of subgroup chains in the symmetric group, Comm. Algebra 14 (1986), 1729–
1736.
[4] R.W. Baddeley, Primitive permutation groups with a regular nonabelian normal subgroup, Proc. London Math. Soc. 67 (1993), 547–595.
[5] G.W. Bell, On the cohomology of the finite special linear groups, I, J. Algebra 54 (1978), 216–238.
[6] J.N. Bray, D.F. Holt and C.M. Roney-Dougal, The maximal subgroups of the low-dimensional finite
classical groups, London Math. Soc. Lecture Note Series, vol. 407, Cambridge University Press, 2013.
[7] B. Brewster, M. Ward and I. Zimmermann, Finite groups having chain difference one, J. Algebra 160
(1993), 179–191.
[8] T.C. Burness, M.W. Liebeck and A. Shalev, The depth of a finite simple group, Proc. Amer. Math.
Soc., to appear.
[9] T.C. Burness, M.W. Liebeck and A. Shalev, The length and depth of algebraic groups, submitted
(arXiv:1712.08214).
[10] P.J. Cameron, R. Solomon and A. Turull, Chains of subgroups in symmetric groups, J. Algebra 127
(1989), 340–352.
[11] A.M. Cohen, M.W. Liebeck, J. Saxl, and G.M. Seitz, The local maximal subgroups of exceptional
groups of Lie type, finite and algebraic, Proc. London Math. Soc. 64 (1992), 21–48.
[12] J.H. Conway, R.T. Curtis, S.P. Norton, R.A. Parker and R.A. Wilson, Atlas of Finite Groups, Oxford
University Press, 1985.
[13] D.A. Craven, Alternating subgroups of exceptional groups Lie type, Proc. Lond. Math. Soc. 115 (2017),
449–501.
[14] A. Gamburd and I. Pak, Expansion of product replacement graphs, Combinatorica 26 (2006), 411–429.
[15] H. Halberstam and H.-E. Richert, Sieve methods, LMS Monographs, Academic Press, London, 1974.
[16] K. Harada, Finite simple groups with short chains of subgroups, J. Math. Soc. Japan 20 (1968),
655-672.
[17] M.A. Hartenstein and R.M. Solomon, Finite groups of chain difference one, J. Algebra 229 (2000),
601–622.
[18] K. Iwasawa, Über die endlichen Gruppen und die Verbände ihrer Untergruppen, J. Fac. Sci. Imp. Univ.
Tokyo. Sect. I. 4 (1941), 171–199.
[19] Z. Janko, Finite groups with invariant fourth maximal subgroups, Math. Z. 82 (1963), 82–89.
[20] Z. Janko, Finite simple groups with short chains of subgroups, Math. Z. 84 (1964), 428–437.
[21] P. Kleidman and M.W. Liebeck, The subgroup structure of the finite classical groups, London Math.
Soc. Lecture Note Series, vol. 129, Cambridge University Press, 1990.
ON THE LENGTH AND DEPTH OF FINITE GROUPS
27
[22] J. Kohler, A note on solvable groups, J. London Math. Soc. 43 (1968), 235–236.
[23] M.W. Liebeck, C.E. Praeger and J. Saxl, A classification of the maximal subgroups of the finite
alternating and symmetric groups, J. Algebra 111 (1987), 365–383.
[24] M.W. Liebeck, C.E. Praeger and J. Saxl, On the O’Nan-Scott Theorem for finite primitive permutation
groups, J. Austral. Math. Soc. 44 (1988), 389–396.
[25] M.W. Liebeck, J. Saxl, and G.M. Seitz, Subgroups of maximal rank in finite exceptional groups of Lie
type, Proc. London Math. Soc. 65 (1992), 297–325.
[26] M.W. Liebeck and G.M. Seitz, A survey of of maximal subgroups of exceptional groups of Lie type,
in Groups, combinatorics & geometry (Durham, 2001), 139–146, World Sci. Publ., River Edge, NJ,
2003.
[27] M.W. Liebeck and G.M. Seitz, On finite subgroups of exceptional algebraic groups, J. reine angew.
Math. 515 (1999), 25–72.
[28] J. Petrillo, On the length of finite simple groups having chain difference one, Arch. Math. 88 (2007),
297–303.
[29] J.W. Porter, Some numerical results in the Selberg sieve method, Acta Arith. 20 (1972), 417–421.
[30] G.M. Seitz, R. Solomon and A. Turull, Chains of subgroups in groups of Lie type, II, J. London Math.
Soc. 42 (1990), 93–100.
[31] J. Shareshian and R. Woodroofe, A new subgroup lattice characterization of finite solvable groups, J.
Algebra 351 (2012), 448–458.
[32] R. Solomon and A. Turull, Chains of subgroups in groups of Lie type, I, J. Algebra 132 (1990),
174–184.
[33] R. Solomon and A. Turull, Chains of subgroups in groups of Lie type, III. J. London Math. Soc. 44
(1991), 437–444.
[34] R.A. Wilson, Maximal subgroups of sporadic groups, in Finite Simple Groups: Thirty Years of the
Atlas and Beyond, 57–72, Contemp. Math., 694, Amer. Math. Soc., Providence, RI, 2017.
[35] R.A. Wilson, The uniqueness of PSU3 (8) in the Monster, Bull. London Math. Soc. 49 (2017), 877–880.
T.C. Burness, School of Mathematics, University of Bristol, Bristol BS8 1TW, UK
E-mail address: [email protected]
D.R. Heath-Brown, Mathematical Institute, Radcliffe Observatory Quarter, Woodstock
Road, Oxford OX2 6GG, UK
E-mail address: [email protected]
M.W. Liebeck, Department of Mathematics, Imperial College, London SW7 2BZ, UK
E-mail address: [email protected]
A. Shalev, Institute of Mathematics, Hebrew University, Jerusalem 91904, Israel
E-mail address: [email protected]
| 4math.GR
|
A hybrid exact-ACO algorithm for the joint
scheduling, power and cluster assignment in
cooperative wireless networks ⋆
Fabio D’Andreagiovanni1,2
arXiv:1704.06684v1 [math.OC] 21 Apr 2017
1
2
Department of Optimization,
Zuse-Institut Berlin (ZIB), 14195 Berlin, Germany
Department of Computer, Control and Management Engineering,
Sapienza Università di Roma, 00185 Roma, Italy
[email protected],
Home page: http://www.dis.uniroma1.it/∼fdag/
Abstract. Base station cooperation (BSC) has recently arisen as a
promising way to increase the capacity of a wireless network. Implementing BSC adds a new design dimension to the classical wireless network
design problem: how to define the subset of base stations (clusters) that
coordinate to serve a user. Though the problem of forming clusters has
been extensively discussed from a technical point of view, there is still
a lack of effective optimization models for its representation and algorithms for its solution. In this work, we make a further step towards
filling such gap: 1) we generalize the classical network design problem by
adding cooperation as an additional decision dimension; 2) we develop a
strong formulation for the resulting problem; 3) we define a new hybrid
solution algorithm that combines exact large neighborhood search and
ant colony optimization. Finally, we assess the performance of our new
model and algorithm on a set of realistic instances of a WiMAX network.
Keywords: cooperative wireless networks, binary linear programming,
exact local search, ant colony optimization, hybridization
1
Introduction
Wireless communications have experienced an astonishing expansion over the
last years and network operators have consequently faced the challenge of providing higher capacity with the same limited amount of radio resources. In
⋆
This is the author’s final version of the paper published in G. Di Caro, G. Theraulaz (eds.), BIONETICS 2012: Bio-Inspired Models of Network, Information, and
Computing Systems. Lecture Notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering (LNICST, vol. 134, pp. 3-17.
Springer, Heidelberg, 2014, DOI: 10.1007/978-3-319-06944-9 1 ). The final publication is available at Springer via http://dx.doi.org/10.1007/978-3-319-06944-9 1. This
work was partially supported by the German Federal Ministry of Education and Research (BMBF), project ROBUKOM [4,27], grant 03MS616E. The Author thanks
Antonella Nardin for fruitful discussions.
2
Fabio D’Andreagiovanni
this context, the capacity increase has been mainly pursued through interference avoidance: by limiting interference, a higher number of users can indeed
be served, thus increasing network capacity. Interference avoidance can be first
obtained by a careful setting of the power emissions of the base stations constituting the network. In recent years, it was showed that an additional reduction
in interference can be obtained by allowing cooperation among the base stations:
the central feature of cooperation is that service can be provided by a number
of base stations forming a cluster, in contrast to classical systems where service
is provided by a single base station. These new wireless systems are called cooperative wireless networks (CWN). The advantage of a CWN is clear: a user that
requires a single powerful signal can be instead served by a (small) number of
signals of lower power, that reduce interference towards other users. Remarkable
gain in capacity are thus possible in wireless technologies supporting cooperation, such as LTE [33] and WiMAX [2]. On the other hand, cooperation has
a cost and must be kept low: cooperating base stations must indeed coordinate
and need to exchange a non-negligible amount of information. For a detailed discussion of the potentialities and advantages of adopting cooperation in wireless
networks, we refer the reader to the recent work [21].
Cooperation introduces an additional decision dimension in the problem of
designing a wireless network, thus making an already complicated problem even
more complicated. In this context, Mathematical Optimization has arisen as a
very effective methodology to improve the quality of design, as proven in successful industrial collaborations with major wireless providers [12,29,31]. Optimization is furthermore able to overcome the limitations of the trial-and-error
approach commonly adopted by network engineers, by pursuing a much more
efficient usage of the scarce radio resources of a network.
Tough CWNs have attracted a lot of attention from a technical and theoretical point of view, there is still a lack of effective optimization models and
algorithms for their design. Moreover, available models and algorithms do not
treasure previous experience about how tackling specific computational issues
that affect the general design problem. The main objective of our work is to
make further steps towards closing such gap. Specifically, our original contributions are: 1) a new binary linear model for the design of cooperative wireless
networks, that jointly optimizes activated service links, power emissions and
cooperative clusters; 2) a strengthening procedure that generates a stronger optimization model, in which sources of numerical issues are completely eliminated;
3) a hybrid solution algorithm, based on the combination of a special exact large
neighborhood search called RINS [11] with ant colony optimization [20] (note
that tough generic heuristics and pure and hybrid nature-inspired metaheuristics
are not a novelty in (wireless) network design - see e.g., [3,5,9,14] - to the best of
our knowledge such combination has not yet been investigated). This algorithm
is built to effectively exploit the precious information coming from the stronger
model and can rapidly find solutions of very satisfying quality.
To our best knowledge, our model is the first that considers the joint optimization of activated service links, power emissions and cooperative clusters.
A hybrid exact-ACO algorithm for Cooperative Wireless Network Design
3
Models available in literature have indeed focused attention on more specific versions of the problem. In [34], a set covering formulation for selecting clusters is
introduced and solved by a greedy heuristic. A weakness of this work is that the
feasible clusters must be explicitly listed and their number is in general exponential. In [22], a linear binary formulation for the problem of forming cooperative
clusters and assigning users to base stations is proposed and a limited computational experience on small-sized instances is presented. In [24] a non-linear model
for the joint optimal clustering and beamforming in CWN is investigated. Given
the intrinsic hardness of the resulting non-linear problem, the Authors propose
an iterative heuristic approach that makes use of mean square errors. In contrast
to our work, all these works do not consider the computational issues associated
with the coefficients representing signal propagation and thus their models are
weak and their solution approaches may lead to coverage errors, as explained in
[8,12,26].
The remainder of this paper is organized as follows: in Section 2, we introduce our new model for the design of cooperative wireless networks and discuss
its strength; in Section 3 we present our hybrid metaheuristic; in Section 4 we
present preliminary computational results over a set of realistic WiMAX instances and, finally, in Section 5, we derive some conclusions.
2
Cooperative Wireless Network Design
We consider the design of a wireless network made up of a set B of base stations
(BSs) that provide a telecommunication service to a set T of user terminals
(UTs). A UT t ∈ T is said to be covered (or served ), if it receives the service
within a minimum level of quality. The BSs and the UTs are characterized by
a number of parameters (e.g., geographical location, power emission, frequency
used for transmitting). The Wireless Network Design Problem (WND) consists
in setting the values of the base station parameters, with the aim of maximizing
a revenue function associated with coverage (e.g., number of covered UTs). For
an exhaustive introduction to the WND, we refer the reader to [12,13,26].
An optimization model for the WND typically focuses attention only on a
subset of the parameters characterizing a BS. In particular, the majority of
the models considers the power emission and the assignment of served UTs to
BSs as the main decision variables. These are indeed two critical decisions that
must be taken by a network administrator, as indicated in several real studies
(e.g., [1,6,7,12,29,31]). Other parameters that are commonly considered are the
frequency and the transmission scheme used to serve a terminal (e.g., [10,17,30]).
Different versions of the WND are discussed in [13], where a hierarchy of WND
problems is also presented.
In this work, we consider a generalization of the Scheduling and Power Assignment Problem (SPAP), a version of the WND that is known to be HP-Hard
[29]. In the SPAP, two decisions must be taken: 1) setting the power emission of
each BS and 2) assigning served UTs to activated cluster of BSs (note that this
corresponds to identify a subset of service links BS-UT that can be scheduled
4
Fabio D’Andreagiovanni
simultaneously without interference, so we use the term scheduling). Since we
address the problem of designing a cooperative network, in our generalization we
also include a third decision: for each served UT, we want to select a subset of
BSs forming the cluster that serves the UT. We call the resulting generalization
by the name Scheduling, Power and Cluster Assignment Problem (SPCAP).
To model these three decisions, we introduce three types of decision variables:
1. a non-negative discrete variable pb ∈ P = {P1 , . . . , P|P| }, with P|P| = Pmax
and Pl > Pl−1 > 0, for l = 2, . . . , |P|, representing the power emission of
a BS b. Such variable can be represented as the linear combination of the
power values Pl multiplied by binary variables: to this end, for each b ∈ B
we introduce one binary variable zbl (power variable) that is equal to 1 if b
emits at power Pl and 0 otherwise. Additionally, we must express the fact
that each BS may emit only at a single power level. If we denote the set of
feasible
of the binary power variables is:
P
P power levels by L, the definition
pb = P l∈L Pl zbl , ∀b ∈ B, with l∈L zbl ≤ 1. We remark that a BS b such
that l∈L zbl = 0 has null power (i.e., pb = 0) and therefore is not deployed
in the network. Our model thus also decides the localization of the BSs to
be deployed, though this feature is not explicitly pointed out by defining a
specific decision variable.
2. a binary service assignment variable xt ∈ {0, 1}, ∀ t ∈ T , that is equal to 1
if UT t ∈ T is served and to 0 otherwise.
3. a binary cluster assignment variable ytb ∈ {0, 1}, ∀ t ∈ T, b ∈ B, that is
equal to 1 if BS b belongs to the cluster that serves UT t and to 0 otherwise.
Every UT t ∈ T picks up signals from each BS b ∈ B in the network and the
power Ptb that t gets from b is proportional to the emitted power pb by a factor
atb ∈ [0, 1], i.e. Ptb = atb · pb . The factor atb is an attenuation coefficient that
summarizes the reduction in power experienced by a signal propagating from b
to t [31]. The BSs whose signals are picked up by t can be subdivided into useful
and interfering, depending on whether they contribute either to guarantee or to
destroy the service. Given a cluster of BSs Ct ⊆ B that cooperate to serve t,
the BSs in Ct provide useful signals to t, while all the BSs that do not belong
to Ct are interferers. For a given service cluster Ct , a UT t is considered served
if the ratio of the sum of the useful powers to the sum of the interfering powers
(signal-to-interference ratio or SIR) is above a threshold δt > 0, which depends
on the desired quality of service [31]:
P
b∈Ct atb pb
P
≥ δt .
(1)
N + b∈B\Ct atb pb
Note that in the denominator, we highlight the presence of the system noise
N > 0 among the interfering terms. By simple linear algebra operations, we
can reorganize the ratio (1) into the following inequality, commonly called SIR
inequality:
X
X
atb pb ≥ δt N .
(2)
atb pb − δt
b∈Ct
b∈B\Ct
A hybrid exact-ACO algorithm for Cooperative Wireless Network Design
5
We now explain how to modify the basic SIR inequality (2), in order to build our
optimization problem. As first step, we need to introduce the service assignment
variable xt into (2), as we want to decide which UTs t ∈ T are served:
X
atb pb − δt
b∈Ct
X
atb pb + M (1 − xt ) ≥ δt N .
(3)
b∈B\Ct
This inequality contains a large positive constant M (the so-called big-M coefficient [12,32]), that in combination with the binary variable xt either activates
or deactivates the constraint. It is straightforward to check that if xt = 1, then
(3) reduces to (2) and the SIR inequality must be
Psatisfied. If instead xt = 0 and
M is sufficiently large (for example, M = +δt
b∈B\Ct atb Pmax + δt N ), then
(3) becomes redundant, as it is satisfied by any power configuration of the BSs.
As second step, we need to introduce the cluster assignment variables ytb into
(3), as we want to decide which BSs b ∈ B form the cluster Ct serving t:
X
atb pb ytb − δt
b∈B
X
atb pb (1 − ytb ) + M (1 − xt ) ≥ δt N .
(4)
b∈B
In this case, we extend the summation of the inequalities to all the BSs b ∈ B
and the service cluster Ct is made up of the BSs b such that ytb = 1. The BSs
that are not in the cluster act as interferers and this is expressed by including
the complement
1 − ytb of the cluster variables into the interfering summation.
P
By pb = l∈L Pl · zbl and by simply reorganizing the terms, we finally obtain:
(1 + δh )
XX
b∈B l∈L
atb Pl zbl ytb − δt
XX
atb Pl zbl + M (1 − xt ) ≥ δt N . (5)
b∈B l∈L
This last inequality constitutes the core of our model and is called SIR constraint.
It is actually non-linear as it includes the product zbl ytb of binary variables.
However, this does not constitute an issue, as this is a special non-linearity that
can be linearized in a standard way, by introducing a new binary variable vtbl =
zbl ytb for each triple (t, b, l) : t ∈ T, b ∈ B, l ∈ L, and three new constraints: (c1)
vtbl ≤ zbl ; (c2) vtbl ≤ ytb , (c3) vtbl ≥ zbl + ytb − 1 [23]. If zbl = 0 or ytb = 0,
then by any of (c1),(c2) we have vtbl = 0. If instead zbl = ytb = 1, then by (c3)
we have vtbl = 1. After these considerations, we can finally state the following
6
Fabio D’Andreagiovanni
pure binary linear formulation for the SPCAP:
!
X
X
X
max
rt xt −
ct
ytb − xt
t∈T
(1 + δh )
t∈T
XX
atb Pl vtbl − δt
b∈B l∈L
X
zbl ≤ 1
(big-M SPCAP)
b∈B
XX
atb Pl zbl + M (1 − xt ) ≥ δt N
b∈B l∈L
t ∈ T,
(6)
b ∈ B,
(7)
t ∈ T, b ∈ B, l ∈ L,
(8)
l∈L
vtbl ≤ zbl
vtbl ≤ ytb
t ∈ T, b ∈ B, l ∈ L,
(9)
vtbl ≥ zbl + ytb − 1
xt ∈ {0, 1}
t ∈ T, b ∈ B, l ∈ L,
t ∈ T,
(10)
(11)
zbl ∈ {0, 1}
ytb ∈ {0, 1}
b ∈ B, l ∈ L,
t ∈ T, b ∈ B,
(12)
(13)
vtbl ∈ {0, 1}
t ∈ T, b ∈ B, l ∈ L .
(14)
The objective function maximizes the difference between the total revenue obtained by serving UTs (each served UT t ∈ T grants a revenue rt > 0) and
the total cost incurred to establish cooperation between BSs (for each UT t, the
cooperation cost ct > 0 arises for each cooperating BS when at least two BSs
are cooperating; no cost arises when the service is provided by a single BS). The
linear SIR constraints (6) express the coverage conditions under cooperation.
The constraints (7) ensure that each BS emits at a single power level, while
the constraints (8)-(10) operate the linearization of the product zbl ytb . Finally,
(11)-(14) define the decision variables of the problem.
Strengthening the formulation for the PSCAP. The formulation (big-M
SPCAP) constitutes a natural optimization model for the SPCAP. However,
because of the presence of the big-M coefficients, it is known to be really weak,
i.e. its linear relaxation produces very low quality bounds [32]. Moreover, as
the fading coefficients typically vary in a very wide range, the coefficient matrix
may be very ill-conditioned and this leads to heavy numerical instability in
the solution process. As a consequence, the effectiveness of standard solution
algorithms provided by state-of-the-art commercial solvers, such as IBM ILOG
Cplex [25], may be greatly reduced and solutions may even contain coverage
errors, as pointed out in several works (e.g., [8,12,13,15,26,29]).
In order to tackle these computational issues, we extend a very effective
strengthening approach that we proposed in [12,13,16]:
P the basic idea is to exploit
the generalized upper bound
(GUB)
constraints
l∈L vtbl ≤ 1, implied by the
P
other GUB constraints l∈L zbl ≤ 1 (as vtbl ≤ zbl ), to replace the SIR constraints
(6) with a set of GUB cover inequalities. For an exhaustive introduction to the
A hybrid exact-ACO algorithm for Cooperative Wireless Network Design
7
cover inequalities and to their GUB version, we refer the reader to the book [32]
and to [35]. Here, we briefly recall the well-known
main results about the general
P
a
cover inequalities: a knapsack constraint
j∈J j xj ≤ b with aj , b ∈ R+ and
xj ∈ {0, 1}, ∀j ∈ J,
can
be
replaced
with
a
(in
general exponential) number of
P
cover inequalities j∈C xj ≤ |C| − 1, where C is a cover. A cover is a subset of
indices C ⊆ J such that the summation ofP
the corresponding coefficients aj , j ∈
C violates the knapsack constraint, i.e.
j∈C aj > b. The cover inequalities
thus represent combinations of binary variables xj that cannot be activated at
the same time (therefore, we can activate at most |C| − 1 variables in each
cover C). The GUB cover inequalities are a stronger version of the simple cover
inequalities,
that can be defined when there are additional constraints of the
P
form j∈K⊆J xj ≤ 1 (GUB constraints), that allows to activate at most one
variable in the subset K.
By applying the definition of [35] and reasoning similarly to [12], we can
define the general form of the GUB cover inequalities (GCIs) needed to replace
the SIR constraint (6):
xt +
|∆| λi
X
X
vtbl +
|Γ | |L|
X
X
zbl ≤ |∆| + |Γ | ,
(15)
i=1 l=qi
i=1 l=1
with t ∈ T , ∆ ⊆ B, λ = (λ1 , . . . , λ|∆| ) ∈ L|∆| , Γ ⊆ B\∆, (q1 , . . . , q|Γ | ) ∈
LI (t, ∆, λ, Γ ), with LI (t, ∆, λ, Γ ) ⊆ L|Γ | representing the subset of interfering
levels of BSs in Γ that destroy the service of t provided by the cluster ∆ of BSs,
emitting with power levels λ = (λ1 , . . . , λ|∆| ). Intuitively, for fixed UT, subset
of serving BSs and subset of interfering BSs, a GCI is built by fixing a power
setting of the serving BSs and defining a power setting of the interfering BSs
that deny the coverage of the considered UT.
If we replace the big-M SIR constraints (6) with the GCIs (15) in the big-M
formulation (big-M SPCAP), we obtain what we call a Power-Indexed formulation (PI-SPCAP). The Power-Indexed formulation (PI-SPCAP) does not contain
big-M and fading coefficients and is very strong and completely stable. On the
other hand, it generally contains an exponential number of constraints and must
be solved through a typical branch-and-cut approach [32]: initially we define a
starting formulation containing only a suitable subset of GCIs (15), then we insert additional GCIs if needed, through the solution of an auxiliary separation
problem (we refer the reader to [12] for details about the separation of the GCIs
of a Power-Indexed formulation). In our case, the starting formulation contains
the GCIs:
xt +
λ
X
l=1
vtβl +
|L|
X
zbl ≤ 2.
(16)
l=q
Such GCIs are obtained by considering a relaxed version of the SIR constraints
(6), which contain only a single-server and a single-interferer (i.e., there are no
summations over multiple serving and interfering BSs). This relaxation comes
8
Fabio D’Andreagiovanni
from the practical observation that it is common to have a server and an interferer BSs that are sensibly stronger than all the other and coverage of the user
just depends on the power configuration of them [12,15]. Computational experience also shows that the relaxed SIR constraints provide a good approximation
of (big-M SPCAP) [12]. The GCIs (16) of the relaxed SIR constraints can be
used to define a relaxed Power-Indexed formulation denoted by (PI0 -SPCAP),
that constitutes a very good starting point for a branch-and-cut algorithm used
to solve (PI-SPCAP), as reported in [12].
Following the features of the improved ANT algorithm proposed in [28], in
our ANT algorithm presented in the next section, we make use of two lower
bounds for the SPCAP: 1) the one obtained by solving the linear relaxation of
(PI-SPCAP), that we will denote by PI-bound ; 2) the one obtained by solving the linear relaxation of (big-M SPCAP), strengthened with the GCIs of
(PI0 -SPCAP), that we will denote by BM-bound ; We remark that PI-bound is
consistently better than BM-bound, as it comes from a stronger formulation.
However, its computation takes more time than that of BM-bound, as it requires
to separate additional GCIs (we recall that (PI-SPCAP) initially corresponds to
(PI0 -SPCAP) and its solution generally requires to generate additional GCIs).
3
A hybrid exact-ACO algorithm for the SPCAP
Ant Colony Optimization (ACO) is a metaheuristic approach to combinatorial
optimization problems that was originally proposed by Dorigo and colleagues,
in a series of works from the 1990s (e.g., [20]), and it was later extended to
integer and continuous optimization problems (e.g., [19]). For an overview of
the theory and applications of ACO, we refer the reader to [5,18,19]. It is now
common knowledge that the algorithm draws its inspiration from the foraging
behaviour of ants. The basic idea of an ACO algorithm is to define a loop where
a number of feasible solutions are iteratively built in parallel, exploiting the
information about the quality of solutions built in previous executions of the
loop. The general structure of an algorithm can be depicted as follows:
UNTIL an arrest condition is not satisfied DO
1. Ant-based solution construction
2. Daemon actions
3. Pheromone trail update
(Gen-ACO)
We now describe the details of each phase presented above for our hybrid exactACO algorithm for the SPCAP. Our approach is hybrid since the canonical
ACO step 1 is followed by a daemon action phase, where we exactly explore
a large neighborhood of the generated feasible solutions, by exactly solving a
Mixed-Integer Program, as explained in Subsection 3.2.
3.1
Ant-based solution construction
In step 1 of the loop, m ∈ Z+ computational agents called ants are defined and
each ant iteratively builds a feasible solution for the optimization problem. At
A hybrid exact-ACO algorithm for Cooperative Wireless Network Design
9
every iteration, the ant is in a so-called state, associated with a partial solution
to the problem, and can complete the solution by selecting a move among a set
of feasible ones. The move is probabilistically chosen on the basis of its associated pheromone trail values. For a more detailed description of the elements
and actions of step 1, we refer the reader to the paper by Maniezzo [28]. The
paper proposes an improved ANT algorithm (ANTS), which we take as reference
for our work. We were particularly attracted by the improvements proposed in
ANTS, as they are based on the attempt of better exploiting the precious information that comes from a strong Linear Programming formulations of the
original discrete optimization problem. Moreover, ANTS also uses a reduced
number of parameters and adopts mathematical operations of higher computational efficiency (e.g., multiplications instead of powers with real exponents).
Before describing the structure and the behaviour of our ants, we make some
preliminary considerations. Our formulation for the SPCAP is based on four
types of variables: 1) power variables zbl ; 2) cluster variables ytb ; 3) service variables xt ; 4) linearization variables vtbl . Once that the power variables and the
cluster variables are fixed: i) the value of the linearization variables is immediately determined, because of constraints (8)-(10), and ii) the objective function
can be easily computed, as the served UTs can be identified by simply checking
which SIR inequalities (6) are satisfied. As a consequence, in the ant-construction
phase we can limit our attention to the cluster and service variables and we introduce the following two concepts of power and cluster states.
Definition 1. Power state (PS): let L0 = L ∪ {0} be the set of power levels plus
the null power level. A power state represents the activation of a subset of BSs
on some power level l ∈ L0 and excludes that the same BS is activated on two
power levels. Formally: P S ⊆ B × L0 :6 ∃(b1 , l1 ), (b2 , l2 ) ∈ P S : b1 = b2 .
We say that a power state PS is complete when it specifies the power configuration of every BS in B (thus |P S| = |B|). Otherwise the PS is called partial and we
have |P S| < |B|. Furthermore, for a given power state PS, we denote by B(P S)
the subset of BSs whose power is fixed in PS (we call such BSs configured ), i.e.
B(P S) = {b ∈ B : ∃(b, l) ∈ P S}.
Definition 2. Cluster state (CS): let B̄ = B and let T × B̄ be the set of couples
(t, b) representing the non-assignment of BS b to the cluster serving UT t. A
cluster state represents the assignment or non-assignment of a subset of BSs to
the clusters serving a subset of UTs and excludes that the same BS is at the
same time assigned and non-assigned to the cluster of a UT. Formally: CS ⊆
T ×B ∪ T ×B̄ : 6 ∃(t1 , b1 ), (t2 , b2 ) ∈ CS : b1 ∈ B, b2 ∈ B̄ and t1 = t2 , b1 = b2 = b.
We say that a cluster state CS is complete when it specifies the cluster assignment
or non-assignment of every BS in B to every UT in T (thus |CS| = |T ||B|).
Otherwise the CS is called partial and we have |CS| < |T ||B|). Moreover, for
a given cluster state CS and UT t, we denote by Bt (CS) the subset of BSs
that are either assigned or not assigned to the service cluster of t in CS, i.e.
Bt (CS) = {b ∈ B ∪ B̄ : ∃(t, b) ∈ CS}.
10
Fabio D’Andreagiovanni
In our ANT algorithm, we decided to first establish the value of the power
variables and then the cluster variables. So an ant first passes through a sequence
of partial power states, till a complete one is reached (power construction phase).
Then it passes through a sequence of partial cluster states, till a complete one is
reached (cluster construction phase). More formally, in the power phase, an ant
moves from a partial power state P Si to a partial power state P Sj such that:
P Sj = P Si ∪ {(b, l)} with (b, l) ∈ B × L0 : b 6∈ B(P Si ) .
Note that by the definition of power state, the added couple (b, l) may not contain
a BS whose power is already fixed in a previous power state.
In the cluster phase, an ant moves from a partial cluster state CSi to a partial
cluster state CSj such that:
CSj = CSi ∪ {(t, b)} with {(t, b)} ∈ T × B ∪ T × B̄ .
Note that by the definition of cluster state, the added couple (t, b) may not contain a BS that has been already either assigned or not assigned to the same UT.
Moreover, note that the definitions of power and cluster state can be immediately traced back to a sequence of fixing of the decision variables, thus relaxing
the concept of state and reducing a move to the fixing of a decision variable j
after the fixing of another decision variable i, as it is done in [28].
Every move adds a single new element to the partial solution. Once that the
construction phases terminate, the value of the decision variables (z, y) is fully
established and, as previously noted, we can immediately derive the value of the
other variables (x, v), thus obtaining a complete feasible solution (x, z, y, v) for
the SPCAP.
The probability that an ant k moves from a power (cluster) state i to a more
complete power (cluster) state j, chosen among a set F of feasible power (cluster)
states, is defined by the improved formula of [28]:
pkij = P
α τij + (1 − α) ηij
,
f ∈F α τif + (1 − α) ηif
where α ∈ [0, 1] is the parameter establishing the relative importance of trail and
attractiveness. Of course, the probability of infeasible moves is set to zero. As
discussed in [28], the trail values τij and the attractiveness values ηij should be
provided by suitable lower bounds of the considered optimization problem. In our
particular case: 1) τij is derived from the values of the variables in the solution
associated with the bound PI-bound, provided by the linear relaxation of the
Power-Indexed formulation (PI-SPCAP) (see the next subsection for the specific
setting of τij ); 2) ηij is equal to the optimal solution of the linear relaxation of
the big-M formulation (big-M SPCAP) strengthened with the GCIs of (PI0 SPCAP) and including additional constraints to fix the value of the variables
fixed in the considered state j. We denote the latter bound by strongBM-bound
and we recall that this can be quickly computed and its computation becomes
faster and faster as we move towards a complete state (the number of fixed
variables indeed increases move after move).
A hybrid exact-ACO algorithm for Cooperative Wireless Network Design
11
As previously explained, once that an ant has finished its construction, we
have a vector (z, y) that can be used to derive the value of the other variables
(x, v) and define a complete feasible solution (x, z, y, v) for the SPCAP.
3.2
Daemon actions: Relaxation Induced Neighborhood Search
We refine the quality of the feasible solutions found through the ant-construction
phase by an exact local search in a large neighborhood, made for each feasible solution generated by the ants. Specifically, we adopt a modified relaxation induced
neighborhood search (RINS) (see [11] for a detailed discussion of the method).
The main steps of RINS are 1) defining a neighborhood by exploiting information
about some continuous relaxation of the discrete optimization problem, and 2)
exploring the neighborhood through a (Mixed) Integer Programming problem,
that is optimally solved through an effective commercial solver.
Let S ANT be a feasible solution to the SPCAP built by an ant and let S PI
be the optimal solution to the linear relaxation of (PI-SPCAP). Additionally,
let SjANT , SjPI denote the j-th component of the vectors. Our modified RINS
(mod-RINS) solves a sub-problem of the big-M formulation (big-M SPCAP)
strengthened with the GCIs of (PI0 -SPCAP) where:
1. we fix the variables whose value in S ANT and S PI differs of at most ǫ > 0,
(i.e., Sj = 0 if SjANT = 0 ∩ S PI ≤ ǫ, Sj = 1 if SjANT = 1 ∩ S PI ≥ 1 − ǫ;
2. set an objective cutoff based on the value of S ANT ;
3. impose a solution time limit of T ;
The time limit is set as the problem may be in general difficult to solve, so the
exploration of the feasible set may need to be truncated. Note that in point 1
we generalize the fixing rule of RINS, in which ǫ = 0.
3.3
Pheromone trail update
At the end of each construction phase t of the ants, the pheromone trails τij (t−1)
are updated according to the following improved formula (see [28] for a detailed
discussion of its elements):
τij (t) = τij (t − 1) +
m
X
k=1
τijk
with
τijk
k
− LB
zcurr
, (17)
= τij (0) · 1 −
z̄ − LB
where, to set the values τij (0) and LB, we solve the linear relaxation of (PISPCAP) and then we set τij (0) equal to the values of the corresponding optimal
decision variables and LB equal to the optimal value of the relaxation. Addik
tionally, zcurr
is the value of the solution built by ant k and z̄ is the moving
average of the values of the last ψ feasible solutions built. As noticed in [28],
formula (17) substitutes a very sensible parameter, the pheromone evaporation
factor, with the moving average ψ whose setting is much less critical.
12
Fabio D’Andreagiovanni
The overall structure of our original hybrid exact-ACO algorithm is presented
in Algorithm 1. The algorithm includes an outer loop repeated r times. At each
execution of the loop, an inner loop defines m ants to build the solutions. Once
that an ant finishes to build its solution, mod-RINS is applied in an attempt at
finding an improvement. Pheromone trail updates are done at the end of each
execution of the inner loop.
Algorithm 1. Hybrid exact-ACO for the SPCAP.
1. Compute the linear relaxation of (PI-SPCAP) and use it to initialize the
values τij (0).
2. FOR t := 1 TO r DO
(a) FOR µ := 1 TO m DO
i. build a complete power state;
ii. build a complete cluster state;
iii. derive a complete feasible solution to the SPCAP;
iv. apply mod-RINS to the feasible solution.
END FOR
(b) Update τij (t) according to (17).
END FOR
4
Computational experiments
We tested the performance of our hybrid algorithm on a set of 15 realistic instances of increasing size, defined in collaboration with the Technical Strategy
& Innovations Unit of British Telecom Italia (BT Italia). The experiments were
made on a machine with a 1.80 GHz Intel Core 2 Duo processor and 2 GB of
RAM and using the commercial solver IBM ILOG Cplex 11.1. All the instances
refer to a WiMAX Network [2] and lead to the definition of very large and hard
to solve (big-M SPCAP) formulations. Even when strengthened with the GCIs
of (PI0 -SPCAP), (big-M SPCAP) continues to constitute a very hard problem
and the simple identification of feasible solutions may be a hard task even for
Cplex. In particular, for most instances it was not possible to find feasible solutions within one hour of computations and, when solutions were found, they were
anyway of low value (up to the 35% of covered UTs). Our heuristic algorithm
was instead able to find good quality solutions.
After a series of preliminary tests, we found that a good setting of the parameters of the heuristic is: α = 0.5 (balance between attractiveness and trail level),
m = |B|/2 (number of ants equal to half the number of BSs), ψ = m = |B|/2
(width of the moving average equal to the number of ants), ǫ = 0.01 (tolerance
of fixing in mod-RINS), T = 10 minutes (time limit in mod-RINS). Moreover,
the construction loop was executed 50 times. In Table 4, for each instance we
report its ID and size and the number |T ∗ | of covered UTs in the best solution
found by mod-RINS (showing also the percentage coverage Cov%) and in the
corresponding ant solution. We also report the maximum size of a cluster in the
best solution. The solutions found by the hybrid algorithm have a much higher
A hybrid exact-ACO algorithm for Cooperative Wireless Network Design
13
value than those found by Cplex directly applied to (big-M SPCAP) and guarantee a good level of coverage ranging from 45 to 80%. Moreover, we note that
the execution of mod-RINS is able to increase the value of the ant solution from
5 to 13%. Finally, it is interesting to note that the size of the clusters keeps in
general low, presenting a maximum dimension of 5. We consider the overall performance highly satisfying, considering that our real aim is to use the solutions
generated by the algorithm to favour a warm start in an exact cutting-plane
algorithm applied to the Power-Indexed formulation (PI-SPCAP). Moreover, we
are confident that refinements of the components of the heuristic and further
tuning of the procedure can lead to the generation of solutions of higher quality.
Table 1. Experimental results
ID
I1
I2
I3
I4
I5
I6
I7
I8
I9
I10
I11
I12
I13
I14
I15
5
|T |
100
100
121
121
150
150
150
169
169
169
196
196
225
225
225
|B|
9
12
9
15
12
15
18
12
15
18
15
21
9
15
18
|T ∗ | (ACO)
55
62
52
72
94
96
103
80
103
121
136
140
125
142
152
|T ∗ | (ACO+RINS) Cov%
60
0.60
67
0.67
55
0.45
80
0.66
106
0.71
106
0.71
112
0.75
87
0.51
116
0.69
133
0.79
144
0.73
156
0.80
142
0.63
149
0.66
163
0.72
Max size cluster
3
2
2
4
3
3
3
2
4
2
3
5
3
3
4
Conclusions
Cooperative wireless networks have recently attracted a lot of attention, since
cooperation among base stations may lead to remarkable increases in the capacity of a network and enhance the service experience of the users. Though
base station cooperation has been extensively discussed from a theoretical and
technical point of view, there is still a lack of effective optimization models and
algorithms for its evaluation and implementation. To make a further step towards filling such gap, in this work we have presented a new model and solution
algorithm for the problem of designing a cooperative wireless network. In particular, we have generalized the classical model for wireless network design, in
order to include cluster definition and assignment. We have then showed how to
14
Fabio D’Andreagiovanni
strengthen the model, through the use of a special class of valid inequalities, the
GUB cover inequalities, that eliminate all the sources of numerical problems.
Finally, we have defined a hybrid heuristic based on the combination of ant
colony optimization and relaxation induced neighborhood search, that exploits
the important information provided by the relaxation of a strong formulation.
Computational experiments on a set of realistic instances showed that our heuristic can find solutions of good quality, which could be used for a warm start in an
exact branch-and-cut algorithm. Future work will consist in refining the components of the heuristic (for example, by better integrating the power and cluster
state moves) and in integrating the heuristic with a branch-and-cut algorithm,
in order to find solutions of higher value and whose quality is precisely assessed.
References
1. Amaldi, E., Capone, A., Malucelli, F., Mannino, C.: Optimization problems and
models for planning cellular networks. In: Resende, M., Pardalos, P. (eds.) Handbook
of Optimization in Telecommunication, pp. 917–939. Springer, Heidelberg (2006)
2. Andrews, J.G., Ghosh, A., Muhamed, R.: Fundamentals of WiMAX. Prentice Hall,
Upper Saddle River (2007)
3. Atzori, L., D’Andreagiovanni, F., Mannino, C., Onali, T.: An algorithm for routing
optimization in DiffServ-aware MPLS networks. DIS Tech. Rep. 4-2010, Sapienza
Università di Roma, Roma (2010)
4. Bley, A., D’Andreagiovanni, F., Hanemann, A.: Robustness in Communication Networks: Scenarios and Mathematical Approaches. In: Proc. of the ITG Symposium
on Photonic Networks 2011, pp. 1–8. VDE Verlag, Berlin (2011)
5. Blum., C., Puchinger, J., Raidl, G.R., Roli., A.: Hybrid metaheuristics in combinatorial optimization: A survey. Appl. Soft Comp. 11 (6), 4135–4151 (2011)
6. Büsing, C., D’Andreagiovanni, F.: New Results about Multi-band Uncertainty in
Robust Optimization. In: Klasing, R. (ed.) Experimental Algorithms - SEA 2012,
LNCS, vol. 7276, pp. 63-74. Springer, Heidelberg (2012)
7. Büsing, C., D’Andreagiovanni, F.: A new theoretical framework for Robust Optimization under multi-band uncertainty. To appear in Operations Research Proceedings 2012 - OR 2012. Springer, Heidelberg (2012)
8. Capone, A., Chen, L., Gualandi, S., Yuan, D.: A New Computational Approach
for Maximum Link Activation in Wireless Networks under the SINR Model. IEEE
Trans. Wireless Comm. 10 (5), 1368–1372 (2011)
9. Choi, Y.S., Kim, K.S., Kim, N.: The Displacement of Base Station in Mobile Communication with Genetic Approach. EURASIP J. Wireless Comm. Net. (2008)
doi:10.1155/2008/580761
10. Classen, G., Koster, A.M.C.A., Schmeink, A.: A robust optimisation model and
cutting planes for the planning of energy-efficient wireless networks. Computers &
OR 40 (1), 80–90 (2013)
11. Danna, E., Rothberg, E., Le Pape, C.: Exploring relaxation induced neighborhoods
to improve MIP solutions. Math. Program. 102, 71–90 (2005)
12. D’Andreagiovanni, F., Mannino, C., Sassano, A.: GUB Covers and PowerIndexed Formulations for Wireless Network Design. Management Sci. (2012) doi:
10.1287/mnsc.1120.1571
13. D’Andreagiovanni, F.: Pure 0-1 Programming approaches to Wireless Network
Design. Ph.D. Thesis. 4OR-Q. J. Oper. Res. (2011) doi: 10.1007/s10288-011-0162-z
A hybrid exact-ACO algorithm for Cooperative Wireless Network Design
15
14. D’Andreagiovanni, F.: On Improving the Capacity of Solving Large-scale Wireless
Network Design Problems by Genetic Algorithms. In: Di Chio, C., et al. (eds.)
EvoApplications 2011. LNCS, vol. 6625, pp. 11–20. Springer, Heidelberg (2011)
15. D’Andreagiovanni, F., Mannino, C., Sassano, A.: Negative Cycle Separation in
Wireless Network Design. In: Pahl, J., Reiners, T., Voss, T. (eds.) INOC 2011.
LNCS, vol. 6701, pp. 51–56. Springer, Heidelberg (2011)
16. D’Andreagiovanni, F., Mannino, C., Sassano, A.: A power-indexed formulation for
wireless network design. DIS Tech. Rep. 14-2009, Sapienza Università di Roma,
Roma (2009)
17. D’Andreagiovanni, F., Mannino, C.: A MILP formulation for WiMAX Network
Planning. DIS Tech. Rep. 02-2008, Sapienza Università di Roma, Roma (2008)
18. Dorigo, M., Blum, C.: Ant colony optimization theory: A survey. Theoret. Comp.
Sci. 344 (2-3), 243–278 (2005)
19. Dorigo, M., Di Caro, G., Gambardella, L.M.: Ant algorithms for discrete optimization. Artificial Life 5 (2), 137–172 (1999)
20. Dorigo, M., Maniezzo, V., Colorni, A.: Ant System: Optimization by a colony of
cooperating agents. IEEE Trans. Syst. Man Cybern. B 26 (1), 29–41 (1996)
21. Gesbert, D., Hanly, S., Huang, H., Shamai Shitz, S., Simeone, O., Wei Yu: MultiCell MIMO Cooperative Networks: A New Look at Interference. IEEE J. Sel. Areas
Comm. 28 (9), 1380–1408 (2010)
22. Giovanidis, A., D’Andreagiovanni, F., Krolikowski, J., Tanzil, V.H., Brueck, S.:
A 0-1 Program for Minimum Clustering in Downlink Base Station Cooperation.
ZIB-Report 11-19, Zuse-Institut Berlin (2011)
23. Hammer, P.L., Rudeanu, S.: Boolean methods in operations research and related
areas. Springer, Heidelberg (1968)
24. Hong, M., Sun, R., Baligh, H., Luo, Z.: Joint Base Station Clustering and Beamformer Design for Partial Coordinated Transmission in Heterogeneous Networks.
Technical Report, arXiv:1203.6390 (2012)
25. IBM ILOG Cplex, http://www-01.ibm.com/software/integration/optimization/cplex-optimizer.
26. Kennington, J., Olinick, E., Rajan, D: Wireless Network Design: Optimization
Models and Solution Procedures. Springer, Heidelberg (2010)
27. Koster, A.M.C.A., Helmberg, C., Bley, A., Grötschel, M., Bauschert, T.: BMBF
Project ROBUKOM: Robust Communication Networks. In: Proc. of the ITG Workshop “Visions of Future Generation Networks” - EuroView2012, pp. 1–2. VDE Verlag, Berlin (2012)
28. Maniezzo, V.: Exact and Approximate Nondeterministic Tree-Search Procedures
for the Quadratic Assignment Problem. INFORMS J. Comp. 11 (4), 358–369 (1999)
29. Mannino, C., Rossi, F., Smriglio, S.: The Network Packing Problem in Terrestrial
Broadcasting. Oper. Res. 54 (6), 611–626 (2006)
30. Montemanni, R., Smith, D.H., Allen, S.M.: An improved algorithm to determine
lower bounds for the fixed spectrum frequency assignment problem. Europ. J. Oper.
Res. 156 (3), 736–751 (2004)
31. Nawrocki, M., Aghvami, H., Dohler, M.: Understanding UMTS Radio Network
Modelling, Planning and Automated Optimisation: Theory and Practice. John Wiley & Sons, Hoboken (2006)
32. Nehmauser, G., Wolsey, L. Integer and Combinatorial Optimization. John Wiley
& Sons, Hoboken (1988)
33. Sesia, S., Toufik, I., Baker, M.: LTE-The UMTS Long Term Evolution: From Theory to Practice. John Wiley & Sons, Hoboken (2009)
16
Fabio D’Andreagiovanni
34. Weber, R., Garavaglia, A., Schulist, M., Brueck, S., Dekorsy, A.: Self-Organizing
Adaptive Clustering for Cooperative Multipoint Transmission. Proc. IEEE
VTC2011-Spring, 1–5 (2011)
35. Wolsey, L.: Valid inequalities for 0-1 knapsacks and mips with generalised upper
bound constraints. Discrete Appl. Math. 29 (2-3), 251–261 (1990)
| 8cs.DS
|
EXACT UPPER AND LOWER BOUNDS ON THE
MISCLASSIFICATION PROBABILITY
arXiv:1712.00812v4 [math.ST] 9 Feb 2018
IOSIF PINELIS
Abstract. Exact lower and upper bounds on the best possible misclassification probability for a finite number of classes are obtained in terms of the
total variation norms of the differences between the sub-distributions over the
classes. These bounds are compared with the exact bounds in terms of the
conditional entropy obtained by Feder and Merhav.
Contents
1. Introduction, summary and discussion
2. Proofs
References
1
11
15
1. Introduction, summary and discussion
Let X and Y be random variables (r.v.’s) defined on the same probability space
(Ω, F , P), X with values in a set S (endowed with a sigma-algebra Σ) and Y with
values in the set [k] := {1, . . . , k}, where k is a natural number; to avoid trivialities,
assume k > 2.
The sets Ω and [k] may be regarded, respectively, as the population of objects of
interest and the set of all possible classification labels for those objects. For each
“object” ω ∈ Ω, the corresponding values X(ω) ∈ S and Y (ω) ∈ [k] of the r.v.’s
X and Y may be interpreted as the (correct) description of ω and the (correct)
classification label for ω, respectively.
Alternatively, Y (ω) may be interpreted as the signal entered at the input side of
a device – with its possibly corrupted, output version X(ω).
The problem is to find a good or, better, optimal way to reconstruct, for each ω ∈
Ω, the correct label (or input signal) Y (ω) based on the description (or, respectively,
the output signal) X(ω). To solve this problem, one uses a measurable function
f : S → [k], referred to as a classification rule or, briefly, a classifier, which assigns
a label (or an input signal) f (x) ∈ [k] to each possible description (or, respectively,
to each possible output signal) x ∈ S. Then
pf := P(f (X) 6= Y )
is the misclassification probability for the classifier f .
2010 Mathematics Subject Classification. Primary: 60E15, 62H30, 91B06. Secondary: 26D15,
26D20, 62C10, 68T10.
Key words and phrases. Classification problem, misclassification probability, total variation
norm, Bayes estimators, maximum likelihood estimators.
1
2
IOSIF PINELIS
For each y ∈ [k], let µy be the sub-probability measure on Σ defined by the
condition
(1)
µy (B) := P(Y = y, X ∈ B)
for B ∈ Σ, so that
µ := µ1 + · · · + µk
(2)
is the probability measure that is the distribution of X in S, and let
dµy
ρy :=
,
dµ
the density of µy with respect to µ.
The value y ∈ [k] may be considered a parameter, so that the problem may be
be viewed as one of Bayesian estimation (of a discrete parameter, with values in
the finite set [k]). If the r.v. X is discrete as well, then of course
ρy (x) = P(Y = y|X = x)
for each x ∈ S with P(X = x) 6= 0. So, for each such x, the function y 7→ ρy (x)
may be referred to as the probability mass function of the posterior distribution of
the parameter corresponding to the observation x.
The following proposition is, essentially, a well-known fact of Bayesian estimation:
Proposition 1. For each x ∈ S, let f∗ (x) := min argmaxy ρy (x), where
argmaxy ρy (x) := {y ∈ [k] : ρy (x) = maxz∈[k] ρz (x)}; thus, f∗ (x) is the smallest
maximizer of ρy (x) in y ∈ [k]. Then the function f∗ is a classifier, and
Z
k
max ρy (x) µ(dx) 6 pf
p∗ := pf∗ = 1 −
S y=1
for any classifier f , so that p∗ is the smallest possible misclassification probability.
The proofs of all statements that may need a proof are deferred to Section 2.
Let
X
X
(3)
∆ :=
kµy − µz k =
|ρy − ρz |dµ,
16y<z6k
16y<z6k
where k · k is the total variation norm.
In the “population” model, the measure µy conveys two kinds of information:
kµy k
(i) the relative size kµk
= kµy k (of the set of all individual descriptions) of the yth
subpopulation (of the entire population Ω) consisting of the objects that carry the
µ
label y and (ii) the (conditional) probability distribution kµyy k of the object descriptions in this yth subpopulation, assuming the size kµy k of the yth subpopulation
is nonzero. Everywhere here, y and z are in the set [k]. Thus, ∆ is a summary
characteristic of the pairwise differences between the k subpopulations, which takes
into account both of the two just mentioned kinds of information.
In the input-output model, the kµy k’s are interpreted as the prior probabilities
of the possible input signals y ∈ [k] – whereas, for each y ∈ [k], the (conditional)
µ
probability distribution kµyy k is the distribution of the output signal corresponding
to the given input y. Thus, here ∆ is a summary characteristic of the pairwise
differences between the k sets of possible outputs corresponding to the k possible
inputs.
EXACT UPPER AND LOWER BOUNDS ON THE MISCLASSIFICATION PROBABILITY
3
Remark 2. By (3),
X
06∆6
(kµy k + kµz k) = (k − 1)
16y<z6k
k
X
kµy k = k − 1.
1
Moreover, the extreme values 0 and k − 1 of ∆ are attained, respectively, when
the measures µy are the same for all y ∈ [k] and when these measures are pairwise
mutually singular.
The main result of this paper provides the following upper and lower bounds on
the smallest possible misclassification probability p∗ in terms of ∆:
Theorem 3. One has
(4)
L(∆) 6 p∗ 6 U (∆) 6 Usimpl (∆),
where
1+∆
,
k
k + 1 + ∆ − 2d∆e
U (∆) := Uk (∆) := 1 −
,
(5)
(k − d∆e)(k + 1 − d∆e)
1
,
Usimpl (∆) := Uk;simpl (∆) := 1 −
k−∆
and d·e is the ceiling function, so that d∆e is the smallest integer that is no less
than ∆.
L(∆) := Lk (∆) := 1 −
Theorem 3 is complemented by
Proposition 4. For each possible value of ∆ in the interval [0, k − 1], the lower
and upper bounds L(∆) and U (∆) on p∗ are exact: For each ∆ ∈ [0, k − 1], there
are r.v.’s X and Y as described in the beginning of this paper for which one has
the equality p∗ = L(∆); similarly, with U (∆) in place of L(∆). More specifically,
the first (respectively, second) inequality in (4) turns into the equality if and only
if there is a set S0 ∈ Σ such that µ(S0 ) = 0 and for each x ∈ S \ S0 the values
ρ1 (x), . . . , ρk (x) constitute a permutation of numbers a1 , . . . , ak as in (17) (respectively, in (18)) with d = ∆. The simpler/simplified upper bound Usimpl (∆) is exact
only for the integral values of ∆.
Remark 5. In view of Remark 2, the functions L, U , and Usimpl , introduced in
Theorem 3, are well defined on the interval [0, k − 1]. Moreover, U (∆) is the linear
interpolation of Usimpl (∆) over the possible integral values 0, . . . , k − 1 of ∆. Thus,
each of the functions L, U , and Usimpl is concave and strictly decreasing (from 1 − k1
to 0) on the interval [0; k − 1]; moreover, the function L is obviously affine. We
see that, the greater is the characteristic ∆ of the pairwise differences between the
k subpopulations, the smaller are the lower and upper bounds L(∆), U (∆), and
Usimpl (∆) on the misclassification probability p∗ . Of course, this quite corresponds
to what should be expected of good bounds on p∗ . It also follows that one always
has
1
(6)
0 6 p∗ 6 1 − ,
k
and the extreme values 0 and 1 − k1 of the misclassification probability p∗ are
attained when, respectively, ∆ = k − 1 and ∆ = 0. The bounds L, U , and Usimpl
are illustrated in Figure 1.
4
IOSIF PINELIS
0.8
0.6
0.4
0.2
1
2
3
4
Figure 1. Graphs of the bounds L (green), U (green), and Usimpl
(dark green) for k = 5.
Feder and Merhav [3] obtained the following exact upper and lower bounds of
the optimal misclassification probability in terms of the conditional entropy H:
LFM (H) 6 p∗ 6 UFM (H),
where
(7)
H := H(Y |X) := − E
k
X
Z
ρy (X) ln ρy (X) = −
S
y=1
(8)
LFM (H) := Φ−1 (H),
µ(dx)
Φ(p) := p ln(k−1)+h2 (p),
k
X
ρy (x) ln ρy (x),
y=1
h2 (p) := −p ln p−(1−p) ln(1−p)
for p ∈ (0, 1), h2 (0) := 0, h2 (1) := 0,
(9)
UFM (H) :=
1
H − ln e(H)
e(H) − 1
+
,
e(H)
e(H)(e(H) + 1) ln(1 + 1/e(H))
and
(10)
e(H) := deH e − 1.
Note that Φ(p) strictly and continuously increases from 0 to ln k as p increases
from 0 to 1 − k1 . Therefore and because all the values of the conditional entropy
H lie between 0 and ln k, the expression Φ−1 (H) is well defined, and its values lie
between 0 and 1 − k1 – which is in accordance with (6).
Throughout this paper, we use only natural, base-e logarithms. In [3], the bounds
are stated in terms of binary, base-2 logarithms. To rewrite LFM (H) and UFM (H)
in terms of binary logarithms, replace all the instances of ln = loge in (7)–(9) by
log2 and, respectively, replace eH in (10) by 2H . An advantage of using natural
logarithms is that then the expressions for the corresponding derivatives, used in
our proofs, are a bit simpler; also, ln is a bit shorter in writing than log2 or even
log.
Note also that, in the notation in [3], the roles of X and Y are reversed: there,
X denotes the input and Y the output. Our notation in this paper is in accordance
with the standard convention in machine learning; cf. e.g. [5, 4].
Let us compare, in detail, our “∆-bounds” L(∆), U (∆), and Usimpl (∆) with the
“H-bounds” LFM (H) and UFM (H). We shall be making the comparisons only in
EXACT UPPER AND LOWER BOUNDS ON THE MISCLASSIFICATION PROBABILITY
5
the “pure” settings, when the set {ρ1 (x), . . . , ρk (x)} is the same for all x ∈ S, that
is, when for each x ∈ S the k-tuple (ρ1 (x), . . . , ρk (x)) is a permutation of one and
the same k-tuple (a1 , . . . , ak ) (of nonnegative real numbers a1 , . . . , ak such that
a1 + · · · + ak = 1). A reason for doing so is that one may expect the comparisons
to be of greater contrast in the “pure” settings than in “mixed”, non-“pure” ones.
Thus, focusing on “pure” settings will likely allow us to see the differences between
the “∆-bounds” and the “H-bounds” more clearly, while taking less time and effort.
We shall see that, even though the “H-bounds” LFM (H) and UFM (H) and the
“∆-bounds” L(∆) and U (∆) are exact in terms of H and ∆, respectively, they have
rather different properties.
Remark 6. Typically, the lower H-bound LFM (H) on p∗ appears to be better (that
is, larger) than the lower ∆-bound L(∆), whereas the upper H-bound UFM (H) on
p∗ appears to be worse (that is, larger) than the upper ∆-bound U (∆) and even
its simplified but less accurate version Usimpl (∆).
However, in some rather exceptional cases these relations are reversed.
In particular, if the best possible misclassification probability p∗ is large enough,
then the lower ∆-bound L(∆) may be better than the lower H-bound LFM (H), for
each k > 3.
On the other hand, if k is large enough and p∗ is small enough, then the upper
∆-bound U (∆) may be worse than the upper H-bound UFM (H). However, I have
not been able to find cases with U (∆) (or even Usimpl (∆)) worse than UFM (H) when
there are at most k = 9 classes.
More specifically, we have the following propositions. (As usual, I{·} will denote
the indicator function.)
Proposition 7. Suppose that k > 3 and for each x ∈ S the vector (ρ1 (x), . . . , ρk (x))
is a permutation of the vector (a1 , . . . , ak ), where
1
ai = I{1 6 i 6 `}
`
for some natural ` > 2 in the set {k − 3, k − 2, k − 1} and for all i = 1, . . . , k; one
may also allow ` = k − 4 if k ∈ {6, 7, 8, 9}. Then L(∆) > LFM (H).
Proposition 8. Fix any ν ∈ (1, ∞). Suppose that for each x ∈ S the vector
(ρ1 (x), . . . , ρk (x)) is a permutation of the vector (a1 , . . . , ak ), where k > ν and
ν − 1
ν−1
ai = 1 −
I{i = 1} +
I{2 6 i 6 k}
k
k(k − 1)
for all i = 1, . . . , k. Then U (∆) > UFM (H) for all large enough k (depending on
the value of ν).
Note that in Proposition 7 the best possible misclassification probability p∗ =
1 − 1` is large, especially when ` is large (and hence so is k). In contrast, in Proposition 8 p∗ = ν−1
k is small for the large values of k, assumed in that proposition.
Either of these two kinds of situations, especially the second one, may be considered somewhat atypical: it usually should be difficult to make the misclassification
probability p∗ small when the number k of possible classes is large; on the other
hand, when k is not very large, one may hope that the best possible misclassification
probability is small enough.
Concerning the case of two classes, we have
6
IOSIF PINELIS
Proposition 9. Suppose that k = 2. Then U (∆) = L(∆) = LFM (H) = p∗ for all
pairs of r.v.’s (X, Y ). So, one can say that the bounds U (∆), L(∆), and LFM (H)
always perfectly estimate the best possible misclassification probability p∗ – if k = 2.
On the other hand, here UFM (H) > Usimpl (∆) > p∗ unless there is a set S0 ∈ Σ
such that µ(S0 ) = 0 and for each x ∈ S \ S0 either ρ1 (x) = ρ2 (x) = 1/2 or
{ρ1 (x), ρ2 (x)} = {0, 1} – that is, the values ρ1 (x) and ρ2 (x) constitute a permutation of the numbers 0 and 1. Thus, in the case k = 2, with the mentioned trivial
exceptions, even the simplified upper ∆-bound Usimpl (∆) on p∗ is strictly better than
the upper H-bound UFM (H), but still Usimpl (∆) is not a perfect estimate of p∗ .
An important case is that of three classes, so that k = 3. Here, in the “pure”
setting, for each x ∈ S the triple (ρ1 (x), ρ2 (x), ρ2 (x)) is a permutation of the
triple (1 − p, p − ε, ε), where p := p∗ ∈ [0, 1 − 1/3] and 1 − p > p − ε > ε > 0
or, equivalently, p ∈ [0, 2/3] and (2p − 1)+ 6 ε 6 p/2, where u+ := max(0, u).
Each of the 6 pictures in Figure 2 presents the graphs of the decimal logarithms
of the bounds L(∆), U (∆), Usimpl (∆), LFM (H), and UFM (H) as functions of ε ∈
[(2p − 1)+ , p/2] with the misclassification probability p = p∗ taking a fixed value
in the set {0.01, 0.1, 0.3, 0.5, 0.6, 0.64}. We see that in all these cases the upper ∆bound U (∆) and even its simplified (but worse) version Usimpl (∆) are better than
the upper H-bound UFM (H), over the entire range of values of ε. For small values
of the best possible misclassification probability p∗ , the lower H-bound LFM (H)
is significantly better than L(∆) over all values of ε; however, this comparison is
reversed if p∗ is large enough but ε is small enough (especially in the case p∗ = 0.5).
An interesting series of cases is given by what may be called the binomial model
(with a parameter q ∈ (0, 1)), in which k = 2m for a natural m, and for each x ∈ S
the vector (ρ1 (x), . . . , ρk (x)) is a permutation of a vector (a1 , . . . , ak ), where each
ai is of the form (1 − q)j q m−j for some j∈ {0, . . . , m}, and the multiplicity of the
form (1 − q)j q m−j among the ai ’s is m
for each j ∈ {0, . . . , m}. Clearly then,
j
Pm
j m−j
= 1. In
all the ai ’s are nonnegative, and a1 + · · · + ak = j=0 m
j (1 − q) q
particular, for m = 1 we have k = 2, and then we may take (a1 , a2 ) = (1 − q, q).
For m = 2 we have k = 4, and then we may take
(a1 , a2 , a3 , a4 ) = (1 − q)2 , (1 − q)q, (1 − q)q, q 2 .
p
Choosing, in the latter case, S = {1, 2, 3, 4} and q = Q( 2Eb /N0 ), where Q(x) :=
R ∞ 1 −u2 /2
√
e
du is the tail probability for the standard normal distribution, Eb is
x
2π
the energy per bit, and N0 /2 is the noise power spectral density (PSD), we see that
the resulting particular case of the binomial model covers the so-called quadrature
phase-shift keying (QPSK) digital communication scheme over an additive white
Gaussian noise (AWGN) channel (cf. e.g. [2, page 313]), which in fact provided the
motivation for the general binomial model.
Another interesting series of cases is given by what may be called the exponential model (with a parameter q ∈ (0, 1)), in which for each x ∈ S the
vector (ρ1 (x), . . . , ρk (x)) is a permutation of a vector (a1 , . . . , ak ), where ai :=
Pk
(1 − q)i−1 q k−i /cq and cq := ck,q := i=1 (1 − q)i−1 q k−i , so that all the ai ’s are nonnegative and a1 + · · · + ak = 1. Informally, the exponential model can be obtained
from the binomial one by removing the multiplicities.
Graphical comparisons of the “∆-bounds” with the “H-bounds” (as functions of
the parameter q) for the cases k = 2, 4, 8 of the binomial and exponential models
EXACT UPPER AND LOWER BOUNDS ON THE MISCLASSIFICATION PROBABILITY
7
are presented in Figure 3. Note here that, by symmetry, it is enough to consider
q ∈ (0, 1/2]. Obviously, for k = 2 the binomial and exponential models are the
same, and in this case they are the same as the essentially unique general “pure”
model for k = 2, fully considered in Proposition 9. Accordingly, the pictures in
the first row in Figure 3 are identical to each other, and the graphs of the bounds
U (∆), L(∆), and LFM (H) are the same as that of p∗ . The cases k = 4, 8 in Figure 3
illustrate the first sentence in Remark 6. It appears that the comparisons in the
exponential model are somewhat more favorable to the ∆-bounds than they are in
the binomial model.
8
IOSIF PINELIS
- 0.2
- 1.4
- 0.3
- 1.6
- 0.4
- 1.8
- 0.5
- 2.0
- 0.6
0.001
0.002
0.003
0.004
0.005
0.05
0.10
0.15
0.20
0.25
0.22
0.24
0.26
0.28
0.30
- 0.18
- 0.6
- 0.7
- 0.20
- 0.8
- 0.22
- 0.9
- 0.24
- 1.0
- 0.26
- 1.1
0.01
0.02
0.03
0.04
0.05
- 0.3
- 0.18
- 0.4
- 0.19
- 0.5
- 0.20
- 0.6
- 0.21
0.02 0.04 0.06 0.08 0.10 0.12 0.14
0.29
0.30
0.31
0.32
Figure 2. Graphs of log10 L(∆) (green), log10 U (∆) (green),
log10 Usimpl (∆) (dark green), log10 LFM (H) (blue), log10 UFM (H)
(blue), and log10 p∗ (dashed) for k = 3 and p∗ ∈ {0.01, 0.1, 0.3,
0.5, 0.6, 0.64}.
EXACT UPPER AND LOWER BOUNDS ON THE MISCLASSIFICATION PROBABILITY
- 0.5
- 0.5
- 1.0
- 1.0
- 1.5
- 1.5
0.1
0.2
0.3
0.4
0.5
- 0.2
- 0.4
- 0.6
- 0.8
- 1.0
- 1.2
- 0.4
- 0.6
0.3
0.4
0.5
0.1
0.2
0.3
0.4
0.5
0.1
0.2
0.3
0.4
0.5
- 1.0
- 1.5
0.1
0.2
0.3
0.2
0.3
0.4
0.4
0.5
0.5
- 2.0
- 0.5
- 1.0
- 0.8
- 1.5
- 1.0
- 2.0
- 1.2
0.2
- 0.5
0.1
- 0.2
0.1
9
- 2.5
Figure 3. Graphs of log10 L(∆) (green), log10 U (∆) (green),
log10 Usimpl (∆) (dark green), log10 LFM (H) (blue), log10 UFM (H)
(blue), and log10 p∗ (dashed) for k = 2, 4, 8. Left column: Binomial
model. Right column: Exponential model.
Upper and lower bounds on the best possible misclassification probability in
terms of Renyi’s conditional entropy were announced in [1], where the input and
10
IOSIF PINELIS
output r.v.’s were denoted by M and W , respectively, and both M and W were
assumed to take values in the same set {1, . . . , k}. Renyi’s conditional entropy used
in [1] was defined by the formula
(11)
Hβ (W |M ) :=
k
X
P(M = m)Hβ (W |m),
m=1
where
k
(12)
X
1
Hβ (W |m) :=
log
P(W = w|M = m)β ,
1−β
w=1
log := log2 ,
and β ∈ (0, 1) ∪ (1, ∞), so that Shannon’s conditional entropy
(13)
H1 (W |M ) := H(W |M ) = −
k
X
P(W = w) log P(W = w|M = m)
m=1
may be viewed as a limit case of Renyi’s: Hβ (W |M ) → H(W |M ) as β → 1.
Note that Hβ (W |M ) is the conditional entropy of the output W given the input
M . Thus, for some reason, the standard roles of the input and output r.v.’s were
reversed in [1]; cf. e.g. the conditional entropy of the input given the output used
in [3, formula (3)].
The mentioned upper and lower bounds announced in [1] were given by inequalities of the form
Hβ (W |M ) − HS (e)
Hα (W |M ) − HS (e)
6 P (e) 6
,
(14)
N1
N2
where β < 1 6 α, N1 , N2 are some positive expressions,
HS (e) := −P (e) log P (e) − (1 − P (e)) log(1 − P (e)),
and P (e) is the “the classification error probability”, which doers not seem to be
explicitly defined in [1]. A proof of these bounds was offered later in [2, Appendix],
from which it appears (see [2, formula (A.1)]) that P (e) is understood as P(W 6=
M ). However, the proof in [2] of the upper bound on P (e) in (14) appears to be
mistaken, and the upper bound itself may be negative and thus false in general.
Indeed, the second inequality in (14) appears to be obtained in [2] by multiplying
the expressions in [2, formula (A.7)] by p(m
P k ), then summing in k (in the notations
there), and finally using the inequality k p(mk )HS (e|mk )[= HS (e|M )] > HS (e).
However, in general the reverse inequality is true: HS (e|M ) 6 HS (e); cf. even the
derivation of (A.6) from (A.5) in [2]. Also, the proof in [2] does not use the condition
that P (e) is the smallest possible classification error probability, and without such
a condition no reasonable upper bound on P (e) is possible.
More importantly, as mentioned above, the upper bound on P (e) in (14) is false
in general. For a very simple counterexample, suppose that k = 2, P(W = 1, M =
1) = P(W = 1, M = 2) = 1/2, and P(W = 2, M = 1) = P(W = 2, M = 2) = 0.
Then P (e) = 1/2, Hβ (W |M ) = 0 for all β, and HS (e) = 1, so that the presumed
upper bound on P (e) in (14) is negative and thus false.
Another two pairs of upper and lower bounds were announced in [1], in terms of
the joint Renyi’s joint entropy Hβ (W, M ) of (W, M ) and Renyi’s mutual information
Iβ (W, M ) between W and M , rather than Renyi’s conditional entropy, with no
apparent proofs for these additional bounds. However, the same simple example
EXACT UPPER AND LOWER BOUNDS ON THE MISCLASSIFICATION PROBABILITY 11
given above will quite similarly show that these additional upper bounds are false
in general, too.
As stated in the abstract in [2], the mentioned bounds in [1] on P (e) “were
practically incomputable”, because P (e) itself appears in those bounds. Therefore,
an effort was made in [2] to modify the bounds in [1] – by making the upper bounds
greater and the lower bounds smaller – to make them computable. However, in view
of what has been said, the modified upper bounds in [2] remain without a valid
proof. As for the modified lower bounds, it is stated in [2, page 313] that, in the
examples considered there, “the modified lower bounds are not depicted because
they turn out to be negative”.
For all these reasons, we shall not attempt to compare our bounds with ones in
[1, 2].
2. Proofs
Proof of Proposition 1. Clearly, f∗ is a map from S to [k]. Also, f∗ is measurable,
Sy−1
Tk
since f∗−1 ({y}) = By \ z=1 Bz ∈ Σ for each y ∈ [k], where By := z=1 By,z and
By,z := {x ∈ S : ρy (x) > ρz (x)} ∈ Σ. Thus, f∗ is a classifier. Moreover, for any
classifier f ,
1 − pf = P(f (X) = Y ) =
k
X
P(Y = y, f (X) = y)
y=1
=
k Z
X
y=1
=
k Z
X
y=1
=
I{f (x) = y} µy (dx)
S
I{f (x) = y} ρy (x) µ(dx)
S
Z X
k
I{f (x) = y} ρy (x) µ(dx)
S y=1
Z
6
k
max ρy (x) µ(dx) = 1 − pf∗ .
S y=1
This completes the proof of Proposition 1.
In view of Proposition 1 and (3), Theorem 3 and Proposition 4 follow immediately
by the lemma below, with ρi (x) in place of ai .
Lemma 10. Suppose that
(15)
a1 , . . . , ak are nonnegative real numbers such that
k
X
1
Then
(16)
k
L(δ) 6 1 − max ai 6 U (δ) 6 Usimpl (δ),
1
where
δ :=
X
|ai − aj |
16i<j6k
and the functions L, U , and Usimpl are defined as in Theorem 3.
ai = 1.
12
IOSIF PINELIS
Under the stated conditions on the ai ’s, one always has 0 6 δ 6 k − 1; cf.
Remark 2.
k
The bounds L(δ) and U (δ) on 1 − max ai are exact for each possible value of δ:
1
(i) For each d ∈ [0, k − 1], if
(17)
1+d
k
1
d
−
,
k k(k − 1)
k
then condition (15) holds, δ = d, max ai = a1 , and the first inequality in (16)
a1 =
and
a2 = · · · = ak =
1
turns into the equality. If the ai ’s satisfy condition (15) but do not constitute
a permutation of the ai ’s as in (17) with d = δ, then the first inequality in
(16) is strict.
(ii) For each d ∈ [0, k − 1], if
(18)
dde − d
I{i = k + 1 − dde}
k + 1 − dde
k
for all i ∈ [k], then condition (15) holds, δ = d, max ai = a1 , and the second
ai = (1 − U (d)) I{i 6 k − dde} +
1
inequality in (16) turns into the equality. If the ai ’s satisfy condition (15) but
do not constitute a permutation of the ai ’s as in (18) with d = δ, then the
second inequality in (16) is strict.
Proof. It is quite easy to see that Usimpl (d) is concave in d ∈ [0, k − 1]. Moreover, as
noted in Remark 5, U (d) is the linear interpolation of Usimpl (d) over d = 0, . . . , k −1.
Thus, we have the last inequality in (16).
k
It remains to establish the lower bound L(δ) and upper bound U (δ) on 1 − max
1
and to show that these bounds are attained, with δ = d, if and only if the ai ’s are
as in (17) and (18), respectively.
By symmetry, without loss of generality (w.l.o.g.) a1 > · · · > ak . Then, letting
hi := ai − ai+1 for i ∈ [k] (with ak+1 := 0), we have
h1 > 0, . . . , hk > 0,
k
X
k
max ai = a1 =
1
δ=
X
16i<j6k
(ai − aj ) =
X
hi ,
1
j−1
X
16i<j6k q=i
hq =
k−1
X
X
hq
q=1
=
X
k−1
X
hq q(k − q) =
q=1
1=
k
X
1
aj =
k X
k
X
hi =
k
X
j=1 i=j
Take now indeed any d ∈ [0, k − 1]. Introducing
pi := ihi
1
16i6q q+16j6k
i=1
k
X
i=1
ihi .
i(k − i)hi ,
EXACT UPPER AND LOWER BOUNDS ON THE MISCLASSIFICATION PROBABILITY 13
for i ∈ [k], we further restate the conditions on the ai ’s (with δ equal the prescribed
value d ∈ [0, k − 1], as desired):
(19)
p1 > 0, . . . , pk > 0,
k
X
pi = 1,
i=1
k
X
(20)
k
X
(k − i)pi = d or, equivalently,
i=1
ipi = k − d,
i=1
and
k
max ai = a1 =
k
X
1
g(i)pi ,
1
where g(i) := 1i ; here and in the rest of the proof of Lemma 10, i is an arbitrary
number in the set [k].
Introduce also
g U (i) := g(k − m − 1) + [g(k − m) − g(k − m − 1)][i − (k − m − 1)]
and
pU
i := (d − m) I{i = k − m − 1} + (m + 1 − d) I{i = k − m} ,
where
m := dde − 1;
here and in the rest of the proof of Lemma 10, i is an arbitrary number in the set
[k]. One may note at this point that m ∈ {0, . . . , k − 2}. Note that the function
g is strictly convex on the set [k], the function g U is affine, g U = g on the set
{k − m − 1, k − m}, and hence g > g U on [k] \ {k − m − 1, k − m}. Moreover,
conditions (19) and (20) hold with pU
i in place of pi . So,
(21)
k
X
g(i)pi >
k
X
1
1
g U (i)pi = g U
k
X
ipi
1
= g U (k − d) = g U
k
X
1
ipU
i
=
k
X
g U (i)pU
i =
1
k
X
g(i)pU
i ;
1
the inequality here holds because g > g U ; the first and fourth equalities follow
because the function g U is affine; the second and third equalities hold because
of the condition (20) for the pi ’s and pU
i ’s; and the last equality follows because
g U (i) = g(i) for i in the set {k − m − 1, k − m}, whereas pU
i = 0 for i not in this
k
set. We conclude that, under conditions (19) and (20), max ai is minimized – or,
1
k
equivalently, 1 − max ai is maximized – if and only if pi = pU
i for all i; that is, if
1
and only if hi = pU
/i or all i; that is, if and only if the ai ’s – related to the pi ’s by
iP
k
the formula ai = j=i 1i pi – are as in (18). This concludes the proof of the part
of Lemma 10 concerning the upper bound U (·).
The proof of the part of Lemma 10 concerning the lower bound L(·) is similar
and even easier. Here let
i−1
g L (i) := g(1) + [g(k) − g(1)]
k−1
14
IOSIF PINELIS
and
d
d
I{i = k} .
I{i = 1} + 1 −
k−1
k−1
Recall that the function g is strictly convex on the set [k]. Note that the function
g L is affine, g L = g on the set {1, k}, and g L > g on the set [k] \ {1, k}, so that
g 6 g L on [k]. Moreover, conditions (19) and (20) hold with pL
i in place of pi . So,
pL
i :=
k
X
g(i)pi 6
1
k
X
g L (i)pi = g L
k
X
1
ipi
1
L
= g (k − d) = g
L
k
X
ipL
i
=
k
X
1
g
L
1
(i)pL
i
=
k
X
g(i)pL
i ;
1
k
cf. (21). So, under conditions (19) and (20), max ai is maximized – or, equivalently,
1
k
1 − max ai is minimized – if and only if pi = pL
i for all i; that is, if and only if
1
hi = pL
i /i or all i; that is, if and only if the ai ’s are as in (17). This concludes
the proof of the part of Lemma 10 concerning the upper bound L(·) and hence the
entire proof of the lemma.
Proof of Proposition 7. We have to show that, under the conditions of this proposition, L(∆) > LFM (H). Recalling (8) and the fact that the function Φ is increasing,
we can rewrite inequality L(∆) > LFM (H) as Φ(L(∆)) > H. In view of (7), (3),
and (5), H = ln `, ∆ = k − `, and L(∆) = `−1
k . So, inequality Φ(L(∆)) > H can
be in turn rewritten as
` − 1
(?)
(22)
d(`) := dk (`) := Φ
− ln ` > 0
k
for k and ` as in the conditions in Proposition 7. For k ∈ {6, 7, 8, 9} and ` = k − 4,
as well as for k ∈ {3, 4, 5} and ` > 2 in the set {k − 3, k − 2, k − 1}, inequality
(22) can be verified by direct calculations. So, without loss of generality k > 6 and
` ∈ [k − 3, k), where we may allow ` to take non-integral values as well. Note that
d00 (`)(` − 1)`2 (k + 1 − `) = −1 + 2` − 2`2 + (` − 1)k
6 −1 + 2` − 2`2 + (` − 1)(` + 3) = −(` − 2)2 < 0
for ` > 2. Hence, d(`) is strictly concave in ` > 2 such that ` ∈ [k − 3, k]. Also,
d(k) = 0. So, to complete the proof of Proposition 7, it suffices to show that
(?)
˜ := k dk (k − 3) = (k − 4) ln k(k−1) + 4 ln k − k ln(k − 3) > 0
(23)
d(k)
k−4
4
for k > 6. We find that
d˜00 (k) = −
36
<0
(k − 4)(k − 3)2 (k − 1)2 k
˜
˜ = 0.446 · · · > 0
for k > 6, and so, d(k)
is strictly concave in k > 6. Moreover, d(6)
˜
and d(k) → 6 − 8 ln 2 = 0.454 · · · > 0 as k → ∞. Thus, inequality (23) indeed holds
for k > 6, and the proof of Proposition 7 is now complete.
Proof of Proposition 8. Under the conditions of this proposition,
ν − 1
ν − 1 ν − 1
ν−1
ln
→0
H =− 1−
ln 1 −
−
k
k
k
k(k − 1)
EXACT UPPER AND LOWER BOUNDS ON THE MISCLASSIFICATION PROBABILITY 15
and hence, by (9) and (10), UFM (H) → 0 as k → ∞.
On the other hand, here ∆ = k − ν. Therefore and because U (∆) is decreasing
in ∆,
1
1
U (∆) > U (d∆e) = Usimpl (d∆e) = 1 −
=1−
,
k − d∆e
bνc
which latter is a positive constant with respect to k and hence does not go to 0 as
k → ∞. Thus, the conclusion of Proposition 8 follows.
Proof of Proposition 9. Here w.l.o.g. {ρ1 (x), ρ2 (x)} = {1 − p, p} for each x ∈ S,
where p := p∗ ∈ [0, 1/2]. Then the equalities U (∆) = L(∆) = LFM (H) = p∗ follow
immediately from the definitions.
It remains to show that UFM (H) > Usimpl (∆) > p for p ∈ (0, 1/2). The second
2p
. To verify that
inequality here is obvious, since in this case Usimpl (∆) = 1+2p
UFM (H) > Usimpl (∆) for p ∈ (0, 1/2), consider d(p) := UFM (H) − Usimpl (∆) =
2p
− 21 (1 − p) log2 (1 − p) − 12 p log2 p − 1+2p
. It is easy to see that
d00 (p)(1 − p)p(1 + 2p)3 ln 4 = −1 + p(16 ln 2 − 6) − p2 (12 + 16 ln 2) − 8p3 < 0
for p ∈ (0, 1/2), so that d is strictly concave on (0, 1/2). Also, d(0+) = d(1/2) = 0.
So, d > 0 on (0, 1/2), which completes the proof of Proposition 9.
In conclusion, let us mention a sample of other related results found in the
literature. In [6], for k = 2, sharp lower bounds on the misclassification probabilities
for three particular classifiers in terms of characteristics generalizing the Kullback–
Leibler divergence and the Hellinger distance we obtained. Lower and upper bounds
on the misclassification probability based on Renyi’s information were given in [2].
Upper and lower bounds on the risk of an empirical risk minimizer for k = 2 were
obtained in [5] and [4], respectively.
References
[1] D. Erdogmus and J. Principe. Information transfer through classifiers and its relation to probability of error. In Proceedings. IJCNN ’01. International Joint Conference on Neural Networks
2001, (Washington, DC, 2001), pages 50–54, 2001.
[2] D. Erdogmus and J. C. Principe. Lower and upper bounds for misclassification probability
based on Renyi’s information. Journal of VLSI signal processing systems for signal, image
and video technology, 37:305–317, 2004.
[3] M. Feder and N. Merhav. Relations between entropy and error probability. IEEE Transactions
on Information Theory, 40(1):259–266, 1994.
[4] A. Kontorovich and I. Pinelis. Exact lower bounds for the agnostic probably-approximatelycorrect (PAC) machine learning model. arXiv:1606.08920 [cs.LG], 2016.
[5] P. Massart and E. Nédélec. Risk bounds for statistical learning. Ann. Statist., 34(5):2326–2366,
2006.
[6] G. T. Toussaint. On the divergence between two distributions and the probability of misclassification of several decision rules. In Proceedings of the Second International Joint Conference
on Pattern Recognition, pages 27–34, 1974.
Department of Mathematical Sciences, Michigan Technological University, Houghton, Michigan 49931, USA, E-mail: [email protected]
| 10math.ST
|
Where the Rubber Meets the Sky:
Bridging the Gap between Databases and Science
Jim Gray
Alex Szalay
October 2004
Technical Report
MSR-TR-2004-110
Microsoft Research
Microsoft Corporation
One Microsoft Way
Redmond, WA 98052
1
Where the Rubber Meets the Sky: Bridging the Gap between Databases and Science
Jim Gray
Microsoft Research
[email protected]
Alex Szalay
Johns Hopkins University
[email protected]
Abstract: Scientists in all domains face a data avalanche
– both from better instruments and from improved
simulations. We believe that computer science tools and
computer scientists are in a position to help all the
sciences by building tools and developing techniques to
manage, analyze, and visualize peta-scale scientific
information. This article is summarizes our experiences
over the last seven years trying to bridge the gap between
database technology and the needs of the astronomy
community in building the World-Wide Telescope.
scientists are adding to the data avalanche. Now we are
seeing the emergence of X-info departments to deal with
the avalanche (e.g., bio-informatics). The X-info
community is trying to put all the data and literature on the
public internet, federate it all, and provide tools that make
it easy for domain scientists to find information, and make
it easy for scientists to add their data and findings to the
corpus. This will redefine the notion of “publishing” and
has deep implications for how the world’s scientific
enterprise is organized.
1. Introduction
The vision is eScience in which the unified experimental,
theoretical, and simulation data and literature are at your
fingertips. You can explore all the world’s scientific data
and literature looking for patterns and anomalies. Tools
that encapsulate the best statistical algorithms, the best
machine learning algorithms and the best data visualization
are at your disposal to make you more productive and a
better scientist. When you want to publish your findings,
all your work has been recorded, so that others can follow
your arguments and extend your work. That’s the vision;
but, the reality is far from that vision.
If you are reading this you are probably a “database
person”, and have wondered why our “stuff” is widely
used to manage information in commerce and government
but seems to not be used by our colleagues in the sciences.
In particular our physics, chemistry, biology, geology, and
oceanography colleagues often tell us: “I tried to use
databases in my project, but they were just to [slow | hardto-use | expensive | complex ]. So, I use files.” Indeed,
even our computer science colleagues typically manage
their experimental data without using database tools.
What’s wrong with our database tools? What are we doing
wrong? We have been trying to answer those two
questions in the context of astronomy data – that is, we
have been working for the last seven years to try to bridge
the gap between database products and the community of
professional astronomers.
In the process we have come to realize that science is
being transformed – most scientific branches are becoming
data-rich. To deal with the data avalanche, each discipline
is growing a data analysis arm that collects large datasets
representing mankind’s knowledge of the field. These
datasets are augmented with applications that encapsulated
mankind’s understanding of the data and our
understanding of that branch of science. Each field is
being objectified – it is building an object model that
represents the facts and building methods that allow easy
selection and recombination of the facts. This represents
one of the most exciting branches of each of these sciences
– and database and data mining technologies are the heart
of each of these developments.
Put more succinctly: each science was empirical at first
and then grew an analytic branch. Recently the analytic
branches have spawned a computational branch that
simulates the analytic models. So the X department of
almost all universities has a computational-X faculty (e.g.:
computational-linguistics, computational-genomics,
computational-chemistry, etc.). These computational
Historically, scientists gathered and analyzed their own
data. But technology has created functional specialization
where some scientists gather or generate data, and others
analyze it. Technology allows us to easily capture vast
amounts of empirical data and to generate vast amounts of
simulated data. Technology also allows us to store these
bytes almost indefinitely. But there are few tools to
organize scientific data for easy access and query, few
tools to curate the data, and few tools to federate science
archives.
Domain scientists, notably NCBI (ncbi.nih.gov) and the
World Wide Telescope (WWT, ivoa.net), are making
heroic efforts to address these information management
problems. But it is a generic problem that cuts across all
scientific disciplines. A coordinated effort by the
computer science community to build generic tools could
help all the sciences. Our current database products are a
start, but much more is needed.
We have been participating directly in the WWT effort.
This is a personal and anecdotal account of what we
learned working with the astronomers and advice to others
on how to do similar things for astronomy or other
disciplines.
2. Science is Changing
For each science discipline X, X-info and comp-X are
controversial. There is a long-tradition of using primary
2
sources -- gathering and analyzing your own data rather
than using second-hand data or synthetic data. The
traditional primary-data approach is still valid, but
scientists increasingly need to compare their data to large
shared databases (e.g. is this DNA sequence already in
Genbank? If so, what is known about it?) In some cases
advances come from analyzing existing data sources in
new ways – the pentaquark was found in the archives once
theoreticians told us what to look for.
Traditionally, data analysis has followed an ftp-grep
model. In the ftp-grep model, the scientist first gets a
copy of the relevant data from an archive: that is she uses
ftp (file transfer protocol) to copy the relevant data sets
from a science archive on the internet to a local
computational server. The next step is that the scientist
uses an analysis package to scan through the data looking
for patterns. We use the name grep (generalized regular
expressions) to describe this process. In fact the analysis
program is often a “dusty deck” written long ago in
Fortran invoked with a scripting language like Python.
The ftp-grep data analysis approach is breaking down as
datasets grow to terabyte and petabyte scale. First,
although you can ftp a megabyte in a second, it takes days
or years to ftp a terabyte to your local server – and most
local servers do not have spare terabytes of space. But
datasets are growing to the petabyte range. So, scientists
must to learn to move the questions-and-answers across
the internet, and do the data analysis near the data
whenever possible.
Section 4 explores this data-archive and portal architecture
in more depth. But here, we want to make the point that
the skills and the workflow of next generation science will
be different. The new breed will need new tools and need
the skills to use them.
Traditionally science data was included in tables and
graphs that were part of the science article. This metaphor
is breaking down or has already broken down – the data is
just too large to be printed. Rather the data is placed in a
project website (like BaBar) or is deposited with a science
archive like Genbank or Astrophysics Data System. So,
the concept of publishing has morphed so that the “paper”
is about three things: (1) “how to read the data”, (2) “here
is what we found”, and (3) “here are some things we are
going to look for next in the data and things you might
want to look at.”
In this new world, the data is published on the Internet and
the literature mostly summarizes the data, is a guide to it,
and presents some interesting conclusions. Not all science
is like this, but the trend is definitely increasing.
Even more interesting to us is the fact that new results
come from new combinations of existing data. It appears
that there is a “Metcalfe’s law” for datasets: the utility of
N independent data sets seems to increase super-linearly.
One can make N(N-1) ~N2 comparisons among the datasets
and so the utility is approximately N2.
The increased utility of being able to combine multiple
independent datasets is one of the drivers for the digital
libraries being built by the various science disciplines.
The biology community has impressive portals in Entrez,
DDJB, and Esembl – they integrate all the data and the
literature. They are adding new datasets rapidly. They
represent an architecture where one group manages both
the data and literature all in one location. An alterative
approach, typified by SkyQuery.Net lets the data reside at
the data sources and a web portal (SkyQuery.Net in this
case) federates them via web services.
We discuss these architectural alternatives in Section 4.
The point here is that the World Wide Digital Library for
each discipline X is quickly evolving – mostly driven by
the domain scientists. With time we expect these many
libraries to grow together to form one giant library. The
architecture for these federated libraries and for the
federation of libraries stands as a challenge for the
computer science field. If we fail in our responsibility to
address that challenge, the domain scientists will certainly
come up with solutions and we computer scientists will
have missed an opportunity to do something really
valuable.
One problem the large science experiments face is that
software is an out-of-control expense. They budget 25%
or so for software and end up paying a lot more. The extra
software costs are often hidden in other parts of the project
– the instrument control system software may be hidden in
the instrument budget. The data pipeline software budget
may be hidden in some of the science budgets. But there is
no easy way to hide the cost of the archive in some other
budget. So the archive software often does not get written
at all. In this scenario, the data is deposited in a poorly
documented ftp-server and is never connected to an easy
public access.
The funding agencies want scientists to publish their data,
but the agencies cannot afford to pay for the data archives.
One thing we computer scientists could do is to make it
much easier and less expensive to publish scientific data.
3. Rules of Engagement – Working With X
How can you, a computer scientist, engage with a domain
scientist (or group)? You might try going native – switch
from computer science to bio or eco or geo or X science –
stop reading all those old CS journals and give up going to
all those computer science conferences.
Certainly, we have seen going native “work” but it has two
shortcomings (1) you are giving up a lot, (2) you are
unlikely to be able to help domains Y and Z who need you
just as much as X. So, our advice is to stick to your
expertise. Keep your office, your CS appointment and your
3
magazine subscriptions; but, start a collaboration with
some scientists and over time broaden that collaboration to
work cross multiple disciplines.
So, how do you engage? Well, first you have to find
someone in domain X who is desperate for help. Frankly,
most scientists are quite happy with the status quo. They
are collecting their data; they are publishing their papers;
and they are generally working the way they and their
colleagues have worked for decades if not centuries. All
this stuff about a data avalanche is not really their problem.
There is however a small group of pioneers in most fields
who are absolutely desperate for help. They are
pioneering a new instrument, or have a new simulator that
is indeed producing the data avalanche. The answer is
somewhere inside the terabytes if they could only get the
data processed, organized, and then be able to query it in
finite time.
You will recognize these people when you meet them –
they are the ones with the jobs that take weeks or months
to run their Python scripts. Their delay from question to
answer is days or weeks. They are the ones who are doing
batch processing on their data. They envy people who have
interactive access to data and envy people who can explore
their data. They are the ones who are desperate for help.
Desperate enough to even spend time on a long-shot like
working with you.
OK, so suppose you locate a desperate person. How can
you help? How can you communicate? Well, first you
have to learn a bit of their language. This generally
involves mastering the introductory text for that domain –
it is painful and you can skip this step if you are in a hurry,
but you will end up doing this work in any case. Doing it
early is the most efficient way. In parallel you have to
form a working relationship with the domain experts
(scientists.) You need to put in enough face time so that
they are not surprised to see you. This goes hand-in hand
with developing a common language. The converse of
this, the domain scientists you are working with need to
explore some of the things done in computer science and
in other disciplines so that they have a sense of what is
possible and what is almost possible.
Once you have established communication, you need to
find problems that leverage your skills. This requires you
do some ethnography of Domain X. Do a time-andmotion study of how the scientists are currently working,
and look for some low-hanging-fruit; simple tools you
could build or deploy that would be a major productivity
boost. You are looking for a project that will take a few
weeks or months of your time, and that would save them
years of their time.
To give a specific example, in the context of the Sloan
Digital Sky Survey data system, we developed a set of 20
questions each of which would have taken a week or more
of programming and a week of execution with the thenexisting tools. We also developed 12 visualization tasks
(scenarios) that typify the kind of data visualization that
scientists want to do with the Sloan Digital Sky Survey
data [Szalay00].
These 20 questions then became the requirements
document for the design of the SDSS Archive server. It
had to be easy (less than an hour) to ask each of these
questions and the answers had to come back within
minutes. This was a contract that both sides (the
Astronomers and the Computer Scientists) could
understand.
Since that time, we have seen this 20-questions approach
applied to many other branches in astronomy. Typically
the new group adopts about half of the 20 questions from
the original set but discards about half and adds ten more
that are unique to that branch or instrument.
Our rough observation is that scientists spend a lot of time
loading, scrubbing, and conditioning data. Once they
have the data in acceptable form, they spend their time
looking for haystacks and for needles in haystacks – that is
they spend their time looking for global properties and
global trends and patterns (the haystacks), or they spend
their time looking for anomalies (needles in the haystacks).
Computer scientists can help automate the data handling
processes, they can help with the data analysis, and they
can help with the scientists ask questions and understand
answers.
It seems Computer Scientists can be roughly divided into
three groups: (1) Data Visualization experts who are good
at capturing user’s questions and good at visualizing
answers in novel ways. (2) Data Analysis experts who are
skilled at statistical analysis, data mining algorithms, and
machine learning algorithms, or just algorithms in general.
(3) Software Systems people who are good at data
management or systems management or workflow
management. The domain scientists will likely not
understand these broad distinctions, but computer
scientists know that these three communities are largely
segregated – they each have their own conference and
journals.
Given these observations, it seems obvious that the
Visualization and Analysis computer scientists can help
improve existing systems, but that the logical progression
is for the systems people to build a “rationalized base” and
then for the Algorithms and Visualization people to build
new tools atop that base. So, we have been working away
to build both individual astronomy archives and to federate
that as described in the next sections. We are certainly not
done building the base, but we think we are ready to
accommodate any Analysis or Visualization experts who
want to apply their techniques to this corpus.
4
A rough chronology of our collaboration over the last
seven years is that the twenty queries suggested a logical
and physical database design – and also showed the need
for a spatial data access method. About 25% of the queries
involved spatial clustering of objects. Addressing these
issues took a year or two and naturally led to an “outreach
website (SkyServer.SDSS.org) that provides both public
access to the data and is also the basis for online
astronomy education for pre-college students. There are
now 150 hours of online astronomy instruction in several
languages that teach both astronomy and teach tools and
skills to analyze a large online scientific database. This
education component comprises more than 10% of the
traffic on the SkyServer website.
At this point, data loading had become an onerous task –
the data kept coming and being reprocessed. So, we
automated the loading process with a workflow system
built atop SQLserver and its Data Transformation Service
(DTS.) With the SkyServer in place we began exploring
architectures for the World Wide Telescope (the Virtual
Observatory) federating all the world’s astronomy
archives.
A pixel server (image cutout) web-service was the first
experiment with web services. It led to the SkyQuery
(SkyQuery.Net) architecture.
In parallel, several other groups did their own 20-questions
exercise and cloned the basic SkyServer database design
and spatial access methods for their archives. Recently we
have been refining and extending our spatial access
techniques , redoing our design in light of the revolution
in database architectures (the unification of data and
methods), and also preparing for the next-generation
surveys which will be capturing a thousand times more
data.
In summary, it has been a very productive collaboration. It
has in part answered the question: “Why don’t they use our
stuff?” The answer is indeed that databases are hard to
use. Paradoxically, it is difficult to get data into them –
data scrubbing and loading is labor intensive. Organizing
data is challenging. Once the data is loaded, cognoscenti
can do wonders, but “normals” do not have the data
analysis skills to pose queries and to visualize answers.
As the community evolves, and as some scientists master
these skills, they communicate successful patterns to their
colleagues – that is how scientific techniques evolve. It is
an organic process.
In closing this section on “how to collaborate” we confess
that we are very data centric. Computer scientists might
contribute to domain X by bringing new ways to think
about dynamic systems or new notations for dynamic
systems. Differential equations are the standard way to
describe dynamics 00 but they are not particularly good at
describing discrete event systems. Computer scientists
have struggled with and modeled discrete event system
with tools like simulation languages, process algebras, and
temporal logics. It is quite possible that these new ways of
describing discrete events could provide breakthroughs in
process-oriented sciences as diverse as ecology, metabolicpathways, or social interactions.
4. Architecture: Services and Portals.
The broad architecture of the World Wide Telescope and
of other digital libraries has emerged over the last decade.
They will be federations of data servers, each providing
access to its data resources. The federation will be unified
by one or more portals, typically collocated with one of the
mega-servers. Each portal provides high-level tools to
access the “local” archive and also to correlate that data
with the other data sources in the internet and with the
scientific literature.
This pattern is emerging in other disciplines – for example
Entrez, DDJB, and Esembl show the same pattern in the
bio-informatics domain. A mega-archive co-located with a
portal providing tools that integrate and federate that
archive with the other important archives.
The Service-Portal architecture requires that data from
different data services be interchangeable – that there be
standard terms and concepts that apply across all the
archives. This requires the objectification of the
discipline – defining an object model that encapsulates
facts and represents them in standard ways. The
astronomy community has embraced the UCD
methodology as a core vocabulary for units and facts
[UCD]. From a database perspective – this is the schema.
Built atop this, the astronomers have defined a few web
services that, given a query, return the answer. These
services are defined in terms of XML Schema and SOAP
messages. The first service returned all objects within a
circle (cone) [Cone]; the second returned an image of an
area of the sky [SIAP]. There is also a spectrogram
service. A dozen generic services have been defined so far
[VO-services].
The portals build on top of these simple services providing
ways to combine data from multiple archives. This is
nascent now – there are only a few portals, but each of
them is evolving rapidly.
One unanticipated benefit of developing the SkyServer
architecture and placing it in the public domain is that it
has been copied about ten times by other archives. The
server has a “spine schema consisting of the spatial data
organization and the general organization of optical and
spectroscopic data. Other surveys have discarded the
SDSS-specific attributes and replaced them with their own.
The result is that the Royal Observatory of Edinburgh,
Space Telescope Institute, Cornell, Caltech, and NOAO
have each been able to leverage this software base. In
5
general, a survey can be converted to the spine schema in a
matter of hours if it is small (gigabytes) or a matter of days
if it is large (terabytes).
The common spine schema and web services interface has
also enabled us and others to add many datasets to the
SkyQuery interface. Once the data is online, the web
services can be defined and it can be registered as a
member of the SkyQuery federation in a few hours [VO
Registry].
The portal itself has evolved in two un-anticipated ways:
(1) users can define personal databases that represent their
“workspace” on the server, and (2) users can submit longrunning queries to a batch system that schedules these jobs
and deposits the answers into the user’s personal database.
Both these features are described in [Nieto-Santisteban 04]
and are briefly sketched here.
Users routinely perform data analysis in steps – using the
results of previous steps in subsequent analysis. The often
start the analysis with datasets that they have developed
over the years. To facilitate this, the portal lets a user
define a database that is private to that user. The user can
create tables in that database and upload private data to it.
The user can deposit query answers to existing or new
tables in this database, and can use tables from this
database in subsequent queries. Access control is based
on username and password (better security mechanisms are
being discussed). The database owner can grant read or
write authority to the personal database. This is
convenient for collaborators who share a workspace and is
a convenient way to publish data to your colleagues. The
databases are currently limited to 1GB, and so far no one
has asked for more – but we see no reason to limit the
databases to such a small size in the future.
The portal has limits on how much processing a query can
consume (elapsed time) and how large the outputs can be.
For the public site the limits are 90 seconds elapsed and
1,000 rows of output. For the collaboration site the limits
are 90 minutes elapsed and 500,000 rows of output. Both
categories occasionally need to exceed these limits, and it
is often the case that the user does not want to wait for the
job to finish. This obviously suggests a batch job
scheduler. Users can submit jobs to this scheduler which
then runs them in the background. If a job exceeds its
quota, there is an option to double the quota and re-run the
job (up to some limit). There is also a simple web
interface to check on the status of jobs. The easiest way
to understand this system is to just try it: create a personal
database and run some sample queries [CAS].
5. Was It Worth the Effort?
After all this work, what are the benefits? Are the
scientists better able to access their data? What
breakthroughs has this enabled? At best, we can answer
these questions with anecdotal evidence.
The SkyServer website has served about 60M web hits.
More than 10% of that traffic is to the educational part of
the site – so the “outreach” has been very successful. The
collaboration SkyServer website is currently processing
more than 100,000 SQL queries per month, the public site
processes many more and the CAS Job system is
processing nearly 10,000 jobs per month. So, the system
is being used. The astronomers have wrapped the web
services in Emacs and Python scripts. Many use them as a
matter of course in their work.
There are two side-by-side comparisons of the files-vsdatabase approach. One compared a “Grid-Condor”
design for finding galaxy clusters to a database-indexparallelism approach. The DB did the job better, simpler,
and 50x faster [Nieto-Santisteban 04a]. In a second
experiment, a colleague had written a program to find
near-earth objects (asteroids). The program was about 20
pages of code and ran for three days. We wrote a 50-line
SQL program in a few hours that got the same answer in
about 30 minutes. When we added an index to support the
query, it ran in 3 minutes. When questions and answers
can be done in minutes it changes the kinds of questions
you ask – this seems to us to be an example of the
breakthrough we are talking about.
The SDSS is a side-by-side optical and spectroscopic
survey. Spectrograms give precise redshifts and hence
approximate distances based on Hubble’s constant. We
use object spectra as a “training set” to estimate even more
galaxy distances based on photometry alone – the so called
photometric redshifts. We found that one category of
objects was significantly under-sampled by the
spectroscopic survey and so our photometric redshift
estimator was less sensitive than it might be. The code to
schedule and “drill” the fiber plate to take spectra was
huge and poorly understood – so the astronomers could not
change it easily to address this custom problem. In an
evening we wrote the queries to select a set of observations
and generated a drilling list (holes in an observing plate).
The observation was scheduled and a few months later we
got the data from the observation. Within hours the data
was loaded into the database and the redshift estimator was
recalibrated with the larger dynamic range and
substantially improved accuracy.
Our colleagues have much better access to the SDSS data
than they had to any previous survey. Not just that, having
a database has improved data quality. Foreign keys, by
having integrity constraints and by doing integrity checks
6
In summary, we think that astronomers are finding that this
is the easiest way to publish their observational data. It
could be better, but it is a substantial advance over what
existed. Now the challenge is to make the process even
more automatic, deal 1,000x more data volumes, and
provide the missing visualization and analysis tools.
6. Some Lessons Learned From the WWT
Astronomy is a good example of the data avalanche. It is
becoming a data-rich science. The computationalAstronomers are riding the Moore’s Law curve, producing
larger and larger datasets each year. The observational
astronomers are also riding Moore’s law, deploying denser
CCDs and more instruments each year. The total data
volume seems to be doubling every year.
We have learned some interesting lessons from managing
the Sloan Digital Sky Survey dataset. It is about 40 TB of
raw image data which places it somewhere between
Genbank (50GB) and BaBar (1PB) in the spectrum of
scientific datasets.
Data Inflation Paradox: You might guess that the derived
data products, the processed astronomy data, would be
10% of the size of the raw instrument data. After all, the
night sky is mostly black. Your intuition is approximately
correct. The Data Inflation Paradox is that after a few
years the derived data products form most of the data
volume.
NASA has a useful classification for data products: Level
0 is the raw instrument data, Level 1 is the calibrated data,
and Level 2 are derived data products built from Level 1
datasets. As time passes, the Level 0 data arrives at an
approximately constant rate. But, a growing number of
Level 1 and higher data products are derived from the raw
data – star catalogs, QSO catalogs, cluster catalogs,
materialized views that are particularly convenient. The
derived products are indeed much smaller – but they are
also much more numerous.
Versions are a second data inflation driver. The derived
data products are produced by software. Software comes
in versions. As the software improves, the entire Level 0
archive is reprocessed to produce a new edition of the data.
The 5-Year peak: When the archive starts, data arrives at
a certain rate, let’s say 10TB per year. After 5 years, you
will have 50 TB of Level 0 data and perhaps 50 TB more
of Level 1 data products. In the mean time, Moore’s law
will have made storage and processing about ten times less
expensive. But in addition, you need to replace
equipment that is more than 3 years old. So, the cash flow
is a fairly complex combination of these three forces: data
inflation, Moore’s law, and depreciation. Figure 1 plots
these three forces. Ignoring data inflation, the peak is at 2.5
years, but with data inflation it is at 5 years.
Overpower by 6x: However much processing and storage
you think you need, you need six times as much. The first
factor of two is easy to explain: if you store data, you need
multiple copies just to be safe. If you do some
computation, you should plan to use 50% of the processing
power because things inevitably break, and mistakes are
inevitably made. So, it is better to build in a factor of 2.
What about that other factor of 3? Well, for storage, you
need to have the “production version” and the “nextproduction version” As you are building the next version
you often find yourself sorting the data or indexing it or
doing some operation that doubles or triples the size of the
data (old version, intermediate version, target version.)
You can generally find a way round this, by being careful
(and lucky), but it is generally a lot more economical to
optimize people’s time rather than computer hardware.
So, that is where the factor of 3 comes from.
180
Yearly Demand
160
140
Depreciated Inflated Demand
120
Inflated Demand
Naive Demand
100
Yearly Capital Cost
4.0
3.5
Marginal Capital Cost
The SkyServer embeds the data documentation in the DDL
(using the UCD notation and terms [UCD].) A tool scans
this documentation and automatically generates online
documentation for the website. So, once the astronomer
has defined the DDL, he can use the automated loading
tools (DTS) to load the database, and the website comes up
with access to the data and with online documentation –
immediately. This has been a boon to several projects –
saving months if not years of development.
Once this edition is published, it becomes part of the
science record and must remain published (unchanged)
forever. Suppose you have an archive growing at one
terabyte per year at level 0 with a derived data product that
is 10x smaller. In year 1, the data product is 10% of the
storage. In year 5 you have five terabytes of level 0 data
and .1+.2+.3+.4+.5 = 1.5 TB of level 1 data. Over the Nyear life of the project, the level 1 data volume grows as N2
while the level 0 data grows as N. Even if there is only
one data product, the derived data volume will eventually
overwhelm the Level 0 data volume. But, there are many
Level 1 products – and their number grows with time.
Yearly Demand ( R )
in the data loading process have uncovered a myriad of
problems which have fed back into the pipeline.
80
60
40
3.0
2.5
2.0
1.5
1.0
0.5
20
0.0
0
0
2
4
Year
6
8
10
0
2
4
Year
6
8
10
Figure 1: Archive expense (yearly outlay) for an
archive where equipment depreciates over 3 years,
where Moore’s law improves prices by 60% per year,
and where data arrives at a constant rate, where the
data inflation of ten data products is quadratic.
7
For example, when you are reprocessing all the old stuff
and you are processing the current stuff, it is important to
have a substantial processing, IO, and storage overcapacity so that you can redo the work before it is time to
redo the redo.
Automate Data Loading: A substantial part of the
project’s energy will go into data loading. The database
will be reloaded several times – perhaps once or twice a
year. So it is essential that the load process be automatic
and that there be automatic ways to detect and reconcile
data quality errors. People told us this before we started –
but we did not believe them. They were right. We
invested in this technology late and so wasted a lot of time.
At this point, data loading is fairly automatic.
Performance Regression Tests: As the database evolves,
and grows, it is essential to have a regression test. The
ones we settled on were (1) we should be able to reload the
entire database in a week, and (2) none of the 20 queries
should take more than 10 minutes to run. Every major
release causes us to fix one or another problem and to buy
much better hardware in order to meet these performance
goals.
Data Pyramid: It is very convenient to have a 1GB, 10
GB, and 100 GB subsets of the database. The smallest
database fits on a laptop. It facilitates quick experiments
on one or another design issue. The larger databases of
course require more equipment to support them, but each
has its uses.
7. Summary
We began by commenting that science is changing and that
computer-science is in a position to help its sister sciences
manage the impending data avalanche. One question we
sought to answer was: “Can databases help?” The simple
answer is: “Yes, if…” The astronomers for example
needed a few things not part of the standard packages
(spherical spatial data access, statistical functions,…), and
they needed database expertise to help them bridge the gap
between their problem and our technology.
The sciences also face data mining and data visualization
challenges that we have not addressed here (beyond
pointing out how important they are.).
This is literally a golden opportunity for Computer Science
to make a contribution if we will take the time to work
with our colleagues in X for almost any science X.
Acknowledgments
The SkyServer evolved from Tom Barclay’s TerraServer
design. The SkyServer, SkyQuery, and many other ideas
described here were developed by the team consisting of
Tamas Budavari, Gyorgy Fekete, Nolan Li, Maria NietoSantisteban, William O'
Mullane, Peter Kuntz, Don Slutz,
Chris Stoughton, and Ani Thakar. A.S. was supported by
grants from NSF, NASA, the W.M. Keck Foundation, and
the Gordon and Betty Moore Foundation.
One particularly useful aspect of the 1GB DB and the
corresponding web server is that you can give it to
someone as a “double-click-to-install” application. OK,
now they have a fully functional personal SkyServer. Now
they can throw away the SDSS data and replace it with
their data. This has been a very easy way for people to get
started with the code.
Capture the Bits: The Sloan Digital Sky Survey is a
federation of more than 25 institutions and more than 200
scientists. Much of the system’s design has been done
with teleconferences and email and source-code-control
systems and a bug-tracking database. Today, you can
answer almost any question by asking the experts. But the
time is coming when the experts will have forgotten or will
no longer be with us. So, it is important to curate both
the data and all the discussions about how it was gathered
and processed. We wish we had the resources to carefully
curate all the SDSS design decisions and software. We do
not. Short of that, what we are doing is capturing all the
bits we can – all the eMail logs, all the software versions,
all the bug logs, all the meeting minutes …everything we
can lay our hands on. This is all going into a big
repository. Our hope is that future software will be able to
organize and mine this information.
8
References
[CAS] http://skyserver.sdss.org/cas/en/
[Cone] “Cone Search definition”,
http://www.us-vo.org/pubs/index.cfm
[Gray02a] “ Online Scientific Data Curation, Publication, and
Archiving,” J. Gray, A. S. Szalay, A. R. Thakar, C.
Stoughton, J. vandenBerg, SPIE, 22-28 August 2002,
Waikoloa, Hawaii,
[Gray02b]“The World Wide Telescope: An Archetype for Online
Science ,” J. Gray, A. S. Szalay, MSR TR 2002-75, pp 4,
CACM, Vol. 45, No. 11, pp. 50-54, Nov. 2002
[Gray03] “Data Mining the SDSS SkyServer Database,” J. Gray,
A.S. Szalay, A. Thakar, P. Kunszt, C. Stoughton, D. Slutz, J.
vandenBerg, Distributed Data & Structures, pp 189-210
Carleton Scientific 2003, ISBN 1-894145-13-5
[Nieto-Santisteban 04a] “Extending the SDSS Batch Query
System to the National Virtual Observatory Grid”, M. A.
Nieto-Santisteban, W. O'
Mullane, J. Gray, N. Li, T.
Budavari, A. S. Szalay, A. R. Thakar, MSR-TR-2004-12,
[Nieto-Santisteban 04b]“When Database Systems Meet The
Grid,” M. A. Nieto-Santisteban, A. S. Szalay, A. R. Thakar,
W. J. O’Mullane, J. Gray, J. Annis, MSR-TR-2004-81,
[SIAP] “Simple Image Access Protocol,”
http://www.us-vo.org/pubs/index.cfm
[Szalay00] A.S. Szalay, J. Gray, P. Kunszt, A. Thakar
“Designing and Mining Multi-Terabyte Astronomy
Archives: The Sloan Digital Sky Survey ,” Proc. ACM
SIGMOD, June 2000, pp. 451-462.
[Szalay02a] “Petabyte Scale Data Mining: Dream or Reality?,”
A. S. Szalay, J. Gray, J. vandenBerg, SIPE, 22-28 August
2002, Waikoloa, Hawaii,
[Szalay02b] “Web Services for the Virtual Observatory,” A. S.
Szalay, T. Budavária, T. Malika, J. Gray, A. R. Thakar,
SPIE, 22-28 August 2002, Waikoloa, Hawaii,
[UCD] “Unified Content Descriptors (UCDs)”, IVOA Proposed
Recommendation 26 October 2004,
http://www.ivoa.net/Documents/
[VO Registry] http://nvo.stsci.edu/voregistry/index.aspx
[VO-services] “Web Services for the Virtual Observatory,”
http://www.voservices.org/
[VOtable] “VOTable: A Proposed XML Format for
Astronomical Tables,”
http://www.us-vo.org/pubs/index.cfm
9
| 5cs.CE
|
STRONG PSEUDOPRIMES TO TWELVE PRIME BASES
arXiv:1509.00864v1 [math.NT] 2 Sep 2015
JONATHAN SORENSON AND JONATHAN WEBSTER
Abstract. Let ψm be the smallest strong pseudoprime to the first m prime
bases. This value is known for 1 ≤ m ≤ 11. We extend this by finding ψ12
and ψ13 . We also present an algorithm to find all integers n ≤ B that are
strong pseudoprimes to the first m prime bases; with a reasonable heuristic
assumption we can show that it takes at most B 2/3+o(1) time.
1. Introduction
Fermat’s Little Theorem states that if n is prime and gcd(a, n) = 1, then
an−1 ≡ 1
(mod n).
A composite number for which this congruence holds is called a pseudoprime to
the base a. We can restate Fermat’s Little Theorem by algebraically factoring
(repeatedly) the difference of squares that arises in an−1 − 1. In which case, if n is
an odd prime, writing n − 1 = 2s d with d odd, and gcd(a, n) = 1, then either
ad ≡ 1
(mod n)
or
k
a2
d
≡ −1
(mod n)
for some integer k with 0 ≤ k < s. If a composite n satisfies the above, we call
n a strong pseudoprime to the base a. Unlike Carmichael numbers, which are
pseudoprimes to all bases, there do not exist composite numbers that are strong
pseudoprimes to all bases. This fact forms the basis of the Miller-Rabin probabilistic
primality test [13]. Define ψm to be the smallest integer that is a strong pseudoprime
to the first m prime bases. (See A014233 at oeis.org.)
The problem of finding strong pseudoprimes has a long history. Pomerance, Selfridge, and Wagstaff [12] computed ψm for m = 2, 3, and 4. Jaeschke [7] computed
ψm for m = 5, 6, 7, and 8. By looking at a narrow class of numbers, Zhang [18] gave
upper bounds on ψm for 9 ≤ m ≤ 19. Recently, Jian and Deng [8] verified some of
Zhang’s conjectures and computed ψm for m = 9, 10, and 11. We continue in this
effort with the following.
Date: September 4, 2015.
2010 Mathematics Subject Classification. Primary 11Y16, 11Y16; Secondary 11A41, 68W40,
68W10.
This work was supported in part by a grant from the Holcomb Awards Committee. We thank
Frank Levinson for his generous support of Butler University’s cluster supercomputer, Big Dawg.
We would also like to thank Andrew Shallue for his feedback on a draft of this paper, and Paul
Pollack for his help with some details regarding an earlier proof approach to the running time
bound for λ-sieving.
1
2
JONATHAN SORENSON AND JONATHAN WEBSTER
Theorem 1.1.
ψ12 = 3186 65857 83403 11511 67461.
ψ13 = 33170 44064 67988 73859 61981.
We also verified Jian and
p Deng’s work in finding ψ9 = ψ10 = ψ11 . Note that the
ERH implies ψm ≥ exp( m/2), if ψm exists [1].
The proof of our theorem relies primarily on running an improved algorithm for
finding strong pseudoprimes:
Theorem 1.2. Given a bound B > 0 and an integer m > 0 that grows with B,
there is an algorithm to find all integers ≤ B that are products of exactly two primes
and are strong pseudoprimes to the first m prime bases using at most B 2/3+o(1)
arithmetic operations.
We have a heuristic argument extending this B 2/3+o(1) running time bound to
integers with an arbitrary number of prime divisors. This result is an improvement
over a heuristic O(B 9/11 ) from [2]. Our algorithm uses the work of Jian and Deng
as a starting point, and our specific improvements are these:
• We applied the results in Bleichenbacher’s thesis [2] to speed the computation of GCDs. See §2.2.
• We applied the space-saving sieve [4, 15] which enabled us to sieve using
full prime signature information. See §2.3.
• We used a simple table of linked-lists indexed by a hash function based on
prime signatures. We used this to store small primes by signature for very
fast lookup. See §2.4.
• When it made sense to do so, we parallelized our code and made use of Big
Dawg, Butler University’s cluster supercomputer.
Our changes imply that adding more prime bases for a pseudoprime search (that
is, making m larger) speeds up the algorithm.
The rest of this paper is organized as follows. Section 2 contains a description
of the algorithm and its theoretical underpinnings. Section 3 contains the proof of
the running time. Section 4 contains implementation details.
2. Algorithmic Theory and Overview
To find all n ≤ B that are strong pseudoprimes to the first m prime bases, we use
some well-known conditions to limit how many such integers n we need to check.
This is outlined in the first subsection below.
We construct all possibilities by generating n in factored form, n = p1 p2 . . . pt−1 pt ,
where t is the number of prime divisors of n, and pi < pi+1 . It turns out that t
cannot be too big; we only had to test up to t = 6. So we wrote separate programs
for each possible t and the case t = 2 dominates the running time.
Let k = p1 p2 · · · pt−1 . Our algorithms generate all candidate k values, and
then for each k search for possible primes pt to match with k, to form a strong
pseudoprime candidate n = kpt .
For the bound B, we choose a cutoff value X. For k ≤ X, we use a GCD
computation to search for possible primes pt , if they exist. For k > X, we use
sieving to find possible primes pt . Again, we use separate programs for GCDs and
for sieving.
STRONG PSEUDOPRIMES TO TWELVE PRIME BASES
3
Once each n is formed, we perform a strong pseudoprime test for the first m
prime bases to see if we have, in fact, found what we are looking for.
2.1. Conditions on (Strong) Pseudoprimes. For the purposes of this paper,
it suffices to check square-free odd integers. A computation by Dorais and Klyve
[3] shows that a pseudoprime to the bases 2 and 3 must be square-free if it is less
than 4.4 · 1031 . Since the values Zhang conjectured as candidates for ψ12 and ψ13
are less than 4.4 · 1031 , we restrict our search to square-free numbers.
In general, to rule out integers divisible by squares, it suffices to find all Wieferich
primes ≤ B 1/2 , and for each such prime, perform appropriate tests. This can easily
be done in well under O(B 2/3 ) time. See §3.6.1 in [2].
To each prime p we associate a vector called its signature. Let ν = (a1 , a2 , . . . , am )
for ai distinct positive integers. Then the signature of p is
σpν = (v(ordp (a1 )), v(ordp (a2 )), . . . , v(ordp (am ))) ,
where v denotes valuation with respect to 2 and ordp (a) is the multiplicative order
of a modulo p.
Example 2.1. Let p be the prime 151121. Then σpν = (3, 4, 0, 4, 2, 1, 2, 4) if ν =
(2, 3, 5, 7, 11, 13, 17, 19).
Theorem 2.2 (Proposition 1 in [7]). Let n = p1 · · · pt with t distinct primes,
ν = (a1 , a2 , . . . , am ) with different integers such that pi - aj for all 1 ≤ i ≤ t and
1 ≤ j ≤ m. Then n is a strong pseudoprime to each base in ν if and only if n is a
pseudoprime to each base in ν and σpν1 = · · · = σpνt .
If t > 2, then this theorem greatly limits the number of candidate k values we
need to check. In the case t = 2 the initial sub-product k is just a single prime,
so this theorem does not, at first, appear to help. However, it does play a role in
sieving for pt . For the rest of this paper, we use ν as the vector containing the first
m = 11 or 12 prime numbers.
2.2. Greatest Common Divisors. As above, let k = p1 · · · pt−1 . We explain in
this subsection how to find all possible choices for pt to form n = kpt as a strong
pseudoprime to the first m prime bases using a simple GCD computation.
Let b be one of the ai from ν. The pseudoprimality condition bkpt −1 ≡ 1
(mod kpt ) implies that
bkpt −1 ≡ bk(pt −1)+k−1 ≡ bk−1 ≡ 1
(mod pt ),
i.e., that pt divides bk−1 − 1. Considering multiple bases, we know that pt must
divide
gcd ak−1
− 1, ak−1
− 1, . . . , ak−1
m −1 .
1
2
However, since we are concerned with strong pseudoprimes, we can consider only
the relevant algebraic factor of bk−1 − 1. Following Bleichenbacher’s notation [2],
let k − 1 = u2d for u odd and
h(b, k) =
bu − 1
c−1
bu2
+1
if v(ordk (b)) = 0
.
if v(ordk (b)) = c > 0
4
JONATHAN SORENSON AND JONATHAN WEBSTER
Theorem 2.3 (Theorem 3.23 of [2]). A necessary condition for an integer n of the
from n = kpt , for a given k with pt prime, to be a strong pseudprime to the bases
a1 , . . . , am is that pt divide gcd(h(a1 , k), . . . , h(am , k)).
This theorem allows us to construct the product of all possible pt values for a
given k by computing a GCD. To carry this out, we figure out which of the h(ai , k)
values would be the smallest, compute that, and then compute the next smallest
h value modulo the smallest, perform a GCD, and repeat until we get an answer
≤ k, or use up all the ai base values.
Example 2.4. Let k = 151121, which is prime. We have
v(ordk (2)) = 3 =⇒ 237780 + 1 ≈ 8.189 × 1011372 ,
v(ordk (3)) = 4 =⇒ 375560 + 1 ≈ 1.914 × 1036051 ,
v(ordk (5)) = 0 =⇒ 59445 − 1 ≈ 5.911 × 106601 .
In this case, we would begin the computations with 59445 − 1 since it is the smallest.
We then compute gcd(237780 + 1 mod (59445 − 1), 59445 − 1) = 151121. This tells
us that any n of the form 151121 · pt is not a strong pseudoprime to the vector
ν = (2, 3, 5).
These computations start off easy as k is small and progressively get harder (in
an essentially linear fashion) as k grows. As a practical matter, these computations
only need to be done once. As an example, [8] found
84 98355 74122 37221 = 206135341 · 412270681
to be a strong pseudoprime to the first eight prime bases. When considering p1 =
206135341, they sieved (which we explain in the next section) for p2 = 412270681.
It is reasonable to ask whether there is a larger prime that can be paired with p1
that would form a strong pseudoprime to eight bases. In our computation we check
by computing a GCD to answer “no.”
Algorithm 1: Using GCDs to rule out k values
Input: Integers t and k, where t is the number of prime divisors and k is the product of t − 1 primes with matching signature, and the base vector ν containing
the m smallest primes.
1: Let b ∈ ν give the smallest estimated value for h(b, k) as defined above.
2: If b = a1 , let i = 2 else let i = 1.
3: Compute h(b, k).
4: Compute x = h(ai , k) mod h(b, k) using modular exponentiation.
5: Set x = gcd(h(b, k), x);
6: while x > k and i < m do
7:
Set i = i + 1; if ai = b set i = i + 1 again.
8:
Compute y = h(ai , k) mod x.
9:
Set x = gcd(x, y).
10: end while
11: if x < k then
12:
We can rule out k.
13: else
14:
We factor x and check each prime pt | x with pt > k to see if kpt is a strong
pseudoprime.
STRONG PSEUDOPRIMES TO TWELVE PRIME BASES
15:
5
end if
Computing h(b, k) is the bottleneck of this algorithm. When using m = 11 or
higher, we never found anything to factor in the last step.
2.3. Sieving. As above, let k = p1 · · · pt−1 and we try to construct n = kpt by
finding pt . In order to search up to a bound B, we need to search primes pt in
the interval [pt−1 , B/k]. We combine two types of sieving to search this interval
quickly. The first we call λ-sieving, and the second we call signature sieving. For
each prime p define
λp = lcm {ordp (a) : a ∈ ν} .
By the r-rank Artin Conjecture
[11], we expect with very high probability that
λp = p − 1. Let λ = lcm λp1 , . . . , λpt−1 . Since akpt −1 ≡ 1 (mod n) for any a ∈ ν
this means kpt − 1 is a multiple of λ. So, pt ≡ k −1 (mod λ). This tells us that
pt lies in a specific residue class modulo λ, and we call searching for pt using this
value λ-sieving.
We can further narrow down the residue classes that pt may lie in with respect
to primes in ν. The following two propositions appeared originally in [7].
Proposition 2.5.
p and q, if v(p − 1) = v(q − 1) and v(ordp (a)) =
For primes
a
a
v(ordq (a)) then p = q .
Proposition 2.6.
For primes p and q, if v(p − 1) < v(q − 1) and v(ordp (a)) =
v(ordq (a)) then aq = 1.
One may consider higher reciprocity laws, and Jaeschke did so [7]. We found
quadratic reciprocity sufficient.
The authors of [8] consider searching for pt modulo lcm{λ, 9240}. They considered 30 residue classes modulo 9240 = 8 · 3 · 5 · 7 · 11 that arose from the propositions
above. With the inclusion of each additional prime from ν, we effectively rule out
half of the cases to consider, thereby doubling the speed of the sieving process.
Ideally, we would like to include every prime in ν for the best performance, but
with normal sieving, the sieve modulus becomes quite large and requires we keep
track of too many residue classes to be practical. Instead, we adapted the spacesaving wheel datastructure [16, 15], which had been used successfully to sieve for
pseudosquares. Sieving for primes pt with specified quadratic character modulo a
list of small primes is the same algorithmicPproblem. This wheel sieve uses a data
structure that takes space proportional to a∈ν a instead of space proportional to
Q
a∈ν a as used in traditional sieving. It does, however, produce candidate primes
pt out of order. This is not an issue, so long as we make sure the sieve modulus
does not exceed B. In practice, we dynamically include primes from ν so that
1000 · k · lcm{λ, 8, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31} < ψ13 , and we favor the inclusion
of smaller primes.
Algorithm 2: Sieving
Input: k in factored form as k = p1 p2 · · · pt−1 , λpi for i < t, search bound B, and
base vector ν of length m.
1: Compute λk = lcm{λpi }.
2: Set w = 1. Compute wheel modulus w as follows:
6
JONATHAN SORENSON AND JONATHAN WEBSTER
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
for i = 2, . . . , m do
if kλk ai w < B/1000 then
if ai does not divide λk then
Set w = w · ai .
end if
end if
end for
if λk divisible by 4 then
Set λ = λk
else
Set λ = λk /2 and
Set w = 8w
end if
Let σ be the signature of p1 .
Build a wheel with modulus w so that for each prime power q | w we have primes
p generated by the wheel with (p/q) consistent with σ as in Propositions 2.5
and 2.6.
for each residue r modulo w generated by the wheel do
Use the Chinese Remainder Theorem to compute x mod wλ such that x ≡
r mod w and x ≡ k −1 mod λ.
Sieve for primes in the interval [k + 1, B/k] that are ≡ x mod wλ.
For each such probable prime pt found, perform a strong pseudoprime test
on kpt .
end for
Note that if k consists entirely of primes that are ≡ 3 mod 4, then all primes pt
found by the sieve, when using all of ν, will have the correct signature.
Example 2.7. Consider p = 3 · 108 + 317. Now, λ = p − 1 = 22 · 7 · 11 · 23 · 42349
(factorization is obtained via sieving). We may use signature information for the
primes 3, 5, 13, 17, 19, 29, 31, and 37. However, since B = 4 · 1024 we may only
use the first 4 primes in the signature before the combined modulus is too large.
The signature information for 3 implies that q ≡ 5, 7 (mod 12) however, the case
q ≡ 7 (mod 12) need not be considered since we know (because of λ) that q ≡ 1
(mod 4). Given the signature information for the primes 5 and 13, we search in
{2, 3 (mod 5)} and {2, 5, 6, 7, 8, 11 (mod 13)}. At this point we sieve for primes
q ≡ 1 (mod λ). To this we add the following 12 new congruence conditions and
sieve for each individual condition.
(1) Let q ≡ 2 (mod 5).
(a) Let q ≡ 2 (mod 13).
(b) Let q ≡ 5 (mod 13).
(c) Let q ≡ 6 (mod 13).
(d) Let q ≡ 7 (mod 13).
(e) Let q ≡ 8 (mod 13).
(f) Let q ≡ 11 (mod 13).
(2) Let q ≡ 3 (mod 5).
(a) Let q ≡ 2 (mod 13).
(b) Let q ≡ 5 (mod 13).
(c) Let q ≡ 6 (mod 13).
STRONG PSEUDOPRIMES TO TWELVE PRIME BASES
7
(d) Let q ≡ 7 (mod 13).
(e) Let q ≡ 8 (mod 13).
(f) Let q ≡ 11 (mod 13).
2.4. Hash Table Datastructure. In order to compute GCDs or sieve as described
above, it is necessary to construct integers k = p1 p2 · · · pt−1 in an efficient way,
where all the primes pi have matching signatures. We do this by storing all small
primes in a hash table datastructure that supports the following operations:
• Insert the prime p with its signature σp . We assume any prime being
inserted is larger than all primes already stored in the datastructure. (That
is, insertions are monotone increasing.) We also store λp with p and its
signature for use later, thereby avoiding the need to factor p − 1.
• Fetch a list of all primes from the datastructure, in the form of a sorted
array s[ ], whose signatures match σ. The fetch operation brings the λp
values along as well.
We then use this algorithm to generate all candidate k values for a given t and
ν:
Algorithm 3: Generating k Values
Input: t > 2, search bound B, cutoff X < B, base vector ν of length m.
1: Let T denote our hash table datastructure
2: Let am+1 denote the smallest prime not in ν.
q
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
for each prime p ≤ B/at−2
m+1 (with p − 1 in factored form) do
Compute λp from the factorization of p − 1;
s[ ] := T .fetch(σp ); List of primes with matching signature
for 0 ≤ i1 < · · · < it−2 ≤ s.length do
Form k = s[i1 ] · · · s[it−2 ] · p;
if k ≤ X then
Use k in a GCD computation to find matching pt values;
else
Use k for a sieving computation to find matching pt values;
end if
end for
1/3
if p ≤ (B/at−3
then
m+1 )
T .insert(p,σp ,λp );
end if
end for
Note that the inner for-loop above is coded using t−2 nested for-loops in practice.
This is partly why we wrote separate programs for each value of t. To make this
work as a single program that handles multiple t values, one could use recursion on
t.
Also note that we split off the GCD and sieving computations into separate
programs, so that the inner for-loop was coded to make sure we always had either
k ≤ X or X < k < B/p as appropriate.
We implemented this datastructure as an array or table of 2m linked lists, and
used a hash function on the prime signature to compute which linked list to use.
The hash function h : σ → 0..(2m − 1) computes its hash value from the signature
as follows:
8
JONATHAN SORENSON AND JONATHAN WEBSTER
• If σ = (0, 0, 0, . . . , 0), the all-zero vector, then h(σ) = 0, and we are done.
• Otherwise, compute a modified copy of the signature, σ 0 . If σ contains only
0s and 1s, then σ 0 = σ. If not, find the largest integer entry in σ, and
entries in σ 0 are 1 if the corresponding entry in σ matches the maximum,
and 0 otherwise.
For example, if σ = (3, 4, 0, 4, 2, 1, 2, 4) (our example from earlier), then
σ 0 = (0, 1, 0, 1, 0, 0, 0, 1).
• h(σ) is the value of σ 0 viewed as an integer in binary.
So for our example, h((3, 4, 0, 4, 2, 1, 2, 4)) = 010100012 = 64 + 16 + 1 =
81.
Computing a hash value takes O(m) time.
Each linked list is maintained in sorted order; since the insertions are monotone,
we can simply append insertions to the end of the list, and so insertion time is
dominated by the time to compute the hash value, O(m) time. Note that the
signature and λp is stored with the prime for use by the fetch operation.
The fetch operation computes the hash value, and then scans the linked list,
extracting primes (in order) with matching signature. If the signature to match is
binary (which occurs roughly half the time), we expect about half of the primes
in the list to match, so constructing the resulting array s[ ] takes time linear in
m multiplied by the number of primes fetched. Note that λ values are brought
along as well. Amortized over the full computation, fetch will average time that is
linear in the size of the data returned. Also note that the average cost to compare
signatures that don’t match is constant.
The total space used by the data structure is an array of length 2m , plus one
linked list node for each prime ≤ B 1/3 . Each linked list node holds the prime, its λ
value, and its signature, for a total of O(2m log B + mB 1/3 ) bits, since each prime
is O(log B) bits.
2.5. Algorithm Outline. Now that all the pieces are in place, we can describe
the algorithm as a whole.
Given an input bound B, we compute the GCD/sieve cutoff X (this is discussed
in §3 below).
We then compute GCDs as discussed above using candidate k values ≤ X for
t = 2, 3, . . .. For some value of t, we’ll discover that there are no candidate k values
with t − 1 prime factors, at which point we are done with GCD computations.
Next we sieve as discussed above, using candidate k values > X for t = 2, 3, . . ..
Again, for some t value, we’ll discover there are no candidate k values with t − 1
prime factors, at which point we are done.
We wrote a total of eight programs; GCD programs for t = 2, 3, 4 and sieving
programs for t = 2, 3, 4, 5, 6. Also, we did not bother to use the wheel and signature
sieving, relying only on λ sieving, for t > 3.
3. Algorithm Analysis
3.1. Model of Computation. We assume a RAM model, with potentially infinite
memory. Arithmetic operations on integers of O(log B) bits (single-precision) and
basic array indexing and control operations all are unit cost. We assume fast, FFTbased algorithms for multiplication and division of large integers. For n-bit inputs,
multiplication and division cost M (n) = O(n log n log log n) operations [14]. From
STRONG PSEUDOPRIMES TO TWELVE PRIME BASES
9
this, an FFT-based GCD algorithm takes O(M (n) log n) = O(n(log n)2 log log n)
operations; see, for example, [17].
Q
3.2. Helpful
Number Theory Results. Let our modulus q = ai ∈ν ai so that
Q
φ(q) = ai ∈ν φ(ai ); each base ai is an odd prime with φ(ai ) = a1 − 1, except for
a1 = 8 with φ(a1 ) = 4. We use our two propositions from §2.3, together with
quadratic reciprocity to obtain a list of (ai − 1)/2 residue classes for each odd
prime ai and two residue classes modulo 8. The total number of residue classes is
2−m φ(q). A prime we are searching for with a matching signature must lie in one
of these residue classes.
Let us clarify that when we state in sums and elsewhere that a prime p has
signature equivalent to σ, we mean that the quadratic character (p/ai ) is consistent
with σ as stated in Propositions 2.5 and 2.6. We write σ ≡ σp to denote this
consistency under quadratic character, which is weaker than σ = σp .
Q
Lemma 3.1. Given a signature σ of length m, and let q = i ai as above. Let
> 0. If q < x1− , then number of primes p ≤ x with σp ≡ σ is at most
x
.
O
2m log x
Proof. This follows almost immediately from the Brun-Titchmarsh Theorem. See,
for example, Iwaniec [6]. We simply sum over the relevant residue classes mod q
based on quadratic character using Propositions 2.5 and 2.6.
We will choose m proportional to log B/ log log B so that q = B c for some
constant 0 < c < 1.
Lemma 3.2. We have
X
p≤x,σp ≡σ
1
= O(2−m log log x).
p
This follows easily from Lemma 3.1 by partial summation. We could make this
tighter by making use of Theorem 4 from [9], and observe that the constants C(q, a)
(in their notation) cancel with one another in a nice fashion when summing over
different residue classes modulo q. See also section 6 of that paper.
Let πt,σ (x) denote the number of integers ≤ x with exactly t distinct prime
divisors, each of which has signature equivalent to σ of length m.
Lemma 3.3.
πt,σ (x)
x(log log x)t−1
.
2tm log x
Proof. Proof is by induction on t. t = 1 follows from Lemma 3.1.
For the general case, we sum over the largest prime divisor p of n,
X
πt,σ (x) =
πt−1,σ (x/p)
p≤x,σp =σ
X
p≤x,σp ≡σ
x(log log(x/p))t−2
p2(t−1)m log(x/p)
x(log log x)t−1
.
2tm log x
10
JONATHAN SORENSON AND JONATHAN WEBSTER
We used Lemma 3.2 for the last step. (See also the proof of Theorem 437 in [5,
§22.18].)
Let λp,m denote the value of λp for the prime p using a vector ν with the first
m prime bases. We need the following lemma from [10, Cor. 2.4]:
Lemma 3.4. There exists a constant τ > 0 such that
1
X
p≤x
λp,m
x1/(m+1)
.
(log x)1+τ /(m+1)
Using this, we obtain the following.
Theorem 3.5. Let 0 < θ < 1. Then
X
1
1
θ−1/(m+1)
.
pλp,m
x
log x
θ
x <p≤x
Proof. We have
X
xθ <p≤x
1
pλp,m
1
xθ
≤
X
xθ <p≤x
1
λp,m
1
x1/(m+1)
.
xθ (log x)1+τ /(m+1)
Note that τ > 0.
3.3. A Conjecture. In order to prove our B 2/3+o(1) running time bound for t > 2,
we make use of the conjecture below. Let K = K(t, σ) be the set of integers with
exactly t distinct prime divisors each of which has signature matching σ. In other
words, πt,σ (x) counts integers in K bounded by x.
Conjecture 3.6. Let B, σ, t, m, ν be as above. Then
X
k≤x,k∈K(t,σ)
1
xo(1) .
λk,m
Theorem 3.7. Assuming Conjecture 3.6 is true, we have
X
xθ ≤k≤x,k∈K(t,σ)
1
x−θ+o(1) .
kλk,m
Proof. The proof is essentially the same as that of Theorem 3.5. Simply substitute
Conjecture 3.6 for Lemma 3.4.
3.4. Proof of Theorem 1.2. We now have the pieces we need to prove our running
time bound.
As above, each pseudoprime candidate we construct will have the form n = kpt ,
where k is the product of t − 1 distinct primes with matching signature. Again,
ν = (2, 3, 5, . . .) is our base vector of small primes of length m.
STRONG PSEUDOPRIMES TO TWELVE PRIME BASES
11
3.4.1. t = 2. In this case k is prime.
GCD Computation. For each prime k = p ≤ X we perform a GCD computation as described in §2.2. The bottleneck of this computation is computing GCDs
of O(p)-bit integers. This gives a running time proportional to
X
M (p) log p π(X)X(log X)2 log log X
p≤X
X 2 log X log log X.
Sieving Computation. For each prime k = p with X < p < B/p, we sieve
the interval [p + 1, B/p] for primes that are ≡ 1 mod λp . We also employ signature
sieving, but for the simplicity of analysis, we omit that for now. Using the methods
from [15], we can sieve an arithmetic progression of length ` in O(` log B) operations.
We do not need proof of primality here, so a fast probabilistic test works fine. This
gives a running time proportional to
X
X B log B
1
B log B
.
pλp
pλp
X<p≤B/X
X<p≤B/p
At this point we know X 2 ≤ B and, to keep the GCD and sieving computations
balanced, X ≥ B 1/4 , say. This means Theorem 3.5 applies; we set xθ = X and
x = B/X to obtain
B log B
X
X<p≤B/X
1
pλp
=
B log B
(B/X)1/(m+1)
X log(B/X)
B
(B/X)1/(m+1)
X
(B/X)1+o(1)
assuming that m → ∞ with B.
Minimizing X 2+o(1) + (B/X)1+o(1) implies X = B 1/3 and gives the desired
running time of B 2/3+o(1) . This completes our proof.
.
3.4.2. t > 2. In this case k is composite.
GCD Computation. For t > 2 we construct integers k = rp for computing
GCDs, with r consisting of exactly t − 2 prime divisors less than p, with signatures
matching σp . For each prime p, we perform a hash table lookup and fetch the list of
such primes; this is Step 5 of Algorithm 3. This overhead cost is O(m)+O(πσp (p)) =
O(m + 2−m p/ log p). Summing this over all primes p ≤ X gives O(2−m X 2 / log X).
Next, for each prime p ≤ X we construct at most πt−2,σp (X/p) values for r (Step
7 of Algorithm 3), and using Lemma 3.3 and multiplying by the cost of computing
the GCD, this gives us
t−2
X
log log X
2
(3.1)
πt−2,σp (X/p) · M (X) log X X
log X
2m
p≤X
for the total running time.
Sieving Computation. Again, the main loop enumerates all choices for the
second-largest prime p = pt−1 . First, we construct all possible k-values with k > X,
k < B/p, and p | k, so k/p = r = p1 · · · pt−2 with all the pi < pi+1 and pt−2 < p.
We also have pt−2 < B 1/3 . This implies p > X 1/(t−1) .
12
JONATHAN SORENSON AND JONATHAN WEBSTER
For a given p, fetching the list of primes below u := min{p, B/p2 } with matching
signatures
takes O(m + u/(2m log u)) time (Algorithm 3 Step 5). Summing over all
√
p ≤ B, splitting the sum at p = B 1/3 , this takes 2−m B 2/3+o(1) time.
As above, we write r = k/p. We claim that the total number of k values is
2−m(t−2) B 2/3+o(1) . Let u be as above. There are at most πt−2,σp (u) values of r for
each p. Again, splitting this at B 1/3 , we have
X
X
p(log log p)t−3
πt−2,σp (p)
2(t−2)m log p
1/(t−1)
1/3
1/(t−1)
1/3
X
<p≤B
X
<p≤B
2−m(t−2) B 2/3+o(1) .
and
X
√
B 1/3 <p≤ B
πt−2,σp (B/p2 )
X
√
B 1/3 <p≤ B
B(log log B)t−3
p2 2(t−2)m log B
1
B(log log B)t−3
B 1/3 log B 2(t−2)m log B
2−m(t−2) B 2/3+o(1) .
This covers work done at Step 7 of Algorithm 3.
Finally, the cost of sieving is at most
X B log B
X
√
X 1/(t−1) <p≤ B
r
rpλrp
.
If we use Theorem 3.7, based on our conjecture, and using the lower bound X ≤
k = rp, this leads to the bound
t−2 1+1/(m+1)+o(1)
log log B
B
.
(3.2)
2m
X
Without the conjecture, we use Lemma 3.2 together with the argument outlined in
Hardy and Wright [5] in deriving (22.18.8), we obtain that
X 1 log log B t−2
∼
.
r
2m
r
We then use
1
1
≤
.
λrp
λp
As above, this leads to the bound
t−2 1+1/(m+1)
log log B
B
(3.3)
.
2m
X 1/(t−1)
To balance the cost of GCDs with sieving, we want to balance (3.1) with (3.2) or
(3.3), depending on whether we wish to assume our conjecture or not. Simplifying
a bit, this means balancing X 2+o(1) with either (B/X)1+o(1) or B 1+o(1) /X 1/(t−1) .
The optimal cutoff point is then either X = B 1/3 under the assumption of Conjecture 3.6, or X = B (t−1)/(2t−1) unconditionally, for a total time of
t−2
log log B
B 2/3+o(1)
2m
STRONG PSEUDOPRIMES TO TWELVE PRIME BASES
13
with our conjecture, or
log log B
2m
t−2
1
B 1− 2t−1 +o(1)
without. We have proven the following.
Theorem 3.8. Assuming Conjecture 3.6, our algorithm takes B 2/3+o(1) time to
find all integers n ≤ B that are strong pseudoprimes to the first m prime bases, if
m → ∞ with B.
If we choose m so that q is a fractional power of B, then Lemma 3.3 implies
that there is a constant c > 0 such that if t > c log log B, then there is no work
to do. This explains why our computation did not need to go past t = 6. It also
explains why, in practice, the t = 2 case is the bottleneck of the computation, and
is consistent with the implications of our conjecture that the work for each value
of t decreases as t increases.
4. Implementation Details
In this final section, we conclude with some details from our algorithm implementation.
4.1. Strong Pseudoprimes Found. In addition to our main results for ψ12 and
ψ13 , we discovered the following strong pseudoprimes. This first table contains
pseudoprimes found while searching for ψ12
n Factorization
3825123056546413051 747451· 149491 ·34233211
230245660726188031 787711· 214831· 1360591
360681321802296925566181 424665351661·849330703321
164280218643672633986221 286600958341·573201916681
318665857834031151167461 399165290221·798330580441
7395010240794120709381 60807114061·121614228121
164280218643672633986221 286600958341·573201916681
This second table contains pseudoprimes found while verifying ψ13 .
n Factorization
318665857834031151167461 399165290221·798330580441
2995741773170734841812261 1223875355821·2447750711641
667636712015520329618581 577770158461·1155540316921
3317044064679887385961981 1287836182261·2575672364521
3317044064679887385961981 1247050339261·2494100678521
552727880697763694556181 525703281661·1051406563321
360681321802296925566181 424665351661·849330703321
7395010240794120709381 60807114061·121614228121
3404730287403079539471001 1304747157001·2609494314001
164280218643672633986221 286600958341·573201916681
4.2. Hash Table Datastructure. It is natural to wonder how much time is
needed to manage the hash table datastructure from §2.4. Specifically, we measured the time to create the table when inserting all primes up to a set bound.
For this measurement, we used a vector ν with the first 8 prime bases. We also
14
JONATHAN SORENSON AND JONATHAN WEBSTER
Figure 1. GCD timings
measured how long it took, for each prime up to the bound listed, to also perform
the fetch operation, which returns a list of primes with matching signature with λ
values.
Largest Prime Table Entries Creation Time Fetching Time
106
78490
1.13
2.58
107
664571
11.49
2:44.49
108
5761447
1:56.22
3:30:50.2
The times are given in seconds, or minutes:seconds, or hours:minutes:seconds as
appropriate.
We felt the data structure was a success, in that it was fast and small enough
that it was not noticed in the larger computation.
4.3. Almost-Linear GCDs. In Figure 1, we present the times to perform Algorithm 1 on selected prime values for k (so this is for the t = 2 case). The data
points were obtained by averaging the time needed to carry out Algorithm 1 for the
first ten primes exceeding n · 107 for 1 ≤ n ≤ 35. Averaging was required because
the size of the h values is not uniformly increasing in k. This explains the variance
in the data; for example, the first ten primes after 35 · 107 happened to have smaller
h values on average than the first ten primes after 34 · 107 .
It should be clear from the data that our GMP algorithm for computing GCDs
was using an essentially linear-time algorithm. Note that if we chose to extend the
computation to large enough k values, memory would become a concern and we
would expect to see paging/thrashing degrade performance.
4.4. GCD/Sieving Crossover. In Figure 2, we present a comparison of the timings for computing GCDs (Algorithm 1, diamonds on the graph) with signature
sieving (Algorithm 2, squares on the graph). For this graph, we are looking at
choosing the crossover point X, and again, we are focusing on the t = 2 case.
• One would expect that the signature sieving curve should behave as an
inverse quadratic; the time to sieve for one prime k = p is proportional to
B log B
pλp
STRONG PSEUDOPRIMES TO TWELVE PRIME BASES
15
Figure 2. Signature Sieving and GCD timing
and we expect λp ≈ p − 1 most of the time. However, utilizing signatures
obscures this to some degree, since the algorithm cannot make use of all
prime bases if they divide λ, hence the minor variations in the curve. Let
us elaborate this here.
The data points for signature sieving in Figure 2 represent the average
sieve time for the first 50 primes past n · 107 for 10 ≤ n ≤ 60. There is
a lot of variance in these timing because of the inclusion and exclusion of
primes from the signature depending if they are relatively prime to λ. For
example, p = 174594421 has λ = 22 · 33 · 5 · 11 · 13 · 17 · 19 and therefore has
few primes from the base set to use in signature sieving. The inability to
add extra primes shows up in the timing data; a careful inspection of the
data shows a strange jump when primes transition from 28 · 107 to 29 · 107 .
The data points to a steady decrease, then an almost two fold increase in
time followed by a steady decrease again. We believe this is because, on
average, one less base prime may be used in signature sieving. Using one
less base prime results in approximately twice as much work.
• In the computation for ψ12 , our actual crossover point X was 30 · 107 , even
though the timing data would suggest the optimal crossover point is around
12 · 107 . From a cost-efficiency point of view, we chose poorly. However,
the choice was made with two considerations. One is that the program
to compute GCDs (Algorithm 1) is simple to write, so that program was
up and running quickly, and we let it run while we wrote the sieving code
(Algorithm 2). Second is that, initially, we did not know which ψm value
we would ultimately be able to compute. Since the results from GCD
computations apply to all larger values of m, we opted to err in favor of
computing more GCDs.
4.5. Signature Sieving. Figure 3 shows the impact that signature sieving makes.
Here the squares in the graph give λ-sieving times with no signature information
16
JONATHAN SORENSON AND JONATHAN WEBSTER
Figure 3. Signature Sieving and λ-Sieving
used, and the diamonds show the same λ-sieving work done while taking advantage
of signature information using the space-saving wheel. Since there is relatively little
variance involved in λ-sieving, each point represents the first prime after n · 107 for
10 ≤ n ≤ 60. On the same interval signature sieving practically looks like it takes
constant time compared to λ-sieving. If we had opted to not incorporate signature
sieving in, then the expected crossover point given these timings would occur when
the primes are around 42 · 107 .
References
1. Eric Bach, Explicit bounds for primality testing and related problems, Math. Comp. 55 (1990),
no. 191, 355–380. MR 1023756 (91m:11096)
2. Daniel Bleichenbacher, Efficiency and security of cryptosystems based on number theory,
Ph.D. thesis, Swiss Federal Institute of Technology Zurich, 1996.
3. François G. Dorais and Dominic Klyve, A Wieferich prime search up to 6.7 × 1015 , J. Integer
Seq. 14 (2011), no. 9, Article 11.9.2, 14. MR 2859986
4. Brian Dunten, Julie Jones, and Jonathan Sorenson, A space-efficient fast prime number sieve,
Inform. Process. Lett. 59 (1996), no. 2, 79–84. MR 1409956 (97g:11141)
5. G. H. Hardy and E. M. Wright, An introduction to the theory of numbers, 5th ed., Oxford
University Press, 1979.
6. Henryk. Iwaniec, On the Brun-Titchmarsh theorem, J. Math. Soc. Japan 34 (1982), no. 1,
95–123.
7. Gerhard Jaeschke, On strong pseudoprimes to several bases, Math. Comp. 61 (1993), no. 204,
915–926. MR 1192971 (94d:11004)
8. Yupeng Jiang and Yingpu Deng, Strong pseudoprimes to the first eight prime bases, Math.
Comp. (2014), 1–10 (electronic).
9. A. Languasco and A. Zaccagnini, A note on Mertens’ formula for arithmetic progressions,
Journal of Number Theory 127 (2007), no. 1, 37 – 46.
10. Francesco Pappalardi, On the order of finitely generated subgroups of q*(mod p) and divisors
of p1, Journal of Number Theory 57 (1996), no. 2, 207 – 222.
11. Francesco Pappalardi, On the r-rank Artin conjecture, Math. Comp. 66 (1997), no. 218,
853–868. MR 1377664 (97f:11082)
12. Carl Pomerance, J. L. Selfridge, and Samuel S. Wagstaff, Jr., The pseudoprimes to 25 · 109 ,
Math. Comp. 35 (1980), no. 151, 1003–1026. MR 572872 (82g:10030)
STRONG PSEUDOPRIMES TO TWELVE PRIME BASES
17
13. Michael O. Rabin, Probabilistic algorithm for testing primality, J. Number Theory 12 (1980),
no. 1, 128–138. MR 566880 (81f:10003)
14. A. Schönhage and V. Strassen, Schnelle Multiplikation großer Zahlen, Computing (Arch.
Elektron. Rechnen) 7 (1971), 281–292, MR 45 #1431. MR 45 #1431
15. Jonathan P. Sorenson, The pseudosquares prime sieve, Algorithmic number theory, Lecture Notes in Comput. Sci., vol. 4076, Springer, Berlin, 2006, pp. 193–207. MR 2282925
(2007m:11168)
16. Jonathan P. Sorenson, Sieving for pseudosquares and pseudocubes in parallel using doublyfocused enumeration and wheel datastructures, Proceedings of the 9th International Symposium on Algorithmic Number Theory (ANTS-IX) (Nancy, France) (Guillaume Hanrot, Francois Morain, and Emmanuel Thomé, eds.), Springer, July 2010, LNCS 6197, ISBN 978-3-64214517-9, pp. 331–339.
17. Damien Stehlé and Paul Zimmermann, A binary recursive GCD algorithm, Sixth International
Algorithmic Number Theory Symposium (Burlington, Vermont, USA) (Duncan Buell, ed.),
Springer, June 2004, LNCS 3076, pp. 411–425. MR MR2138011 (2006e:11194)
18. Zhenxiang Zhang, Two kinds of strong pseudoprimes up to 1036 , Math. Comp. 76 (2007),
no. 260, 2095–2107 (electronic). MR 2336285 (2008h:11114)
Department of Computer Science and Software Engineering, Butler University
E-mail address: [email protected]
Department of Mathematics and Actuarial Science, Butler University
E-mail address: [email protected]
| 8cs.DS
|
arXiv:1705.08320v1 [cs.AI] 23 May 2017
Explaining Transition Systems
through Program Induction
Svetlin Penkov
School of Informatics
The University of Edinburgh
[email protected]
Subramanian Ramamoorthy
School of Informatics
The University of Edinburgh
[email protected]
Abstract
Explaining and reasoning about processes which underlie observed black-box
phenomena enables the discovery of causal mechanisms, derivation of suitable
abstract representations and the formulation of more robust predictions. We propose
to learn high level functional programs in order to represent abstract models which
capture the invariant structure in the observed data. We introduce the π-machine
(program-induction machine) – an architecture able to induce interpretable LISPlike programs from observed data traces. We propose an optimisation procedure for
program learning based on backpropagation, gradient descent and A* search. We
apply the proposed method to three problems: system identification of dynamical
systems, explaining the behaviour of a DQN agent and learning by demonstration
in a human-robot interaction scenario. Our experimental results show that the πmachine can efficiently induce interpretable programs from individual data traces.
1
Introduction
Learning models of transition systems has been a core concern within machine learning, with
applications ranging from system identification of dynamical systems [1] and inference of human
choice behaviour [2, 3] to reverse engineering the behaviour of a device or computer program from
observations and traces [4]. With the increasing use of these learnt models in the inner loops of
decision making systems, e.g., in robotics and human-machine interfaces, it has become necessary
to ensure not only that these models are accurate predictors of behaviour, but also that their causal
mechanisms are exposed to the system designer in a more interpretable manner. There is also the
need to explain the model in terms of counterfactual reasoning [5], e.g., what would we expect the
system to do if a certain variable were changed or removed, or model checking [6] of longer term
properties including safety and large deviations in performance. We address these needs through a
program induction based framework for explainability.
We propose to learn high level functional programs in order to represent abstract models which
capture the invariant structure in the observed data. Recent works have demonstrated the usefulness
of program representations in capturing human-like concepts [7]. Used in this way, programbased representations boost generalisation and enable one-shot learning. Also, and arguably more
importantly, they are significantly more amenable to model checking and human interpretability.
In this paper, we introduce the π-machine (program-induction machine), an architecture which is able
to induce LISP-like programs from observed transition system data traces in order to explain various
phenomena. Inspired by differentiable neural computers [8, 9], the π-machine, as shown in Figure 1,
is composed of a memory unit and a controller capable of learning programs from data by exploiting
the scalability of stochastic gradient descent. However, the final program obtained after training is
not an opaque object encoded in the weights of a controller neural network, but a LISP-like program
which provides a rigorous and interpretable description of the observed phenomenon. A key feature
of our approach is that we allow the user to specify the properties they are interested in understanding
as well as the context in which the data is to be explained - by providing a set of predicates of interest.
By exploiting the equivalence between computational graphs and functional programs we describe a
hybrid optimisation procedure based on backpropagation, gradient descent, and A* search which is
used to induce programs from data traces.
We evaluate the performance of the π-machine on three different problems. Firstly, we apply it to data
from physics experiments and show that it is able to induce programs which represent fundamental
laws of physics. The learning procedure has access to relevant variables, but it does not have any
other prior knowledge regarding physical laws which it has discovered in the same sense as in [1]
although far more computationally tractably. We then study the use of the proposed procedure in
explaining control policies learnt by a deep Q-network (DQN). Starting from behaviour traces of
a reinforcement learning agent that has learnt to play the game of Pong, we demonstrate how the
π-machine learns a functional program to describe that policy. Finally, we consider the domain of
learning by demonstration in human-robot interaction, where the π-machine successfully induces
programs which capture the structure of the human demonstration. In this domain, the learnt program
plays a key role in enabling the grounding of abstract knowledge (e.g., in natural language commands)
in the embodied sensory signals that robots actually work with, as in [10].
2
Related work
Explainability and interpretability. The immense success of deep neural network based learning
systems and their rapid adoption in numerous real world application domains has renewed interest in
the interpretability and explainability of learnt models 1 . There is recognition that Bayesian rule lists
[11, 12], decision trees and probabilistic graphical models are interpretable to the extent that they
impose strong structural constraints on models of the observed data and allow for various types of
queries, including introspective and counterfactual ones. In contrast, deep learning models usually are
trained ‘per query’ and have numerous parameters that could be hard to interpret. Zeiler and Fergus
[13] introduced deconvolutional networks in order to visualise the layers of convolutional networks
and provide a more intuitive understanding of why they perform well. Similar approaches can be
seen in [14, 15], in the context of autonomous driving. Zahavy et al. [16] describe Semi-Aggregated
Markov Decision Process (SAMDP) in order to analyse and understand the behaviour of a DQN
based agent. Methods for textual rationalisation of the predictions made by deep models have also
been proposed [17–19]. While all of these works provide useful direction, we need many more
methods, especially generic approaches that need not be hand-crafted to explain specific aspects of
individual models. In this sense, we follow the model-agnostic explanation approach of Ribeiro et al.
[20], who provide “textual or visual artefacts” explaining the prediction of any classifier by treating
it as a black-box. Similarly to the way in which [20] utilise local classifiers composed together to
explain a more complex model, we present an approach to incrementally constructing functional
programs that explain a complex transition system from more localised predicates of interest to the
user.
The π-machine treats the process which has generated the observed data as a black-box and attempts
to induce LISP-like program which can be interpreted and used to explain the data. We show that the
proposed method can be applied both to introspection of machine learning models and to the broader
context of autonomous agents.
Program learning and synthesis. Program learning and synthesis has a long history, with the
long-standing challenge being the high complexity deriving from the immense search space. Following classic and pioneering work such as by Shapiro [21] who used inductive inference in a logic
programming setting, others have developed methods based on a variety of methods ranging from
SAT solvers [22] to genetic algorithms [1], which tend to scale poorly hence often become restricted
to a narrow class of programs. Recently, deep neural networks have been augmented with a memory
unit resulting in models similar to the original von Neumann architecture. These models can induce
programs through stochastic gradient descent by optimising performance on input/output examples
[8, 9, 23] or synthetic execution traces [24–26]. Programs induced with such neural architectures
are encoded in the parameters of the controller network and are, in general, not easily interpretable
(particularly from the point of view of being able to ask counterfactual questions or performing
1
See, for instance, the end user concerns that motivate the DARPA Explainable AI Programme: http:
//www.darpa.mil/program/explainable-artificial-intelligence
2
Figure 1: Overall architecture of the π-machine. The current best candidate solution (1) is used to
propose new, structurally more complex candidates (2). Each one of the new candidate programs ρ
is optimised (3) through gradient descent by comparing its execution trace to the observation trace
(4). The observation trace in this case is a demonstration of a tower building task. During execution,
the program has access to memory (5) which stores both state variables and induced parameters.
All new candidate programs are scored based on their performance and complexity and are inserted
in the candidate solutions priority queue (6). Once the execution trace of a candidate matches the
observation trace the final solution is returned (7).
model checking). Interestingly, paradigms from functional programming such as pure functions
and immutable data structures have been shown to improve performance of neural interpreters [27].
Another approach is to directly generate the source code of the output program which yields consistent
high level programs. Usually, these types of approaches require large amounts of labelled data - either
program input/output examples [28, 29] or input paired with the desired output program code [30].
Determining how many input/output examples or execution traces are required in order to generalise
well is still an open research problem. However, in this paper, we focus attention more on the
explanatory power afforded by programs rather than on the broader problems of generalisation in
the space of programs. While these characteristics are of course related, we take a view similar to
that of [20], arguing that it is possible to build from locally valid program fragments which provide
useful insight into the black-box processes generating the data. By combining gradient descent and
A* search the π-machine is able to learn informative and interpretable high-level LISP-like programs,
even just from a single observation trace.
3
Problem definition
Consider the labelled transition system Ω(S, A, δ) where S is a non-empty set of states, A is a nonempty set of actions, each parametrised by θ ∈ RD , and δ : S ×A → S is the state transition function.
We define an observation trace T as a sequence of observed state-action pairs (st , at (θt )) ∈ S × A
generated by the recursive relationship st+1 = δ (st , at ( θt )) for 1 ≤ t ≤ T . We are interested in
inducing a LISP-like functional program ρ which when executed by an abstract machine is mapped
to an execution trace Tρ such that Tρ and T are equivalent according to an input specification.
We represent the abstract machine as another labelled transition system Π(M, I, ε) where M is the
set of possible memory state configurations, I is the set of supported instructions and ε : M×I → M
specifies the effect of each instruction. We consider two types of instructions – primitive actions which
emulate the execution of a ∈ A or arithmetic functions f ∈ F such that I = A ∪ F. Furthermore, a
set of observed state variables Mv ⊆ S, which vary over time, are stored in memory together with a
set of induced free parameters Mp . The variables in Mv form a context which the program will be
built on. A custom detector Dv , operating on the raw data stream, could be provided for each variable,
thus enabling the user to make queries with respect to different contexts and property specifications.
3
The execution of a program containing primitive actions results in a sequence of actions. Therefore,
we represent a program ρ as a function which maps a set of input variables xv ⊂ Mv and a set of
free parameters xp ⊂ Mp to a finite sequence of actions â1 (θˆ1 ), . . . âT 0 (θˆT 0 ). We are interested in
inducing a program which minimises the total error between the executed and the observed actions:
min(T,T 0 )
L(ρ) =
X
σact (ât , θˆt , at , θt ) + σlen (T, T 0 )
(1)
t=1
The error function σact determines the difference between two actions, while σlen compares the
lengths of the generated and observed action traces. By providing the error functions σact and σlen
one can target different aspects of the observation trace to be explained as they specify when two
action traces are equivalent.
4
Method
The proposed program induction procedure is based on two major steps. Firstly, we explain how
a given functional program can be optimised such that the loss L(ρ) is minimised. Secondly, we
explain how the space of possible program structures can be searched efficiently by utilising gradient
information. An architectural overview of the π-machine is provided in Figure 1.
4.1
Program optimisation
Functional programs as computational graphs. Neural networks are naturally expressed as
computational graphs which are the most fundamental abstraction in computational deep learning
frameworks [31–33]. Optimisation within a computational graph is usually performed by pushing the
input through the entire graph in order to calculate the output (forward pass) and then backpropagating
the error signal to update each parameter (backward pass). A key observation for the development
of the π-machine is that computational graphs and functional programs are equivalent as both
describe arbitrary compositions of pure functions applied to input data. For example, Figure 2
shows how a logistic regression classifier can be represented both as a computational graph and as a
functional program. Therefore, similarly to a computational graph, a functional program can also be
optimised by executing the program (forward pass), measuring the error signal and then performing
backpropagation to update the program (backward pass).
Figure 2: Logistic regression expressed as: computational graph (left); functional program (right).
Forward pass. When a program is executed it is interpreted to a sequence of instructions
i1 , . . . , in ∈ I which are executed by recursively calling ε(. . . ε(ε(M1 , i1 ), i2 ) . . . , in ). M1 is
the initial memory state initialised with the observed variables from s1 and any induced parameters.
The π-machine keeps a time counter t which is initialised to 1 and is automatically incremented
whenever a primitive action instruction is executed. If the instruction ik is a primitive action, ik ∈ A,
then the π-machine automatically sets ât = ik and invokes the error function σact (ât , θˆt , at , θt ),
where θˆt has been calculated by previous instructions. If the error is above a certain threshold emax
the program execution is terminated and the backward pass is initiated. Otherwise, the time counter
is incremented and the values of the variables in Mv are automatically updated to the new observed
state. Essentially, the π-machine simulates the execution of each action reflecting any changes it
has caused in the observed state. Alternatively, if the currently executed instruction ik is a function,
ik ∈ F, then the resulting value is calculated and ik , together with its arguments, is added to a
detailed call trace χ maintained by the π-machine. Importantly, each function argument is either
a parameter or a variable read from memory at time t or the result of another function. All this
information is kept in χ which eventually contains the computational tree for the entire program.
4
Backward pass. The gradients of the loss function L(ρ) with respect to the program inputs xv and
xp are required to perform a gradient descent step. Crucially, programs executed by the π-machine
are automatically differentiated. The π-machine performs reverse-mode automatic differentiation,
similarly to Autograd [34], by traversing the call trace χ, and post-multiplying Jacobian matrices.
We assume that the Jacobian matrix with respect to every input argument of any function f ∈ F or
any specified error function σact is known a priori. Let f ∈ F be a function whose output needs to
be differentiated with respect to the input arguments. There are three types of derivatives, which need
to be considered in order to traverse backwards the entire tree of computations:
1. Let g ∈ F, then ∂f
∂g is the Jacobian matrix of f with respect to the output of g and can be
directly calculated.
2. Let p ∈ xp , then the gradient ∂f
∂p is calculated by multiplying the corresponding Jacobian
matrix of f with the value of p.
3. Let v ∈ xv , then the gradient
∂f
∂v
t=tr
is calculated by multiplying the corresponding
Jacobian of f with the value of the variable at the time it was read from memory tr .
Gradient descent step. Once the gradient ∇p L(ρ) of the loss function with respect to each input
parameter p ∈ xp is calculated we utilise AdaGrad [35] to update the values of all parameters after
each program execution. The gradient ∇v L(ρ) with respect to each input variable v ∈ xv is also
available. However, a variable cannot be simply updated in the direction of the gradient as it represents
a symbol, not just a value. Variables can only take values from memory which is automatically
updated according to the observation trace during execution. Nevertheless, the gradient provides
important information about the direction of change which we utilise to find variables that minimise
the loss. Whenever the memory state is automatically updated, a KD-tree is built for each type of
variable stored in memory. We assume that the variables in memory are real vectors with different
length. So, we represent the KD-tree which stores all D-dimensional variables in memory at time t
as KtD . If a d-dimensional variable v is to be optimised it is replaced with a temporary parameter
ptemp initialised with vt which is the value of v read from memory at the respective time step t.
The temporary parameter ptemp is also updated with AdaGrad [35]. After each descent step, the
nearest neighbour of the updated value p0temp is determined by querying the KD-tree with Ktd (p0temp ).
If the result of the query is a different d-dimensional variable u then the temporary parameter is
immediately set to ptemp = ut . As this often shifts the solution to a new region of the error space
the gradient history for all parameters p ∈ xp is reset. Eventually, when a solution is to be returned,
the temporary parameters are substituted with their closest variables according to the respective KtD .
The forward and backward passes are repeated until the error is below the maximum error threshold
emax or a maximum number of iterations is reached. After that the optimised program ρ∗ is scored
according to its error and complexity, and pushed to a priority queue holding potential solutions.
4.2
Structure search
We represent the space of possible program structures as a graph G = (T AST , E) where each node
Ti ∈ T AST is a valid program abstract syntax tree (AST). There is an edge from Ti to Tj if and only
if Tj can be obtained by replacing exactly one of the leaves in Ti with a subtree Ts of depth 1. The
program induction procedure always starts with an empty program. So, we frame structure search as
a path finding problem, solved through the use of A* search.
Score function. The total cost function we use is ftotal (ρ) = C(ρ) + L(ρ), where L(ρ) is the
loss function defined in equation (1) and C(ρ) is a function which measures the complexity of the
program ρ. C(ρ) can be viewed as the cost to reach ρ and L(ρ) as the distance to the desired goal.
The complexity function C(ρ) is the weighted sum of (i) maximum depth of the program AST; (ii) the
number of free parameters; (iii) the number of variables used by the program; the weights of which
we set to wC = [10, 5, 1]. These choices have the effect that short programs maximally exploiting
structure from the observation trace are preferred.
Neighbours expansion. When the current best candidate solution is popped from the priority
queue, we check if it matches the observation trace according to the input specification. If so, the
candidate can be returned as the final solution, otherwise it is used as a seed to propose new candidate
5
Figure 3: The π-machine applied to explaining the behaviour of a pendulum (top) and a linear
oscillator(bottom). The best 3 solutions for each system are shown in the middle.
solutions. Typically in A* search, all neighbouring nodes are expanded and pushed to the priority
queue, which is not feasible in our case, though. Therefore, we utilise the available gradients in order
to perform a guided proposal selection. Each leaf in the abstract syntax tree Tρ of a seed candidate
solution ρ corresponds to a parameter or a variable. According to the definition of G we need to
select exactly 1 leaf to be replaced with a subtree Ts of depth 1. We select leaf l ∈ Tρ according to:
l = arg max k∇x L(ρ)k2
(2)
x∈xp ∪xv
After that, all possible replacement subtrees are constructed. An AST subtree Ts of depth 1 represents
a function call. We prune the number of possible functions in F by ensuring type consistency. Each
leaf of Ts can be a parameter or a variable. So, all possible combinations are considered. New variable
leaves are initialised to a random variable with suitable type from memory, while new parameter
leaves are sampled from the multivariate normal distribution N (0, 0.1). As a result, if nf functions
are type compatible with l and each function takes na arguments at most, then there are 2na · nf
replacement subtrees, resulting in that many new candidates. All newly proposed candidates are
optimised in parallel, scored by ftotal and pushed to the priority queue.
5
Experimental results
In order to evaluate the effectiveness of the π-machine, we apply it to three different scenarios.
Firstly, we consider model learning for a physical system. We show that the π-machine successfully
induces basic physical laws from observational data. Secondly, we demonstrate how a program can
be induced in order to explain the behaviour of the policy learnt by a DQN network. This experiment
is based on our view that the core deep neural network based policy learner and the explanation layer
play complementary roles. As is well known by now, there are numerous advantages to performing
end-to-end policy learning, such as DQN-learning from raw video. Having done this, there is also a
need to explain the behaviour of the learnt policy with respect to user-defined properties of interest,
hence as a program defined in terms of user-defined detectors, Dv . Lastly, we apply the π-machine to
a learning-by-demonstration task, where demonstrated behaviour of physical object manipulation
is explained through a functional program that describes the abstract structure of that planning task.
All experiments are run on an Intel Core i7-4790 processor with 32GB RAM. The π-machine is
implemented in Clojure, which is a LISP dialect supporting powerful data structures and homoiconic
syntax. We will be releasing the code once this paper has been accepted for publication.
Given the nature of the experimental tasks, we use the following list of supported functions, F, for
all of our experiments: vector addition, subtraction and scaling.
5.1
Physical systems
The transition dynamics of a second order dynamical system is written as ẍ(t) = k1 x(t) + k2 ẋ(t),
where x(t) is the state of the system at time t and k1 , k2 are system coefficients. We have recreated
an experiment described in Schmidt and Lipson [1], where the authors show the learning of physical
laws associated with classical mechanical systems including the simple pendulum and linear oscillator.
A diagram of these two systems is shown in Figure 3 (left). We set A = {accel(θ)} where θ ∈ R for
both experiments. The observation trace for each system is generated by simulating the dynamics
6
Figure 4: The π-machine applied to explaining the behaviour of a DQN playing ATARI Pong. The
best 3 solutions are shown in the middle.
for 1s at 100Hz. We specify the action error function as σact = kθ̂ − θk2 and set σlen = 0. In
the pendulum experiment, x ∈ R is the angular position of the pendulum, while v = ẋ ∈ R is the
angular velocity. In the linear oscillator experiment, x and v are the linear position and velocity,
respectively.
The three best solutions found by the π-machine for each system are shown in Figure 3 (middle).
The best solution for each system correctly represents the underlying laws of motion. The program
describing the behaviour of the pendulum was induced in 18 iterations, while the linear oscillator
program needed 146 iterations. The total number of possible programs with AST depth of 2, given the
described experimental setup, is approximately 1.7 × 104 . The average duration of an entire iteration
(propose new programs, optimise and evaluate) was 0.6s. Schmidt and Lipson [1] achieve similar
execution times, but distributed over 8 quad core computers (32 cores in total). The experimental
results demonstrate that the π-machine can efficiently induce programs representing fundamental
laws of physics.
5.2
Deep Q-network
We consider explaining the behaviour of a DQN trained to play the ATARI Pong game. We are
interested in the question: how does the network control the position of the paddle in order to hit
the ball when it is in the right side of the screen. A diagram of the experimental setup is shown in
figure 4 (left). The behaviour of the DQN is observed during a single game. Since the environment is
deterministic, the state transition function, which generates the observation trace for this experiment,
is the policy π(s) that the DQN has learnt. We would like to explain the behaviour of the DQN in
terms of the position of the opponent, the ball and the DQN agent (so, not just in terms of RAM
memory values, for instance). Therefore, the observation trace contains those positions which are
extracted from each frame by a predefined detector. We set A = {move(θ)} where θ ∈ R and
represent the discrete actions of the network left, right, nop as move(1), move(−1), move(0)
respectively. We specify the action error function as σact = kθ̂ − θk2 and set σlen = 0.
The best 3 programs found by the π-machine are shown in Figure 4 (middle), where it took 38
iterations for the best one (average iteration duration 3.2s). By inspecting the second solution
it becomes clear that the neural network behaviour can be explained as a proportional controller
minimising the vertical distance between the agent and the ball. However, the best solution reveals
even more structure in the behaviour of the DQN. The coefficient in front of the agent position is
slightly larger than the one in front of the ball position which results in a small amount of damping
in the motion of the paddle. Thus, it is evident that the DQN not only learns the value of each
game state, but also the underlying dynamics of controlling the paddle. Furthermore, we have tested
the performance of an agent following a greedy policy defined by the induced program. In our
experiments over 100 games this agent achieved a score of 11.1(±0.17). This is not quite the score
of 18.9(±1.3) obtained by an optimised DQN, but it is better than human performance 9.3 [36].
This difference of course emanates from the predefined detector not capturing all aspects of what
the perceptual layers in DQN have learnt, so improved detector choices should yield interpretable
programs that also attain performance closer to the higher score of the black-box policy.
7
Figure 5: Program induced by the π-machine explaining a human demonstration of tower building
(induced for 27 iterations and 7.2s).
5.3
Learning by demonstration
Work in collaborative human-robot interaction [10] suggests that programmatic description of the
task enables robots to better ground symbols to their physical instances, improving their perceptual
capabilities. We consider a learning by demonstration scenario where a person demonstrates how to
build a tower in a virtual simulated 2D environment. A typical demonstration is depicted in Figure 5
(left). Our goal is to learn a program describing the demonstration such that it could later be utilised
by the robot.
We are interested in how a person moves the cubes through the entire demonstration. So, we set
A = {pick(θ), place(θ)}, where θ ∈ R2 is a 2D location. We specify the action and length error
functions as
(
kθ̂ − θk2 if â = a
σact (â, θ̂, a, θ) =
σlen (T 0 , T ) = (T 0 − T )2
(3)
dmax
otherwise
where dmax is the maximum distance within the simulated 2D environment. The states in the
observation trace contain the 2D position of every cube in the environment.
We tested the π-machine on 300 demonstrations and it successfully induced a program for each one
of them. On average, 67 iterations (average iteration duration 0.4s) were needed per demonstration.
The solution for one of the demonstrations is shown in Figure 5. From inspection, it can be seen that
the program not only describes which cube was picked up and put where, but also what the spatial
relations are when stacked. The π-machine induced and optimised a free parameter which is used to
encode the relation “above” as (+ loc [-0.05 1.17]). This type of information could speed up
task learning and improve the robot’s ability to understand its environment.
6
Discussion
The π-machine can be viewed as a framework for automatic network architecture design [37, 38],
as different models can be expressed as concise LISP-like programs (see Figure 2). Deep learning
methods for limiting the search space of possible programs, which poses the greatest challenge, have
been proposed [29], but how they can be applied to more generic frameworks such as the π-machine
is an open question. The specification of variable detectors not only addresses this issue, but enables
the user to make targeted and well grounded queries about the observed data trace. Such detectors
can also be learnt from raw data in an unsupervised fashion [39, 15].
7
Conclusion
We propose a novel architecture, the π-machine, for inducing LISP-like functional programs from
observed data traces by utilising backpropagation, stochastic gradient descent and A* search. The experimental results demonstrate that the π-machine can efficiently and successfully induce interpretable
programs from short data traces.
8
References
[1] Michael Schmidt and Hod Lipson. Distilling free-form natural laws from experimental data.
Science, 324(5923):81–85, 2009.
[2] Paul Glimcher. Foundations of Neuroeconomic Analysis. Oxford University Press, 2011.
[3] William Brendel and Sinisa Todorovic. Learning spatiotemporal graphs of human activities. In
Computer vision (ICCV), 2011 IEEE international conference on, pages 778–785. IEEE, 2011.
[4] Frits Vaandrager. Model learning. Commun. ACM, 60(2):86–95, January 2017.
[5] Léon Bottou, Jonas Peters, Joaquin Quinonero Candela, Denis Xavier Charles, Max Chickering,
Elon Portugaly, Dipankar Ray, Patrice Y Simard, and Ed Snelson. Counterfactual reasoning
and learning systems: the example of computational advertising. Journal of Machine Learning
Research, 14(1):3207–3260, 2013.
[6] Christel Baier and Joost-Pieter Katoen. Principles of Model Checking. The MIT Press, 2008.
[7] Brenden M. Lake, Ruslan Salakhutdinov, and Joshua B. Tenenbaum. Human-level concept
learning through probabilistic program induction. Science, 350(6266):1332–1338, 2015.
[8] Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint
arXiv:1410.5401, 2014.
[9] Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwińska, Sergio Gómez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou,
et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538
(7626):471–476, 2016.
[10] Svetlin Penkov, Alejandro Bordallo, and Subramanian Ramamoorthy. Physical symbol grounding and instance learning through demonstration and eye tracking. In Robotics and Automation,
2017 IEEE International Conference on, Singapore, June 2017.
[11] Benjamin Letham, Cynthia Rudin, Tyler H McCormick, David Madigan, et al. Interpretable
classifiers using rules and bayesian analysis: Building a better stroke prediction model. The
Annals of Applied Statistics, 9(3):1350–1371, 2015.
[12] Hongyu Yang, Cynthia Rudin, and Margo Seltzer. Scalable bayesian rule lists. arXiv preprint
arXiv:1602.08610, 2016.
[13] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In
European conference on computer vision, pages 818–833. Springer, 2014.
[14] Mariusz Bojarski, Philip Yeres, Anna Choromanska, Krzysztof Choromanski, Bernhard Firner,
Lawrence Jackel, and Urs Muller. Explaining how a deep neural network trained with end-to-end
learning steers a car. arXiv preprint arXiv:1704.07911, 2017.
[15] Jinkyu Kim and John Canny. Interpretable learning for self-driving cars by visualizing causal
attention. arXiv preprint arXiv:1703.10631, 2017.
[16] Tom Zahavy, Nir Ben-Zrihem, and Shie Mannor. Graying the black box: Understanding dqns.
arXiv preprint arXiv:1602.02658, 2016.
[17] Brent Harrison, Upol Ehsan, and Mark O Riedl. Rationalization: A neural machine translation
approach to generating natural language explanations. arXiv preprint arXiv:1702.07826, 2017.
[18] Lisa Anne Hendricks, Zeynep Akata, Marcus Rohrbach, Jeff Donahue, Bernt Schiele, and
Trevor Darrell. Generating visual explanations. In European Conference on Computer Vision,
pages 3–19. Springer, 2016.
[19] Tao Lei, Regina Barzilay, and Tommi Jaakkola. Rationalizing neural predictions. arXiv preprint
arXiv:1606.04155, 2016.
9
[20] Marco Ribeiro, Sameer Singh, and Carlos Guestrin. Why should i trust you?: Explaining
the predictions of any classifier. In ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining (KDD), 2016.
[21] Ehud Y. Shapiro. Algorithmic Program Debugging. MIT Press, Cambridge, MA, USA, 1983.
[22] Armando Solar-Lezama, Liviu Tancau, Rastislav Bodik, Sanjit Seshia, and Vijay Saraswat.
Combinatorial sketching for finite programs. ACM SIGOPS Operating Systems Review, 40(5):
404–415, 2006.
[23] Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to
transduce with unbounded memory. In Advances in Neural Information Processing Systems,
pages 1828–1836, 2015.
[24] Scott Reed and Nando De Freitas.
arXiv:1511.06279, 2015.
Neural programmer-interpreters.
arXiv preprint
[25] Jonathon Cai, Richard Shin, and Dawn Song. Making neural programming architectures
generalize via recursion. In International Conference on Learning Representations (ICLR),
April 2017.
[26] Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunsom. Program Induction by Rationale Generation:Learning to Solve and Explain Algebraic Word Problems. arXiv preprint
arXiv:1705.04146, 2017.
[27] John Kser, Marc Brockschmidt, Alexander Gaunt, and Daniel Tarlow. Differentiable functional
program interpreters. arXiv preprint arXiv:1611.01988v2, 2017.
[28] Jacob Devlin, Jonathan Uesato, Surya Bhupatiraju, Rishabh Singh, Abdel-rahman Mohamed,
and Pushmeet Kohli. Robustfill: Neural program learning under noisy i/o. arXiv preprint
arXiv:1703.07469, 2017.
[29] Matej Balog, Alexander L Gaunt, Marc Brockschmidt, Sebastian Nowozin, and Daniel Tarlow.
Deepcoder: Learning to write programs. arXiv preprint arXiv:1611.01989, 2016.
[30] Pengcheng Yin and Graham Neubig. A syntactic neural model for general-purpose code
generation. arXiv preprint arXiv:1704.01696, 2017.
[31] Seiya Tokui, Kenta Oono, Shohei Hido, and Justin Clayton. Chainer: a next-generation open
source framework for deep learning. In Proceedings of workshop on machine learning systems
(LearningSys) in the twenty-ninth annual conference on neural information processing systems
(NIPS), 2015.
[32] James Bergstra, Olivier Breuleux, Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, Guillaume
Desjardins, Joseph Turian, David Warde-Farley, and Yoshua Bengio. Theano: A cpu and gpu
math compiler in python. In Proc. 9th Python in Science Conf, pages 1–7, 2010.
[33] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro,
Greg S Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. Tensorflow: Large-scale
machine learning on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467,
2016.
[34] Dougal Maclaurin, David Duvenaud, and Ryan P Adams. Autograd: Effortless gradients in
numpy. In ICML 2015 AutoML Workshop, 2015.
[35] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning
and stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011.
[36] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G
Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al.
Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.
[37] Barret Zoph and Quoc Le. Neural architecture search with reinforcement learning. In International Conference on Learning Representations (ICLR), April 2017.
10
[38] Renato Negrinho and Geoff Gordon. Deeparchitect: Automatically designing and training deep
architectures. arXiv preprint arXiv:1704.08792, 2017.
[39] Marta Garnelo, Kai Arulkumaran, and Murray Shanahan. Towards deep symbolic reinforcement
learning. arXiv preprint arXiv:1609.05518, 2016.
11
| 2cs.AI
|
A practical efficient and effective method for the Hamiltonian
cycle problem that runs on a standard computer
Eric Lewin Altschuler*, Timothy J. Williams**
*Department of Physical Medicine and Rehabilitation Temple University School of
Medicine, 3401 N. Broad St., Philadelphia, PA 19140, USA
**Leadership Computing Facility, Argonne National Laboratory 9700 S. Cass Ave, Bldg.
240, Argonne, IL 60439, USA
1
Abstract
Given N cities and R < N2-N directed (unidirectional/one way) roads does there exist a
tour of all N cities stopping at each city exactly once using the given roads (a
Hamiltonian cycle)? This Hamiltonian cycle problem (HCP) is an NP-complete problem,
for which there is no known polynomial time solution algorithm. The HCP has important
practical applications, for example, to logistical problems. It was claimed that an
adiabatic quantum computer could solve an NP-complete problem faster than classical
algorithms, but claim appears to have been debunked. Here we demonstrate an algorithm
which runs on a standard computer that efficiently and effectively solves the HCP for at
least up to 500 cities: We first optimized a simulated annealing based algorithm used for
smaller sized HCP problems. Then we found that when a tour was deliberately inserted in
a list of otherwise randomly chosen roads, crucially, if "extra" random roads are added to
bring the total number of roads up to 0.58∗N∗loge N or more there is a 100% chance our
algorithm will find a HC, but conversely when a list of roads does not include a preinserted tour random roads have to be added until there are 0.9∗N∗loge N roads to have a
chance of finding a HC. We found similarly for a set of roads non-randomly chosen.
Thus, the presence of a HC in a set of roads induces "connectivity" throughout the roads
and a HC can be found with an insertion of a modest number of extra roads. Our
algorithm also shows that only weakly non-local information is needed to find an HCP
that is a global state.
Given N cities and R < N2-N directed (unidirectional/one way) roads does there exist a
tour of all N cities stopping at each city exactly once using the given roads (a
Hamiltonian cycle)? This Hamiltonian cycle problem (HCP) is an NP-complete problem,
for which there is no known polynomial time solution algorithm.1 The HCP has important
practical applications, for example, to logistical problems. While there was a claim2 that
an adiabatic quantum computer could solve an NP-complete problem faster than classical
algorithms, this claim appears to have been debunked3. Here we demonstrate an
algorithm which runs on a standard computer that efficiently and effectively solves the
HCP for at least up to 500 cities: We first optimized a simulated annealing4 based
algorithm5 used for smaller sized HCP problems. Then we found that when a tour was
deliberately inserted in a list of otherwise randomly chosen roads, crucially, if "extra"
random roads are added to bring the total number of roads up to 0.58*N*loge N or more
there is a 100% chance our algorithm will find a HC, but conversely when a list of roads
does not include a pre-inserted tour random roads have to be added until there are
0.9*N*loge N roads to have a chance of finding a HC. We found similarly for a set of
roads non-randomly chosen. Thus, the presence of a HC in a set of roads induces
"connectivity" throughout the roads and a HC can be found with an insertion of a modest
number of extra roads. Our algorithm also shows that only weakly non-local information
is needed to find a HC that is a global state.
Simulated annealing4 is an approach inspired by statistical mechanics that by analogy
views the values of a multivariable numeric problem as physical states of particles.
Simulated annealing is very useful practically for problems such as the traveling
2
salesperson problem and other so called NP-complete problems that have no known
polynomial time numeric solutions.
To employ simulated annealing one needs first to define a "cost" or energy function and
also "moves" to perturb the state of the system. We take a tour as a permutation of the
numbers 1,2,3…N. The tour length is the cost function we use. We assign the distance
between two cities 0 if there exists a road between the cities, and 1 if not. A tour of length
0 thus corresponds to a Hamiltonian cycle. When perturbing the energy function in using
simulated annealing ideally one wants to keep much of a configuration the same and only
change a small portion of the configuration. The idea is to retain the "good" parts of the
configuration and find and modify the high-energy parts. If a tour is not a Hamiltonian
cycle we make a new tour randomly using one of two “moves” to modify the tour: The
first move we use is the “transport” move of Lin and Kernighan6 (Figure 1a) which
transports a subsegment of the tour to a new location in the tour. We do not use the
“reversal” move of Lin and Kernighan6, because, as shown in Figure 1b, this would
change the direction of travel between potentially many cities, but since the roads are one
way a road between the same cities but in the opposite direction might not be in the given
set of roads so cannot be used to form a Hamiltonian cycle. Instead of the reversal move
we use a “swap” move5 (Figure 1c) that interchanges the positions of two cities in the
tour. At most this affects only four directional roads and thus keeps more of the original
tour intact than the reversal move. We now compute the length of the new tour. If the
length of the new tour is shorter than the prior tour, it is accepted as the new tour. Even if
the new tour is longer, we occasionally accept the longer tour using the Metropolis
algorithm.7
3
(a)
…
37
8
42
90
27
23
65
13
2
55
…
…
37
8
42
90
27
23
65
13
2
55
…
…
37
8
42
90
27
23
65
13
2
55
…
…
37
8
23
65
13
42
90
27
2
55
…
(b)
…
37
8
42
90
27
23
65
13
2
55
…
…
37
8
42
90
27
23
65
13
2
55
…
…
37
8
23
27
90
42
65
13
2
55
…
…
37
8
42
90
27
23
65
13
2
55
…
…
37
8
42
90
27
23
65
13
2
55
…
…
37
8
23
90
27
42
65
13
2
55
…
(c)
Figure 1: System perturbing moves for simulated annealing algorithm
for the HCP. a, Lin-Kernighan “transport” move.6 b, Lin- Kernighan
“reversal” move.6 c, “Swap” move5.
4
At each annealing step, try some number of configurations—rearrangements of the order
of cities in the tour, using the moves detailed above to generate the configurations. The
chance of accepting a configuration resulting in a longer tour is reduced as time goes on;
by analogy, the problem cools according to an annealing schedule. From an initial
“temperature” value T0, given a cooling factor Fc, we iterate through Na annealing steps,
each time reducing the temperature by the cooling factor. That is, moving from step i to
step i+1,
T{i + 1} = Ti * Fc
The Metropolis condition is
exp(-|dL|/kT)
where dL is the increase/decrease in tour length after a given move, and k is a Boltzmannlike constant. The number of configurations tried in each annealing step is equal to MN2,
where M is a multiplicative constant—a tunable parameter. The roads connecting pairs of
cities are initialized randomly, and there are mN loge N, where m is a tunable
multiplicative constant.
Certainly, if a set of roads does not contain a Hamiltonian cycle our algorithm will
correctly report that there is no HC. But since it cannot be known a priori (for any
sizeable set of roads, e.g., 40 cities) that the roads do not contain a HC to study an
algorithm one needs to use a set of roads into which a HC has been placed and see if the
algorithm will report that there is a HC, additionally one can see if the algorithm has
found the "planted" HC or another one.
We first optimized an algorithm5 for the HCP that was effective in finding HCs for N up
to 50 cities but then began to become significantly less effective. We took the case of N
= 150 and with 0.5*150*loge 150 roads as this number of roads for a given N seems to be
theoretically the most difficult case.8 This extreme difficulty of the HCP around 0.5N loge
N roads seems related though not exactly the same as phase transition like behavior
Kirkpatrick and Selman found9 in studying the k-satisfiability (k-SAT) problem: given a
randomly generated Boolean expression of N (Boolean) variables built of clauses having
k variables per clause, can the expression be satisfied (made true) with a single
combination of true/false values for all the variables. For k=2, k-SAT has a polynomial
time solution but for k >2 the problem is known to be NP-complete. It was found that for
random (2 + e)-SAT problems the ability to find solutions numerically the probability of
finding a solution goes through a phase transition.
The Boltzmann constant k changes the threshold for accepting “bad” moves, which
increase the length of the trip rather than decrease it (that is, they increase the number of
nodes in the current tour with no graph edge connecting them). Smaller values of k
decrease the threshold, so that “bad” moves are less frequent. Figure 2 shows the effect of
k, and the “sweet spot” at k=0.4 maximizing the number of Hamiltonian cycles found out
of 128 trials.
5
N=150 - Dependence of HC found on
Boltzmann constant k
140
HC Found
120
100
80
60
40
20
0
0
1
2
3
4
5
k
Figure 2 Dependence of the number of HCs found on k. Horizontal line is 128, the
maximum possible.
Next using this optimized value for k we evaluated the algorithm starting with varying
numbers of random roads. The number of roads is m*N*loge N, where m is a
multiplicative constant. Results are shown in Figure 3. We see that a HC is found in all
cases for 0.58*N*loge N or more roads. We also see that for m=0.5 or greater the inserted
tour is never found and that for m=0.54 or greater all 128 HCs found are unique.
6
N=200 k=0.4 How "hard" it is to find cycles vs.
the number of roads in the system
Found Inserted Cycle
Unique HC Found
HC Found
HC found
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
m
1
Figure 3 Probability of finding a HC as a function of number of roads. The number of
roads is m*N loge N. The dotted horizontal line is 128, the maximum possible value.
We next asked what would happen if we did not insert a tour. As shown in Figure 4 for
N=200, no HC was found until there were 0.9*N*loge N total roads. Thus the presence of
even one HC (e.g., an inserted tour) seems to induce sufficient “connectivity” in the
system so as to generate more HCs.
7
N=200 k=0.4 How "hard" it is to find cycles vs.
number of roads, with no inserted cycle
HC Found
Unique HC Found
HC found
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
0
1
2
3
4
5
6
7
8
9 10 11
m
Figure 4 Probability of finding a HC as a function of number of roads when there is no
pre-inserted HC. The number of roads is m*N loge N. At m=4 and higher, 128 HCs are
found, the maximum possible number.
What about systems in which the edges are nonrandom? We ran some experiments with
MN2 deterministic edges in the graph, guaranteeing no Hamiltonian cycle. We generate
the edges starting with stride 1, connecting every other pair of nodes; the sequence is
{0®1, 2®3, 3®4, 5®6,….} until we reach node N (or N-1, depending on whether N is
even or odd), or until we have the desired number of edges. If stride-1 reaches N before
we have the specified number of edges, we move on to stride-2: {0®2, 3®5, 6®8,
9®11,….}. If we have not hit the quota of edges yet within this sequence, we move on to
stride-3: {0®3, 4®7, 8®11, 12®15,….}. We continue on up through larger strides as
far as needed to generate the specified number of edges. To this we add, as before, one
8
random-order Hamiltonian cycle. Results are shown in Figure 5. As for the case of roads
added randomly to a HC, when there are 0.58*N*loge N total roads a HC is always found.
N=200 k=0.4 How "hard" it is to find cycles vs. the
number of roads: nonrandom roads
HC found
Found inserted cycle
HC Found
Unique HC found
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1
m
Figure 5 Probability of finding a HC as a function of number of roads when there is an
inserted HC and other roads are added in a non-random fashion. The number of roads is
m*N loge N. See text for method of adding roads non-randomly.
9
Again, as in the case when a collection of roads is generated randomly without an
inserted HC, for a non-randomly generated set of even 0.9*N*loge N roads (in the same
fashion as above) no HC is found (Table 1).
m HC found
0.1
0
0.2
0
0.3
0
0.4
0
0.42
0
0.44
0
0.46
0
0.48
0
0.5
0
0.52
0
0.54
0
0.6
0
0.7
0
0.8
0
0.9
0
1
0
Table 1 Probability of finding an HC as a function of number of roads for a set of
non-randomly generated roads, with no cycle inserted. The number of roads is m*N
loge N. Because of the way the non-random roads are generated, they are
guaranteed not to contain an HC.
We thus suggest the following practical algorithm for the HCP: To the given set of roads
add randomly generated roads until there are at least 0.58*N*loge N total roads and run
our annealing based algorithm to look for a HC. Given the results in Figure 3 and Figure
4, if a HC is found then there is an HC in the original set of roads. If a HC is not found,
the original set of roads does not contain an HC. To test this we tried the cases of N =
300, 400 and 500. Results are shown in Table 2.
10
N
M
Cycles Found
200
0.5
125
200
0.7
128
300
0.5
125
300
0.7
128
400
0.5
124
400
0.7
128
500
0.5
126
500
0.7
128
Table 2 Probability of finding and HC for large N problems. In all cases, all the HC
found are unique, and in no case is the specifically-inserted HC found.
The energy function and the perturbing moves we use are purely local affecting only four
roads at a time. Yet, this energy function depending only on local information with some
global jumps from the annealing yields effective results for a factorially growing
problem. It seems that local information is essentially enough to solve a vast, complex
global problem. It will be interesting to see at what size N, if any, our algorithm ceases to
be effective. Or is global “knowledge” never needed? Why would a local algorithm
work? For a few or many roads (R << 0.5*N*loge N or R >> 0.5*N*loge N) the problem
is easily solved8: For small R there are just not enough roads to find a tour, and for large
R there is so much connectivity there almost always is a tour. The addition of the extra
roads must provide our predominantly local algorithm with just enough connectivity or
“information” to permit solution in the difficult intermediate case.
We do not have a proof that our algorithm will in all cases determine if a set of roads
contains a HC, but practically the method is efficient, runs on a standard computer, and as
far as can be tested is effective. Our results also bring up a new “constructive HCP”
problem: Can a given HC be found in a set of roads? Finally, our approach and findings
may be helpful in guiding proofs on the question of whether or not there is a polynomial
time algorithm for NP-complete problems.
This research used resources of the Argonne Leadership Computing Facility, which is a
DOE Office of Science User Facility supported under Contract DE-AC02-06CH11357.
11
References
1
Karp, R. Complexity of computer computations. R. Miller, J. Thatcher, J. Bohlinger,
eds., The IBM Research Symposia Series (Springer US, 1972), pp. 85–103.
2
Johnson, M.W., et al., Quantum annealing with manufactured spins. Nature 473, 194198 (2011).
3
Rnnow, T. F. et al., Defining and detecting quantum speedup. Science 345, 420-424
(2014).
4
Kirkpatrick, S. Gelatt, Jr., C. D. & Vecchi, M. P. Optimization by simulated annealing.
Science 220, 671-680 (1983).
5
Altschuler, E. L., Lades, M., & Stong, R. Finding Hamiltonian cycles. Science 273, 413
(1996).
6
Lin, S. & Kernighan, B. W. An effective heuristic algorithm for the traveling salesman
problem. Operations Research 21, 498-516 (1973).
7
Metropolis, N. Rosenbluth, A. W., Rosenbluth, M. N., Teller, A. H. & Teller, E.
Equation of state calculations by fast computing machines. The Journal of Chemical
Physics 21, 1087-1092 (1953).
8
Bollobas, B. Fenner, T. & Frieze, A. An algorithm for finding Hamilton paths and
cycles in random graphs Combinatorica 7, 327-342 (1987).
9
Kirkpatrick, S. & Selman, B. Critical behavior in the satisfiability of random Boolean
expressions. Science 264, 1297-1301 (1994).
12
| 8cs.DS
|
NULL-FINITE SETS IN TOPOLOGICAL GROUPS AND THEIR APPLICATIONS
arXiv:1706.08155v5 [math.GN] 28 Nov 2017
TARAS BANAKH AND ELIZA JABLOŃSKA
Abstract. In the paper we introduce and study a new family of “small” sets which is tightly connected with
two well known σ-ideals: of Haar-null sets and of Haar-meager sets. We define a subset A of a topological
group X to be null-finite if there exists a convergent sequence (xn )n∈ω in X such that for every x ∈ X the set
{n ∈ ω : x + xn ∈ A} is finite. We prove that each null-finite Borel set in a complete metric Abelian group is
Haar-null and Haar-meager. The Borel restriction in the above result is essential as each non-discrete metric
Abelian group is the union of two null-finite sets. Applying null-finite sets to the theory of functional equations
and inequalities, we prove that a mid-point convex function f : G → R defined on an open convex subset G of
a metric linear space X is continuous if it is upper bounded on a subset B which is not null-finite and whose
closure is contained in G. This gives an alternative short proof of a known generalization of Bernstein-Doetsch
theorem (saying that a mid-point convex function f : G → R defined on an open convex subset G of a metric
linear space X is continuous if it is upper bounded on a non-empty open subset B of G). Since Borel Haar-finite
sets are Haar-meager and Haar-null, we conclude that a mid-point convex function f : G → R defined on an
open convex subset G of a complete linear metric space X is continuous if it is upper bounded on a Borel
subset B ⊂ G which is not Haar-null or not Haar-meager in X. The last result resolves an old problem in the
theory of functional equations and inequalities posed by Baron and Ger in 1983.
Introduction
In 1920 Steinhaus [29] proved that for any measurable sets A, B of positive Haar measure in a locally compact
Polish group X the sum A + B := {a + b : a ∈ A, b ∈ B} has non-empty interior in X and the difference
B − B := {a − b : a, b ∈ B} is a neighborhood of zero in X.
In [8] Christensen extended the “difference” part of the Steinhaus results to all Polish Abelian groups proving
that for a Borel subset B of a Polish Abelian group X the difference B − B is a neighborhood of zero if B is
not Haar-null. Christensen defined a Borel subset B ⊂ X to be Haar-null if there exists a Borel σ-additive
probability measure µ on X such that µ(B + x) = 0 for all x ∈ X.
A topological version of Steinhaus theorem was obtained by Pettis [25] and Piccard [26] who proved that for
any non-meager Borel sets A, B in a Polish group X the sum A + B has non-empty interior and the difference
B − B is a neighborhood of zero.
In 2013 Darji [9] introduced a subideal of the σ-ideal of meager sets in a Polish group, which is a topological
analog to the σ-ideal of Haar-null sets. Darji defined a Borel subset B of a Polish group X to be Haar-meager if
there exists a continuous map f : K → X from a non-empty compact metric space K such that for every x ∈ X
the preimage f −1 (B + x) is meager in K. By [18], for every Borel subset B ⊂ X which is not Haar-meager in
a Polish Abelian group X the difference B − B is a neighborhood of zero.
It should be mentioned that in contrast to the “difference” part of the Steinhaus and Piccard–Pettis theorems,
the “additive” part can not be generalized to non-locally compact groups: by [23], [18] each non-locally compact
Polish Abelian group X contains two closed sets A, B whose sum A + B has empty interior but A, B are neither
Haar-null nor Haar-meager in X. In the Polish group X = Rω , for such sets A, B we can take the positive
cone Rω
+.
Steinhaus-type theorems has significant applications in the theory of functional equations and inequalities.
For example, the “additive” part of Steinhaus Theorem can be applied to prove that a mid-point convex
function f : Rn → R is continuous if it is upper bounded on some measurable set B ⊂ Rn of positive Lebesgue
measure (see e.g. [20, p.210]). We recall that a function f : G → R defined on a convex subset of a linear space
f (x)+f (y)
for any x, y ∈ G. Unfortunately, due to the example of Matous̆ková
≤
is mid-point convex if f x+y
2
2
and Zelený [23] we know that the “additive” part of the Steinhaus theorem does not extend to non-locally
1991 Mathematics Subject Classification. 39B52, 28C10, 18B30, 54E52.
Key words and phrases. Automatic continuity, additive function, mid-point convex function, null-finite set, Haar-null set,
Haar-meager set.
1
2
TARAS BANAKH AND ELIZA JABLOŃSKA
compact Polish Abelian groups. This leads to the following natural problem whose “Haar-null” version was
posed by Baron and Ger in [3, P239].
Problem (Baron, Ger, 1983). Is the continuity of an additive or mid-point convex function f : X → R on a
Banach space X equivalent to the upper boundedness of the function on some Borel subset B ⊂ X which is not
Haar-null or not Haar-meager?
In this paper we give the affirmative answer to the Baron–Ger Problem applying a new concept of a nullfinite set, which will be introduced in Section 2. In Section 3 we show that non-null-finite sets A in metric
groups X possess a Steinhaus-like property: if a subset A of a metric group X is not null-finite, then A − Ā
is a neighborhood of θ and for some finite set F ⊂ X the set F + (A − A) is a neighborhood of zero in X.
In Sections 5 and 6 we shall prove that Borel null-finite sets in Polish Abelian
groups are Haar-meager and
Q
Haar-null. On the other hand, in Example 7.1 we show that the product n∈N Cn of cyclic
Q groups contains a
closed Haar-null set which is not null-finite and in Example 7.2 we show that the product n∈N C2n contains a
null-finite Gδ -set which cannot be covered by countably many closed Haar-null sets. In Section 8 we prove that
each non-discrete metric Abelian group is the union of two null-finite sets and also observe that sparse sets
(in the sense of Lutsenko–Protasov [22]) are null-finite. In Section 9 we apply null-finite sets to prove that an
additive functional f : X → R on a Polish Abelian group is continuous if it is upper bounded on some subset
B ⊂ X which is not null-finite in X. In Section 10 we generalize this result to additive functions with values
in Banach and locally convex spaces. Finally, in Section 11, we prove a generalization of Bernstein–Doetsch
theorem and next apply it in the proof of a continuity criterion for mid-point convex functions on complete
metric linear spaces, thus answering the Baron–Ger Problem. In Section 12 we pose some open problems
related to null-finite sets.
1. Preliminaries
All groups considered in this paper are Abelian. The neutral element of a group will be denoted by θ. For a
group G by G∗ we denote the set of non-zero elements in G. By Cn = {z ∈ C : z n = 1} we denote the cyclic
group of order n.
By a (complete) metric group we understand an Abelian group X endowed with a (complete) invariant
metric k · − · k. The invariant metric k · − · k determines (and can be recovered from) the prenorm k · k defined
by kxk := kx − θk. So, a metric group can be equivalently defined as a group endowed with a prenorm.
Formally, a prenorm on a group X is a function k · k : X → R+ := [0, ∞) satisfying three axioms:
• kxk = 0 iff x = θ;
• k − xk = kxk;
• kx + yk ≤ kxk + kyk
for any x, y ∈ X; see [1, §3.3].
Each metric group is a topological group with respect to the topology, generated by the metric.
All linear spaces considered in this paper are over the field R of real numbers. By a metric linear space we
understand a linear space endowed with an invariant metric.
A non-empty family I of subsets of a set X is called ideal of sets on X if I satisfies the following conditions:
• X∈
/ I;
• for any subsets J ⊂ I ⊂ X the inclusion I ∈ I implies J ∈ I;
• for any sets A, B ∈ I we have A ∪ B ∈ I.
S
An ideal I on X is called a σ-ideal if for any countable subfamily C ⊂ I the union C belongs to X.
An ideal I on a group X is called invariant if for any I ∈ I and x ∈ X the shift x + I of I belongs to the
ideal I.
For example, the family [X]<ω of finite subsets of an infinite group X is an invariant ideal on X.
A topological space is Polish if it is homeomorphic to a separable complete metric space. A topological
space is analytic if it is a continuous image of a Polish space.
2. Introducing null-finite sets in topological groups
In this section we introduce the principal new notion of this paper.
A sequence (xn )n∈ω in a topological group X is called a null-sequence if it converges to the neutral element
θ of X.
NULL-FINITE SETS IN TOPOLOGICAL GROUPS AND THEIR APPLICATIONS
3
Definition 2.1. A set A of a topological group X is called null-finite if there exists a null-sequence (xn )n∈ω
in X such that for every x ∈ X the set {n ∈ ω : x + xn ∈ A} is finite.
Null-finite sets in metrizable topological groups can be defined as follows.
Proposition 2.2. For a non-empty subset A of a metric group X the following conditions are equivalent:
(1) A is null-finite;
(2) there exists an infinite compact set K ⊂ X such that for every x ∈ X the intersection K ∩ (x + A) is
finite;
(3) there exists a continuous map f : K → X from an infinite compact space K such that for every x ∈ X
the preimage f −1 (A + x) is finite.
Proof. The implications (1) ⇒ (2) ⇒ (3) are obvious, so it is enough to prove (3) ⇒ (1). Assume that
f : K → X is a continuous map from an infinite compact space K such that for every x ∈ X the preimage
f −1 (A + x) is finite. Fix any point a ∈ A. It follows that for every x ∈ X the fiber f −1 (a + x) ⊂ f −1 (A + x) is
finite and hence the image f (K) is infinite. So, we can choose a sequence (yn )n∈ω of pairwise distinct points
in f (K). Because of the compactness and metrizability of f (K), we can additionally assume that the sequence
(yn )n∈ω converges to some point y∞ ∈ f (K). Then the sequence (xn )n∈ω of points xn := yn − y∞ , n ∈ ω, is a
null-sequence witnessing that the set A is null-finite.
Null-finite sets are “small” in the following sense.
Proposition 2.3. For any null-finite set A in a metric group X and any finite subset F ⊂ X the set F + A
has empty interior in X.
Proof. Since A is null-finite, there exists a null-sequence (xn )n∈ω in X such that for every x ∈ X the set
{n ∈ ω : x + xn ∈ A} is finite.
To derive a contradiction, assume that for some finite set F the set F + A has non-empty interior U in X.
Then for any point u ∈ U the set {n ∈ ω : u+xn ∈ U } ⊂ {n ∈ ω : u+xn ∈ F +A} is infinite. By the Pigeonhole
Principle, for some y ∈ F the set {n ∈ ω : u + xn ∈ y + A} is infinite and so is the set {n ∈ ω : −y + u + xn ∈ A}.
But this contradicts the choice of the sequence (xn )n∈ω .
Now we present some examples of closed sets which are (or are not) null-finite. We recall that for a group
G by G∗ we denote the set G \ {θ} of non-zero elements of G.
Q
Example 2.4. Let (Gn )n∈ω be a Q
sequence of finite groups. The set A := n∈ω G∗n is null-finite in the compact
metrizable topological group G = n∈ω Gn if and only if limn→∞ |Gn | 6= ∞.
Proof. First, we assume that limn→∞ |Gn | = ∞ and show that the set A is not null-finite in G. Given a
null-sequence (xn )n∈ω in the compact topological group G, we should find an element a ∈ G such that the set
{n ∈ ω : a + xn ∈ A} is infinite.
L
It will be convenient to think of elements of the group G as functions x : ω → n∈ω Gn such that x(k) ∈ Gk
for all k ∈ ω.
Taking into account that limk→∞ |Gk | = ∞ and limn→∞ xn = θ ∈ G, we can inductively construct an
increasing number sequence (nk )k∈ω such that n0 = 0 and for every k ∈ N and m ≥ nk the following two
conditions are satisfied:
(1) xm (i) = θ for all i < nk−1 ;
(2) |Gm | ≥ k + 3.
We claim that for every i ≥ n1 the set {xnk (i) : k ∈ ω} has cardinality < |Gi |. Indeed, given any i ≥ n1 , we
can find a unique j ≥ 1 such that nj ≤ i < nj+1 and conclude that xnk (i) = θ for all k ≥ j + 1 (by the condition
(1)). Then the set {xnk (i) : k ∈ ω} = {θ} ∪ {xnk (i) : k ≤ j} has cardinality ≤ 1 + (j + 1) < j + 3 ≤ |Gi |. The
last inequality follows from i ≥ nk and condition (2). Therefore, the set {xnk (i) : k ∈ ω} has cardinality < |Gi |
and we can choose a point ai ∈ Gi \ {−xnk (i) : k ∈ ω} and conclude that ai + xnk (i) ∈ G∗i for all k ∈ ω. Then
the element a = (ai )i∈ω ∈ G has the required property as the set {n ∈ ω : a + xn ∈ A} ⊃ {nk }k∈ω is infinite.
Next, assuming that limk→∞ |Gk | 6= ∞, we shall prove that the set A is Haar-finite in G. Since limk→∞ |Gk | 6=
∞, for some l ∈ ω the set Λ = {k ∈ ω : |Gk | = l} is infinite.
Since limn→∞ (1 − 1l )n = 0, there exists n ∈ N such that (1 − 1l )n < 1l and hence l(l − 1)n < ln . This implies
that for every k ∈ Λ the set Gk ·(G∗k )n := {(x + a1 , · · · x + an ) : x ∈ Gk , a1 , . . . , an ∈ G∗k } has cardinality
|Gk ·(G∗k )n | ≤ l(l − 1)n < ln = |Gnk |. Consequently, the (compact) set G·An = {(x + a1 , . . . , x + an ) : x ∈
4
TARAS BANAKH AND ELIZA JABLOŃSKA
G, a1 , . . . , an ∈ A} is nowhere dense in the compact topological group Gn . Now Theorem 2.5 ensures that A
is null-finite in G.
Theorem 2.5. A non-empty subset A of a Polish group X is null-finite in X if for some n ∈ N the set
X·An := {(x + a1 , . . . , x + an ) : x ∈ X, a1 , . . . , an ∈ A} is meager in X n .
Proof. Let K(X) be the space of non-empty compact subsets of X, endowed with the Vietoris topology. For
a compact subset K ⊂ X let K∗n = {(x1 , . . . , xn ) ∈ K n : |{x1 , . . . , xn }| = n} the set of n-tuples consisting of
pairwise distinct points of K.
By Mycielski-Kuratowski Theorem [19, 19.1], the set W∗ (A) := {K ∈ K(X) : K∗n ∩ (X·An ) = ∅} is comeager
in K(X) and hence contains some infinite compact set K. We claim that for every x ∈ X the set K ∩ (x + A)
has cardinality < n. Assuming the opposite, we could find n pairwise distinct points a1 , . . . , an ∈ A such that
x + ai ∈ K for all i ≤ n. Then (x + a1 , . . . , x + an ) ∈ K∗n ∩ X · An , which contradicts the inclusion K ∈ W∗ (A).
By Proposition 2.2, the set A is null-finite.
3. A Steinhaus-like properties of sets which are not null-finite
In [29] Steinhaus proved that for a subset A of positive Lebesgue measure in the real line, the set A − A
is a neighborhood of zero. In this section we establish three Steinhaus-like properties of sets which are not
null-finite.
Theorem 3.1. If a subset A of a metric group X is not null-finite, then
(1) the set A − Ā is a neighborhood of θ in X;
(2) each neighborhood U ⊂ X of θ contains a finite subset F ⊂ U such that F + (A − A) is a neighborhood
of θ.
(3) If X is Polish (and A is analytic), then A − A is not meager in X (and (A − A) − (A − A) is a
neighborhood of θ in X).
Proof. 1. Assuming that A − Ā is not a neighborhood of θ, we could find a null-sequence (xn )n∈ω contained
in X \ (A − Ā). Since A is not null-finite, there exists a ∈ X such that the set Ω = {n ∈ ω : a + xn ∈ A} is
infinite. Then a ∈ {a + xn }n∈Ω ⊂ Ā and hence xn = (a + xn ) − a ∈ A − Ā for all n ∈ Ω, which contradicts the
choice of the sequence (xn )n∈ω .
2. Fix a decreasing neighborhood base (Un )n∈ω at zero in X such that U0 ⊂ U . For the proof by contradiction, suppose that for any finite set F ⊂ U the set F + (A − A) is not
S a neighborhood of zero. Then
we can inductively construct a null-sequence (xn )n∈ω such that xn ∈ Un \ 0≤i<n (xi + A − A) for all n ∈ ω.
Observe that for each z ∈ X the set {n ∈ ω : z + xn ∈ A} contains at most one point. Indeed, in the opposite
case we could find two numbers k < n with z + xk , z + xn ∈ A and conclude that z ∈ −xk + A and hence
xn ∈ −z + A ⊂ xk − A + A, which contradicts the choice of the number xn . The sequence (xn )n∈ω witnesses
that the set A is null-finite in X.
3. Finally assume that the group X is Polish. To derive a contradiction, assume that A − A is meager.
Since the map δ : X × X → X, δ : (x, y) 7→ x − y, is open, the preimage δ −1 (A − A) is meager in X × X.
Observe that δ −1 (A − A) = X·A2 where X·A2 = {(x + a, x + b) : x ∈ X, a, b ∈ A}. By Theorem 2.5, the set
A is null-finite, which is a desired contradiction. This contradiction shows that the set A − A is not meager
in X. If the space A is analytic, then so is the subset A − A of X. Being analytic, the set A − A has the
Baire Property in X; see [19, 29.14]. So, we can apply the Pettis-Piccard Theorem [19, 9.9] and conclude that
(A − A) + (A − A) is a neighborhood of θ in X.
The last statement of Theorem 3.1 does not hold without any assumptions on A and X.
Example 3.2. Let f : C2ω → C2 be a discontinuous homomorphism on the compact Polish group X = C2ω
(thought as the vector space C2ω over the two-element field Z/2Z). Then the subgroup A = f −1 (θ) is not
null-finite in X (by Proposition 2.3), but A − A = A = (A − A) − (A − A) is not a neighborhood of θ in X.
4. A combinatorial characterization of null-finite sets in compact metric groups
In compact metric groups null-finite sets admit a purely combinatorial description.
Proposition 4.1. A non-empty subset A of a compact metric group X is null-finite
if and only if there exists
T
an infinite set I ⊂ X such that for any infinite subset J ⊂ I the intersection x∈J (A − x) is empty.
NULL-FINITE SETS IN TOPOLOGICAL GROUPS AND THEIR APPLICATIONS
5
Proof. To prove the “only if” part, assume that A ⊂ X is null-finite. So, there exists a null-sequence (xn )n∈ω
such that for every x ∈ X the set {n ∈ ω : x + xn ∈ A} is finite. It follows that for every x ∈ X the set
{n ∈ ω : xn = x} is finite and hence the set I := {xn }n∈ω is infinite. We claim that
T this set has the required
property. Indeed, assuming that for some infinite subset J ⊂ I the intersection x∈J (A − x) contains some
point a ∈ X, we conclude that the set {n ∈ ω : a + xn ∈ A} contains the set {n ∈ ω : xn ∈ J} and hence is
infinite, which contradicts the choice of the sequence (xn )n∈ω .
To prove the “if” part,
T assume that there exists an infinite set I ⊂ X such that for every infinite set
J ⊂ I the intersection x∈J (A − x) is empty. By the compactness of the metric group X, some sequence
(xn )n∈ω of pairwise distinct points of the infinite set I converges to some point x∞ ∈ X. Then the nullsequence (zn )n∈ω consisting of the points zn = xn − x∞ , n ∈ ω, witnesses that A is null-finite. Assuming
the opposite, we would find a point a ∈ X such that the set {n ∈ ω : a + zn ∈ A} is infinite. Then the set
J := {x ∈ I : a − x∞ + x ∈ A} ⊂ {xn : n ∈ ω, a − x∞ + xn ∈ A} is infinite, too, and the intersection
T
x∈J (A − x) contains the point a − x∞ and hence is not empty, which contradicts the choice of the set I.
Following Lutsenko and Protasov [22] we define a subset T
A of an infinite group X to be sparse if for any
infinite set I ⊂ X there exists a finite set F ⊂ I such that x∈F (x + A) is empty. By [22, Lemma 1.2] the
family of sparse subsets of a group is an invariant ideal on X (in contrast to the family of null-finite sets).
Proposition 4.2. Each sparse subset A of a non-discrete metric group X is null-finite.
Proof. Being non-discrete, the metric group X contains a null-sequence (xn )n∈ω consisting of pairwise distinct
points. Assuming that the sparse set A is not null-finite, we can find a point a ∈ X such that the set
Ω := {n ∈ ω : a + xn ∈ A} is infinite. Since
T the infinite set I := {−xn : n ∈ Ω} there exists a
T A is sparse, for
finite set F ⊂ I such that the intersection x∈F (x + A) ⊃ m∈Ω (−xn + A) ∋ a is empty, which is not possible
as this intersection contains the point a.
5. Null-finite Borel sets are Haar-meager
In this section we prove that each null-finite set with the universal Baire property in a complete metric
group is Haar-meager.
A subset A of a topological group X is defined to have the universal Baire Property (briefly, A is a uBP-set)
if for any function f : K → X from a compact metrizable space K the preimage f −1 (A) has the Baire Property
in K, which means that for some open set U ⊂ K the symmetric difference U △f −1 (A) is meager in K. It is
well-known that each Borel subset of a topological group has the universal Baire Property.
A uBP-set A of a topological group X is called Haar-meager if there exists a continuous map f : K → X
from a compact metrizable space K such that f −1 (x + A) is meager in K for all x ∈ X. By [9], for a complete
metric group X the family HMX of subsets of Haar-meager uBP-sets in X is an invariant σ-ideal on X. For
more information on Haar-meager sets, see [9], [11], [10], [18].
Theorem 5.1. Each null-finite uBP-set in a compete metric group is Haar-meager.
Proof. To derive a contradiction, suppose that a null-finite uBP-set A in a complete metric group (X, k · k) is
not Haar-meager. Since A is null-finite, there exists a null-sequence (an )n∈ω such that for every x ∈ X the set
{n ∈ ω : x + an ∈ A} is finite.
Replacing (an )n∈ω by a suitable subsequence, we can assume that kan k < 21n for all n ∈ ω. For every n ∈ ω
consider the compact set Kn := {θ} ∪ {am }m≥n ⊂ X. The metric restriction on the sequence (an )n∈ω implies
that the function
Y
X
Σ:
Kn → X, Σ : (xn )n∈ω 7→
xn ,
n∈ω
n∈ω
is well-defined and continuous (the proof of this fact can be found in [18]; see the proof of Theorem 2).
−1
Since the set A is not Haar-meager and has uBP, there exists a point z ∈ X such
Q that B := Σ (z + A)
is not meager and has the Baire Property in the compact metrizable space K := n∈ω Kn . Consequently,
there exits a non-empty open set U ⊂ K such Q
that U ∩ B is comeager in U . Replacing U by a smaller
subset,
Q
we can assume that U is of basic form {b} × m≥j Km for some j ∈ ω and some element b ∈ m<j Km . It
Q
follows that for every x ∈ Kj \ {θ} the set {b} × {x} × m>j Km is closed and open in K and hence the set
T
Q
Q
Cx := {y ∈ m>j Km : {b}×{x}×{y} ⊂ U ∩B} is comeager in m>j Km . Then the intersection x∈Kn\{θ} Cx
is comeager and hence contains some point c. For this point we get the inclusion {b} × (Kj \ {θ}) × {c} ⊂ B.
Let v ∈ K be a unique point such that {v} = {b} × {θ} × {c}. For any m ≥ j the point am belongs to Kj and
6
TARAS BANAKH AND ELIZA JABLOŃSKA
the inclusion {b} × {am} × {c} ⊂ B implies that Σ(v) + am ∈ z + A. Consequently, the element u := −z + Σ(v)
has the property u + am ∈ A for all m ≥ j, which implies that the set {n ∈ ω : u + an ∈ A} is infinite. But
this contradicts the choice of the sequence (an )n∈ω .
6. Null-finite Borel sets are Haar-null
In this section we prove that each universally measurable null-finite set in a complete metric group is Haarnull.
A subset A of a topological group X is defined to be universally measurable if A is measurable with respect to
any σ-additive Radon Borel probability measure on X. It is clear that each Radon Borel subset of a topological
group is universally measurable. We recall that a measure µ on a topological space X is Radon if for any ε > 0
there exists a compact subset K ⊂ X such that µ(X \ K) < ε.
A universally measurable subset A of a topological group X is called Haar-null if there exists a σ-additive
Radon Borel probability measure µ on X such that µ(x + A) = 0 for all x ∈ X. By [8], for any complete metric
group X the family HNX of subsets of universally measurable Haar-null subsets of X is an invariant σ-ideal in
X. For more information on Haar-null sets, see [8], [16], [17] and also e.g. [13], [15], [23], [28].
Theorem 6.1. Each universally measurable null-finite set A in a complete metric group X is Haar-null.
Proof. To derive a contradiction, suppose that the null-finite set A is not Haar-null in X. Since A is null-finite,
there exists a null-sequence (an )n∈ω such that for each x ∈ X the set {n ∈ ω : x + an ∈ A} is finite. Replacing
(an )n∈ω by a suitable subsequence, weQ
can assume that kan k ≤ 41n for all n ∈ ω.
Consider the compact space Π := n∈ω {0, 1, . . . , 2n } endowed with the product measure λ of uniformly
distributed measures on the finite discrete spaces {0, 1, 2, 3, . . . , 2n }, n ∈ ω. Consider the map
∞
X
p i ai .
Σ : Π → X, Σ : (pi )i∈ω 7→
i=0
P
Since kpi ai k ≤ 2i kai k ≤ 2i 41i = 21i , the series ∞
i=0 pi ai is convergent and the function Σ is well-defined and
continuous.
Since the set A is not Haar-null, there exists an element z ∈ X such that the preimage Σ−1 (z + A) has
positive λ-measure and hence contains a compact subset K of positive measure. For every n ∈ ω consider
Qn−1
Q∞
the subcube Πn := i=0 {0, 1, 2, . . . , 2i } × i=n {1, 2, . . . , 2i } of Π and observe that λ(Πn ) → 1 as n → ∞.
Replacing K by K ∩ Πl for a sufficiently large l, we can assume that K ⊂ Πl . For every m ≥ l let sm : Πl → Π
be the “back-shift” defined by the formula sm ((xi )i∈ω ) := (yi )i∈ω where yi = xi for i 6= m and yi = xi − 1 for
i = m.
Claim 6.2. For any compact set C ⊂ Πl of positive measure λ(C) and any ε > 0 there exists k ≥ l such that
for any m ≥ k the intersection C ∩ sm (C) has measure λ(C ∩ sm (C)) > (1 − ε)λ(C).
Proof. By the regularity of the measure λ, the set C has a neighborhood O(C) ⊂ Π such that λ(O(C) \ C) <
ελ(K). By the compactness of C, there exists k ≥ l such that for any m ≥ k the shift sm (C) is contained in
O(C). Hence λ(sm (C) \ C) ≤ λ(O(C) \ C) < ελ(C) and thus λ(sm (C) ∩ C) = λ(sm (C)) − λ(sm (C) \ C) >
λ(C) − ελ(C) = (1 − ε)λ(C).
Using Claim 6.2 we can choose an increasing number sequence (mk )k∈ω such that m0 > l and the set
T
K∞ := k∈ω smk (K) has positive measure and hence contains a point ~b := (bi )i∈ω . It follows that for every
−1
~
k ∈ ω the point ~bk := s−1
(z + A).
mk (b) belongs to K ⊂ Σ
~
~
Observe that Σ(bk ) = Σ(b) + amk ∈ z + A and hence −z + Σ(~b) + amk ∈ A for all k ∈ ω, which contradicts
the choice of the sequence (an )n∈ω .
Remark 6.3. After writing the initial version of this paper we discovered that a result similar to Theorem 6.1
was independently found by Bingham and Ostaszewski [6, Theorem 3].
7. The σ-ideal generated by (closed) Borel null-finite sets
In this section we introduce two new invariant σ-ideals generated by (closed) Borel null-finite sets and study
the relation of these new ideals to the σ-ideals of Haar-null and Haar-meager sets.
Namely, for a complete metric group X let
• σNFX be the smallest σ-ideal containing all Borel null-finite sets in X;
NULL-FINITE SETS IN TOPOLOGICAL GROUPS AND THEIR APPLICATIONS
7
• σNFX be the smallest σ-ideal containing all closed null-finite sets in X;
• σHNX be the smallest σ-ideal containing all closed Haar-null sets in X.
Theorems 5.1 and 6.1 imply that σNFX ⊂ σHNX and σNFX ⊂ HNX ∩ HMX . So, we obtain the following
diagram in which an arrow A → B indicates that A ⊂ B.
σNFX
/ σHNX
σNFX
/ HNX ∩ HMX
In Examples 7.1 and 7.2 we show that the σ-ideal σNFX is strictly smaller than σHNX and the ideal σNFX is
not contained in σHNX .
Q
Q
Example 7.1. The closed set A = n≥2 Cn∗ in the product X = n≥2 Cn of cyclic groups is Haar-null but
cannot be covered by countably many closed null-finite sets. Consequently, A ∈ σHNX \ σNFX .
Q
∗
Proof. The set A = ∞
n=2 Cn has Haar measure
∞
∞
Y
Y
n−1
|Cn∗ |
=
=0
|Cn | n=2 n
n=2
and hence is Haar-null in the compact Polish group X.
Next, we show S
that A cannot be covered by countably many closed null-finite sets. To derive a contradiction,
assume that A = n∈ω An where each An is closed and null-finite in X. By the Baire Theorem, for some n ∈ ω
Qm−1
the setQAn has non-empty interior in A. Consequently,
we can find m > 2 and
a ∈ n=2 Cn∗ such that
Q
Q
∞
∞
∞
{a} × n=m Cn∗ ⊂ A
. By Example 2.4, the set n=m Cn∗ is not null-finite in n=m Cn , which implies that
Qn∞
∗
the set An ⊃ {a} × n=m Cn is not null-finite in the group X. But this contradicts the choice of An .
Our next example shows that σNFX 6⊂ σHNX for some compact Polish group X.
Q
Example 7.2. For any function f : ω → [2, ∞) with n∈ω f (n)−1
f (n) > 0, the compact metrizable group X =
Q
C
contains
a
null-finite
G
-set
A
⊂
X
which
cannot
be
covered by countably many closed Haar-null
δ
n∈N f (n)
sets in X. Consequently, A ∈ NFX \ σHNX .
Proof.QFor every n ∈ ω let gn be a generator of the cyclic group Cf (n) . In the compact metrizable group
X = n∈ω Cf (n) consider the null-sequence (xn )n∈ω defined by the formula
(
θ if i ≤ n
xn (i) =
gi if i > n.
Q
Consider the closed subset B = n∈ω Cf∗(n) in X and observe that it has positive Haar measure, equal to the
infinite product
Y f (n) − 1
> 0.
f (n)
n∈ω
It is easy to see that for every n ∈ ω the set Cf∗(n) = Cf (n) \ {θ} is not equal to Cf∗(n) ∩ (gn + Cf∗(n) ), which
S
implies that the intersection B ∩ (xn + B) is nowhere dense in B. Consequently, the set A = B \ n∈ω (xn + B)
is a dense Gδ -set in B.
We claim that the set A is null-finite. Given any a ∈ X we should prove that the set {n ∈ ω : a + xn ∈ A} is
finite. If a ∈
/ B, then the open set X \ B is a neighborhood of a in X. Since the sequence (a + xn )n∈ω converges
to a ∈ X \ B, the set {n ∈ ω : a + xn ∈ B} ⊃ {n ∈ ω : a + xn ∈ A} is finite.
If a ∈ B, then {n ∈ ω : a + xn ∈ A} ⊂ {n ∈ ω : (B + xn ) ∩ A 6= ∅} = ∅ by the definition of the set A.
Next, we prove that the Gδ -set
S A cannot be covered by countably many closed Haar-null sets. To derive a
contradiction, assume that A ⊂ n∈ω Fn for some closed Haar-null sets Fn ⊂ X. Since the space A is Polish,
we can apply the Baire Theorem and find n ∈ ω such that
Q the set A ∩ Fn has non-empty interior in A and
hence its closure A ∩ Fn has non-empty interior in B = n∈ω Cf∗(n) . It is easy to see that each non-empty
open subset of B has positive Haar measure in X. Consequently, the set A ∩ Fn has positive Haar measure,
which is not possible as this set is contained in the Haar-null set Fn .
8
TARAS BANAKH AND ELIZA JABLOŃSKA
Remark 7.3. Answering a question posed in a preceding version of this paper, Adam Kwela [21] constructed
two compact null-finite subsets A, B on the real line, whose union A ∪ B is not null-finite. This means that
the family of subsets of (closed) Borel null-finite subsets on the real line is not an ideal, and the ideal σNFR
contains compact subsets of the real line, which fail to be null-finite.
8. Decomposing non-discrete metric groups into unions of two null-finite sets
By Theorem 6.1 and the countable additivity of the family of Borel Haar-null sets [8], the countable union
of Borel null-finite sets in a complete metric group X is Haar-null in X and hence is not equal to X. So, a
complete metric group cannot be covered by countably many Borel null-finite sets. This result dramatically
fails for non-Borel null-finite sets.
Theorem 8.1. Each non-discrete metric group X can be written as the union X = A ∪ B of two null-finite
subsets A, B of X.
Proof. Being non-discrete, the metric group X contains a non-trivial null-sequence, which generates a nondiscrete countable subgroup Z in X.
Let Z = {zn }n∈ω be an enumeration of the countable infinite group Z such that z0 = θ and zn 6= zm for
any distinct numbers n, m ∈ ω. By induction we can construct sequences (un )n∈ω and (vn )n∈ω in Z such that
u0 = v0 = θ and for every n ∈ N the following two conditions are satisfied:
/ {−zi + zj + vk : i, j ≤ n, k < n} ∪ {−zi + zj + uk : i, j ≤ n, k < n};
(1) kun k ≤ n1 and un ∈
1
(2) kvn k ≤ n and vn ∈
/ {−zi + zj + uk : i, j ≤ n, k ≤ n} ∪ {−zi + zj + vk : i, j ≤ n, k < n}.
At the n-th step of the inductive construction, the choice of the points un , vn is always possible as the ball
{z ∈ Z : kzk ≤ n1 } is infinite and un , vn should avoid finite sets.
After completing the inductive construction, we obtain the null-sequences (un )n∈ω and (vn )n∈ω of pairwise
distinct points of Z such that for any points x, y ∈ Z the intersection {x + un }n∈ω ∩ {y + um }m∈ω is finite and
for every distinct points x, y ∈ Z the intersections {x + un }n∈ω ∩ {y + um }m∈ω and {x + vn }n∈ω ∩ {y + vm }m∈ω
are finite.
Using these facts, for every n ∈ ω we can choose a number in ∈ ω such that the set {zn + vk : k ≥ in } is
disjoint with the set {zi + vm : i < n, m ∈ ω} ∪ {zi + um : i ≤ n, m ∈ ω} and the set {zn + uk : k ≥ in } is
disjoint with the set {zi + um S
: i < n, m ∈ ω} ∪ {zi + vm : i ≤ n, m ∈ ω}.
We claim that the set A := n∈ω {zn + um : m > in } is null-finite in Z. This will follow as soon as we verify
the condition:
(3) for any z ∈ Z the set {n ∈ ω : z + vn ∈ A} is finite.
Find j ∈ ω such that z = zj and observe that for every n ≥ j the choice of the number
S in guarantees that
{zn + um }m>in ∩ {zj + vn }n∈ω = ∅, so {n ∈ ω : z + vn ∈ A} is contained in the finite set n<j {zn + vm }m∈ω ∩
{zj + um }m∈ω and hence is finite.
Next, we show that the set B := Z \ A is null-finite in Z. This will follow as soon as we verify the condition
(4) for any z ∈ Z the set {m ∈ ω : z + um ∈ B} is finite.
Find a number n ∈ ω such that z = zn and observe that {m ∈ ω : z + um ∈ B} = {m ∈ ω : zn + um ∈
/ A} ⊂
{m ∈ ω : m ≤ in } is finite. Therefore, the sets A and B = Z \ A are null-finite in Z.
Using Axiom of Choice, choose a subset S ⊂ X such that for every x ∈ X the intersection S ∩ (x + Z) is
a singleton. It follows that X = S + Z = (S + A) ∪ (S + B). We claim that the sets S + A and S + B are
null-finite in X. This will follow as soon as for any x ∈ X we check that the sets {n ∈ ω : x + vn ∈ S + A}
and {n ∈ ω : x + un ∈ S + B} are finite. Since X = S + Z, there exist elements s ∈ S and z ∈ Z such that
x = s + z. Observe that if for n ∈ ω we get x + vn ∈ S + A, then s + z + vn = t + a for some t ∈ S and
a ∈ A ⊂ Z. Subtracting a from both sides, we get t = s + z + vn − a ∈ s + Z and hence t = s as S ∩ (s + Z)
is a singleton containing both points t and s. So, the inclusion x + vn = s + z + vn ∈ S + A is equivalent to
z + vn ∈ A. By analogy we can prove that x + un ∈ S + B is equivalent to z + un ∈ B. Then the sets
{n ∈ ω : x + vn ∈ S + A} = {n ∈ ω : z + vn ∈ A} and {n ∈ ω : x + un ∈ S + B} = {n ∈ ω : z + un ∈ B}
are finite by the properties (3) and (4) of the null-finite sets A, B.
Proposition 2.3 and Theorem 8.1 imply the following corollary.
Corollary 8.2. Let X be a non-discrete metric group.
NULL-FINITE SETS IN TOPOLOGICAL GROUPS AND THEIR APPLICATIONS
9
(1) Each null-finite subset of X is contained in some invariant ideal on X;
(2) For any ideal I on X there exists a null-finite set A ⊂ X such that A ∈
/ I.
Proof. 1. For any null-finite set A in X the family
IA = {I ⊂ X : ∃F ∈ [X]<ω I ⊂ F A}
is an invariant ideal on X whose elements have empty interiors in X by Proposition 2.3. We recall that by
[X]<ω we denote the ideal of finite subsets of X.
2. By Theorem 8.1, the group X can be written as the union X = A ∪ B of two null-finite sets. Then for
any ideal I on X one of the sets A or B does not belong to I.
9. Applying null-finite sets to additive functionals
In this section we apply null-finite sets to prove a criterion of continuity of additive functionals on metric
groups.
A function f : X → Y between groups is called additive if f (x + y) = f (x) + f (y) for every x, y ∈ X. An
additive function into the real line is called an additive functional.
Theorem 9.1. An additive functional f : X → R on a metric group X is continuous if it is upper bounded on
a set B ⊂ X which is not null-finite.
Proof. Suppose that the functional f is not continuous. Then there exists ε > 0 such that f (U ) 6⊂ (−ε, ε)
for each neighborhood U ⊂ X of zero. It follows that for every n ∈ ω there is a point xn ∈ X such that
kxn k ≤ 41n and |f (xn )| > ε. Observe that, k2n xn k ≤ 2n · kxn k ≤ 21n and |f (2n xn )| = 2n |f (xn )| > 2n ε. Choose
εn ∈ {1, −1} such that εn f (2n xn ) is positive and put zn := εn 2n xn . Then f (zn ) > 2n ε and kzn k ≤ 21n .
We claim that the null-sequence (zn )n∈ω witnesses that the set B is null-finite. Given any point x ∈ X we
need to check that the set Ω := {n ∈ ω : x + zn ∈ B} is finite.
Let M := sup f (B) and observe that for every n ∈ Ω we have
M ≥ f (x + zn ) = f (x) + f (zn ) ≥ f (x) + 2n ε,
which implies 2n ≤ 1ε (M − f (x)). Hence the set Ω is finite. Therefore, the null-sequence (zn )n∈ω witnesses
that the set B is null-finite, which contradicts our assumption.
Theorems 9.1 and 5.1 imply
Corollary 9.2. An additive functional f : X → R on a complete metric group is continuous if and only if it
is upper bounded on some uBP set B ⊂ X which is not Haar-meager.
An analogous result for Haar-null sets follows from Theorems 9.1 and 6.1.
Corollary 9.3. An additive functional f : X → R on a complete metric group is continuous if and only if it
is upper bounded on some universally measurable set B ⊂ X which is not Haar-null.
10. Applying null-finite sets to additive functions
In this section we prove some continuity criteria for additive functions with values in Banach spaces or
locally convex spaces.
Corollary 10.1. An additive function f : X → Y from a complete metric group X to a Banach space Y is
continuous if for any linear continuous functional y ∗ : Y → R the function y ∗ ◦ f : X → R is upper bounded
on some set B ⊂ X which is not null-finite in X.
Proof. Assuming that f is discontinuous, we can find ε > 0 such that for each neighborhood U ⊂ X of θ
we get f (U ) 6⊂ Bε := {y ∈ Y : kyk < ε}. Then for every n ∈ ω we can find a point xn ∈ X such that
kxn k ≤ 41n and kf (xn )k ≥ ε. Since k2n xn k ≤ 2n kxn k ≤ 21n , the sequence (2n xn )n∈ω is a null-sequence in X.
On the other hand, kf (2n xn )k = 2n kf (xn )k ≥ 2n ε for all n ∈ ω, which implies that the set {f (2n xn )}n∈ω is
unbounded in the Banach space Y . By the Banach-Steinhaus Uniform Boundedness Principle [14, 3.15], there
exists a linear continuous functional y ∗ : Y → R such that the set {y ∗ ◦ f (2n xn ) : n ∈ ω} is unbounded in
R. By our assumption, the additive functional y ∗ ◦ f : X → R is upper bounded on some set B ⊂ X which
∗
is not null-finite
in X. By Theorem 9.1, the additive functional y ◦ f is continuous and thus the sequence
∗
n
y ◦ f (2 xn ) n∈ω converges to zero and hence cannot be unbounded in R. This is a desired contradiction,
completing the proof.
10
TARAS BANAKH AND ELIZA JABLOŃSKA
Corollary 10.1 admits a self-generalization.
Theorem 10.2. An additive function f : X → Y from a complete metric group X to a locally convex space Y
is continuous if for any linear continuous functional y ∗ : Y → R the function y ∗ ◦ f : X → R is upper bounded
on some set B ⊂ X which is not null-finite in X.
Proof. This theorem follows from Corollary 10.1 and the well-known fact [27, p.54] that each locally convex
space is topologically isomorphic to a linear subspace of a Tychonoff product of Banach spaces.
Combining Theorem 10.2 with Theorems 5.1 and 6.1, we derive two corollaries.
Corollary 10.3. An additive function f : X → Y from a complete metric group X to a locally convex space Y
is continuous if for any linear continuous functional y ∗ : Y → R the function y ∗ ◦ f : X → R is upper bounded
on some uBP-set B ⊂ X which is not Haar-meager in X.
Corollary 10.4. An additive function f : X → Y from a complete metric group X to a locally convex space Y
is continuous if for any linear continuous functional y ∗ : Y → R the function y ∗ ◦ f : X → R is upper bounded
on some universally measurable set B ⊂ X which is not Haar-null in X.
11. Applying null-finite sets to mid-point convex functions
In this section we apply null-finite sets to establish a continuity criterion for mid-point convex functions on
linear metric spaces.
A function f : C → R on a convex subset C of a linear space is called mid-point convex if
f ( x2 + y2 ) ≤ 12 f (x) + 12 f (y)
for any x, y ∈ C. Mid-point convex functions are alternatively called Jensen convex.
Theorem 11.1. A mid-point convex function f : G → R defined on an open convex subset G ⊂ X of a metric
linear space X is continuous if and only if f is upper bounded on some set B ⊂ G which is not null-finite in
X and whose closure B is contained in G.
Proof. The “only if” part is trivial. To prove the “if” part, assume that f is upper bounded on some set B ⊂ G
such that B ⊂ G and B is not null-finite in X. We need to check the continuity of f at any point c ∈ G.
Shifting the set G and the function f by −c, we may assume that c = θ. Also we can replace the function f
by f − f (θ) and assume that f (θ) = 0. In this case the mid-point convexity of f implies that f ( 21n x) ≤ 21n f (x)
for every x ∈ G and n ∈ N.
To derive a contradiction, suppose that the function f is not continuous at θ. Then there exists ε > 0 such
that f (U ) 6⊂ (−ε, ε) for any open neighborhood U ⊂ G. For every n ∈ ω consider the neighborhood Un := {x ∈
G : kxk ≤ 41n , x ∈ ( 21n G) ∩ (− 21n G)} and find a point xn ∈ Un such that |f (xn )| ≥ ε. Replacing xn by −xn ,
if necessary, we can assume that f (xn ) ≥ ε (this follows from the inequality 12 (f (xn ) + f (−xn )) ≥ f (θ) = 0
ensured by the mid-point convexity of f ).
n
Since xn ∈ 21n G, for every k ≤ n the point 2k xn belongs to G and has prenorm k2k xn k ≤ 2k kxn k ≤ 24n = 21n
in the metric linear space (X, k · k). This implies that (2n xn )n∈ω is a null-sequence in X. Since the set B is
not null-finite, there exists a ∈ X such that the set Ω := {n ∈ N : a + 2n xn ∈ B} is infinite. Then a ∈ B ⊂ G.
Choose k ∈ N so large that 2−k+1 a ∈ −G (such number k exists as −G is a neighborhood of θ).
Observe that for every number n ∈ Ω with n > k the mid-point convexity of f ensures that
−k+1
−k+1
a+2n−k+1 xn
2n−k ε ≤ 2n−k f (xn ) ≤ f (2n−k xn ) = f −2 2 a + 2
≤
2
≤ 21 f (−2−k+1 a) + 12 f (2−k+1 a + 2n−k+1 xn ) ≤ 21 f (−2−k+1 a) + 2−k f (a + 2n xn )
and hence
sup f (B) ≥ sup f (a + 2n xn ) ≥ sup (2n ε − 2k−1 f (−2−k+1 a)) = ∞,
n∈Ω
k<n∈Ω
which contradicts the upper boundedness of f on B.
Theorem 11.1 implies the following generalization of the classical Bernstein-Doetsch Theorem [5], due to
Mehdi [24]; see also a survey paper of Bingham and Ostaszewski [7].
Corollary 11.2. A mid-point-convex function f : G → R defined on an open convex subset G of a metric
linear space X is continuous if and only if f is upper-bounded on some non-empty open subset of G.
NULL-FINITE SETS IN TOPOLOGICAL GROUPS AND THEIR APPLICATIONS
11
Combining Theorem 11.1 with Theorems 5.1 and 6.1, we obtain the following two continuity criteria, which
answer the Problem of Baron and Ger [3, P239].
Corollary 11.3. A mid-point convex function f : G → R defined on a convex subset G of a complete metric
linear space X is continuous if and only if it is upper bounded on some uBP-set B ⊂ G which is not Haar-meager
in X.
Proof. Assume that f is upper bounded on some
S uBP-set B ⊂ G which is not Haar-meager in X.
Write the open set G of X as the union U = n∈ω Fn of
Sclosed subsets of X. By [9], the countable union of
uBP Haar-meager sets is Haar-meager. Since the set B = n∈ω B ∩ Fn is not Haar-meager, for some n ∈ N the
subset B ∩ Fn of B is not Haar-meager. By Theorem 5.1, B ∩ Fn is not null-finite. Since B ∩ Fn ⊂ Fn ⊂ G and
f is upper bounded on B ∩ Fn , we can apply Theorem 11.1 and conclude that the mid-point convex function
f is continuous.
Corollary 11.4. A mid-point convex function f : G → R defined on a convex subset G of a complete metric
linear space X is continuous if and only if it is upper bounded on some universally measurable set B ⊂ G which
is not Haar-null in X.
Proof. The proof of this corollary runs in exactly the same way as the proof of Corollary 11.3 and uses the
well-known fact [8] that the countable union of universally measurable Haar-null sets in a complete metric
group is Haar-null.
12. Some Open Problems
In this section we collect some open problems related to null-finite sets.
It is well-known that for a locally compact Polish group X, each Haar-meager set in X can be enlarged
to a Haar-meager Fσ -set and each Haar-null set in X can be enlarged to a Haar-null Gδ -set in X. Those
“enlargement” results dramatically fail for non-locally compact Polish groups, see [13], [10]. We do not know
what happens with null-finite sets in this respect.
Problem 12.1. Is each Borel null-finite subset A of a (compact) Polish group X contained in a null-finite set
B ⊂ X of low Borel complexity?
By [28] (resp. [12, 4.1.6]), each analytic Haar-null (resp. Haar-meager) set in a Polish group is contained
in a Borel Haar-null (resp. Haar-meager) set. On the other hand, each non-locally compact Polish group
contains a coanalytic Haar-null (resp. Haar-meager) set which cannot be enlarged to a Borel Haar-null (resp.
Haar-meager) set, see [13] (resp. [10]).
Problem 12.2. Is each (co)analytic null-finite set A in a Polish group X contained in a Borel null-finite set?
Our next problem ask about the relation of the σ-ideal σNFX generated by null-finite sets to other known
σ-ideals.
Problem 12.3. Let X be a Polish group.
(1) Is σHNX ⊂ σNFX ?
(2) Is σNFX = HNX ∩ HMX ?
The negative answer to both parts of Problem 12.3 would follow from the negative answer to the following
concrete question.
Q∞
Q∞
Problem 12.4. Can the closed Haar-null set n=2 Cn∗ in the group X = n=2 Cn be written as a countable
union of Borel null-finite sets in X?
Problem 12.5. For an infinite Polish group X and an ideal I ∈ {σNFX , σNFX } evaluate the standard cardinal
characteristics of I:
S
• add(I) = min{|J | : J ⊂ I and J ∈
/ I};
• non(I) = min{|A| : A ⊂ X and S
A∈
/ I};
• cov(I) = min{|J | : J ⊂ I and J = X};
• cof(I) = min{|J | : ∀I ∈ I ∃J ∈ J (I ⊂ J)}.
The cardinal characteristics of the σ-ideals HNX , HMX , σHNX on Polish groups are well-studied [4], [2] and
play an important role in Modern Set Theory.
12
TARAS BANAKH AND ELIZA JABLOŃSKA
References
[1] A. Arhangel’skii, M. Tkachenko, Topological groups and related structures, Atlantis Studies in Mathematics, 1. Atlantis Press,
Paris; World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2008.
[2] T. Banakh, Cardinal characteristics of the ideal of Haar null sets, Comment. Math. Univ. Carolin. 45:1 (2004), 119–137.
[3] K. Baron, R. Ger, Problem (P239), in: The 21st International Symposium on Functional Equations, August 6-13, 1983,
Konolfingen, Switzerland, Aequationes Math. 26 (1984) 225–294.
[4] T. Bartoszynski, H. Judah, Set theory. On the structure of the real line, A K Peters, Ltd., Wellesley, MA, 1995.
[5] F. Bernstein, G. Doetsch, Zur Theorie der konvexen Funktionen, Math. Ann. 76:4 (1915) 514–526.
[6] N.H. Bingham, A.J. Ostaszewski, The Steinhaus-Weil property: its converse, Solecki amenability and subcontinuity, preprint
(https://arxiv.org/abs/1607.00049).
[7] N.H. Bingham, A.J. Ostaszewski, Category-measure duality: convexity, mid-point convexity and Berz sublinearity, preprint
(https://arxiv.org/abs/1607.05750).
[8] J.P.R. Christensen, On sets of Haar measure zero in abelian Polish groups, Israel J. Math. 13 (1972) 255–260.
[9] U.B. Darji, On Haar meager sets, Topology Appl. 160 (2013) 2396–2400.
[10] M. Doležal, V. Vlasǎk, Haar meager sets, their hulls, and relationship to compact sets, J. Math. Anal. Appl. 446:1 (2017)
852–863.
[11] M. Doležal, V. Vlasǎk, B. Vejnar, M. Rmoutil, Haar Meager Sets revisited, J. Math. Anal. Appl. 440:2 (2016) 922–939.
[12] M. Elekes, D. Nagy, Haar null and Haar meager sets: a survey and new results, preprint (https://arxiv.org/abs/1508.02053).
[13] M. Elekes, Z. Vidnyánszky, Haar null sets without Gδ hulls, Israel J. Math. 209 (2015) 199–214.
[14] M. Fabian, P. Habala, P. Hajek, V. Montesinos, J. Pelant, V. Zizler, Functional analysis and infinite-dimensional geometry,
Springer-Verlag, New York, 2001.
[15] P. Fischer, Z. Slodkowski, Christensen zero sets and measurable convex functions, Proc. Amer. Math. Soc. 79 (1980) 449–453.
[16] B.R. Hunt, T. Sauer, J.A. Yorke, Prevalence: a translation-invariant “almost every” on infinite-dimensional spaces, Bull.
Amer. Math. Soc. 27 (1992) 217–238.
[17] B.R. Hunt, T. Sauer, J.A. Yorke, Prevalence: an addendum, Bull. Amer. Math. Soc. 28 (1993) 306–307.
[18] E. Jablońska, Some analogies between Haar meager sets and Haar null sets in abelian Polish groups, J. Math. Anal. Appl.
421 (2015) 1479–1486.
[19] A. Kechris, Classical Descriptive Set Theory, Springer-Verlag, New York, 1995.
[20] M. Kuczma, An introduction to the theory of functional equations and inequalities. Cauchy’s Equation and Jensen’s Inequality,
PWN & Uniwersytet Śla̧ski w Katowicach, Warszawa-Kraków-Katowice, 1985.
[21] A. Kwela, Haar-smallest sets, (https://arxiv.org/abs/1711.09753).
[22] Ie. Lutsenko, I.V. Protasov, Sparse, thin and other subsets of groups, Internat. J. Algebra Comput. 19:4 (2009) 491–510.
[23] E. Matoušková, M. Zelený, A note on intersections of non–Haar null sets, Colloq. Math. 96 (2003) 1–4.
[24] M.R. Mehdi, On convex functions, J. London Math. Soc. 39 (1964) 321–326.
[25] B.J. Pettis, Remarks on a theorem of E. J. McShane, Proc. Amer. Math. Soc. 2 (1951) 166–171.
[26] S. Piccard, Sur les ensembles de distances des ensembles de points d’un espace Euclidien, Mem. Univ. Neuchâtel, vol. 13,
Secrétariat Univ., Neuchâtel, 1939.
[27] H. Schaefer, Topological vector spaces, Springer–Verlag, New York–Berlin, 1971.
[28] S. Solecki, On Haar null sets, Fund. Math. 149 (1996) 205–210.
[29] H. Steinhaus, Sur les distances des points des ensembles de mesure positive, Fund. Math. 1 (1920) 99–104.
T. Banakh: Ivan Franko National University of Lviv (Ukraine) and Jan Kochanowski University in Kielce (Poland)
E-mail address: [email protected]
E. Jablońska: Department of Discrete Mathematics, Rzeszów University of Technology (Poland)
E-mail address: [email protected]
| 4math.GR
|
Generalization in Adaptive Data Analysis and Holdout Reuse
Cynthia Dwork∗
Vitaly Feldman†
Moritz Hardt‡
Omer Reingold¶
Toniann Pitassi§
Aaron Rothk
arXiv:1506.02629v2 [cs.LG] 25 Sep 2015
September 28, 2015
Abstract
Overfitting is the bane of data analysts, even when data are plentiful. Formal approaches to understanding this problem focus on statistical inference and generalization of individual analysis procedures.
Yet the practice of data analysis is an inherently interactive and adaptive process: new analyses and
hypotheses are proposed after seeing the results of previous ones, parameters are tuned on the basis of
obtained results, and datasets are shared and reused. An investigation of this gap has recently been
initiated by the authors in [DFH+ 14], where we focused on the problem of estimating expectations of
adaptively chosen functions.
In this paper, we give a simple and practical method for reusing a holdout (or testing) set to validate
the accuracy of hypotheses produced by a learning algorithm operating on a training set. Reusing a
holdout set adaptively multiple times can easily lead to overfitting to the holdout set itself. We give an
algorithm that enables the validation of a large number of adaptively chosen hypotheses, while provably
avoiding overfitting. We illustrate the advantages of our algorithm over the standard use of the holdout
set via a simple synthetic experiment.
We also formalize and address the general problem of data reuse in adaptive data analysis. We
show how the differential-privacy based approach given in [DFH+ 14] is applicable much more broadly
to adaptive data analysis. We then show that a simple approach based on description length can also
be used to give guarantees of statistical validity in adaptive settings. Finally, we demonstrate that
these incomparable approaches can be unified via the notion of approximate max-information that we
introduce. This, in particular, allows the preservation of statistical validity guarantees even when an
analyst adaptively composes algorithms which have guarantees based on either of the two approaches.
1
Introduction
The goal of machine learning is to produce hypotheses or models that generalize well to the unseen instances
of the problem. More generally, statistical data analysis is concerned with estimating properties of the
underlying data distribution, rather than properties that are specific to the finite data set at hand. Indeed, a
large body of theoretical and empirical research was developed for ensuring generalization in a variety of
settings. In this work, it is commonly assumed that each analysis procedure (such as a learning algorithm)
operates on a freshly sampled dataset – or if not, is validated on a freshly sampled holdout (or testing) set.
Unfortunately, learning and inference can be more difficult in practice, where data samples are often
reused. For example, a common practice is to perform feature selection on a dataset, and then use the
features for some supervised learning task. When these two steps are performed on the same dataset, it is no
∗ Microsoft
Research
Almaden Research Center. Part of this work done while visiting the Simons Institute, UC Berkeley
‡ Google Research
§ University of Toronto
¶ Samsung Research America
k Department of Computer and Information Science, University of Pennsylvania
† IBM
1
longer clear that the results obtained from the combined algorithm will generalize. Although not usually
understood in these terms, “Freedman’s paradox” is an elegant demonstration of the powerful (negative) effect
of adaptive analysis on the same data [Fre83]. In Freedman’s simulation, variables with significant t-statistic
are selected and linear regression is performed on this adaptively chosen subset of variables, with famously
misleading results: when the relationship between the dependent and explanatory variables is non-existent,
the procedure overfits, erroneously declaring significant relationships.
Most of machine learning practice does not rely on formal guarantees of generalization for learning
algorithms. Instead a dataset is split randomly into two (or sometimes more) parts: the training set and the
testing, or holdout, set. The training set is used for learning a predictor, and then the holdout set is used
to estimate the accuracy of the predictor on the true distribution1 . Because the predictor is independent
of the holdout dataset, such an estimate is a valid estimate of the true prediction accuracy (formally, this
allows one to construct a confidence interval for the prediction accuracy on the data distribution). However,
in practice the holdout dataset is rarely used only once, and as a result the predictor may not be independent
of the holdout set, resulting in overfitting to the holdout set [Reu03, RF08, CT10]. One well-known reason
for such dependence is that the holdout data is used to test a large number of predictors and only the best
one is reported. If the set of all tested hypotheses is known and independent of the holdout set, then it is
easy to account for such multiple testing or use the more sophisticated approach of Ng [Ng97].
However such static approaches do not apply if the estimates or hypotheses tested on the holdout are
chosen adaptively: that is, if the choice of hypotheses depends on previous analyses performed on the
dataset. One prominent example in which a holdout set is often adaptively reused is hyperparameter tuning
(e.g.[DFN07]). Similarly, the holdout set in a machine learning competition, such as the famous ImageNet
competition, is typically reused many times adaptively. Other examples include using the holdout set for
feature selection, generation of base learners (in aggregation techniques such as boosting and bagging),
checking a stopping condition, and analyst-in-the-loop decisions. See [Lan05] for a discussion of several subtle
causes of overfitting.
The concrete practical problem we address is how to ensure that the holdout set can be reused to perform
validation in the adaptive setting. Towards addressing this problem we also ask the more general question
of how one can ensure that the final output of adaptive data analysis generalizes to the underlying data
distribution. This line of research was recently initiated by the authors in [DFH+ 14], where we focused on
the case of estimating expectations of functions from i.i.d. samples (these are also referred to as statistical
queries). They show how to answer a large number of adaptively chosen statistical queries using techniques
from differential privacy [DMNS06](see Sec. 1.3 and Sec. 2.2 for more details).
1.1
Our Results
We propose a simple and general formulation of the problem of preserving statistical validity in adaptive
data analysis. We show that the connection between differentially private algorithms and generalization from
[DFH+ 14] can be extended to this more general setting, and show that similar (but sometimes incomparable)
guarantees can be obtained from algorithms whose outputs can be described by short strings. We then
define a new notion, approximate max-information, that unifies these two basic techniques and gives a new
perspective on the problem. In particular, we give an adaptive composition theorem for max-information,
which gives a simple way to obtain generalization guarantees for analyses in which some of the procedures
are differentially private and some have short description length outputs. We apply our techniques to the
problem of reusing the holdout set for validation in the adaptive setting.
1.1.1
A Reusable Holdout
We describe a simple and general method, together with two specific instantiations, for reusing a holdout
set for validating results while provably avoiding overfitting to the holdout set. The analyst can perform
any analysis on the training dataset, but can only access the holdout set via an algorithm that allows the
1 Additional
averaging over different partitions is used in cross-validation.
2
analyst to validate her hypotheses against the holdout set. Crucially, our algorithm prevents overfitting to the
holdout set even when the analysts hypotheses are chosen adaptively on the basis of the previous responses
of our algorithm.
Our first algorithm, referred to as Thresholdout, derives its guarantees from differential privacy and the
results in [DFH+ 14, NS15]. For any function φ : X → [0, 1] given by the analyst, Thresholdout uses the
holdout set to validate that φ does not overfit to the training set, that is, it checks that the mean value of φ
evaluated on the training set is close to the mean value of φ evaluated on the distribution P from which the
data was sampled. The standard approach to such validation would be to compute the mean value of φ on the
holdout set. The use of the holdout set in Thresholdout differs from the standard use in that it exposes very
little information about the mean of φ on the holdout set: if φ does not overfit to the training set, then the
analyst receives only the confirmation of closeness, that is, just a single bit. On the other hand, if φ overfits
then Thresholdout returns the mean value of φ on the training set perturbed by carefully calibrated noise.
Using results from [DFH+ 14, NS15] we show that for datasets consisting of i.i.d. samples these modifications
provably prevent the analyst from constructing functions that overfit to the holdout set. This ensures
correctness of Thresholdout’s responses. Naturally, the specific guarantees depend on the number of samples
n in the holdout set. The number of queries that Thresholdout can answer is exponential in n as long as the
number of times that the analyst overfits is at most quadratic in n.
Our second algorithm SparseValidate is based on the idea that if most of the time the analysts procedures
generate results that do not overfit, then validating them against the holdout set does not reveal much
information about the holdout set. Specifically, the generalization guarantees of this method follow from
the observation that the transcript of the interaction between a data analyst and the holdout set can be
described concisely. More formally, this method allows the analyst to pick any Boolean function of a dataset
ψ (described by an algorithm) and receive back its value on the holdout set. A simple example of such a
function would be whether the accuracy of a predictor on the holdout set is at least a certain value α. (Unlike
in the case of Thresholdout, here there is no need to assume that the function that measures the accuracy
has a bounded range or even Lipschitz, making it qualitatively different from the kinds of results achievable
subject to differential privacy). A more involved example of validation would be to run an algorithm on
the holdout dataset to select an hypothesis and check if the hypothesis is similar to that obtained on the
training set (for any desired notion of similarity). Such validation can be applied to other results of analysis;
for example one could check if the variables selected on the holdout set have large overlap with those selected
on the training set. An instantiation of the SparseValidate algorithm has already been applied to the problem
of answering statistical (and more general) queries in the adaptive setting [BSSU15]. We describe the formal
guarantees for SparseValidate in Section 4.2.
In Section 5 we describe a simple experiment on synthetic data that illustrates the danger of reusing
a standard holdout set, and how this issue can be resolved by our reusable holdout. The design of this
experiment is inspired by Freedman’s classical experiment, which demonstrated the dangers of performing
variable selection and regression on the same data [Fre83].
1.2
Generalization in Adaptive Data Analysis
We view adaptive analysis on the same dataset as an execution of a sequence of steps A1 → A2 → · · · → Am .
Each step is described by an algorithm Ai that takes as input a fixed dataset S = (x1 , . . . , xn ) drawn from
some distribution D over X n , which remains unchanged over the course of the analysis. Each algorithm Ai
also takes as input the outputs of the previously run algorithms A1 through Ai−1 and produces a value in
some range Yi . The dependence on previous outputs represents all the adaptive choices that are made at step
i of data analysis. For example, depending on the previous outputs, Ai can run different types of analysis on
S. We note that at this level of generality, the algorithms can represent the choices of the data analyst, and
need not be explicitly specified. We assume that the analyst uses algorithms which individually are known to
generalize when executed on a fresh dataset sampled independently from a distribution D. We formalize
this by assuming that for every fixed value y1 , . . . , yi−1 ∈ Y1 × · · · × Yi−1 , with probability at least 1 − βi
over the choice of S according to distribution D, the output of Ai on inputs y1 , . . . , yi−1 and S has a desired
property relative to the data distribution D (for example has low generalization error). Note that in this
3
assumption y1 , . . . , yi−1 are fixed and independent of the choice of S, whereas the analyst will execute Ai on
values Y1 , . . . , Yi−1 , where Yj = Aj (S, Y1 , . . . , Yj−1 ). In other words, in the adaptive setup, the algorithm
Ai can depend on the previous outputs, which depend on S, and thus the set S given to Ai is no longer
an independently sampled dataset. Such dependence invalidates the generalization guarantees of individual
procedures, potentially leading to overfitting.
Differential privacy: First, we spell out how the differential privacy based approach from [DFH+ 14] can
be applied to this more general setting. Specifically, a simple corollary of results in [DFH+ 14] is that for a
dataset consisting of i.i.d. samples any output of a differentially-private algorithm can be used in subsequent
analysis while controlling the risk of overfitting, even beyond the setting of statistical queries studied in
[DFH+ 14]. A key property of differential privacy in this context is that it composes adaptively: namely if each
of the algorithms used by the analyst is differentially private, then the whole procedure will be differentially
private (albeit with worse privacy parameters). Therefore, one way to avoid overfitting in the adaptive setting
is to use algorithms that satisfy (sufficiently strong) guarantees of differential-privacy. In Section 2.2 we
describe this result formally.
Description length: We then show how description length bounds can be applied in the context of
guaranteeing generalization in the presence of adaptivity. If the total length of the outputs of algorithms
A1 , . . . , Ai−1 can be described with k bits then there are at most 2k possible values of the input y1 , . . . , yi−1
to Ai . For each of these individual inputs Ai generalizes with probability 1 − βi . Taking a union bound over
failure probabilities implies generalization with probability at least 1 − 2k βi . Occam’s Razor famously implies
that shorter hypotheses have lower generalization error. Our observation is that shorter hypotheses (and the
results of analysis more generally) are also better in the adaptive setting since they reveal less about the
dataset and lead to better generalization of subsequent analyses. Note that this result makes no assumptions
about the data distribution D. We provide the formal details in Section 2.3. In Section B we also show that
description length-based analysis suffices for obtaining an algorithm (albeit not an efficient one) that can
answer an exponentially large number of adaptively chosen statistical queries. This provides an alternative
proof for one of the results in [DFH+ 14].
Approximate max-information: Our main technical contribution is the introduction and analysis of
a new information-theoretic measure, which unifies the generalization arguments that come from both
differential privacy and description length, and that quantifies how much information has been learned about
the data by the analyst. Formally, for jointly distributed random variables (S, Y ), the max-information is
the maximum of the logarithm of the factor by which uncertainty about S is reduced given the value of
.
| Y =y]
Y , namely I∞ (S, Y ) = log max P[S=S
, where the maximum is taken over all S in the support of S
P[S=S]
and y in the support Y . Informally, β-approximate max-information requires that the logarithm above be
bounded with probability at least 1 − β over the choice of (S, Y ) (the actual definition is slightly weaker, see
Definition 10 for details).In our use, S denotes a dataset drawn randomly from the distribution D and Y
denotes the output of a (possibly randomized) algorithm on S. We prove that approximate max-information
has the following properties
• An upper bound on (approximate) max-information gives generalization guarantees.
• Differentially private algorithms have low max-information for any distribution D over datasets. A
stronger bound holds for approximate max-information on i.i.d. datasets. These bounds apply only to
so-called pure differential privacy (the δ = 0 case).
• Bounds on the description length of the output of an algorithm give bounds on the approximate
max-information of the algorithm for any D.
• Approximate max-information composes adaptively.
• Approximate max-information is preserved under post-processing.
4
Composition properties of approximate max-information imply that one can easily obtain generalization
guarantees for adaptive sequences of algorithms, some of which are differentially private, and others of which
have outputs with short description length. These properties also imply that differential privacy can be used
to control generalization for any distribution D over datasets, which extends its generalization guarantees
beyond the restriction to datasets drawn i.i.d. from a fixed distribution, as in [DFH+ 14].
We remark that (pure) differential privacy and description length are otherwise incomparable – low
description length is not a sufficient condition for differential privacy, since differential privacy precludes
revealing even a small number of bits of information about any single individual in the data set. At the same
time differential privacy does not constrain the description length of the output. Bounds on max-information
or differential privacy of an algorithm can, however, be translated to bounds on randomized description length
for a different algorithm with statistically indistinguishable output. Here we say that a randomized algorithm
has randomized description length of k if for every fixing of the algorithm’s random bits, it has description
length of k. Details of these results and additional discussion appear in Sections 3 and A.
1.3
Related Work
This work builds on [DFH+ 14] where we initiated the formal study of adaptivity in data analysis. The
primary focus of [DFH+ 14] is the problem of answering adaptively chosen statistical queries. The main
technique is a strong connection between differential privacy and generalization: differential privacy guarantees
that the distribution of outputs does not depend too much on any one of the data samples, and thus,
differential privacy gives a strong stability guarantee that behaves well under adaptive data analysis. The
link between generalization and approximate differential privacy made in [DFH+ 14] has been subsequently
strengthened, both qualitatively — by [BSSU15], who make the connection for a broader range of queries —
and quantitatively, by [NS15] and [BSSU15], who give tighter quantitative bounds. These papers, among
other results, give methods for accurately answering exponentially (in the dataset size) many adaptively
chosen queries, but the algorithms for this task are not efficient. It turns out this is for fundamental reasons
– Hardt and Ullman [HU14] and Steinke and Ullman [SU14] prove that, under cryptographic assumptions,
no efficient algorithm can answer more than quadratically many statistical queries chosen adaptively by an
adversary who knows the true data distribution.
Differential privacy emerged from a line of work [DN03, DN04, BDMN05], culminating in the definition
given by [DMNS06]. There is a very large body of work designing differentially private algorithms for various
data analysis tasks, some of which we leverage in our applications. See [Dwo11] for a short survey and [DR14]
for a textbook introduction to differential privacy.
The classical approach in theoretical machine learning to ensure that empirical estimates generalize to
the underlying distribution is based on the various notions of complexity of the set of functions output by
the algorithm, most notably the VC dimension(see e.g. [SSBD14] for a textbook introduction). If one has a
sample of data large enough to guarantee generalization for all functions in some class of bounded complexity,
then it does not matter whether the data analyst chooses functions in this class adaptively or non-adaptively.
Our goal, in contrast, is to prove generalization bounds without making any assumptions about the class
from which the analyst can choose query functions. In this case the adaptive setting is very different from
the non-adaptive setting.
An important line of work [BE02, MNPR06, PRMN04, SSSSS10] establishes connections between the
stability of a learning algorithm and its ability to generalize. Stability is a measure of how much the output
of a learning algorithm is perturbed by changes to its input. It is known that certain stability notions are
necessary and sufficient for generalization. Unfortunately, the stability notions considered in these prior
works do not compose in the sense that running multiple stable algorithms sequentially and adaptively may
result in a procedure that is not stable. The measure we introduce in this work (max information), like
differential privacy, has the strength that it enjoys adaptive composition guarantees. This makes it amenable
to reasoning about the generalization properties of adaptively applied sequences of algorithms, while having
to analyze only the individual components of these algorithms. Connections between stability, empirical risk
minimization and differential privacy in the context of learnability have been recently explored in [WLF15].
5
Freund gives an approach to obtaining data-dependent generalization bounds that takes into account the
set of statistical queries that a given learning algorithm can produce for the distribution from which the
data was sampled [Fre98]. A related approach of Langford and Blum also allows to obtain data-dependent
generalization bounds based on the description length of functions that can be output for a data distribution
[LB03]. Unlike our work, these approaches require the knowledge of the structure of the learning algorithm
to derive a generalization bound. More importantly, the focus of our framework is on the design of new
algorithms with better generalization properties in the adaptive setting.
Finally, inspired by our work, Blum and Hardt [BH15] showed how to reuse the holdout set to maintain
an accurate leaderboard in a machine learning competition that allows the participants to submit adaptively
chosen models in the process of the competition (such as those organized by Kaggle Inc.). Their analysis also
relies on the description length-based technique we used to analyze SparseValidate.
2
Preliminaries and Basic Techniques
In the discussion below log refers to binary logarithm and ln refers to the natural logarithm. For simplicity
we restrict our random variables to finite domains (extension of the claims to continuous domains is
straightforward using the standard formalism). For two random variables X and Y over the same domain X
the max-divergence of X from Y is defined as
D∞ (XkY ) = log max
x∈X
P[X = x]
.
P[Y = x]
δ-approximate max-divergence is defined as
δ
D∞
(XkY ) = log
max
O⊆X , P[X∈O]>δ
P[X ∈ O] − δ
.
P[Y ∈ O]
We say that a real-valued function over datasets f : X n → R has sensitivity c for all i ∈ [n] and
x1 , x2 , . . . , xn , x0i ∈ X , f (x1 , . . . , xi , . . . , xn )−f (x1 , . . . , x0i , . . . , xn ) ≤ c. We review McDiarmid’s concentration
inequality for functions of low-sensitivity.
Lemma 1 (McDiarmid’s inequality). Let X1 , X2 , . . . , Xn be independent random variables taking values
in the set X . Further let f : X n → R be a function of sensitivity c > 0. Then for all α > 0, and
µ = E [f (X1 , . . . , Xn )],
h
i
−2α2
P f (X1 , . . . , Xn ) − µ ≥ α ≤ exp
.
n · c2
¯
. Pn
For a function φ : X → R and a dataset S = (x1 , . . . , xn ), let ES [φ] = n1 i=1 φ(xi ). Note that if the
range of φ is in some interval of length α then f (S) = ES [φ] has sensitivity α/n. For a distribution P over X
.
and a function φ : X → R, let P[φ] = Ex∼P [φ(x)].
2.1
Differential Privacy
On an intuitive level, differential privacy hides the data of any single individual. We are thus interested in
pairs of datasets S, S 0 that differ in a single element, in which case we say S and S 0 are adjacent.
Definition 2. [DMNS06, DKM+ 06] A randomized algorithm A with domain X n for n > 0 is (ε, δ)δ
differentially private if for all pairs of datasets that differ in a single element S, S 0 ∈ X n : D∞
(A(S)kA(S 0 )) ≤
ε
log(e ). The case when δ = 0 is sometimes referred to as pure differential privacy, and in this case we may
say simply that A is ε-differentially private.
Differential privacy is preserved under adaptive composition. Adaptive composition of algorithms is
a sequential execution of algorithms on the same dataset in which an algorithm at step i can depend on
the outputs of previous algorithms. More formally, let A1 , A2 , . . . , Am be a sequence of algorithms. Each
6
algorithm Ai outputs a value in some range Yi and takes as an input dataset in X n as well as a value in
.
Ȳi−1 = Y1 × · · · × Yi−1 . Adaptive composition of these algorithm is the algorithm that takes as an input
a dataset S ∈ X n and executes A1 → A2 → · · · → Am sequentially with the input to Ai being S and the
outputs y1 , . . . , yi−1 of A1 , . . . , Ai−1 . Such composition captures the common practice in data analysis of
using the outcomes of previous analyses (that is y1 , . . . , yi−1 ) to select an algorithm that is executed on S.
For an algorithm that in addition to a dataset has other input we say that it is (ε, δ)-differentially private
if it is (ε, δ)-differentially private for every setting of additional parameters. The basic property of adaptive
composition of differentially private algorithms is the following result (e.g.[DL09]):
algorithm for i ∈ [m].
Theorem 3. Let Ai : X n × Y1 × · · · × Yi−1 → Yi be an (εi , δi )-differentiallyPprivate P
m
m
Then the algorithm B : X n → Ym obtained by composing Ai ’s adaptively is ( i=1 εi , i=1 δi )-differentially
private.
A more sophisticated argument yields significant improvement when ε < 1 (e.g.[DR14]).
Theorem 4. For all ε, δ, δ 0 ≥ 0, the adaptive composition of m arbitrary (ε, δ)-differentially private algorithms
is (ε0 , mδ + δ 0 )-differentially private, where
p
ε0 = 2m ln(1/δ 0 ) · ε + mε(eε − 1).
Another property of differential privacy important for our applications is preservation of its guarantee
under post-processing (e.g.[DR14, Prop. 2.1]):
Lemma 5. If A is an (, δ)-differentially private algorithm with domain X n and range Y, and B is any,
possibly randomized, algorithm with domain Y and range Y 00 , then the algorithm B ◦ A with domain X n and
range Y 0 is also (, δ)-differentially private.
2.2
Generalization via Differential Privacy
Generalization in special cases of our general adaptive analysis setting can be obtained directly from results in
[DFH+ 14] and composition properties of differentially private algorithms. For the case of pure differentially
private algorithms with general outputs over i.i.d. datasets, in [DFH+ 14] we prove the following result.
Theorem 6. Let A be an ε-differentially private algorithm with range Y and let S be a random variable
drawn from a distribution P n over X n . Let Y = A(S) be the corresponding output distribution. Assume
q that
for each element y ∈ Y there is a subset R(y) ⊆ X n so that maxy∈Y P[S ∈ R(y)] ≤ β. Then, for ε ≤
√
we have P[S ∈ R(Y )] ≤ 3 β.
ln(1/β)
2n
An immediate corollary of Thm. 6 together with Lemma 1 is that differentially private algorithms that
output low-sensitivity functions generalize.
Corollary 7. Let A be an algorithm that outputs a c-sensitive function f : X n → R. Let S be a random
dataset chosen according to distribution P n over X n and let f = A(S). If A is τ /(cn)-differentially private
then P[f (S) − P n [f ] ≥ τ ] ≤ 3 exp (−τ 2 /(c2 n)).
By Theorem 3, pure differential privacy composes adaptively. Therefore, if in a sequence of algorithms
A1 , A2 , . . . , Am algorithm Ai is εi -differentially private
P for all i ≤ m − 1 then composition of the first i − 1
i−1
0
0
algorithms is εi−1 -differentially private for εi−1 =
j=1 εj . Theorem 6 can be applied to preserve the
generalization guarantees of the last algorithm Am (that does not need to be differentially private). For
example, assume that for every fixed setting of ȳm−1 , Am has the property that it outputs a hypothesis
2
function h such that, P[ES [L(h)] − P[L(h)] ≥ τ ] ≤ e−nτ /d , for some notion of dimension d and a real-valued
loss function L. Generalization bounds of this type follow from uniform convergence arguments based on
various notions of complexity of hypotheses classes such as VC dimension, covering numbers, fat-shattering
dimension and Rademacher complexity (see [SSBD14] for examples). Note that, for different settings of ȳm−1 ,
7
different sets of hypotheses and generalization techniques might be used. We define R(ȳp
m−1 ) be all datasets
S for which Am (S, ȳm−1 ) outputs h such that ES [L(h)] − P[L(h)] ≥ τ . Now if ε0m−1 ≤ τ 2 /(2d), then even
for the hypothesis output in the adaptive execution of Am on a random i.i.d. dataset S (denoted by h) we
2
have P [ES [L(h)] − P[L(h)] ≥ τ ] ≤ 3e−τ n/(2d) .
For approximate (ε, δ)-differential privacy, strong preservation of generalization results are currently
known only for algorithms that output a function over X of bounded range (for simplicity we use range [0, 1])
[DFH+ 14, NS15]. The following result was proved by Nissim and Stemmer [NS15] (a weaker statement is
also given in [DFH+ 14, Thm. 10]).
Theorem 8. Let A be an (ε, δ)-differentially private algorithm that outputs a function from X to [0, 1]. For
a random variable S distributed according to P n we let φ = A(S). Then for n ≥ 2 ln(8/δ)/ε2 ,
2δ
2
P [|P[φ] − ES [φ]| ≥ 13ε] ≤
ln
.
ε
ε
Many learning algorithms output a hypothesis function that aims to minimize some bounded loss function
L as the final output. If algorithms used in all steps of the adaptive data analysis are differentially private and
the last step (that is, Am ) outputs a hypothesis h, then generalization bounds for the loss of h are implied
directly by Theorem 8. We remark that this application is different from the example for pure differential
privacy above since there we showed preservation of generalization guarantees of arbitrarily complex learning
algorithm Am which need not be differentially private. In Section 4 we give an application of Theorem 8 to
the reusable holdout problem.
2.3
Generalization via Description Length
Let A : X n → Y and B : X n × Y → Y 0 be two algorithms. We now give a simple application of bounds on
the size of Y (or, equivalently, the description length of A’s output) to preserving generalization of B. Here
generalization can actually refer to any valid or desirable output of B for a given given dataset S and input
y ∈ Y. Specifically we will use a set R(y) ⊆ X n to denote all datasets for which the output of B on y and S
is “bad” (e.g. overfits). Using a simple union bound we show that the probability (over a random choice of a
dataset) of such bad outcome can be bounded.
Theorem 9. Let A : X n → Y be an algorithm and let S be a random dataset over X n . Assume that
n
R : Y → 2X is such that for every y ∈ Y, P[S ∈ R(y)] ≤ β. Then P[S ∈ R(A(S))] ≤ |Y| · β.
Proof.
P[S ∈ R(A(S))] ≤
X
P[S ∈ R(y)] ≤ |Y| · β.
y∈Y
The case of two algorithms implies the general case since description length composes (adaptively).
Namely, let A1 , A2 , . . . , Am be a sequence of algorithms such that each algorithm Ai outputs a value in some
range Yi and takes as an input dataset in X n as well as a value in Ȳi−1 . Then for every i, we can view
the execution of A1 through Ai−1 as the first algorithm Āi−1 with an output in Ȳi−1 and Ai as the second
algorithm. Theorem 9 implies that if for every setting of ȳi−1 = y1 , . . . , yi−1 ∈ Ȳi−1 , R(ȳi−1 ) ⊆ X n satisfies
that P[S ∈ R(ȳi−1 )] ≤ βi then
P[S ∈ R(Āi−1 (S))] ≤ |Ȳi−1 | · βi =
i−1
Y
|Yj | · βi .
j=1
In Section A we describe a generalization of description length bounds to randomized algorithms and
show that it possesses the same properties.
8
3
Max-Information
Consider two algorithms A : X n → Y and B : X n × Y → Y 0 that are composed adaptively and assume that for
every fixed input y ∈ Y, B generalizes for all but fraction β of datasets. Here we are speaking of generalization
informally: our definitions will support any property of input y ∈ Y and dataset S. Intuitively, to preserve
generalization of B we want to make sure that the output of A does not reveal too much information about
the dataset S. We demonstrate that this intuition can be captured via a notion of max-information and its
relaxation approximate max-information.
For two random variables X and Y we use X × Y to denote the random variable obtained by drawing
X and Y independently from their probability distributions.
Definition 10. Let X and Y be jointly distributed random variables. The max-information between X and
Y , denoted I∞ (X; Y ), is the minimal value of k such that for every x in the support of X and y in the
support of Y we have P[X = x | Y = y] ≤ 2k P[X = x]. Alternatively, I∞ (X; Y ) = D∞ ((X, Y )kX × Y ).
β
β
The β-approximate max-information is defined as I∞
(X; Y ) = D∞
((X, Y )kX × Y ).
β
β
It follows immediately from Bayes’ rule that for all β ≥ 0, I∞
(X; Y ) = I∞
(Y ; X). Further, I∞ (X; Y ) ≤ k
if and only if for all x in the support of X, D∞ (Y | X = x k Y ) ≤ k. Clearly, max-information upper
bounds the classical notion of mutual information: I∞ (X; Y ) ≥ I(X; Y ).
In our use (X, Y ) is going to be a joint distribution (S, A(S)), where S is a random n-element dataset
and A is a (possibly randomized) algorithm taking a dataset as an input. If the output of an algorithm on any
distribution S has low approximate max-information then we say that the algorithm has low max-information.
More formally:
Definition 11. We say that an algorithm A has β-approximate max-information of k if for every distribution
β
S over n-element datasets, I∞
(S; A(S)) ≤ k, where S is a dataset chosen randomly according to S. We
β
denote this by I∞ (A, n) ≤ k.
An alternative way to define the (pure) max-information of an algorithm is using the maximum of the
infinity divergence between distributions on two different inputs.
Lemma 12. Let A be an algorithm with domain X n and range Y. Then I∞ (A, n) = maxS,S 0 ∈X n D∞ (A(S)kA(S 0 )).
Proof. For the first direction let k = maxS,S 0 ∈X n D∞ (A(S)kA(S 0 )). Let S be any random variable over
n-element input datasets for A and let Y be the corresponding output distribution Y = A(S). We will argue
that I∞ (Y ; S) ≤ k, that I∞ (S; Y ) ≤ k follows immediately from the Bayes’ rule. For every y ∈ Y, there
must exist a dataset Sy such that P[Y = y | S = Sy ] ≤ P[Y = y]. Now, by our assumption, for every S,
P[Y = y | S = S] ≤ 2k · P[Y = y | S = Sy ]. We can conclude that for every S and every y, it holds that
P[Y = y | S = S] ≤ 2k P[Y = y]. This yields I∞ (Y ; S) ≤ k.
For the other direction let k = I∞ (A, n), let S, S 0 ∈ X n and y ∈ Y. For α ∈ (0, 1), let S be the random
variable equal to S with probability α and to S 0 with probability 1 − α and let Y = A(S). By our assumption,
I∞ (Y ; S) = I∞ (S; Y ) ≤ k. This gives
P[Y = y | S = S] ≤ 2k P[Y = y] ≤ 2k (α P[Y = y | S = S] + (1 − α) P[Y = y | S = S 0 ])
and implies
P[Y = y | S = S] ≤
2k (1 − α)
· P[Y = y | S = S 0 ].
1 − 2k α
This holds for every α > 0 and therefore
P[Y = y | S = S] ≤ 2k · P[Y = y | S = S 0 ].
Using this inequality for every y ∈ Y we obtain D∞ (A(S)kA(S 0 )) ≤ k.
9
Generalization via max-information: An immediate corollary of our definition of approximate maxinformation is that it controls the probability of “bad events” that can happen as a result of the dependence
of A(S) on S.
Theorem 13. Let S be a random dataset in X n and A be an algorithm with range Y such that for some
β
β ≥ 0, I∞
(S; A(S)) = k. Then for any event O ⊆ X n × Y,
P[(S, A(S)) ∈ O] ≤ 2k · P[S × A(S) ∈ O] + β.
In particular, P[(S, A(S)) ∈ O] ≤ 2k · maxy∈Y P[(S, y) ∈ O] + β.
We remark that mutual information between S and A(S) would not suffice for ensuring that bad events
happen with tiny probability. For example mutual information of k allows P[(S, A(S)) ∈ O] to be as high as
k/(2 log(1/δ)), where δ = P[S × A(S) ∈ O].
Composition of max-information: Approximate max-information satisfies the following adaptive composition property:
β1
Lemma 14. Let A : X n → Y be an algorithm such that I∞
(A, n) ≤ k1 , and let B : X n × Y → Z be an
algorithm such that for every y ∈ Y, B(·, y) has β2 -approximate max-information k2 . Let C : X n → Z be
β1 +β2
defined such that C(S) = B(S, A(S)). Then I∞
(C, n) ≤ k1 + k2 .
Proof. Let D be a distribution over X n and S be a random dataset sampled from D. By hypothesis,
β1
I∞
(S; A(S)) ≤ k1 . Expanding out the definition for all O ⊆ X n × Y:
P[(S, A(S)) ∈ O] ≤ 2k1 · P[S × A(S) ∈ O] + β1 .
We also have for all Q ⊆ X n × Z and for all y ∈ Y:
P[(S, B(S, y)) ∈ Q] ≤ 2k2 · P[S × B(S, y) ∈ Q] + β2 .
For every O ⊆ X n × Y, define
µ(O) = P[(S, A(S)) ∈ O] − 2k1 · P[S × A(S) ∈ O] + .
Observe that µ(O) ≤ β1 for all O ⊆ X n × Y. For any event Q ⊆ X n × Z, we have:
10
P[(S, C(S)) ∈ Q]
= P[(S, B(S, A(S))) ∈ Q]
X
=
P[(S, B(S, y)) ∈ Q] · P[S = S, A(S) = y]
S∈X n ,y∈Y
≤
X
min
2k2 · P[S × B(S, y) ∈ Q] + β2 , 1 · P[S = S, A(S) = y]
S∈X n ,y∈Y
≤
X
min 2k2 · P[S × B(S, y) ∈ Q], 1 + β2 · P[S = S, A(S) = y]
S∈X n ,y∈Y
≤
X
min 2k2 · P[S × B(S, y) ∈ Q], 1 · P[S = S, A(S) = y] + β2
S∈X n ,y∈Y
≤
X
min 2k2 · P[S × B(S, y) ∈ Q], 1 · 2k1 · P[S = S] · P[A(S) = y] + µ(S, y) + β2
S∈X n ,y∈Y
≤
X
X
µ(S, y) + β2
S∈X n ,y∈Y
S∈X n ,y∈Y
≤
X
min 2k2 · P[S × B(S, y) ∈ Q], 1 · 2k1 · P[S = S] · P[A(S) = y] +
min 2k2 · P[S × B(S, y) ∈ Q], 1 · 2k1 · P[S = S] · P[A(S) = y] + β1 + β2
S∈X n ,y∈Y
≤ 2k1 +k2 ·
X
P[S × B(S, y) ∈ Q] · P[S = S] · P[A(S) = y] + β1 + β2
S∈X n ,y∈Y
=
2k1 +k2 · P[S × B(S, A(S)) ∈ Q] + (β1 + β2 ) .
β1 +β2
Applying the definition of max-information, we see that equivalently, I∞
(S; C(S)) ≤ k1 + k2 , which is
what we wanted.
This lemma can be iteratively applied, which immediately yields the following adaptive composition
theorem for max-information:
Theorem 15. Consider an arbitrary sequence of algorithms A1 , . . . , Ak with ranges Y1 , . . . , Yk such that for
all i, Ai : X n × Y1 × . . . × Yi−1 → Yi is such that Ai (·, y1 , . . . , yi−1 ) has βi -approximate max-information ki
for all choices of y1 , . . . , yi−1 ∈ Y1 × . . . × Yi−1 . Let the algorithm B : X n → Yk be defined as follows:
B(S):
1. Let y1 = A1 (S).
2. For i = 2 to k: Let yi = Ai (S, y1 , . . . , yi−1 )
3. Output yk
P
P
Then B has ( i βi )-approximate max-information ( i ki ).
Post-processing of Max-information: Another useful property that (approximate) max-information
shares with differential privacy is preservation under post-processing. The simple proof of this lemma is
identical to that for differential privacy (Lemma 5) and hence is omitted.
Lemma 16. If A is an algorithm with domain X n and range Y, and B is any, possibly randomized, algorithm
with domain Y and range Y 00 , then the algorithm B ◦ A with domain X n and range Y 0 satisfies: for every
β
β
random variable S over X n and every β ≥ 0, I∞
(S; B ◦ A(S)) ≤ I∞
(S; A(S)).
11
3.1
Bounds on Max-information
We now show that the basic approaches based on description length and (pure) differential privacy are
captured by approximate max-information.
3.1.1
Description Length
Description length k gives the following bound on max-information.
Theorem 17. Let A be a randomized algorithm taking as an input an n-element dataset and outputting a
β
value in a finite set Y. Then for every β > 0, I∞
(A, n) ≤ log(|Y|/β).
We will use the following simple property of approximate divergence (e.g.[DR14]) in the proof. For a
random variable X over X we denote by p(X) the probability distribution associated with X.
Lemma 18. Let X and Y be two random variables over the same domain X . If
P[X = x]
P
≥ 2k ≤ β
x∼p(X) P[Y = x]
β
(XkY ) ≤ k.
then D∞
Proof Thm. 17. Let S be any random variable over n-element input datasets and let Y be the corresponding
β
output distribution Y = A(S). We prove that for every β > 0, I∞
(S; Y ) ≤ log(|Y|/β).
For y ∈ Y we say that y is “bad” if exists S in the support of S such that
P[Y = y | S = S]
≥ |Y|/β.
P[Y = y]
Let B denote the set of all “bad” y’s. From this definition we obtain that for a “bad” y, P[Y = y] ≤ β/|Y|
and therefore P[Y ∈ B] ≤ β. Let B = X n × B. Then
P[(S, Y ) ∈ B] = P[Y ∈ B] ≤ β.
For every (S, y) 6∈ B we have that
P[S = S, Y = y] = P[Y = y | S = S] · P[S = S] ≤
and hence
P
(S,y)∼p(S,Y )
|Y|
· P[Y = y] · P[S = S],
β
P[S = S, Y = y]
|Y|
≥
≤ β.
P[S = S] · P[Y = y]
β
β
This, by Lemma 18, gives that I∞
(S; Y ) ≤ log(|Y|/β).
.
We note that Thms. 13 and 17 give a slightly weaker form of Thm. 9. Defining
p event O = {(S, y) | S ∈ R(y)},
0
the assumptions of Thm. 9 imply that P[S × A(S) ∈ O] ≤ β. For β = |Y|β, by Thm. 17, we have
p that
β0
I∞
(S; A(S)) ≤ log(|Y|/β 0 ). Now applying, Thm. 13 gives that P[S ∈ R(A(S))] ≤ |Y|/β 0 · β + β 0 = 2 |Y|β.
In Section A we introduce a closely related notion of randomized description length and show that it also
provides an upper bound on approximate max-information. More interestingly, for this notion a form of the
reverse bound can be proved: A bound on (approximate) max-information of an algorithm A implies a bound
on the randomized description length of the output of a different algorithm with statistically indistinguishable
from A output.
12
3.1.2
Differential Privacy
We now show that pure differential privacy implies a bound on max information. We start with a simple
bound on max-information of differentially private algorithms that applies to all distributions over datasets.
In particular, it implies that the differential privacy-based approach can be used beyond the i.i.d. setting in
[DFH+ 14].
Theorem 19. Let A be an -differentially private algorithm. Then I∞ (A, n) ≤ log e · n.
Proof. Clearly, any two datasets S and S 0 differ in at most n elements. Therefore, for every y we have
P[Y = y | S = S] ≤ en P[Y = y | S = S 0 ] (this is a direct implication of Definition 2 referred to as group
privacy [DR14]), or equivalently, D∞ (A(S)kA(S 0 )) ≤ log e · n. By Lemma 3 we obtain the claim.
Finally, we prove a stronger bound on approximate max-information for datasets consisting of i.i.d. samples
using the technique from [DFH+ 14]. This bound, together with Thm. 13, generalizes Thm. 6.
Theorem 20. Let A be an ε-differentially private algorithm with range Y. For a distribution P over X , let
S be a random variable drawn from P n . Let Y = A(S)
p denote the random variable output by A on input S.
β
Then for any β > 0, I∞
(S; A(S)) ≤ log e(ε2 n/2 + ε n ln(2/β)/2).
Proof. Fix y ∈ Y. We first observe that by Jensen’s inequality,
E n [ln(P[Y = y | S = S])] ≤ ln
E n [P[Y = y | S = S]] = ln(P[Y = y]).
S∼P
S∼P
Further, by definition of differential privacy, for two databases S, S 0 that differ in a single element,
P[Y = y | S = S] ≤ eε · P[Y = y | S = S 0 ].
=y | S=S]
Now consider the function g(S) = ln P[Y P[Y
. By the properties above we have that E[g(S)] ≤
=y]
ln(P[Y = y]) − ln(P[Y = y]) = 0 and |g(S) − g(S 0 )| ≤ ε. This, by McDiarmid’s inequality (Lemma 1), implies
that for any t > 0,
2
2
P[g(S) ≥ t] ≤ e−2t /(nε ) .
(1)
p
. 2
i
For an integer i ≥ 1, let ti = ε n/2 + ε n ln(2 /β)/2 and let
.
Bi = {S | ti < g(S) ≤ ti+1 } .
Let
[
.
By = {S |g(S) > t1 } =
Bi .
i≥1
By inequality (1), we have that for i ≥ 1,
2
p
√
P[g(S) > ti ] ≤ exp −2 ε n/2 + ln(2i /β)/2
.
By Bayes’ rule, for every S ∈ Bi ,
P[S = S | Y = y]
P[Y = y | S = S]
=
= exp(g(S)) ≤ exp(ti+1 ).
P[S = S]
P[Y = y]
13
Therefore,
P[S ∈ Bi | Y = y] =
X
P[S = S | Y = y]
S∈Bi
≤ exp(ti+i ) ·
X
P[S = S]
S∈Bi
≤ exp(ti+i ) · P[g(S) ≥ ti ]
√
2
p
p
2
i+1
i
= exp ε n/2 + ε n ln(2 /β)/2 − 2 ε n/2 + ln(2 /β)/2
p
p
p
≤ exp ε n/2
ln(2i+1 /β) − 2 ln(2i /β) − ln(2i /β)
< exp(− ln(2i /β)) = β/2i .
An immediate implication of this is that
P[S ∈ By | Y = y] =
X
P[S ∈ Bi | Y = y] ≤
i
X
β/2i ≤ β.
i≥1
Let B = {(S, y) | y ∈ Y, S ∈ By }. Then
P[(S, Y ) ∈ B] = P[(S, Y ) ∈ BY ] ≤ β.
(2)
For every (S, y) 6∈ B we have that
P[S = S, Y = y] = P[S = S | Y = y] · P[Y = y] ≤ exp(t1 ) · P[S = S] · P[Y = y],
and hence by eq.(2) we get that
P
(S,y)∼p(S,Y )
P[S = S, Y = y]
≥ exp(t1 ) ≤ β.
P[S = S] · P[Y = y]
This, by Lemma 18, gives that
p
β
I∞
(S; Y ) ≤ log(exp(t1 )) = log e(ε2 n/2 + ε n ln(2/β)/2).
Applications: We give two simple examples of using the bounds on max-information obtained from
differential privacy to preserve bounds on generalization error that follow from concentration of measure
inequalities. Strong concentration of measure results are at the core of most generalization guarantees in
machine learning. Let A be an algorithm that outputs a function f : X n → R of sensitivity c and define the
“bad event” Oτ is when the empirical estimate of f is more than τ away from the expectation of f (S) for S
distributed according to some distribution D over X n . Namely,
Oτ = {(S, f ) : f (S) − D[f ] ≥ τ } ,
(3)
where D[f ] denotes ES∼D [f (S)].
By McDiarmid’s inequality (Lem. 1) we know that, if S is distributed according to P n then supf :X n →R P[(S, f ) ∈
Oτ ] ≤ exp(−2τ 2 /(c2 n)). The simpler bound in Thm. 19 implies following corollary.
Corollary 21. Let A be an algorithm that outputs a c-sensitive function f : X n → R. Let S be a
random dataset chosen according to distribution P n over X n and let f = A(S). If for β ≥ 0 and τ > 0,
β
I∞
(S; f ) ≤ log e · τ 2 /c2 , then P[f (S) − P n [f ] ≥ τ ] ≤ exp (−τ 2 /(c2 n)) + β. In particular, if A is τ 2 /(c2 n2 )differentially private then P[f (S) − P n [f ] ≥ τ ] ≤ exp (−τ 2 /(c2 n)).
14
Note that for f (S) = ES [φ], where φ : X → [0, 1] this result requires ε = τ 2 . The stronger bound allows to
preserve concentration of measure even when ε = τ /(cn) which corresponds to τ = ε when f (S) = ES [φ].
Corollary 22. Let A be an algorithm that outputs a c-sensitive function f : X n → R. Let S be a random
dataset chosen according to distribution P n over X n and let f = A(S). If A is τ /(cn)-differentially private
then P[f (S) − P n [f ] ≥ τ ] ≤ exp (−3τ 2 /(4c2 n)).
Proof. We apply Theorem 20 with β = 2 exp (−τ 2 /(c2 n)) to obtain that
p
√
β
I∞
(S; f ) ≤ log e · (ε2 n/2 + ε n ln(2/β)/2)) ≤ log e · (τ 2 /(c2 n)/2 + τ 2 /(c2 n)/ 2).
Applying Thm. 13 to McDiarmid’s inequality we obtain that
√
P[f (S) − P n [f ] ≥ τ ] ≤ exp ((1/2 + 1/ 2)τ 2 /(c2 n)) · exp (−2τ 2 /(c2 n)) + 2 exp (−τ 2 /(c2 n))
≤ exp (−3τ 2 /(4c2 n)),
where the last inequality holds when τ 2 /(c2 n) is larger than a fixed constant.
4
Reusable Holdout
We describe two simple algorithms that enable validation of analyst’s queries in the adaptive setting.
4.1
Thresholdout
Our first algorithm Thresholdout follows the approach in [DFH+ 14] where differentially private algorithms are
used to answer adaptively chosen statistical queries. This approach can also be applied to any low-sensitivity
functions 2 of the dataset but for simplicity we present the results for statistical queries. Here we address an
easier problem in which the analyst’s queries only need to be answered when they overfit. Also, unlike in
[DFH+ 14], the analyst has full access to the training set and the holdout algorithm only prevents overfitting
to holdout dataset. As a result, unlike in the general query answering setting, our algorithm can efficiently
validate an exponential in n number of queries as long as a relatively small number of them overfit.
Thresholdout is given access to the training dataset St and holdout dataset Sh and a budget limit B. It
allows any query of the form φ : X → [0, 1] and its goal is to provide an estimate of P[φ]. To achieve this
the algorithm gives an estimate of ESh [φ] in a way that prevents overfitting of functions generated by the
analyst to the holdout set. In other words, responses of Thresholdout are designed to ensure that, with high
probability, ESh [φ] is close to P[φ] and hence an estimate of ESh [φ] gives an estimate of the true expectation
P[φ]. Given a function φ, Thresholdout first checks if the difference between the average value of φ on the
training set St (or ESt [φ]) and the average value of φ on the holdout set Sh (or ESh [φ]) is below a certain
threshold T + η. Here, T is a fixed number such as 0.01 and η is a Laplace noise variable whose standard
deviation needs to be chosen depending on the desired guarantees (The Laplace distribution is a symmetric
exponential distribution.) If the difference is below the threshold, then the algorithm returns ESt [φ]. If the
difference is above the threshold, then the algorithm returns ESh [φ] + ξ for another Laplacian noise variable
ξ. Each time the difference is above threshold the “overfitting” budget B is reduced by one. Once it is
exhausted, Thresholdout stops answering queries. In Fig. 1 we provide the pseudocode of Thresholdout.
We now establish the formal generalization guarantees that Thresholdout enjoys. As the first step we state
what privacy parameters are achieved by Thresholdout.
p
Lemma 23. Thresholdout satisfies (2B/(σn), 0)-differential privacy. Thresholdout also satisfies ( 32B ln(2/δ)/(σn), δ)differential privacy for any δ > 0.
2 Guarantees based on pure differential privacy follow from the same analysis. Proving generalization guarantees for
low-sensitivity queries based on approximate differential privacy requires a modification of Thresholdout using techniques in
[BSSU15].
15
Algorithm Thresholdout Input: Training set St , holdout set Sh , threshold T, noise rate σ, budget B
1. sample γ ∼ Lap(2 · σ); T̂ ← T + γ
2. For each query φ do
(a) if B < 1 output “⊥”
(b) else
i. sample η ∼ Lap(4 · σ)
ii. if |ESh [φ] − ESt [φ]| > T̂ + η
A. sample ξ ∼ Lap(σ), γ ∼ Lap(2 · σ)
B. B ← B − 1 and T̂ ← T + γ
C. output ESh [φ] + ξ
iii. else output ESt [φ].
Figure 1: The details of Thresholdout algorithm
Proof. Thresholdout is an instantiation of a basic tool from differential privacy, the “Sparse Vector Algorithm”
([DR14, Algorithm 2]), together with the Laplace mechanism ([DR14, Defn. 3.3]). The sparse vector algorithm
takes as input a sequence of c sensitivity 1/n queries3 (here c = B, the budget), and for each query, attempts
to determine whether the value of the query, evaluated on the private dataset, is above a fixed threshold T or
below it. In our instantiation, the holdout set Sh is the private data set, and each function φ corresponds to
the following query evaluated on Sh : fφ (Sh ) := |ESh [φ] − ESt [φ]|. (Note that the training set St is viewed
as part of the definition of the query). Thresholdout then is equivalent to the following procedure: we run
the sparse vector algorithm [DR14, Algorithm 2] with c = B, queries fφ for each function φ, and noise
rate 2σ. Whenever an above-threshold query is reported by the sparse vector algorithm, we release its
value using the Laplace mechanism [DR14, Defn. 3.3] with noise rate σ (this is what occurs every time
Thresholdout answers by outputting ESh [φ] + ξ). By the privacy guarantee of the sparse vector algorithm
([DR14, Thm. 3.25]), the√sparse vector portion of Thresholdout satisfies (B/(σn), 0)-differential privacy, and
simultaneously satisfies (
8B ln(2/δ)
, δ/2)-differential
σn
privacy. The Laplace mechanism portion of Thresholdout
√
8B ln(2/δ)
, δ/2)satisfies (B/(σn), 0)-differential privacy by [DR14, Thm. 3.6], and simultaneously satisfies (
σn
differential privacy by [DR14, Thm. 3.6] and [DR14, Cor. 3.21]. Finally, the composition of two mechanisms,
the first of which is (1 , δ1 )-differentially private, and the second of which is (2 , δ2 )-differentially private is
itself (1 + 2 , δ1 + δ2 )-differentially private (Thm. 3). Adding the privacy parameters of the Sparse Vector
portion of Thresholdout and the Laplace mechanism portion of Thresholdout yield the parameters of our
theorem.
We note that tighter privacy parameters are possible (e.g. by invoking the parameters and guarantees of
the algorithm “NumericSparse” ([DR14, Algorithm 3]), which already combines the Laplace addition step) –
we chose simpler parameters for clarity.
Note the seeming discrepancy between the guarantee provided by Thresholdout and generalization
guarantees in Theorem 8 and Corollary 7: while Theorem 8 promises generalization bounds for functions that
are generated by a differentially private algorithm, here we allow an arbitrary data analyst to generate query
functions in any way she chooses, with access to the training set and differentially private estimates of the
means of her functions on the holdout set. The connection comes from preservation of differential privacy
guarantee under post-processing (Lem. 5).
We can now quantify the number of samples necessary to achieve generalization error τ with probability
3 In fact, the theorems for the Sparse Vector algorithm in Dwork and Roth are stated for sensitivity 1 queries – we use them
for sensitivity 1/n queries of the form ESh [φ], which results in all of the bounds being scaled down by a factor of n.
16
at least 1 − β.
Lemma 24. Let τ, β, T, B > 0. Let S denote the holdout dataset of size n drawn i.i.d. from a distribution P.
Consider an algorithm that is given access to St and adaptively chooses functions φ1 , . . . , φm : X → [0, 1]
while interacting with Thresholdout which is given datasets S, St and parameters σ, B, T . If
.
n ≥ n0 (B, σ, τ, β) = max{2B/(στ ), ln(6/β)/τ 2 }
or
. 80 ·
n ≥ n1 (B, σ, τ, β) =
p
B ln(1/(τ β))
τσ
then for every i ∈ [m], P [|P[φi ] − ES [φi ]| ≥ τ ] ≤ β.
Proof. Consider the first guarantee of Lemma 23. In order to achieve generalization error τ via Corollary 7
2
(i.e. in order to guarantee that for every function φ we have: P [|P[φi ] − ES [φi ]| ≥ τ ] ≤ 6e−τ n ) we need
to have n large enough to achieve (ε, 0)-differential privacy for ε = τ . To achieve this it suffices to have
2
n ≥ 2B/(στ ). By ensuring that n ≥ ln(6/β)/τ 2 we also have that 6e−τ n ≤ β.
We can also make use of the second guarantee in Lemma 23 together with the results of Nissim and
Stemmer [NS15] (Thm. 8). In order to achieve generalization error τ with probability 1 − β (i.e. in order
to guarantee
for every function φ we have: P [|P[φi ] − ES [φi ]| ≥ τ ] ≤ β), we can apply Thm. 8 by setting
p
= 32B ln(2/δ)/(σn) = τ /13 and δ = 26 lnβτ
(26/τ ) . We can obtain these privacy parameters from Lemma 23
√
80· B ln(1/(τ β))
by choosing any n ≥
(for sufficiently small β and τ ). We remark that a somewhat worse bound
τσ
√
2048 ln(8/β)
of n1 (B, σ, τ, β) =
follows by setting = τ /4 and δ = (β/8)4/τ in [DFH+ 14, Thm. 10].
τ 3/2 σ
Both settings lead to small generalization error and so we can pick whichever gives the larger bound.
The first bound has grows linearly with B but is simpler can be easily extended to other distributions over
datasets and to low-sensitivity functions. The second bound has quadratically better dependence on B at
the expense of a slightly worse dependence on τ . We can now apply our main results to get a generalization
bound for the entire execution of Thresholdout.
Theorem 25. Let β, τ > 0 and m ≥ B > 0. We set T = 3τ /4 and σ = τ /(96 ln(4m/β)). Let S denote a
holdout dataset of size n drawn i.i.d. from a distribution P and St be any additional dataset over X . Consider
an algorithm that is given access to St and adaptively chooses functions φ1 , . . . , φm while interacting with
Thresholdout which is given datasets S, St and values σ, B, T . For every i ∈ [m], let ai denote the answer of
Thresholdout on function φi : X → [0, 1]. Further, for every i ∈ [m], we define the counter of overfitting
.
Zi = |{j ≤ i : |P[φj ] − ESt [φj ]| > τ /2}| .
Then
P [∃i ∈ [m], Zi < B & |ai − P[φi ]| ≥ τ ] ≤ β
p
whenever n ≥ min{n0 (B, σ, τ /8, β/(2m)), n1 (B, σ, τ /8, β/(2m))} = O ln(m/β)
·min{B, B ln(ln(m/β)/τ )}.
τ2
Proof. There are two types of error we need to control: the deviation between ai and the average value of
φi on the holdout set ES [φi ], and the deviation between the average value of φi on the holdout set and the
expectation of φi on the underlying distribution, P[φi ]. Specifically, we decompose the error as
P[ai 6= ⊥ & |ai − P[φi ]| ≥ τ ] ≤ P [ai 6= ⊥ & |ai − ES [φi ]| ≥ 7τ /8] + P [|P[φi ] − ES [φi ]| ≥ τ /8] .
(4)
To control the first term we need to bound the values of noise variables used by Thresholdout. For the second
term we will use the generalization properties of Thresholdout given in Lemma 24.
We now deal with the errors introduced by the noise variables. For i ∈ [m], let ηi , ξi and γi denote the
random variables η, ξ and γ, respectively, at step i of the execution of Thresholdout. We first note that each of
17
these variables is chosen from Laplace distribution at most m times. By properties of the Laplace distribution
with parameter 4σ, we know that for every t > 0, P[|ηi | ≥ t · 4σ] = e−t/2 . Therefore for t = 2 ln(4m/β) we
obtain
β
P[|ηi | ≥ 2 ln(4m/β) · 4σ] ≤ e−t/2 =
.
4m
By the definition of σ, 8 ln(4m/β) · σ = τ /12. Applying the union bound we obtain that
P[∃i, |ηi | ≥ τ /12] ≤ β/4,
where by ∃i we refer to ∃i ∈ [m] for brevity. Similarly, ξi and γi are obtained by sampling from the Laplace
distribution and each is re-randomized at most B times. Therefore
P[∃i, |γi | ≥ τ /24] ≤ B · β/4m ≤ β/4
and
P[∃i, |ξi | ≥ τ /48] ≤ B · β/4m ≤ β/4.
For answers that are different from ⊥, we can now bound the first term of Equation (4) by considering two
cases, depending on whether Thresholdout answers query φi by returning ai = ES [φi ] + ξi or by returning
ai = ESt [φi ]. First, consider queries whose answers are returned using the former condition. Under this
condition, |ai − ES [φi ]| = |ξi |. Next, we consider the second case, those queries whose answers are returned
using ESt [φi ]. By definition of the algorithm, we have
|ai − ES [φi ]| = |ESt [φi ] − ES [φi ]| ≤ T + γi + ηi ≤ 3τ /4 + |γi | + |ηi |.
Combining these two cases implies that
P [∃i, ai 6= ⊥ & |ai − ES [φi ]| ≥ 7τ /8] ≤ max{P [∃i, |ξi | ≥ 7τ /8] , P [∃i, |γi | + |ηi | ≥ τ /8]}.
Noting that τ /24 + τ /12 = τ /8 and applying our bound on variables ηi , ξi and γi we get
P [∃i, ai 6= ⊥ & |ai − ES [φi ]| ≥ 7τ /8] ≤ β/2.
(5)
By Lemma 24, for n ≥ min{n0 (B, σ, τ /8, β/2m), n1 (B, σ, τ /8, β/2m)},
P [|P[φi ] − ES [φi ]| ≥ τ /8] ≤ β/2m.
Applying the union bound we obtain
P [∃i, |P[φi ] − ES [φi ]| ≥ τ /8] ≤ β/2.
Combining this with Equation (5) and using in Equation (4) we get that
P[∃i, ai 6= ⊥ & |ai − P[φi ]| ≥ τ ] ≤ β.
To finish the proof we show that under the conditions on the noise variables and generalization error
used above, we have that if Zi < B then ai 6= ⊥. To see this, observe that for every j ≤ i that reduces
Thresholdout’s budget, we have
|P[φj ] − ESt [φj ]| ≥ |ES [φj ] − ESt [φj ]| − |P[φj ] − ES [φj ]|
≥ |T + γj + ηj | − |P[φj ] − ES [φj ]|
≥ T − |γj | − |ηj | − |P[φj ] − ES [φj ]|.
This means that for every j ≤ i that reduces the budget we have |P[φj ]−ESt [φj ]| ≥ 3τ /4−τ /24−τ /12−τ /8 =
τ /2 and hence (when the conditions on the noise variables and generalization error are satisfied) for every i,
if Zi < B then Thresholdout’s budget is not yet exhausted and ai 6= ⊥. We can therefore conclude that
P[∃i, Zi < B & |ai − P[φi ]| ≥ τ ] ≤ β.
18
Note that in the final bound on n, the term O ln(m/β)
is equal (up to a constant factor) to the number
τ2
of samples that are necessary to answer m non-adaptively chosen queries with tolerance
√ τ and confidence 1 − β.
In particular, as in the non-adaptive setting, achievable tolerance τ scales as 1/ n (up to the logarithmic
factor). Further, this bound allows m to be exponentially large in n as long as B grows sub-quadratically in
n (that is, B ≤ n2−c for a constant c > 0).
Remark 26. In Thm. 25 Sh is used solely to provide a candidate estimate of the expectation of each query
function. The theorem holds for any other way to provide such estimates. In addition, the one-sided version
of the algorithm can be used when catching only the one-sided error is necessary. For example, in many cases
overfitting is problematic only if the training error estimate is larger than the true error. This is achieved by
using the condition ESh [φ] − ESt [φ] > T̂ + η to detect overfitting. In this case only one-sided errors will be
caught by Thresholdout and only one-sided overfitting will decrease the budget.
4.2
SparseValidate
We now present a general algorithm for validation on the holdout set that can validate many arbitrary queries
as long as few of them fail the validation. The algorithm which we refer to as SparseValidate only reveals
information about the holdout set when validation fails and therefore we use bounds based on description
length to analyze its generalization guarantees.
More formally, our algorithm allows the analyst to pick any Boolean function of a dataset ψ (or even any
algorithm that outputs a single bit) and provides back the value of ψ on the holdout set ψ(Sh ). SparseValidate
has a budget m for the total number of queries that can be asked and budget B for the number of queries
that returned 1. Once either of the budgets is exhausted, no additional answers are given. We now give a
general description of the guarantees of SparseValidate.
Theorem 27. Let S denote a randomly chosen holdout set of size n. Let A be an algorithm that is given
access to SparseValidate(m, B) and outputs queries ψ1 , . . . , ψm such that each ψi is in some set Ψi of functions
from X n to {0, 1}. Assume that for every i ∈ [m] and ψi ∈ Ψi , P[ψi (S) = 1] ≤ βi . Let ψi be the random
Pmin{i−1,B} i
B
variable equal to the i’th query of A on S. Then P[ψi (S) = 1] ≤ `i · βi , where `i = j=0
j ≤m .
Proof. Let B denote the algorithm that represents view the interaction of A with SparseValidate(m, B) up
until query i and outputs the all the i − 1 responses of SparseValidate(m, B) in this interaction. If there are B
responses with value 1 in the interaction then all the responses after the last one are meaningless and can be
assumed to be equal to 0. The number of binary strings of length i − 1 that contain at most B ones is exactly
Pmin{i−1,B} i
`i = j=0
j . Therefore we can assume that the output domain of B has size `i and we denote it
by Y. Now, for y ∈ Y let R(y) be the set of datasets S such that ψi (S) = 1, where ψi is the function that
A generates when the responses of SparseValidate(m, B) are y and the input holdout dataset is S (for now
assume that A is deterministic). By the conditions of the theorem we have that for every y, P[S ∈ R(y)] ≤ βi .
Applying Thm. 9 to B, we get that P[S ∈ R(B(S))] ≤ `i βi , which is exactly the claim. We note that to
address the case when A is randomized (including dependent on the random choice of the training set) we
can use the argument above for every fixing of all the random bits of A. From there we obtain that the claim
holds when the probability is taken also over the randomness of A.
We remark that the proof can also be obtained via a more direct application of the union bound over all
strings in Y. But the proof via Thm. 9 demonstrates the conceptual role that short description length plays
in this application.
In this general formulation it is the analyst’s responsibility to use the budgets economically and pick
query functions that do not fail validation often. At the same time, SparseValidate ensures that (for the
appropriate values of the parameters) the analyst can think of the holdout set as a fresh sample for the
purposes of validation. Hence the analyst can pick queries in such a way that failing the validation reliably
indicates overfitting. To relate this algorithm to Thresholdout, consider the validation query function that
is the indicator of the condition |ESh [φ] − ESt [φ]| > T̂ + η (note that this condition can be evaluated using
19
an algorithm with access to Sh ). This is precisely the condition that consumes the overfitting budget of
2
Thresholdout. Now, as in Thresholdout, for every fixed φ, P[|ESh [φ] − P[φ]| ≥ τ ] ≤ 2e−2τ n . If B ≤ τ 2 n/ ln m,
then we obtain that for every query φ generated by the analyst, we still have strong concentration of the mean
2
on the holdout set around the expectation: P[|ESh [φ] − P[φ]| ≥ τ ] ≤ 2e−τ n . This implies that if the condition
|ESh [φ] − ESt [φ]| > T̂ + η holds, then with high probability also the condition |P[φ] − ESt [φ]| > T̂ + η − τ holds,
indicating overfitting. One notable distinction of Thresholdout from SparseValidate is that SparseValidate does
not provide corrections in the case of overfitting. One way to remedy that is simply to use a version of
SparseValidate that allows functions with values in {0, 1, . . . , L}. It is easy to see that for such functions
we would obtain the bound of the form `i · Lmin{B,i−1} · βi . To output a value in [0, 1] with precision τ ,
L = b1/τ c would suffice. However, in many cases a more economical solution would be to have a separate
dataset which is used just for obtaining the correct estimates.
An example of the application of SparseValidate for answering statistical and low-sensitivity queries that is
based on our analysis can be found in [BSSU15]. The analysis of generalization on the holdout set in [BH15]
and the analysis of the Median Mechanism we give in Section B also rely on this sparsity-based technique.
An alternative view of this algorithm is as a general template for designing algorithms for answering some
specific type of adaptively chosen queries. Generalization guarantees specific to the type of query can then be
obtained from our general analysis. For example, an algorithm that fits a mixture of Gaussians model to the
data could define the validation query to be an algorithm that fits the mixture model to the holdout and
obtains a vector of parameters Θh . The validation query then compares it with the vector of parameters Θt
obtained on the training set and outputs 1 if the parameter vectors are “not close” (indicating overfitting).
Given guarantees of statistical validity of the parameter estimation method in the static setting one could
then derive guarantees for adaptive validation via Thm. 27.
5
Experiments
We describe a simple experiment on synthetic data that illustrates the danger of reusing a standard holdout
set and how this issue can be resolved by our reusable holdout. In our experiment the analyst wants to
build a classifier via the following common strategy. First the analyst finds a set of single attributes that
are correlated with the class label. Then the analyst aggregates the correlated variables into a single model
of higher accuracy (for example using boosting or bagging methods). More formally, the analyst is given
a d-dimensional labeled data set S of size 2n and splits it randomly into a training set St and a holdout
set Sh of equal size. We denote an element of S by a tuple (x, y) where x is a d-dimensional vector and
y ∈ {−1, 1} is the corresponding class label. The analyst wishes to select variables to be included in her
classifier. For various values of the number of variables to select k, she picks k variables with the largest
absolute correlations with the label. However, she verifies the correlations (with the label) on the holdout set
and uses only those variables whose correlation agrees in sign with the correlation on the training set and
both correlations are larger than some threshold in absolute value. She then creates a simple linear threshold
classifier on the selected variables using only the signs of the correlations of the selected variables. A final
test evaluates the classification accuracy of the classifier on both the training set and the holdout set.
Formally, the algorithm is used to build a linear threshold classifier:
1. For each
the correlation with the label on the training and holdout sets:
P attribute i ∈ [d] compute
P
wit = (x,y)∈St xi y and wih = (x,y)∈Sh xi y. Let
√
√
W = i | wit · wih > 0; |wit | ≥ 1/ n; |wih | ≥ 1/ n|
√
that is the set of variables for which wit and wih have the same sign and both are at least 1/ n in
absolute value (this is the standard deviation of the correlation in our setting). Let Vk be the subset of
variables in W with k largest values of |wit |.
P
t
2. Construct the classifier f (x) = sgn
i∈Vk sgn(wi ) · xi .
20
In the experiments we used an implementation of Thresholdout that differs somewhat from the algorithm
we analyzed theoretically (given in Figure 1). Specifically, we set the parameters to be T = 0.04 and τ = 0.01.
This is lower than the values necessary for the proof (and which are not intended for direct application)
but suffices to prevent overfitting in our experiment. Second, we use Gaussian noise instead of Laplacian
noise as it has stronger concentration properties (in many differential privacy applications similar theoretical
guarantees hold mechanisms based on Gaussian noise).
No correlation between labels and data: In our first experiment, each attribute is drawn independently
from the normal distribution N (0, 1) and we choose the class label y ∈ {−1, 1} uniformly at random so that
there is no correlation between the data point and its label. We chose n = 10, 000, d = 10, 000 and varied
the number of selected variables k. In this scenario no classifier can achieve true accuracy better than 50%.
Nevertheless, reusing a standard holdout results in reported accuracy of over 63% for k = 500 on both the
training set and the holdout set (the standard deviation of the error is less than 0.5%). The average and
standard deviation of results obtained from 100 independent executions of the experiment are plotted in
Figure 2 which also includes the accuracy of the classifier on another fresh data set of size n drawn from
the same distribution. We then executed the same algorithm with our reusable holdout. The algorithm
Thresholdout was invoked with T = 0.04 and τ = 0.01 explaining why the accuracy of the classifier reported
by Thresholdout is off by up to 0.04 whenever the accuracy on the holdout set is within 0.04 of the accuracy
on the training set. Thresholdout prevents the algorithm from overfitting to the holdout set and gives a valid
estimate of classifier accuracy.
accuracy
0.65
Standard holdout
0.70
training
holdout
fresh
0.65
accuracy
0.70
0.60
0.55
0.50
0
100
200
300
number of variables
400
500
Thresholdout
training
holdout
fresh
0.60
0.55
0.50
0
100
200
300
number of variables
400
500
Figure 2: No correlation between class labels and data points. The plot shows the classification accuracy of the
classifier on training, holdout and fresh sets. Margins indicate the standard deviation.
High correlation between labels and some of the variables: In our second experiment, the class
labels are correlated with some of the variables. As before the label is randomly chosen from {−1, 1} and
each of the attributes is drawn from N (0, 1) aside from 20 attributes which are drawn from N (y · 0.06, 1)
where y is the class label. We execute the same algorithm on this data with both the standard holdout and
Thresholdout and plot the results in Figure 3. Our experiment shows that when using the reusable holdout,
the algorithm still finds a good classifier while preventing overfitting. This illustrates that the reusable
holdout simultaneously prevents overfitting and allows for the discovery of true statistical patterns.
In Figures 2 and 3, simulations that used Thresholdout for selecting the variables also show the accuracy
on the holdout set as reported by Thresholdout. For comparison purposes, in Figure 4 we plot the actual
accuracy of the generated classifier on the holdout set (the parameters of the simulation are identical to those
used in Figures 2 and 3). It demonstrates that there is essentially no overfitting to the holdout set. Note
that the advantage of the accuracy reported by Thresholdout is that it can be used to make further data
dependent decisions while mitigating the risk of overfitting.
Discussion of the results: Overfitting to the standard holdout set arises in our experiment because the
analyst reuses the holdout after using it to measure the correlation of single attributes. We first note that
neither cross-validation nor bootstrap resolve this issue. If we used either of these methods to validate the
21
accuracy
0.65
Standard holdout
0.70
training
holdout
fresh
0.65
accuracy
0.70
0.60
0.55
0.50
0
100
200
300
number of variables
400
0.60
0.55
0.50
0
500
Thresholdout
training
holdout
fresh
100
200
300
number of variables
400
500
400
500
Figure 3: Some variables are correlated with the label.
accuracy
0.65
Uncorrelated
0.70
training
holdout
fresh
0.65
accuracy
0.70
0.60
0.55
0.50
0
100
200
300
number of variables
400
500
Correlated
training
holdout
fresh
0.60
0.55
0.50
0
100
200
300
number of variables
Figure 4: Accuracy of the classifier produced with Thresholdout on the holdout set.
correlations, overfitting would still arise due to using the same data for training and validation (of the final
classifier). It is tempting to recommend other solutions to the specific problem we used in our experiment.
Indeed, a significant number of methods in the statistics and machine learning literature deal with inference
for fixed two-step procedures where the first step is variable selection (see [HTF09] for examples). Our
experiment demonstrates that even in such simple and standard settings our method avoids overfitting
without the need to use a specialized procedure – and, of course, extends more broadly. More importantly,
the reusable holdout gives the analyst a general and principled method to perform multiple validation steps
where previously the only known safe approach was to collect a fresh holdout set each time a function depends
on the outcomes of previous validations.
6
Conclusions
In this work, we give a unifying view of two techniques (differential privacy and description length bounds)
which preserve the generalization guarantees of subsequent algorithms in adaptively chosen sequences of data
analyses. Although these two techniques both imply low max-information – and hence can be composed
together while preserving their guarantees – the kinds of guarantees that can be achieved by either alone are
incomparable. This suggests that the problem of generalization guarantees under adaptivity is ripe for future
study on two fronts. First, the existing theory is likely already strong enough to develop practical algorithms
with rigorous generalization guarantees, of which Thresholdout is an example. However additional empirical
work is needed to better understand when and how the theory should be applied in specific application
22
scenarios. At the same time, new theory is also needed. As an example of a basic question we still do
not know the answer to: even in the simple setting of adaptively reusing a holdout set for computing the
expectations of boolean-valued predicates, is it possible to obtain stronger generalization guarantees (via any
means) than those that are known to be achievable via differential privacy?
References
[BDMN05] Avrim Blum, Cynthia Dwork, Frank McSherry, and Kobbi Nissim. Practical privacy: the SuLQ
framework. In PODS, pages 128–138, 2005.
[BE02]
Olivier Bousquet and André Elisseeff. Stability and generalization. JMLR, 2:499–526, 2002.
[BH15]
Avrim Blum and Moritz Hardt. The ladder: A reliable leaderboard for machine learning
competitions. CoRR, abs/1502.04585, 2015.
[BNS13]
Amos Beimel, Kobbi Nissim, and Uri Stemmer. Characterizing the sample complexity of private
learners. In ITCS, pages 97–110, 2013.
[BSSU15]
Raef Bassily, Adam Smith, Thomas Steinke, and Jonathan Ullman. More general queries and
less generalization error in adaptive data analysis. CoRR, abs/1503.04843, 2015.
[CT10]
Gavin C. Cawley and Nicola L. C. Talbot. On over-fitting in model selection and subsequent
selection bias in performance evaluation. Journal of Machine Learning Research, 11:2079–2107,
2010.
[DFH+ 14] Cynthia Dwork, Vitaly Feldman, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Aaron Roth.
Preserving statistical validity in adaptive data analysis. CoRR, abs/1411.2664, 2014. Extended
abstract in STOC 2015.
[DFN07]
Chuong B. Do, Chuan-Sheng Foo, and Andrew Y. Ng. Efficient multiple hyperparameter learning
for log-linear models. In NIPS, pages 377–384, 2007.
[DKM+ 06] Cynthia Dwork, Krishnaram Kenthapadi, Frank McSherry, Ilya Mironov, and Moni Naor. Our
data, ourselves: Privacy via distributed noise generation. In EUROCRYPT, pages 486–503, 2006.
[DL09]
C. Dwork and J. Lei. Differential privacy and robust statistics. In Proceedings of the 2009
International ACM Symposium on Theory of Computing (STOC), 2009.
[DMNS06] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity
in private data analysis. In Theory of Cryptography, pages 265–284. Springer, 2006.
[DN03]
Irit Dinur and Kobbi Nissim. Revealing information while preserving privacy. In PODS, pages
202–210. ACM, 2003.
[DN04]
Cynthia Dwork and Kobbi Nissim. Privacy-preserving datamining on vertically partitioned
databases. In CRYPTO, pages 528–544, 2004.
[DR14]
Cynthia Dwork and Aaron Roth. The algorithmic foundations of differential privacy. Foundations
and Trends in Theoretical Computer Science, 9(34):211–407, 2014.
[Dwo11]
Cynthia Dwork. A firm foundation for private data analysis. CACM, 54(1):86–95, 2011.
[Fre83]
David A. Freedman. A note on screening regression equations. The American Statistician,
37(2):152–155, 1983.
[Fre98]
Yoav Freund. Self bounding learning algorithms. In COLT, pages 247–258, 1998.
23
[HTF09]
Trevor Hastie, Robert Tibshirani, and Jerome H. Friedman. The Elements of Statistical Learning:
Data Mining, Inference, and Prediction. Springer series in statistics. Springer, 2009.
[HU14]
Moritz Hardt and Jonathan Ullman. Preventing false discovery in interactive data analysis is
hard. In FOCS, pages 454–463, 2014.
[Lan05]
John Langford. Clever methods of overfitting. http://hunch.net/?p=22, 2005.
[LB03]
John Langford and Avrim Blum. Microchoice bounds and self bounding learning algorithms.
Machine Learning, 51(2):165–179, 2003.
[MMP+ 10] Andrew McGregor, Ilya Mironov, Toniann Pitassi, Omer Reingold, Kunal Talwar, and Salil P.
Vadhan. The limits of two-party differential privacy. In FOCS, pages 81–90. IEEE Computer
Society, 2010.
[MNPR06] Sayan Mukherjee, Partha Niyogi, Tomaso Poggio, and Ryan Rifkin. Learning theory: stability
is sufficient for generalization and necessary and sufficient for consistency of empirical risk
minimization. Advances in Computational Mathematics, 25(1-3):161–193, 2006.
[Ng97]
Andrew Y. Ng. Preventing ”overfitting” of cross-validation data. In ICML, pages 245–253, 1997.
[NS15]
Kobbi Nissim and Uri Stemmer. On the generalization properties of differential privacy. CoRR,
abs/1504.05800, 2015.
[PRMN04] Tomaso Poggio, Ryan Rifkin, Sayan Mukherjee, and Partha Niyogi. General conditions for
predictivity in learning theory. Nature, 428(6981):419–422, 2004.
[Reu03]
Juha Reunanen. Overfitting in making comparisons between variable selection methods. Journal
of Machine Learning Research, 3:1371–1382, 2003.
[RF08]
R. Bharat Rao and Glenn Fung. On the dangers of cross-validation. an experimental evaluation.
In International Conference on Data Mining, pages 588–596. SIAM, 2008.
[RR10]
Aaron Roth and Tim Roughgarden. Interactive privacy via the median mechanism. In 42nd
ACM STOC, pages 765–774. ACM, 2010.
[SSBD14]
Shai Shalev-Shwartz and Shai Ben-David. Understanding Machine Learning: From Theory to
Algorithms. Cambridge University Press, 2014.
[SSSSS10] Shai Shalev-Shwartz, Ohad Shamir, Nathan Srebro, and Karthik Sridharan. Learnability, stability
and uniform convergence. The Journal of Machine Learning Research, 11:2635–2670, 2010.
[SU14]
Thomas Steinke and Jonathan Ullman. Interactive fingerprinting codes and the hardness of
preventing false discovery. arXiv preprint arXiv:1410.1228, 2014.
[WLF15]
Yu-Xiang Wang, Jing Lei, and Stephen E. Fienberg. Learning with differential privacy: Stability,
learnability and the sufficiency and necessity of ERM principle. CoRR, abs/1502.06309, 2015.
A
From Max-information to Randomized Description Length
In this section we demonstrate additional connections between max-information, differential privacy and
description length. These connections are based on a generalization of description length to randomized
algorithms that we refer to as randomized description length.
Definition 28. For a universe Y let A be a randomized algorithm with input in X and output in Y. We say
that the output of A has randomized description length k if for every fixed setting of random coin flips of A
the set of possible outputs of A has size at most 2k .
24
We first note that just as the (deterministic) description length, randomized description length implies
generalization and gives a bound on max-information.
Theorem 29. Let A : X n → Y be an algorithm with randomized description length k and let S be a
n
random dataset over X n . Assume that R : Y → 2X is such that for every y ∈ Y, P[S ∈ R(y)] ≤ β. Then
P[S ∈ R(A(S))] ≤ 2k · β.
Theorem 30. Let A be an algorithm with randomized description length k taking as an input an n-element
β
dataset and outputting a value in Y. Then for every β > 0, I∞
(A, n) ≤ log(|Y|/β).
Proof. Let S be any random variable over n-element input datasets and let Y be the corresponding output
β
distribution Y = A(S). It suffices to prove that for every β > 0, I∞
(S; Y ) ≤ k + log(1/β).
Let R be the set of all possible values of the random bits of A and let R denote the uniform distribution
over a choice of r ∈ R. For r ∈ R, let Ar denote A with the random bits set to r and let Yr = Ar (S). Observe
that by the definition of randomized description length, the range of Ar has size at most 2k . Therefore, by
β
Theorem 17, we obtain that I∞
(S; Yr ) ≤ log(2k /β).
n
For any event O ⊆ X × Y we have that
P[(S, Y ) ∈ O] = E [P[(S, Yr ) ∈ O]]
r∼R
k
2
≤ E
· P[S × Yr ∈ O] + β
r∼R β
2k
· E [P[S × Yr ∈ O]] + β
β r∼R
2k
· P[S × Y ∈ O] + β.
=
β
=
β
(S; Y ) ≤ log(2k /β).
By the definition of β-approximate max-information, we obtain that I∞
We next show that if the output of an algorithm A has low approximate max-information about its input
then there exists a (different) algorithm whose output is statistically close to that of A while having short
randomized description. We remark that this reduction requires the knowledge of the marginal distribution
A(S).
Lemma 31. Let A be a randomized algorithm taking as an input a dataset of n points from X and outputting
β
a value in Y. Let Z be a random variable over Y. For k > 0 and a dataset S let βS = min{β | D∞
(A(S)kZ) ≤
0
n
0
k}. There exists an algorithm A that given S ∈ X , β, k and any β > 0,
1. the output of A0 has randomized description length k + log ln(1/β 0 ).
2. for every S, ∆(A0 (S), A(S)) ≤ βS + β 0 .
βS
(A(S)kZ) ≤ k. By the properties of
Proof. Let S denote the input dataset. By definition of βS , D∞
βS
approximate divergence (e.g.[DR14]), D∞ (A(S)kZ) ≤ k implies that there exists a random variable Y such
that ∆(A(S), Y ) ≤ βS and D∞ (Y kZ) ≤ k.
For t = 2k ln(1/β 0 ) the algorithm A0 randomly and independently chooses t samples from Z. Denote
=yi ]
them by y1 , y2 , . . . , yt . For i = 1, 2, . . . , t, A0 outputs yi with probability pi = 2kP[Y
and goes to the next
·P[Z=yi ]
sample otherwise. Note that pi ∈ [0, 1] and therefore this is a legal choice of probability. When all samples
are exhausted the algorithm outputs y1 .
We first note that by the definition of this algorithm its output has randomized description length
log t = k + log ln(1/β 0 ). Let T denote the event that at least one of the samples was accepted. Conditioned
on this event the output of A0 (S) is distributed according to Y . For each i,
X P[Y = yi ]
P[Y = yi ]
1
E [pi ] = E
=
= k.
2k
2
yi ∼p(Z)
yi ∼p(Z) 2k · P[Z = yi ]
yi ∈Y
25
k
This means that the probability that none of t samples will be accepted is (1 − 2−k )t ≤ et/2 ≤ β 0 . Therefore
∆(A0 (S), Y ) ≤ β 0 and, consequently, ∆(A0 (S), A(S)) ≤ βS + β 0 .
We can now use Lemma 31 to show that if for a certain random choice of a dataset S, the output of A has
low approximate max-information then there exists an algorithm A0 whose output on S has low randomized
description length and is statistically close to the output distribution of A.
Theorem 32. Let S be a random dataset in X n and A be an algorithm taking as an input a dataset in
β
X n and having a range Y. Assume that for some β ≥ 0, I∞
(S; A(S)) = k. For any β 0 > 0, there exists an
0
n
algorithm A taking as an input a dataset in X such that
1. the output of A0 has randomized description length k + log ln(1/β 0 );
2. ∆((S, A0 (S)), (S, A(S)) ≤ β + β 0 .
β
Proof. For a dataset S let βS = min{β | D∞
(A(S)kA(S)) ≤ k}. To prove this result it suffices to observe
that E[βS ] ≤ β and then apply Lemma 31 with Z = A(S) To show
S that E[βS ] ≤ β let OS ⊆ Y denote an
event such that P[A(S) ∈ OS ] = 2k · P[A(S) ∈ OS ] + βS . Let O = S∈X n {(S, OS )}. Then,
P[(S, A(S) ∈ O] =
=
E
S∼p(S)
E
S∼p(S)
[P[(S, A(S) ∈ OS ]]
k
2 · P[A(S) ∈ OS ] + βS
= 2k · P[S × A(S) ∈ O] + E[βS ].
0
If E[βS ] > β then it would hold for some k 0 > k that P[(S, A(S) ∈ O] = 2k · P[S × A(S) ∈ O] + β
β
contradicting the assumption I∞
(S; A(S)) = k. We remark that, it is also easy to see that E[βS ] = β.
It is important to note that Theorem 32 is not the converse of Theorem 30 and does not imply equivalence
between max-information and randomized description length. The primary difference is that Theorem 32
defines a new algorithm rather than arguing about the original algorithm. In addition, the new algorithm
requires samples of A(S), that is, it needs to know the marginal distribution on Y. As a more concrete
example, Theorem 32 does not allow us to obtain a description-length-based equivalent of Theorem 20 for all
i.i.d. datasets. On the other hand, any algorithm that has bounded max-information for all distributions over
datasets can be converted to an algorithm with low randomized description length.
Theorem 33. Let A be an algorithm over X n with range Y and let k = I∞ (A, n). For any β > 0, there
exists an algorithm A0 taking as an input a dataset in X n such that
1. the output of A0 has randomized description length k + log ln(1/β);
2. for every dataset S ∈ X n , ∆(A0 (S), A(S)) ≤ β.
Proof. Let S0 = (x, x, . . . , x) be an n-element dataset for an arbitrary x ∈ X . By Lemma 3 we know that for
every S ∈ X n , D∞ (A(S)kA(S0 )) ≤ k. We can now apply Lemma 31 with Z = A(S0 ), and β 0 = β to obtain
the result.
The conditions of Theorem 33 are satisfied by any ε-differentially private algorithm with k = log e · εn.
This immediately implies that the output of any ε-differentially private algorithm is β-statistically close to
the output of an algorithm with randomized description length of log e · εn + log ln(1/β). Special cases of
this property have been derived (using a technique similar to Lemma 31) in the context of proving lower
bounds for learning algorithms [BNS13] and communication complexity of differentially private protocols
[MMP+ 10].
26
B
Answering Queries via Description Length Bounds
In this section, we show a simple method for answering any adaptively chosen sequence of m statistical queries,
using a number of samples that scales only polylogarithmically in m. This is an exponential improvement
over what would be possible by naively evaluating the queries exactly on the given samples. Algorithms that
achieve such dependence were given in [DFH+ 14] and [BSSU15] using differentially private algorithms for
answering queries and the connection between generalization and differential privacy (in the same way as
we do in Section 4.1). Here we give a simpler algorithm which we analyze using description length bounds.
The resulting bounds are comparable to those achieved in [DFH+ 14] using pure differential privacy but are
somewhat weaker than those achieved using approximate differential privacy [DFH+ 14, BSSU15, NS15].
The mechanism we give here is based on the Median Mechanism of Roth and Roughgarden [RR10]. A
differentially private variant of this mechanism was introduced in [RR10] to show that it was possible to
answer exponentially many adaptively chosen counting queries (these are queries for an estimate of the
empirical mean of a function φ : X → [0, 1] on the dataset). Here we analyze a noise-free version and establish
its properties via a simple description length-based argument. We remark that it is possible to analogously
define and analyze the noise-free version of the Private Multiplicative Weights Mechanism of Hardt and
Rothblum [?]. This somewhat more involved approach would lead to better (but qualitatively similar) bounds.
Recall that statistical queries are defined by functions φ : X → [0, 1], and our goal is to correctly estimate
their expectation P[φ]. The Median Mechanism takes as input a dataset S and an adaptively chosen sequence
of such functions φ1 , . . . , φm , and outputs a sequence of answers a1 , . . . , am .
Algorithm Median Mechanism
Input: An upper bound m on the total number of queries, a dataset S and an accuracy parameter τ
1. Let α = τ3 .
2. Let Consistent0 = X
log m
α2
3. For a query φi do
(a) Compute apub
= median({ES 0 [φi ] : S 0 ∈ Consistenti−1 }).
i
(b) Compute apriv
= ES [φi ].
i
(c) If apub
− apriv
≤ 2α Then:
i
i
i. Output ai = apub
.
i
ii. Let Consistenti = Consistenti−1 .
(d) Else:
i. Output ai = bapriv
cα .
i
ii. Let Consistenti = {S 0 ∈ Consistenti−1 : |ai − ES 0 [φi ]| ≤ 2α}.
Figure 5: Noise-free version of the Median Mechanism from [RR10]
The guarantee we get for the Median Mechanism is as follows:
Theorem 34. Let β, τ > 0 and m ≥ B > 0. Let S denote a dataset of size n drawn i.i.d. from a distribution P
over X . Consider any algorithm that adaptively chooses functions φ1 , . . . , φm while interacting with the
Median Mechanism which is given S and values τ , β. For every i ∈ [m], let ai denote the answer of the
Median Mechanism on function φi : X → [0, 1]. Then
P [∃i ∈ [m], |ai − P[φi ]| ≥ τ ] ≤ β
27
whenever
81 · log |X | · log m · ln(3m/τ ) 9 ln(2m/β)
+
.
2τ 4
2τ 2
Proof. We begin with a simple lemma which informally states that for every distribution P, and for every set
of m functions φ1 , . . . , φm , there exists a small dataset that approximately encodes the answers to each of
the corresponding statistical queries.
n ≥ n0 =
Lemma 35 ([DR14] Theorem 4.2). For every dataset S over X , any set C = {φ1 , . . . , φm } of m functions
m
φi : X → [0, 1], and any α ∈ [0, 1], there exists a data set S 0 ∈ X t of size t = log
α2 such that:
max |ES [φi ] − ES 0 [φi ]| ≤ α.
φi ∈C
Next, we observe that by construction, the Median Mechanism (as presented in Figure 5) always returns
answers that are close to the empirical means of respective query functions.
Lemma 36. For every sequence of queries φ1 , . . . , φm and dataset S given to the Median Mechanism, we
have that for every i:
|ai − ES [φi ]| ≤ 2α.
Finally, we give a simple lemma from [RR10] that shows that the Median Mechanism only returns answers
computed using the dataset S in a small number of rounds – for any other round i, the answer returned is
computed from the set Consistenti−1 .
Lemma 37 ([RR10], see also Chapter 5.2.1 of [DR14]). For every sequence of queries φ1 , . . . , φm and a
dataset S given to the Median Mechanism:
{i : ai 6= apub
i } ≤
log |X | log m
.
α2
log m/α2
Proof. We simply note several facts. First, by construction, |Consistent0 | = |X |
. Second, by Lemma
35, for every i, |Consistenti | ≥ 1 (because for every set of m queries, there is at least one dataset S 0 of
size log m/α2 that is consistent up to error α with S on every query asked, and hence is never removed
from Consistenti on any round). Finally, by construction, on any round i such that ai 6= apub
, we have
i
|Consistenti | ≤ 21 · |Consistenti−1 | (because on any such round, the median dataset S 0 – and hence at least
half of the datasets in Consistenti−1 were inconsistent with the answer
given, and hence removed.) The
lemma follows from the fact that there can therefore be at most log |X |
log m/α2
many such rounds.
Our analysis proceeds by viewing the interaction between the data analyst and the Median Mechanism
(Fig. 5) as a single algorithm A. A takes as input a dataset S and outputs a set of queries and answers
A(S) = {φi , ai }m
i=1 . We will show that A’s output has short randomized description length (the data analyst
is a possibly randomized algorithm and hence A might be randomized).
Lemma 38. Algorithm A has randomized description length of at most
log |X | · log m
1
b≤
·
log
m
+
log
α2
α
bits.
Proof. We observe that for every fixing of the random bits of the data analyst the entire sequence of queries
asked by the analyst, together with the answers he receives, can be reconstructed from a record of the
indices i of the queries φi such that ai 6= apub
, together with their answers ai (i.e. it is sufficient to encode
i
M :={(i, ai ) | ai 6= apub
}.
Once
this
is
established,
the lemma follows because by Lemma 37, there are at
i
most
with
log |X | log m
α2
1
log α bits.
such queries, and for each one, its index can be encoded with log m bits, and its answer
28
To see why this is so, consider the following procedure for reconstructing the sequence (φ1 , a1 , . . . , φm , am )
of queries asked and answers received. For every fixing of the random bits of the data analyst, her queries
can be expressed as a sequence of functions (f1 , . . . , fm ) that take as input the queries previously asked to
the Median Mechanism, and the answers previously received, and output the next query to be asked. That is,
we have:
f1 () := φ1 , f2 (φ1 , a1 ) := φ2 , f3 (φ1 , a1 , φ2 , a2 ) := φ3 , . . . , fm (φ1 , a1 , . . . , φm−1 , am−1 ) := φm .
Assume inductively that at stage i, the procedure has successfully reconstructed (φ1 , a1 , . . . , φi−1 , ai−1 , φi ),
and the set Consistenti−1 (This is trivially satisfied at stage i = 1). For the inductive case, we need to recover
ai , φi+1 , and Consistenti . There are two cases we must consider at stage i. In the first case, i is such that
ai 6= apub
. But in this case, (i, ai ) ∈ M by definition, and so we have recovered ai , and we can compute
i
φi+1 = fi+1 (φ1 , a1 , . . . , φ1 , ai ), and can compute Consistenti = {S 0 ∈ Consistenti−1 : |ai − ES 0 [φi ]| ≤ 2α}. In
the other case, ai = apub
. But in this case, by definition of apub
, we can compute ai = median({ES 0 [φi ] :
i
i
0
S ∈ Consistenti−1 }), φi+1 = fi+1 (φ1 , a1 , . . . , φi , ai ), and Conisistenti = Consistenti−1 . This completes the
argument – by induction, M is enough to reconstruct the entire query/answer sequence.
Finally, we can complete the proof. By Hoeffding’s concentration inequality and the union bound we know
that for any every sequence of queries φ1 , . . . , φm and a dataset S of size n drawn from the distribution P n :
P [∃i, |ES [φi ] − P[φi ]| ≥ α] ≤ 2m · exp −2nα2 .
Applying Theorem 29 to the set R(φ1 , a1 , . . . , φm , am ) = {S | ∃i, |ES [φi ] − P[φi ]| ≥ α} we obtain that for
the queries φ1 , . . . , φm generated on the dataset S and corresponding answers of the Median Mechanism
a1 , . . . , am we have
P [∃i, |ES [φi ] − P[φi ]| ≥ α] ≤ 2b · 2m · exp −2nα2 .
log |X | log m
·log(m/α)
α2
· 2m · exp −2nα2 .
≤ 2
Solving, we have that whenever:
n≥
log |X | · log m
ln(2m/β)
· ln(m/α) +
,
2α4
2α2
we have: P [∃i, |ES [φi ] − P[φi ]| ≥ α] ≤ β. Combining this with Lemma 36 we have:
P [∃i ∈ [m], |ai − P[φi ]| ≥ 3α] ≤ β.
Plugging in τ = 3α gives the theorem.
29
| 8cs.DS
|
A BLOWUP ALGEBRA OF HYPERPLANE ARRANGEMENTS
arXiv:1701.03470v1 [math.AC] 12 Jan 2017
MEHDI GARROUSIAN, ARON SIMIS AND ŞTEFAN O. TOHĂNEANU
A BSTRACT. It is shown that the Orlik-Terao algebra is graded isomorphic to the special fiber of the ideal I
generated by the (n − 1)-fold products of the members of a central arrangement of size n. This momentum
is carried over to the Rees algebra (blowup) of I and it is shown that this algebra is of fiber-type and CohenMacaulay. It follows by a result of Simis-Vasconcelos that the special fiber of I is Cohen-Macaulay, thus giving
another proof of a result of Proudfoot-Speyer about the Cohen-Macauleyness of the Orlik-Terao algebra.
I NTRODUCTION
The central theme of this paper is to study the ideal theoretic aspects of the blowup of a projective space
along a certain scheme of codimension 2. To be more precise, let A = {ker(ℓ1 ), . . . , ker(ℓn )} be an
arrangement of hyperplanes in Pk−1 and consider the closure of the graph of the following rational map
Pk−1 99K Pn−1 ,
x 7→ (1/ℓ1 (x) : · · · : 1/ℓn (x)).
Rewriting the coordinates of the map as forms of the same positive degree in the source Pk−1 =
Proj(k[x1 , . . . , xk ]), we are led to consider the corresponding graded k[x1 , . . . , xk ]-algebra, namely, the
Rees algebra of the ideal of k[x1 , . . . , xk ] generated by the (n − 1)-fold products of ℓ1 , . . . , ℓn .
It is our view that bringing into the related combinatorics a limited universe of gadgets and numerical
invariants from commutative algebra may be of help, specially regarding the typical operations with ideals
and algebras. This point of view favors at the outset a second look at the celebrated Orlik-Terao algebra
k[1/ℓ1 , . . . , 1/ℓn ] which is regarded as a commutative counterpart to the combinatorial Orlik-Solomon algebra. The fact that the former, as observed by some authors, has a model as a finitely generated graded
k-subalgebra of a finitely generated purely transcendental extension of the field k, makes it possible to
recover it as the homogeneous coordinate ring of the image of a certain rational map.
This is our departing step to naturally introduce other commutative algebras into the picture. As shown
in Theorem 2.4, the Orlik-Terao algebra now becomes isomorphic, as graded k-algebra, to the special fiber
algebra (also called fiber cone algebra or central algebra) of the ideal I generated by the (n−1)-fold products
of the members of the arrangement A. This algebra is in turn defined as a residue algebra of the Rees algebra
of I, so it is only natural to look at this and related constructions. One of these constructions takes us to the
symmetric algebra of I, and hence to the syzygies of I. Since I turns out to be a perfect ideal of codimension
2, its syzygies are rather simple and allow us to further understand these algebras.
As a second result along this line of approach, we show that a presentation ideal of the Rees algebra of I
can be generated by the syzygetic relations and the Orlik-Terao ideal (see Theorem 4.2). This property has
been coined fiber type property in the recent literature of commutative algebra.
2010 Mathematics Subject Classification. Primary 13A30, 14N20; Secondary 13C14, 13D02, 13D05.
Key words and phrases. Rees algebra, special fiber algebra, Orlik-Terao algebra, Cohen-Macaulay.
The second author has been partially supported by a CNPq grant (302298/2014-2) and by a CAPES-PVNS Fellowship
(5742201241/2016); he is currently holding a one year Senior Visiting Professorship (2016/08929-7) at the Institute for Mathematical and Computer Sciences, University of São Paulo, São Carlos, Brazil.
Garrousian’s Address: Department of Mathematics, University of Western Ontario, London, ON N6A 5B7, Canada, Email:
[email protected].
Simis’ Address: Departamento de Matemática, Centro de Ciências Exatas e da Natureza, Universidade Federal de Pernambuco,
50740-560, Recife, Pernambuco, Brazil, Email: [email protected].
Tohaneanu’s Address: Department of Mathematics, University of Idaho, Moscow, Idaho 83844-1103, USA, Email: [email protected], Phone: 208-885-6234, Fax: 208-885-5843.
1
2
MEHDI GARROUSIAN, ARON SIMIS AND ŞTEFAN O. TOHĂNEANU
The third main result of this work, as an eventual outcome of these methods, is a proof of the CohenMacaulay property of the Rees algebra of I (see Theorem 4.9).
The typical argument in the proofs is induction on the size or rank of the arrangements. Here we draw
heavily on the operations of deletion and contraction of an arrangement. In particular, we introduce a variant
of a multiarrangement that allows repeated linear forms to be tagged with arbitrarily different coefficients.
Then the main breakthrough consists in getting a precise relation between the various ideals or algebras
attached to the original arrangement and those attached to the minors.
One of the important facts about the Orlik-Terao algebra is that it is Cohen-Macaulay, as proven by
Proudfoot-Speyer [12]. Using a general result communicated by the second author and Vasconcelos, we
recover this result as a consequence of the Cohen-Macaulay property of the Rees algebra.
The structure of this paper is as follows. The first section is an account of the needed preliminaries from
commutative algebra. The second section expands on highlights of the settled literature about the OrlikTerao ideal as well as a tangential discussion on the so-called non-linear invariants of our ideals such as the
reduction number and analytic spread. The third section focuses on the ideal of (n − 1)-fold products and
the associated algebraic constructions. The last section is devoted to the statements and proofs of the main
theorems where we draw various results from the previous sections to establish the arguments.
1. I DEAL
THEORETIC NOTIONS AND BLOWUP ALGEBRAS
The Rees algebra of an ideal I in a ring R is the R-algebra
M
R(I) :=
I i.
i≥0
This is a standard R-graded algebra with R(I)0 = R, where multiplication is induced by the internal
multiplication rule I r I s ⊂ I r+s . One can see that there is a graded isomorphism R[It] ≃ R(I), where
R[It] is the homogeneous R-subalgebra of the standard graded polynomial R[t] in one variable over R,
generated by the elements at, a ∈ I, of degree 1.
Quite generally, fixing a set of generators of I determines a surjective homomorphism of R-algebras from
a polynomial ring over R to R[It]. The kernel of such a map is called a presentation ideal of R[It]. In this
generality, even if R is Noetherian (so I is finitely generated) the notion of a presentation ideal is quite loose.
In this work we deal with a special case in which R = k[x1 , . . . , xk ] is a standard graded polynomial
ring over a field k and I = hg1 , . . . , gn i is an ideal generated by forms g1 , . . . , gn of the same degree. Let
T = R[y1 , . . . , yn ] = k[x1 , . . . , xk ; y1 , . . . , yn ], a standard bigraded k-algebra with deg xi = (1, 0) and
deg yj = (0, 1). Using the given generators to obtain an R-algebra homomorphism
ϕ : T = R[y1 , . . . , yn ] −→ R[It], yi 7→ gi t,
yields a presentation ideal I which is bihomogeneous in the bigrading of T . Therefore, R[It] acquires the
corresponding bigrading.
Changing k-linearly independent sets of generators in the same degree amounts to effecting an invertible
k-linear map, so the resulting effect on the corresponding presentation ideal is pretty much under control. For
this reason, we will by abuse talk about the presentation ideal of I by fixing a particular set of homogeneous
generators of I of the same degree. Occasionally, we may need to bring in a few superfluous generators into
a set of minimal generators.
Since the given generators have the same degree they span a linear system defining a rational map
Φ : Pk−1 99K Pn−1 ,
(1)
by the assignment x 7→ (g1 (x) : · · · : gn (x)), when some gi (x) 6= 0.
The ideal I is often called the base ideal (to agree with the base scheme) of Φ. Asking when Φ is birational
onto its image is of interest and we will briefly deal with it as well. Again note that changing to another set
of generators in the same degree will not change the linear system thereof, defining the same rational map
up to a coordinate change at the target.
A BLOWUP ALGEBRA OF HYPERPLANE ARRANGEMENTS
3
The Rees algebra brings along other algebras of interest. In the present setup, one of them is the special
fiber F(I) := R[It] ⊗R R/m ≃ ⊕s≥0 I s /mI s , where m = hx1 , . . . , xk i ⊂ R. The Krull dimension of the
special fiber ℓ(I) := dim F(I) is called the analytic spread of I.
The analytic spread is a significant notion in the theory of reductions of ideals. An ideal J ⊂ I is said
to be a reduction of I if I r+1 = JI r for some r. Most notably, this is equivalent to the condition that the
natural inclusion R[Jt] ֒→ R[It] is a finite morphism. The smallest such r is the reduction number rJ (I)
with respect to J. The reduction number of I is the infimum of all rJ (I) for all minimal reductions J of I;
this number is denoted by r(I).
Geometrically, the relevance of the special fiber lies in the following result, which we isolate for easy
reference:
Lemma 1.1. Let Φ be as in (1) and I its base ideal. Then the homogeneous coordinate ring of the image of
Φ is isomorphic to the special fiber F(I) as graded k-algebras.
To see this, note that the Rees algebra defines a biprojective subvariety of Pk−1 × Pn−1 , namely the
closure of the graph of Φ. Projecting down to the second coordinate recovers the image of Φ. At the level
of coordinate rings this projection corresponds to the inclusion k[Id t] = k[g1 t, . . . , gn t] ⊂ R[It], where
g1 , . . . , gn are forms of the degree d, this inclusion is a split k[Id t]-module homomorphism with mR[It] as
direct complement. Therefore, one has an isomorphism of k-graded algebras k[Id ] ≃ k[Id t] ≃ F(I).
As noted before, the presentation ideal of R[It]
M
I=
I(a,b) ,
(a,b)∈N×N
is a bihomogeneous ideal in the standard bigrading of T . Two basic subideals of I are hI(0,−) i and hI(−,1) i.
The two come in as follows.
Consider the natural surjections
T
ϕ
// R[It]
⊗R R/m
// F(I)
44
ψ
where the kernel of the leftmost map is the presentation ideal I of R[It]. Then, we have
F(I) ≃
T
T
k[y1 , . . . , yn ]
≃
≃
.
ker ψ
hker ϕ|(0,−) , mi
hI(0,−) i
Thus, hI(0,−) i is the homogeneous defining ideal of the special fiber (or, as explained in Lemma 1.1, of the
image of the rational map Φ).
As for the second ideal hI(−,1) i, one can see that it coincides with the ideal of T generated by the biforms
s1 y1 + · · · + sn yn ∈ T , whenever (s1 , . . . , sn ) is a syzygy of g1 , . . . , gn of certain degree in R. Thinking
about the one-sided grading in the y’s there is no essential harm in denoting this ideal simply by I1 . Thus,
T /I1 is a presentation of the symmetric algebra S(I) of I. It obtains a natural surjective map of R-graded
algebras
S(I) ≃ T /I1 ։ T /I ≃ R(I).
As a matter of calculation, one can easily show that I = I1 : I ∞ , the saturation of I1 with respect to I.
The ideal I is said to be of linear type provided I = I1 , i.e., when the above surjecton is injective. It is
said to be of fiber type if I = I1 + hI(0,−) i = hI1 , I(0,−) i.
A basic homological obstruction for an ideal to be of linear type is the so-called G∞ condition of ArtinNagata [2], also known as the F1 condition [7]. A weaker condition is the so-called Gs condition, for a
suitable integer s. All these conditions can be stated in terms of the Fitting ideals of the given ideal or,
equivalently, in terms of the various ideals of minors of a syzygy matrix of the ideal. In this work we will
have a chance to use condition Gk , where k = dim R < ∞. Given a free presentation
ϕ
Rm −→ Rn −→ I → 0
4
MEHDI GARROUSIAN, ARON SIMIS AND ŞTEFAN O. TOHĂNEANU
of an ideal I ⊂ R, the Gk condition for I means that
ht(Ip (ϕ)) ≥ n − p + 1, for p ≥ n − k + 1,
(2)
where It (ϕ) denotes the ideal generated by the t-minors of ϕ. Note that nothing is required about the values
of p strictly smaller than n − k + 1 since for such values one has n − p + 1 > k = dim R, which makes the
same bound impossible.
A useful method to obtain new generators of I from old generators (starting from generators of I1 ) is
via Sylvester forms (see [8, Proposition 2.1]), which has classical roots as the name indicates. It can be
defined in quite generality as follows: let R := k[x1 , . . . , xk ], and let T := R[y1 , . . . , yn ] as above. Given
F1 , . . . , Fs ∈ I, let J be the ideal of R generated by all the coefficients of the Fi – the so-called R-content
ideal. Suppose J = ha1 , . . . , aq i, where ai are forms of the same degree. Then we have the matrix equation:
a1
F1
a2
F2
.. = A · .. ,
.
.
Fs
aq
where A is an s × q matrix with entries in T .
If q ≥ s and if the syzygies on Fi′ s are in mT , then the determinant of any s × s minor of A is an
element of I. These determinants are called Sylvester forms. The main use in this work is to show that the
Orlik-Terao ideal is generated by such forms (Proposition 3.5).
The last invariant we wish to comment on is the reduction number r(I). For convenience, we state the
following result:
Proposition 1.2. With the above notation, suppose that the special fiber F(I) is Cohen-Macaulay. Then the
reduction number r(I) of I coincides with the Castelnuovo-Mumford regularity reg(F(I)) of F(I).
Proof. By [19, Proposition 1.85], when the special fiber is Cohen-Macaulay, one can read r(I) off the
Hilbert series. Write
1 + h1 s + h2 s2 + · · · + hr sr
HS(F(I), s) =
,
(1 − s)d
with hr 6= 0 and d = ℓ(I), the dimension of the fiber (analytic spread). Then, r(I) = r.
Since F(I) ≃ S/hI(0,−) i, where S := k[y1 , . . . , yn ], we have that F(I) has a minimal graded S-free
resolution of length equal to m := ht hI(0,−) i, and reg(F(I)) = α − m, where α is the largest shift in
the minimal graded free resolution, occurring also at the end of this resolution. These last two statements
mentioned here come from the Cohen-Macaulayness of F(I).
The additivity of Hilbert series under short exact sequences of modules, together with the fact that
sv
HS(S u (−v), s) = u
gives that r + m = α = m + reg(F(I)), so r(I) = reg(F(I)).
(1 − s)n
2. H YPERPLANE A RRANGEMENTS
Let A = {H1 , . . . , Hn } ⊂ Pk−1 be a central hyperplane arrangement of size n and rank k. Here
Hi = ker(ℓi ), i = 1, . . . , n, where each ℓi is a linear form in R := k[x1 , . . . , xk ] and hℓ1 , . . . , ℓn i =
m := hx1 , . . . , xk i. From the algebraic viewpoint, there is a natural emphasis on the linear forms ℓi and the
associated ideal theoretic notions.
Deletion and contraction are useful operations upon A. Fixing an index 1 ≤ i ≤ n, one introduces two
new minor arrangements:
A′ = A \ {Hi } (deletion),
A′′ = A′ ∩ Hi := {Hj ∩ Hi | 1 ≤ j ≤ n , j 6= i} (contraction).
Clearly, A′ is a subarrangement of A of size n − 1 and rank at most k, while A′′ is an arrangement of size
≤ n − 1 and rank k − 1.
A BLOWUP ALGEBRA OF HYPERPLANE ARRANGEMENTS
5
Contraction comes with a natural multiplicity given by counting the number of hyperplanes of A′ that
give the same intersection. A modified version of such a notion will be thoroughly used in this work.
The following notion will play a substantial role in some inductive arguments throughout the paper: ℓi is
′
called
T a coloop if the rank of the deletion A withk respect to ℓi is k − 1, ie. drops by one. This simply means
that j6=i Hj is a line rather than the origin in A . Otherwise, we say that ℓi is a non-coloop.
2.1. The Orlik-Terao algebra. One of our motivations is to clarify the connections between the Rees
algebra and the Orlik-Terao algebra which is an important object in the theory of hyperplane arrangements.
We state the definition and review some of its basic properties below.
Let A ⊂ Pk−1 be a hyperplane arrangement as above. Suppose ci1 ℓi1 + · · · + cim ℓim = 0 is a linear
dependency among m of the linear forms defining A, denoted D. Consider the following homogeneous
polynomial in S := k[y1 , . . . , yn ]:
m
m
Y
X
cij
y ik .
(3)
∂D :=
j=1
j6=k=1
Note that deg(∂D) = m − 1.
The Orlik-Terao algebra of A is the standard graded k-algebra
OT(A) := S/∂(A),
where ∂(A) is the ideal of S generated by {∂D|D a dependency of A}, with ∂D as in (3) – called the OrlikTerao ideal. This algebra was introduced in [10] as a commutative analog of the classical combinatorial
Orlik-Solomon algebra, in order to answer a question of Aomoto.
The following remark states a few important properties of this algebra.
Remark 2.1.
i. Recalling that a circuit is a minimally dependent set, one has that ∂(A) is generated by ∂C, where C
runs over the circuits of A ([11]). In addition, these generators form an universal Gröbner basis for
∂(A) ([12]).
ii. OT(A) is Cohen-Macaulay ([12]).
iii. OT(A) ≃ k[1/ℓ1 , . . . , 1/ℓn ], a k-dimensional k-subalgebra of the field of fractions k(x1 , . . . , xk )
([15, 17]). The corresponding projective variety is called the reciprocal plane and it is denoted by L−1
A .
iv. Although the Orlik-Terao algebra is sensitive to the linear forms defining A, its Hilbert series only
depends on the underlying combinatorics ([17]). Let
X
π(A, s) =
µA (F )(−s)r(F ) .
F ∈L(A)
be the Poincaré polynomial where µA denotes the Möbius function, r is the rank function and F runs
over the flats of A. Then, we have
s
).
HS(OT(A), s) = π(A,
1−s
See [10] for details and [17] and [3] for proofs of the above statement.
2.2. Ideals of products from arrangements. Let A = {ℓ1 , . . . , ℓn } denote a central
Q arrangement in R :=
k[x1 , . . . , xk ], n ≥ k. Denoting [n] := {1, . . . , n}, if S ⊂ [n], then we set ℓS := i∈S ℓi , ℓ∅ := 1. Also set
S c := [n] \ S.
Let S := {S1 , . . . , Sm }, where Sj ⊆ [n] are subsets of the same size e. We are interested in studying the
Rees algebras of ideals of the form
IS := hℓS1 , . . . , ℓSm i ⊂ R.
(4)
Example 2.2. (i) (The Boolean case) Let n = k and ℓi = xi , i = 1, . . . , k. Then the ideal IS is monomial
for any S. In the simplest case where e = n − 1, it is the ideal of the partial derivatives of the monomial
x1 · · · xk – also the base ideal of the classical Möbius involution. For e = 2 the ideal becomes the edge ideal
6
MEHDI GARROUSIAN, ARON SIMIS AND ŞTEFAN O. TOHĂNEANU
of a simple graph with k vertices. In general, it gives a subideal of the ideal of paths of a given length on the
complete graph and, as such, it has a known combinatorial nature.
(ii) ((n − 1)-fold products) Here one takes S1 := [n] \ {1}, . . . , Sn := [n] \ {n}. We will designate the
corresponding ideal by In−1 (A). This case will be the main concern of the paper and will be fully examined
in the following sections.
(iii) (a-fold products) This is a natural extension of (ii), where Ia (A) is the ideal generated by all distinct
a-products of the linear forms defining A. The commutative algebraic properties of these ideals connect
strongly to properties of the linear code built on the defining linear forms (see [1]). In addition, the dimensions of the vector spaces generated by a-fold products give a new interpretation to the Tutte polynomial of
the matroid of A (see [3]).
We can naturally introduce the reciprocal plane algebra
1
1
−1
,...,
LS := k
c
ℓS1c
ℓ Sm
(5)
as a generalized version of the notion mentioned in Remark 2.1 (iii).
Proposition 2.3. In the above setup there is a graded isomorphism of k-algebras
1
1
k[ℓS1 , . . . , ℓSm ] ≃ k
,...,
.
c
ℓS1c
ℓ Sm
Proof. Consider both algebras as homogeneous k-subalgebras of the homogeneous total quotient ring of the
standard polynomial ring R, generated in degrees e and −(d − e), respectively. Then multiplication by the
total product ℓ[d] gives the required isomorphism:
·ℓ[d]
1
1
,...,
−→ k[ℓS1 , . . . , ℓSm ].
k
c
ℓS1c
ℓ Sm
A neat consequence is the following result:
Theorem 2.4. Let A denote a central arrangement of size n, let S := {S1 , . . . , Sm } be a collection of
subsets of [n] of the same size and let IS be as in (4). Then the reciprocal plane algebra L−1
S is isomorphic
to the special fiber of the ideal IS as graded k-algebras. In particular, the Orlik-Terao algebra OT(A) is
graded isomorphic to the special fiber F(I) of the ideal I = In−1 (A) of (n − 1)-fold products of A.
Proof. It follows immediately from Proposition 2.3 and Lemma 1.1.
Remark 2.5. In the case of the Orlik-Terao algebra, the above result gives an answer to the third question at
the end of [14]. Namely, let k ≥ 3 and consider the rational map Φ as in (1). Then Theorem 2.4 says that the
projection of the graph of Φ onto the second factor coincides with the reciprocal plane L−1
A (see Remark 2.1
(iii)). In addition, the ideal I := In−1 (A) has a similar primary decomposition as obtained in [14, Lemmas
3.1 and 3.2], for arbitrary k ≥ 3. By [1, Proposition 2.2], one gets
\
I=
I(Y )µA (Y ) ,
Y ∈L2 (A)
Theorem 2.4 contributes additional information on certain numerical invariants and properties in the strict
realm of commutative algebra and algebraic geometry.
Corollary 2.6. Let I := In−1 (A) denote the ideal generated by the (n − 1)-fold products coming from a
central arrangement of size n and rank k. One has:
(a) The special fiber F(I) of I is Cohen-Macaulay.
(b) The analytic spread is ℓ(I) = k.
(c) The map Φ is birational onto its image.
(d) The reduction number is r(I) = k − 1.
A BLOWUP ALGEBRA OF HYPERPLANE ARRANGEMENTS
7
Proof. (a) It follows from from Theorem 2.4 via Remark 2.1 (ii).
(b) It follows by the same token from 2.1 (iii).
(c) This follows from and [5, Theorem 3.2] since the ideal I is linearly presented (see proof of Lemma
3.1), and ℓ(I) = k, maximum possible.
(d) Follows from Part (a), Proposition 1.2, and [14, Theorem 3.7].
It may be interesting to remark that, because of this value, in particular the Orlik-Terao algebra is the
homogeneous coordinate ring of a variety of minimal degree if and only if k = 2, in which case it is the
homogeneous coordinate ring of the rational normal curve.
3. I DEALS
OF
(n − 1)- FOLD
PRODUCTS AND THEIR BLOWUP ALGEBRAS
As mentioned in Example 2.2, a special case of the ideal IS , extending the case of the ideal generated by
the (n − 1)-fold products, is obtained by fixing a ∈ {1, . . . , n} and considering the collection of all subsets
of [n] of cardinality a. Then the corresponding ideal is
Ia (A) := hℓi1 · · · ℓia |1 ≤ i1 < · · · < ia ≤ ni ⊂ R
and is called the ideal generated by the a-fold products of linear forms of A. The projective schemes
defined by these ideals are known as generalized star configuration schemes. Unfortunately, only few things
are known about these ideals: if d is the minimum distance of the linear code built from the linear forms
defining A and if 1 ≤ a ≤ d, then Ia (A) = ma (cf. [18, Theorem 3.1]); and the case when a = n is trivial.
In the case where a = n − 1, some immediate properties are known already, yet the more difficult
questions in regard to the blowup and related algebras have not been studied before. These facets, to be
throughly examined in the subsequent sections, is our main endeavor in this work.
Henceforth, we will be working with the following data: A is an arrangement with n ≥ k and for every
1 ≤ i ≤ n, we consider the (n − 1)-fold products of the n linear forms defining the hyperplanes of A
fi := ℓ1 · · · ℓˆi · · · ℓn ∈ R,
and write
I := In−1 (A) := hf1 , . . . , fn i.
Let T = k[x1 , . . . , xk , y1 , . . . , yn ] = R[y1 , . . . , yn ] as before and denote by I(A, n − 1) ⊂ T the presentation ideal of the Rees algebra R[It] corresponding to the generators f1 , . . . , fn .
3.1. The symmetric algebra. Let I1 (A, n − 1) ⊂ T stand for the subideal of I(A, n − 1) presenting the
symmetric algebra S(I) of I = In−1 (A).
Lemma 3.1. With the above notation, one has:
(a) The ideal I = In−1 (A) is a perfect ideal of codimension 2.
(b) I1 (A, n − 1) = hℓi yi − ℓi+1 yi+1 |1 ≤ i ≤ n − 1i.
(c) I1 (A, n − 1) is an ideal of codimension k; in particular, it is a complete intersection if and only if
n = k.
Proof. (a) This is well-known, but we give the argument for completeness. Clearly, I has codimension 2.
The following reduced Koszul like relations are syzygies of I: ℓi yi − ℓi+1 yi+1 , 1 ≤ i ≤ n − 1. They alone
form the following matrix of syzygies of I:
ℓ1
−ℓ2
ℓ2
..
.
−ℓ3
ϕ=
.
..
. ℓn−1
−ℓn
8
MEHDI GARROUSIAN, ARON SIMIS AND ŞTEFAN O. TOHĂNEANU
Since the rank of this matrix is n − 1, it is indeed a full syzygy matrix of I; in particular, I has linear
resolution
ϕ
0 −→ R(−n)n−1 −→ R(−(n − 1))n −→ I −→ 0.
(b) This is an expression of the details of (a).
(c) Clearly, I1 (A, n − 1) ⊂ mT , hence its codimension is at most k. Assuming, as we may, that
{ℓ1 , . . . , ℓk } is k-linearly independent, we contend that the elements s := {ℓi yi − ℓi+1 yi+1 , 1 ≤ i ≤ k},
form a regular sequence. To see this, we first apply a k-linear automorphism of R to assume that ℓi = xi ,
for 1 ≤ i ≤ k – this will not affect the basic ideal theoretic invariants associated to I. Then note
that in the set of generators of I1 (A, n − 1) the elements of s can be replaced by the following ones:
{xi yi − ℓk+1 yk+1 , 1 ≤ i ≤ k}. Clearly, this is a regular sequence – for example, because hxi yi , 1 ≤ i ≤ ki
is the initial ideal of the ideal generated by this sequence, in the revlex order.
There are two basic ideals that play a distinguished role at the outset. In order to capture both in one
single blow, we consider the Jacobian matrix of the generators of I1 (A, n − 1) given in Lemma 3.1 (b).
Its transpose turns out to be the stack of two matrices, the first is the Jacobian matrix with respect to the
variables y1 , . . . , yn – which coincides with the syzygy matrix φ of I as described in the proof of Lemma 3.1
(a) – while the second is the Jacobian matrix B = B(φ) with respect to the variables x1 , . . . , xk – the socalled Jacobian dual matrix of [16]. The offspring are the respective ideals of maximal minors of these
stacked matrices, the first retrieves I, while the second gives an ideal Ik (B) ⊂ S = k[y1 , . . . , yn ] that will
play a significant role below (see also Proposition 4.1) as a first crude approximation to the Orlik-Terao
ideal.
Proposition 3.2. Let S(I) ≃ T /I1 (A, n − 1) stand for the symmetric algebra of the ideal I of (n − 1)-fold
products. Then:
(i) depth(S(I)) ≤ k + 1.
(ii) As an ideal in T , every minimal prime of S(I) is either mT , the Rees ideal I(A, n − 1) or else has
the form (ℓi1 , . . . , ℓis , yj1 , . . . , yjt ), where 2 ≤ s ≤ k − 1, t ≥ 1, {i1 , . . . , is } ∩ {j1 , . . . , jt } = ∅,
and ℓi1 , . . . , ℓis are k-linearly independent.
(iii) The primary components relative to the minimal primes m = (x)T and I(A, n − 1) are radical; in
addition, with the exception of mT , every minimal prime of S(I) contains the ideal Ik (B).
Proof. (i) Since I(A, n − 1) is a prime ideal which is a saturation of I1 (A, n − 1) then it is an associated
prime of S(I). Therefore, depth(S(I)) ≤ dim R(I) = k + 1.
(ii) Since I(A, n − 1) is a saturation of I1 (A, n − 1) by I, one has I(A, n − 1) I t ⊂ I1 (A, n − 1), for
some t ≥ 1. This implies that any (minimal) prime of S(I) in T contains either I or I(A, n − 1). By the
proof of (i), I(A, n − 1) is an associated prime of S(I), hence it must be a minimal prime thereof since a
minimal prime of S(I) properly contained in it would have to contain I, which is absurd.
Now, suppose P ⊂ T is a minimal prime of S(I) containing I. One knows by Lemma 3.1 that m = (x)T
is a minimal prime of S(I). Therefore, we assume that mT 6⊂ P . Since any minimal prime of I is a complete
intersection of two distinct linear forms of A then P contains at least two, and at most k − 1, linearly
independent linear forms of A. On the other hand, since I1 (A, n − 1) ⊂ P , looking at the generators of
I1 (A, n − 1) as in Lemma 3.1 (b), by a domino effect principle we finally reach the desired format for P as
stated.
(iii) With the notation prior to the statement of the proposition, we claim the following equality:
I1 (A, n − 1) : Ik (B)∞ = mT
It suffices to show for the first quotient as mT is a prime ideal. The inclusion m Ik (B) ⊂ I1 (A, n − 1) is a
consequence of the Cramer rule. The reverse inclusion is obvious because I1 (A, n − 1) ⊂ mT implies that
I1 (A, n − 1) : Ik (B) ⊂ mT : Ik (B) = mT , as mT is a prime ideal.
Note that, as a very crude consequence, one has Ik (B) ⊂ I(A, n − 1).
A BLOWUP ALGEBRA OF HYPERPLANE ARRANGEMENTS
9
Now, let P(mT ) denote the primary component of mT in I1 (A, n − 1). Then
mT = I1 (A, n − 1) : Ik (B)∞ ⊂ P(mT ) : Ik (B)∞ = P(mT ).
The same argument goes through for the primary component of I(A, n − 1) using the ideal I instead of
Ik (B).
To see the last statement of the item, let P denote the primary component of one of the remaining minimal
primes P of S(I). Since P : Ik (B)∞ is P -primary and m 6⊂ P , then by the same token we get that
Ik (B) ⊂ P .
Remark 3.3. (a) It will be shown in the last section that the estimate in (i) is actually an equality.
As a consequence, every associated prime of S(I) viewed in T has codimension at most n − 1. This
will give a much better grip on the minimal primes of the form hℓi1 , . . . , ℓis , yj1 , . . . , yjt i. Namely, one
must have in addition that s + t ≤ n − 1 and, moreover, due to the domino effect principle, one must have
s = k − 1, hence t ≤ n − k.
(b) We conjecture that S(I) is reduced.
The property (R0 ) of Serre’s is easily verified due to the format of the Jacobian matrix as explained before
the above proposition. The problem is, of course, the property (S1 ), the known obstruction for the existence
of embedded associated primes. The case where n = k + 1, is easily determined. Here the minimal primes
are seen to be m, hx1 , . . . , xk−1 , yk i and the Rees ideal hI1 (A, k), ∂i, where ∂ is the relation corresponding
to the unique circuit. A calculation will show that the three primes intersect in I1 (A, k). As a side, this fact
alone implies that the maximal regular sequence in the proof of Lemma 3.1 (c) generates a radical ideal. For
n ≥ k + 2 the calculation becomes sort of formidable, but we will prove later on that the Rees ideal is of
fiber type.
(c) The weaker question as to whether the minimal component of S(I) is radical seems pliable.
If the conjectural statement in Remark 3.3 (b) is true then, for any linear form ℓ = ℓi the following basic
formula holds
I1 (A, n − 1) : ℓ = I(A, n − 1) ∩
\
ℓ∈P
/
P
!
,
where P denotes a minimal prime other that mT and I(A, n − 1), as described in proposition 3.2 (i). Thus
one would recover sectors of the Orlik-Terao generators inside this colon ideal.
Fortunately, this latter virtual consequence holds true and has a direct simple proof. For convenience of
later use, we state it explicitly. Let ∂(A|ℓ ) denote the subideal of ∂(A) generated by all polynomial relations
∂ corresponding to minimal dependencies (circuits) involving the linear form ℓ ∈ A.
Lemma 3.4. ∂(A|ℓ ) ⊂ I1 (A, n − 1) : ℓ.
Proof. Say, ℓ = ℓ1 . Let D : a1 ℓ1 + a2 ℓ2 + · · · + as ℓs = 0 be a minimal dependency involving ℓ1 , for some
3 ≤ s ≤ n. In particular, ai 6= 0, i = 1, . . . , s. The corresponding generator of ∂(A|ℓ1 ) is
∂D := a1 y2 y3 · · · ys + a2 y1 y3 · · · ys + · · · + as y1 y2 · · · ys−1 .
The following calculation is straightforward.
10
MEHDI GARROUSIAN, ARON SIMIS AND ŞTEFAN O. TOHĂNEANU
ℓ1 ∂D =
+
=
+
=
+
=
+
=
a1 ℓ1 y2 y3 · · · ys + (ℓ1 y1 − ℓ2 y2 )(a2 y3 · · · ys + · · · + as y2 · · · ys−1 )
ℓ2 y2 (a2 y3 · · · ys + · · · + as y2 · · · ys−1 )
(a1 ℓ1 + a2 ℓ2 )y2 y3 · · · ys + (ℓ1 y1 − ℓ2 y2 )(a2 y3 · · · ys + · · · + as y2 · · · ys−1 )
ℓ2 y2 (a3 y2 y4 · · · ys + · · · + as y2 y3 · · · ys−1 )
(−a3 ℓ3 − · · · − as ℓs )y2 y3 · · · ys + (ℓ1 y1 − ℓ2 y2 )(a2 y3 · · · ys + · · · + as y2 · · · ys−1 )
ℓ2 y22 (a3 y4 · · · ys + · · · + as y3 · · · ys−1 )
(ℓ1 y1 − ℓ2 y2 )(a2 y3 · · · ys + · · · + as y2 · · · ys−1 ) + y2 (ℓ2 y2 − ℓ3 y3 )a3 y4 · · · ys
· · · + y2 (ℓ2 y2 − ℓs ys )as y3 · · · ys−1
a2 y3 · · · ys (ℓ1 y1 − ℓ2 y2 ) + a3 y2 y4 · · · ys (ℓ1 y1 − ℓ3 y3 ) + · · · + as y2 · · · ys−1 (ℓ1 y1 − ℓs ys ).
Hence the result.
3.2. Sylvester forms. The Orlik-Terao ideal ∂(A) has an internal structure of classical flavor, in terms of
Sylvester forms.
Proposition 3.5. The generators ∂(A) of the Orlik-Terao ideal are Sylvester forms obtained from the generators of the presentation ideal I1 (A, n − 1) of the symmetric algebra of I.
Q
Proof. Let D be a dependency ci1 ℓi1 + · · · + cim ℓim = 0 with all coefficients cij 6= 0. Let f = ni=1 ℓi .
Evaluating the Orlik-Terao element ∂D on the products we have
∂D(f1 , . . . , fn ) =
m
X
j=1
m
cij
X
f m−1
f m−2
f m−1
cij m
=
ℓ ij =
(ci ℓi + · · · + cim ℓim ) = 0.
m
Πj6=k=1 ℓik
Πk=1 ℓik
ℓ i1 · · · ℓ im 1 1
j=1
Therefore, ∂D ∈ I(A, n − 1), and since ∂D ∈ S := k[y1 , . . . , yn ], then ∂D ∈ hI(A, n − 1)(0,−) i.
For the second part, suppose that the minimal generators of I1 (A, n − 1) are
∆1 := ℓ1 y1 − ℓ2 y2 , ∆2 := ℓ2 y2 − ℓ3 y3 , . . . , ∆n−1 := ℓn−1 yn−1 − ℓn yn .
Without loss of generality suppose ℓj = c1 ℓ1 + · · · + cj−1 ℓj−1 is some arbitrary dependency D. We have
y1
−y2
0
···
0
0
∆1
ℓ1
0
y
−y
·
·
·
0
0
2
3
∆2
ℓ2
..
..
..
..
..
=
..
.
· ..
.
.
.
.
.
.
.
0
0
0
···
yj−2
−yj−1
∆j−1
ℓj−1
−c1 yj −c2 yj −c3 yj · · · −cj−2 yj yj−1 − cj−1 yj
The determinant of the (j − 1) × (j − 1) matrix we see above is ±∂D.
3.3. A lemma on deletion. In this and the next parts we build on the main tool of an inductive procedure.
Let A′ = A \ {ℓ1 }, and denote n′ := |A′ | = n − 1. We would like to investigate the relationship between
the Rees ideal I(A′ , n′ − 1) of In′ −1 (A′ ) and the Rees ideal I(A, n − 1) of In−1 (A), both defined in terms
of the naturally given generators.
To wit, we will denote the generators of In′ −1 (A′ ) as
f12 := ℓ[n]\{1,2} , . . . , f1n := ℓ[n]\{1,n} .
One can move between the two ideals in a simple manner, which is easy to verify:
In−1 (A) : ℓ1 = In′ −1 (A′ ).
A BLOWUP ALGEBRA OF HYPERPLANE ARRANGEMENTS
11
Note that the presentation ideal I(A′ , n′ − 1) of the Rees algebra of In′ −1 (A′ ) with respect to these generators lives in the polynomial subring T ′ := R[y2 , . . . , yn ] ⊂ T := R[y1 , y2 , . . . , yn ]. From Lemma 3.1,
we know that
I1 (A′ , n′ − 1)T = hℓ2 y2 − ℓ3 y3 , ℓ3 y3 − ℓ4 y4 , . . . , ℓn−1 yn−1 − ℓn yn iT ⊂ I1 (A, n − 1).
Likewise, for the Orlik-Terao ideal (which is an ideal in S ′ := k[y2 , . . . , yn ] ⊂ S := k[y1 , y2 , . . . , yn ]), it
obtains via Theorem 2.4:
∂(A′ )S = hI(A′ , n′ − 1)(0,−) iS ⊂ ∂(A) = hI(A, n − 1)(0,−) i.
Lemma 3.6. One has
I(A, n − 1) = hℓ1 y1 − ℓ2 y2 , I(A′ , n′ − 1)i : ℓ∞
1 .
Proof. The inclusion hℓ1 y1 − ℓ2 y2 , I(A′ , n′ − 1)i : ℓ∞
1 ⊂ I(A, n − 1) is clear since we are saturating
a subideal of a prime ideal by an element not belonging to the latter. We note that the codimension of
hℓ1 y1 − ℓ2 y2 , I(A′ , n′ − 1)i exceeds by 1 that of I(A′ , n′ − 1) since the latter is a prime ideal even after
extending to the ambient ring T . Therefore, by a codimension counting it would suffice to show that the
saturation is itself a prime ideal.
Instead, we choose a direct approach. Thus, let F ∈ I(A, n − 1) be (homogeneous) of degree d in
variables y1 , . . . , yn . We can write
F = y1u Gu + y1u−1 Gu−1 + · · · + y1 G1 + G0 , 0 ≤ u ≤ d,
where Gj ∈ k[x1 , . . . , xk ][y2 , . . . , yn ], are homogeneous of degree d − j in y2 , . . . , yn for j = 0, . . . , u.
Evaluating yi = fi , i = 1, . . . , n we obtain
0 = F (f1 , . . . , fn )
u d−u
d
= ℓu2 f12
ℓ1 Gu (f12 , . . . , f1n ) + · · · + ℓ2 f12 ℓd−1
1 G1 (f12 , . . . , f1n ) + ℓ1 G0 (f12 , . . . , f1n ).
This means that
ℓu2 y2u Gu (y2 , . . . , yn ) + · · · + ℓu−1
ℓ2 y2 G1 (y2 , . . . , yn ) + ℓu1 G0 (y2 , . . . , yn ) ∈ I(A′ , n′ − 1).
ℓd−u
1
1
{z
}
|
F′
By writing ℓ1 y1 = ℓ1 y1 − ℓ2 y2 + ℓ2 y2 , it is not difficult to see that
ℓu1 F ≡ F ′ mod hℓ1 y1 − ℓ2 y2 i,
hence the result.
3.4. Stretched arrangements with coefficients. Recall the notion of contraction and the inherent idea of
a multiarrangement, as mentioned in Section 2. Here we wish to consider such multiarrangements, allowing
moreover the repeated individual linear functionals corresponding to repeated hyperplanes to be tagged with
a nonzero element of the ground field. For lack of better terminology, we call such a new gadget a stretched
arrangement with coefficients. Note that, by construction, a stretched arrangement with coefficients B has
a uniquely defined (simple) arrangement A as support. Thus, if A = {ℓ1 , . . . , ℓn } is a simple arrangement,
then a stretched arrangement with coefficients B is of the form
{b1,1 ℓ1 , . . . , b1,m1 ℓ1 , b2,1 ℓ2 , . . . , b2,m2 ℓ2 , . . . , bn,1 ℓn , . . . , bn,mn ℓn },
{z
} |
{z
}
{z
}
|
|
H1 =ker ℓ1
H2 =ker ℓ2
Hn =ker ℓn
where 0 6= bi,j ∈ k and Hi = ker(ℓi ) has multiplicity mi for any 1 ≤ i ≤ n, and for convenience, we
assume that bi,1 = 1. Set m := m1 + · · · + mn . We emphasize the ingredients of a stretched arrangement
by writing B = (A, m).
Proceeding as in the situation of a simple arrangement, we introduce the collection of (m − 1)-products
of elements of B and denote Im−1 (B) the ideal of R generated by them. As in the simple case, we consider
12
MEHDI GARROUSIAN, ARON SIMIS AND ŞTEFAN O. TOHĂNEANU
the presentation ideal I(B, m − 1) of Im−1 (B) with respect to its set of generators consisting of the (m − 1)products. The next lemma relates this ideal to the previously considered presentation ideal I(A, n − 1) of
In−1 (A) obtained by taking the set of generators consisting of the (n − 1)-products of elements of A.
Lemma 3.7. Let A denote an arrangement and let B = (A, m) denote a multiarrangement supported on
A, as above. Let G ∈ R stand for the gcd of the (m − 1)-products of elements of B. Then:
(i) The vector of the (m − 1)-products of elements of B has the form G · PA , where PA denotes
the vector whose coordinates are the (n − 1)-products of the corresponding simple A, each such
product repeated as many times as the stretching in B of the corresponding linear form deleted in
the expression of the product, and further tagged with certain coefficient
(ii) I(B, m − 1) = hI(A, n − 1), DA i, where DA denotes the k-linear dependency relations among
elements of PA .
Proof. (i) This follows from the definition of a stretched arrangement vis-à-vis its support arrangement.
(ii) By (i), the Rees algebra of Im−1 (B) is isomorphic to the Rees algebra of the ideal with generating set
PA . By the nature of the latter, the stated result is now clear.
4. T HE
MAIN THEOREMS
We keep the previous notation as in (3.3), where In−1 (A) is the ideal of (n − 1)-fold products of a
central arrangement A of size n and rank k. We had T := R[y1 , . . . , yn ], with R := k[x1 , . . . , xk ], S :=
k[y1 , . . . , yn ], and I1 (A, n − 1) ⊂ I(A, n − 1) ⊂ T denote, respectively, the presentation ideals of the
symmetric algebra and of the Rees algebra of I. Recall that from Theorem 2.4, the Orlik-Terao ideal ∂(A)
coincides with the defining ideal (I(A, n − 1)(0,−) )S of the special fiber algebra of I.
4.1. The case of a generic arrangement. Simple conceptual proofs can be given in the case where A is
generic (meaning that any k of the defining linear forms are linearly independent), as follows.
Proposition 4.1. If A = {ℓ1 , . . . , ℓn } ⊂ R = k[x1 , . . . , xk ] is a generic arrangement, one has:
(a) I := In−1 (A) is an ideal of fiber type.
(b) The Rees algebra R[It] is Cohen-Macaulay.
(c) The Orlik-Terao ideal of A is the 0-th Fitting ideal of the Jacobian dual matrix of I (i.e., the ideal
generated by the k × k minors of the Jacobian matrix of the generators of I1 (A, n − 1) with respect
to the variables of R).
(d) Let k = n, ie. the case of Boolean arrangement. Under the standard bigrading deg xi = (1, 0) and
deg yj = (0, 1), the bigraded Hilbert series of R[It] is
HS(R[It]; u, v) =
(1 − uv)k−1
.
(1 − u)k (1 − v)k
Proof. As described in the proof of Lemma 3.1, I is a linearly presented codimension 2 perfect ideal with
syzygy matrix of the following shape
ℓ1
−ℓ2
ℓ2
−ℓ3
ϕ=
..
..
.
.
. ℓn−1
−ℓn
The Boolean case n = k is well-known, so we assume that µ(I) = n > k.
We claim that I satisfies the Gk condition. For this purpose we check the requirement in (2). First note
that, for p ≥ n − k + 1, one has
Ip (ϕ) = Ip (A),
A BLOWUP ALGEBRA OF HYPERPLANE ARRANGEMENTS
13
where the rightmost ideal is the ideal generated by all p-fold products of the linear forms defining A, as in
our earlier notation. Because A is generic, it is the support of the codimension (n − p + 1)-star configuration
Vn−p+1 (see [6]). By [6, Proposition 2.9(4)], the defining ideal of Vn−p+1 is a subset of Ip (A), hence
ht(Ip (A)) ≥ n − p + 1. By [18], any minimal prime of Ip (A) can be generated by n − p + 1 elements.
Therefore, ht(Ip (A)) ≤ n − p + 1, and hence equality.
The three statements now follow from [9, Theorem 1.3], where (a) and (c) are collected together by saying
that R[It] has a presentation ideal of the expected type – quite stronger than being of fiber type. Note that,
as a bonus, [9, Theorem 1.3] also gives that ℓ(I) = k and r(I) = k − 1, which are parts (b) and (d) in
Corollary 2.6, when A is generic.
Part (d) follows from an immediate application of [13, Theorem 5.11] to the (k − 1) × k matrix
x1 0 . . .
0
xk
0 x2 . . .
0
xk
M = ..
.
..
..
.
.
.
x
k
0
0
...
xk−1 xk
One can verify that the codimension of It (M ), the ideal of size t minors of M , is k − t + 2. Note that
their setup of [13] is different in that they set deg yj = (n − 1, 1), whereas for us deg yj = (0, 1). To get
our formula we make the substitution in their formula: a ↔ u, and an−1 b ↔ v.
4.2. The fiber type property. In this part we prove one of the main assertions of the section and state a
few structural consequences.
Theorem 4.2. Let A be a central arrangement of rank k ≥ 2 and size n ≥ k. The ideal In−1 (A) of
(n − 1)-fold products of A is of fiber type:
I(A, n − 1) = hI1 (A, n − 1)i + hI(A, n − 1)(0,−) i,
as ideals in T , where hI(A, n − 1)(0,−) iS = ∂(A) is the Orlik-Terao ideal.
Proof. We first consider the case where n = k. Then In−1 (A) is an ideal of linear type by Lemma 3.1, that
is to say, I(A, n − 1) = I1 (A, n − 1). This proves the statement of the theorem since ∂(A) = 0 in this
case.
We now prove the statement by induction on the pairs (n, k), where n > k ≥ 2. In the initial induction
step, we deal with the case k = 2 and arbitrary n > 2 (the argument will even be valid for n = 2). Here one
claims that In−1 (A) = hx1 , x2 in−1 . In fact, since no two forms of the arrangement are proportional, the
generators of In−1 (A) are k-linearly independent – because, e.g., dehomogenizing in one of the variables
yields the first n powers of the other variable up to elementary transformations. Also, since these forms have
degree n − 1, they forcefully span the power hx1 , x2 in−1 .
Now, any hx1 , x2 i-primary ideal in k[x, y] automatically satisfies the property G2 (see (2)). Therefore,
the Rees ideal is of fiber type, and in fact it is of the expected type and Cohen-Macaulay by [9, Theorem
1.3]. In any case, the Rees ideal has long been known in this case, with the defining ideal of the special fiber
generated by the 2-minors of the generic 2 × (n − 1) Hankel matrix, i.e., by the homogeneous defining ideal
of the rational normal curve in Pn−1 (see [4]).
For the main induction step, suppose n > k > 3 and let A′ := A \ {ℓ1 } stand for the deletion of ℓ1 , a
subarrangement of size n′ := n − 1. Applying a change of variables in the base ring R – which, as already
remarked, does not disturb the ideal theoretic properties in sight – we can assume that ℓ1 = x1 and ℓ2 = x2 .
The following extended ideals I(A′ , n′ − 1)T, ∂(A′ )S, I1 (A′ , n′ − 1)T will be of our concern.
The following equalities of ideals of T are easily seen to hold:
I1 (A, n − 1) = hx1 y1 − x2 y2 , I1 (A′ , n′ − 1)i as ideals in T,
∂(A) = h∂(A|x1 ), ∂(A′ )i, as ideals in S.
Let F ∈ I(A, n − 1) be bihomogeneous with degy (F ) = d.
(6)
14
MEHDI GARROUSIAN, ARON SIMIS AND ŞTEFAN O. TOHĂNEANU
Suppose that M = xa1 y1b N ∈ T is a monomial that appears in F , where x1 , y1 ∤ N . If a ≥ b, we can
write
b
M = xa−b
1 (x1 y1 − x2 y2 + x2 y2 ) N,
and hence
b b
M ≡ xa−b
1 x2 y2 N mod hx1 y1 − x2 y2 i.
If a < b, we have
M = (x1 y1 − x2 y2 + x2 y2 )a y1b−a N,
and hence
M ≡ xa2 y2a y1b−a N mod hx1 y1 − x2 y2 i.
Denote R′ := k[x2 , . . . , xk ] ⊂ R, T ′′ := R′ [y2 , . . . , yn ] ⊂ T ′ := R[y2 , . . . , yn ] ⊂ T .
In any case, one can write
m2
mu
1
F = (x1 y1 − x2 y2 )Q + xm
1 P1 + x1 P2 + · · · + x1 Pu + Pu+1 , m1 > · · · > mu ≥ 1,
for certain forms Q ∈ T , P1 , . . . , Pu ∈ T ′′ , and Pu+1 ∈ R′ [y1 , . . . , yn ] = T ′′ [y1 ] of degree d in the variables
y1 , . . . , yn .
Also
Pu+1 = y1v Gv + y1v−1 Gv−1 + · · · + y1 G1 + G0 ,
where Gj ∈ T ′′ and deg(Gj ) = d − j, j = 0, . . . , v.
Let us use the following generators (3.3) for In′ −1 (A′ ):
f12 := ℓ[n]\{1,2} , . . . , f1n := ℓ[n]\{1,n} .
Since evaluating F ∈ I(A, n − 1) at
y1 7→ f1 = x2 ℓ3 · · · ℓn , y2 7→ f2 = x1 f12 , . . . , yn 7→ fn = x1 f1n
vanishes, upon pulling out the appropriate powers of x1 , it yields
1 +d
0 = xm
P1 (f12 , . . . , f1n ) + · · · + x1mu +d Pu (f12 , . . . , f1n )
1
d−1
d
+ f1v xd−v
1 Gv (f12 , . . . , f1n ) + · · · + f1 x1 G1 (f12 , . . . , f1n ) + x1 G0 (f12 , . . . , f1n ).
Suppose first that the rank of A′ is k − 1, i.e. x1 is a coloop. This means that x2 = ℓ2 , ℓ3 , . . . , ℓn are
actually forms in the subring R′ = k[x2 , . . . .xk ]. Since m1 + d > · · · > mu + d > d > · · · > d − v,
Pi (f12 , . . . , f1n ) = 0, i = 1, . . . , u,
Gj (f12 , . . . , f1n ) = 0, j = 0, . . . , v.
Therefore, Pi , Gj ∈ I(A′ , n′ − 1), and hence F ∈ hx1 y1 − x2 y2 , I(A′ , n′ − 1)i. This shows that
I(A, n − 1) = hx1 y1 − x2 y2 , I(A′ , n′ − 1)i
and the required result follows by the inductive hypothesis as applied to I(A′ , n′ − 1).
Suppose now that the rank of A′ does not drop, i.e. x1 is a non-coloop.
Case 1. v = 0. In this case, after canceling xd1 , we obtain
mu
1
0 = xm
1 P1 (f12 , . . . , f1n ) + · · · + x1 Pu (f12 , . . . , f1n ) + G0 (f12 , . . . , f1n ).
Thus,
m2
mu
′ ′
1
xm
1 P1 + x1 P2 + · · · + x1 Pu + Pu+1 ∈ I(A , n − 1).
in the above equation. This will give
Case 2. v ≥ 1. In this case we cancel the factor xd−v
1
x1 |Gv (f12 , . . . , f1n ).
At this point we resort to the idea of stretched arrangements with coefficients as developed in Subsection 3.4. Namely, we take the restriction (contraction) of A to the hyperplane x1 = 0. Precisely, say
ℓi = ai x1 + ℓ̄i , where ℓ¯i ∈ R′ , ai ∈ k,
for i = 2, . . . , n. Note that a2 = 0 since ℓ2 = x2 .
A BLOWUP ALGEBRA OF HYPERPLANE ARRANGEMENTS
15
Write
Ā = {ℓ̄2 , . . . , ℓ̄n } ⊂ R′ .
a stretched arrangement of total multiplicity n̄ = n − 1 with support A′′ of size n′′ ≤ n̄.
Likewise, let
f¯12 := ℓ̄3 · · · ℓ̄n , . . . , f¯1n := ℓ̄2 · · · ℓ̄n−1
denote the (n̄ − 1)-products of this stretched arrangement. Then, Gv vanishes on the tuple (f¯12 , . . . , f¯1n )
and since its is homogeneous it necessarily belong to I(Ā, n̄ − 1).
From Lemma 3.7, we have
I(Ā, n̄ − 1) = hI(A′′ , n′′ − 1), DĀ i,
where DĀ is a linear ideal of the form hyi − bi,j yj i2≤i,j≤n .
Let us analyze the generators of I(Ā, n̄ − 1).
• A generator yi − bi,j yj , i, j ≥ 2 of DĀ comes from the relation ℓ̄j = bi,j ℓ̄i , bi,j ∈ k. Thus, back in A
we have the minimal dependency
ℓj − aj x1 = bi,j (ℓi − ai x1 ),
yielding an element of ∂(A|ℓ1 ):
y1 (yi − bi,j yj ) + (bi,j ai − aj ) yi yj .
{z
}
|
ci,j
• Since gcd(ℓ̄i , ℓ̄j ) = 1, for 2 ≤ i < j ≤ n′′ + 1, a typical generator of I1 (A′′ , n′′ − 1) is ℓ̄i yi − ℓ̄j yj ,
that we will rewrite as
ℓ̄i yi − ℓ̄j yj = (ℓi yi − ℓj yj ) − x1 (ai yi − aj yj ).
• A typical generator of ∂(A′′ ) is of the form b1 yi2 · · · yis + · · · + bs yi1 · · · yis−1 coming from a minimal
dependency
b1 ℓ̄i1 + · · · + bs ℓ̄is = 0, ij ∈ {2, . . . , n′′ + 1}.
Since ℓ̄ij = ℓij − aij x1 , we obtain a dependency
b1 ℓi1 + · · · + bs ℓis − (b1 ai1 + · · · + bs ais )x1 = 0.
|
{z
}
α
If α = 0, then
b1 yi2 · · · yis + · · · + bs yi1 · · · yis−1 ∈ ∂(A′ ),
whereas if α 6= 0, then
−αyi2 · · · yis + y1 (b1 yi2 · · · yis + · · · + bs yi1 · · · yis−1 ) ∈ ∂(A|ℓ1 ).
We have that
X
X
X
Gv =
Es,t (ys − bs,t yt ) +
Ai,j (ℓ̄i yi − ℓ̄j yj ) +
Bi1 ,...,is (b1 yi2 · · · yis + · · · + bs yi1 · · · yis−1 ),
where Es,t , Ai,j , Bi1 ,...,is ∈ T ′′ and s, t, i, j, ik ≥ 2. Then, by using the expressions in the three bullets
above and splicing according to the equality x1 y1 = (x1 y1 − x2 y2 ) + x2 y2 , we get:
16
MEHDI GARROUSIAN, ARON SIMIS AND ŞTEFAN O. TOHĂNEANU
X
X
y1v Gv = y1v−1 (
Es,t (y1 (ys − bs,t yt ) + cs,t ys yt ) −
Es,t cs,t ys yt
|
{z
} |
{z
}
∈T ′′
∈∂(A|ℓ1 )
+
+
X
|
X
|
X
Ai,j y1 (ℓi yi − ℓj yj ) −
{z
} |
∈I(A′ ,n′ −1)
X
Ai,j (x1 y1 − x2 y2 )(ai yi − aj yj ) −
Ai,j x2 y2 (ai yi − aj yj )
{z
} |
{z
}
∈T ′′
∈hx1 y1 −x2 y2 i
X
Bi1 ,...,is (y1 (b1 yi2 · · · yis + · · · + bs yi1 · · · yis−1 ) − αyi2 · · · yis ) +
{z
} |
∈∂(A|ℓ1 )
Thus, y1v Gv = y1v−1 G′v−1 + W , where
G′v−1 ∈ T ′′ ,
Bi1 ,...,is αyi2 · · · yis ).
{z
}
∈T ′′
W ∈ hx1 y1 − x2 y2 , ∂(A|ℓ1 ), I(A′ , n′ − 1)i.
Then returning to our original F , it obtains
mu
v−1
v−2
′
1
F = ∆ + xm
1 P1 + · · · + x1 Pu + y1 (Gv−1 + Gv−1 ) + y1 Gv−2 + · · · + G0 ,
|
{z
}
′′
G′′
v−1 ∈S
where ∆ ∈ hx1 y1 − x2 y2 , ∂(A|ℓ1 ), I(A′ , n′ − 1)i ⊂ I(A, n − 1).
The key is that modulo the ideal hx1 y1 − x2 y2 , ∂(A|ℓ1 ), I(A′ , n′ − 1)i the power of y1 dropped from v
to v − 1 in the expression of F . Iterating, with F (f1 , . . . , fn ) = 0 = ∆(f1 , . . . , fn ), will eventually drop
further the power of y1 to v − 2. Recursively we end up with v = 0, which is Case 1 above.
This way, we eventually get
I(A, n − 1) = hx1 y1 − x2 y2 , ∂(A|ℓ1 ), I(A′ , n′ − 1)i.
By the inductive hypothesis as applied to I(A′ , n′ − 1) and from the two equalities in (6), one gets the
stated result.
Corollary 4.3. I(A′ , n′ − 1) = I(A, n − 1) ∩ T ′ as ideals in T ′ = R[y2 , . . . , yn ].
Proof. Recall the notation T ′ := R[y2 , . . . , yn ] ⊂ T := R[y1 , . . . , yn ] = T ′ [y1 ] as in the proof of the
previous theorem. Denote J := I(A, n − 1) ∩ T ′ . We show that J ⊆ I(A′ , n′ − 1), the other inclusion
being obvious.
Let F ∈ J. Then F ∈ T ′ and F ∈ I(A, n − 1). By Theorem 4.2, we can write
F = (ℓ1 y1 − ℓ2 y2 )P + Q + G, where P ∈ T, Q ∈ ∂(A|ℓ1 )T, G ∈ I(A′ , n′ − 1)T.
By Lemma 3.4,
ℓ1 Q ∈ I1 (A, n − 1) = hℓ1 y1 − ℓ2 y2 , I1 (A′ , n′ − 1)i.
Therefore,
ℓ1 F = (ℓ1 y1 − ℓ2 y2 )P ′ + G′ ,
(7)
′
′
′
′
for suitable P ∈ T, G ∈ I(A , n − 1)T .
We write P ′ = y1u Pu + · · · + y1 P1 + P0 , Pi ∈ T ′ , and G′ = y1v Gv + · · · + y1 G1 + G0 , Gj ∈ T ′ .
Since G′ ∈ I(A′ , n′ − 1) ⊂ T ′ , setting y1 = 0 in the expression of G′ gives that G0 ∈ I(A′ , n′ − 1).
Therefore, G−G0 = y1 (y1v−1 Gv +· · ·+G1 ) ∈ I(A′ , n′ −1), and hence y1v−1 Gv +· · ·+G1 ∈ I(A′ , n′ −1)
since I(A′ , n′ − 1) is prime. Setting again y1 = 0 in this expression we obtain that G1 ∈ I(A′ , n′ − 1), and
so on, eventually obtaining
Gj ∈ I(A′ , n′ − 1), j = 0, . . . , v.
Suppose u ≥ v. Then, by grouping the powers of y1 we have
ℓ1 F
= (−ℓ2 y2 P0 + G0 ) + (ℓ1 P0 − ℓ2 y2 P1 + G1 )y1 + · · · + (ℓ1 Pv−1 − ℓ2 y2 Pv + Gv )y1v
+ (ℓ1 Pv − ℓ2 y2 Pv+1 )y1v+1 + · · · + (ℓ1 Pu−1 − ℓ2 y2 Pu )y1u + ℓ1 Pu y1u+1 .
A BLOWUP ALGEBRA OF HYPERPLANE ARRANGEMENTS
17
Since F ∈ T ′ , then ℓ1 F ∈ T ′ . Thus, the “coefficients” of y1 , y12 , . . . , y1u+1 must vanish. It follows that
Pu = · · · = Pv = 0 and ℓ1 Pv−1 = −Gv ∈ I(A′ , n′ − 1).
Since I(A′ , n′ − 1) is a prime ideal, we have Pv−1 ∈ I(A′ , n′ − 1), and therefore
ℓ1 Pv−2 = ℓ2 y2 Pv−1 − Gv−1 ∈ I(A′ , n′ − 1).
Recursively we get that
Pv−1 , Pv−2 , . . . , P1 , P0 ∈ I(A′ , n′ − 1).
If u < v, a similar analysis will give the same conclusion that P ′ ∈ I(A′ , n′ − 1)T .
Therefore, (7) gives ℓ1 F ∈ I(A′ , n′ − 1)T, and hence F ∈ I(A′ , n′ − 1)T by primality of the extended
ideal. But then F ∈ I(A′ , n′ − 1)T ∩ T ′ = I(A′ , n′ − 1), as required.
The next two corollaries help compute the Rees ideal from the symmetric ideal via a simple colon of
ideals.
Corollary 4.4. Let ℓi ∈ A and yi be the corresponding external variable. Then
I(A, n − 1) = I1 (A, n − 1) : ℓi yi .
Proof. Without loss of generality, assume i = 1.
The inclusion ⊇ is immediate, since I1 (A, n − 1) ⊂ I(A, n − 1), and the Rees ideal I(A, n − 1) is a
prime ideal not containing ℓ1 nor y1 .
Now we show the inclusion ⊆. Let F ∈ I(A, n − 1). Then, from Theorem 4.2,
X
PD ∂D,
F =G+
D
where the sum is taken over all minimal dependencies D, and G ∈ I1 (A, n − 1).
Obviously, ℓ1 y1 G ∈ I1 (A, n − 1). Also, if ∂D ∈ ∂(A|ℓ1 ), then, from Lemma 3.4, ℓ1 ∂D, hence ℓ1 y1 ∂D
belongs to I1 (A, n − 1).
Suppose ∂D ∈
/ ∂(A|ℓ1 ). Since D is a minimal dependency among the hyperplanes of A, there exists
j ∈ {1, . . . , n} such that ∂D ∈ ∂(A|ℓj ). Thus, ℓ1 y1 ∂D = (ℓ1 y1 − ℓj yj )∂D + ℓj yj ∂D belongs to the
ideal I1 (A, n − 1) since each summand belongs to I1 (A, n − 1) – the first trivially and the second due to
Lemma 3.4.
Since the rank of A is k, after a reordering of the linear forms ℓ1 , . . . , ℓn that define A, we can assume
that the last k linear forms ℓn−k+1 , . . . , ℓn are linearly independent. With this proviso, one has:
Qn−k
Corollary 4.5. I(A, n − 1) = I1 (A, n − 1) : i=1
ℓi .
Proof. Since ℓn−k+1 , . . . , ℓn are k linearly independent linear forms, any minimal dependency that involves
at least one of them, must involve also a linear form ℓj , where j ∈ {1, . . . , n − k}. So
∂(A) = ∂(A|ℓ1 ) + · · · + ∂(A|ℓn−k ).
Qn−k
We obviously have I1 (A, n − 1) ⊆ I1 (A, n − 1) : i=1
ℓi , and from Lemma 3.4,
∂(A|ℓj ) ⊂ I1 (A, n − 1) :
n−k
Y
ℓi , for all j = 1, . . . , n − k.
i=1
Then, from Theorem 4.2, one has
I(A, n − 1) ⊆ I1 (A, n − 1) :
n−k
Y
ℓi .
i=1
The reverse inclusion comes from the fact that I1 (A, n − 1) ⊆ I(A, n − 1), and from I(A, n − 1) being
a prime ideal with ℓi ∈
/ I(A, n − 1).
18
MEHDI GARROUSIAN, ARON SIMIS AND ŞTEFAN O. TOHĂNEANU
In the next statement we denote the extended ideal (I1 (A′ , n′ − 1))T by hI1 (A′ , n′ − 1)i.
Lemma 4.6. Let A′ = A \ {ℓ1 } and n′ = |A′ | = n − 1. We have
hI1 (A′ , n′ − 1)i : (ℓ1 y1 − ℓ2 y2 ) = hI1 (A′ , n′ − 1)i : ℓ1 .
In particular, when ℓ1 is a coloop, the biform ℓ1 y1 − ℓ2 y2 is a nonzerodivisor on hI1 (A′ , n′ − 1)i.
Proof. For convenience, let us change coordinates to have ℓ1 = x1 and ℓ2 = x2 . Let f ∈ hI1 (A′ , n′ − 1)i :
(x1 y1 − x2 y2 ). Then f (x1 y1 − x2 y2 ) ∈ hI1 (A′ , n′ − 1)i ⊂ hI(A′ , n′ − 1)i. Since hI(A′ , n′ − 1)i is a
prime ideal not containing x1 y1 − x2 y2 , we obtain f ∈ hI(A′ , n′ − 1)i, and by Theorem 4.2, we have
f = g + h,
g ∈ hI1 (A′ , n′ − 1)i,
h ∈ h∂(A′ )i.
By multiplying this by x1 y1 − x2 y2 , we get that
(x1 y1 − x2 y2 )h ∈ hI1 (A′ , n′ − 1)i.
By Corollary 4.4, since h ∈ h∂(A′ )i ⊂ hI(A′ , n′ − 1)i, and x2 ∈ A′ , we have x2 y2 h ∈ hI1 (A′ , n′ − 1)i. So
h ∈ hI1 (A′ , n′ −1)i : x1 y1 , and together with f = g+h with g ∈ hI1 (A′ , n′ −1)i ⊂ hI1 (A′ , n′ −1)i : x1 y1 ,
gives
f ∈ hI1 (A′ , n′ − 1)i : x1 y1 .
Conversely, let ∆ ∈ hI1 (A′ , n′ − 1)i : x1 y1 . Then x1 y1 ∆ ∈ hI1 (A′ , n′ − 1)i ⊆ hI(A′ , n′ − 1)i. The
ideal hI(A′ , n′ − 1)i is a prime ideal, and x1 y1 ∈
/ hI(A′ , n′ − 1)i, so ∆ ∈ hI(A′ , n′ − 1)i. So, by Corollary
′
′
4.4, x2 y2 ∆ ∈ hI1 (A , n − 1)i. Therefore
(x1 y1 − x2 y2 )∆ = x1 y1 ∆ − x2 y2 ∆ ∈ hI1 (A′ , n′ − 1)i.
Thus far, we have shown that hI1 (A′ , n′ − 1)i : (ℓ1 y1 − ℓ2 y2 ) = hI1 (A′ , n′ − 1)i : x1 y1 . Clearly, the right
hand side is the same as hI1 (A′ , n′ − 1)i : x1 since y1 is a nonzero divisor on hI1 (A′ , n′ − 1)i.
4.3. The Cohen-Macaulay property. In this part the goal is to prove that the Rees algebra is CohenMacaulay. Since we are in a graded setting, this is equivalent to showing that its depth with respect to the
maximal graded ideal hm, y1 , . . . , yn i is (at least) k + 1 = dim R[It].
This will be accomplished by looking at a suitable short exact sequence, where two of the modules
will be examined next. We state the results in terms of depth since this notion is inherent to the CohenMacaulay property, yet the proofs will take the approach via projective (i.e., homological) dimension. By
the Auslander-Buchsbaum equality, we are home anyway.
Throughout, pdimT will denote projective dimension over the polynomial ring T . Since we are in a
graded situation, this is the same as the projective dimension over the local ring T(m,y1 ,...,yn ) , so we may
harmlessly proceed.
The first module is obtained by cutting the binomial generators of I1 (A, n − 1) into its individual terms.
The result may have interest on itself.
Lemma 4.7. Let ℓ1 , . . . , ℓn ∈ k[x1 , . . . , xk ] be linear forms, allowing some of them to be mutually proportional. Let T := k[x1 , . . . , xk ; y1 , . . . , yn ]. Then
T
≥ k.
depth
hℓ1 y1 , ℓ2 y2 , . . . , ℓn yn i
Proof. If k = 1, the claim is clearly satisfied, since hx1 y1 , . . . , x1 yn i = x1 hy1 , . . . , yn i, and {y1 , . . . , yn }
is a T -regular sequence. Assume k ≥ 2.
We will use induction on n ≥ 1 to show that the projective dimension is at most n + k − k = n.
If n = 1, the ideal hℓ1 y1 i is a principal ideal, hence the claim is true. Suppose n > 1. We may apply a klinear automorphism on the ground variables, which will not disturb the the projective dimension. Thus, say,
ℓ1 = x1 and this form is repeated s times. Since nonzero coefficients from k tagged to x1 will not change
the ideal in question, we assume that ℓi = x1 , for 1 ≤ i ≤ s, and gcd(x1 , ℓj ) = 1, for s + 1 ≤ j ≤ n. Write
ℓj = cj x1 + ℓ̄j , for s + 1 ≤ j ≤ n, with cj ∈ k, and 0 6= ℓ̄j ∈ k[x2 , . . . , xk ].
A BLOWUP ALGEBRA OF HYPERPLANE ARRANGEMENTS
19
Denoting J := hx1 y1 , . . . , x1 ys , ℓs+1 ys+1 , . . . , ℓn yn i, we claim that
J : x1 = hy1 , . . . , ys , ℓs+1 ys+1 , . . . , ℓn yn i.
(8)
This is certainly the expression of a more general result, but we give a direct proof here.
One inclusion is obvious. For the reverse inclusion, let F ∈ hx1 y1 , . . . , x1 ys , ℓs+1 ys+1 , . . . , ℓn yn i : x1 .
Then, say,
n
s
X
X
Pj ℓj yj ,
Pi yi +
x1 F = x1
i=1
j=s+1
for certain Pi , Pj ∈ T . Rearranging we have
x1 (F −
s
X
n
X
Pi yi −
j=s+1
i=1
cj Pj yj ) =
n
X
Pj ℓ¯j yj .
(9)
j=s+1
Since x1 is a nonzero divisor in T /hℓ̄s+1 ys+1 , . . . , ℓ̄n yn i, the second factor of the left hand side in (9)
must be of the form
n
X
Qj ℓ¯j yj , Qj ∈ T.
j=s+1
P
P
Substituting in (9) we find Pj = x1 Qj , s + 1 ≤ j ≤ n, and hence F = si=1 Pi yi + nj=s+1 Qj ℓj yj , as
claimed.
Computing projective dimensions with respect to T and T ′ = k[x1 , . . . , ys+1 , . . . , yn ] and applying the
inductive hypothesis, one has
T
T′
pdimT
= s + pdimT ′
≤ s + (n − s) = n.
J : x1
hℓs+1 ys+1 , . . . , ℓn yn i
At the other end, we have hx1 , Ji = hx1 , ℓ̄s+1 ys+1 , . . . , ℓ̄n yn i. Applying the inductive hypothesis this
time around gives
T
≤ 1 + (n − s) ≤ n.
pdimT
hx1 , Ji
From the short exact sequence of T -modules
·x
1
0 −→ T /(J : x1 ) −→
T /J −→ T /hx1 , Ji −→ 0,
knowingly the projective dimension of the middle term does not exceed the maximum of the projective
dimensions of the two extreme terms. Therefore, pdimT (T /J) ≤ n, as was to be shown..
The difficult result of this section is the following exact invariant of the symmetric algebra S(I) ≃
T /hI1 (A, n − 1)i:
Proposition 4.8. Let I = In−1 (A) as before. Then depth(S(I)) = k + 1.
Proof. By Proposition 3.2 (i), it suffices to prove the lower bound depth(S(I)) ≥ k + 1.
As in the proof of Theorem 4.2, we argue by induction on all pairs n, k, with n ≥ k ≥ 2, where n and k
are, respectively, the size and the rank of A.
If k = 2 and n > 2, let R = k[x, y]. As seen in that proof, one has I = hx, yin−1 , and hence
I1 (A, n − 1) = hxy1 − yy2 , xy2 − yy3 , . . . , xyn−1 − yyn i.
A direct calculation shows that {y1 , x + yn , y + yn−1 } is a regular sequence modulo I1 (A, n − 1).
If n = k, the ideal I1 (A, n − 1) is a complete intersection by Lemma 3.1.
Thus, for the main inductive step suppose n > k > 3.
We will equivalently show that pdimT (S(I)) ≤ n − 1. First apply a change of ground variables so as to
have ℓ1 = x1 and ℓ2 = x2 .
20
MEHDI GARROUSIAN, ARON SIMIS AND ŞTEFAN O. TOHĂNEANU
Let A′ := A \ {x1 } denote deletion. Since I1 (A, n − 1) = hI1 (A′ , n′ − 1), x1 y1 − x2 y2 i, we have the
following short exact sequence of T -modules
T
T
T
·(x1 y1 −x2 y2 )
0→
−→
→ 0.
(10)
−−−−−−−−→
′
′
′
′
hI1 (A , n − 1)i : (x1 y1 − x2 y2 )
hI1 (A , n − 1)i
I1 (A, n − 1)
We consider separately the cases where ℓ1 is a coloop or a non-coloop.
x1 IS A COLOOP.
Here the rank of A′ is k − 1 and x1 is altogether absent in the linear forms of the deletion. Thus, the
natural ambient ring of I1 (A′ , n′ − 1) is T ′ := k[x2 , . . . , xk ; y2 , . . . , yn ].
In this case, by Lemma 4.6, the left most nonzero term of (10) becomes
T /hI1 (A′ , n′ − 1)i =
T′
[x1 , y1 ],
I1 (A′ , n′ − 1)
hence
pdimT (T /hI1 (A′ , n′ − 1)i) = pdimT ′ (T ′ /I1 (A′ , n′ − 1)) ≤ n′ − 1,
by the inductive hypothesis applied to S(In′ −1 (A′ )) ≃ T ′ /I1 (A′ , n′ − 1).
Then, from (10) we have
pdimT (T /I1 (A, n − 1)) ≤ max{pdimT (T /hI1 (A′ , n′ − 1)i) + 1, pdimT (T /hI1 (A′ , n′ − 1)i)}
≤ (n′ − 1) + 1 = n′ = n − 1.
x1 IS A NON - COLOOP.
This case will occupy us for the rest of the proof. Here T ′ := k[x1 , . . . , xk ; y2 , . . . , yn ] is the natural
ambient ring of the deletion symmetric ideal.
By Lemma 4.6, the left most nonzero term of (10) is T /(hI1 (A′ , n′ − 1)i : x1 ). Thus, multiplication by
x1 gives a similar exact sequence to (10):
T
T
T
−→
−→ 0.
(11)
−→
0 −→
′
′
′
′
hI1 (A , n − 1)i : x1
hI1 (A , n − 1)i
hx1 , I1 (A′ , n′ − 1)i
Suppose for a minute that one has
pdimT
Then (11) implies
pdimT
T
hx1 , I1 (A′ , n′ − 1)i
T
I1 (A′ , n′ − 1) : x1
≤ n′ .
(12)
≤ max{n′ − 1, n′ − 1} = n′ − 1.
Back to (10) would finally give
T
pdimT
≤ max{(n′ − 1) + 1, n′ − 1} = n′ = n − 1,
I1 (A, n − 1)
proving the required statement of the theorem.
Thus, it suffices to prove (12).
For this, one sets hx1 , I1 (A′ , n′ − 1)i = hx1 , x2 y2 − ℓ̄3 y3 , . . . , ℓ̄n−1 yn−1 − ℓ̄n yn i, where we have written
ℓj = cj x1 + ℓ̄j , with cj ∈ k, ℓ̄j ∈ k[x2 , . . . , xk ], for 3 ≤ j ≤ n. Then,
T′
T
= 1 + pdimT ′
.
(13)
pdimT
hx1 , I1 (A′ , n′ − 1)i
hx2 y2 − ℓ̄3 y3 , . . . , ℓ̄n−1 yn−1 − ℓ̄n yn i
Let Ā = {x2 , ℓ̄3 , . . . , ℓ̄n } denote the corresponding stretched arrangement and set
J := hx2 y2 − ℓ̄3 y3 , . . . , ℓ̄n−1 yn−1 − ℓ̄n yn i ⊂ T ′ := k[x2 , . . . , xk ; y2 , . . . , yn ].
C LAIM . pdimT ′ (T ′ /J) ≤ n′ − 1.
A BLOWUP ALGEBRA OF HYPERPLANE ARRANGEMENTS
21
If the size of Ā is = n − 1 = n′ (i.e., no two linear forms of Ā are proportional), then J = I1 (Ā, n′ − 1),
and by the inductive hypotheses pdimT ′ (T ′ /I1 (Ā, n′ − 1)) ≤ n′ − 1.
Otherwise, suppose s − 1 ≥ 2 of the linear forms of Ā are mutually proportional. Without loss of
generality, say
ℓ̄3 = d3 x2 , . . . , ℓ̄s = ds x2 , di ∈ k \ {0}, 3 ≤ i ≤ s
and
ℓ̄j = dj x2 + Lj , dj ∈ k, 0 6= Lj ∈ k[x3 , . . . , xk ], 4 ≤ j ≤ n.
Then
J = hx2 (y2 − d3 y3 ), . . . , x2 (y2 − ds ys ), x2 y2 − ℓ̄s+1 ys+1 , . . . , x2 y2 − ℓ̄n yn i.
We now provide the following estimates:
(a) pdimT ′ (T ′ /hx2 , Ji) ≤ 1 + (n − s).
(b) pdimT ′ (T ′ /hJ : x2 i) ≤ n′ − 1.
For (a), note that hx2 , Ji = hx2 , Ls+1 ys+1 , . . . , Ln yn i, while from the proof of Lemma 4.7 we have
pdimT ′ (T ′ /hx2 , Ji) ≤ 1 + (n − s),
since x2 is a nonzero divisor in T ′ /hLs+1 ys+1 , . . . , Ln yn i.
As for (b), we first claim that J : x2 = hy2 − d3 y3 , . . . , y2 − ds ys , x2 y2 − ℓ̄s+1 ys+1 , . . . , x2 y2 − ℓ̄n yn i.
The proof is pretty much the same as that of the equality in (8), hence will be omitted.
This equality implies that
T′
T ′′
pdimT ′
,
= s − 2 + pdimT ′′
J : x2
hx2 y2 − ℓ̄s+1 ys+1 , . . . , x2 y2 − ℓ̄n yn i
where T ′′ := k[x2 , . . . , xk ; y2 , ys+1 , . . . , yn ].
Let B := {x2 , ℓ̄s+1 , . . . , ℓ̄n }. With same reasoning for B as for Ā (i.e., removing proportional linear
forms), we obtain
T ′′
≤ (n − s + 1) − 1 = n − s,
pdimT ′′
hx2 y2 − ℓ̄s+1 ys+1 , . . . , x2 y2 − ℓ̄n yn i
and therefore
T′
≤ s − 2 + n − s = n − 2 = n′ − 1.
pdimT ′
J : x2
Drawing on the estimates (a) and (b) above, the exact sequence of T ′ -modules
0 −→ T ′ /(J : x2 ) −→ T ′ /J −→ T ′ /hx2 , Ji −→ 0,
gives that
pdimT ′ (T ′ /J) ≤ max{n′ − 1, 2 + n′ − s} ≤ n′ − 1,
since s ≥ 3.
Rolling all the way back to (13), we have proved that
T
pdimT
≤ 1 + (n′ − 1) = n′ ,
hx1 , I1 (A′ , n′ − 1)i
as required.
The main result now follows quite smoothly.
22
MEHDI GARROUSIAN, ARON SIMIS AND ŞTEFAN O. TOHĂNEANU
Theorem 4.9. The Rees algebra of In−1 (A) is Cohen-Macaulay.
Proof. From Corollary4.4 we have the following short exact sequence of T -modules
T
T
T
0 −→
−→
−→
−→ 0.
I(A, n − 1)
I1 (A, n − 1)
hI1 (A, n − 1), ℓ1 y1 i
By Proposition 4.8, the depth of the middle module is k + 1, while by Lemma 4.7 that of the right most
module is at least k – in fact, by the domino effect one has hI1 (A, n − 1), ℓ1 y1 i = hℓ1 y1 , ℓ2 y2 , . . . , ℓn yn i.
By standard knowledge, the depth of the left most module is at least that of the middle module, namely,
k + 1.
A consequence is an alternative proof of a result of Proudfoot-Speyer ([12]):
Corollary 4.10. Let A be any central arrangement. Then the associated Orlik-Terao algebra is CohenMacaulay.
Proof. As we have seen, the Orlik-Terao algebra is the special fiber of the ideal In−1 (A). Since the latter is
a homogeneous ideal generated in one single degree, then the Cohen-Macaulay property of the Rees algebra
transfers to its direct summand, the special fiber – this result has been communicated by W. Vasconcelos
and the second author as part of an ongoing joint work.
R EFERENCES
[1] B. Anzis, M. Garrousian and Ş. Tohǎneanu, Generalized star configurations and the Tutte polynomial, arXiv preprint
arXiv:1604.01311 (2016).
[2] M. Artin and M. Nagata, Residual intersections in Cohen-Macaulay rings, J. Math. Kyoto Univ. 12 (1972), 307–323.
[3] A. Berget, Products of linear forms and Tutte polynomials, European J. Combin. 31 (2010), 1924–1935.
[4] P. Corsini, Sulle potenze di un ideale generato da una A-successione, Rend. Sem. Mat. Univ. Padova 39 (1967), 190–204.
[5] A.V. Doria, S.H. Hassanzadeh and A. Simis, A characteristic-free criterion of birationality, Adv. in Math. 230 (2012), 390–
413.
[6] A.V. Geramita, B. Harbourne and J. Migliore, Star configurations in Pn , J. Algebra 376 (2013), 279–299.
[7] J. Herzog, A. Simis, W.V. Vasconcelos, Koszul homology and blowing-up rings, in: Commutative Algebra, Proceedings,
Trento, 1981, in: Lect. Notes Pure Appl. Math., vol.84, Marcel Dekker, 1983, pp.79-169.
[8] J. Hong, A. Simis and W. V. Vasconcelos, The equations of almost complete intersections, Bull. Braz. Math. Soc. (New Series)
43 (2012), 171–199.
[9] S. Morey and B. Ulrich, Rees algebras of ideals of low codimension, Proc. Amer. Math. Soc. 124 (1996), 3653–3661.
[10] P. Orlik and H. Terao, Arrangements of Hyperplanes, Grundlehren Math. Wiss., Bd. 300, Springer-Verlag, Berlin-HeidelbergNew York, 1992.
[11] P. Orlik and H. Terao, Commutative algebras for arrangements, Nagoya Math. J. 134 (1994), 65–73.
[12] N. Proudfoot and D. Speyer, A broken circuit ring, Beiträge Algebra Geom. 47 (2006), 161–166.
[13] L. Robbiano and G. Valla, Hilbert-Poincaré series of bigraded algebras, Bollettino U.M.I. (8) 1-B (1998), 521–540.
[14] H. Schenck, Resonance varieties via blowups of P2 and scrolls, Int. Math. Res. Not. 20 (2011), 4756–4778.
[15] H. Schenck and Ş. Tohǎneanu, The Orlik-Terao algebra and 2-formality, Math. Res. Lett. 16 (2009), 171–182.
[16] A. Simis, B. Ulrich and W. Vasconcelos, Jacobian dual fibrations, Amer. J. Math. 115 (1993), 47–75.
[17] H. Terao, Algebras generated by reciprocals of linear forms, J. Algebra 250 (2002), 549–558.
[18] Ş. Tohǎneanu, On the De Boer-Pellikaan method for computing minimum distance, J. Symbolic Comput. 45(2010), 965–974.
[19] W. Vasconcelos, Integral closure, Springer Monographs in Mathematics, Springer-Verlag, Berlin, 2005, Rees algebras, multiplicities, algorithms.
| 0math.AC
|
SIMULATING PATHO-REALISTIC ULTRASOUND IMAGES USING DEEP GENERATIVE
NETWORKS WITH ADVERSARIAL LEARNING
Francis Tom, Debdoot Sheet
Indian Institute of Technology Kharagpur, India
arXiv:1712.07881v2 [cs.CV] 8 Jan 2018
ABSTRACT
Ultrasound imaging makes use of backscattering of waves
during their interaction with scatterers present in biological tissues. Simulation of synthetic ultrasound images is a
challenging problem on account of inability to accurately
model various factors of which some include intra-/inter
scanline interference, transducer to surface coupling, artifacts
on transducer elements, inhomogeneous shadowing and nonlinear attenuation. Current approaches typically solve wave
space equations making them computationally expensive and
slow to operate. We propose a generative adversarial network
(GAN) inspired approach for fast simulation of patho-realistic
ultrasound images. We apply the framework to intravascular
ultrasound (IVUS) simulation. A stage 0 simulation performed using pseudo B-mode ultrasound image simulator
yields speckle mapping of a digitally defined phantom. The
stage I GAN subsequently refines them to preserve tissue
specific speckle intensities. The stage II GAN further refines
them to generate high resolution images with patho-realistic
speckle profiles. We evaluate patho-realism of simulated
images with a visual Turing test indicating an equivocal
confusion in discriminating simulated from real. We also
quantify the shift in tissue specific intensity distributions of
the real and simulated images to prove their similarity.
Fig. 1: Overview of the proposed framework for ultrasound simulation using stacked generative adversarial networks (GAN).
Index Terms— Adversarial learning, deep convolutional
neural network, generative adversarial network, intravascular
ultrasound, ultrasound simulation.
nator. In an adversarial learning strategy, the generator learns
while attempting to generate realistic ultrasound images. The
discriminator simultaneously learns while attempting to discriminate between true images and those simulated by the
generator. This learning rule manifests as a two player minimax game. Our formulation is elucidated in Sec. 3.
The rest of the paper has Sec. 2 describing the prior art
of ultrasound image simulation. Sec. 4 presents the results
of our experiments on the publicly available border detection in IVUS challenge1 dataset along with discussion. Sec. 5
presents the conclusions of the work.
1. INTRODUCTION
2. PRIOR ART
Ultrasound images are formed by reflection of ultrasonic
acoustic waves from body structures. This makes the task
of simulating realistic appearing medical ultrasound images
computationally complex [1]. Such simulators could be used
as a learning aid for doctors to be able to simulate and visualize co-morbid pathologies which occur rarely in reality.
They can also simulate images to augmented datasets for
supervised learning algorithms. Fig. 1 summarizes the contribution of this paper where we present a generative adversarial network (GAN) based framework for fast and realistic
simulation of ultrasound images. The GAN framework [2]
consists of two neural networks: a generator and a discrimi-
The different approaches for simulating ultrasound images
can broadly be grouped as employing (i) numerical modeling
of ultrasonic wave propagation or (ii) image convolution operation based. An early approach employed convolutional operators leveraging the linearity, separability and space invariance properties of the point spread function (PSF) of imaging
system to simulate pseudo B-mode images [3]. Numerical
modeling based approaches have employed ray-based model
to simulate ultrasound from volumetric CT scans [4], and with
the Born approximation [5] have also simulated images using
1 http://www.cvc.uab.es/IVUSchallenge2011/dataset.html
histopathology based digital phantoms employing wave equations [6, 7, 8]. Further Rayleigh scattering model [9], physical reflection, transmission and absorption have also been included in simulation [10]. Commercially available solutions
like Field II simulator2 have used a finite element model for
the purpose. These simulators that model the tissue as a collection of point scatterers fail to model appearance of artifacts, and the high computational complexity challenge their
deployment. Recent approaches employ spatially conditioned
GAN [11], yet are limited in resolution and the appearance of
unrealistic artefacts in the generated images.
3. METHODOLOGY
Our framework employs (1) simulation of ultrasound images
from the tissue echogenicity maps using a physics based simulator, (2) generation of low resolution images from the earlier simulated images by learning Stage I GAN, (3) generation
of high resolution patho-realistic images from the low resolution earlier stage by a Stage II GAN. Stage I GAN does the
structural and intensity mapping of the speckle maps generated by the physics based simulator. Stage II GAN models
patho-realistic speckle PSF and also corrects any defects in
the images simulated by the stage I GAN. Generation of high
resolution images by GANs is difficult due to training instability [12]. Hence the complex task of ultrasound simulation
is decomposed into sub-problems.
3.1. Stage 0 simulation from tissue echogenicity map
The ground truth label of lumen and external elastic lamina
boundary contour available in the dataset is used to generate
the tissue echogenicity map (Fig. 2b). The initial simulation
is performed by a pseudo B-mode ultrasound image simulator3 [3, 13] that assumes a linear and space invariant point
spread function (PSF). Image is formed with assumption of
wave propagating vertically along the echogenicity map. The
images are generated in the polar domain (Fig. 2c).
3.2. Stage I GAN
Refinement of the output of the pseudo B-mode simulator is
performed by a two stage stacked GAN [14]. The Stage I
GAN is trained to generate a low resolution image of 64 × 64
px. size in the polar domain using the simulated + unsupervised learning approach [15]. This involves learning of a refiner network GI (x) to map the speckle map generated by the
pseudo B-mode simulator x, with θ as the function parameters. The discriminator network (DIφ ) is trained to classify
whether the images are real or refined, where φ denotes the
discriminator network’s parameters. The refined image is de∼
noted as x (Fig. 2d). The set of real images yi ∈ Υ (Fig. 2a)
2 http://field-ii.dk/
3 https://in.mathworks.com/matlabcentral/fileexchange/34199
are also used for learning the parameters. The loss function
consists of a realism loss term (lreal ) and an adversarial loss
term. The realism loss term ensures that the ground truth label remains intact in the low resolution refined image which
is used to condition the Stage II GAN. During training of the
Stage I GAN, the discriminator loss (LDI ) and generator loss
(LGI ) are minimized alternately where,
LGI (θ) = −
X
log(1 − DI φ (GI θ (xi )) + λlreg (θ; xi ) (1)
i
LDI (φ) = −
X
log(DI φ (GI θ (xi )))
i
−
X
log(1 − DI φ (yj ))
(2)
j
Self-regularization here minimizes per-pixel difference between the refined image and the synthetic image as
∼
lreg = kx − xk1
(3)
A buffer of refined images generated over the previous steps
was used to improve the realism of the artifacts in the refined
images as in [15].
Model architecture: The Stage I GAN is a residual network [16] with 4 residual blocks. The simulated image from
the pseudo B-mode simulator of size 64 × 64 is first passed
through the generator with convolution layer with 3 × 3 sized
kernel and 64 feature maps on output. The refined image is
generated by a convolution layer with 1 × 1 sized kernel. The
discriminator comprises of 5 convolution layers and 2 max
pooling layers as in [15].
3.3. Stage II GAN
The generator of the Stage II GAN accepts the low resolution
images from the Stage I GAN generator and generates high
resolution images adding realistic artifacts. While training
the Stage II GAN, the discriminator loss (LDII ) and generator
loss (LGII ) are minimized alternately where,
LGII (θ) = −
X
log(1 − DII φ (GII θ (GI (xi )))
(4)
i
LDII (φ) = −
X
log(DII φ (GII θ (GI (xi )))
i
−
X
log(1 − DII φ (yj ))
(5)
j
The high resolution images generated are in polar domain
(Fig. 2e) and scan converted to cartesian coordinate domain
(Fig. 2j) for visual inspection.
(a) Real image
(b) Tissue map
(c) Pseudo B-Mode
(d) Stage I GAN
(e) Stage II GAN
(f) Real image
(g) Tissue map
(h) Pseudo B-Mode
(i) Stage I GAN
(j) Stage II GAN
Fig. 2: Comparison of images obtained at different stages of the proposed framework. The first row displays images in the
Polar domain and the second row displays the corresponding figures in the Cartesian domain.
Model architecture: The Stage II GAN generator consists of a downsampling block that comprises of two maxpooling layers to downsample the 64 × 64 image by a factor
of four. The feature maps are then fed to residual blocks,
followed by upsampling the image to 256 × 256. The discriminator consists of downsampling blocks that reduce the
dimension to 4 × 4 followed by a 1 × 1 convolution layer and
a fully connected layer to generate the prediction.
4. EXPERIMENTS, RESULTS AND DISCUSSION
In this section, experimental validation of our proposed
framework is presented. Images from the dataset released
as part of the border detection in IVUS challenge4 were used
in our work. The images were obtained from clinical recordings using a 20 MHz probe. Of the 2, 175 available images,
435 had been manually annotated. We used 2, 025 real images for training (Patient Id. 1-9) and 150 images were held
out for testing (Patient Id. 10 ). The ground truth labeled
dataset containing 435 images were augmented by rotating
each image by 30◦ over 12 steps. Each such image was then
warped by translating up and down by 2% in the polar domain
to yield an augmented dataset of 15, 660 tissue maps to be
processed with the pseudo B-mode simulator.
The Stage I GAN was trained with a learning rate of 0.001
with the Adam optimizer over 20 epochs with a batch size of
512. While training the Stage II GAN, the weights of Stage I
GAN were held fixed. Training was done with the Adam optimizer with an initial learning rate of 0.0002, and a learning
4 http://www.cvc.uab.es/IVUSchallenge2011/dataset.html
rate decay of 0.5 per 100 epochs over 1, 200 epochs with a
batch size of 64.
Images obtained at different stages of the pipeline are provided in Fig. 2. The Stage I GAN generates low resolution (64
x 64) images and are blurry with many defects. As shown, the
Stage II GAN refines the image generated by Stage I GAN,
adding patho-realistic details at 4× higher resolution.
Qualitative evaluation of the refined images: In order
to evaluate the visual quality of the adversarially refined images, we conduct a Visual Turing Test (VTT). A set of 20 real
images were randomly selected from the 150 images in test
set. Each real image was paired with simulated image and
presented to 13 evaluators who had experience in clinical or
research ultrasonography. Each evaluator was presented with
an independently drawn 20 pairs and they had to identify the
real image in each pair. Cumulatively they were able to correctly identify 147 real IVUS out of 260 present with 56%
chance of correct identification.
Quantitative evaluation of the refined images: The histograms of the real and adversarially refined images in the
three regions i.e., lumen, media and externa for 30 randomly
chosen annotated images were obtained and the probability
mass functions were computed from the histograms. A random walk based segmentation algorithm [17] was used for
the annotation of the refined images. The Jensen-Shannon
(JS) divergence between the probability mass functions of the
three regions in real vs. simulated images were computed and
summarized in Table. 1. It is observed that result of Stage
II GAN are realistically closer to real IVUS as compared to
Stage 0 results. The pairwise JS-divergence between lumen
and media, media and externa and lumen and externa were
Table 1: JS-divergence between speckle distribution of different regions in simulated vs. real IVUS.
Stage II GAN
Stage 0
lumen
0.0458
0.0957
media
0.1159
0.5245
externa
0.3343
0.6685
Table 2: JS-divergence between pair of tissue types in real
and simulated IVUS.
Real
Stage II GAN
lumen-media
0.2843
0.1471
media-ext.
0.1460
0.1413
lumen-ext.
0.2601
0.3394
computed and summarized in Table. 2. The divergence between pair of tissue classes in Stage II GAN are also realistically closer to real IVUS.
5. CONCLUSION
Here we have proposed a stacked GAN based framework for
the fast simulation of patho-realistic ultrasound images using
a two stage GAN architecture to refine images synthesized
from an initial simulation performed with a pseudo B-model
ultrasound image generator. Once training is completed, images are generated by providing a tissue map. The quality of
the simulated images was evaluated through a visual Turing
test that evoked equivocal visual response across experienced
raters with 50% chance of identifying the real from simulated IVUS. The quality of the simulated images was quantitatively evaluated using JS divergence between the distributions of the real and simulated images. Similarity of the
real and simulated images is also quantified by computing the
shift in tissue specific speckle intensity distributions. Conclusively these evaluations substantiate ability of our approach
to simulate patho-realistic IVUS images, converging closer to
real appearance as compared to prior-art, while simulating an
image in under 10 ms during deployment.
6. REFERENCES
[1] Martino Alessandrini, Mathieu De Craene, Olivier Bernard,
Sophie Giffard-Roisin, Pascal Allain, Irina Waechter-Stehle,
Jürgen Weese, Eric Saloux, Hervé Delingette, Maxime Sermesant, et al., “A pipeline for the generation of realistic 3d synthetic echocardiographic sequences: methodology and openaccess database,” IEEE transactions on medical imaging, vol.
34, no. 7, pp. 1436–1451, 2015.
[2] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. WardeFarley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” in Proc. Adv., Neur. Inf. Process. Sys., 2014,
pp. 2672–2680.
[3] J. C. Bamber and R. J. Dickinson, “Ultrasonic b-scanning: a
computer simulation,” Phys., Med., Biol., vol. 25, no. 3, pp.
463, 1980.
[4] O. Kutter, R. Shams, and N. Navab, “Visualization and gpuaccelerated simulation of medical ultrasound from ct images,”
Comp. Methods, Prog., Biomed., vol. 94, no. 3, pp. 250–266,
2009.
[5] M. D. R. Ramı́rez, P. R. Ivanova, J. Mauri, and O. Pujol, “Simulation model of intravascular ultrasound images,” in Int. Conf.
Med. Image Comput., Comp.-Assist. Interv., 2004, pp. 200–
207.
[6] S. C. Groot, R. Hamers, F. H. Post, C. P. Botha, and N. Bruining, “Ivus simulation based on histopathology,” in Proc. Int.
Conf. Comp. Cardiol., 2006, pp. 681–684.
[7] S. Kraft, A. Karamalis, D. Sheet, E. Drecoll, E. J. Rummeny,
N. Navab, P. B. Noël, and A. Katouzian, “Introducing nuclei
scatterer patterns into histology based intravascular ultrasound
simulation framework,” in Proc. SPIE Med. Imaging, 2013,
vol. 8675, pp. 86750Y–6.
[8] S. Kraft, S. Conjeti, P. B. Noël, S. Carlier, N. Navab, and
A. Katouzian, “Full-wave intravascular ultrasound simulation
from histology,” in Int. Conf. Med. Image Comput., Comp.Assist. Interv., 2014, pp. 627–634.
[9] C. Abkai, N. Becherer, J. Hesser, and R. Männer, “Real-time
simulator for intravascular ultrasound (ivus),” in Proc. SPIE
Med. Imaging, 2007, vol. 6513, pp. 65131E–1–6.
[10] F. M. Cardoso, M. C. Moraes, and S. S. Furuie, “Realistic ivus
image generation in different intraluminal pressures,” Ultrasound, Med., Biol., vol. 38, no. 12, pp. 2104–2119, 2012.
[11] Yipeng Hu, Eli Gibson, Li-Lin Lee, Weidi Xie, Dean C Barratt, Tom Vercauteren, and J Alison Noble, “Freehand ultrasound image simulation with spatially-conditioned generative
adversarial networks,” in Molecular Imaging, Reconstruction
and Analysis of Moving Body Organs, and Stroke Imaging and
Treatment, pp. 105–115. Springer, 2017.
[12] Martin Arjovsky and Léon Bottou, “Towards principled
methods for training generative adversarial networks,” arXiv
preprint arXiv:1701.04862, 2017.
[13] Y. Yu and S. T. Acton, “Speckle reducing anisotropic diffusion,” IEEE Trans. Image Process., vol. 11, no. 11, pp. 1260–
1270, 2002.
[14] H. Zhang, T. Xu, H. Li, S. Zhang, X. Huang, X. Wang, and
D. Metaxas, “Stackgan: Text to photo-realistic image synthesis
with stacked generative adversarial networks,” arXiv preprint
arXiv:1612.03242, 2016.
[15] A. Shrivastava, T. Pfister, O. Tuzel, J. Susskind, W. Wang, and
R. Webb, “Learning from simulated and unsupervised images
through adversarial training,” Proc. IEEE/CVF Conf. Comp.
Vis., Patt. Recog., pp. 2107–2116.
[16] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning
for image recognition,” in Proc. IEEE/CVF Conf. Comp. Vis.,
Patt. Recog., 2016, pp. 770–778.
[17] L. Grady, “Random walks for image segmentation,” IEEE
Trans. Patt. Anal., Mac. Intell., vol. 28, no. 11, pp. 1768–1783,
2006.
| 1cs.CV
|
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
1
Single-Anchor Localization and Orientation
Performance Limits using Massive Arrays:
MIMO vs. Beamforming
arXiv:1702.01670v2 [cs.IT] 25 May 2017
Anna Guerra, Member, IEEE, Francesco Guidi, Member, IEEE, Davide Dardari Senior Member, IEEE
Abstract—Next generation cellular networks will experience
the combination of femtocells, millimeter-wave (mm-wave) communications and massive antenna arrays. Thanks to the beamforming capability as well as the high angular resolution provided
by massive arrays, only one single access point (AP) acting as
an anchor node could be used for localization estimation, thus
avoiding over-sized infrastructures dedicated to positioning. In
this context, our paper aims at investigating the localization
and orientation performance limits employing massive arrays
both at the AP and mobile side. Thus, we first asymptotically
demonstrate the tightness of the Cramér-Rao bound (CRB) in
massive array regime, and in the presence or not of multipath.
Successively, we propose a comparison between MIMO and
beamforming in terms of array structure, time synchronization
error and multipath components. Among different array configurations, we consider also random weighting as a trade-off
between the high diversity gain of MIMO and the high directivity
guaranteed by phased arrays. By evaluating the CRB for the
different array configurations, results show the interplay between
diversity and beamforming gain as well as the benefits achievable
by varying the number of array elements in terms of localization
accuracy.
Index Terms—Position Error Bound, Orientation Error Bound,
Millimeter-wave, Massive array, 3D localization, 5G.
I. I NTRODUCTION
The widespread use of personal devices generates new
challenges while opening new appealing scenarios for future
applications, such as, for example, those entailing device-todevice (D2D) interactions or Big Data management issues.
To meet these new trends, different disruptive technologies
have been recently proposed for the next fifth generation (5G)
wireless communications networks [1], [2]. In particular, largescale antenna arrays at base stations (BSs) or femtocells APs
allow to smartly direct the power flux towards intended users
thus increasing data rates, whereas mm-wave communication
provides a less crowded and larger spectrum [3]–[5].
In next years, it is expected that personal devices localization and communication capabilities will play a crucial
Anna Guerra and Davide Dardari are with the Dipartimento di Ingegneria dell’Energia Elettrica e dell’Informazione “Guglielmo Marconi” - DEI,
University of Bologna, via Venezia 52, 47521 Cesena, ITALY.
(e-mail: [email protected], [email protected]).
Francesco Guidi is with CEA, LETI, MINATEC Campus, 38054 Grenoble,
France. (e-mail: [email protected]).
This research was supported in part by the XCycle project (Grant 635975)
and by the IF-EF Marie-Curie project MAPS (Grant 659067). Authors would
like to thank Nicoló Decarli, Raffaele D’Errico and Henk Wymeersch for
fruitful discussions.
role [6]: in fact, the possibility of localizing nodes in indoor
environments will be an essential feature of future devices.
In this context, the AP could be used as a single-anchor
node, i.e., a node whose position is a-priori known, in a
radio-localization perspective permitting the mobile users to
be aware of their own position. Furthermore, the adoption of
more than one antenna at the transmitter (Tx) and receiver
(Rx), will enable the user orientation estimation at an accuracy
higher than that provided by compass and gyroscopes. Such
feature could play a key role in applications beyond 5G as for
example augmented reality and simultaneous localization and
mapping (SLAM), where trajectory errors, comprising both
position and orientation estimation inaccuracies, dramatically
affect the performance [7]. Contrarily to traditional scenarios
where dedicated multiple anchor nodes are necessary to allow
classic triangulation/multilateration techniques [8], here the
possibility to centralize both communication and localization
capabilities in a single multi-antenna AP working at mm-wave
frequencies is envisioned with the advantage of drastically
decreasing the overall system complexity and cost. Moreover,
when operating at such high frequencies, not only APs but
also user terminals could adopt massive arrays thanks to the
reduced wavelength [9], thus increasing even more the localization accuracy given the potential huge set of measurements
[10]–[13].
While at microwave frequencies the antenna array technology is quite mature, at mm-wave severe technological constraints are still present and must be taken into account when
designing positioning systems. Recently, massive antennas
prototypes have been proposed with electronic beamsteering
capabilities. In order to reduce the complexity, they adopt
simple switches and thus, the resulting non-perfect signals
phasing operations could impact the array radiation characteristics [14]–[16]. In such a scenario, it becomes of great interest
to understand the fundamental limits on localization error with
massive antenna arrays both at AP and mobile terminals using
only a single reference node.
Concerning the ultimate localization performance evaluation, a rich literature has been produced for the analysis
of wideband multiple anchors systems. Specifically, in [17],
[18] authors explore the localization accuracy for a wideband
sensors network composed of several independent anchors.
Their results are further discussed in [19] where a more
realistic Rx architecture able to exploit the carrier phases
information has been taken in consideration while deriving
the localization performance. Differently from these works,
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
where anchors send orthogonal waveforms, we consider a
signal model dependent on the particular arrays architecture
chosen where both orthogonal and non-orthogonal waveforms
can be transmitted. Moreover, our work is not focused on
a specific Rx structure, as in [19], but it aims to compare
different Tx array architectures. In [20], [21], a joint delayangle estimation is reported considering different array technologies and frequency bandwidths. Nevertheless, these works
analyze the performance in terms of CRB on delay and angular
information rather than directly on localization, and neither
a comparison between different array schemes, nor the time
synchronization issue and the impact of multipath are treated.
In our previous work [12], [22], some preliminary results on
positioning accuracy considering only beamforming strategies
have been presented, but the comparison with multiple-input
multiple-output (MIMO), as well as the impact of multipath
components (MPCs), was not considered.
Stimulated by this framework, in this paper we conduct
a CRB-based analysis of a localization system exploiting
the next 5G technologies potentialities. Differently from the
state-of-the-art, we adopt a 1-step approach in which the
Tx position and orientation are directly inferred from the
received signals and, thus, without the need of estimating
intermediate parameters (e.g., time-of-arrival (TOA)-directionof-arrival (DOA)) or applying geometrical methods which do
not ensure the optimality of the approach [23].
The main contributions of this work can be summarized as
follows:
•
•
•
•
•
•
Derivation of the theoretical performance limits on the
localization and orientation error for different array configurations in a single-anchor scenario;
Proposal of a signal model valid for any antenna array geometry, configuration (i.e., MIMO, phased, timed arrays),
and frequency band. As a case study, in the numerical
results the focus is on the adoption of mm-wave massive
arrays due to their expected attractiveness in next 5G
applications;
Introduction of low-complexity random weighting approach, i.e., randomly chosen beamforming weights, and
analysis of its performance compared to that of classical
beamforming and MIMO solutions;
Investigation of the CRB tightness in massive array
regime (i.e., letting the number of antennas → ∞) for
any signal-to-noise ratio (SNR) condition;
Analysis of the trade-off between SNR enhancement
obtained via beamforming and diversity gain of MIMO
considering the impact of different types of uncertainties,
as, for example, the MPCs, beamforming weights and
time synchronization errors;
Demonstration that in the massive array regime (i.e., array
antennas → ∞), the effect of multipath can be made
negligible on average.
The rest of the paper is organized as follows. Sec. II
describes the geometry of the localization system. Then,
Sec. III introduces the signal model taking into account different array structures. In Sec. IV the localization performance
limits derivation is reported. Sec. V analyzes the asymptotic
2
conditions for which the CRB can be considered a tight
bound. Sec. VI derived compact formulas for a ideal freespace case. The multipath impact on localization performance
are investigated in Sec. VII. Finally, Sec. VIII presents the
localization performance results and Sec. IX concludes the
work.
Notation: Lower case and capital letters in bold denote
T
∗
vectors and matrices, respectively. The subscripts [·] , [·] and
H
[·] indicate the transpose, the conjugate and the Hermitian
operators. k·k2 is the Euclidean norm, A B indicates
that the matrix A − B is non-negative definite, and diag (·)
represents the diagonal operator. The subscripts (·)t and (·)r
refer to quantities related to the transmitting and receiving
array, respectively, while the subscript (·)tr to elements that
can be referred to both the Tx and the Rx. (·)FS indicates
the free-space scenario. F (·) denotes the Fourier transform
operation, U (a, b) a uniform distribution in the interval [a, b],
and CN µ, σ 2 a circularly symmetric Gaussian distribution
with mean µ and variance σ 2 .
The notations of frequently-used symbols are listed as
follows.
Ntx , Nrx
Number of Tx-Rx array antennas
L
Number of MPCs
At , Ar
Area of the Tx-Rx array
pt , ϑt
Tx centroid position and orientation
pr , ϑr
Rx centroid position and orientation
d
Distance between Tx-Rx centroids
S = Ar /d2
Ratio between the Rx array area and the
squared inter-array distance
Tx/Rx antenna position
pti , prm
dant
Inter-antenna spacing
d (θ)
Direction cosine
κ
Multipath parameters vector
θ1
Direct path wave direction
θl
lth path wave direction
θ0
Steering direction
τim1 , τiml
Propagation delay relative to the direct
and lth path between the ith Tx-mth Rx
antenna
τ1 , τl
Propagation delay relative to the direct
and lth path between centroids
a1 , αl
Direct path amplitude and lth complex
channel coefficient
Inter-antenna delay between the ith Tx
τit (θlt , ϑt )
antenna and the relative array centroid
r
τm
(θlr , ϑr )
Inter-antenna delay between the mth Rx
antenna and the relative array centroid
fc , W, β
Transmitted signal carrier frequency,
bandwidth, baseband effective bandwidth
Tobs
Observation interval
Etot , E
Total and normalized energy at each antenna element
N0
Single-side noise power spectral density
NF
Receiver noise figure
SNR1
SNR relative to the direct path
Si (f ), s(f )
Equivalent low-pass signal at the ith Tx
antenna and transmitted signals vector in
the frequency domain
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
3
Pi (f )
Equivalent low-pass unitary-energy signal at the ith Tx antenna in the frequency
domain
Rm (f ), r(f )
Received signal at the mth Rx antenna
and received signals vector in the frequency domain
Xm (f ), x(f )
Useful Rx signal at the mth Rx antenna
and useful Rx signals vector in the frequency domain
Nm (f ), n(f )
Noise component at the mth Rx antenna
and noise vector in the frequency domain
ωi , B(f, θ0 )
Beamforming weight and matrix
µti (θ0 ), τit (θ0 ), νi Beamforming phase, TDL and random
weight
δit , ∆τit
Beamforming phase and TDL errors
ω̃i , Q(f )
Beamforming weight and matrix with errors
ǫs
Time synchronization error
ψ
Estimation parameter vector
Bayesian FIM, FIM relative to data, aJψ , Jdψ , Jpψ
priori FIM
CRB (q)
CRB on position and orientation
CRB0
Single-antenna CRB on ranging error
II. A NTENNA A RRAY G EOMETRIC C ONFIGURATION
A. Geometric Relationships
We consider a 3D localization scenario, as the one reported
in Fig. 1, consisting of a single AP acting as reference receiving node equipped with an antenna array, with Nrx antennas,
and a transmitting mobile terminal with a Ntx -antenna array.
The localization process aims at directly inferring:1
T
t
• the position of the Tx centroid p
= [xt0 , y0t , z0t ] =
T
[x, y, z] ;
t
t
t T
• the orientation of the Tx ϑ = [ϑ , ϕ ]
T
T
when the Rx centroid position pr = [xr0 , y0r , z0r ] = [0, 0, 0]
T
and orientation ϑr = [ϑr , ϕr ] are known.2 With reference
T
to Fig. 1, pti (ϑt ) = [xti , yit , zit ] indicates the position of
the ith transmitting antenna relative to the Tx geometric
center and dependent on the Tx orientation, and prm (ϑr ) =
r
r T
[xrm , ym
, zm
] the position of the mth receiving antenna
relative to the Rx geometric center. Considering spherical
coordinates, we have
tr
(1)
ptri/m (ϑtr ) = ρtri/m R ϑtr dT θi/m
with the direction cosine is expressed as
d (θ) = [sin(θ) cos(φ), sin(θ) sin(φ), cos(θ)]
(2)
iT
h
tr
tr
and ρtri/m = kptri/m (ϑtr ) − ptr k2 and θi/m
= θi/m
, φtri/m
being the distance and the couple of angles between the con1 As previously stated, we consider the Tx position and orientation with
respect to the relative centroid (see (1)-(3) in the following) as we adopt a
1-step approach in which the Tx position and orientation are directly inferred
from the received signals. Thus, we do not estimate neither the DOA (i.e.,
angle between arrays centroids) nor the direct path TOA.
2 Without loss of generality, the Rx is assumed located at the origin of the
coordinates system.
z ′′
Transmitter
D
Tx centroid
Rx centroid
pt
D ≪ cτ
′′
x
y ′′
pti
z′
τim
τ
Receiver
z
pant
θ
p0
prm
pr
y′
φ
ρ
y
x
x′
Fig. 1. Multi-antenna system configuration.
sidered array antenna from the correspondent array centroid.3
The rotational matrix R(ϑtr ) is given by
R(ϑtr ) = Rz (ϕtr ) Rx (ϑtr )
(3)
where Rz (ϕtr ) and Rx (ϑtr ) define the counter-clock wise
rotation around the z-axis and the clock wise rotation around
T
the x-axis, respectively. Finally θ1 = [θ1 , φ1 ] designates the
angle of incidence between arrays centroids (direct path) and
T
θ0 = [θ0 , φ0 ] represents the intended pointing direction of the
steering process when applied.
The diameter D of the transmitting and receiving arrays
is assumed much smaller than the inter-array distance d =
kpr − pt k2 , i.e., D ≪ d. Note that this hypothesis is especially
verified at mm-wave where the array dimensions are very
small thanks to the reduced wavelength. Moreover the arrays
are supposed to be sufficiently far from the surrounding
scatterers thus obtaining identical angles of incidence for both
direct and MPCs at each antenna element.
We take L MPCs into consideration as nuisance parameters
in the localization process and the first path is assumed
always experiencing a line-of-sight (LOS) propagation condition. For what the MPCs parameters are concerned, we
follow the same notation introduced in [19]. In particular, let
T
T
θlt = [θlt , φtl ] = [θ1 + ∆θlt , φ1 + ∆φtl ] and θlr = [θlr , φrl ]T =
T
[θ1 + ∆θlr , φ1 + ∆φrl ] , with l = 1, 2, . . . , L , indicate the angles of departure from the transmitting array and of incidence
at the Rx side of the l th path, respectively. The angular biases
T
T
[∆θlt , ∆φtl ] and [∆θlr , ∆φrl ] are the displacement with respect
to the direct path at the Tx and Rx side. Obviously, it is
T
T
T
[∆θ1t , ∆φt1 ] = [∆θ1r , ∆φr1 ] = [0, 0] , when direct path is
considered.
Let τ1 , kpr − pt k2 /c = d/c and τim1 , kprm − pti k2 /c
being the propagation delay related to the direct path between
the transmitting and receiving centroids and between the ith
and mth antenna, respectively, where c is the speed of light.
Considering the multipath, the lth propagation delay between
array centroids is defined as τl = τ1 + ∆τl where ∆τl is the
non-negative delay bias of the lth path with ∆τ1 = 0 [19].
3 Note that the elevation angle in all the text is indicated with θ and it can
assume values in the interval [0, π). Contrarily the azimuthal angle is denoted
with φ and it ranges between [0, 2 π).
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
z
Receiver
Transmitter
τim1
mth
ith
θ1t
θ1r
τ1
r
τm
φr1
τit
φt1
y
x
Fig. 2. Array geometric configuration.
According to the geometric assumption previously described,
the TOA and amplitude between each couple of transmitting
and receiving antennas can be expressed using the following
approximations [19], [21]
1) τiml ≈ τl +
τit (θlt , ϑt )
−
r
τm
(θlr , ϑr )
2) aiml ≈ al
(4)
where aiml is the amplitude of the lth path between the mth
r
receiving and the ith transmitting antenna, and τm
(θlr , ϑr )
t t
t
and τi (θl , ϑ ) are, respectively, the receiving and transmitting
inter-antenna propagation delays defined as
1
tr
(5)
(θltr , ϑtr ) = d θltr ptri/m ϑtr
τi/m
c
Fig. 2 reports a graphical explanation of the considered system
and delays. As it can be seen, the approximation in (4) permits
to write the TOA as the summation of the inter-antenna delays
and the delay between the array centroids.
B. Special Case: Planar Array Geometry
Planar array configurations appear to be the most suitable
when considering the integration of massive arrays in portable
devices or in small spaces. For this reason, in addition to the
general analysis valid for any geometric configuration (i.e.,
any antennas spatial deployment and arrays orientation), some
compact specialized equations will be derived in the following
sections for squared arrays of area At = d2ant Ntx (Ar = d2ant Nrx ),
with the antennas equally spaced of dant . Both arrays are
considered lying on the XZ-plane and being located one in
T
T
front of the other with pr = [0, 0, 0] and pt = [0, y, 0] with
y > 0, so that d = y and thus τ1 = y/c. In this case, the
antenna coordinates in (1) becomes
tr T
= R ϑtr [mx dant , 0, mz dant ]T (6)
ptrm ϑrt = xtrm , 0, zm
√
√
√
where mx = mz = − N2−1 , − N−1
+ 1, . . . , N2−1 are the
2
antenna indexes along the x− and z−axis respectively, and N
indicates the number of antennas.
We assume for now a free-space propagation
T condition so
T
that θ1t = θ1r = θ1 = [θ1 , φ1 ] = π2 , − π2 and d(θ1 ) =
[0, −1, 0]. Consequently it is possible to specialize (5) as
dant
(mx sin (ϕr ) + mz cos (ϕr ) sin (ϑr ))
c
dant
(7)
ix sin ϕt + iz cos ϕt sin ϑt
τit (θ1 , ϑt ) = −
c
Note that, in the special case in which the Rx and Tx
T
orientation is ϑr = ϑt = [0, 0] , the inter-antenna delays
r
τm
(θ1 , ϑr ) = −
r
are zeros, i.e. τm
(θ1 , ϑr ) = τit (θ1 , ϑt ) = 0 ∀m, i, as the
antennas are aligned to the array centroids, thus the incident
wave impinges simultaneously at all the antennas.
III. A NTENNA A RRAY S CHEMES
pt
pr
4
AND
S IGNAL M ODEL
In this section, different types of antenna array schemes are
analyzed starting from a unified signal model with the purpose
to highlight their beamforming and diversity gain properties.
Specifically, the four array structures reported in Fig. 3 will
be analysed from a signal processing point-of-view and by
focusing on how the different signaling schemes translate into
different localization capabilities. Table I reports a comparison
in terms of arrays complexity, capabilities and cost.
A. Transmitted Signal Model
The transmitted signal
at the ith transmitting antenna is
denoted with gi (t) = ℜ si (t) ej 2 πfc t where si (t) represents
the equivalent low-pass signal and fc the carrier frequency.
We consider a constraint on the total transmitted energy
Etot which Ris uniformly allocated among antennas, thus E=
Etot /Ntx = Tobs |si (t)|2 dt, i = 1, 2, . . . , Ntx , represents the
normalized energy at each antenna element. We introduce
the Fourier transform of si (t) as Si (f ) = F {si (t)}, with
F {·} denoting the Fourier transform operation in a suitable
observation interval Tobs containing the signal support. For
T
further convenience, the vector s(f ) = [S1 (f ), . . . , SNtx (f )]
contains all the baseband transmitted signals. In the following,
the signal model for each array configuration will be further
detailed with reference to Fig. 3.
1) Timed and phased arrays: In multi-antenna systems,
beamforming is obtained by applying a progressive time delay
at each array element so that the emitted signals result summed
up coherently towards the intended steering direction. Considering the signal bandwidth W , when the condition W ≪ fc
holds, this process can be well approximated using only phase
shifters (PSs) (phased arrays). On the contrary, when W ≈ fc ,
phase shifts are no longer sufficient to align all the signals.
As a consequence, to avoid distortion and beamsteering4
degradation (squinting effect), timed arrays consisting of PSs
and time delay lines (TDLs) must be introduced. The following
analysis considers both array structures in order to preserve
the generality of the model. Nevertheless, in Sec. VIII, arrays
operating at mm-wave frequencies with W ≪ fc (narrowband)
will be adopted in simulations.5 In [12], different fractional
bandwidths have been taken into account in the results.
Moreover, differently from [24], [25], where multiple beams
are generated, here we consider a single-beam scenario in
order to maximize the SNR in the desired steering direction
and to reduce the processing time.
Given these array schemes, the transmitted signal is the
same
for all transmitting antennas, i.e., si (t) = s(t) =
√
E p(t) ∀i = 1, . . . , Ntx , with p(t) being the unitary energy
normalized version of s(t), and beamforming weights are
applied to each branch of the array to focus the power flux in
4 The
terms beamsteering and beamforming are used as synonymous.
expected, since W ≪ fc , the localization performance of timed and
phased arrays coincides.
5 As
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
(b) Timed Array
(a) Phased Array
µt1 (θ0 )
+
5
δ1t
τ1t (θ0 ) + ∆τ1t
+
δ1t
TDL
S(f )
µti (θ0 ) + δit
t
µtNtx (θ0 ) + δN
tx
S(f )
τit (θ0 ) + ∆τit
(d) Random Weighting
(c) MIMO
µt1 (θ0 )
υ1
S1 (f )
µti (θ0 ) + δit
υi
S(f )
Si (f )
TDL
t
t
t
µtNtx (θ0 ) + δN
(θ0 ) + ∆τN
τN
tx
tx
tx
TDL
υNtx
SNtx (f )
Fig. 3. From the left to the right: Phased, timed, MIMO and random weighting array schemes.
TABLE I
A RRAY SCHEMES COMPARISON .
Signal Design
Array Implementation Complexity
Cost
Beamforming capabilities
Timed
Low: same signal for all antenna branches
High: TDLs needed when W ≫ fc
High
Yes
Phased
Low: same signal for all antenna branches
Medium: only PSs
Medium
Yes
MIMO
High: one different signal for all antenna branches
High: a RF chain for each branch
High
No
Random
Low: same signal for all antenna branches
Low: only PSs
Low
No
a precise direction in space. Specifically, when no quantization
errors are present in the weights, the ideal beamforming matrix
can be defined as
B (f, θ0 ) = diag (ω1 , ω2 , . . . , ωi , . . . ωNtx ) .
(8)
The ith beamforming weight is ωi = bi (f ) bci having indit
t
cated bi (f ) = ej2 π f τi (θ0 ) and bci = ejµi (θ0 ) , where µti (θ0 ) =
t
t
2 π fc τi (θ0 ) and τi (θ0 ) are the transmitting steering phase and
delay related to the ith PS and TDL of the array, respectively.
The main difference between phased and timed array is the
way in which the beamsteering process is performed: in the
former only PSs are present (i.e., τit (θ0 )=0 ∀i = 1, . . . , Ntx ,
refer to Fig. 3-(a)) while in the latter TDLs and PSs are
both employed to counteract the beamsquinting effect caused
by a larger W/fc ratio (see Fig. 3-(b)). Nevertheless, some
technological issues could induce errors in the beamforming
vector. Firstly, when digitally controlled PSs are used in place
of their high-resolution analog counterparts, the presence of
quantization errors has to be accounted for [7]. As shown in
[14], [15], where some massive arrays prototypes working in
the X- and V-bands have been proposed, PSs can be realized
by simply adopting switches, or by rotating patch antennas.
Therefore, continuous phase shifts ranging from 0◦ to 360◦
are not realizable in practice and the quantization errors
generated by the consequent discretization of phases should
be taken into account when considering real massive arrays.
Secondly, time synchronization between the Tx and the Rx is
required to estimate the position. There are several techniques
to accomplish this task [26], with the two-way ranging being
one of the most used. Unfortunately due to several factors
such as clock drift, a residual time synchronization error is
always present and it is accounted by the term ǫs in our
model. In the presence of such non-perfect weights and time
synchronization error, a matrix accounting for all the nonidealities is introduced
s
Q(f ) = e−j 2 π (f +fc )ǫ diag (ς1 , ς2 , . . . , ςi , . . . , ςNtx ) (10)
where ςi takes into account the i th beamforming weight
t
t
quantization error, i.e., ςi = ej (2 π f ∆τi +δi ) with δit being the
phase error and ∆τit the TDL error. For further convenience,
t
t
let indicate with ω̃i = b̃i (f ) b̃ci where b̃i (f )=ej2 π f (τi (θ0 )+∆τi )
t
t
and b̃ci=ej(µi (θ0 )+δi ) the quantized weights. After the transmitting beamforming process, the signal at each antenna element
can be written as Q(f ) B(f, θ0 ) s(f ).
2) MIMO arrays: Contrarily to timed or phased arrays,
which perform accurate beamforming, MIMO arrays take
advantage of the diversity gain provided by multiple different
waveforms [27], [28] (see Fig. 3-(c)).6 To make the Rx able
to discriminate the signal components coming from each single transmitting antenna, orthogonal waveforms are typically
adopted [28]–[31]. As an example, in [28] a class of signals
(i.e., frequency spread signals) are demonstrated to maintain
orthogonality for time delays and frequency Doppler shifts.
This comes at the expense of a large bandwidth or symbol
duration time and of a higher complexity. In MIMO arrays,
the normalized baseband transmitted signals are
R indicated with
Pi (f ) = F {pi (t)} = √1E F {si (t)}, where Tobs |pi (t)|2 dt =
1, i = 1, 2, . . . , Ntx . We consider orthogonal waveforms, such
6 Note that here we refer to MIMO as done in radar literature rather than
in communications.
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
r(f ) =
L
X
6
ar (f, θlr , ϑr ) c(f, τl ) At (f, θlt , ϑt ) Q(f ) B(f, θ0 ) s(f ) + n(f ) = x(f ) + n(f )
(9)
l=1
that the correlation function is
Z
(l,k)
(l,k)
=
Pi (f ) Pj∗ (f ) e−j2πf ∆τij df
∆τ
Rp
ij
(W
0
i 6= j
=
∀l, k = 1, . . . , L (11)
6= 0
i=j
(l,k)
= τiml − τjmk with m = 1, . . . , Nrx and
where ∆τij
i, j = 1, . . . , Ntx . The possibility to provide orthogonal waveforms permits to increase the diversity gain, as it will be
detailed in next sections, but it requires a greater bandwidth
demand and a more complex Tx structure. In MIMO, the
matrix in (8) is an identity matrix B (f, θ0 ) = B = INtx .
In presence of the time synchronization error, (10) becomes
s
Q(f ) = e−j 2 π (f +fc )ǫ INtx .
3) Random Weighting: To avoid the complexity of MIMO,
we propose a strategy relying on the same structure of phased
arrays, i.e., with only PSs at each antenna branch (see Fig. 3(d)), with the fundamental difference that the value assigned
to each PS is randomly chosen. The beamforming matrix of
(8) becomes
B (f, θ0 ) = B = diag ej υ1 , ej υ2 , . . . , ej υi , . . . , ej υNtx
(12)
with υi ∼ U (0, 2π). Note that in this configuration the matrix
in (12) does not depend on the frequency and on the steering
direction, thus resulting in an array pattern with a random
shape [12]. In the simplest implementation, random weighting
could be realized using switches as discrete PSs randomly
changing their status [15]. An important aspect is that, for both
MIMO and random weighting, the rank of B is maximum and
equal to Ntx .
B. Received Signal Model
In this section, a general framework for the received signal
model is illustrated. The received signals are collected in
T
a vector r(f ) = [R1 (f ), . . . , Rm (f ), . . . , RNrx (f )] , where
Rm (f ) = F {rm (t)} is evaluated in Tobs and rm (t) is
the equivalent low-pass received signal at the mth receiving
antenna. Specifically, the received signal can be written as in
(9). The receiving and transmitting direction matrices for the
inter-antennas delays and Tx orientation are given by
iT
h r
r
r
(13)
ar (f, θlr ) = ej γ1 , . . . , ej γm , . . . , ej γNrx
t
t
t
At (f, θlt , ϑt ) = diag e−j γ1 , . . . , e−j γi , . . . , e−j γNtx (14)
tr
tr
where γi/m
= 2 π (f + fc ) τi/m
(θltr , ϑtr ); while c(f, τl ) =
cl 11×Ntx is the 1 × Ntx channel vector whose generic element
is cl = al e−j 2 π (f +fc ) τl = αl e−j 2 π f τl . Specifically, the
dominant LOS component related to direct path (i.e., l=1) is
considered deterministic while, for l > 1, αl ∼ CN 0, σl2
is a circularly symmetric Gaussian random variable (RV)
statistically modelling the lth MPC [32]. Finally, x(f ) =
[X1 (f ), . . . , Xm (f ), . . . , XNrx (f )]T is the set of useful reT
ceived signals and n(f )=[N1 (f ), . . . , Nm (f ), . . . , NNrx (f )]
is the noise vector with Nm (f ) = F {nm (t)}, with nm (t) ∼
CN (0, N0 ) being a circularly symmetric, zero-mean, complex Gaussian noise. For further convenience, define νt =
Etot /N0 = ν Ntx , with ν = E/N0 . The (total) SNR at
each receiving antenna element is SNRt = Ntx SNR1 , where
2
SNR1 = (a1 ) ν represents the SNR component related to
the direct path between a generic couple of TX-RX antenna
elements.
IV. P OSITION
AND
O RIENTATION E RROR B OUND
A. Unknown Parameters
The aim of the system is to estimate the position pt of
the Tx and its orientation ϑt starting from the set of received
waveforms r(f ).
In this context, MPCs and the residual time synchronization error represent nuisance parameters when evaluating the
ultimate performance of the estimator. Thus, the unknown
parameters vector is defined as
ψ = [qT , κT , ǫs ]T
(15)
where the parameters of interest related
to localization
and
T
T T
orientation are collected in q = (pt ) , (ϑt ) , and the
T
multipath parameters in κ = [κT1 , κT2 , . . . , κTl , . . . , κTL ] , with
(
[a1 ]T
if l = 1
(16)
κl = ℜ ℑ T
if l > 1.
αl , αl
ℑ
The terms αℜ
l = ℜ {αl } and αl = ℑ {αl } indicate the real
and imaginary part of the complex channel coefficient, respectively [33], [34]. The time synchronization error is modeled as
independent Gaussian zero-mean RV with standard deviation
T
σǫ2 . Note that the nuisance parameters ψr = [κT2 , . . . , κTL , ǫs ]
are described statistically (a-priori information available)
T
whereas ψnr = [qT , a1 ] are treated as deterministic (no apriori information available).
In the following, we will discern among two different cases
based on the orientation awareness. Specifically, we refer to
the orientation-unaware case for indicating the situation in
which the Tx orientation is not known at Rx. Contrarily, the
orientation-aware case is the opposite situation in which the
orientation is exactly known at the Rx side and it can be removed from the list of unknown parameters in (15). Moreover,
we suppose that an initial search is conducted by the Tx in
order to coarsely estimate its own position and orientation with
respect to the Rx. Consequently, the beamforming weights can
be communicate to the Rx by exploiting the communication
link dedicated to data exchange.
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
B. CRB General Expression
b=ψ
b (r (f ))
The performance of any unbiased estimator ψ
can be bounded by the hybrid CRB defined as [35]
nh
ih
iT o
−1
b−ψ ψ
b−ψ
Er,ψr ψ
Jψ
= CRB (ψ)
(17)
where Jψ is the Bayesian Fisher Information Matrix (FIM)
defined as
Jψ , −Er,ψr ∇2ψψ ln f (r, ψr ) = Jdψ + Jpψ
d
Jqq
Jdqκ
Jdqǫs
= AH C . (18)
Jdκǫs
= Jdκq Jdκκ + Jpκκ
C D
Jdǫs q
Jdǫs κ
Jdǫs ǫs + Jpǫs ǫs
The symbol ∇2ψψ = ∂ 2 /∂ψ∂ψ denotes the second partial
derivatives with respect to the elements in ψ and
Jdψ = −Er,ψr ∇2ψψ ln f (r|ψr )
Jpψ = −Eψr ∇2ψψ ln f (ψr )
(19)
are the FIM related to data and the FIM containing the a-priori
statistical information on the parameters, respectively.
Since the observations at each receiving antenna element
are independent, the log-likelihood function ln f (r|ψr ) can
be written as
Nrx Z
1 X
|Rm (f ) − Xm (f )|2 df. (20)
ln f (r|ψr ) ∝ −
N0 m=1 W
Moreover, based on the statistical information of ψr , it is
possible to derive the a-priori probability density function of
parameters ψr whose expression is reported in Appendix A.
All FIM elements are reported in details in Appendixes A and
B.
Finally, by using the Schur complement, the CRB expression related to the localization and orientation estimation error
can be easily derived as
CRB (q) = A − C D−1 CH
−1
.
(21)
Equation (21) is a general bound valid for different setup (MIMO, timed, phased and random weighting arrays) and
accounting for signal weights quantization effects, time synchronization mismatch and multipath. Specialized expressions
can be derived from (21) for specific cases to get insights on
the key parameters affecting the performance as will be done
in Sec. VI.
V. O N
THE
CRB
TIGHTNESS IN MASSIVE ARRAY REGIME
It is well known that the CRB is a meaningful metric when
the global ambiguities are negligible [35]. Such a condition
is satisfied when operating at high SNR (asymptotic SNR
regime) but, unfortunately, the required high SNRs cannot be
in general obtained, especially at high frequencies.
Therefore, in the following, we demonstrate that the global
ambiguities can be made negligible without imposing the SNR
to be very large by letting the antenna array being massive
(massive array regime). In particular, we aim to show that,
under random Rx array orientations, the number of geometric
configurations in which the ambiguities are not negligible
7
vanishes as the number of receiving antennas increases.
To this purpose, the ambiguity function (AF) is a powerful
tool to investigate the presence of ambiguities especially used
in radar systems and, it can be derived from the maximum
likelihood (ML) discarding the thermal noise component [36].
Let define the normalized AF as
2
Z
Tobs
H
(22)
x (f, p)x(f, p̃) df
AF (p, p̃) =
Ntx Nrx W
where p is the true Tx position, p̃ is a test position and, x
is the useful signal vector reported in (9). Asymptotically for
Nrx → ∞ (massive array regime), for the weak law of the
large number [37], we can write
2
Z
Tobs
P
H
(23)
E [ x (f, p)x(f, p̃)] df
AF (p, p̃) −→
Ntx Nrx W
P
where the operator −
→ indicates the convergence in probability.
In the following, we will consider the free-space and the
multipath cases, separately, in order to show how the sidelobes
level behaves in the massive array regime. The analysis in nonmassive array regime is considered in Sec. VIII.
A. Free-space Scenario
Here we focus our attention to the free-space scenario (i.e.,
l = k = 1). In this case, the expectation term in (23) becomes
1
1
E [ xH (f, p)x(f, p̃)] ∝
E [H(p, p̃)]
Ntx Nrx
Ntx Nrx
(24)
where H(p, p̃), is a Ntx × Ntx matrix whose generic element
is given by
[H(p, p̃)]i,j = |a1 |2 e−j 2 π (f +fc )∆τ1 (p,p̃) ω̃i ω̃j∗
Nrx
i
h (1,1)
X
(1,1)
r
× ejΨij (p,p̃)
E ejΨm (p,p̃,ϑ )
=
(
m=1
(1,1)
|a1 | Nrx ejΨij
0
2
(p,p̃)
p = p̃
otherwise
(25)
where we have defined ∆τ1 (p, p̃) = τ1 (p) − τ1 (p̃),
(1,1)
(1,1)
Ψij (p, p̃) = γit (p, ϑt ) − γjt (p̃, ϑt ) and Ψm (p, p̃, ϑr ) =
r
r
−γm
(p, ϑr ) + γm
(p̃, ϑr ) which depends on the Rx array ori(1,1)
entation. Note that Ψm (p, p̃, ϑr ) = 0 for p = p̃ regardless
the Rx orientation. On the other side, when p 6= p̃, in the
presence of a large number of antenna elements (Nrx → ∞)
and considering random Rx orientations, the inter-antenna
phase terms Ψm (p, p̃, ϑr ) can be modeled as independent RVs
uniformly distributed in [0, 2π). In fact, different geometric
configurations permit to span all the angles especially when
large arrays are considered 7 .
This means that the percentage of geometrical configurations of the Rx for which the ambiguities are not negligible
(i.e., AF(p, p̃) → 0 when p 6= p̃), vanishes as Nrx increases.
In other words, the conditions that permit to operate in the
non-ambiguity region during the CRB evaluation are twofold:
7 The goodness of the fitting with a uniform distribution has been validated
through simulations.
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
the first is to increase the SNR (high-SNR regime) by keeping
the number of antennas fixed, whereas the second fixes the
SNR (even not extremely large) and let the number of antennas
explode.
=
Tobs
Ntx Nrx
Z
2
xH (f, p)x(f, p̃) df
W
Z
(x1 (f, p) + xl>1 (f, p))
H
W
2
× (x1 (f, p̃) + xl>1 (f, p̃)) df
Z
L X
L
XX
(l,k)
Si (f ) Sj (f ) ω̃i ω̃j∗ e−j Ψij
(p,p̃)
mij l=2 k=2
This section aims at investigating if the CRB still remains
a meaningful metric in the presence of multipath. To this
purpose, we consider the normalized AF by putting in evidence
the multipath contribution, as
Tobs
Ntx Nrx
considerations are valid for E xH
l>1 (f, p)x1 (f, p̃) .
Finally, consider the last term in (28), i.e.
E xH
l>1 (f, p)xl>1 (f, p̃) =
=
B. Multipath Scenario
AF (p, p̃) =
8
2
i
h
(l,k)
× αl α∗k e−j 2 π f ∆τlk E e−j Ψm (p,p̃)
(30)
(l,k)
r
(θl , p, ϑr ) −
where ∆τlk = τl − τk , Ψm (p, p̃) = γm
(l,k)
t
r
r
t
γm (θk , p̃, ϑ ), Ψij (p, p̃) = γi (θl , p) − γi (θk , p̃).
In this case, since it holds
h
i
(l,k)
E e−j Ψm (p,p̃)
0 if l 6= k, ∀p̃
1 h if l = k, pi= p̃
(31)
=
E e−j Ψ(l,l)
m (p,p̃)
= 0 if l = k, p 6= p̃ ,
where x1 (f, p) and xl>1 (f, p) indicate the expected received
(noise-free) signal due to the direct path and multipath, respectively. Given the expression in (26), the following asymptotic
analysis aims at verifying that the number of times the
multipath impacts on the AF shape is negligible compared
to the number of times it has not an effect at all, provided that
the number of Rx antennas goes to infinity and that random
array orientations are considered. More precisely, recalling the
weak law of the large numbers, it is
it follows that (30) is equal to 0 for p 6= p̃, i.e. in all those
cases in which a global ambiguity can arise.
The obtained result shows that the global ambiguities due
to the multipath are, on average, negligible. Nevertheless, the
effect of multipath still remains in the correspondence of the
true peak of the AF, i.e., that for p = p̃, as reported in
(31) for l = k. Consequently, even if we can state that the
CRB is a valid metric in establishing the ultimate performance
provided that Nrx is sufficiently large, the effect of multipath
on the localization accuracy necessitates to be investigated.
Specifically, Sec. VII analyzes the effect of multipath from a
localization accuracy point-of-view.
fMP (p, p̃) P
1
E [fMP (p, p̃)]
−−→
Ntx Nrx
Ntx Nrx
VI. F REE -S PACE L OCALIZATION B OUND
=
W
fAWGN (p, p̃) fMP (p, p̃)
+
df
Ntx Nrx
Ntx Nrx
(26)
(27)
where we aim at verifying that the right-hand side of (27)
is 0 for p 6= p̃, meaning that AF sidelobes depending on
multipath disappear when Nrx is large and random orientations
are considered.
The expectation argument in (27) is given by
E [fMP (p, p̃)] = E xH
l>1 (f, p)x1 (f, p̃)
+ E xH
1 (f, p)xl>1 (f, p̃)
+ E xH
(28)
l>1 (f, p)xl>1 (f, p̃) .
Treating separately the terms in (28), we have
E xH
1 (f, p)xl>1 (f, p̃) =
=
L
XX
(1,k)
α∗1 αk Si (f )Sj∗ (f ) ω̃i ω̃j∗ e−j2 π f τk e−j Ψij
(p,p̃)
mij k=2
i
h (1,k)
× E ejΨm (p,p̃) = 0
Here we provide an example on how the general expression
(21) can be simplified in absence of beamforming weights
errors and MPCs. Specifically, in free-space conditions, (18)
can be reduced to
FS
Jqq
0
Jqq Jqa1
d
Jψ = Jψ =
=
(32)
Ja1 q Ja1 a1
0
Ja 1 a 1
where its elements are reported in Appendix B and where
the superscript d is omitted as in this case all the parameters
to estimate are deterministic. For readability convenience, we
report here the expression of the FIM related to the localization
parameters, that is:
o
X n
(1,1) (1,1)
ℜ b̃cij ξij χij (2)
Jqb qa =8 π 2 ν a21
mij
∀p̃
× ∇qa (τim1 ) ∇qb (τjm1 )
(29)
PNrx PNtx PNtx
P
(1,k)
=
where
=
mij
m=1
i=1
j=1 , Ψm (p, p̃)
(1,k)
r
r
−γm (θh1 , ϑ )+γm (θk , ϑ ), Ψij i = −γi (θ1 (p))+γj (θ1 (p̃)),
(1,k)
and E e−j (2 π f τk + Ψm (p,p̃)) = 0 as the phases are
assumed uniformly distributed between 0 and 2 π. Similar
(33)
where qa/b are two elements in the set {x, y, z, ϑt , ϕt }, and
Z
(1,1)
(1,1)
χij (2)= b̃ij (f ) (f + fc )2 e−j 2 π f ∆τij Pi (f ) Pj∗ (f ) df
W
(34)
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
(1,1)
(1,1)
(1,1)
= e−j 2 π fc ∆τij ,
= τim1 − τjm1 , ξij
∗
b̃ij (f ) = b̃i (f ) b̃∗j (f ), and b̃cij = b̃ci b̃cj . In (33), the derivatives translate the TOA and DOA in position and orientation
information. In particular, for the position we have
1
c ∇p (τ1 ) + ∇p (θ1 ) pti ϑt − prm (ϑr ) .
∇p (τim1 ) =
c
(35)
with ∆τij
The term ∇p (τ1 ) expresses the dependence of the position
from the direct path TOA; while
T
cos(φ1 )
∇p (θ1 ) =∇p (θ1 ) cos(θ1 ) sin(φ1 )
− tan(φ1 )
T
− sin(φ1 )
+ ∇p (φ1 ) sin(θ1 ) cos(φ1 )
(36)
0
includes the dependence of the position from the DOA information. Finally, for what the orientation information is
regarded, we have
1
∇ϑt (τim1 ) = ∇ϑt (τit θ1 , ϑt ) = d (θ1 ) ∇ϑt pti ϑt .
c
(37)
By further analyzing (33), one can notice the dependence
of the FIM from the beamforming weights given by the
coefficients b̃cij and b̃ij (f ).
Given the FIM in (33) and starting from (32), it can be easily
found that for beamforming and MIMO it is
−1 FS −1
CRBFS (q) = JFS
= J̆qq G
(38)
qq
where we have separated the effect of signal design J̆FS
qq , i.e.,
that related to (34), from that of the geometry G, i.e., that
related to (35)-(37). Specifically for timed arrays, we have
X
2
2
2
∇qq (τim1 , τjm1 )
J̆FS
qq = 8π SNR1 β + fc , G =
mij
(39)
where ∇qq (τim1 , τjm1 ) is a 5 × 5 matrix whose entries
are given by ∇qa (τim1 ) ∇qb (τjm1 ), and β is the baseband
effective bandwidth of p(t), defined as
Z
12
2
2
β=
f |P (f )| df
.
(40)
W
Similarly, for MIMO arrays, it is possible to find
X
2
2
2
∇qq (τim1 , τim1 )
J̆FS
qq = 8π SNR1 βi + fc , G =
mi
(41)
P
PNrx PNtx
2
β
2
where
mi =
m=1
i=1 and βi = Ntx is the squared
baseband effective bandwidth of pi (t).
The matrix G provides, through derivatives, the relationship
between the TOA at each TX-RX antenna element couple and
the Tx position and orientation.
To improve the comprehension of (38)-(39), in the next
sections two particular cases of planar MIMO and timed arrays
will be discussed considering a fixed Tx and Rx orientation
9
T
i.e., ϑt = ϑr = [0, 0] . Note that the overall CRB analysis
is still valid for any orientation. In Secs. VI-1 and VI-2, we
choose a specific case just to provide some insights on how
the number of transmitting and receiving antennas can impact
the performance. In Appendix C, the matrix G is evaluated,
considering this specific array geometry.
1) Special Case: Planar MIMO Array: For the planar
geometric configuration and in the orientation-unaware case,
the diagonal elements in the position and orientation CRB
matrix derived starting from (38)-(39) and from (55), are given
by
CRB (x) = CRB (z) = CRB0
CRB (y) =
CRB0
Nrx
12
S (Nrx − 1)
CRB ϑt = CRB ϕt = CRB0
12 (Ntx + Nrx − 2)
(42)
Arx (Ntx − 1)(Nrx − 1)
where CRB0 = c2 / 8π 2 SNRt βi2 + fc2 is the CRB of the
ranging error one would obtain using single antenna, and
S = Ar /y 2 represents the ratio between the Rx array area
and the squared Tx-Rx distance. Note that CRB0 depends on
the carrier frequency fc , on the shape of the pulse through βi2 ,
on the received SNR, and it does not depend on the number
of transmitting antennas. The analytical derivation is reported
in Appendix C. From (42), it is possible to remark that the
CRB of the estimation error in the y-coordinate is inversely
proportional to the number of the receiving antenna elements
accounting for the number of independent measurements available at the Rx. Regarding the other two coordinates, a key
parameter on the estimation accuracy is S which is related
to the ratio between the dimension of the Rx array and the
distance between the arrays: as this ratio becomes smaller
(i.e., as the distance between the arrays becomes larger with
respect to the array size), the positioning accuracy degrades.
From (42) it is also possible to notice that the accuracy in
estimating the orientation depends both on the transmitting
and receiving antennas. Specifically both Ntx and Nrx must
be greater than one to make the orientation possible, whereas
for the positioning, the constraint is only on the number of
receiving elements that must be larger than 1. Moreover, nonzero off-diagonal elements remark a correlation between the
error on the estimation of position and orientation. Specifically
we have
CRB z, ϑt = CRB ϑt , z = CRB x, ϕt = CRB ϕt , x
12
.
(43)
= CRB0
S y (1 − Nrx )
Contrarily in the orientation aware case, it can be found
CRB (x) = CRB (z) = CRB0
CRB (y) =
CRB0
.
Nrx
12
S (Ntx + Nrx − 2)
(44)
Note that when passing from a condition of orientationunawareness to that of orientation-awareness the positioning
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
accuracy increases, thanks to the additional information provided. In fact, the CRB on x and z coordinates now depends
also on the number of transmitting antennas.
2) Special Case: Planar Timed Array: Differently from
MIMO, here in the orientation-unaware case, the equivalent
FIM for position and orientation is singular meaning that it is
not possible to jointly localize and determine the orientation
using beamforming strategies. Nevertheless, when multiple
beams are generated [24], [25], such singularity can be solved
thus allowing the localization process, but at the prize of
an increased scanning time if the beams are sequentially
generated in time, or of a decreased SNR if such beams are
simultaneously formed. The investigation of this trade-off is
out-of-the-scope of this paper.
If the Tx orientation is a known parameter (orientation
aware case) and it is discarded from the estimation parameters
vector, the elements of the position CRB matrix result from
(56)
1
12
CRB (x) = CRB (z) = CRB0
S Ntx (Nrx − 1)
CRB0
CRB (y) =
.
(45)
Ntx Nrx
From (45) it is possible to remark that the CRB of the
estimation error in the y-coordinate is inversely proportional
to Ntx and Nrx : in fact, the Ntx term accounts for the SNR
enhancement due to the beamforming process while the Nrx
term accounts for the number of independent measurements
available at the Rx (receiver diversity). Note that when Nrx =
1, the localization along the x and z axes is not possible (only
ranging in the y direction), as for MIMO. Refer to Appendix C
for more details related to the derivation of (45).
VII. M ULTIPATH EFFECT ON
LOCALIZATION ACCURACY
Once verified that the CRB is a meaningful metric in different propagation conditions in Sec. V-B, we now investigate
the impact of MPCs on the localization performance for the
considered scenario. In [18], it is demonstrated that only the
information related to the first-contiguous cluster, i.e. the set
of MPCs overlapped to the first path, is relevant from a
localization perspective in the asymptotic SNR regime. Here
we show that under the asymptotic massive antenna regime, all
the MPCs can be made negligible, included those belonging
to the first-contiguous cluster.
The FIM in presence of multipath can be written as follows
Jqq Jqκ
Jψ =
(46)
Jκq Jκκ
where Jκκ contains also the a-priori information on MPCs
statistics reported in Appendix A. Consequently, the CRB for
the multipath scenario can be formulated as
−1
(47)
CRB (q) = Jqq − Jqκ J−1
κκ Jκq
where all multipath information is gathered in Jqκ J−1
κκ Jκq .
Considering the average over different geometric configurations (e.g., the average over different Rx orientations) and for
large values of Nrx , it is possible to show that the number of
configurations where the multipath impacts the localization
10
performance compared to the number of configurations in
which it does not influence the accuracy is negligible regardless the array architecture chosen.
Considering (47), for the weak law of the large number (i.e.,
for Nrx → ∞), it holds
1
1
P
Jqκ −−→
E [Jqκ ]
Nrx Ntx
Nrx Ntx
(48)
where we aim at demonstrating that
1
E [Jqκ ] = 0 .
Nrx Ntx
(49)
In the presence of a large number of antenna elements and
considering random Rx orientations, the inter-antenna phase
terms can be modeled as RVs uniformly distributed in [0, 2π).
Under this assumption, we have
E [Jq a1 ] = Jq a1 = 0
io
h
i
h
X n
(k,1) (k,1)
c
ℑ
(1)
χ
b̃
E
ξ
=
−4
π
a
ν
E Jq αℜ
1
ij
ij
ij
k
mij
× ∇q (τjm1 ) = 0
io
h
h
i
X n
(k,1) (k,1)
c
ℜ
(1)
χ
b̃
E
ξ
E Jq αℑ
=
4
π
a
ν
1
ij
ij
ij
k
mij
× ∇q (τim1 ) = 0
(50)
where
i
h
r
(k,1) (k,1)
E ξij χij (1) ∝ E[e−j 2 π (f +fc ) (∆τm (θ1 ,θk )) ] = 0 (51)
r
r
r
with ∆τm
(θ1 , θk ) = τm
(θ1 , ϑr ) − τm
(θk , ϑr ). Following
similar considerations, it is straightforward to prove that the
expectation of the Jκq elements is zero.
The result in (48) leads to the important conclusion that letting
the antennas array be massive, i.e., large Nrx , makes the set
of geometric configurations significantly impacted by MPCs
negligible, and the performance converges to that of the free
space case. As a consequence, the CRB converges to the CRB
averaged over the RX orientations for massive antenna arrays.
VIII. N UMERICAL R ESULTS
In this section, numerical results are reported considering
different array schemes, multipath conditions and system nonidealities. Four array structures are analyzed: timed arrays
equipped with TDLs and PSs, phased and random weighting
arrays using only PSs, and finally the MIMO array in which
orthogonal waveforms are transmitted and neither PSs nor
TDLs are present. For what the antennas spatial deployment
is regarded, planar arrays are considered as they represent
the most conventional structure to be integrated in APs and
mobiles. Differently from Secs. VI-1 and VI-2, here we
consider the results averaged over the Rx orientations if not
otherwise indicated, and thus it will be possible to appreciate
the impact of the array rotational angle on the localization
performance. In the following figures, we indicate with Q
the presence of quantization errors, with S the presence of a
residual time synchronization error. Moreover, we designate
with:
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
Fixed orientation: the array configuration with the Tx and
T
the Rx parallel to each other (i.e., ϑt = ϑr = [0, 0] ), as
described in Sec. VI;
• Averaged orientation: the geometric configuration in
which, for each Monte Carlo iteration, a different 3D Rx
array orientation is generated, and the CRB is computed
as the average over all the partial CRB results computed
at each cycle.
Finally, we recall that:
• Orientation-aware indicates the case in which the Tx
orientation is known at the Rx side and, thus, it is not
considered in the parameters vector to be estimated;
• Orientation-unaware: the case in which the Tx orientation
is unknown at Rx side and it has to be estimated together
with the position. In the next figures, we will denote with
O, the orientation-unawareness case.
11
•
Timed - Fixed Orien.
Phased - Fixed Orien.
Timed - Q - Fixed Orien.
Phased - Q - Fixed Orien.
MIMO - Fixed Orien.
MIMO - Averaged Orien.
Timed - Averaged Orien.
Phased - Averaged Orien.
Phased - Q - Averaged Orien.
Timed - Q - Averaged Orien.
PEB [m]
10-1
10-2
10-3
10-4
4 10
20
30
40
50
60
70
80
90
100
Ntx
A. System Configuration
We consider a scenario with a single AP equipped with
a massive array, with the centroid placed in pr = [0, 0, 0]T ,
and a transmitting array antenna whose centroid is located in
pt = [0, 5, 0]T (d = 5 m).
As in the mathematical model, the Rx has a perfect knowledge of the Tx steering direction, and the results are obtained
for fc = 60 GHz and W = 1 GHz (the signal duration is
τp = 1.6 ns) in free-space and multipath conditions. Root
raised cosine (RRC) transmitted pulses centered at frequency
fc = 60 GHz and roll-off factor of 0.6 are adopted, being
compliant with Federal Communications Commission (FCC)
mask at 60 GHz [38]. A receiver noise figure of NF = 4 dB
and a fixed transmitted power of Pt = 10 mW are considered,
if not otherwise indicated.
The performance is evaluated in terms of Position Error
Bound (PEB) and Orientation Error Bound (OEB) averaged
over Ncycle = 500 Monte Carlo iterations. For each cycle, a
T
different 3D Rx array orientation , i.e., ϑr = [ϑr , ϕr ] , and
multipath scenario are generated. Specifically, the receiving
(transmitting) antennas are spaced apart of dant = λL /2, where
λL = c/fL and fL = fc − W/2. When present, the PSs
quantization errors are δit ∼ U (−π/4, π/4) while the TDLs
errors are ∆τit ∼ U (0, dant /c). The standard deviation of the
time synchronization error is set to σǫ = 1 ns.
When operating at mm-wave frequencies, the path arrival
time distributions can be described by a Poisson process and
the inter-arrival times by an exponential probability density
function [39]. The paths arrival rate is set to 4 [1/ns] while
the paths azimuth and elevation angle-of-arrival (AOA) are
modeled as uniformly distributed between (0, 2π] and (0, π],
respectively. Note that these values are also in line with those
found in [40], [41] where a mm-wave measurements campaign
using massive arrays for radar-based mapping purposes have
been described.
Before analyzing the MIMO and beamforming localization
performance, it is necessary to ensure that the comparison
based on CRB can be considered fair in terms of SNR working
regimes when operating with non-massive arrays. To this
purpose, a threshold in terms of SNR is derived in order to
Fig. 4. PEB vs. Ntx , Nrx = 25 and orientation aware.
understand if ambiguities are significant or not and, hence,
whether the CRB can still be used as a performance metric
for comparison. In this perspective, we still consider the AF as
a tool to investigate the performance of maximum likelihood
estimator (MLE) as a function of the probability of ambiguity,
i.e., secondary lobes higher than the main lobe. In fact, the AF
main lobe determines the mean square error (MSE) behaviour
in the high SNR region that is described by the CRB, while
the AF sidelobes might generate ambiguities in case of large
noise in the low SNR region and that are not taken into account
by the CRB. The details of the aforementioned analysis are
reported in Appendix D.
In our numerical results, for each tested configuration we
verified that the SNR level is above a threshold calculated to
guarantee that the probability of ambiguity is less than 10−2 .
B. Results
The results of this section have been obtained as a function
of Ntx and Nrx ; the array structure (i.e. timed, phased, random,
MIMO); the presence and absence of array beamforming and
of a residual time synchronization error; and the multipath
overlapping effect.
1) Free space scenario:
a) Results as a function of Ntx : Figure 4 reports the PEB
performance as a function of Ntx and of the Rx orientation in
free-space. MIMO, timed and phased arrays (with and without
quantization errors) are compared in the orientation-aware
case when the number of receiving antennas is kept fixed to
Nrx = 25.
It can be observed that MIMO arrays, relying on different
transmitted waveforms, outperform timed and phased arrays in
averaged orientations cases; whereas, in fixed Rx orientation,
as described in Sec. VI, arrays operating beamforming exhibit
a better performance. This is due to the fact that beamforming
strategies (e.g., those adopted in timed and phased arrays) fail
in preserving the same accuracy for any geometric configuration (i.e., for any Rx orientation). Contrarily, thanks to the
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
12
10-1
10-1
MIMO - O - Averaged Orien.
MIMO - O - Fixed Orien.
Timed
Phased
Phased - Q
MIMO
MIMO - O
Random
Timed - Q
10-2
OEB [deg.]
PEB [m]
10-2
MIMO - S - O - Averaged Orien.
10-3
10-3
10-4
20
40
60
80
100
10-4
4 10
20
30
40
Nrx
Fig. 5. PEB vs. Nrx , Ntx = 25 and fixed receiver orientations.
Fig. 7.
50
60
Nrx
70
80
90
100
OEB vs. Nrx , Ntx = 25.
100
MIMO
Timed
Phased
Phased - Q
Timed - Q
Random
MIMO - O
Timed - S
MIMO - S
MIMO - S - O
PEB [m]
10-1
10-2
10-3
4 10
20
30
40
50
60
70
80
90
100
Nrx
Fig. 6. PEB vs. Nrx , Ntx = 25 and averaged receiver orientation.
diversity gain characterizing MIMO arrays, Rx orientations
have a less significant effect on positioning accuracy.
For what the beamforming arrays are concerned, it can be
observed that, as expected, timed and phased arrays results
coincide as W/fc ≪ 1. In fact, phased arrays are the best
candidate to be adopted in narrowband systems where there
is no need to compensate for delays to perform beamsteering
operation. We refer the reader to our previous work [12] to
better appreciate the impact of the fractional bandwidth on the
timed/phased array performance.
Another important outcome from Fig. 4 is that array
quantization errors, once characterized, slightly affect the
localization performance. This implies that we can rely on
simpler array structures (i.e., using switches instead of multipliers) without severely affecting the performance. Finally,
with Ntx ≥ 25, the performance improvement becomes less
important if a sufficiently high Nrx is considered. This implies
that Ntx can be relaxed to shrink the array dimensions and to
enable the integration on mobiles [9]. Consequently, in the
following, the number of transmitting antennas will be fixed
to Ntx = 25.
b) Results as a function of Nrx : In Fig. 5, the PEB
performance are reported for both orientation-aware and unaware cases as a function of Nrx in free-space propagation
condition, Ntx = 25 and fixed Tx and Rx orientation ϑt =
T
ϑr = [0, 0] . The results have been obtained using the analytic
expressions (42), (44) and (45) and reveal that arrays performing beamforming outperform the performance of MIMO for
the particular steering and geometric configuration conditions
chosen, as already observed in Fig. 4. We ascribe the effect to
an increased SNR in the considered direction. Nevertheless,
with arrays operating single-beam beamforming, orientation
estimation is not always possible and consequently the FIM
results to be singular. Matrix singularity or ill-conditioning
are, here, synonymous of the impossibility to estimate the
position/orientation given the collected measurements.
Figure 6 shows the average PEB performance when the
Rx orientation randomly changes at each Monte Carlo iteration. For this analysis, we consider also random weighting,
quantization errors as well as time synchronization mismatch
between the Tx and the Rx. The positioning performance
are shown for the orientation-aware case if not otherwise
indicated. As in Figs. 4-5, the performance of timed and
phased arrays coincide due to the narrow fractional bandwidth
(i.e., W/fc ≈ 0.016). Differently from Fig. 5, in Fig. 6
MIMO achieves a higher positioning accuracy with respect to
arrays employing beamforming strategies due to the fact that
results are averaged over different Rx orientations. In fact, with
MIMO, a reduction in the received SNR is experienced, but
the number of independent measurements is maximized (i.e.,
Ntx Nrx ).
For what random weighting arrays are regarded, they share
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
13
100
20
0.03
L=1
L=2
L=5
18
14
0.02
y [m]
12
10
0.015
PEB [m]
0.025
16
10-1
8
0.01
6
4
10-2
0.005
2
0
-10
-5
0
x [m]
5
10
0
10-3
4
10
20
30
40
50
Nrx
20
0.03
10-1
L=1
L=2
L=5
18
0.025
16
14
y [m]
10
0.015
8
OEB [deg]
0.02
12
10-2
0.01
6
4
0.005
2
0
-10
-5
0
x [m]
5
10
0
Fig. 8. Phased (top) and MIMO (bottom) array PEB, Nrx = 100, Ntx = 25
and averaged receiver orientations.
the structure simplicity of phased arrays but they neither
perform beamforming nor achieve the diversity gain of MIMO,
and thus, the positioning accuracy results degraded with
respect to other structures. Nevertheless, if the localization
accuracy required by the application of interest is not so
stringent, they could be an interesting option to guarantee both
a sub-centimeter positioning accuracy (e.g., for Nrx = 50 and
Ntx = 25, PEB ≈ 7 mm) and an easy implementation in future
mobiles and AP operating at mm-wave frequencies.
Note that when the Tx orientation is one of the parameters to
be estimated (orientation-unaware case), only MIMO results
in a non-singular FIM. Obviously, in this case, given the
reduced information available at the Rx side, the positioning
accuracy worsen with respect to the orientation-aware case.
In all cases, the residual time synchronization error degrades
the localization performance.
Figure 7 reports the OEB as a function of Nrx . In this
case, only the performance of MIMO is present because of
the singularity problem arising in timed, phased and random
weighting arrays. An interesting results is that, in this case,
time synchronization error does not impact the orientation
accuracy.
c) Grid results: Figure 8 reports the PEB results for the
orientation-aware case when the mobile moves in a grid of
points spaced of 0.5 m considering phased (Fig. 8-top) and
MIMO (Fig. 8-bottom) arrays, respectively. We considered a
10-3
4
10
20
30
40
50
Nrx
Fig. 9. PEB and OEB vs. Nrx in multipath propagation scenario, Ntx = 25,
W = 1 GHz, averaged receiver orientation. Diamond marked lines refer to
phased array, circle marked lines to MIMO, and square marked lines to MIMO
orientation unaware.
3D indoor scenario of 10×10×3 m3 where the mobile and the
AP centroids are at the same height. The Rx, equipped with
T
Nrx = 100 antennas, is placed in pr = [0, 0, 0] with orientation changing at each Monte Carlo iteration. On the other side,
the mobile array is equipped with Ntx = 25 and its orientation
and its steering direction are fixed to ϑt = [0, 0]T and to
the broadside direction, respectively. Grid results confirm that
MIMO arrays localization performance does not depend on
the Rx orientation and on mobile position in space but only
on the distance between the Tx and the Rx. Indeed, when
comparing Fig. 8-top and Fig. 8-bottom, it can be seen that
if the mobile steering is fixed, the localization accuracy is
higher in a privileged direction in space corresponding to the
best geometric configuration conditions.
2) Multipath scenario:
a) Results as a function of Nrx : Figure 9 investigates the
multipath effect by analysing the PEB and OEB averaged over
different Rx orientations and as a function of the number of
MPCs for phased and MIMO arrays. We consider the statistical
multipath parameters described in Sec. VIII-A. As foreseen in
the asymptotic analysis in Sec. VII, when increasing Nrx , the
MPCs effect becomes negligible and the performance tends
to coincide with that obtained in free-space. Moreover, it is
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
possible to remark that phased arrays are more sensitive to
multipath with respect to MIMO arrays, especially when the
number of receiving antennas is small. In fact, for phased
arrays at least Ntx = 25 and Nrx = 25 are necessary to make
the MPCs fully resolvable. In any case, it is interesting to note
that to make the impact of MPCs negligible the antenna arrays
are not required to be significantly massive.
IX. C ONCLUSION
In this paper, we have considered a new scenario where a
single-anchor localization exploiting mm-wave massive arrays
has been put forth for next 5G applications. The theoretical
localization performance has been evaluated by deriving the
position and orientation CRB for different array configurations. Phase quantization, the residual time synchronization
mismatch, and multipath have been considered as nuisance
parameters. The comparison between MIMO and beamforming has been carried out by analyzing the impact of the number
of antenna elements on the performance. From analytical and
simulation results, the main conclusions and achievements
emerged in this paper can be summarized as follows:
• We show through an asymptotic analysis (i.e., Nrx → ∞)
that the considered CRB is a tight bound regardless
the propagation condition and the array configuration
considered.
• A beamforming with steering capability is desirable to
maximize the SNR towards a specific direction, as happens for phased and timed arrays; but it reduces the
diversity gain (typical of MIMO) exploitable for retrieving positioning information, especially in orientationunaware situations.
• Quantization errors slightly impact the localization performance of both timed and phased arrays. Consequently,
the array design requirements can be relaxed in favor of
lower complexity and cost.
• Both for MIMO and beamforming, a time synchronization mismatch between the Tx and the Rx significantly
degrades the positioning performance. From this pointof-view, it is important to minimize the clock mismatch
between the arrays. Contrarily, in the orientation-unaware
case, the time synchronization error does not affect the
performance of the estimation;
• The adoption of multiple antennas makes the positioning
insensitive to multipath for most of geometric configurations. This is true even when the number of antennas is
not extremely high (i.e., Ntx , Nrx > 20); we demonstrated
such point also through an asymptotic analysis.
• Finally, random weighting results to be a good lowcomplexity strategy to achieve centimeter level accuracy.
A PPENDIX A
Considering the sub-set
of random
parameters ψr =
T
T
[κT2 , . . . , κTL , ǫs ] = αℜ T , αℑ T , ǫs , and treating them as
independent RVs we can write
ln (f (ψr )) = ln f (αℜ T ) + ln f (αℑ T ) + ln (f (ǫs )) .
(52)
Therefore, the elements of the a-priori FIM are
14
Jǫps ǫs
1
= 2,
σǫ
Jαp ℜ αℜ
k l
=
Jαp ℑ αℑ
k l
=
(
1
σl2
if l = k
0
if l 6= k
.
(53)
A PPENDIX B
In this Appendix we derive the elements of the data FIM
reported in (19). To accomplish this task, we introduce the
following quantities
Z
(l,k)
(l,k)
p
χij (p)= b̃ij (f ) (f + fc ) e−j 2 π f ∆τij Pi (f ) Pj∗ (f )df
W
Z
p
b̃ij (f ) e−j 2 π f ∆τ Pi (f ) Pj∗ (f ) df
Rij (∆τ ) =
W
Z
p̈
Rij
(∆τ ) =
b̃ij (f ) f 2 e−j 2 π f ∆τ Pi (f ) Pj∗ (f ) df (57)
W
(l,k)
where ∆τij = τiml − τjmk , b̃ij (f )
∗
b̃ci b̃cj . The elements of Jdqq can
The elements of Jdκκ are
Ja 1 a 1 = 2 ν
X
mij
be expressed as in (33).
o
n
(1,1)
(1,1) p
∆τij
ℜ b̃cij ξij Rij
= JaH1 αℜ = 2 ν
Jαℜ
k a1
k
X
mij
= JaH1 αℑ = 2 ν
Jαℑ
k a1
k
o
n
(1,k)
(1,k) p
∆τij
ℜ b̃cij ξij Rij
X
n
o
(1,k) p
(1,k)
ℑ b̃cij ξij Rij
∆τij
X
ℜ
n
X
ℑ
n
mij
Jαℜ
ℜ = Jαℑ αℑ = 2 ν
k αl
k
l
mij
Jαℑ
= JαH ℜ αℑ = 2 ν
αℜ
k
l
l
= b̃i (f ) b̃∗j (f ), and b̃cij =
k
mij
(l,k)
b̃cij ξij
(l,k)
b̃cij ξij
o
(l,k)
p
∆τij
Rij
o
(l,k)
p
∆τij
Rij
s
r
t
(l,k)
(1,k)
= e−j 2 π fc (τim1 +ǫ +τm (θk )−τj (θk )) and ξij =
where ξij
r
r
t
t
e−j 2 π fc (−τm (θl )+τm (θk )+τi (θl )−τj (θk )) .
The elements of Jdqκ are
o
X n
(1,1) (1,1)
ℑ b̃cij ξij χij (1) ∇q (τim1 ) = 0
Jq a 1 = 4 π a 1 ν
mij
Jq αℜ
= −4 π a1 ν
k
Jq αℑ
= 4 π a1 ν
k
X
mij
X
mij
ℜ
ℑ
n
n
o
(k,1) (k,1)
χij (1)
b̃cij ξij
o
(k,1) (k,1)
χij (1)
b̃cij ξij
∇q (τjm1 )
∇q (τim1 ) (58)
= e−j 2 π fc (−τjm1 −ǫ −τm (θk )+τi (θk )) . Note that
where ξij
d
dH
Jκq = Jqκ .
Now, if we consider the presence of a residual time synchronization error, the FIM Jdǫs ǫs is
X
h
(1,1) (1,1)
b̃cij a21 ξij χij (2) +
Jǫ s ǫ s = 8 π 2 ν ℜ
mij
#)
L
X
t
t
(k,k)
+
σ 2 Rp̈ ∆τ
e−j 2 π fc (τi (θk )−τj (θk ))
.
s
(k,1)
l
k=2
ij
ij
r
t
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
d2
∇qq (τim1 , τjm1 )= ant 2
(c y)
(mx − ix ) (mx − jx )
...
...
...
...
y
dant (mx
y2
d2ant
− ix ) (mx − ix )(jz − mz )
1
c2
CRB (q)= 2
8π Ntx SNR1 (βi2 + fc2 ) S
12
(Nrx −1)
0
0
0
1
c2
diag
2
8π Ntx SNR1 (β 2 + fc2 ) S
Jαℜ
s = 4 π a1 ν
k ǫ
X
mij
Jαℑ
s = 4 π a1 ν
k ǫ
X
mij
0
S
Nrx
0
0
y (ix − mx ) jz . . . y (ix − mx ) jx
2
− dyant jz
y (mz − iz ) jz
y 2 iz jz
...
0
12
(Nrx −1)
12
y (1−Nrx )
0
0
12
y (1−Nrx )
12 (Nrx +Ntx −2)
y 2 (Ntx −1) (Nrx −1)
0
n
o
(1,k) (1,k)
ℑ b̃cij ξij χij (1)
n
o
(1,k) (1,k)
ℜ b̃cij ξij χij (1) .
As before the elements of Jdǫs κ could be found as Jdǫs κ = JdκǫHs
Finally, the elements of Jdqǫs are
o
X n
(1,1) (1,1)
ℜ b̃cij ξij χij (2) ∇q (τjm1 )
Jqǫs = 8 π 2 ν a21
mij
and Jdǫs q = JdqǫHs .
2
− dyant jx
y (mz − iz ) jx
y 2 iz jx
2
y ix jx
(54)
12
y(1−Nrx )
0
0
12
S
12
,
,
Ntx (Nrx − 1) Ntx Nrx Ntx (Nrx − 1)
The elements of Jdκǫs are
o
X n
(1,1) (1,1)
ℑ b̃cij ξij χij (1)
Ja 1 ǫ s = 4 π a 1 ν
mij
− mz )
(iz − mz ) (jz − mz )
...
...
0
12
y(1−Nrx )
CRB pt =
y
dant (jz
...
...
...
15
0
0
0
12 (Nrx +Ntx −2)
y 2 (Ntx −1) (Nrx −1)
(55)
(56)
TABLE II
MIMO vs. BEAMFORMING COMPARISON
MIMO/Timed
Nrx
γ [dB]
σthr [mV]
σsim [mV]
MIMO
4
-36.9
0.062
0.022
MIMO
36
-32.1
0.187
0.022
MIMO
100
-29.9
0.313
0.022
Phased
4
-33.5
0.136
0.022
Phased
36
-28.7
0.406
0.022
Phased
100
-26.5
0.677
0.022
A PPENDIX C
In this Appendix we will specialize the expression of the
symmetric matrix G reported in (38)-(39). To this end, we
explicit the geometric relationship relating the TOA between
each TX-RX antennas couple and the considered localization
(position or orientation) parameter, i.e., ∇qa qb (τiml , τjml ) =
∇qa (τiml )∇qb (τjml ). For the particular antenna configuration
chosen described in Sec. II-B, in which the array antennas are
spaced of dant , and considering ϑt = [0, 0]T , we can compute
a simplified version of (35)-(37). Specifically, it is possible to
obtain:
1
∇p (τim1 ) = [c ∇p (τ1 ) + dant ((ix − mx )∇p (φ1 )
c
+(mz − iz )∇p (θ1 ))]
(59)
dant
dant
∇ϑt (τim1 ) = −
iz , ∇ϕt (τim1 ) = −
ix
(60)
c
c
√
√
√
−1
−1
, − N2rx√−1 + 1, . . . , N2rx√
and
with mx = mz = − N2rx√
ix = iz = jx = jz = − N2tx −1 , − N2tx −1 + 1, . . . , N2tx −1 .
From (59)-(60), it is straightforward to derive (54). Then,
by considering the summations present in G, it is possible
to obtain the CRB matrices for MIMO and timed arrays
respectively as in (55)-(56) where S = Ar /y 2 .
A PPENDIX D
In this Appendix we consider the AF as defined in (22).
The AF for the position coordinates shows a main peak
in correspondence of the true Tx position and secondary
sidelobes peaks relative to “wrong” positions. An ambiguity
problem arises when one of these sidelobes overcomes or
becomes comparable to the main beam due to noise. Consequently, to determine whether ambiguities are negligible in
the non-massive array case, we have derived a threshold on
the noise standard deviation in order to keep the ambiguity
probability fixed to a desired low value.
By comparing the threshold obtained with the value used
in the numerical results, we can demonstrate that we operate
at a high SNR regime where the CRB is tight even if a nonmassive array is adopted.
To this end, we define the ambiguity probability as [35]
γ
1
(61)
PA = erfc √
2
4 σ2
where σ is the noise standard deviation and γ is the gap
between the main lobe of the AF and the highest secondary
SUBMITTED FOR REVIEW TO IEEE TRANSACTIONS ON SIGNAL PROCESSING
sidelobe. Then, given a certain gap γ, it is possible to compute
the noise threshold as
γ
1
.
(62)
σthr =
−1
2 erfc (2 P∗A )
In Table II, we report the obtained simulation results. We have
considered the Tx moving in a grid of points spaced apart of
0.2 m in a cube of dimension 8×8×8 m3. The target ambiguity
probability has been fixed to 10−2 . The gap γ has been set
to the minimum side-lobe level considering the three spatial
coordinates (that is, to the worst case scenario). σsim represents
the noise standard deviation used in the numerical results of
the paper.
As one can notice, in all the tested configurations the noise
standard deviation used in the numerical results is always
much lower than the threshold σthr above that the ambiguity
effect is not anymore negligible.
The proposed method is a useful tool to test whether a
specific scenario can be considered in the asymptotic regime
and, hence, the CRB can be a meaningful metric.
R EFERENCES
[1] W. H. Chin, Z. Fan, and R. Haines, “Emerging technologies and research
challenges for 5G wireless networks,” IEEE Wireless Commun., vol. 21,
no. 2, pp. 106–112, 2014.
[2] V. Khaitan, P. Tinnakornsrisuphap, and M. Yavuz, “Indoor positioning
using femtocells,” in Proc. IEEE Veh. Technol. Conf. (VTC Fall), 2011,
pp. 1–5.
[3] E. G. Larsson et al., “Massive mimo for next generation wireless
systems,” IEEE Commun. Mag., vol. 52, no. 2, pp. 186–195, 2014.
[4] F. Rusek et al., “Scaling up MIMO: Opportunities and challenges with
very large arrays,” IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–60,
2013.
[5] A. L. Swindlehurst et al., “Millimeter-wave massive MIMO: The next
wireless revolution?” IEEE Commun. Mag., vol. 52, no. 9, pp. 56–62,
2014.
[6] R. Di Taranto et al., “Location-aware communications for 5G networks:
How location information can improve scalability, latency, and robustness of 5G,” IEEE Signal Process. Mag., vol. 31, no. 6, pp. 102–112,
2014.
[7] F. Guidi, A. Guerra, and D. Dardari, “Personal mobile radars with
millimeter-wave massive arrays for indoor mapping,” IEEE Trans.
Mobile Comput., vol. 15, no. 6, pp. 1471–1484, 2016.
[8] D. Dardari, P. Closas, and P. M. Djurić, “Indoor tracking: Theory,
methods, and technologies,” IEEE Trans. Veh. Technol., vol. 64, no. 4,
pp. 1263–1278, 2015.
[9] W. Hong et al., “Study and prototyping of practically large-scale
mmwave antenna systems for 5G cellular devices,” IEEE Commun.
Mag., vol. 52, no. 9, pp. 63–69, 2014.
[10] S. M. Razavizadeh, M. Ahn, and I. Lee, “Three-dimensional beamforming: A new enabling technology for 5G wireless networks,” IEEE Signal
Process. Mag., vol. 31, no. 6, pp. 94–101, 2014.
[11] K. Witrisal et al., “High-accuracy localization for assisted living: 5G
systems will turn multipath channels from foe to friend,” IEEE Signal
Process. Mag., vol. 33, no. 2, pp. 59–70, 2016.
[12] A. Guerra, F. Guidi, and D. Dardari, “Position and orientation error
bound for wideband massive antenna arrays,” in Proc. IEEE Int. Conf.
on Commun. Workshop (ICCW), 2015, pp. 853–858.
[13] N. Garcia et al., “Direct localization for massive MIMO,” arXiv preprint
arXiv:1607.00946, 2016.
[14] H. Kaouach et al., “Wideband low-loss linear and circular polarization
transmit-arrays in V-band,” IEEE Trans. Antennas Propag., vol. 59,
no. 7, pp. 2513–2523, 2011.
[15] A. Clemente et al., “Wideband 400-element electronically reconfigurable
transmitarray in X band,” IEEE Trans. Antennas Propag., vol. 61, no. 10,
pp. 5017–5027, 2013.
[16] F. Guidi et al., “Joint energy detection and massive array design for
localization and mapping,” IEEE Trans. Wireless Commun., vol. 16,
no. 3, pp. 1359–1371, March 2017.
16
[17] Y. Shen and M. Z. Win, “On the accuracy of localization systems using
wideband antenna arrays,” IEEE Trans. Commun., vol. 58, no. 1, pp.
270–280, 2010.
[18] ——, “Fundamental limits of wideband localizationpart I: A general
framework,” IEEE Trans. Inf. Theory, vol. 56, no. 10, pp. 4956–4980,
2010.
[19] Y. Han et al., “Performance limits and geometric properties of array
localization,” IEEE Trans. Inf. Theory, vol. 62, no. 2, pp. 1054–1075,
2016.
[20] A. Shahmansoori et al., “5G position and orientation estimation through
millimeter wave MIMO,” in Proc. IEEE Global Conf. on Commun.
(GLOBECOM) Workshops, 2015, pp. 1–6.
[21] A. Mallat and L. Vandendorpe, “CRBs for the joint estimation of TOA
and AOA in wideband MISO and MIMO systems: comparison with
SISO and SIMO systems,” in Proc. IEEE Int. Conf. on Commun., 2009,
pp. 1–6.
[22] A. Guerra, F. Guidi, and D. Dardari, “On the Impact of Beamforming
Strategy on mm-Wave Localization Performance Limits,” in Proc. IEEE
Int. Conf. on Commun. Workshop (ICCW).
[23] T. M. Cover and J. A. Thomas, Elements of information theory. John
Wiley & Sons, 2012.
[24] A. Alkhateeb et al., “Channel estimation and hybrid precoding for
millimeter wave cellular systems,” IEEE J. Sel. Topics Signal Process.,
vol. 8, no. 5, pp. 831–846, 2014.
[25] N. Garcia et al., “Location-aided mm-wave channel estimation for
vehicular communication,” in 2016 IEEE 17th Int. Workshop on Signal
Process. Advances in Wireless Commun. (SPAWC). IEEE, 2016, pp.
1–5.
[26] D. Dardari et al., “Ranging with ultrawide bandwidth signals in multipath environments,” Proc. IEEE, vol. 97, no. 2, pp. 404–426, 2009.
[27] J. Li and P. Stoica, “MIMO radar with colocated antennas,” IEEE Signal
Process. Mag., vol. 24, no. 5, pp. 106–114, 2007.
[28] E. Fishler et al., “Spatial diversity in radars-models and detection
performance,” IEEE Trans. Signal Process., vol. 54, no. 3, pp. 823–838,
2006.
[29] A. Hassanien and S. A. Vorobyov, “Phased-MIMO radar: A tradeoff
between phased-array and MIMO radars,” IEEE Trans. Signal Process.,
vol. 58, no. 6, pp. 3137–3151, 2010.
[30] Q. He et al., “Target velocity estimation and antenna placement for
MIMO radar with widely separated antennas,” IEEE J. Sel. Topics Signal
Process., vol. 4, no. 1, pp. 79–100, 2010.
[31] A. M. Haimovich, R. S. Blum, and L. J. Cimini, “MIMO radar with
widely separated antennas,” IEEE Signal Process. Mag., vol. 25, no. 1,
pp. 116–129, 2008.
[32] A. F. Molisch, Wireless communications. John Wiley & Sons, 2007.
[33] H. Godrich, A. M. Haimovich, and H. V. Poor, “An analysis of
phase synchronization mismatch sensitivity for coherent MIMO radar
systems,” in Proc. 3rd IEEE Int. Workshop Comput. Adv. in Multi-Sensor
Adaptive Process. (CAMSAP), 2009, pp. 153–156.
[34] K. Witrisal and P. Meissner, “Performance bounds for multipath-assisted
indoor navigation and tracking (MINT),” in Proc. IEEE Int. Conf.
Commun. (ICC), 2012, pp. 4321–4325.
[35] H. L. Van Trees, Detection, estimation, and modulation theory. John
Wiley & Sons, 2004.
[36] G. San Antonio, D. R. Fuhrmann, and F. C. Robey, “MIMO radar
ambiguity functions,” IEEE J. Sel. Topics in Signal Process., vol. 1,
no. 1, pp. 167–177, 2007.
[37] A. W. Van der Vaart, Asymptotic statistics. Cambridge university press,
2000, vol. 3.
[38] FCC, “Revision of Part 15 of the Commissions Rules Regarding
Operation in the 57-64 GHz Band. August 2013.”
[39] C. Gustafson et al., “On mm-wave multipath clustering and channel
modeling,” IEEE Trans. Antennas Propag., vol. 62, no. 3, pp. 1445–
1455, 2014.
[40] A. Guerra et al., “Delay spread characterization of millimeter-wave
indoor backscattering channel,” in Proc. 10th European Conf. Antennas
and Propag. (EuCAP), 2016, pp. 1–2.
[41] F. Guidi, A. Clemente, and R. D’Errico, “Multipath estimation technique for wideband mm-wave backscattering channels,” in Proc. 11th
European Conf. Antennas and Propag. (EuCAP). IEEE, 2017.
| 7cs.IT
|
LOW DIMENSIONAL BOUNDED COHOMOLOGY AND EXTENSIONS OF
GROUPS
arXiv:1703.08802v1 [math.GR] 26 Mar 2017
NICOLAUS HEUER
Abstract. Let G, N be groups and let ψ : G → Out(N ) be a homomorphism. An extension
for these data is a group E which fits into the short exact sequence 1 → N → E → G → 1
such that the action of E on N by conjugation induces ψ. It is classical that extensions can
be characterised using ordinary group cohomology in dimension 2 and 3. We prove that an
analogous statement is true in bounded cohomology if we restrict to extensions which admit a
quasihomomorphic section.
1. Introduction
Let 1 → N → E → G → 1 be a short exact sequence of groups. For any set-theoretic
section s : G → E, the map φ : G → Aut(N ) defined via φ(g) : n 7→ s(g)ns(g)−1 induces a
homomorphism ψ : G → Out(N ) independent of s. In this case, we say that E is an extension
of G by N inducing ψ. Two extensions E1 and E2 of G by N which induce ψ are equivalent
(E1 ∼ E2 ) if there is an isomorphism I : E1 → E2 that commutes with the maps of the short
exact sequences i.e. if
1
N
id
1
N
E1
I
E1
G
1
.
id
G
1
is a commutative diagram. The relation ∼ imposes an equivalence relation among such extensions. We write E(G, N, ψ) for the set of all extensions of G by N inducing ψ : G → Out(N ) under
this equivalence relation. It is classical that this set can be fully characterised using ordinary
cohomology (cf. Theorem 6.6 of [Bro82]):
Theorem. Let G and N be groups and let ψ : G → Out(N ) be a homomorphism. There is a
cocycle o(G, N, ψ) ∈ C 3 (G, Z) such that [o(G, N, ψ)] = 0 in H 3 (G, Z) if and only if E(G, N, ψ) 6=
∅, where Z = Z(N ). In this case there is a bijection between H 2 (G, Z) and E(G, N, ψ).
Here Z(N ) denotes the centre of N . The aim of this paper is to show the following analogous
result for bounded cohomology:
Theorem 1.1. Let G and N be groups such that Z = Z(N ) is finitely generated and let ψ : G →
Out(N ) be a homomorphism with finite image. There is a cocycle ob (G, N, ψ) ∈ Cb3 (G, Z) such
that [ob (G, N, ψ)] = 0 in Hb3 (G, Z) if and only if Eb (G, N, ψ) 6= ∅ and [ob (G, N, ψ)] = 0 in
H 3 (G, Z) if and only if E(G, N, ψ) 6= ∅. If Eb (G, N, ψ) 6= ∅, there is a bijection from H 2 (G, Z)
to E(G, N, ψ) which restricts to a bijection from im(c2 ) ⊂ H 2 (G, Z) to Eb (G, N, ψ).
Here Eb (G, N, ψ) ⊂ E(G, N, ψ) is the subset of extensions of G by N inducing ψ consisting of
all extensions E such that there is a section f : G → E which is a quasihomomorphism in the sense
of Fujiwara and Kapovich (see Definition 1.2) and which induces a finite map φ : G → Aut(N )
via conjugation in E. The obstruction ob (G, N, ψ) is defined in the proof of Theorem 1.1 in
1
Section 2, Equation (1). The conditions of Theorem 1.1 may seem artificial but are natural as
we will see at the end of Subsection 1.2.
This paper is organised as follows: Subsections 1.1 and 1.2 introduce bounded cohomology
and quasihomomorphisms respectively. In Section 2 we will prove Theorem 1.1. In Section 3
we characterise the set of classes arising as obstructions ob as in Theorem 1.1 and give examples
to show that the assumptions of the theorem are necessary. For the sake of readability some
calculations will be postponed to the Appendix (Section 4).
1.1. Bounded Cohomology. Let G be a group, let R be Z or R and let R[G] be the associated
group ring. Let V be a normed R[G]-module, i.e. a module equipped with a function k · k : V →
R+ , the norm which is positive definite and satisfies the triangle inequality. Furthermore, it
satisfies that the action of G on V is an isometry and that kr · vk ≤ |r| · kvk for all r ∈ R
and v ∈ V . Throughout this paper V will be finitely generated as a Z module. For k ≥ 1,
let C k (G, V ) be the set of V -valued functions on Gk and Cbk (G, V ) ⊂ C k (G, V ) the subset of
bounded functions with respect to the supremum norm. We set C 0 (G, V ) = Cb0 (G, V ) = V . The
function δ k : C k (G, V ) → C k+1 (G, V ) defined via
δ 0 (v) : g1 7→ g1 · v − v
n
δ (f ) : (g1 , . . . gn+1 ) 7→ g1 · f (g2 , . . . , gn+1 )
n
X
(−1)i f (g1 , . . . , gi gi+1 , . . . , gn+1 )
+
i=1
+(−1)n+1 f (g1 , . . . , gn )
for n ≥ 1 is called coboundary operator. This operator restricts to a map δ k : Cbk (G, V ) →
Cbk+1 (G, V ) for k ≥ 0. It is classical that (C ⋆ (G, V ), δ ⋆ ) is a cochain complex. We define the
(ordinary) cohomology of G with V coefficients to be the homology of this cochain complex and
denote it by H ⋆ (G, V ). Similarly we define the bounded cohomology of G with V coefficients to
be the homology of the chain complex (Cb⋆ (G, V ), δ ⋆ ), and denote it by Hb⋆ (G, V ). Note that
the inclusion Cb⋆ (G, V ) ֒→ C ⋆ (G, V ) commutes with the coboundary operator and hence induces
a well defined map c⋆ : Hb⋆ (G, V ) → H ⋆ (G, V ), called the comparison map. For a thorough
treatment of ordinary and bounded cohomology see [Bro82] and [Fri16].
It can be seen from the classical version of Theorem 1.1 that the ordinary cohomology in
dimensions 2 and 3 can be interpreted in terms of group extensions. This paper provides an
analogous interpretation for bounded cohomology. The cohomology of a group G in dimension
1 can be seen to be the group of homomorphisms from G to the coefficients V under pointwise
addition. Similarly, the first bounded cohomology with V coefficients is the group of bounded
homomorphisms G → V . Hence, if V is torsion free, the first bounded cohomology is trivial.
1.2. Quasihomomorphisms. A function f : G → R where R = Z, R satisfying that there is a
constant C ≥ 0 such that
|δ 1 f (g, h)| = |f (g) + f (h) − f (gh)| ≤ C
for all g, h ∈ G is called quasimorphism. The smallest C satisfying the above inequality is called
defect of φ and is denoted by D(φ). Note that the set of quasimorphisms form a module under
pointwise addition and scalar multiplication. Quasimorphisms have multiple applications, are
well studied and understood, and intimately related to bounded cohomology. Homomorphisms
and bounded functions are quasimorphisms and the module generated by both is called the
space of trivial quasimorphisms. The quotient of the space of quasimorphisms by the space
of trivial quasimorphisms can be identified with ker(c2 ), the kernel of the comparison map
2
c2 : Hb2 (G, R) → H 2 (G, R) where the coeficients have a trivial G action. For a thorough treatment
of quasimorphisms see [Cal09].
It is not obvious how to generalise quasimorphisms to maps where the image is not commutative. This paper exclusively deals with the generalisation of Fujiwara and Kapovich ([FK16]).
However, we note that there are other generalisations, for example one by Hartnick and Schweitzer
([HS16]). These are considerably more general than the one we are concerned with; see Subsection 3.2.
Definition 1.2. (Fujiwara, Kapovich [FK16]) Let G and H be arbitrary groups. A function
f : G → H is called quasihomomorphism if the defect set
D(f ) = {d(g, h)|g, h ∈ G}
is finite, where d(g, h) = f (g)f (h)f (gh)−1 . The group ∆f < H generated by D(f ) is called the
defect group.
When there is no danger of ambiguity we will write D = Df and ∆ = ∆f . This definition
slightly differs from the original definition; see Remark 1.5. However, it is easier to work with in
the context of group extensions and equivalent. Just like in the case of ordinary quasimorphisms,
both homomorphisms and bounded functions f : G → H are quasihomomorphisms. Similarly, for
any cyclic subgroup C < H and any ordinary quasimorphism f : G → Z the map G → H defined
by precomposing f with the identification Z ∼
= C is (trivially) a quasihomomorphism. Fujiwara
and Kapovich showed that if the target group H is torsion free torsion free and hyperbolic these
are in fact all possible quasihomomorphisms. We recall basic properties of quasihomomorphisms.
Proposition 1.3. Let f : G → H be a quasihomomorphism, D and ∆ be as above and H0 < H
the subgroup of H generated by f (G). Then ∆ is normal in H0 . The function φ : G → Aut(∆)
defined via φ(g) : δ 7→ f (g)δf (g)−1 has finite image and it’s quotient ψ : G → Out(∆) is a
homomorphism with finite image. Moreover, the pair (d, φ) satisfies
φ(g)
d(h, i) d(g, hi) = d(g, h) d(gh, i)
for all g, h, i ∈ G.
Here and throughout the paper, for two terms a, b in a group H and an automorphism α ∈
Aut(H), α ab denotes the element α(a) · b.
Proof. Calculate:
d(g, h) d(gh, i) = f (g)f (h)f (i)f (ghi)−1 =f (g) d(h, i) d(g, hi) =φ(g) d(h, i) d(g, hi)
so (d, φ) satisfies the identity of the proposition. In particular
f (g)
d(h, i) = d(g, h) d(gh, i) d(g, hi)−1
so f (g) conjugates any d(h, i) ∈ D into the finite set D · D · D−1 . Here for two sets A, B ⊂ H,
we write A · B = {a · b ∈ H | a ∈ A, b ∈ B} and A−1 denotes the set of inverses of A. This also
shows that ∆ is a normal subgroup of H0 (D generates ∆) and that φ : G → Aut(∆) has finite
image. It is straightforward to see that it’s quotient map ψ : G → Out(∆) is a homomorphism
with finite image.
Definition 1.4. Let G and N be groups and let e : G × G → N and φ : G → Aut(N ) be
functions such that
φ(g)
e(h, i) e(g, hi) = e(g, h) e(gh, i)
and for all g, h, i ∈ G, n ∈ N ,
e(g, h)n e(g, h)−1 =φ(g)φ(h)φ(gh)
−1
Then the pair (e, φ) satisfies the non-abelian cocycle condition.
3
n.
The last proposition showed in particular that the pair (d, φ) satisfies the non-abelian cocycle
condition for N = ∆. Pairs that satisfy the non-abelian cocycle condition are intimately related
to group extensions of G by N .
Remark 1.5. (Equivalent definitions of quasihomomorphism) f (e) = d(e, e) ∈ D and d(g −1 , g) =
f (g −1 )f (g)f (e)−1 and hence f (g −1 ) ∈ Df (g −1 ). Using this as well as a multiple application of
Proposition 1.3 we see that {f (g)−1 f (h)−1 f (gh) | g, h ∈ G} is a finite set. This was in fact the
original definition of [FK16] for quasihomomorphisms. Analogous calculations show that in fact
their definition is equivalent to ours.
In light of Proposition 1.3 the extra assumption in Theorem 1.1 that the conjugation by the
quasihomomorphism induces a finite image in Aut(N ) is natural: Given a short exact sequence
1 → N → E → G → 1 that obtains a quasihomomorphic section f : G → E it is quickly seen
that 1 → ∆ → E0 → G → 1 is a short exact sequence where ∆ ֒→ N , E0 ֒→ E and the map to
Aut(N ) has finite image. In fact this assumption is necessary as Example 3.4 shows.
1.3. Acknowledgements. The author would like to thank his supervisor Martin Bridson for his
helpful comments and support. The author was funded by the Oxford-Cocker Graduate Scholarship. Part of this work was written at the Isaac Newton Institue in the programme Non-Positive
Curvature Group Actions and Cohomology, supported by the EPSRC Grant EP/K032208/1.
2. Extensions and proof of Theorem 1.1
We follow the outline of the proof of the classical analogue of Theorem 1.1, as found in
[Bro82]. Let 1 → N → E → G → 1 be a short exact sequence of groups and let s : G → E be
a set-theoretic section. Let e(g, h) = s(g)s(h)s(gh)−1 be the defect of s and φ : G → Aut(N )
be defined via φ(g) : n 7→ f (g)nf (g)−1 . The calculations of Proposition 1.3 show that the pair
(e, φ) satisfies the non-abelian cocycle condition. In this case we say that (e, φ) is induced by the
extension E. The key observation in characterising group extensions is that the converse holds:
Every pair (e, φ) that satisfies the non-abelian cocycle condition gives rise to a group extension.
We define E(e, φ) to be the group on the set N × G and with group multiplicaion
(n1 , g1 ) · (n2 , g2 ) = (n1
φ(g1 )
n2 e(g1 , g2 ), g1 g2 )
We claim that the maps N → E(e, φ) via n 7→ (n, 1) and E(e, φ) → G via (n, g) 7→ g are
homomorphisms and that
1 → N → E(e, φ) → G → 1
is a short exact sequence.
Proposition 2.1. Let G, N be groups.
(1) Let e : G × G → N and φ : G → Aut(N ) be functions such that (e, φ) satisfies the nonabelian cocycle condition. Then E(e, φ) is a group which is an extension of G by N with
maps N → E(e, φ) via n 7→ (n, 1) and E(e, φ) → G via (n, g) 7→ g. Furthermore, this
extension induces (e, φ).
(2) Let E be an extension of G by N that induces (e, φ). Then E(e, φ) is equivalent to E.
(3) Let E be an extension of G by N and let ψ : G → Out(N ) be the induced homomorphism.
For every set-theoretic lift φ : G → Aut(N ) (i.e. not necessarily a homomorphism) there
is a map e : G × G → N such that (e, φ) satisfies the non-abelian cocycle condition and
E is equivalent to E(e, φ) as an extension.
(4) Let ψ : G → Out(N ) be a homomorphism with finite image and let φ : G → Aut(N ) be
a set-theoretic lift with finite image. Let E be an extension of G by N which induces ψ.
Then there is a section s : G → E which is a quasihomomorphism and which induces a
4
finite map to Aut(N ) if and only if there is a function e : G × G → N with finite image
such that (e, φ) satisfies the non-abelian cocycle condition and E is equivalent to E(e, φ).
Note that this proposition reduces the problem of finding extensions of G by N inducing a
homomorphism ψ to the problem of finding functions e such that (e, φ) satisfy the non-abelian
cocycle condition for a fixed lift φ.
Proof. E(e, φ) is a group by a straightforward calculation performed in the Appendix (Subsection
4.0.1) and it is easily seen to be an extension of G by N . The section s : G → E(e, φ) via
g 7→ (1, g) induces (e, φ) which satisfies the cocycle condition by the same calculation as in the
proof of Proposition 1.3 which shows (1). Let E be an extension of G by N and choose a section
s : G → E such that (e, φ) is enduced via this section. Let i : N → E be the injection and
p : E → G be the surjection. Then the maps I : E → E(e, φ) via e 7→ (e · s ◦ p(e)−1 , p(e)) and
E(e, φ) → E via (n, g) 7→ i(n) · s(g) are seen to be inverse homomorphisms and it is easy to
see that E and E(e, φ) are equivalent, which shows (2). To see (3) let ψ : G → Out(N ) be the
homomorphism. For any lift φ : G → Aut(N ) we can find a section s : G → E such that φ acts
on N as conjugation by s since we can modify s by an arbitrary map G → N . This section then
gives to the desired function e.
Let ψ and φ be as in the assumptions of (4). Suppose there is a quasimorphic section s : G → E
which induces a finite image to Aut(N ). Then there is a finite-valued function n : G → N such
that s′ : G → E induces φ where s′ (g) = s(g)n(g). Because of the assumptions, s′ can also be
seen to be a quasihomomorphism. On the other hand if (e, φ) satisfies the non-abelian cocycle
condition with e : G × G → N finite image then g 7→ (1, g) necessarily is a quasihomomorphism
which induces the finite valued function φ.
Lemma 2.2. Let ψ : G → Out(N ) be a homomorphism and let φ : G → Aut(N ) be a settheoretic lift. Let e : G × G → N be such that (e, φ) satsifies the non-abelian cocycle condition.
(1) Let e′ : G × G → N be a function representing the defect of φ i.e. the function (g, h) 7→
φ(g)φ(h)φ(gh)−1 . Then (e′ , φ) satisfies the non-abelian cocycle condition if and only if
there is a function c : G × G → Z where Z denotes the centre of N such that
e(g, h) = e′ (g, h) c(g, h)
for all g, h ∈ G.
(2) Let e′′ : G × G → N be another function such that (e′′ , φ) satisfies the non-abelian cocycle
condition. Then E(e, φ) and E(e′′ , φ) are equivalent if and only if there is a function
z : G → Z where Z is the centre of N such that
e(g, h) = e′′ (g, h)δ 1 z(g, h)
for all g, h ∈ G.
Proof. To see the first part, note that both e(g, h) and e′ (g, h) have to represent φ(g)φ(h)φ(gh)−1
and hence only differ by an element in the centre. For the second part let (n, g) and (n, g)′′
denote the coordinates of E(e, φ) and E(e′′ , φ) respectively. An isomorphism E(e, φ) → E(e′′ , φ)
which commutes with the maps of the short exact sequences neccesarily has to have the form
(n, g) 7→ (n z(g), g). The function z : G → N maps to the centre as for both extensions the
sections s : g 7→ (1, g) and s′′ : g 7→ (1, g)′′ induce the map φ. The statement follows by calculating
the defect of s and s′′ .
We can now give a proof to Theorem 1.1.
Proof. Recall that Eb (G, N, ψ) denotes the subset of extensions E of G by N that induce the
homomorphism ψ : G → Out(N ) with finite image such that there is a section s : G → E which
5
is a quasihomomorphism that induces a finite valued map G → Aut(N ) via conjugation in E.
Fix a (set theoretic) lift φ : G → Aut(N ) of ψ with finite image. Recall that we assume Z, the
centre of N , to be finitely generated.
Suppose that Eb (G, N, ψ) 6= 0 and fix E ∈ Eb (G, N, ψ). By Proposition 2.1 (4) there is a
function e0 : G × G → N with finite image such that E is equivalent to E(e0 , φ). We claim that
the map
H 2 (G, Z) → E(G, N, ψ)
[ω] 7→ E(e0 ω, φ)
is a bijection which restricts to a bijection H 2 (G, Z) ⊃ im(c2 ) → Eb (G, N, ψ). Note that for any
cocycle ω ∈ C 2 (G, Z) the map e defined via e(g, h) = e0 (g, h)ω(g, h) necessarily satisfies the nonabelian cocycle condition as well. Lemma 2.2 shows that the two pairs (e0 ω, φ), (e0 ω ′ , φ) that
satisfy the non-abelian cocycle condition give rise to an equivalent extension if and only if there
is a map z ∈ C 1 (G, Z) such that ω = ω ′ + δ 1 z. This shows that the map H 2 (G, Z) → E(G, N, ψ)
is injective.
By Proposition 2.1 any group extension E arises as a group extension E(e, φ), where (e, φ)
satisfies the non-abelian cocycle condition. By the lemma there is a function c ∈ C 2 (G, Z) such
that e(g, h) = e0 (g, h) c(g, h). By comparing the cocycle conditions we see that c is an honest
cocycle and hence the map H 2 (G, Z) → E(G, N, ψ) is a bijection. By the same argument and
restricting to classes where e : G × G → N is bounded we see that this map induces a bijection
im(c2 ) → Eb (G, N, ψ).
In view of Proposition 2.1, Eb (G, N, ψ) is not empty if and only if there is a finite map
e : G × G → N such that (e, φ) satisfies the non-abelian cocycle condition. Now we define the obstruction ob (G, N, ψ). Pick for each g, h ∈ G an element e(g, h) ∈ N such that
−1
e(g, h)n e(g, h) =φ(g)φ(h)φ(gh) n and such that e(g, h) has finite image. Define the function
(g, h, i) 7→ ob (g, h, i) ∈ N via
φ(g)
(1)
e(h, i) e(g, hi) = ob (g, h, i) e(g, h) e(gh, i)
and observe that ob necessarily has finite image as e : G×G → N and φ : G → Aut(N ) have finite
image. We claim that ob maps to Z and that ob ∈ Cb3 (G, Z) is a cocycle, i.e. δ 3 ob = 0. Moreover
we claim that [ob ] ∈ Hb3 (G, Z) is independent of the choices made. The proof of these claims
are calculations which are postponed to Appendix 4.0.2. We call this class [ob ] ∈ Hb3 (G, Z) the
obstruction of (G, N, ψ) and denote it by ob (G, N, ψ).
If [ob (G, N, ψ)] = 0 ∈ Hb3 (G, Z), then there is a c ∈ Cb2 (G, Z) such that
ob (g, h, i) =φ(g) c(h, i) − c(gh, i) + c(g, hi) − c(g, h)
for all g, h, i ∈ G. Define e′ : G × G → N via e′ (g, h) = e(g, h) c(g, h)−1 . Then e′ is bounded and
(e′ , φ) satisfies the non-abelian cocycle condition and hence E(e′ , φ) ∈ Eb (G, N, ψ).
For e as above, each e′ : G×G → N such that (e′ , φ) satisfies the non-abelian cocycle condition
has to have the form e′ (g, h) = e(g, h) c(g, h) where c(g, h) ∈ Z. If Eb (G, N, ψ) = ∅ there is no
e′ : G × G → N bounded such that (e′ , φ) satisfies the non-abelian cocycle condition. Hence there
is no c ∈ Cb2 (G, Z) such that
φ(g)
e(h, i)φ(g) c(h, i) e(g, hi) c(g, hi) = e(g, h) c(g, h) e(gh, i) c(gh, i)
φ(g)
e(h, i) e(g, hi) = δ 2 c(g, h, i)−1 e(g, h) e(gh, i)
which shows in particular that [ob ] 6= 0 ∈ Hb3 (G, Z). By precisely the same argument without requireing that e′ is bounded we see that [ob (G, N, ψ)] = 0 ∈ H 3 (G, Z) if and only if
E(G, N, ψ) 6= ∅.
6
3. The set of obstructions and examples
Theorem 1.1 gives us a machinery to construct classes [o] ∈ Hb3 (G, Z), for Z finitely generated
and the action of G on Z finite. One may wonder how many such classes [o] ∈ Hb3 (G, Z) arise as
an obstruction. In the classical theory, every cocycle in H 3 (G, Z) arises as such an obstruction.
Theorem. (Section IV [Bro82]) Let G be a group, Z a G-module, then for every ω ∈ H 3 (G, Z)
there is a group N with Z(N ) = Z and a homomorphism ψ : G → Out(N ) extending the action
of G on Z such that ω = [o(G, N, ψ)] ∈ H 3 (G, Z).
For Z a finitely generated abelian group with a finite G action define the set of obstructions
Ob (G, Z) = {[ob (G, N, ψ)] ∈ Hb3 (G, Z) | Z = Z(N ),ψ(g) z = g · z, ψ : G → Out(N ) finite}
Observe that in the construction of ob (G, N, ψ) we could choose ob as a pullback of the finite
group Out(N ) via the homomorphism ψ : G → Out(N ). This motivates the definition
F (G, Z) := {f ∗ k ∈ Hb3 (G, Z) | f : G → M a homomorphism, M a finite group, k ∈ H 3 (M, Z)}
and f ∗ k denotes the pullback of k via the homomorphism f . As M is finite, k is necessarily
bounded.
Theorem 3.1. Let G be a group and Z be a finitely generated abelian group with finite G-action.
Then
Ob (G, Z) = F (G, Z)
as subsets of Hb3 (G, Z).
This fully characterises the obstructions. In particular it shows that no ”exotic” cocycles arise.
Proof. We have just seen that O(G, Z) ⊂ F (G, Z) above. Hence, for every finite group M
and a class Z ∈ H 3 (M, Z) we wish to construct a group N and a homomorphism ψ : M →
Out(N ) which induces k as a cocycle. We recall a construction of [Mac67]. We may assume that
k(1, g, h) = k(g, 1, h) = k(g, h, 1) = 0 for all g, h ∈ G.
Define the abstract symbols hg, hi for each 1 6 g, h ∈ M and set hg, 1i = h1, gi = h1, 1i = 1 for
the abstract symbol 1. Let F be the free group on these symbols and set 1 to be the identity
element. Set N = Z × F . Define the function φ : M → Aut(N ) so that for g ∈ M the action of
φ(g) on F is given by
φ(g)
hh, ii = k(g, h, i)hg, hihgh, iihg, hii−1
and so that the action of φ(g) on Z is given by the M -action on Z. A direct calculation yields
that for each g ∈ M , φ(g) : N → N indeed defines a isomorphism. Here, we need the assumption
k(1, g, h) = k(g, 1, h) = k(g, h, 1) = 0. Also it can be seen that for all n ∈ N and g1 , g2 ∈ F
φ(g1 )φ(g2 )
n =τhg1 ,g2 i ◦φ(g1 g2 ) n
where we have to use the fact the k is a cocycle. Here, for a group G and an element g ∈ G,
the map τg : G → G denotes the left conjugation τg : h 7→ ghg −1 . Hence, φ : M → Aut(N ) is
well defined and induces a homomorphism ψ : M → Out(N ). It is easy to see that ψ induces
the M -action on Z. If M ∼
6 Z2 , the centre of N is Z. In this case, to calculate ob (M, N, ψ)
=
we choose as representatives for φ(g)φ(h)φ(gh)−1 simply hg, hi and then see by definition that
ob (M, N, ψ) is precisely k.
If M = Z2 then the centre of N as constructed above is not Z But all of N . However, we can
enlarge M by setting M̃ = M × Z2 . We have both a homomorphism p : M̃ → M via (m, z) 7→ m
and a homomorphism i : M → M̃ via m 7→ (m, 1) such that p ◦ i = idM . Let k̃ ∈ H 3 (M̃ , Z)
be the pullback of k via p. Let Ñ be the group constructed as above with this cocycle and let
7
φ̃ : M̃ → Aut(Ñ ) and ψ̃ : M̃ → Out(Ñ ) be the corresponding functions. The centre of Ñ is Z.
Set ψ : M → Out(Ñ ) via ψ = ψ̃ ◦ i. Then the obstruction ob (M, Ñ , ψ) can be seen to be k.
3.1. Examples. The subset Eb (G, N, ψ) of E(G, N, ψ) is generally neither empty nor all of
E(G, N, ψ). For any hyperbolic group we have Eb (G, N, ψ) = E(G, N, ψ) as the comparison map
is surjective ([Min02]). We give different examples where the inclusion Eb (G, N, ψ) ⊂ E(G, N, ψ)
is strict:
Example 3.2. Let G = Z2 and N = Z and ψ be trivial. The direct product
1 → Z → Z × Z2 → Z2 → 1
clearly has a quasihomomorphic section that induces a finite map to Aut(N ) hence ∅ =
6 Eb (Z2 , Z, id).
2
2
H (Z , Z) is generated by the Heisenberg cocycle ω, which does not admit a bounded representative. Hence im(c2 ) = 0 and so by Theorem 1.1 {Z × Z2 } = Eb (Z2 , Z, id). However the Heisenberg
extension
1 → Z → H3 → Z2 → 1
is another non-equivalent extension, which hence can’t admit a quasihomomorphic section. So
∅ 6= Eb (Z2 , Z, id) ( E(Z2 , Z, id).
Example 3.3. Eb (G, N, ψ) can be empty and E(G, N, ψ) nonempty. By Theorem 3.1 this happens if and only if there is a cocycle k such that 0 6= [k] ∈ Hb3 (G, Z), 0 = [k] ∈ H 3 (G, Z) and k
factors through a finite group.
Let G = Z2 and let Z = Z. Define the map o ∈ C 2 (Z2 , Z) via o(v, w) = ⌊ 21 det(v, w)⌋ and
k = δ 2 o ∈ Cb3 (Z2 , Z). Obviously 0 = [k] ∈ H 3 (G, Z). By exploiting the linear and quadratic
growth of o, one sees that [k] does not have a bounded coboundary and hence 0 6= [k] ∈ Hb3 (G, Z).
Furthermore, one can see that k factors through the finite group Z2 .
Example 3.4. The assumption that the quasihomomorphism f : G → E has to induce a finite
map to Aut(N ) may seem artificial, as the induced homomorphism to Out(N ) already has finite
image. However it is necessary as the following example shows.
Set C = Z and Z = Z2 . The Heisenberg Group H3 is the central extension C → H3 → Z
corresponding to the determinant-cocycle ω ∈ H 2 (Z, C). Since it is a central extension of Z, it’s
inner automorphism group is isomorphic to Z ∼
= Z2 . Let G be another copy of Z2 and let φ be
∼
the isomorphism φ : G → Inn(H3 ) = Z. Let E = H3 ⋊φ G, the semi-direct product of H3 and G
via φ.
The group G acts on H3 via inner automorphisms hence the induced map to the outer autormorphisms is trivial. We will see that E 6∈ Eb (G, H3 , id). Clearly Eb (G, H3 , id) is not empty as it
contains the direct product H3 ×G which is not isomorphic to E. By Theorem 1.1, Eb (G, H3 , id)
is in a bijective correspondence to im(c2 ) for c2 : Hb2 (G, Z) → H 2 (G, Z), where G acts on Z with
trivial action. It is well known that the comparison map has trivial image, hence Eb (G, H3 , id)
only consists of the direct product and in particular does not contain E.
However, there are two very different sort of sections G → E which both satisfy partially the
requirements of being in Eb (G, H3 , id): As E is a semidirect product there are sections which
are quasihomomorphism, in fact homomorphisms. Also, as the homomorphism to Out(H3 ) is
trivial, there is a section which induces a trivial map to Aut(H3 ). However, we can’t have both
properties for sections s : G → E simultaneously.
3.2. Generalisations. One interesting aspect of Theorem 1.1 is that it gives a characterisation
of classes in third bounded cohomology, the obstructions. One may wonder in view of Theorem
3.1 if another definition of ”quasihomomorphism” gives more interesting classes by considering
8
only extensions which obtain this other definiton of quasihohomomorphism. This thought has
some limitations:
Soma ([Som97]) showed that the third bounded cohomology of the free group is uncountable dimensional. However, there should not be ”obstructions” in extensions of the free group
at least under the natural assumption that the class of ”quasimorphisms” considered contains
homomorphisms.
Another much broader definition of quasimorphisms can be found in [HS16] by Hartnick and
Schweitzer. They consider maps f : G → H such that the pullback of any ordinary quasimorphism on H gives a quasimorphism on G. Here, there are even less obstructions: Let G = hS | Ri
be a torsion free group satisfying the small cancelation condition C′ ( 18 ). Let F (S) → G be the
obvious homomorphism. Rips and Sela ([RS95]) gave a construction of a section (normal form)
θ : G → F (S) which can be seen to be a quasimorphism in the sense of Hartnick and Schweitzer.
This shows that for G as above and any surjection E → G there always is a section s : G → E
which is a quasimorphism in the sense of Hartnick and Schweitzer. It would be interesting to
study groups with this property and groups where this property fails to hold.
4. Appendix: Calculations
4.0.1. E(e, φ) is a group. We have to check the group axioms. We start by associativity.
φ(g )
((n1 , g1 ) · (n2 , g2 )) · (n3 , g3 ) = n1 1 n2 e(g1 , g2 ) · (n3 , g3 )
φ(g1 )
= (n1
n2 e(g1 , g2 )φ(g1 g2 ) n3 e(g1 g2 , g3 ), g1 g2 g3 )
φ(g1 ) φ(g1 )φ(g2 )
n2
n3
= (n1
e(g1 , g2 ) e(g1 g2 , g3 )
On the other hand we have:
φ(g2 )
(n1 , g1 ) · ((n2 , g2 ) ⋆ (n3 , g3 )) = (n1 , g2 ) · (n2
=
n3 e(g2 , g3 ), g2 g3 )
φ(g ) φ(g )φ(g2 ) φ(g1 )
(n1 1 n2 1
n3
e(g2 , g3 ) e(g1 , g2 g3 ), g1 g2 g3 )
As (e, φ) satisfies the non-abelian cocycle condition we see that both terms are equal, so · is
associative. The identity can be seen to be (e(1, 1)−1 , 1).
It is a calculation to check that the inverse of (n, g) is
(e(1, 1)−1 e(g −1 , g)−1 )φ(g
−1
) −1
n
, g −1 )
4.0.2. The obstruction ob (G, N, ψ).
Claim. o is a cocycle of Cb3 (G, Z)
Calculate:
φ(g)φ(h)
e(i, k)
φ(g)
e(h, ik) e(g, hik)
=φ(g)φ(h) e(i, k) ob (g, h, ik) e(g, h) e(gh, ik)
= e(g, h)φ(gh) e(i, k) ob (gh, ik) ob (g, h, ik)
= e(g, h) e(gh, i) e(ghi, k) ob (g, h, ik) ob (gh, i, k)
9
and
φ(g)φ(h)
e(i, k)φ(g) e(h, ik) e(g, hik)
=φ(g) (ob (h, i, k) e(h, i) e(hi, k)) e(g, hik)
=φ(g) e(h, i) e(g, hi) e(ghi, k)φ(g) ob (h, i, k) ob (g, hi, k)
= e(g, h) e(gh, i) e(ghi, k) ob (g, h, i)φ(g) ob (h, i, k) ob (g, hi, k)
Finally, comparing these two terms yields
∂ ob (g, h, i, k) =φ(g) ob (h, i, k) − ob (gh, i, k) + ob (g, hi, k) − ob (g, h, ik) + ob (g, h, i) = 0
So ob really is a cocycle.
Claim. [ob ] ∈ Hb3 (G, Z) is well defined i.e. independent of the choice of e and φ.
Let φ, φ′ : G → Aut(N ) be two finite lifts of ψ and let defined as above and choose corresponding functions e, e′ representing the defect of φ and φ′ respectively. There is a finite
function n : G → N such that φ(g) = n̄(g)φ′ (g) where n̄(g) is the element in Inn(N ) ⊂ Aut(N )
represented by n(g).
Hence:
′
φ(g)φ(h)φ(gh)−1 = n̄(g)φ (g) n̄(h) φ′ (g)φ′ (h)φ′ (gh)−1 n̄(gh)−1
As a functions to Inn(N ). Lifting this to N gives:
′
e(g, h) = z(g, h)n(g)φ (g) n(h) e′ (g, h)n(gh)−1
where z : G × G → Z is a finite-valued function to Z, the centre of N . It is a calculation to show
that ob , the obstruction defined via the choices φ and e and o′b , the obstruction defined via the
choices φ′ and e′ differ by δ 2 z and hence define the same class in bounded cohomology.
References
[Bro82] Kenneth S. Brown. Cohomology of groups, volume 87 of Graduate Texts in Mathematics. Springer-Verlag,
New York-Berlin, 1982.
[Cal09] Danny Calegari. scl, volume 20 of MSJ Memoirs. Mathematical Society of Japan, Tokyo, 2009.
[FK16] Koji Fujiwara and Michael Kapovich. On quasihomomorphisms with noncommutative targets. Geom.
Funct. Anal., 26(2):478–519, 2016.
[Fri16] Roberto Frigerio. Bounded cohomology of discrete groups. arXiv:1610.08339, 2016.
[HS16] Tobias Hartnick and Pascal Schweitzer. On quasioutomorphism groups of free groups and their transitivity properties. J. Algebra, 450:242–281, 2016.
[Mac67] Saunders MacLane. Homology. Springer-Verlag, Berlin-New York, first edition, 1967. Die Grundlehren
der mathematischen Wissenschaften, Band 114.
[Min02] Igor Mineyev. Bounded cohomology characterizes hyperbolic groups. Q. J. Math., 53(1):59–73, 2002.
[RS95] E. Rips and Z. Sela. Canonical representatives and equations in hyperbolic groups. Invent. Math.,
120(3):489–512, 1995.
[Som97] Teruhiko Soma. Bounded cohomology and topologically tame Kleinian groups. Duke Math. J., 88(2):357–
370, 1997.
Department of Mathematics, University of Oxford
E-mail address, N. Heuer: [email protected]
10
| 4math.GR
|
Two-stage Algorithm for Fairness-aware Machine Learning
Junpei Komiyama1 and Hajime Shimao2
1
University of Tokyo
2
Purdue University
arXiv:1710.04924v1 [stat.ML] 13 Oct 2017
October 16, 2017
Abstract
Algorithmic decision making process now affects many aspects of our lives. Standard tools for
machine learning, such as classification and regression, are subject to the bias in data, and thus direct
application of such off-the-shelf tools could lead to a specific group being unfairly discriminated.
Removing sensitive attributes of data does not solve this problem because a disparate impact can
arise when non-sensitive attributes and sensitive attributes are correlated. Here, we study a fair
machine learning algorithm that avoids such a disparate impact when making a decision. Inspired
by the two-stage least squares method that is widely used in the field of economics, we propose a
two-stage algorithm that removes bias in the training data. The proposed algorithm is conceptually
simple. Unlike most of existing fair algorithms that are designed for classification tasks, the proposed
method is able to (i) deal with regression tasks, (ii) combine explanatory attributes to remove reverse
discrimination, and (iii) deal with numerical sensitive attributes. The performance and fairness of
the proposed algorithm are evaluated in simulations with synthetic and real-world datasets.
1
Introduction
Algorithmic decision making process now affects many aspects of our lives. Emails are spam-filtered by
classifiers, images are automatically tagged and sorted, and news articles are clustered and ranked. These
days, even decisions regarding individual people are being made algorithmically. For example, computergenerated credit scores are popular in many countries, and job interviewees are sometimes evaluated by
assessment algorithms1 . However, a potential loss of transparency, accountability, and fairness arises
when decision making is conducted on the basis of past data. For example, if a dataset indicates that
specific groups based on sensitive attributes (e.g., gender, race, and religion) are of higher risk in receiving
loans, direct application of machine learning algorithm would highly likely result in loan applicants on
those groups being rejected.
In other words, a machine learning algorithm that utilizes sensitive attributes is subject to biases
in the existing data. This could be viewed as an algorithmic version of disparate treatment [8], where
decisions are made on the basis of these sensitive attributes. However, removing sensitive attributes from
the dataset is not sufficient solution as it has a disparate impact. Disparate impact is a notion that was
born in the 1970s. The U.S. Supreme Court ruled that the hiring decision at the center of the Griggs
v. Duke Power Co. case [9] was illegal because it disadvantaged an application of to a certain race,
even though the decision was not explicitly determined based on the basis of race. Duke Power Co. was
subsequently forced to stop using test scores and diplomas, which are highly correlated with race, in its
1 https://www.hirevue.com/
1
hiring decisions. Later on, the U.S. Equal Employment Opportunity Commission [7] stated that disparate
impact arises when some group has a less than 80% selection probability in hiring, promotion, and other
employment decisions involving race, sex, or ethnic group, which is so-called the 80%-rule. Economists
have argued that not only is disparate impact ethically problematic, in the labor market, it may also
cause inefficiency in social welfare even when the discrimination is based on statistical evidence [19].
In this paper, we are interested in a fair algorithm that prevents disparate impact. The study of
disparate impact is not new in the context of fairness-aware machine learning. As we review in Section
1.1, there are a number of existing studies on this topic. However, there are three major limitations on
the existing algorithms intended to alleviate disparate impact:
• Most of the existing algorithms are built for classification tasks and cannot deal with regression
tasks. Arguably, classification is very important. A fair classifier avoids disparate impact in making
decisions on admission in universities and hiring by companies. Still, there are tasks that require
continuous attributes, such as salaries quoted in a job offer and penalties of criminals. Unfortunately,
only a few algorithms Calders et al. [5, 2] are able to handle regression. Note that data preprocessing algorithms such as Zemel et al. [24, 10] could also be combined with regressors.
• Direct application of fairness could lead to reverse discrimination. To see this, let us take the
example of income [25]. In the Adult dataset 2 , women on average have lower incomes than men.
However, women in the dataset work fewer hours than men per week on average. A fairness-aware
classifier built on the top of this dataset, which equalizes the wages of women and men, leads to
a reverse discrimination that makes the salary-per-hour of men smaller than that of women. Such
discrimination can be avoided by introducing explanatory attributes and this allows us to make a
difference on the basis of the explanatory attributes. In fact, the as in case of Griggs v. Duke Power
Co., promoting decisions that cause disparate impacts is not allowed because they are not based
on a reasonable measure of job performance, which implies (in some cases) decisions can be fair if
they are of reasonable explanatory attributes. Unfortunately, most of the existing studies cannot
utilize explanatory attributes.
• Existing algorithms cannot deal with numerical (continuous) sensitive attributes. Although most
sensitive attributes, such as gender, race, and religions are binary or categorical (polyvalent), some
sensitive attributes are naturally dealt with in terms of numerical values. For example, the Age
Discrimination Act [6] in the U.S. prohibits discrimination in hiring, promotion, and compensation
on the basis of age for workers age 40 or above; here, age is a sensitive attribute that is naturally
dealt with numerically.
Table 1 compares our algorithm with the existing ones.
Contributions: Inspired by the econometrics literature, we propose a two-stage discrimination remover (2SDR) algorithm (Section 3). The algorithm consists of two stages. The first removes disparate
impact, and the second is for prediction. The first stage can be considered to be a data transformation
that makes the linear classifiers of the second stage fair. Theoretical property of 2SDR is analyzed in
Section 4. We verified the practical utility of 2SDR by using real-world datasets (Section 5). We showed
that 2SDR is a fair algorithm that (i) performs quite well in not only regression tasks but also classification tasks and (ii) is able to utilize explanatory attributes to improve estimation accuracy. Moreover
(iii), it reduces discrimination bias in numeric sensitive attributes, which enables us to avoid other classes
of discrimination, such as age discrimination [6].
2 https://archive.ics.uci.edu/ml/machine-learning-databases/adult/
2
1.1
Related work
This section reviews the previous work on fairness-aware machine learning algorithms. These algorithms
can be classified into two categories: Algorithms of the first category process datapoints before or after
putting them into classifier or regressor. Such an algorithm typically transforms training datasets so
as to remove any dependency between the sensitive attribute and target attribute. The advantage of
these algorithms is generality: they can be combined with a larger class of off-the-shelf algorithms for
classification and regression. Moreover, the transformed data can be considered as a “fair representation”
[10] that is free from discrimination. The biggest downside of these algorithms that they treat a classifier
as a black-box, and as a result, they need to change the datapoints drastically, which tends to reduce
accuracy. Regarding the algorithms of this category, Kamiran et al. [15] proposed a data-debiasing
scheme by using a ranking algorithm. They were inspired by the idea that the datapoints close to
the class borderline are prone to discrimination, and they resample datapoints so as to satisfy fairness
constraints. Zliobaite et al. [25] argued that some part of discrimination is explainable by some attributes.
They also proposed resampling and relabelling methods that help in training fair classifiers. Zemel et
al. [24] proposed a method to learn a discrete intermediate fair representation. Feldman et al. [10]
considered a quantile-based transformation of each attribute. Hardt et al. [13] studied the condition of
equalized odds, and provided a post-processing method that fulfills the condition.
Algorithms of the second category directly classify or regress datapoints. Such algorithms tend to
perform well in practice since they do not need to conduct explicit data transformation that loses some
information. The downside of these algorithms is that one needs to modify an existing classifier for each
task. Regarding the algorithms of this approach, Ristanoski et al. [20] proposed a version of support
vector machine (SVM), called SVMDisc, that involves a discrimination loss term. Fish et al. [11] shifted
the decision boundary of the classical AdaBoost algorithm so that fairness is preserved. Goh et al.
[12] considered a constrained optimization that satisfies various constraints including the one of fairness.
Kamishima et al. [17] proposed prejudice index and proposed a regularizer to reduce prejudice. Zafar
et al. [23] considered a constrained optimization for classification tasks that maximizes accuracy (resp.
fairness) subject to fairness (resp. accuracy) constraint.
Our two-stage approach lies somewhere between the data preprocessing approach and direct approach.
The first stage of 2SDR transforms datasets to make the classifier or regressor in the second stage fair.
Table 1: List of fair estimators and their capabilities. “Categorical sensitive attrs” indicates that an
algorithm can deal with more than binary sensitive attributes. “Numeric sensitive attrs” indicates that an
algorithm can deal with continuous sensitive attributes. “explanatory attrs” indicates that an algorithm
utilizes some attributes that justify the treatment (e.g. the effect of working hours on wages) [25]. The
checkmark indicates the capability of the algorithm in the corresponding aspect.
algorithms
Kamiran et al. [15]
Zliobaite et al. [25]
Kamishima et al. [17]
Calders et al. [5]
Zemel et al. [24]
Fish et al. [11]
Feldman et al. [10]
Zafar et al. [23]
This paper
categorical
sensitive
attrs
7
7
7
7
3
7
3
3
3
numeric
sensitive
attrs
7
7
7
7
7
7
7
3
3
3
explanatory
attrs
classification
regression
7
3
7
3
7
7
7
7
3
3
3
3
3
3
3
3
3
3
7
7
7
3
3
7
3
7
3
Unlike most data preprocessing algorithms, the transformation of the first stage in 2SDR conducts the
minimum amount of transformation that is primarily intended for linear algorithms, and thus, it does
not degrade the original information by much. Moreover, any class of linear algorithm can be used in the
second stage, and as a result our algorithm can handle more diverse range of tasks and conditions than
the existing algorithms can.
Note that other tasks have been considered in the literature of fairness-aware machine learning. To
name a few, Kamishima et al. [16, 18] considered methods for removing discrimination in recommendation
tasks. Joseph et al. [14] considered fairness in the context of online content selection. Bolukbasi et al.
[4] considered fairness in dense word representation learnt from text corpora.
2
Problem
Each vector in this paper is a column vector and is identified as a d × 1 matrix where d is the dimension of
the vector. Let n be the number of datapoints. The i-th datapoint is comprised of a tuple (si , xi , zi , yi ),
where
• si ∈ Rds is the ”sensitive” attributes of ds dimensions that requires special care (e.g., sex, race, and
age).
• xi ∈ Rdx is the normal non-sensitive attributes of dx dimensions. The difficulty in fairness-aware
machine learning is that xi is correlated with si and requires to be ”fairness adjusted”.
• zi ∈ Rdz is the set of explanatory attributes of dz dimensions that either are not independent of si ,
or not to be adjusted for whatever philosophical reasons. Note that zi can be blank (i.e., dz = 0)
when no explanatory attribute is found.
• yi is the target attribute to predict. In the case of classification, yi ∈ {0, 1}, whereas in the case of
regression, yi ∈ R.
Note that, unlike most existing algorithms, we allow si to be continuous. A fairness-aware algorithm
outputs ŷ(s, x, z), which is an estimator of y that complies with some fairness criteria, which we discuss
in the next section. We also use Y ∈ Rn×1 , X ∈ Rn×dx , Z ∈ Rn×dz , S ∈ Rn×ds to denote a sequence of n
> >
>
datapoints. Namely, the i-th rows of S, X, Z, and Y are s>
i , xi , zi , and yi , respectively.
2.1
Fairness criteria
This section discusses fairness criteria that a fairness-aware algorithm is expected to comply with. We
consider group-level fairness in the sense of preventing disparate impact [7], which benefits some group
disproportionally. For ease of discussion, we assume ds = 1 and s is a binary3 or real single attribute.
Note that our method (Section 3) is capable of dealing with (i) multiple attributes and (ii) categorical s
by expanding dummies (i.e., converting them into multiple binary attributes). Let (s, x, z, y) be a sample
from the target dataset to make a prediction. Let ŷ = ŷ(x, z) be an estimate of y that an algorithm
outputs. For binary s and ŷ, the P%-rule [7, 23] is defined as
min
p
P[ŷ = 1|s = 1] P[ŷ = 1|s = 0]
,
P[ŷ = 1|s = 0] P[y = 1|s = 1]
≥
p
.
100
(1)
The rule states that each group has a positive probability at least p% of the other group. The 100%-rule
implies perfect removal of disparate impact on group-level fairness, and a large value of p is preferred.
3 Although there are several possible definitions, it is not very difficult to extend a fairness measure of binary s to one of
a categorical s.
4
For binary s and continuous ŷ, an natural measure that corresponds to the p%-rule is the mean
distance (MD) [5], which is defined as:
|E[ŷ|s = 1] − E[ŷ|s = 0]| ,
(2)
which is a non-negative real value, and a MD value close to zero implies no correlation between s and y.
Moreover, Calders et al. [5] introduced the area under the receiver operation characteristic curve (AUC)
between ŷ and s:
P
P
i∈{1,2,...,n}:si =1
j∈{1,2,...,n}:sj =0 I[ŷi > ŷj ]
,
(3)
ns=1 × ns=0
where I[x] is 1 if x is true and 0 otherwise, and ns=1 (resp. ns=0 ) is the number of datapoints with s = 1
(resp. s = 0), respectively. The AUC takes value in [0, 1] and is equal to 0.5 if s shows no predictable
effect on y.
Moreover, for continuous s, we use the correlation coefficient (CC) |Cov(s, ŷ)| between s and ŷ as a
fairness measure. Note that, when s is binary, the correlation is essentially equivalent to MD (Eq. (2)
up to a normalization factor.
3
Proposed Algorithm
Here, we start by discussing the two-stage least squares (2SLS), a debiasing method that is widely used
in statistics, econometrics, and many branches of natural science that involve observational bias (Section
3.1). After that, we describe the two-stage discrimination remover (2SDR) for fairness-aware classification
and regression (Section 3.2). Section 3.3 compares 2SDR with existing data preprocessing methods.
3.1
Two-stage least squares (2SLS)
Consider a linear regression model
yi = x>
i β + i ,
where the goal is to predict yi ∈ R from attributes xi ∈ Rdx . If the noise i is uncorrelated with xi , an
ordinary least square β̂OLS = (X > X)−1 X > Y consistently estimates β. However, the consistent property
is lost when xi is correlated with : Namely, it is well-known [21] that, under mild assumption
p
β̂OLS → β +
Cov(x, )
,
σx2
p
where Cov(x, ) is the covariance between x and . σx2 is the variance of xi , and the arrow → indicates a
convergence in probability. To remove the bias term, one can utilize a set of additional attributes zi that
are (i) independent of i , and (ii) correlate with xi . The crux of 2SLS is to project the columns of X in
the column space of Z:
X̂ = Z(Z > Z)−1 Z > X
β̂2SLS = (X̂ > X̂)−1 X̂ > Y.
Unlike the OLS estimator, the 2SLS estimator consistently estimates β. That is,
p
β̂2SLS → β.
5
(4)
Algorithm 1 2-Stage Discrimination Remover (2SDR).
1: Input: Second stage algorithm f (x, z).
2: Using training data (Strain , Xtrain , Ztrain , Ytrain ):
−1 >
3: B̂s ← (S>
Strain Xtrain .
train Strain )
4: Utrain ← Xtrain − Strain B̂s .
5: Train the function f with (Utrain , Ztrain ).
6: for each data point (si , xi , zi , yi ) in testdata do
>
>
7:
Predict u>
i ← xi − si B̂s .
8:
Predict ŷi ← f (ui , zi ).
9: end for
3.2
Proposed algorithm: 2SDR
Our case considers a classification problem with a fairness constraint (Section 2). That is, to estimate
the relationship
yi = x>
i β + i ,
subject to fairness criteria that urges an estimator ŷi to be uncorrelated to si (Section 2.1). The main
challenge here is that xi is correlated with the sensitive attribute si , and thus, simple use of the OLS
estimator yields a dependency between ŷi and si . To resolve this issue, we use U = X − S(S> S)−1 S> X,
which is the residual of X that is free from the effect of S, for predicting Y. In the second stage, we use
U and Z to learn an estimator of Y by using an off-the-shelf regressor or classifier. The entire picture of
the 2SDR algorithm is summarized in Algorithm 1. One may use any algorithm in the second stage: We
mainly intend a linear classifier or regressor for the reason discussed in Section (Theorem 4). Following
the literature of machine learning, we learn the first and the second stage with the training dataset, and
use them in the testing data set.
Note that each column 1, 2, . . . , dx of the first stage corresponds to an OLS estimator, which involves
no hyperparameter. In the case that the second-stage classifier or regressor involves hyperparameters, we
perform standard k-fold cross validation (CV) on the training dataset to optimize the predictive power
of them.
3.3
Comparison with other data preprocessing methods
The first stage of 2SDR (Line 3 of Algorithm 1) learns a linear relationship between S and X. This stage
transforms each datapoint by making the second stage estimator free from the disparate impact, so one
may view 2SDR as a preprocessing-based method that changes the data representation. This section
compares 2SDR with existing methods that transform a dataset before classifying or regressing it. At a
word, there are two classes of data transformation algorithm: An algorithm of the first class utilizes the
decision boundary and intensively resamples datapoints close to the boundary [15]. Such an algorithm
performs well in classifying datasets, but its extension to a regression task is not straightforward. An
algorithm of the second class successfully learns a generic representation that can be used with any
classifier or regressor afterward [24, 10]. Such an algorithm tends to lose information at the cost of
generality: the method proposed by Zemel et al. [24] maps datapoints into a finite prototypes, and the
one in Feldman et al. [10] conducts a quantile-based transformation, and loses the individual modal
structures of the datapoints of s = 0 and s = 1. As a result, these methods tend to lose estimation
accuracy. Moreover, its extension to a numeric s is non-trivial. The first stage in our method can be
considered to be a minimum transformation for making linear regression fair and preserves the original
data structure. Section 5 compares the empirical performance of 2SDR with those of Zemel et al. and
Feldman et al. [24, 10].
6
4
Analysis
This section analyses 2SDR. We first assume the linearity between S and X in the first stage, and derive
the asymptotic independence of U and S (Theorem 1). Although such assumptions essentially follow the
literature of 2SLS and are reasonable, regarding our aim of achieving fairness, a guarantee for any classes
of distribution on x and s is desired: Theorem 2 guarantees the fairness with a very mild assumption
when the second stage is a linear regressor.
Assumption 1. Assume the following data generation model where datapoints are i.i.d. drawn:
yi = x>
i β + i
(5)
>
>
x>
i = si Bs + ηi
(6)
and
where i ∈ R and ηi ∈ Rdx are mean-zero random variables independent of si . Moreover, the covariance
matrix of x is finite and full-rank 4 .
The following theorem states that under Assumption 1, u is asymptotically independent of s.
Theorem 1. (Asymptotic fairness of 2SLS under linear dependency) Let (si , xi , zi , yi ) be samples drawn
from the same distribution as the training dataset, and ui is the corresponding residual learnt from the
training distribution. Under Assumption 1, ui is asymptotically independent of si . Moreover, if zi is
independent of si , ŷi is asymptotically independent of si .
Proof. Under Assumption 1, it is well known (e.g., Thm 5.1 in Wooldridge [21]) that the first-stage
p
p
estimator is consistent. That is, B̂s → Bs as n → ∞, from which we immediately obtain ui → ηi . By
the assumption that ηi is independent of si , ui is asymptotically independent of si . The independence of
ŷi and si follows from the fact that ŷi is a function of ui and zi that are asymptotically independent of
si .
From Theorem 1, we see that 2SDR combined with any classifier or regressor in the second-stage is
fair (i.e., achieves a p%-rule for any p < 100 (resp. any MD > 0) in classification (resp. regression)
with a sufficiently large dataset. Essentially, Theorem 1 states that if the relation between u and s is
linear, the first-stage OLS estimator is able to learn the relationship between them, and as a result u is
asymptotically equivalent to η, which is the fraction of u that cannot be explained by s.
Heteroskedasticity in x: As long as Assumption 1 holds, x is asymptotically independent of s.
However, some of the assumptions may not hold for some attributes in a dataset. In particular, Eq. (6)
implies that x is linear to s, and thus, the distribution of x conditioned on s = 1 and s = 0 is identical
after correcting the bias E[x|s = 1] − E[x|s = 0] 5 . Figure 1 shows some attributes where the distribution
of x is very different among s = 1 and s = 0. Taking these attributes into consideration, we would like
to seek some properties that hold regardless of the linear assumption in the first stage. The following
theorem states that 2SDR has a plausible property that makes ŷ fair under very mild assumptions.
Theorem 2. (Asymptotic fairness of 2SLS under general distributions) Assume that each training and
testing datapoint is i.i.d. drawn from the same distribution. Assume that the covariance matrix of x
and s are finite and full-rank. Assume that the covariance matrix between x and s is finite. Then, the
covariance vector Cov(s, u) ∈ Rds ×dx converges to 0 in probability as n → ∞, where 0 denotes a zero
matrix.
4 Note
5 For
that this is a sufficient condition for the “no perfect collinearity” condition in Wooldridge [21].
the ease of discussion, let s be binary value here.
7
(a) Distribution of PctUnemployed in the C&C
dataset
(b) Distribution of age in the Adult dataset
Figure 1: The first histogram (Figure (a)) shows the percentage of people in the labor force and unemployed (PctUnemployed) in each community in the C&C dataset, where the horizontal axis is PctUnemployed and vertical axis is the number of corresponding communities. The communities are categorized
into the ones with a large portion of black people (C1) and the others (C2). One can see that PctUnemployed in C2 is sharply centered around 0.25, whereas the value in C1 shows a broader spectrum: As a
result the variance of PctUnemployed is greatly differ among the two categories. The second histogram
(Figure (b)) shows the number of people of different age in the Adult dataset, where the horizontal axis
is the age and the vertical axis is the number of people. One can see that not only the variances but also
the form of distributions are different between women and men, as majority of the women in the dataset
are of the youngest category. The details of these datasets are provided in Section 5.
Proof. Let (s, x, z, y) be a sample from the identical distribution. The OLS estimator in the first stage is
explicitly written as
−1 >
B̂s = (S>
Strain Xtrain ,
train Strain )
which, by the law of large numbers, converges in probability to Cov−1 (s, s)Cov(s, x), where Cov−1 (s, s) ∈
Rds ×ds is the inverse of the covariance matrix of s, and Cov(s, x) ∈ Rds ×dx is the covariance matrix
between s and x. Then,
Cov(s, u) = Cov(s, x) − Cov(s, B̂>
s s)
p
→ Cov(s, x) − Cov(s, s)Cov−1 (s, s)Cov(s, x)
= Cov(s, x) − Cov(s, x) = 0.
(7)
Asymptotic fairness of regressor: Notice that a linear regressor in the second stage outputs ŷ as
8
(a) CC as a func. of n
(b) CC as a func. of
dx
(c) CC as a func. of
σηs
(d) CC as a func. of
σs
(e) RMSE as a func.
of n
(f) RMSE as a func.
of dx
(g) RMSE as a func.
of σηs
(h) RMSE as a func.
of σs
Figure 2: Correlation coefficient (CC) with different parameters (Figures (a)-(d)). Figure (a) is the result
with different datasize, Figure (b) is the result with different dimension of x, Figure (c) is the result with
different strength of correlation between x and s, and Figure (d) is the result with different variance of s.
One can see that with sufficient large n (n ≥ 1, 000), 2SDR has consistently removes correlation between
s and ŷ. Figures (e)-(h) shows the root mean square error (RMSE) with the same setting as Figures
(a)-(d), where RMSE is defined as the squared empirical mean of (ŷ − y)2 . The larger dx , Cov(x, s), or σs
is, the gap of RMSE between 2SDR and OLS is larger. This is because (i) the correlation between x and
s causes a disparate impact of OLS, (ii) whereas 2SDR, which keeps ŷ fair, forces a large bias correction
when the correlation is large. For each set of parameters the result is averaged over 100 independent
runs.
a linear combination of the elements of u and z. Theorem 2 implies that a regressor is asymptotically fair
in the sense of MD (for binary s) or correlation coefficient (for continuous s). Unfortunately, it does not
necessarily guarantee a fair classification under heteroskedasticity: A linear classifier divides datapoints
into two classes by a linear decision boundary (i.e. ŷ is whether a linear combination of u and z is positive
or negative), and no correlation between u, z and s does not necessarily implies no correlation property
between ŷ and s. Still, later in Section 5 we empirically verify the fairness property of 2SDR in both
classification and regression.
Generalization and finite-time analysis: The analysis in this section is very asymptotic and lacks
a finite time bound. As OLS is a parametric model, the standard central limit theorem can be applied to
obtain the asymptotic properties of the 2SDR estimator: Like the 2SLS estimator, the 2SDR estimator
√
is expected to converge at a rate of O(1/ n).
5
Experiments
In the previous section, we provided results suggesting that 2SDR achieves fairness in an asymptotic
sense. To verify the actual performance of 2SDR, we conducted computer simulations. We first describe
its results for a synthetic dataset (Section 5.1), and then describes its results for five real-world datasets
(Section 5.2). Our simulation was implemented in Python by using the scikit-learn library6 . Each of the
simulations took from several seconds to several minutes on a modern PC.
6 http://scikit-learn.org/
9
5.1
Synthetic dataset
This section compares 2SDR with the standard OLS estimator on synthetically-generated datasets. Each
data point (si , xi , zi , yi ) was generated from the following process, which is the standard assumption in
the two-stage regression problem (Section 3.1):
>
yi = x>
i βx + zi βz +
xi =
s>
i βs
(8)
+ ηi
(9)
zi ∼ N (0, σz )
(10)
∼ N (0, σ )
(ηi , si ) ∼ N
(11)
0,
ση
σηs
σηs
σs
!!
.
Obviously, xi and si are correlated, and thus, a naive algorithm that tries to learn Eq. (8) suffers a
disparate impact, whereas 2SDR tries to untangle this dependency by learning the relationship (9) in the
first stage. Unless specified, we set each parameters as follows: dx = dz = 5 and ds = 1. σ = 3.0. ση ,
σz , and σs are diagonal matrices with each diagonal entry is 1.0, and σηs is a matrix with each entry is
0.3. Each entry of βx and βz are 0.5, and each entry of βs is 0.2. The number of datapoint n is set to
1, 000, and 2/3 (resp. 1/3) of the datapoints are used as training (resp. testing) datasets, respectively.
Figure 2 shows the correlation coefficient as as measure of fairness and root mean squared error (RMSE)
as a measure of prediction power for various values of parameters. 2SDR is consistently fair regardless
of the strength of the correlation between x and s.
5.2
Real-world datasets
This section examines the performance of 2SDR in real-world datasets. The primary goal of Section 5.2.1
and 5.2.2 is to compare the results of 2SDR with existing results. We tried to reproduce the settings of
existing papers [5, 10] as much as possible. Section 5.2.3 provides the results with numerical s. Section
5.2.4 provides additional results for other datasets. Results with several other settings are shown in
Section 5.2.5.
Table 2 provides statistics on these datasets. Unless explicitly described, we only put the intercept
attribute (i.e., a constant 1 for all datapoints) into z. We used OLS in each attribute of the first stage,
and OLS (resp. the Ridge classifier) in the regression (resp. classification) of the second stage. Note
that the ridge classifier is a linear model that imposes l2-regularization to avoid very large coefficients,
which performs better when the number of samples is limited. The strength of the regularizer of the
Ridge classifier is optimized by using the ten-fold cross-validation (CV). For binary s, (i) we removed the
attributes of variance conditioned on s = 0 or s = 1 being zero because such a attribute gives a classifier
information that is very close to s, and (ii) we conducted a variance correction after the first stage that
makes the variance of U conditioned on s = 1 and s = 0 identical.
5.2.1
Regression results for C&C dataset
We first show the results of a regression on the Communities and Crime7 dataset that combines socioeconomic data and crime rate data on communities in the United States. Following Calders et al. [5],
we made a binary attribute s as to the percentage of black population, which yielded 970 instances of
s = 1 with a mean crime rate y = 0.35 and 1, 024 instances of s = 0 with a mean crime rate y = 0.13.
7 http://archive.ics.uci.edu/ml/datasets/communities+and+crime
10
Table 2: List of regression or classification datasets. D is the number of binary or numeric attributes
(after expanding unordered categorical attributes into dummies (i.e., set of binary dummy attributes)),
and N is the number of datapoints.
datasets
Adult
Communities & Crime (C&C)
Compas
German
LSAC
Regression
or Classification
Classification
Regression
Classification
Classification
Classification
D
N
49
101
12
47
24
45,222
1,994
5,855
1,000
20,798
Table 3: Regression Results. The scores are averaged result over 10-fold cross validation [5]. The results
of SEM-S and SEM-MP are the ones reported in Calders et al. [5]. “2SDR with explanatory attrs” shows
the result of 2SDR with two explanatory attributes (”FemalePctDiv”,”PctImmigRecent”]) added to z.
A smaller MD indicates better fairness, and an AUC close to 0.5 indicates a very fair regressor. Smaller
RMSE indicates better regression accuracy.
Algorithm
OLS
2SDR
2SDR with explanatory attrs
SEM-S
SEM-MP
MD
0.22
0.02
0.12
0.01
0.17
AUC
0.85
0.48
0.69
0.50
0.76
RMSE
0.14
0.18
0.15
0.20
0.14
Table 4: Classification results for the Adult dataset. The column “Accuracy” presents the classification
accuracy. Unlike OLS, which does not take fairness into consideration, 2SDR complies with the 80%-rule.
Algorithm
OLS
2SDR
P%-rule
0.30
0.83
Accuracy
0.84
0.82
Note that these figures are consistent with the ones reported in Calders et al. [5]. Table 3 shows the
results of the simulation. At a word, 2SDR removes discrimination while minimizing the increase of the
root mean square error (RMSE). One can see that in the sense of RMSE, OLS and SEM-MP [5] perform
the best, although these algorithms do not comply with the two fairness criteria. On the other hand,
2SDR and SEM-S [5] comply with the fairness criteria, and with 2SDR performing better in the sense
of regression among the two algorithms. Furthermore, we put two attributes (“percentage of divorced
females” and “percentage of immigrants in the last three years”) into explanatory attributes z, whose
results are shown as “2SDR with explanatory attrs” in Table 3. One can see that the RMSE of 2SDR
with these explanatory attributes is significantly improved and very close to OLS.
5.2.2
Classification result with Adult and German datasets
This section shows the result of classification with the Adult and German datasets. The adult dataset
is extracted from the 1994 census database, where the target binary attribute indicates whether each
person’s income exceeds 50,000 dollars or not. German is a dataset that classifies people into good or
11
Table 5: Classification results for the German dataset. Unlike OLS, 2SDR complies with the 80%-rule.
The result is averaged over 100 random splits over the training and testing datasets, where two-thirds of
the datapoints are assigned to the training dataset at each split.
Algorithm
OLS
2SDR
P%-rule
0.47
0.81
Accuracy
0.73
0.73
Table 6: Results in the case s is median income (C&C) or age (Adult and German). Note that age was
not binarized in the result of this table.
Algorithm (Dataset)
OLS (C&C)
2SDR (C&C)
OLS (Adult)
2SDR (Adult)
OLS (German)
2SDR (German)
CC
0.50
0.04
0.22
0.07
0.11
0.05
Accuracy
0.84
0.83
0.76
0.75
RMSE
0.14
0.17
-
bad credit risks 8 . Following Zemel et al. and Feldman et al. [24, 10], we used sex (resp. age) in the
Adult (resp. German) datasets. Age in the German dataset is binarized into Young and Old at the age
of 25 [5]. Some sparse attributes in Adult are summarized to reduce dimensionality [22].
Let us compare the results shown in Tables 4 and 5 with the ones reported in previous papers. In a
nutshell, 2SDR, which complies with the 80%-rule, outperforms two data preprocessing methods on the
Adult dataset, and performs as well as them on the German dataset: Zemel et al. [24] reported that their
data transformation combined with a naive Bayes classifier has ∼ 80% (resp. ∼ 70%) accuracy on the
Adult (resp. German) datasets. Moreover, Feldman et al. [10] reported that their data transformation
combined with a Gaussian Naive Bayes classifier had accuracy of 79 ∼ 80% (resp. 70 ∼ 76%) on the
Adult (resp. German) datasets. The method by Zemel et al. [24] coarse-grains the data by mapping them
into a finite space, which we think the reason why its performance is not as good as ours. Meanwhile,
the quantile-based method by Feldman et al. [10] performed impressively well in the German dataset
but not very well in the Adult dataset: In the Adult dataset, it needed to discard most of the attributes
that are binary or categorical, which we consider as the reason for the results.
5.2.3
Numeric s
Next, we considered numeric sensitive attributes. Table 6 shows the accuracy and correlation coefficient
in OLS and 2SDR. On the C&C dataset, 2SDR reduced correlation coefficient (CC) with a minimum
deterioration to its RMSE. In other words, 2SDR was a very efficient at removing the correlation between
ŷ and s.
5.2.4
Other Datasets
Furthermore, we conducted additional experiments on two other datasets (Table 7). The ProPublica
Compas dataset [1] is a collection of criminal offenders screened in Broward County, Florida during
2013-2014, where x is a demographic and criminal record of offenders and y is whether or not a person
recidivated within two years after the screening. We set sex as the sensitive attribute s. The Law School
8 https://archive.ics.uci.edu/ml/datasets/Statlog+(German+Credit+Data)
12
Table 7: Results for the Compas and LSAC datasets. We balanced training data by resampling in
LSAC dataset to cope with class inbalance problem. Compas-R is a version of the Compas dataset where
predictive attributes are dropped: In this version, we dropped the attributes of the original dataset whose
correlation with y was stronger than 0.3. This significantly reduces the prediction accuracy and fairness
of the OLS estimator which tries to utilize the available information as much as possible. Unlike OLS,
the fairness of 2SLS does not decrease even if these attributes are dropped.
Algorithm (Dataset)
OLS (Compas)
2SDR (Compas)
OLS (Compas-R)
2SDR (Compas-R)
OLS (LSAC)
2SDR (LSAC)
P%-rule
0.59
0.92
0.19
0.93
0.21
0.86
Accuracy
0.73
0.73
0.65
0.65
0.75
0.73
Table 8: Performance of 2SDR on the Adult dataset where s is (sex, age). We show p%-rule (resp.
correlation coefficient, CC) with respect to sex (resp. age). Both fairness criteria are improved by using
2SDR.
Algorithm
OLS
2SDR
P%-rule
0.30
0.65
CC
0.22
0.10
Accuracy
0.84
0.82
Admissions Council (LSAC) dataset 9 is a survey among students attending law schools in the U.S. in
1991, where y indicates whether each student passed the first bar examination. We set whether or not
the race of the student is black as the sensitive attribute. Similar to the German dataset, we used 2/3
(resp. 1/3) of the datapoints as training (resp. testing) datasets, and results are averaged over 100
runs. The results, shown in Table 7 implies that 2SDR complies with the 80%-rule with an insignificant
deterioration in classification performance on these datasets.
5.2.5
Other settings
This section shows results with several other settings.
Multiple s: Here, we report the result for multiple sensitive attributes. Table 8 lists the results for
the Adult dataset, where s is sex (binary) and age (numeric). One can see that (i) 2SDR reduces discrimination for both of sensitive attributes with a very small deterioration on the classification performance,
and (ii) the power of removing discrimination is weaker than in the case of applying 2SDR to a single s.
Effect of ordinal transformation: The method proposed by Feldman et al. [10] conducts a
quantile-based transformation. We have also combined the transformation with 2SDR. Let xi,(k) be the
k-th attribute in xi . A quantile-based transformation maps each attribute xi,(k) into its quantile rank
among its sensitive attributes si :
P
j∈{1,2,...,n}:si =sj I[xi,(k) > xi,(k) ]
.
(12)
Ranki,k =
|{j ∈ {1, 2, . . . , n} : si = sj }|
Feldman et al. [10] showed that the dependence between xi,(k) and s can be removed by using such a
quantile-based transformation (c.f. Figure 1 in Feldman et al. [10]). Table 9 and 10 list the results of
9 http://www2.law.ucla.edu/sander/Systemic/Data.htm
13
Table 9: Classification results for the Adult dataset, with or without the ordinal transformation of
Eq. (12). “With ordinal trans.” (resp. “Without ordinal trans.”) indicates an ordinal transformation is
conducted (resp. is not conducted) for each attribute. ”cont. only” indicates that non-numeric attributes
in x are discarded beforehand.
Algorithm
OLS
2SDR
OLS (cont. only)
2SDR (cont. only)
Without ordinal trans.
P%-rule
Accuracy
0.30
0.84
0.83
0.82
0.22
0.81
0.88
0.79
With ordinal trans.
P%-rule Accuracy
0.29
0.83
0.82
0.82
0.06
0.80
0.87
0.78
Table 10: Regression results for the C&C dataset.
Algorithm
OLS
2SDR
Without ordinal trans.
MD
RMSE
0.22
0.14
0.02
0.18
With ordinal trans.
MD
RMSE
0.23
0.16
0.02
0.19
Table 11: Classification result of 2SDR combined with logistic regression.
Algorithm
2SDR
2SDR
dataset
Adult
German
P%-rule
0.72
0.80
Accuracy
0.83
0.73
applying the transformation of Eq. (12) for each non-binary attribute. Applying an ordinal transformation slightly decreased accuracy (or increased RMSE in regression), as it discards the modal information
on the original attribute.
Generalized linear models: We also tried logistic regression in the second stage classifier. Logistic
regression is a binary classification model that assumes the following relation between the attributes xi
and target yi :
1
(13)
P[yi = 1|xi ] =
> ,
1 + e−xi β
where β is the model parameter to be learnt. Table 11 shows the results of classification when we replaced
the second-stage classifier with the logistic regression. Compared with a linear model (Ridge classifier),
this yielded a lower p%-rule in the Adult dataset. This fact is consistent with Theorem 2. It states that u
is asymptotically uncorrelated to s: However, a non-linear map such as the sigmoid function in Eq. (13)
can cause bias between the mapped u and s. We should also note that the more involved non-linear second
stage classifiers, such as naive Bayes classifiers, support vector machines, and gradient boosting machines,
resulted in a significantly lower p%-rule than logistic regression because of their strong non-linearity.
6
Conclusion
We studied indirect discrimination in classification and regression tasks. In particular, we studied a twostage method to reduce disparate impact. Our method is conceptually simple and has a wide range of
potential applications. Unlike most of the existing methods, our method is general enough to deal with
14
both classification and regression with various settings. It lies midway between a fair data preprocessing
and a fair estimator: It conducts a minimum transformation so that linear algorithm in the second stage is
fair. Extensive evaluations showed that our method complied the 80%-rule the tested real-world datasets.
The following are possible directions of future research:
• Other criteria of fairness: While the disparate impact considered in this paper is motivated
by the laws in the United States, the notion of fairness is not limited to disparate impact [3]. To
name a few studies, the equalized odds condition [13] and disparate mistreatment [22] have been
considered. Extending our method to other criteria of fairness would be interesting.
• Non-Linear second stage: In this study, we restricted the second-stage algorithm to be linear.
The main reason for doing so is that the first stage in 2SDR is designed to remove the correlation
between ŷ and s, which is very suitable to linear algorithms (Theorem 1 and 2). We have also conducted some experiment with generalized linear model in the second stage (Section 5.2.5), where we
observed a inferior fairness than a linear model. Extending our work to a larger class of algorithms
would boost the accuracy of 2SDR on some datasets where non-linearity is important.
15
References
[1] Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner. Machine Bias: Theres software used
across the country to predict future criminals. May 23, 2016.
[2] Richard Berk, Hoda Heidari, Shahin Jabbari, Matthew Joseph, Michael Kearns, Jamie Morgenstern,
Seth Neel, and Aaron Roth. A convex framework for fair regression. In Fairness, Accountability,
and Transparency in Machine Learning (FATML), 2017.
[3] Richard Berk, Hoda Heidari, Shahin Jabbari, Michael Kearns, and Aaron Roth. Fairness in criminal
justice risk assessments: The state of the art. arXiv preprint arXiv:1703.09207, 2017.
[4] T. Bolukbasi, K.-W. Chang, J. Y. Zou, V. Saligrama, and A. T. Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in Neural Information
Processing Systems 23, 2016.
[5] Toon Calders, Asim Karim, Faisal Kamiran, Wasif Ali, and Xiangliang Zhang. Controlling attribute
effect in linear regression. In 2013 IEEE 13th International Conference on Data Mining, Dallas,
TX, USA, December 7-10, 2013, pages 71–80, 2013.
[6] The United States Civil Rights Center. Age Discrimination Act. November 28, 1975.
[7] The United States Equal Employment Opportunity Commission. Uniform guidelines on employee
selection procedures. March 2, 1979.
[8] The United States Congress. Civil Rights Act. July 2, 1964.
[9] The United States Supreme Court. Case: Griggs v. Duke Power Co. 401 U.S. 424. March 8, 1971.
[10] Michael Feldman, Sorelle A. Friedler, John Moeller, Carlos Scheidegger, and Suresh Venkatasubramanian. Certifying and removing disparate impact. In Proceedings of the 21th ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, Sydney, NSW, Australia, August 10-13, 2015, pages 259–268, 2015.
[11] Benjamin Fish, Jeremy Kun, and Ádám Dániel Lelkes. Fair boosting: a case study. In 2nd Workshop
on Fairness, Accountability, and Transparency in Machine Learning (FATML 2015), 2015.
[12] Gabriel Goh, Andrew Cotter, Maya R. Gupta, and Michael P. Friedlander. Satisfying real-world
goals with dataset constraints. In Advances in Neural Information Processing Systems 29: Annual
Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain,
pages 2415–2423, 2016.
[13] Moritz Hardt, Eric Price, and Nati Srebro. Equality of opportunity in supervised learning. In
Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information
Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 3315–3323, 2016.
[14] Matthew Joseph, Michael Kearns, Jamie H. Morgenstern, and Aaron Roth. Fairness in learning:
Classic and contextual bandits. In Advances in Neural Information Processing Systems 29: Annual
Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain,
pages 325–333, 2016.
[15] Faisal Kamiran and T Calders. Classification with no discrimination by preferential sampling. In
The annual machine learning conference of Belgium and The Netherlands (BENELEARN), 01 2010.
16
[16] Toshihiro Kamishima, Shotaro Akaho, Hideki Asoh, and Jun Sakuma. Enhancement of the neutrality in recommendation. In Proceedings of the 2nd Workshop on Human Decision Making in
Recommender Systems, Dublin, Ireland, September 9, 2012, pages 8–14, 2012.
[17] Toshihiro Kamishima, Shotaro Akaho, Hideki Asoh, and Jun Sakuma. Fairness-aware classifier
with prejudice remover regularizer. In Machine Learning and Knowledge Discovery in Databases European Conference, ECML PKDD 2012, Bristol, UK, September 24-28, 2012. Proceedings, Part
II, pages 35–50, 2012.
[18] Toshihiro Kamishima, Shotaro Akaho, Hideki Asoh, and Issei Sato. Model-based approaches for
independence-enhanced recommendation. In IEEE International Conference on Data Mining Workshops, ICDM Workshops 2016, December 12-15, 2016, Barcelona, Spain., pages 860–867, 2016.
[19] Peter Norman. Statistical discrimination and efficiency. The Review of Economic Studies, 70(3):615–
627, 2003.
[20] Goce Ristanoski, Wei Liu, and James Bailey. Discrimination aware classification for imbalanced
datasets. In 22nd ACM International Conference on Information and Knowledge Management,
CIKM’13, San Francisco, CA, USA, October 27 - November 1, 2013, pages 1529–1532, 2013.
[21] Jeffrey M. Wooldridge. Introductory Econometrics: A Modern Approach. South-Western, Cengage
Learning, 5th edition, 2013.
[22] Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez-Rodriguez, and Krishna P. Gummadi. Fairness
beyond disparate treatment & disparate impact: Learning classification without disparate mistreatment. In Proceedings of the 26th International Conference on World Wide Web, WWW 2017, Perth,
Australia, April 3-7, 2017, pages 1171–1180, 2017.
[23] Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez-Rodriguez, and Krishna P. Gummadi. Fairness
constraints: Mechanisms for fair classification. In Proceedings of the 20th International Conference
on Artificial Intelligence and Statistics, AISTATS 2017, 20-22 April 2017, Fort Lauderdale, FL,
USA, pages 962–970, 2017.
[24] Richard S. Zemel, Yu Wu, Kevin Swersky, Toniann Pitassi, and Cynthia Dwork. Learning fair
representations. In Proceedings of the 30th International Conference on Machine Learning, ICML
2013, Atlanta, GA, USA, 16-21 June 2013, pages 325–333, 2013.
[25] Indre Zliobaite, Faisal Kamiran, and Toon Calders. Handling conditional discrimination. In 11th
IEEE International Conference on Data Mining, ICDM 2011, Vancouver, BC, Canada, December
11-14, 2011, pages 992–1001, 2011.
17
| 2cs.AI
|
Point-primitive, line-transitive generalised quadrangles of
holomorph type
arXiv:1508.02295v4 [math.GR] 4 May 2016
John Bamberg, Tomasz Popiel, Cheryl E. Praeger
Abstract. Let G be a group of collineations of a finite thick generalised quadrangle Γ. Suppose that
G acts primitively on the point set P of Γ, and transitively on the lines of Γ. We show that the
primitive action of G on P cannot be of holomorph simple or holomorph compound type. In joint
work with Glasby, we have previously classified the examples Γ for which the action of G on P is of
affine type. The problem of classifying generalised quadrangles with a point-primitive, line-transitive
collineation group is therefore reduced to the case where there is a unique minimal normal subgroup
M and M is non-Abelian.
1. Introduction
A partial linear space is a point–line incidence geometry in which any two distinct points are
incident with at most one line. All partial linear spaces considered in this paper are assumed to be
finite. A generalised quadrangle Q is a partial linear space that satisfies the generalised quadrangle
axiom: given a point P and line ℓ not incident with P , there is a unique line incident with P and
concurrent with ℓ. This axiom implies, in particular, that Q contains no triangles. If each point of
Q is incident with at least three lines, and each line is incident with at least three points, then Q is
said to be thick. In this case, there exist constants s, t > 2 such that each point (line) is incident with
exactly t + 1 lines (s + 1 points), and (s, t) is called the order of Q. Generalised quadrangles were
introduced by Tits [9], together with the other generalised polygons, in an attempt to find a systematic
geometric interpretation for the simple groups of Lie type. It is therefore very natural to ask which
groups arise as collineation groups of generalised quadrangles.
A topic of particular interest is that of generalised quadrangles admitting collineation groups
M that act regularly on points, where the point set is identified with M acting on itself by right
multiplication. Ghinelli [6] showed that a Frobenius group or a group with non-trivial centre cannot
act regularly on the points of a generalised quadrangle of order (s, t) if s is even and s = t, and
Yoshiara [10] showed that a generalised quadrangle with s = t2 does not admit a point-regular
collineation group. Regular groups arise, in particular, as subgroups of certain primitive groups.
Bamberg et al. [2] showed that a group G acting primitively on both the points and the lines of a
generalised quadrangle must be almost simple. The present authors and Glasby [3, Corollary 1.5]
sought to weaken this assumption to primitivity on points and transitivity on lines, and, using a result
of De Winter and Thas [4], classified the generalised quadrangles admitting such a group in the case
where the primitive action on points is of affine type. (There are only two examples, arising from
hyperovals in PG(2, 4) and PG(2, 16).) In this case, the regular subgroup M of G is Abelian, and hence
left multiplication by any element of M is also a collineation. We consider the situation where M is
non-Abelian but G has a second minimal normal subgroup, which is necessarily the centraliser of M ,
so that all left multiplications are again collineations. In the context of the O’Nan–Scott Theorem [8,
2010 Mathematics Subject Classification. primary 51E12; secondary 20B15, 05B25.
Key words and phrases. generalised quadrangle, primitive permutation group.
The first author acknowledges the support of the Australian Research Council (ARC) Future Fellowship
FT120100036. The second author acknowledges the support of the ARC Discovery Grant DP1401000416. The research
reported in the paper forms part of the ARC Discovery Grant DP1401000416 of the third author.
1
2
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
Section 5] for primitive permutation groups, this means that the action of G on points is of either
holomorph simple (HS) or holomorph compound (HC) type (see Section 2 for definitions). We prove
the following result.
Theorem 1.1. Let G be a collineation group of a finite thick generalised quadrangle with point
set P and line set L. If G acts transitively on L and primitively on P, then G has a unique minimal
normal subgroup; that is, the action of G on P does not have O’Nan–Scott type HS or HC.
The proof of Theorem 1.1 is given in Sections 3 and 4, using some preliminary results established
in Section 2, and the Classification of Finite Simple Groups.
2. Preliminaries
We first recall some definitions and facts about permutation groups. Let G be a group acting on
a set Ω, and denote the image of x ∈ Ω under g ∈ G by xg . The orbit of x ∈ Ω under G is the
set xG = {xg | g ∈ G}, the subgroup Gx = {g ∈ G | xg = x} is the stabiliser of x ∈ Ω, and the
Orbit–Stabiliser Theorem says that |G : Gx | = |xG |. The action of G is transitive if xG = Ω for some
(and hence every) x ∈ Ω, and semiregular if Gx is trivial for all x ∈ Ω. It is regular if it is both
transitive and semiregular. If G acts transitively on Ω and M is normal subgroup of G, then all orbits
of M on Ω have the same length, and in particular it makes sense to speak of M being semiregular.
Given g ∈ G, define ρg , λg , ιg ∈ Sym(Ω) by ρg : x 7→ xg, λg : x 7→ g−1 x, and ιg : x 7→ g −1 xg.
Set GR = {ρg : g ∈ G}, GL = {λg : g ∈ G}, and Inn(G) = {ιg : g ∈ G}. The holomorph Hol(G)
of G is the semidirect product GR ⋊ Aut(G) with respect to the natural action of Aut(G) on GR [1,
Section 2.6]. We have Hol(G) = NSym(G) (GR ), and GL = CSym(G) (GR ). A group H acting on a set ∆
is permutationally isomorphic to G acting on Ω if there is an isomorphism θ : G → H and a bijection
β : Ω → ∆ such that β(ω g ) = β(ω)θ(g) for all g ∈ G and ω ∈ Ω. If a group M acts regularly on Ω, then
there is a permutational isomorphism θ : NSym(Ω) (M ) → Hol(M ) with bijection β : Ω → M , where
β : αg 7→ g for some fixed α ∈ Ω, and θ : τ 7→ β −1 τ β. We have θ(M ) = MR , so the regular action of
M on Ω is permutationally isomorphic to the action of M on itself by right multiplication, and hence
we can identify Ω with M . Furthermore, θ(CSym(Ω) (M )) = ML . If M is a normal subgroup of G, then
G is permutationally isomorphic to a subgroup of Hol(M ). If M ⋊ Inn(M ) 6 G, then G contains ML
because M ⋊ Inn(M ) = hMR , ML i.
A transitive action of G on Ω is said to be primitive if it preserves no non-trivial partition of Ω.
The structure of a primitive permutation group is described by the O’Nan–Scott Theorem [8, Section
5], which splits the primitive permutation groups into eight types. We are concerned with only two
of these types. If M ⋊ Inn(M ) 6 G 6 M ⋊ Aut(M ) with M ∼
= T for some non-Abelian finite simple
group T , then G, being contained in the holomorph of a simple group, is said to have type HS. If
instead M is isomorphic to a compound group T k , k > 2, then G has type HC. In this case, G induces
a subgroup of Aut(M ) ∼
= Aut(T ) ≀ Sk which acts transitively on the set of k simple direct factors of
k
∼
M = T . In either case, G contains MR and ML , as explained above.
If we write S = (P, L, I ) for a partial linear space, then we mean that P is the point set, L is the
line set, and I is the incidence relation. An incident point–line pair is called a flag. A collineation
of S is a permutation of P, together with a permutation of L, such that incidence is preserved. If
S admits a group of collineations M that acts regularly on P, then we identity P with M acting on
itself by right multiplication (as above). A line ℓ is then identified with the subset of M comprising
all of the points incident with ℓ, and hence P I ℓ if and only if P ∈ ℓ. Moreover, the stabiliser Mℓ is
the set of all elements of M that fix ℓ setwise by right multiplication.
Lemma 2.1. Let S = (P, L, I ) be a partial linear space with no triangles, and let G be a group of
collineations of S with a normal subgroup M that acts regularly on P. Let ℓ be a line incident with
the identity 1 ∈ M = P, and suppose that its stabiliser Mℓ is non-trivial. Then
(i) ℓ is a union of left Mℓ -cosets, including the trivial coset;
(ii) if M ⋊ Inn(M ) 6 G, then Mℓ = ℓ.
Proof. (i) Let g ∈ Mℓ . Since 1 I ℓ, namely 1 ∈ ℓ, we have g = 1g I ℓg = ℓ, namely g ∈ ℓ. Therefore,
Mℓ ⊆ ℓ. Now, if h 6∈ Mℓ \{1} is incident with ℓ, then every non-trivial element of Mℓ must map h to
another point incident with ℓ, and hence the whole coset hMℓ is contained in ℓ.
PRIMITIVE GENERALISED QUADRANGLES OF HOLOMORPH TYPE
3
(ii) By (i), Mℓ ⊆ ℓ, so it remains to show the reverse inclusion. Let m ∈ ℓ \ {1}. Since Mℓ is
non-trivial, there exists a non-trivial element h ∈ Mℓ . Since M ⋊ Inn(M ) 6 G, left multiplication
by h−1 is a collineation of S. Since 1 and m are both incident with ℓ, it follows that h−1 and h−1 m
are collinear. On the other hand, h−1 ∈ Mℓ ⊆ ℓ by (i), so h−1 m is collinear with m because right
multiplication by m is a collineation. That is, h−1 m is collinear with two points h−1 , m that are
incident with ℓ, and so h−1 m is itself incident with ℓ because S contains no triangles. Therefore, m
maps two points 1, h−1 incident with ℓ to two points m, h−1 m incident with ℓ, and so m ∈ Mℓ .
Theorem 2.2. Let S = (P, L, I ) be a partial linear space with no triangles. Let G be a group
of collineations of S that acts transitively on L, and suppose that G has a normal subgroup M that
acts regularly on P and satisfies M ⋊ Inn(M ) 6 G 6 M ⋊ Aut(M ). If the action of M on L is not
semiregular, then the lines ℓ1 , . . . , ℓt+1 incident with 1 are a G1 -conjugacy class of subgroups of M ,
and G acts transitively on the flags of S.
Proof. Since M acts transitively on P, we have G = M G1 = G1 M . By assumption, G 6 Hol(M )
and so G1 6 Aut(M ). By Lemma 2.1(ii), the lines ℓ1 , . . . , ℓt+1 can be identified with subgroups of M .
Each g ∈ G1 , acting naturally as an element of Aut(M ), fixes 1 and hence maps ℓ1 to ℓg1 = ℓi for some
i ∈ {1, . . . , t+1}. Conversely, consider the map ϕ : G → Aut(M ) defined by ϕ(g) = ιg . The restriction
of ϕ to G1 is the identity. Moreover, ker(ϕ) = CG (M ), and hence θ(ker(ϕ)) = ML , where θ is the
permutational isomorphism defined above. In particular, ker(ϕ) acts transitively (indeed, regularly)
on P. Hence, ker(ϕ)G1 = G, so Im(ϕ) = ϕ(G1 ) = G1 . Now consider a line ℓi for some i > 1. By
line-transitivity, ℓi = ℓg1 for some g ∈ G. On the other hand, since G = ker(ϕ)G1 , we have g = zg1 for
some z ∈ ker(ϕ) and g1 ∈ G1 , so ℓg1 = ℓg11 . Therefore, ℓ1 , . . . , ℓt+1 are precisely the subgroups of the
form ℓg1 with g ∈ G1 . Since the lines ℓi and ℓj intersect precisely in the point 1 for i 6= j, the t + 1
subgroups ℓ1 , . . . , ℓt+1 are distinct, and they form a single G1 -conjugacy class of subgroups of M . In
particular, G1 acts transitively on {ℓ1 , . . . , ℓt+1 }, so G acts transitively on the flags of S.
Let us draw a corollary in the case where S is a thick generalised quadrangle. In this case, S has
(s + 1)(st + 1) points and (t + 1)(st + 1) lines, where (s, t) is the order of S.
Corollary 2.3. If the partial linear space in Theorem 2.2 is a thick generalised quadrangle of
order (s, t), then s + 1 divides t − 1.
Proof. Begin by observing that Inn(M ) acts on {ℓ1 , . . . , ℓt+1 }. That is, for each g ∈ M , we have
g −1 ℓ1 g = ℓi for some i ∈ {1, . . . , t + 1}. Suppose first that Inn(M ) is intransitive on {ℓ1 , . . . , ℓt+1 }.
Then, without loss of generality, ℓ2 is in a different Inn(M )-orbit to ℓ1 , and so, for every g ∈ M ,
we have g−1 ℓ1 g = ℓi for some i 6= 2. Hence, every double coset ℓ1 gℓ2 , where g ∈ M , has size
|ℓ1 gℓ2 | = |g−1 ℓ1 gℓ2 | = |ℓi ℓ2 | = (s + 1)2 . Here the final equality holds because |ℓi ∩ ℓ2 | = 1 (because
distinct concurrent lines intersect in a unique point, in this case the point 1). Since the double cosets
of ℓ1 and ℓ2 partition M , it follows that (s + 1)2 divides |M | = |P| = (s + 1)(st + 1). Therefore, s + 1
divides st + 1 = (s + 1)t − (t − 1), and hence s + 1 divides t − 1, as claimed.
Now suppose, towards a contradiction, that Inn(M ) is transitive on {ℓ1 , . . . , ℓt+1 }. Consider two
lines incident with 1, say ℓ1 , ℓ2 . Then a double coset ℓ1 gℓ2 , where g ∈ M , has size (s + 1)2 or s + 1
according as g −1 ℓ1 g 6= ℓ2 or g−1 ℓ1 g = ℓ2 . There are exactly |M |/(t + 1) elements g ∈ M for which
g −1 ℓ1 g = ℓ2 , and since ℓ1 hℓ2 = ℓ1 gℓ2 if and only if h ∈ ℓ1 gℓ2 (where h ∈ M ), it follows that there
are precisely |M |/((s + 1)(t + 1)) double cosets of size s + 1. Therefore, (s + 1)(t + 1) must divide
|M | = |P| = (s + 1)(st + 1), and so t + 1 must divide st + 1 = (t + 1)s − (s − 1) and hence s − 1. In
particular, we have s > t + 2 > t, and so [7, 2.2.2(i)] implies that S cannot contain a subquadrangle
of order (s, 1). For a contradiction, we now construct such a subquadrangle.
Consider the subset P ′ = ℓ1 ℓ2 of P = M , let L′ = {g1 ℓ2 | g1 ∈ ℓ1 } ∪ {ℓ1 g2 | g2 ∈ ℓ2 }, and let I ′ be
the restriction of I to (P ′ × L′ ) ∪ (L′ × P ′ ). We claim that S ′ = (P ′ , L′ , I ′ ) is a subquadrangle of S of
order (s, 1). First observe that, for each ℓ ∈ L′ and each P ∈ P ′ not incident with ℓ, the unique point
incident with ℓ and collinear with P lies in P ′ , because ℓ ⊂ P ′ . Hence, S ′ satisfies the generalised
quadrangle axiom. Now, every line in L′ is incident with s + 1 points in P ′ , being a coset of either
ℓ1 or ℓ2 , so it remains to show that every point in P ′ is incident with exactly two lines in L′ . Given
P = g1 g2 ∈ P ′ , where g1 ∈ ℓ1 , g2 ∈ ℓ2 , each line ℓ ∈ L′ incident with P is either of the form h1 ℓ2 for
4
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
some h1 ∈ ℓ1 or ℓ1 h2 for some h2 ∈ ℓ2 , and since P ∈ ℓ, we must have h1 = g1 or h2 = g2 , respectively.
Therefore, P is incident with exactly two lines in L′ , namely g1 ℓ2 and ℓ1 g2 .
We also check that, in the case of a thick generalised quadrangle, the assumption that M is not
semiregular on L is satisifed when |M | is even.
Lemma 2.4. Let Q = (P, L, I ) be a thick generalised quadrangle of order (s, t). Let G be a group
of collineations of Q that acts transitively on L, and suppose that G has a normal subgroup M that
acts regularly on P. If M has even order, then M does not act semiregularly on L.
Proof. If Mℓ is trivial for ℓ ∈ L, then |ℓM | = |M | = |P| = (s + 1)(st + 1) divides |L| = (t + 1)(st + 1),
and hence s + 1 divides t + 1, so [2, Lemma 3.2] implies that gcd(s, t) > 1. However, |M | is even,
so M contains an element of order 2, and because gcd(s, t) > 1, it follows from [2, Lemma 3.4] that
every such element must fix some line, contradicting the assumption that Mℓ is trivial.
3. Proof of Theorem 1.1: HS type
Suppose that Q = (P, L, I ) is a thick generalised quadrangle with a collineation group G that
acts transitively on L and primitively of O’Nan–Scott type HS on P. Then
T ⋊ Inn(T ) 6 G 6 T ⋊ Aut(T )
for some non-Abelian finite simple group T , with T acting regularly on P. Since |T | is even by the
Feit–Thompson Theorem [5], Lemma 2.4 tells us that Q satisfies the hypotheses of Theorem 2.2 and
Corollary 2.3. In particular, s + 1 divides t − 1 (by Corollary 2.3), and we write
t−1
.
(1)
t′ :=
s+1
Since T acts regularly on P, we have |T | = |P| = (s + 1)(st + 1). By Higman’s inequality, t 6 s2 , and
hence t′ 6 s − 1. Therefore,
|T | = (s + 1)2 (st′ + 1) for some 1 6 t′ 6 s − 1.
By Theorem 2.2, G1 6 Aut(T ) acts transitively on the t + 1 lines incident with 1, and hence t + 1
divides | Aut(T )| = |T | · | Out(T )|. Therefore, | Out(T )| is divisible by (t + 1)/ gcd(t + 1, |T |), so
t+1 6 gcd(t+1, |T |)| Out(T )|. Since |T | = (s+1)(st+1) is even, s must be odd; and since s+1 divides
t − 1, we have gcd(t + 1, s + 1) = 2. Moreover, st′ + 1 = t − t′ , so gcd(t + 1, st′ + 1) = gcd(t + 1, t − t′ ) =
gcd(t + 1, t′ + 1), and in particular gcd(t + 1, |T |) 6 22 (t′ + 1). Therefore, t + 1 6 4(t′ + 1)| Out(T )|.
Together with (1), this implies t′ (s + 1) + 2 6 4(t′ + 1)| Out(T )|, and because t′ > 1, it follows that
s 6 8| Out(T )| − 3.
Since |T | 6 (s + 1)(s3 + 1) (by Higman’s inequality), we have
|T | 6 (8| Out(T )| − 2)((8| Out(T )| − 3)3 + 1).
The following lemma therefore completes the proof of Theorem 1.1 in the HS case.
Lemma 3.1. There is no finite non-Abelian simple group T satisfying
(a) |T | = (s + 1)2 (st′ + 1), where 1 6 t′ 6 s − 1;
(b) 2 6 s 6 8| Out(T )| − 3; and
(c) |T | 6 (8| Out(T )| − 2)((8| Out(T )| − 3)3 + 1).
Proof. Since (8x − 2)((8x − 3)3 + 1) ≤ (8x)4 for real x > 1, condition (c) implies that
(2)
|T | 6 212 | Out(T )|4 .
We use (2) instead of (c) to rule out certain possibilities for T .
Case 1. T ∼
= Altn or a sporadic simple group:
If T ∼
= Alt6 , then | Out(T )| = 4 and there is no solution to (a) subject to (b). If T is an alternating
group other than Alt6 , or a sporadic simple group, then | Out(T )| 6 2, and so (c) implies that
|T | 6 (13 + 1)(133 + 1) = 30 772. This rules out everything except T ∼
= Alt5 , Alt7 and M11 , and for
these cases one checks that there is no solution to (a) subject to (b).
PRIMITIVE GENERALISED QUADRANGLES OF HOLOMORPH TYPE
5
Case 2. T ∼
= A1 (q):
Suppose that T ∼
= A1 (q), and write q = pf with p prime and f > 1. Then |T | = q(q 2 −1)/ gcd(2, q−1),
and | Out(T )| = gcd(2, q − 1)f .
Suppose first that q is even, namely that p = 2. Then gcd(2, q − 1) = 1, and (c) implies that
2f (22f − 1) 6 (8f − 2)((8f − 3)3 + 1),
which holds only if f 6 7. If f = 1, then T is not simple; and if f = 2, then T ∼
= Alt5 , which we have
already ruled out. For 3 6 f 6 7, there is no solution to (a) subject to (b).
Now suppose that q = pf is odd. Then gcd(2, q − 1) = 2, and hence | Out(T )| = 2f , so (c) reads
pf (p2f − 1) 6 2(16f − 2)((16f − 3)3 + 1).
If f > 6, then this inequality fails for all p > 3. The inequality holds if and only if
q = pf ∈ {3, 5, 7, 32 , 11, 13, 17, 19, 23, 52 , 33 , 29, 31, 37, 72 , 34 , 53 , 35 }.
If q = 3, then T is not simple; if q = 5, then T ∼
= Alt5 , which we have ruled out; if q = 7, then
∼
T = A2 (2), which is ruled out in Case 3 below; and if q = 9, then T ∼
= Alt6 , which we have ruled out.
For the remaining values of q, there is no solution to (a) subject to (b).
Case 3. T ∼
= An (q), n > 2:
Suppose that T ∼
= An (q), with n > 2 and q = pf . Then
n
Y
q n(n+1)/2
|T | =
(q i+1 − 1),
gcd(n + 1, q − 1)
i=1
and | Out(T )| = 2 gcd(n + 1, q − 1)f .
First suppose that n > 3. Noting that f = logp (q) = ln(q)/ ln(p) 6 ln(q)/ ln(2) and gcd(n + 1, q −
1) 6 q − 1, and applying (2), we find
q
n(n+1)/2
n
Y
(q i+1 − 1) 6
i=1
216
(q − 1)5 ln4 (q).
4
ln (2)
This inequality fails for all q > 2 if n = 4, and therefore fails for all q > 2 for every n > 4 (because
the left-hand side is increasing in n while the right-hand side does not depend on n). It fails for
n = 3 unless q ∈ {2, 3}, but A3 (2) ∼
= Alt8 has already been ruled out, and (c) rules out A3 (3) because
|A3 (3)| = 6 065 280 > 30(293 + 1) = 731 700.
Finally, suppose that n = 2. Noting that gcd(3, q − 1) 6 3 and f 6 ln(q)/ ln(2), (2) gives
q 3 (q 2 − 1)(q 3 − 1) 6
216 35 4
ln (q).
ln4 (2)
This implies that q 6 15. For q ∈ {5, 8, 9, 11, 13}, the sharper inequality (c) fails. For q ∈ {2, 3, 4, 7},
there are no solutions to (a) subject to (b).
Case 4. T ∼
= 2 An (q 2 ):
Suppose that T ∼
= 2 An (q 2 ), where now q 2 = pf for some prime p and f ≥ 1. We have n > 2,
n
|T | =
Y
q n(n+1)/2
(q i+1 − (−1)i+1 ),
gcd(n + 1, q + 1)
i=1
and | Out(T )| = gcd(n + 1, q + 1)f .
First suppose that n > 4. Noting that f = logp (q) = ln(q 2 )/ ln(p) 6 2 ln(q)/ ln(2), and that
gcd(n + 1, q + 1) 6 q + 1, (2) gives
q n(n+1)/2
n
Y
216
(q i+1 − (−1)i+1 ) 6 4 (q + 1)5 ln4 (q).
ln (2)
i=1
This inequality fails for all q > 2 for n = 4, and hence fails for all q > 2 for every n > 4.
6
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
45
Now suppose that n = 3. Then we can replace the (q + 1)5 on the right-hand side above by
= 210 , because gcd(n + 1, q + 1) = gcd(4, q + 1) 6 4. This yields
q 6 (q 2 − 1)(q 3 + 1)(q 4 − 1) 6
226
ln4 (q),
4
ln (2)
which implies that q 6 4. If q ∈ {2, 3}, then there are no solutions to (a) subject to (b). If q = 4,
then (c) fails.
Finally, suppose that n = 2. Then gcd(n + 1, q + 1) 6 3, and hence
q 3 (q 2 − 1)(q 3 + 1) 6
216 35 4
ln (q),
ln4 (2)
which implies that q 6 15. If q = 2, then T ∼
= 2 A2 (22 ) is not simple. If q ∈ {3, 4, 5, 8}, then there are
no solutions to (a) subject to (b). If q ∈ {7, 9, 11, 13}, then (c) fails.
Case 5. Remaining possibilities for T :
We now rule out the remaining possibilities for the finite simple group T .
(i) T ∼
= Bn (q) or Cn (q). First suppose that T ∼
= Cn (q), and write q = pf with p prime and f > 1.
Q
2
n
We have n > 3, |T | = q n / gcd(2, q − 1) · i=1 (q 2i − 1), and | Out(T )| = gcd(2, q − 1)f . Noting that
f 6 ln(q)/ ln(2) and gcd(2, q − 1) 6 2, (2) implies that
q
n2
n
Y
217
ln4 (q).
(q 2i − 1) 6 4
ln
(2)
i=1
However, this inequality fails for all q > 2 if n = 3, and hence fails for all q > 2 for every n > 3.
Now suppose that T ∼
= Bn (q), writing q = pf as before. In this case we have n > 2, and again
Q
2
|T | = q n / gcd(2, q − 1) · ni=1 (q 2i − 1). If n > 3 and q is even, then Bn (q) ∼
= Cn (q). If n > 3 and q
is odd, then | Out(T )| is the same as for Cn (q). We may therefore assume that n = 2. First suppose
that q = 2f . Then | Out(T )| = 2 gcd(2, q − 1)f = 2f , so (2) implies that
24f (22f − 1)(24f − 1) 6 216 f 4 ,
(3)
∼ Alt6 is
and hence f ∈ {1, 2}. For f = 1, B2 (2) is not simple but its derived subgroup B2 (2)′ =
simple and has already been ruled out. For f = 2, (c) fails. Now suppose that q is odd. Then
| Out(T )| = gcd(2, q − 1)f = 2f and f 6 ln(q)/ ln(3), so (2) implies that
q 4 (q 2 − 1)(q 4 − 1) 6
217
ln4 (q),
4
ln (3)
and hence q = 3. However, B2 (3) ∼
= 2 A3 (22 ), which has been dealt with in Case 4.
(ii) T ∼
T ∼ D (q), writing q = pf again. We have n > 4, |T | =
= Dn (q). Suppose that
Qn−1 2i= n
n(n−1)
n
n
q
(q − 1)/ gcd(4, q − 1) · i=1 (q − 1), and
2
6 gcd(2, q − 1) f if n = 4
| Out(T )| = 2 gcd(2, q − 1)2 f if n < 4 and n is even
2 gcd(4, q n − 1)f if n < 4 and n is odd.
If q is odd, then gcd(4, q n − 1) 6 4, | Out(T )| 6 24f , and f 6 ln(q)/ ln(3), so (2) implies that
q
n(n−1)
n
(q − 1)
n−1
Y
(q 2i − 1) 6
i=1
226 34 4
ln (q),
ln4 (3)
which fails for all q > 3 if n = 4, and hence fails for all q > 3 for every n > 4. If q is even, then
gcd(4, q n − 1) = 1, | Out(T )| 6 6f and f = ln(q)/ ln(2), so (2) implies that
q
n(n−1)
n
(q − 1)
n−1
Y
i=1
(q 2i − 1) 6
216 34 4
ln (q),
ln4 (2)
which fails for all q > 2 if n = 4, and hence fails for all q > 2 for every n > 4.
PRIMITIVE GENERALISED QUADRANGLES OF HOLOMORPH TYPE
7
(iii) T ∼
= E6 (q), E7 (q), E8 (q) or F4 (q). Suppose that T is one of E6 (q), E7 (q), E8 (q) or F4 (q), and
write q = pf again. Observe that |Ei (q)| > |F4 (q)| for every i ∈ {6, 7, 8}, for all q > 2. Hence
|T | > |F4 (q)| = q 24 (q 12 − 1)(q 8 − 1)(q 6 − 1)(q 2 − 1) >
q 52
.
24
Since | Out(T )| 6 2 gcd(3, q − 1)f 6 6 ln(q)/ ln(2), (2) implies the following inequality, which fails for
all q > 2:
220 34 4
ln (q).
q 52 6 4
ln (2)
(iv) T ∼
= G2 (q). Suppose that T ∼
= G2 (q), with q = pf . Then |T | = q 6 (q 6 − 1)(q 2 − 1). If p = 3,
then | Out(T )| = 2f , so (c) implies that 36f (36f − 1)(32f − 1) 6 216 f 4 , which fails for all f > 1. If
p 6= 3, then | Out(T )| = f 6 ln(q)/ ln(2), and (2) implies the following inequality, which fails for all
q > 2:
212
q 6 (q 6 − 1)(q 2 − 1) 6 4
ln4 (q).
ln (2)
Note that G2 (2) is not simple, but G2 (2)′ ∼
= 2 A2 (32 ) is simple and has already been ruled out.
2
2
∼
∼
(v) T = Dn (q). Suppose that T = Dn (q 2 ), now writing q 2 = pf . Then n > 4,
n−1
q n(n−1) (q n + 1) Y 2i
(q − 1),
|T | =
gcd(4, q n + 1)
i=1
and | Out(T )| =
gcd(4, q n
+ 1)f . Since f 6 2 ln(q)/ ln(2) and gcd(4, q n + 1) 6 4, (2) implies that
q n(n−1) (q n + 1)
n−1
Y
i=1
(q 2i − 1) 6
226
ln4 (q).
ln4 (2)
This fails for all q > 2 if n = 4, and hence fails for all q > 2 for every n > 4.
(vi) T ∼
= 2 E6 (q 2 ). Suppose that T ∼
= 2 E6 (q 2 ), with q 2 = pf . Then
|T | =
1
q 36 (q 12 − 1)(q 9 + 1)(q 8 − 1)(q 6 − 1)(q 5 + 1)(q 2 − 1),
(3, q + 1)
and | Out(T )| = gcd(3, q + 1)f . Noting that f 6 2 ln(q)/ ln(2) and gcd(3, q + 1) 6 3, (2) implies the
following inequality, which fails for all q > 2:
q 36 (q 12 − 1)(q 9 + 1)(q 8 − 1)(q 6 − 1)(q 5 + 1)(q 2 − 1) 6
35 216 4
ln (q).
ln4 (2)
(vii) T ∼
= 3 D4 (q 3 ). Suppose that T ∼
= 3 D4 (q 2 ), where now q 3 = pf . Then
|T | = q 12 (q 8 + q 4 + 1)(q 6 − 1)(q 2 − 1),
and | Out(T )| = f = 3 ln(q)/ ln(p) 6 3 ln(q)/ ln(2), so (2) implies the following inequality, which fails
for all q > 2:
34 212 4
ln (q).
q 12 (q 8 + q 4 + 1)(q 6 − 1)(q 2 − 1) 6 4
ln (2)
(viii) T ∼
= 2 B2 (q), 2 G2 (q), or 2 F4 (q). Finally, suppose that T is as in one of the lines of Table 1.
Suppose first that n > 1. Then | Out(T )| = 2n + 1 in each case, and (2) therefore implies that |T | 6
212 (2n+1)4 . This inequality holds only in the case T ∼
= 2 B2 (22n+1 ) with n = 1, but |2 B2 (23 )| = 29 120
cannot be written in the form (a) subject to (b). For n = 0, we have that 2 B2 (q) is not simple; 2 G2 (3)
is not simple, but 2 G2 (3)′ ∼
= A1 (8) has been ruled out in Case 2 above; and 2 F4 (2) is not simple, but
2 F (2)′ is simple of order 17 971 200 and has outer automorphism group of order 2, so (2) fails.
4
This completes the proof of Lemma 3.1.
8
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
T
|T |
q
2 B (q)
2
2 G (q)
2
2 F (q)
4
q 2 (q 2 + 1)(q − 1)
22n+1
3
3
q (q + 1)(q − 1)
32n+1
12
6
4
3
q (q + 1)(q − 1)(q + 1)(q − 1) 22n+1
Table 1. Orders of the Suzuki and Ree simple groups.
4. Proof of Theorem 1.1: HC type
Suppose that Q = (P, L, I ) is a thick generalised quadrangle with a collineation group G that
acts transitively on L and primitively of O’Nan–Scott type HC on P. Then
M ⋊ Inn(M ) 6 G 6 M ⋊ Aut(M ),
where M = T1 ×· · ·×Tk , with k > 2 and T1 ∼
= ··· ∼
= Tk ∼
= T for some non-Abelian finite simple group T .
Moreover, M acts regularly on P, and G induces a subgroup of Aut(T ) ≀ Sk which acts transitively on
the set {T1 , . . . , Tk } (see [8, Section 5]). Since |M | = |T |k is even by the Feit–Thompson Theorem [5],
Lemma 2.4 tells us that Q satisfies the hypotheses of Theorem 2.2 and Corollary 2.3. In particular,
s + 1 divides t − 1 (by Corollary 2.3), and we define t′ as in (1).
We first rule out the case k > 3, and then deal with the case k = 2 separately.
4.1. k > 3. Suppose, towards a contradiction, that k > 3. Denote by ℓ1 , . . . , ℓt+1 the lines incident
with the identity 1 ∈ M . By Lemma 2.1(ii), we may identify ℓi with the subgroup of M comprising
all points incident with ℓi . Let us write ℓ := ℓ1 for brevity.
Claim 4.1. M cannot be decomposed in the form M = A × B with ℓ ∩ A 6= {1} and ℓ ∩ B 6= {1}.
Proof. Suppose, towards a contradiction, that M = A × B with ℓ ∩ A 6= {1} and ℓ ∩ B 6= {1}. We
may assume, without loss of generality, that (i) A contains T1 , and (ii) ℓ ∩ A contains an element
x = (x1 , . . . , xk ) that projects non-trivially onto each simple direct factor of A (if not, then change
the decomposition of M to A′ × B ′ with A′ 6 A and B ′ > B). Take also y ∈ ℓ ∩ B with y 6= 1. For
every a ∈ Inn(A) 6 Inn(M ), we have y a = y and hence ℓa = ℓ, because a also fixes the point 1 ∈ ℓ.
In particular, ℓ is fixed by every element of Inn(T1 ), regarded as a subgroup of Inn(A). Therefore,
(z, x2 , . . . , xk ) ∈ ℓ for all z ∈ xT1 1 , and hence ℓ contains the group ℓ0 := h(z, x2 , . . . , xk ) : z ∈ xT1 1 i. Let
π1 denote the projection onto T1 . Then π1 (ℓ0 ) = hz : z ∈ xT1 1 i = T1 , and hence π1 (ℓ) = T1 . Also,
taking z 6= x1 , we see that ℓ ∩ T1 contains (z, x2 , . . . , xk )−1 x = (z −1 x1 , 1, . . . , 1) 6= 1. That is, ℓ ∩ T1 is
non-trivial, and it is normal in the simple group π1 (ℓ) = T1 , so ℓ ∩ T1 = T1 and hence T1 6 ℓ. Now,
G1 acts transitively on both {T1 , . . . , Tk } (because G is transitive on {T1 , . . . , Tk } and G = M G1 ) and
{ℓ1 , . . . , ℓt+1 } (because G is flag-transitive, by Theorem 2.2). Therefore, t + 1 divides k, and, without
loss of generality, ℓ = ℓ1 contains TU1 := T1 ×· · ·×Tk/(t+1) , ℓ2 contains TU2 := Tk/(t+1)+1 ×· · ·×T2k/(t+1) ,
and so on.
Sub-claim. ℓ = TU1 .
Proof of sub-claim. It remains to show that TU1 contains ℓ. Suppose, towards a contradiction,
that there exists w ∈ ℓ \ TU1 . Then there exists i > k/(t + 1) such that the ith component wi of
w is non-trivial, and so there exists σ ∈ Inn(Ti ) such that wiσ 6= wi . Regarding σ as an element of
Inn(M ) 6 G1 , we see that σ fixes ℓ, because it centralises T1 6 ℓ. Hence, wσ ∈ ℓ, and so ℓ contains
w−1 wσ ∈ ℓ ∩ Ti \ {1}. However, Ti 6 TUj 6 ℓj for some j 6= 1, and hence ℓ intersects ℓj in more than
one point, a contradiction, proving the sub-claim.
By the sub-claim, s + 1 = |T |u , where u = k/(t + 1). Since |T |(t+1)u = |M |, we have (s + 1)t+1 =
(s + 1)2 (st′ + 1), where t′ := (t − 1)/(s + 1) 6 s − 1 as before. Since st′ + 1 6 s(s − 1) + 1 < (s + 1)2 ,
this implies that (s + 1)t−1 < (s + 1)2 , so t = 2, and hence s + 1 | t − 1 = 1, a contradiction.
Claim 4.2. ℓ is isomorphic to a subgroup of T .
Proof. Let x ∈ ℓ \ {1} have minimal support U . Suppose, without loss of generality, that x1 :=
π1 (x) 6= 1. Suppose further, towards a contradiction, that there exists y ∈ ℓ \ {1} with π1 (y) = 1.
PRIMITIVE GENERALISED QUADRANGLES OF HOLOMORPH TYPE
9
Then every a ∈ Inn(T1 ) fixes y and hence fixes ℓ, so ℓ contains xa and therefore contains xa x−1 ∈ T1 ∩ℓ.
Taking a not in CT (x1 ) makes xa x−1 non-trivial, and the minimality of the support U of x implies
that U = {1}, so x ∈ T1 . However, the existence of y now contradicts Claim 4.1, because taking
A = T1 and B = T2 × · · · × Tk gives x ∈ ℓ ∩ A and y ∈ ℓ ∩ B. Hence, if x has minimal support U
containing 1, then every non-trivial element of ℓ must project non-trivially onto T1 . Therefore, ℓ is
isomorphic (under projection) to a subgroup of T1 .
We now use Claim 4.2 to derive a contradiction to the assumption that k > 3. By Claim 4.2, s+1 =
|ℓ| divides |T |, so in particular s+1 6 |T |. Writing |M | = (s+1)2 (st′ +1) with t′ := (t−1)/(s+1) 6 s−1
as before, we have (s + 1)2 > s(s − 1) + 1 > st′ + 1 = |M |/(s + 1)2 > |M |/|T |2 = |T |k−2 > (s + 1)k−2 ,
and hence 2 > k − 2, namely k 6 3.
Now suppose, towards a contradiction, that k = 3. Write |T | = n(s + 1). Then st + 1 =
|M |/(s + 1) = |T |3 /(s + 1) = n3 (s + 1)2 , and hence n3 ≡ 1 (mod s). On the other hand, s3 + 1 >
st + 1 = n3 (s + 1)2 > n3 s2 + 1, so n3 < s. Therefore, n = 1, so |T | = s + 1 and t = s + 2. Together
with Claim 4.2, this implies that ℓ is isomorphic to T . Consider first the case where ℓ is a diagonal
subgroup {(t, ta , tb ) : t ∈ T } 6 M for some a, b ∈ Aut(T ). As (c, d) ∈ Inn(T2 )×Inn(T3 ) 6 G1 runs over
all possibilities, we obtain |T |2 distinct images ℓ(c,d) = {(t, tac , tbd ) : t ∈ T } of ℓ. Indeed, if ℓ = ℓ(c,d) ,
then ta = tac for all t ∈ T , or equivalently, u = uc for all u ∈ T ; that is, c is the identity automorphism
of T (and similarly, d is the identity). Hence, s + 3 = t + 1 ≥ (s + 1)2 , a contradiction. Now consider
the case where ℓ is a diagonal subgroup {(t, ta , 1) : t ∈ T } 6 T1 × T2 for some a ∈ Aut(T ). Then 3
divides t + 1 because G1 is transtive on the Ti , and we have exactly (t + 1)/3 lines incident with 1
that are diagonal subgroups of T1 × T2 . As c ∈ Inn(T2 ) 6 G1 runs over all possibilities, we obtain |T |
distinct images ℓc = {(t, tac , 1) : t ∈ T } of ℓ. Hence, (s + 3)/3 = (t + 1)/3 > s + 1, a contradiction.
This leaves only the possibility that ℓ 6 T1 , and hence ℓ = T1 because |ℓ| = s + 1 = |T1 |. This implies
that t + 1 = 3, and hence s = 0 because s + 1 divides t − 1, a contradiction.
4.2. k = 2. Here we argue as in the case where the primitive action of G on P has type HS. That
is, we obtain an upper bound on |T | in terms of | Out(T )|, and consider the possibilities for T case by
case using the Classification of Finite Simple Groups. We have M = T1 × T2 ∼
= T 2 , and
|M | = (s + 1)(st + 1) = (s + 1)2 (st′ + 1),
where 1 6 t′ 6 s − 1.
Therefore,
|T | = (s + 1)(st′ + 1)1/2 , where 1 6 t′ 6 s − 1 and st′ + 1 is a square.
Writing y 2 = st′ + 1, this is equivalent to
|T | = (s + 1)y, where 3 6 y 2 6 s(s − 1) + 1 and s | y 2 − 1.
By Theorem 2.2, G1 6 Aut(M ) ∼
= Aut(T ) ≀ S2 acts transitively on the lines incident with 1, and hence
t + 1 divides |Aut(M )| = 2|T |2 | Out(T )|2 . Therefore, | Out(T )|2 is divisible by
t+1
t+1
=
.
2
gcd(t + 1, 2|T | )
gcd(t + 1, 2(s + 1)2 (st′ + 1))
In particular, t+1 6 gcd(t+1, 2|T |2 )| Out(T )|2 . We have (i) gcd(t+1, s+1) = 2, so gcd(t+1, 2(s+1)2 ) 6
8; and (ii) gcd(t + 1, st′ + 1) = gcd(t + 1, t′ + 1). Hence, gcd(t + 1, 2|T |2 ) 6 8(t′ + 1), and so
t + 1 6 8(t′ + 1)| Out(T )|2 . Re-writing this as t′ (s + 1) + 2 6 8(t′ + 1)| Out(T )|2 , and noting that t′ > 1,
we obtain
s 6 16| Out(T )|2 − 3.
Higman’s inequality then gives
|T |2 = |M | 6 (16| Out(T )|2 − 2)((16| Out(T )|2 − 3)3 + 1).
The following lemma therefore rules out all but two possibilities for T .
Lemma 4.3. Let T be a finite non-Abelian simple group satisfying
(a) |T | = (s + 1)y, where 3 6 y 2 6 s(s − 1) + 1 and s | y 2 − 1;
(b) 2 6 s 6 16| Out(T )|2 − 3; and
(c) |T |2 6 (16| Out(T )|2 − 2)((16| Out(T )|2 − 3)3 + 1).
10
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
Then either (i) T ∼
= Alt6 , s = 19, and y = 18; or (ii) T ∼
= A2 (2), s = 13, and y = 12.
Proof. The right-hand side of (c) is at most (16| Out(T )|2 )4 , so
(4)
|T | 6 28 | Out(T )|4 .
Since (4) implies (2), any group T that was ruled out using (2) in the HS case (that is, in the proof
of Lemma 3.1) is automatically ruled out here. To rule out the remaining possibilities for T , we use
either (4) or (c), or check that (a) has no solution subject to (b). Note that (a) implies y 6 s < y 2 .
Case 1. T ∼
= Altn or a sporadic simple group:
If T is an alternating group other than Alt6 , or a sporadic simple group, then | Out(T )| 6 2 and
so (c) implies that |T | < 3 752. Hence, T is one of Alt5 , Alt6 , or Alt7 . If T ∼
= Alt5 , then by (a),
we have (s + 1)y = 60 and s | y 2 − 1, which is impossible. If T ∼
= Alt7 , then we again apply (a):
(s + 1)y = 2520, s | y 2 − 1, and y 2 6 s(s − 1) + 1, which is again impossible. Finally, we examine the
case T ∼
= Alt6 , where | Out(T )| = 4. Applying (a), we have s = 19, y = 18 as the only valid solution.
Case 2. T ∼
= A1 (q):
Suppose that T ∼
= A1 (q), and write q = pf with p prime and f > 1. Then |T | = q(q 2 − 1)/(2, q − 1),
and | Out(T )| = (2, q − 1)f .
Suppose first that q is even, namely that p = 2. Then gcd(2, q − 1) = 1, and (c) implies that
22f (22f − 1)2 6 (16f 2 − 2)((16f 2 − 3)3 + 1),
which holds only if f 6 7. If f = 1, then T is not simple; and if f = 2, then T ∼
= Alt5 , which we have
already ruled out. For 3 6 f 6 7, there is no solution to (a) subject to (b).
Now suppose that q = pf is odd. Then gcd(2, q − 1) = 2, and hence | Out(T )| = 2f . By (c), we
have
p2f (p2f − 1)2 6 8(32f 2 − 1)((64f 2 − 3)3 + 1),
which implies that either 11 6 p 6 19 and f = 1; 5 6 p 6 7 and f 6 2; or p = 3 and f 6 4. If q = 3,
then T is not simple; if q = 5, then T ∼
= A2 (2),
= Alt5 , which we have ruled out; if q = 7, then T ∼
which is ruled out in Case 3 below; and if q = 9, then T ∼
= Alt6 , which we have already dealt with in
Case 1. Hence, we only need to consider q ∈ {11, 13, 17, 19, 33 , 34 , 52 , 72 }. For each of these values,
there is no solution to (a) subject to (b).
Case 3. T ∼
= An (q), n > 2:
Since (4) implies (2), by comparing with the proof of Case 2 in Lemma 3.1, we see that we only
need to check T ∼
= A3 (3), and T ∼
= A2 (q) for q 6 13. The former is ruled out by (4), because
|A3 (3)| = 6 065 280 > 28 44 = 65 536. For T ∼
= A2 (q), (c) implies that
!
3
576
576
q 6 (q 2 − 1)2 (q 3 − 1)2 6 9
ln2 (q) − 2
ln2 (q) − 3 + 1
ln2 (2)
ln2 (2)
Therefore, q 6 10. For q = 2, there is a unique solution to (a) subject to (b), namely s = 13, t′ = 11.
For q ∈ {3, 4, 5, 7, 8, 9}, there are no solutions to (a) subject to (b).
Case 4. T ∼
= 2 An (q 2 ):
Since (4) implies (2), we only need to check T ∼
= 2 A3 (q 2 ) for 2 6 q 6 4, and T ∼
= 2 A2 (q 2 ) for q 6 13.
If (n, q) = (3, 3) or (3, 4), then (4) fails; and for (n, q) = (3, 2), there are no solutions to (a) subject
to (b). For n = 2, (c) gives
!
3
576
576
q 6 (q 2 − 1)2 (q 3 + 1)2 6 9
ln2 (q) − 2
ln2 (q) − 3 + 1 ,
ln2 (2)
ln2 (2)
and hence q 6 10. If q = 2, then T ∼
= 2 A2 (q 2 ) is not simple. If q ∈ {3, 4, 5, 7, 8, 9}, then there are no
solutions to (a) subject to (b).
Case 5. Remaining possibilities for T :
We only need to check the groups from Case 5 of the proof of Lemma 3.1 that were not ruled out by
(2) or by exceptional isomorphisms to groups that have already been handled. There are only two
such cases. If T ∼
= B2 (2f ) with f = 2, then, using (4) instead of (2), the 216 on the right-hand side
of (3) becomes 212 , and the resulting inequality 24f (22f − 1)(24f − 1) 6 212 f 4 fails when f = 2. If
T ∼
= 2 B2 (22n+1 ) with n = 1, then (4) fails (although (2) does not).
PRIMITIVE GENERALISED QUADRANGLES OF HOLOMORPH TYPE
This completes the proof of Lemma 4.3.
11
y2
st′
It remains to rule out cases (i) and (ii) from Lemma 4.3. Using
=
+ 1, we find that t = 341
in case (i), and t = 155 in case (ii). Both cases are then ruled out because the required divisibility
condition t + 1 | | Aut(M )| = 2|T |2 | Out(T )|2 fails. (Note that | Aut(M )| = 4 147 200 if T ∼
= Alt6 , and
| Aut(M )| = 225 792 if T ∼
= A2 (2).)
Acknowledgements
We thank the referee for a careful reading of the paper, and in particular for bringing to our
attention an error in a previous version of the proof of Corollary 2.3.
References
[1] J. Bamberg, Innately transitive groups, Ph.D. thesis, The University of Western Australia, 2003.
[2] J. Bamberg, M. Giudici, J. Morris, G. F. Royle, and P. Spiga, “Generalised quadrangles with a group of automorphisms acting primitively on points and lines”, J. Combin. Theory Ser. A 119 (2012) 1479–1499.
[3] J. Bamberg, S. P. Glasby, T. Popiel, and C. E. Praeger, “Generalised quadrangles and transitive pseudo-hyperovals”,
J. Combin. Des. 24 (2016) 151–164.
[4] S. De Winter and K. Thas, “Generalized quadrangles with an abelian Singer group”, Des. Codes Cryptogr. 39 (2006)
81–87.
[5] W. Feit and J. G. Thompson, “Solvability of groups of odd order”, Pacific J. Math. 13 (1963) 775–1029.
[6] D. Ghinelli, “Regular groups on generalized quadrangles and nonabelian difference sets with multiplier −1”, Geom.
Dedicata 41 (1992) 165–174.
[7] S. E. Payne and J. A. Thas, Finite generalized quadrangles, Second edition, EMS Series of Lectures in Mathematics,
European Mathematical Society (EMS), Zürich, 2009.
[8] C. E. Praeger, “Finite quasiprimitive graphs”, in: Surveys in combinatorics, 1997 (London), volume 241 of London
Math. Soc. Lecture Note Ser., Cambridge University Press, Cambridge, 1997, pp. 65–85.
[9] J. Tits, “Sur la trialité et certains groupes qui s’en déduisent”, Inst. Hautes Études Sci. Publ. Math. 2 (1959) 13–60.
[10] S. Yoshiara, “A generalized quadrangle with an automorphism group acting regularly on the points”, European J.
Combin. 28 (2007) 653–664.
Centre for the Mathematics of Symmetry and Computation, School of Mathematics and Statistics,
The University of Western Australia, 35 Stirling Highway, Crawley, W.A. 6009, Australia.
Email: {john.bamberg, tomasz.popiel, cheryl.praeger† }@uwa.edu.au,
†
Also affiliated with King Abdulaziz University, Jeddah, Saudi Arabia.
| 4math.GR
|
Trace Typing: An Approach for Evaluating
Retrofitted Type Systems (Extended Version)
Esben Andreasen1 , Colin S. Gordon2 , Satish Chandra3 , Manu
Sridharan3 , Frank Tip3 , and Koushik Sen4
1
2
3
arXiv:1605.01362v1 [cs.PL] 4 May 2016
4
Aarhus University
[email protected]
Drexel University
[email protected]
Samsung Research America
{schandra,m.sridharan,ftip}@samsung.com
UC Berkeley
[email protected]
Abstract
Recent years have seen growing interest in the retrofitting of type systems onto dynamicallytyped programming languages, in order to improve type safety, programmer productivity, or
performance. In such cases, type system developers must strike a delicate balance between
disallowing certain coding patterns to keep the type system simple, or including them at the
expense of additional complexity and effort. Thus far, the process for designing retrofitted type
systems has been largely ad hoc, because evaluating multiple variations of a type system on large
bodies of existing code is a significant undertaking.
We present trace typing: a framework for automatically and quantitatively evaluating variations of a retrofitted type system on large code bases. The trace typing approach involves
gathering traces of program executions, inferring types for instances of variables and expressions occurring in a trace, and merging types according to merge strategies that reflect specific
(combinations of) choices in the source-level type system design space.
We evaluated trace typing through several experiments. We compared several variations of
type systems retrofitted onto JavaScript, measuring the number of program locations with type
errors in each case on a suite of over fifty thousand lines of JavaScript code. We also used trace
typing to validate and guide the design of a new retrofitted type system that enforces fixed
object layout for JavaScript objects. Finally, we leveraged the types computed by trace typing
to automatically identify tag tests — dynamic checks that refine a type — and examined the
variety of tests identified.
1998 ACM Subject Classification F.3.3 Studies of Program Constructs
Keywords and phrases Retrofitted type systems, Type system design, trace typing
1
Introduction
In recent years, there have been a number of efforts to retrofit [20] type systems onto
dynamically-typed languages, to aid developer productivity, correctness, and performance.
These languages are of increasing importance, primarily due to their common use in web
applications on both the client and server side. As more large-scale, complex programs are
written in such languages, greater need arises for static types, due to the resulting benefits
for static error checking, developer tools, and performance. Recent high-profile projects in
type system retrofitting include Closure [15], TypeScript [4] and Flow for JavaScript [12],
© Esben Andreasen, Colin S. Gordon, Satish Chandra, Manu Sridharan, Frank Tip, Koushik Sen;
licensed under Creative Commons License CC-BY
Samsung Research America Technical Report SRA-CSIC-2016-001.
This is an extended version of a paper published in ECOOP 2016.
232
Trace Typing: An Approach for Evaluating Retrofitted Type Systems (Extended Version)
and Hack for PHP [31]. Given the proliferation of dynamically-typed languages, there are
many retrofitted type systems ahead.
The key questions in the design of these type systems involve selecting which features to
include. Richer features may allow more coding patterns to be understood and validated
by the type checker. However, richer type systems come at the cost of greater complexity
and implementation effort, and it is rarely a priori obvious whether a given feature’s benefit
outweighs its implementation burden. Hence, great care must be taken in deciding if a
particular type system feature is worthy of inclusion. Such decision points continue to arise
as a type system evolves: e.g., TypeScript recently added union types [28], and Flow recently
added bounded polymorphism [7].
When retrofitting a type system, there is generally a wealth of existing code to which
one would like to introduce types. When considering a type system feature, it would
be very helpful to know how useful the feature would be for typing an existing body of
code. Unfortunately, truly gauging the feature’s usefulness would require implementing type
checking (and possibly inference) for the feature and adding the necessary annotations to the
code base. This involves a significant effort, which ultimately may not have lasting value if an
alternate type system proves to be more suitable. Some previous work [3, 11, 20, 22] reduces
the first burden by providing a reusable framework for implementing type systems, with some
inference. However, each still requires manual annotation, so large-scale evaluation remains
a substantial undertaking — in one case an expert required over 17 hours to annotate part
of one program with type qualifiers [16]. For this reason, in current practice, decisions are
often made based on intuition and anecdotal evidence, without the aid of carefully collected
quantitative results on the impact of certain features. Anecdotes and other qualitative criteria
may validate inclusion of features that are rarely used but important, but quantitative results
on a large corpus can rapidly guide the high level design.
In this work, we propose a novel framework, called trace typing, for automatic and
quantitative evaluation of retrofitted type system features for existing code bases. Our system
works by collecting detailed, unrolled traces of dynamic executions of the code in question.
Evaluating a type system feature then requires implementing type checking and inference for
the unrolled trace, and also defining a merge strategy (explained shortly) for combining the
types of different occurrences of a variable in the trace into the corresponding static program
entity. The crucial observation here is that type checking and effective inference for the
unrolled trace is far simpler than static type checking / inference for the full source language,
and requires no type annotations due to the simplicity of the trace format. Hence, our system
dramatically reduces the implementation effort required to perform quantitative comparisons
of different type system variants, and allows type system designers to automatically gather
feedback from large code bases by typing the trace.
While type inference and checking for unrolled traces is simpler to implement than a
static type checker, the ultimate goal of the designer is a type system for source programs,
not traces. Our key insight is that–given types for the unrolled trace–many source type
systems can be expressed via a family of merge operations, used to merge the types of runtime
entities that the source type system cannot distinguish. For example, a merge operator can
control whether assigning an integer and an object into the same variable yields a union
type for the variable, or a type error. (We give examples of merging throughout the paper.)
The type errors found by using merged types give an indicative lower bound on the set
of program locations that would be ill-typed under a static type checker, as long as the
merge operator produces useful types. A designer can, therefore, evaluate the usefulness of a
proposed static type system via suitable merge operators in the trace typing framework, and
E. Andreasen et al.
can also compare the relative power of different static type systems.
We have implemented our techniques in a system for evaluating type system variants for
JavaScript. We evaluated our system on a diverse suite of programs, and evaluated several
different variations on a core object-oriented type system. In particular, our experiments
yielded quantitative answers to the following questions for our benchmarks:
How pervasive is the need for supporting union types on top of a type system that
supports subtyping?
If union types are supported, what kind of “tag tests” usually occur to discriminate
between branches of a union?
How pervasive is the need for supporting parametric polymorphism?
How pervasive is the need for intersection types for functions?
The results are given in Section 6. We believe trace typing gives a quick and easy way to
obtain answers to these (and similar) questions on a given code corpus.
We have also used trace typing to guide the design of a type system to support aheadof-time compilation of JavaScript extending previous work [8]. Trace typing allowed us
to validate some of our design choices without implementing a type inference engine that
works on real code. Moreover, it helped point to where we need to extend the type system.
Section 6.4 details this case study.
Trace typing is designed to present a rapid, aggregate, quantitative picture to help a type
system designer decide whether to support certain features and prioritize the development
of a type checker. When features are not supported, a programmer needing said features
must either ignore the warnings from the type system, or refactor her code to fit the type
system restrictions. For example, if trace typing shows that parametric polymorphism is
rarely needed in the code of interest, then the designer can defer supporting polymorphism;
in the interim, developers must work around the limitation, e.g., by code cloning. Of course,
a type system designer will need to consider other factors besides the output of trace typing
when prioritizing, such as idioms used by critical modules, or the difficulty of converting
code to fit within certain restrictions.
Contributions This paper makes the following contributions:
We propose trace typing as an approach for easily evaluating the usefulness of type system
features, allowing large bodies of existing code to inform type system design. Trace typing
requires far less effort than implementing various static type checkers, and potentially
adding type annotations to large bodies of code.
Using the trace typing approach, we systematically compare the precision of a number of
object type systems for JavaScript on some of the most popular packages for node.js,
totaling over 50,000 lines of code. We include union types as well as several variants of
polymorphism in our study.
We describe our experience with trace typing to guide and check choices in the design of
a type system for use by an optimizing JavaScript compiler.
We use trace typing to automatically identify tag tests in our large corpus, and analyze
the relative frequency of different tests and their structure.
2
Trace Typing by Examples
In this section, we show informally how trace typing can be used to carry out different kinds
of quantitative experiments relevant when designing a type system. Section 3 gives a more
formal description of the framework.
233
234
Trace Typing: An Approach for Evaluating Retrofitted Type Systems (Extended Version)
function f ( a ) { a . p = 7; }
function g ( b ) { return b ; }
var x = { p : 3 };
var y = { p : 4 , q : " hi " };
var z = { q : " bye " , r : false };
f ( x ) ; // f .6: { p : Number } -> void
f ( y ) ; // f .7: { p : Number , q : String } -> void
g ( y ) ; // g .8: { p : Number , q : String } -> { p : Number , q : String }
var w = g ( z ) ; // g .9: { q : String , r : Boolean } -> { q : String , r : Boolean }
w . r = true ;
1
2
3
4
5
6
7
8
9
10
Figure 1 A JavaScript program to illustrate polymorphism.
2.1
Polymorphism
Figure 1 gives a small JavaScript example. The program allocates new objects at lines 3–5,
and then calls functions f and g. The f function accesses the p field, while g just returns its
parameters. All field accesses are well-behaved: they access pre-existing fields, and fields are
only updated with a value of a compatible type. Here, we show how trace typing can can
quantify the relative effectiveness of two type systems with different levels of polymorphism
for handling this code.
We focus on the types observed in the dynamic trace for functions f and g. For our
purpose, these types correspond to the parameter and return types observed per function
call site. So, for function f we observe the type {p: Number} -> void for the call at line 6,
and {p: Number, q: String} -> void for line 7. Similarly, for function g we observe
type {p: Number, q: String} -> {p: Number, q: String} for the line 8 call, and {q:
String, r: Boolean} -> {q: String, r: Boolean} for line 9. Note that the types above
are not based on inference on the body of functions f and g.
Type system variants can be distinguished by how their merge strategies combine these
types into a single function type each for f and for g. We consider the following merge
strategies. Tsub merges, separately the argument and return types, in each case taking the
least upper bound in a suitable subtyping lattice. For our example, this strategy yields the
type {p: Number} -> void for f and {q: String} -> {q: String} for g.1 Tpoly merges
the function types by introducing type variables, arriving at a signature with parametric
polymorphism (details in Section 4.3). The type we obtain for f is the same as in the case of
Tsub , but for g, it is (E) -> E; E is a type parameter.
Next, the trace typing system performs type checking using the merged types, and reports
the number of type errors to the user. This type error count identifies program locations
that a static checker for the type system would be unable to validate, a valuable statistic for
determining the usefulness of the variant in practice (fewer errors2 means that the variant
can validate more code patterns).
With Tsub , the types are sufficient for type-checking the bodies of the functions, but g’s
type is still insufficient to typecheck the dereference at line 10, leaving one error remaining.
With Tpoly , the types are sufficient to ensure there is no type error in the program. The type
checking phase also checks conformance of the actual parameters being passed to function
calls at lines 6, 7, 8 and 9; these checks pass for both Tsub and Tpoly .
Tsub produces one type error on this trace, while Tpoly produces 0. These counts are
1
2
Note that we did not compute a type for g via least-upper-bound with respect to function subtyping
(with contravariance in argument position); see Section 3.1 for further details.
Throughout this paper, we use error as a synonym for static type error, not to refer to actual developer
mistakes in a program.
E. Andreasen et al.
1
2
3
4
5
6
7
8
9
function f ( x ) {
... x ... // x .2: A
if ( x instanceof A ) {
... x ... // x .4:
}
... x ... // x .6: A
}
f ( new A () ) ;
f (3) ;
235
| int
// x .3: A | int
A
| int
Figure 2 Tag test example.
underestimates with respect to a (hypothetical) static type checker, both due to our reliance
on dynamic information, and due to the over-approximations inherent in any static type
checking. Nevertheless, across a large code base, such data could be invaluable in finding the
extent to which a particular type system feature, e.g. parametric polymorphism, is useful.
Section 6.2 presents results from an experiment that evaluated the usefulness of parametric
polymorphism and a synthetic type system with “unbounded” polymorphism on a suite of
benchmarks, via comparison of error counts.
2.2
Discriminating Unions
Suppose we wish to design a type system for JavaScript that is equipped to reason about tag
checks at the points where the program refines a union type. Such a construct is shown in the
example of Figure 2. In this example, a union type with more than one case is narrowed to a
specific case guarded by a condition. In general, the guard can be an arbitrary expression,
but that makes type checking difficult — it would need to track an arbitrary number of
predicates in carrying out an analysis. For example, the following “non-local” variation of
the type guard in the example in Figure 2 would be more complex to handle:
1
2
3
var y = x instanceof A
...
if ( y ) { ...
It is simpler to support a limited number of syntactic forms.
Can a type system designer obtain quantitative feedback on what kind of conditional
constructs occur in the code corpora of interest? Are non-local type guards common? What
about conjunctions and disjunctions of predicates? Simple syntactic pattern matching is an
unreliable way to find this out, because it is fundamentally a flow analysis problem.
Trace typing makes this experiment easy. Trace typing ascribes a type to each occurrence
of each variable in the program based on dynamic observation. Given these types, one can
find pairs of successive reads of a variable without an intervening write, where the type
ascribed to the variable in the second read is different from that of the first. This is a
heuristic to identify instances of a variable’s type becoming narrower on an unrolled path,
likely to be a tag test.
Consider the example in Figure 2 again. We examine some of the type ascriptions carried
out by trace typing during an execution of the program. Here, at line 2, x refers to an object
of type A and an int, respectively, in the invocations from line 8 and line 9. Let us label the
two dynamic occurrences of x as x.2.1 and x.2.2, where the first index refers to the line
on which the variable occurs and the second refers to the invocation sequence number. The
same types will be observed for x.3.1 and x.3.2. At line 4, x.4.1 observes A; that is the
only dynamic occurrence of x on this line. Finally at line 6, x.6.1 and x.6.2 observe A and
int respectively.
236
Trace Typing: An Approach for Evaluating Retrofitted Type Systems (Extended Version)
Candidate
type
system
representative
input(s)
Program
Trace
Ascribe
types
to
values
Assign
dynamic
types
to
variables
Assign
static
types
to
variables
Check
for
errors
Type
errors
Figure 3 Trace typing framework overview.
Crucially, trace typing can be configured to merge the types ascribed to variable occurrences on the same line, but across different invocations. Say that our type system
includes union types. In that case, x.2 (the result of merging x.2.1 and x.2.2) gets the
type A | int, which represents a union type consisting of types A and int. For x.3 we get
A | int as well. However, for x.4, we get just A. Hence, for successive reads x.3 and x.4
(with no intervening write), we observe the situation that the second type is a refinement of
the first. This is a clue that the conditional at line 3 (that falls between the first and the
second reads) is a type guard.
Using this technique on a large code corpus, we found that while non-local type guards
are uncommon, boolean combinations of predicates are commonly used in type guards.
Section 6.3 details the experiment we carried out to find the nature of tag tests in real code.
3
Trace typing
Figure 3 gives an overview of the trace typing process. The trace typing framework takes
two inputs:
a type system of interest, specified by defining various parameters in a framework
(Section 3.1); and
a program P of interest, and inputs for exercising P.
Given these inputs, trace typing proceeds in the following four phases:
1. An unrolled trace is produced by executing an instrumented version of P on the provided
inputs (Section 3.2).
2. The types of all values (including objects and functions) and variables in the trace are
inferred (Section 3.3).3
3. The inferred types are merged, coarsening the precise types in the trace to a corresponding
source typing (Section 3.4).
4. The resulting types are used for type checking, to estimate the ability of the type system
to handle the observed behaviors (Section 3.5).
The number of errors reported in the last step is a useful metric for evaluating type
system features (see Section 7 for further discussion). Our framework has proven to be
sufficiently expressive to model even modern industrial proposals for retrofitted type systems
(Section 4.4). Also, the computed types can be used as part of experiments to test richer
type system features, such as tag tests (Section 4.2).
3.1
Trace Type Systems
Here, we describe the interface by which a type system is specified to the trace typing
framework. A type system T has five components:
3
In the remainder of the paper, we use the terms “ascribed” and “inferred” interchangeably.
E. Andreasen et al.
237
T.α: (Value) → Type ascribes a type to a concrete value. For JavaScript, values include
both primitives as well as object instances.
T.t: (Type, Type) → Type computes the least-upper-bound of the input types in the
subtyping lattice for the type system. For termination, we require T.t to be monotonic.
→
T. t: (Function-type. . . ) → Function-type takes a list of types for individual invocations
of a function f and computes a type that can accommodate all invocations. A type for
an invocation is computed by applying T.α to the receiver, argument, and return values,
→
and then combining them into a function type. T. t often differs from applying T.t to the
invocation types; further discussion below.
T.equiv: (FI | FS) × (CI | CS) governs how types for different trace occurrences of the
same source variable are merged to approximate source-level typing (see Section 3.4). The
first component determines whether to maintain a separate type for a local variable at
different program points (flow sensitivity), while the second component determines if separate
types should be maintained across different function invocations (context sensitivity). Note
→
that use of CI vs. CS is usually dictated by the choice of T. t (see Section 6.2). Also, our
implementation supports more fine-grained flow- and context-sensitivity settings; here we
only present two options for simplicity.
T.check: (TypeEnv, Statement) → number is a type checking function that counts the
number of type errors a trace statement causes in a given type environment.
→
Together, T.t, T. t, and T.equiv comprise the type system’s merge strategy. Value
represents the values of our trace language (see Section 3.2), while Type and Function-type
are types defined entirely by T. Defining T.α for JavaScript objects is non-trivial, since the
shape and property types of the object may change over time; we detail the issues involved
in Section 4.1.
→
Note that T. t should not use contravariance when handling function parameters, unlike
→
typical function subtyping. T. t is used to compute a type τ for function f that is consistent
with all observed invocations of f. Hence, each parameter type in τ should accommodate
any of the corresponding parameter types from the individual calls, a covariant handling.
For an example, see the handling of calls to g in Figure 1 (Section 2.1).
Example 1: Simple OO To simulate a basic structural OO type system with width
subtyping:
T.α gives a type to an object by recursively characterizing the types of its properties (see
Section 4.1 for details).
T.t gives the least-upper-bound of two object types (retaining common fields that have
identical types) and returns the same type if given two identical types. Otherwise it
returns >.
→
T. t applies T.t to the individual argument and return types (not using standard function
subtyping; see above) to compute a generalized function type.
T.equiv is (FI,CI).
T.check performs standard assignment compatibility checks for variable and field writes,
property type lookup on property reads, and standard checks at function invocations.
238
Trace Typing: An Approach for Evaluating Retrofitted Type Systems (Extended Version)
htracei
::= hstatementi*
hstatementi ::= hvari = hexpressioni | hvari.hnamei = hvari | delete hvari.hnamei | hmetai
hexpressioni ::= hvari | hvari.hnamei | allocate | null | undefined | hbooli | hnumberi | hstringi
hvari
::= v1 | v2 | v3, . . .
hnamei ::= a | b | . . .
hmetai
::= begin-callhvarihvarihvari* | end-callhvari | end-initializationhvari
Figure 4 TL grammar
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
< tmp0@G > = a l l o c a t e // ... allocation and assignment at line 6
< tmp1@G > = 3
< tmp0@G >. p = < tmp1@G >
end - in itializa tion < tmp0@1 >
< x@G > = < tmp0@G >
begin - call < f@G > < x@G > // ... first call to f , at line 9
< a@f_1 > = < x@G >
< tmp 0 @ f _ 1 > = 7
< a@f_1 >. p = < t m p 0 @ f _ 1 >
end - call
begin - call < f@G > < y@G > // ... second call to f , at line 10
< a@f_2 > = < y@G >
< tmp 0 @ f _ 2 > = 7
< a@f_2 >. p = < t m p 0 @ f _ 2 >
end - call
Figure 5 Excerpts from the trace for the program in Figure 1. The name of variables include
their scope, e.g. <x@G> is the x variable in the global scope, and <a@f_2> is the a variable during
the second call to f.
Example 2: Union Types To extend Example 1 with union types, we modify T.t to
introduce unions when merging different sorts of types (e.g., a number and object type). For
the code ‘var x = 3; x = { f : 4 };’, x would be given the union type number | {f :
number}: Section 4.1 discusses how we perform type checking in the presence of such union
types.
Example 3: Tsub and Tpoly revisited
We referred to Tsub and Tpoly in Section 2.1. Tsub is
→
precisely the Simple OO of Example 1. Tpoly replaces T. t of Tsub to introduce parametric
polymorphism (described further in Section 4.3).
3.2
Traces
The execution of a source code program can be recorded as a finite sequence of primitive
instructions that reproduces the dataflow of the execution. We call such a sequence a trace.
A trace is a simplified version of the original source program:
Every dynamic read or write of a variable or subexpression is given a unique identifier.
Control flow structures are erased: conditionals are reduced to the instructions of the
conditional expression and chosen branch, loops are unrolled, and calls are inlined.
Native operations are recorded as their results; e.g. a call to the runtime’s cosine function
is recorded as the computed value.
We express our JavaScript traces as programs in a simple language TL. TL is an
imperative, object-oriented language without control flow structures or nested expressions.
TL has reads and writes for variables and fields, deletion for fields, and values (objects and
primitive values). The syntax of TL can be seen in Figure 4.
Figure 5 shows excerpts from the trace for executing the code in Figure 1. Many JavaScript
details are omitted here for clearer exposition but handled by our implementation. The
variable names in the trace are generated from source variable names where present, and
E. Andreasen et al.
Γ̂(x) = Γ̂m (x) T. t Γ̂p (x)
239
Γ̂m (x) = T.
G
Γ̂(xi )
xi ∈x̂
rhs = allocate ∨ rhs ∈ primitives
Γ0 (x),
Γ̂(y),
G
rhs = y
Γ̂p (x) =
T.
Γ̂(b).p, rhs = b.p ∧ Γ̂(b) is object with property p
0 x=rhs0 ∈stmts
Γ0 (b).p, rhs = b.p ∧ (Γ̂(b) not an object ∨ Γ̂(b).p not present)
Figure 6 Equations for type merging and propagation.
otherwise fresh temporaries are generated. (Note that fresh temporaries are used within each
call to f.) A source mapping for each trace statement is maintained separately.
The trace contains some meta-statements to aid in later analyses. Each call is demarcated
by begin-call and end-call statements, to aid in recovering the relevant types for the
invocation. Data flow from parameter passing and return values is represented explicitly.
The end-initialization statement marks the end of a constructor or initialization of an
object literal.
3.3
Initial Type Ascription
Given a trace and a type system as specified in Section 3.1, trace typing first ascribes precise
types to variables and values in the trace, without consideration for mapping types back
to the source code. First, types are ascribed to all values. Primitive values and objects
are handled directly using T.α, while functions are handled by combining the type for each
→
invocation with T. t, as described in Section 3.1.
Once types for values have been ascribed, each trace variable is given the type of the
value it holds (recall that each trace variable is only written once). For a given trace T ,
we define the initial type environment Γ0 as: ∀v ∈ Variables(T ). Γ0 (v) = T.α(ValueOf(T, v)).
This typing for trace variables is very precise—every variable access is given its own type,
based solely on the value it holds at runtime. The next phase (Section 3.4) generalizes these
initial types to more closely mimic the desired source-level treatment of variables.
Example Consider the variable <tmp1@G> at line 2 in Figure 5; it gets the value 3, for which
Tsub .α yields the type Number. Similarly the type of <x@G> at line 5 is {p: Number} because
that is the type ascribed to the value in <tmp0@G> at line 5.
3.4
Type Merging and Propagation
To model realistic type systems, we construct a less precise type environment Γ̂ by merging
the types of “equivalent” trace variables in Γ0 . Variable equivalence classes are determined by
the T.equiv operator provided by the type system (see Section 3.1). If the first component
of T.equiv is FI, we employ a flow-insensitive treatment of variables: all variable occurrences
v_i within a single dynamic call corresponding to the same source variable v are made
equivalent. No such equivalences are introduced if the first T.equiv component is FS. If the
second component of T.equiv is CI, we impose a context-insensitive treatment of variables:
each matching v_i across function invocations (i.e., the same read or write of source variable
v) is placed in the same equivalence class.
Given variable equivalence classes and Γ0 , Figure 6 defines Γ̂ in terms of two components:
Γ̂m for merging equivalent variables, and Γ̂p for propagating across assignments. (As Γ̂m
2310
Trace Typing: An Approach for Evaluating Retrofitted Type Systems (Extended Version)
1:
2:
6:
7:
8:
12:
f
g
x
y
z
w
::
::
::
::
::
::
<({ p : Number })
<({ q : String })
{ p : Number }
{ p : Number , q :
{ q : String , r :
{ q : String }
-> Undefined >
-> { q : String }) >
String }
Boolean }
Figure 7 Types ascribed for variable writes in Figure 1 with Tsub .
and Γ̂p are themselves defined in terms of Γ̂, the equations must be solved by computing
a fixed point.) Given trace variable x, Γ̂m (x) computes the least-upper bound (T.t) of all
variables in its equivalence class x̂. However, this merging alone is insufficient for mimicking
source typing, as it does not consider relationships between variables: if the program contains
statement x = y, then the type of y influences the type of x in a source level type system.
Γ̂p in Figure 6 handles type propagation across assignments. The first case handles
assignments of values, using the baseline environment Γ0 . For the second case, x = y, we
generalize Γ̂(x) to include Γ̂(y). This step is important to mimic source-level handling of
the assignment. The final two cases handle object field reads b.p. If Γ̂(b) is an object type
with a property p, the handling is straightforward. However, this may not hold, due to other
approximations introduced in computing Γ̂(b) (e.g., if p were dropped when merging object
types due to width subtyping). If Γ̂(b).p does not exist, we fall back to Γ0 (b).p, i.e., we
use the precise type of b. Without this treatment, type errors could propagate throughout
the remainder of the trace, misleadingly inflating the type error count. Our fallback to the
precise type of b helps to localize type errors, thereby allowing for gathering more useful
information from the rest of the trace than the alternative of producing >. For example, if x
holds a number and later an object with property foo, FI merging will give it type >. A
read of x.foo would produce a type error. Rather than tainting all subsequent uses of the
read result, we ascribe the type of the value observed as the result of that read.
Example Consider type propagation for the a variable in function f of Figure 1, using
type system Tsub from Section 2. The corresponding variables in Figure 5 are <a@f_1>
and <a@f_2>, whose respective types are {p: Number} and {p: Number, q: String} in Γ0
(<y@G>’s initialization (line 4 in Figure 1) is elided in Figure 5). Hence, with the widthsubtyping-based merge mechanism Tsub .t, both <a@f_1> and <a@f_2> are assigned type
{p: Number} in Γ̂. The types in Γ̂ for all variables in Figure 1 (using Tsub ) appear in Figure 7.
3.5
Type Checking
Once type propagation is done, type checking can be performed applying T.check to every
statement in the trace, using Γ̂ as the context. T.check should return, for each statement,
the number of falsified antecedents for the corresponding type rule (returning 0 implies a
well-typed statement).
Note that our type propagation biases the location of type errors to occur more often at
reads instead of writes. Type propagation uses the T.t operator to find a variable type that
handles all writes to that variable. This process may produce a type higher in the subtyping
lattice than what the programmer may have intended, causing type errors to be pushed to
uses of that variable.
Consider the example: ‘x = {name: "Bob"}; x = {nam: "Bill"}; print(x.name);’
Most likely, the programmer intended the type of x to include a name field. If this type were
declared, the type checker would produce an error at statement two, due to the misspelled
nam field. With trace typing under the Tsub type system, we will instead compute a type for
E. Andreasen et al.
2311
τ ::= > | ⊥ | number | boolean | string | O | F |
O ::= PropName → τ
V
F ::= [τ ](τ ) → τ | F
S
τ | µX.τ
Figure 8 Types used in the type systems we model.
shapeMap(o) ≡
T.α(o) ≡
{(pi , vi ) | o.pi = vi at some trace point}
{(pi , τi ) | M = shapeMap(o) ∧ pi ∈ dom(M ) ∧ τi = T.t T.α(v)}
v∈M [pi ]
Figure 9 Type ascription for objects.
x that has no fields by merging the types of the two objects — essentially over-generalizing
due to the presence of what is well described as a static type error. Hence, the type error is
pushed to the access of name at statement three, which is no longer in the type of x.
Intuitively, we want that when mimicking two comparable source-level type systems, trace
typing ends up reporting fewer type errors for the more powerful one. Indeed, our results
show that this is generally the case. However, this is not guaranteed to be the case because
our ascription operation (T.α) can sometimes over-generalize, leading to unpredictable error
counts. See Section 7 for further discussion of this matter.
4
Instantiations
In this section, we detail several instantiations of the trace typing framework of Section 3,
showing its flexibility. We first describe a core type system that allows many of JavaScript’s
dynamic object behaviors. We then show how this system can be used to detect possible tag
tests in programs. Then, we present an extension of this system to test the usefulness of
parametric polymorphism. Finally, we show an instantiation with a different, stricter type
system for objects based on recent work [8], and how to evaluate various features of that
system using trace typing.
4.1
Core Type System
The types we use for most of our experiments appear in Figure 8. The top (>), bottom
S
(⊥), primitive, union ( ), and recursive types (µ) are standard. An object type (O) is a
map from property names to types. Function types (F) are either standard (with receiver,
argument, and return types), or an intersection type over function types. We use intersection
types as a very precise model of call-return behavior, for a limit study of the usefulness of
polymorphism (Section 6.2).
In our implementation, our object types are significantly more complex, mirroring the
many roles objects play in JavaScript. In JavaScript, functions are themselves objects, and
hence can have properties in addition to being invokable. Arrays are also objects with numeric
properties and possibly non-numeric as well. Our implemented object types model these
semantics, but we elide them here for simplicity. JavaScript also has prototype inheritance,
where a read of property p from object o is delegated to o’s prototype if p is not present
on o. For the core type system, object types collapse this prototype chain, so unshadowed
properties from the prototype parent are collapsed into the child object type (more discussion
shortly). The type system of Section 4.4 deals with prototype inheritance more precisely.
2312
Trace Typing: An Approach for Evaluating Retrofitted Type Systems (Extended Version)
Ascribing Object Types As noted in Section 3.1, defining the T.α operation for object
values is non-trivial. Since object properties may point to other objects, types for object
values are inter-dependent. Further, JavaScript objects may be mutated in many ways:
beyond changing property values, properties can also be added and deleted. Even the
prototype of an object may be mutated.
Trace typing is currently limited to object types that do not change over time, i.e., only
one type can be associated with each object instance during the entire execution. So, for an
object o, T.α(o) must merge together the different observed properties and property types
for o into a single object type. Nearly all practical type systems use object types in this
manner, as state-dependent object types require complex reasoning and restrictions around
pointer aliasing.
Figure 9 gives our definition of T.α for objects. We first define shapeMap(o), a map from
property names to sets of values (shown as a set of pairs). shapeMap(o) captures all the
property values that could be read from o at any point in the trace. (Note that we do not
require the property read to actually exist in the trace.) Since JavaScript property reads
may be delegated to the prototype object, shapeMap(o) gives a “flattened” view of o that
includes inherited properties. Consider the following example:
1
2
3
4
5
var
var
x.b
y.d
x.a
y
x
=
=
=
= {};
= { a : 4 } proto y ;
{ c : false };
" hello " ;
10;
// prototype inheritance shorthand from [8]
For this example, the shape map for the object allocated on line 2 is: [a 7→ {4, 10}, b 7→
{{c : false}}, d 7→ {”hello”}].
The object type T.α(o) is written as a set of pairs in Figure 9. The properties in T.α(o)
are exactly those in the domain of shapeMap(o). The type for property pi is obtained by
applying T.α to each value of pi in shapeMap(o), and then combining the resulting types
using T.t. When computing T.α, cyclic data structures must be detected and handled by
introducing a recursive type.
Type Checking The type rules we checked in our implemented type systems are standard.
Property accesses can be checked normally, since the accessed property name is always
evident in the trace. Function invocations are checked normally with one caveat: passing too
few arguments is permitted, as all JavaScript functions are implicitly variadic.
For type systems with union types, we optimistically allow most operations that work for
one case of the union (e.g., property access on a union typed expression) without validating
that the code has checked for the appropriate case, but treat assignment of union-typed
expressions to storage locations soundly (TypeScript and Flow have very similar rules for
unions). Our trace language does not currently include conditionals, making it non-trivial to
use tag tests to eliminate union types [19, 27]. However, our types do allow for discovering
likely tag tests in the program, as we show below.
In the case when T.t causes a type to be >, we count every use of and assignment to a
variable of that type to be an error.
4.2
Detecting Tag Tests
To soundly eliminate union types, a type system must support narrowing the union under a
conditional that checks some tag information of a value. While the theory for such narrowing is
rich and well-developed [19, 27], industrial type systems for JavaScript (Flow and TypeScript)
treat union elimination unsoundly in the same “optimistic” manner we do (Section 4.1). A
E. Andreasen et al.
2313
key issue is that tag tests can vary in complexity, from simple type checks (e.g., JavaScript’s
typeof operator) to complex combinations of user-defined predicates, and it is unclear what
level of complexity must be handled to provide adequate support for tag tests.
Using trace typing, we designed a technique to observe which tag test constructs are
being used most frequently. As described previously in Section 2.2, the technique works by
observing when there are two consecutive reads of a variable v (with no intervening write),
and v’s type is narrower at the second read than the first. This technique requires using the
FS setting in T.equiv, to keep separate types for each variable occurrence.4 For function
→
types, we used T. t from Tsub , thereby generalizing using only subtype polymorphism. In
Section 6.3, we discuss the kinds of tag tests discovered in our benchmarks with this technique
and their prevalence.
4.3
Parametric Polymorphism
→
→
Here, we define T. t poly , an instantiation of the T. t operator that discovers types with
parametric polymorphism. With this approach, one can evaluate the usefulness of adding
parametric polymorphism to a retrofitted type system.
→
T. t poly works by enumerating all possible type parameter replacements in each invocation
type for a function, and finally choosing the most general parameterized type matching all
invocations. The enumeration proceeds by replacing all subsets of occurrences of a concrete
type by a type variable. Consider the identity function, function id(x){ return x; },
and two invocations, id(3) and id(true). The observed invocation types for id are Number
→
-> Number and Boolean -> Boolean. (Applying T. t from Tsub to these types would yield
the unhelpful type (>) → > for id.) For the first type, the enumeration would yield types
X -> Number, Number -> X, and X -> X (along with the original type). After proceeding
→
similarly for the second type, T. t poly returns the type X -> X for id, as it is the most general
type matching both invocations.
→
In general, T. t poly also generates signatures with multiple type parameters. It also
attempts to replace types nested one level into each argument or return type, to discover
types operating over generic data structures. While this enumeration is exponential in the
arity of the function, we did not observe an appreciable slowdown in practice.
After discovering polymorphic types, we must check the invocations against those types
in T.check. The checking can be done by treating each type variable as a direct subtype
of >, distinct from other type variables. Our implementation currently uses an alternate
strategy of modifying the context-sensitivity policy in T.equiv, so that variables in function
invocations with distinct instantiations of the type parameters are not merged. This strategy
corresponds to checking a more powerful form of bounded polymorphism, which may yield a
reduced error count compared to directly checking unbounded type parameters.
4.4
Fixed Object Layout
We now show how to enhance the object types presented in Section 4.1 to support fixed object
layout, as described in recent work by Choi et al. [8]. In general, JavaScript objects is used as
dictionaries, with arbitrary addition and deletion of properties (keys). However, implementing
all objects in this manner would lead to poor performance. Real-world programs often do
not use this dictionary-like behavior. Objects often behave like records, in which the set of
4
Our implementation also disabled the assignment merging shown in Figure 6, to discover more tag tests.
2314
Trace Typing: An Approach for Evaluating Retrofitted Type Systems (Extended Version)
properties is stable: after initialization, new properties are not added or removed. Modern
just-in-time compilers detect the record usage pattern at run time, and use it to optimize
the object representation where possible.
Choi et al. [8] show how to enforce the record usage pattern in a type system, enabling
efficient static compilation of a JavaScript subset. Prototype inheritance complicates the
design of such a type system, since a write of an inherited property creates a shadowing
property in the child object. Consider the following example:
1
2
3
var o1 = { a :3 , m : function ( x ) { this . a = x ;}}
var o2 = { b :5} proto o1 // prototype inheritance shorthand from [8]
o2 . a = o2 . b ; // adds field a to o2
The write at line 3 adds a local a property to o2, rather than updating the inherited property
o1.a. The type system for fixed layout handles this case by distinguishing between read-write
properties local to an object and read-only properties inherited from the prototype chain.
With this distinction, the type system would reject the statement on line 3 above, as it is a
write to a read-only property.5
To track read-only and read-write properties in object types, we extend T.α from Section 4.1 as follows. Instead of tracking a single shape map for each object, we keep a
read-write shape map for locally-declared properties, and a read-only shape map for inherited
properties. In Section 4.1, the domain of shapeMap(o) includes any pi present on o at any
point in the execution. Here, to enforce the fixed layout property, we restrict the domain
of each shape map to properties present before the end of o’s initialization. (Recall from
Section 3.2 that our traces include end-initialization statements to mark these points.)
Given these two shape maps, computation of read-only and read-write properties in each
object type proceeds as in Figure 9. Type checking is enhanced to ensure that property
writes are only performed on read-write properties.
We also check for two additional properties from the Choi et al. type system [8]. The
type system requires any object used as a prototype parent to have a precise type, i.e., no
properties can have been erased from the type using width subtyping. (This is required to
handle subtle cases with inheritance [8].) In trace typing, we check this property by only
flagging types of object literals as precise, so if width subtyping is ever applied, the resulting
type is not flagged. Finally, the type system requires that if a property from a prototype
is shadowed, the parent and child properties have the same type. This condition is easily
checked in T.α once read-only and read-write properties are computed.
5
Implementation
We implemented the trace typing process described in Section 3 as a toolfor typing JavaScript
programs. This section reports on some of the more noteworthy aspects of the implementation.
Tool Architecture Our tool consists of two components: Trace Collector and Trace Typer.
Trace Collector is a Jalangi [25] analysis that obtains a trace of a program of interest
by monitoring the execution using source-level instrumentation; it consists of about 2500
lines of JavaScript code. Trace Typer implements the core trace typing framework. The
framework and the type systems described in Section 4 are implemented in about 4000 lines
of TypeScript, of which about 1000 lines are the type system implementations.
5
The type system of Choi et al. also tracks which properties are written on method receivers, to locally
check validity of inheritance [8]. With trace typing, this information need not be tracked explicitly,
since receiver types are propagated into method calls.
E. Andreasen et al.
Modeling JavaScript Semantics In our implementation, great care is taken to accurately
model JavaScript’s complex primitive operators. Much of the complexity in operator semantics
lies in the implicit conversions performed on operands. For example, the binary && operator
first coerces each of its operands to a boolean value; this coercion can be applied to a
value of any type. However, the value of the && expression is always one of the uncoerced
operands. The implementation creates trace statements that models these conversions
explicitly, enabling accurate data flow tracking and type ascription.
Modeling the Native Environment To handle interactions with the native environment
(e.g., built-in JavaScript library routines), we require a model of the native behavior. We
express native models using trace statements, avoiding the need for a separate modeling
language. In many cases, we can actually infer native models from the concrete states of the
execution, thereby avoiding the significant work of writing models by hand. Model inference
works by associating a new global escape variable with each object that can escape to the
native environment. Passing an object to the native environment is modeled as a write to the
corresponding escape variable, while retrieving the object from the environment is modeled
as a read. With this handling, a model can be automatically inferred for much of the native
environment, including complex functions such as the overloaded Array constructor. Further
details of this inference can be found in Appendix A.
6
Experiments
We report on several experiments we conducted using trace typing. First, we studied the
trends in error counts for six type system variants, generated by varying handling of subtyping
and function types (Section 6.2), yielding useful insights on the relative importance of these
variants. Second, we used trace typing to discover tag tests in our benchmark and characterize
their relative frequency and complexity (Section 6.3). Finally, we studied several questions
around the restrictiveness in practice of a type system for fixed object layout (Section 6.4).
Together, these experiments show the usefulness and versatility of the trace typing framework.
We first describe our benchmarks, and then present the experiments.
6.1
Benchmarks
We use nine popular packages from npm, the node.js package manager, as benchmark programs.
Figure 10 measures the static source code size of each program,6 as well as the number
of statements in the traces we generate and the number of static source lines covered by
our executions. underscore and lazy.js are utility libraries. esprima and escodegen are
a JavaScript parser and code generator, respectively. typescript is the compiler for the
TypeScript language.7 minimist, optparse and qs are parsers for command line options
and query strings. xml2js is an XML serializer/deserializer.
These benchmarks capture a range of interesting JavaScript styles. The parsers, code
generators, and compilers manipulate OO-style AST data structures In contrast, the utility
libraries and option parsers are highly reflective (e.g., constructing or extending objects using
dynamically-computed property names), but are otherwise written in a largely functional
style. Hence, the suite as a whole exercises a variety of features of trace typing.
6
7
Non-comment, non-blank lines as counted by cloc, excluding build scripts, tests and test harness code.
Figure 10 gives lines of JavaScript source (the compiler is bootstrapped).
2315
2316
Trace Typing: An Approach for Evaluating Retrofitted Type Systems (Extended Version)
Benchmark sizes
Fixed-object-layout error rates
Benchmark
LOC
LOC exec
Length
ro/rw
prototypal
inheritance
escodegen
esprima
lazy.js
minimist
optparse
qs
typescript
underscore
xml2js
2132
4610
2557
186
222
726
35456
1098
840
723
1052
1016
151
141
256
8139
727
275
375325
23986
25439
140812
15246
102637
167730
120081
73672
0/53
0/44
114/387
0/14
0/15
0/41
1/2085
2/175
0/42
0/0
0/2
0/17
0/0
0/1
0/0
2/3
0/0
1/2
0/217
0/298
2/470
0/81
0/51
0/62
0/3286
1/294
0/91
Total
52193
14541
1243857
117/2856
3/25
3/4850
Figure 10 Benchmarks sizes and error rates for the fixed-object-layout type system. ‘LOC’ is
the lines of code in the program. ‘LOC exec’ is the lines covered in our execution of the program.
‘Length’ is the total number of trace statements. ‘ro/rw’, ‘prototypal’ and ‘inheritance’ are the
error-rates for different each kind of type check.
To exercise the programs for trace generation, we wrote a separate driver for each
benchmark. For the option parsers and escodegen, this driver comprised extracted tool
invocations from their test suites. We exercised the utility libraries with code snippets from
their official tutorials.
6.2
Comparing Type System Variants
For our first experiment, we compared six type systems, generated by varying the T.t and
→
T. t operators over the core types of Section 4.1. For T.t, we used two options:
subtyping: This uses standard structural (width) subtyping for objects, as in Tsub from
Section 2.1 (see also Example 1 in Section 3.1).
unions: T.t treats two object types as in Subtyping, but also introduces a union type
when applied to an object type and a primitive (see Example 2 in Section 3.1).
→
For T. t, we use three options:
base: T.t is applied individually to argument and return types, as discussed in Section 3.1.
→
poly: This uses T. t poly to discover parametric polymorphism, as discussed in Section 4.3.
intersect: T.t simply combines two function types into an intersection of the two types.
This strategy yields more precise types than any practical type system, and hence is only
used as a limit study of the usefulness of polymorphic function types.
T.equiv uses FI for flow-insensitive variable types in all configurations. Context sensitivity
→
is varied in accordance with T. t: for base, we use CI, while for intersect we use CS (since
each invocation may have its own case in the intersection type). Context sensitivity for poly
was discussed in Section 4.3.
After collecting traces, we ran our trace typer with each of our type system variants,
counting the number of static (source location) type errors that arose. We include only type
errors that occur in the subject program, not in the libraries they depend on.
Experiments were run on a quad-core Intel Core [email protected] with 16GB RAM,
running Ubuntu 15.10 and node.js 5.0. Collecting traces and running all six configurations
required less than 15 minutes.
Our full data set is large—error counts for all configurations on all benchmarks—and is
included in Appendix B. We present two select slices of data here. Figure 11 presents the
E. Andreasen et al.
2317
300
200
100
0
unions
qs
ty
pe
sc
rip
un
t
de
rs
co
re
xm
l2
js
ar
se
tp
op
im
i st
m
in
zy
. js
la
rim
a
es
p
es
co
d
eg
en
subtyping
→
Figure 11 Error counts for subtyping and unions, selecting base for T. t.
300
200
100
0
base
qs
ty
pe
sc
rip
un
t
de
rs
co
re
xm
l2
js
ar
se
tp
op
im
i st
m
in
zy
. js
la
rim
a
es
p
es
co
d
eg
en
poly
intersect
→
Figure 12 Error counts for different T. t settings, selecting subtyping for T.t.
→
number of syntactic locations with errors for subtyping and unions, where T. t is base in
→
either case. Figure 12 shows error counts for the three T. t settings, where T.t is subtyping
in all three cases. The trends shown here are similar across other configurations.
Overall we see trends we would expect, with more powerful type systems yielding fewer
errors. There are two small inversions in the results, where a more precise configuration
(poly) produces more errors than a less precise configuration (base) discussed shortly.
This experiment let us find quantitative answers to the kinds of questions we raised in
Section 1 and Section 2. We remind the reader that the objective of these experiments is not to
pass verdict on the usefulness of certain type system features in an absolute sense. Rather the
objective is to help designers of retrofitted type systems prioritize features based on empirical
evidence rather than by intuition alone. If a type system feature ends up unsupported, then
users will need to work around that limitation in their code (or live with type errors).
How prevalent is the need for union types? While Figure 11 shows that union types do
reduce error counts over just using subtyping (as expected, since unions are more powerful),
the degree of reduction is relatively small. While TypeScript and Flow now both include
unions, TypeScript did not add them until two years after its initial release. The small error
count reduction due to union types in our experiments could help to explain why they were
not needed in TypeScript from the very beginning.8
How prevalent is the need for parametric polymorphism? For intersection types? As
shown in Figure 12, the intersection configuration always reduced errors compared to
poly, while poly almost always reduces errors compared to base9 . While the error count
8
9
TypeScript’s issue tracker has further discussion of their motivation for adding unions; see https:
//github.com/Microsoft/TypeScript/issues/805.
The inversions in esprima and optparse are due mostly to implementation quirks, such as an incomplete
merge operator on recursive types that returns > more than necessary.
2318
Trace Typing: An Approach for Evaluating Retrofitted Type Systems (Extended Version)
reductions for poly are not dramatic in general (typescript is discussed below), parametric
polymorphism is crucial for ascribing types to core routines manipulating data structures
(e.g., array routines), and hence its inclusion in TypeScript and Flow is unsurprising. The
dramatic drop in error count for several programs under the intersection configuration gives
strong evidence that JavaScript requires types that can express non-uniform polymorphism,
such as intersection types. In fact, TypeScript has long supported overloaded functions for
expressing intersection types.10
The typescript benchmark stands out from the others as it was originally implemented
in TypeScript, and hence we would expect it to be mostly well typed. Our results confirm this:
the error count is only 290 in 8139 (JavaScript) lines of code for the least precise configuration,
a much lower error rate than the other programs. Also, the significant reduction in errors
going from base to poly for the benchmark (see Figure 12) corresponds nicely to the use of
parametric polymorphism in the TypeScript code base; see Section 7 for an example.
While in the above experiment we used trace typing to retroactively find evidence of the
need of type system features, the next two experiments are concerned with more exploratory
questions.
6.3
Finding Tag Tests
Here, we provide the first empirical analysis of the use of tag tests in JavaScript which
identifies tag tests based on observed narrowing (as described in Sections 2.2 and 4.2), rather
than syntactic criteria. We post-process the ascription results to automatically identify
occurrences of guard predicates using the techniques of Section 4.2. We then manually
analyze the predicates to identify non-local guards (where the boolean result of a check is
computed outside a syntactic conditional statement), non-atomic guards (i.e., conjunction
and disjunction of guards), and predicate functions (other functions whose result indicates a
certain refinement).
Methodology As mentioned in Section 4.2, we detect tag tests by observing consecutive
reads of a variable where the type is narrowed at the second read. Usually, the first read of
this pair will be the use in the type guard. In cases where the guard expression is saved to a
local variable before being checked, this technique detected spurious additional guards, but
they were easy to recognize and discard manually.
We ran the set of benchmark programs from Section 6.1 with the union-producing
T.t operator of Section 6.2, and a flow sensitive T.equiv policy to keep types of variable
occurrences under conditionals separated.
Results It took 90 minutes to manually classify all of the discovered type guards, the results
of which can be seen in Figure 13. The figure contains rows for different kinds of type guards,
if a type is non-atomic, it can be classified as having multiple kinds. Due to their prevalence,
extra rows have been added for type guards that are applied on the property of the guarded
value.
Overall there are 6 major classes of narrowing checks: typeof, instanceof, checking an
explicit tag field in a data structure11 (mostly in the AST manipulations in our corpus), a
general predicate function, explicit property presence checks (hasOwnProperty), and checks
of objects’ prototype or constructor fields. The most prevalent type guards are typeof,
10
11
See http://www.typescriptlang.org/Handbook#functions-overloads.
A popular requested TypeScript feature: https://github.com/Microsoft/TypeScript/issues/186
E. Andreasen et al.
Measure
2319
Count
Example
Type guards
164
-
Non-local
Non-atomic
6
46
var isObj = typeof x == ’object’; if(isObj)
if(typeof x == ’string’ || typeof x == ’object’)
typeof
instanceof
tag field
predicate function
property presence
.prototype/.constructor
30
9
16
30
35
3
typeof x === ’function’
x instanceof Function
x.kind = Kinds.Template
Array.isArray(x)
x.hasOwnProperty(’prop’), x.prop === undefined
x.constructor === Function
.. test on property
37
typeof x.prop === ’function’
other
23
className === toString.call(x)
Figure 13 Tag test classification results.
use of general predicate functions, and (a variety of) property presence tests. Checking an
explicit tag field, used to mimic algebraic datatypes in JavaScript, are common but highly
program-dependent. Non-local type guards are rare, and would be easy to rewrite if needed.
Boolean combination of type guards is common enough that they probably should not be
ignored by a sound union elimination check.
6.4
Evaluating Fixed Object Layout
We present a case study of evaluating Choi et al.’s type system [8] to enforce fixed object
layout in JavaScript, as described in Section 4.4. We checked the following salient features in
the trace typing framework:
1. Property writes We checked that property writes were only performed on properties
deemed to be “read-write” properties in the type system.
2. Precise types We checked that only objects with a “precise” type (see Section 4.4) were
used as a prototype parent for other objects.
3. Consistent shadowing We checked that when properties were shadowed in the prototype
chain, the parent and child properties had the same type.
We ran trace typing with this type system with the benchmark suite of Section 6.1 (with
the subtyping T.equiv and base policy from Section 6.2, merging all occurrences of a variable
in a dynamic function invocation). Figure 10 shows the errors reported by our trace typing
model. The column ro/rw shows, in absolute numbers, how many of the property writes were
to fields that were inherited from prototype chain (and not shadowed by a local redefinition),
of the total number of property writes. The column prototypal shows, how many of the
assignments to some object’s prototype property was in violation of the restriction mentioned
(second bullet) above, vs total number of such assignments. The column inheritance shows
in how many cases, a property was shadowed with an inconsistent type.
The results add significant confidence to our initial beliefs that some of the restrictions
the type system imposes are not onerous in practice: the numerator numbers are generally
(though not universally) quite small compare to the denominator. We manually found that
most of the ro/rw type errors for lazy.js were due to the way prototype hierarchies were
initialized, and that the type errors could be eliminated by a simple refactoring.
2320
Trace Typing: An Approach for Evaluating Retrofitted Type Systems (Extended Version)
// TypeScript / src / compiler / parser . ts :39
export function forEachChild <T >( node : Node , ...) : T {
... switch ( node . kind ) {
... case SyntaxKind . PropertySignature :
... return visitNode ( cbNode , ( < VarLikeDecl > node ) . propName ) ;
}
}
// TypeScript / src / compiler / type . ts :504
export interface VarLikeDecl extends Declaration {
propName ?: Identifier ; \ n dotDotDotToken ?: Node ;
name : DeclarationName ; \ n questionToken ?: Node ;
type ?: TypeNode ; \ n initializer ?: Expression ;
}
Figure 14 Code fragment from typescript.
Additionally, from the trace typing data (not presented here) we found that support
for optional properties in the type system could have helped eliminate a large number of
type errors, and so would be a very useful feature to add to the type system. Trace typing
finds this information without the need for implementing an actual static type inferencer for
JavaScript that is robust enough to work on large code bases.
7
Discussion
In this section, we discuss the quality of types inferred by trace typing, present limitations
and threats to validity, and then mention possible future generalizations of our approach.
Quality of Types Inferred Because our type inference approach is non-standard and our
error counts are sensitive to over-generalization (Section 3.5), it is worth asking whether the
types we infer are sensible — whether a human developer would write similar types. While
there is no systematic way to verify this, our experience says that the types are sensible.
In the course of developing the framework and type system plugins, we spent a significant
amount of type examining the types inferred. Object (property) types rarely deviated from
what we would have written ourselves.
We discuss one specific example of types inferred by trace typing for the typescript
benchmark, for which the original TypeScript code contains declared types (trace typing
analyzes the transpiled JavaScript version). Figure 14 contains an excerpt from the TypeScript
code. At the return statement the node variable always has type VarLikeDecl (also shown
in the figure). Trace typing finds the relevant properties that we see mentioned in the
interface, as well as the inherited properties. Moreover, it also correctly computes the
structural type of each property of VarLikeDecl such as initializer. Due to space reasons,
we do not show all the inferred types; but we found them reasonable on manual inspection.
Our ascription of parametric polymorphic types (Section 4.3) also works well in practice.
For example, it computes these polymorphic signatures for array methods (from the native
environment):
Array <E >. prototype . indexOf : ( E ) -> Number
Array <E >. prototype . concat : ( Array <E >) -> Array <E >
Array <E >. prototype . push : ( E ) -> Number
Array <E >. prototype . pop : () -> E
These types are similar to the types used for native environment by TypeScript programs.12
12
https://github.com/Microsoft/TypeScript/blob/v1.7.3/lib/lib.core.d.ts#L1008
E. Andreasen et al.
The exceptions to the above arise mostly where type errors occur, and in those cases the
types are not bizarre but simply a different choice among several incompatible options (e.g.,
when a method is overridden with an incompatible type). For the intersect configuration
of Section 6.2, with unbounded intersection types, the gap between inferred types and
what one would write at the source level is larger, due to the synthetic nature of the type
system. Designing a trace typing system that infers a mix of parametric polymorphism and
intersection types closer to what a person would write is future work.
Limitations and Threats to Validity Trace typing works under the assumption that the
input programs are mostly type correct; the error locations generated by our framework
lose accuracy when this assumption is violated. Moreover, we assume that most parts of a
program have a correct type in the system being modeled, which may not always hold. Our
manual inspection of inferred types suggests that the natural formulations for T.α and T.t
generally behave well, but it is possible other pathological cases exist.
A result that type system design A yields fewer type errors than design B in our framework
does not strictly imply that the same will hold for complete source-level implementations,
due to inevitable over-approximations in a static type checker. Constructs inducing such
approximations include reflective constructs like dynamic property accesses (which are
resolved in our trace) and loops. Type system feature interactions can exacerbate this effect.
The best use of trace typing is to compare type systems that differ in only one or two
parameters, and to give less weight to very small differences in the number of errors detected.
Consider this example:
1
2
3
4
5
6
function f ( x ) { return x . p ; }
f ({ p :3 , q :4}) ;
f ({ p :3 , r : true }) ;
var y = { p :3};
y = { p :3 , q :5};
f(y);
Here, type ascription’s over-generalizing in the presence of what would be static type
errors can lead to some unexpected results. For the example above, ascription in the base
polymorphism configuration (see Section 6.2) with no union types would produce the type
{p:Number} -> Number for f, which would allow the trace of the program to type check.
Ascription with the more precise intersect configuration would give the intersection type {p:
Number,q:Number} -> Number ∧ {p:Number,r:Boolean} -> Number. But, the two writes
to y ensures that it is ascribed type {p:Number}, which is not a subtype of either argument
type in the intersection! In such cases, a more powerful trace type system can produce more
errors than a less powerful one. We did not observe this issue in experimental results we
inspected manually, but it may have occurred elsewhere.
Our work is an empirical study using dynamic analysis, and we inherit the normal threats
to validity from such an approach, namely possible sampling bias in the subject programs
and sampling bias from the inputs used to gather traces. For our study, we took inputs from
each program’s often-substantial test suite, which appeared from manual inspection to give
reasonable path and value coverage.
Potential Extensions We have focused thus far on type systems for JavaScript, but the
core ideas of our technique apply to other languages as well. The generalization to other
dynamically typed languages (e.g., Ruby, Python) would require instrumentation of the other
language. The type system implementations on top of the new framework would need to be
tailored to the primitives and idioms of the new target language.
2321
2322
Trace Typing: An Approach for Evaluating Retrofitted Type Systems (Extended Version)
Our existing trace format could be used for additional experiments that analyze the types
ascribed to specific parts of the program. For example, this could inform strategies for typing
challenging constructs like computed property accesses [23].
More metadata could be added to our traces to enable experimentation with other
type-system features, with no changes to our core approach. For example, more detailed
information about the evolving types and layouts of objects could be maintained, to enable
experimentation with alias types [9, 26]. Our hope is that trace typing will be a useful base
for such future extensions.
8
Related work
We focus here primarily on approaches to retrofitting type systems onto existing languages
and on other means of evaluating type systems on large bodies of code, with a secondary
focus on type systems for JavaScript.
Retrofitting Type Systems Retrofitting type systems onto existing languages is not new,13
nor is it restricted to dynamic languages. The notion of soft typing for dynamically typed
languages is now well-established [6, 27, 30]. The notion of pluggable type systems — including
extensions to existing type systems — is also an established idea [5], which continues to
produce useful tools [11, 22]. Both bodies of work suffer the same difficulty in evaluating lost
precision: it requires tremendous manual effort to evaluate on non-trivial amounts of code,
leading to design decisions based on intuition and a handful of carefully-chosen examples.
More recently, two groups of researchers arrived at means to induce a gradual type system
variant of a given base type system, one as a methodology to produce a gradual type system
by careful transformation of a static type system [13] and another as a a tool to compute a
gradual version of a type system expressed as a logic program [13]. Both deterministically
produce one gradual type system from one static type system, aiding with the introduction
of gradual typing, but not of overall design.
Evaluating Type Systems The implementation effort for prototyping similar but slightly
different type systems and applying them is a significant barrier to evaluating completely
new type systems for existing programming languages. TeJaS [20] attempts to remedy this
for JavaScript type systems by carefully architecting a modular framework for implementing
type systems. It relies on bidirectional typing to reduce annotation burden, but this still
requires significant manual annotation.
A smaller but closely-related body of work is evaluating changes to existing type systems.
Wright studied approximately 250,000 lines of SML code [29], cataloguing the changes necessary to existing code in order to type-check under the value-restriction to type generalization.
The study in this case required less implementation effort than the general approach we
propose; SML implementations existed, and could be modified to test the single alternative
type system change. Nonetheless, his careful analysis of the impact of the proposed change
justified a type system change that has withstood the test of time. SML still uses the value
restriction, and OCaml uses a mild relaxation of it [14]. More recently Greenman et al. [17]
performed a similar evaluation of an alternative proposal for F-bounded polymorphism in
Java.
13
The use of the term retrofitting is relatively recent, due to Lerner et al. [20].
E. Andreasen et al.
Dynamic Type Inference Type inference for dynamic languages is another rich area with
strong connections to our trace-based type inference. The closest such work to ours is dynamic
type inference for Ruby [2]. They use Ruby’s reflection capabilities to pass virtualized values
through a program, with each use of the virtualized objects gathering constraints on the
type of the object. Their system has a guarantee that the inferred types will be sound if
every path within each procedure is executed at least once. Like trace typing, they rely on
dynamic information to infer types for program fragments, including for complex source-level
constructs. Unlike our approach, they make no attempt to simulate the effect of type checking
the program: they infer types that hold at method entry and exit, but do not infer types
for local variables, whose types may change arbitrarily. Also, they cannot infer method
polymorphism (though they support annotations), require manual annotation for native code,
and evaluate only a single type system design.
Saftoiu et al. [24] implement dynamic type inference for JavaScript, generating types
for a particular JavaScript type system based on dynamic observation, and find that it is
useful for converting code to their typed JavaScript dialect. Their type system is sound with
respect to λJS [18], and they can type-check the source code to verify the inferred types.
We cannot do this, because our goal is experimenting with type system design, so our type
systems are not adequate to type check source programs. Like us they instrument code via
source-to-source translation. We evaluate significantly more code: their examples total 3799
lines of code.
Other Uses of Traces Coughlin et al. [10] use dynamic traces to identify the spans between
program safety checks and data uses guarded by those checks, e.g., a null check followed
by a dereference. These measurements can inform the design of static analyses, e.g., by
indicating whether or not an analysis should be interprocedural. While similar in philosophy
to our work, we differ substantially in technical details. First, their implementation is specific
to null dereference errors. Second, their goal is to measure aspects of program behavior
quantitatively, and interpret that measurement when designing an analysis; we instead
advocate directly comparing approximations of a real analysis.
Åkerblom and Wrigstad [1] study the receiver polymorphism of a very large base of open
source Python programs, primarily to examine how adequate nominal subtyping might be
for Python (as opposed to structural subtyping). They classify static call sites by whether
they were observed to have multiple receiver types, consider whether call sites might be
dispatched on parametrically-polymorphic targets by clustering call sites in different ways
(roughly similar to our context-sensitivity policies), and do some reasoning about cases
where different static calls on the same syntactic receiver may have different qualities (a
coarse approximation of typing local variables). They do not reason about actual method
types, argument types (except implicitly if methods are invoked on arguments), and do not
attempt to simulate source type checking as we do. ECMAScript 5 — the previous version
of JavaScript — lacks classes, so even code written to ECMAScript 6 — which added classes
— will generally require structural subtyping. Our type ascription machinery could be used in
a style similar to Section 6.3 to implement most of their instrumentation.
Type Systems for JavaScript Most previous work on static typing for JavaScript focuses
on novel approaches to typing its most expressive features; we describe particularly relevant
examples here. Guha et al.’s flow typing [19] allows for refining types in code guarded
by runtime tag checks. DJS [9] implements a sophisticated dependent type system that
includes a type-based analogue [26] of separation logic to describe heap changes, and an
SMT-based refinement type system for characterizing the shape of objects. TeJaS [20] builds
2323
2324
REFERENCES
an expressive framework for experimenting with JavaScript type systems, whose core is an
ω
extension to F<:
.
Throughout we have discussed Flow [12] and TypeScript [4, 21]. They are quite similar
(objects, unions with unsound elimination rules, intersections) with a few small points of
divergence (singleton string types and intentional unsound subtyping in TypeScript, nonnullable fields in Flow). They have very different approaches to type inference and checking:
TypeScript uses limited local type inference, while Flow performs a global (per-module) data
flow analysis. We have already modeled key features — objects with structural subtyping,
union and intersection types — in our framework.
TypeScript and Flow include forms of parametric polymorphism, based on programmer
annotations. Trace typing can approximate some forms of parametric polymorphism (Section 3.4). Both TypeScript and Flow also allow some narrowing of union types based on tag
tests (unsound in general due to aliasing and mutation). Trace typing could be extended to
handle such features in future work, as discussed in Section 7.
Both TypeScript and Flow have added type system features based on user feedback, with
great success. Trace typing provides an alternate, complementary source of information for
deciding which type system features to include, which can be employed without having a
large user base.
9
Conclusions
We presented a framework for quantitatively evaluating variations of a retrofitted type system
on large code bases. Our approach involves gathering traces of program executions, inferring
types for instances of variables and expressions occurring in a trace, merging types according
to merge strategies that reflect options in the source-level type system design space, and
type-checking the trace. By leveraging the simple structure of traces, we dramatically reduce
the effort required to get useful feedback on the relative benefit of type system features.
To evaluate the framework, we considered six variations of a type system retrofitted
onto JavaScript. In each case, we measured the number of type errors reported for a set of
traces gathered from nine JavaScript applications (over 50KLOC) written in a variety of
programming styles. The data offers quantitative validation for some of the design choices in
two popular retrofitted type systems for JavaScript (Typescript and Flow). In a different
experiment, we used the results of trace typing to automatically identify places where type
narrowing occurred dynamically, to gather empirical results on the frequency and variety of
tag tests in JavaScript. In yet another experiment, we evaluated how onerous the restrictions
for a new retrofitted JavaScript type system [8] were, validating intuitions that our restrictions
were mostly reasonable, and identifying priorities for future extensions.
The feasibility of carrying out these experiments is a strong validation of the trace typing
approach.
References
1 Beatrice Åkerblom and Tobias Wrigstad. Measuring polymorphism in Python programs.
In Proceedings of the 11th Symposium on Dynamic Languages (DLS), 2015.
2 Jong-hoon David An, Avik Chaudhuri, Jeffrey S Foster, and Michael Hicks. Dynamic
Inference of Static Types for Ruby. In POPL, 2011.
3 Chris Andreae, James Noble, Shane Markstrum, and Todd Millstein. A Framework for
Implementing Pluggable Type Systems. In OOPSLA, 2006.
REFERENCES
4 Gavin Bierman, Martín Abadi, and Mads Torgersen. Understanding TypeScript. In
ECOOP, 2014.
5 Gilad Bracha. Pluggable Type Systems. In OOPSLA Workshop on Revival of Dynamic
Languages, 2004.
6 Robert Cartwright and Mike Fagan. Soft Typing. In PLDI, 1991.
7 Avik Chaudhuri. Bounded Polymorphism. URL: http://flowtype.org/blog/2015/03/
12/Bounded-Polymorphism.html.
8 Philip Wontae Choi, Satish Chandra, George Necula, and Koushik Sen. SJS: A Type
System for JavaScript with Fixed Object Layout. In SAS, 2015.
9 Ravi Chugh, David Herman, and Ranjit Jhala. Dependent Types for JavaScript. In
OOPSLA, 2012.
10 Devin Coughlin, Bor-Yuh Evan Chang, Amer Diwan, and Jeremy G. Siek. Measuring
enforcement windows with symbolic trace interpretation: What well-behaved programs
say. In ISSTA, 2012.
11 Werner Dietl, Stephanie Dietzel, Michael D Ernst, Kivanç Muşlu, and Todd W Schiller.
Building and Using Pluggable Type-Checkers. In ICSE, 2011.
12 Facebook. Flow: A Static Type Checker for JavaScript. http://flowtype.org/.
13 Ronald Garcia, Alison M. Clark, and Éric Tanter. Abstracting Gradual Typing. In
POPL, 2016.
14 Jacques Garrigue. Relaxing the Value Restriction. In Functional and Logic Programming,
2004.
15 Google. Closure. https://developers.google.com/closure/.
16 Colin S Gordon, Werner Dietl, Michael D Ernst, and Dan Grossman. JavaUI: Effects for
Controlling UI Object Access. In ECOOP, 2013.
17 Ben Greenman, Fabian Muehlboeck, and Ross Tate. Getting F-bounded Polymorphism
into Shape. In PLDI, 2014.
18 Arjun Guha, Claudiu Saftoiu, and Shriram Krishnamurthi. The Essence of JavaScript.
In ECOOP, 2010.
19 Arjun Guha, Claudiu Saftoiu, and Shriram Krishnamurthi. Typing Local Control and
State Using Flow Analysis. In ESOP, 2011.
20 Benjamin S Lerner, Joe Gibbs Politz, Arjun Guha, and Shriram Krishnamurthi. TeJaS:
Retrofitting Type Systems for JavaScript. In DLS, 2013.
21 Microsoft. TypeScript Handbook. http://www.typescriptlang.org/Handbook.
22 Matthew M Papi, Mahmood Ali, Telmo Luis Correa Jr, Jeff H Perkins, and Michael D
Ernst. Practical Pluggable Types for Java. In ISSTA, 2008.
23 Joe Gibbs Politz, Arjun Guha, and Shriram Krishnamurthi. Semantics and Types for
Objects with First-Class Member Names. In FOOL, 2012.
24 Claudiu Saftoiu, Arjun Guha, and Shriram Krishnamurthi. Runtime Type-Discovery for
JavaScript. Technical Report CS-10-05, Brown University, 2010.
25 Koushik Sen, Swaroop Kalasapur, Tasneem Brutch, and Simon Gibbs. Jalangi: A selective
record-replay and dynamic analysis framework for JavaScript. In FSE, 2013.
26 Frederick Smith, David Walker, and Greg Morrisett. Alias Types. In ESOP, 2000.
27 Sam Tobin-Hochstadt and Matthias Felleisen. Logical Types for Untyped Languages. In
ICFP, 2010.
28 Jonathan Turner. Announcing TypeScript 1.4. URL: http://blogs.msdn.com/b/
typescript/archive/2015/01/16/announcing-typescript-1-4.aspx.
29 Andrew K. Wright. Simple Imperative Polymorphism. LISP and Symbolic Computation,
8(4):343–355, 1995.
2325
2326
REFERENCES
30 Andrew K. Wright and Robert Cartwright. A Practical Soft Type System for Scheme.
ACM TOPLAS, 19(1):87–152, January 1997. doi:10.1145/239912.239917.
31 Haiping Zhao, Iain Proctor, Minghui Yang, Xin Qi, Mark Williams, Qi Gao, Guilherme
Ottoni, Andrew Paroski, Scott MacVicar, Jason Evans, and Stephen Tu. The HipHop
Compiler for PHP. In OOPSLA, 2012.
A
Modeling the Native Environment
In Section 5 we briefly mentioned that we automatically construct models of the native
environment. This appendix gives full details.
We rely on three procedures, toNative, fromNative and allocate (see Figure 15 for
pseudo-code14 ) that performs this inference:
toNative: This function models the four cases in which objects escape to native functions:
the object is (i) the this-object or (ii) an argument in a call to a native function, or
(iii) the object is returned to a native function as part of a callback or (iv) the object
is thrown as an exception. An object, o, that escapes to a native function forces a call
to toNative(o), which uses a unique escape-variable for the object to ensure that the
model represents the flow of that object regardless of how it is actually used in the native
function.
fromNative: This function handles the four cases where objects enter the application from
the native environment: the object is (i) the this-object or (ii) an argument in a callback
from native code, or (iii) the object is returned from a native function, or (iv) the
object is caught as an exception. An object, o, that comes from the native environment
forces a call to fromNative(o) which will create a trace read-expression for the special
escape-variable for the object. If the object has not been observed before, it assumed to
be freshly allocated by the native environment, so a call to allocate is also performed.
allocate: An object, o, that is observed during the execution has an associated call to
allocate(o), which will create a trace allocate-expression and the necessary trace
statements to initialize the fields and prototype of o. This process might allocate further
objects recursively. The initial global JavaScript object is modeled with a single call to
allocate(global).
The use of an escape variable per escaping object o may seem useless, but during type
propagation this results in the merge of each type at which the object o escapes (e.g., if
o is thrown from source locations that see different subsets of its fields), and allows that
imprecision to manifest where the object re-enters the program from the native environment
(e.g., a catch of o will see the merge of the types o had at throws) so we can soundly
overapproximate the typing consequences of the native data flow.
As an example, consider the inference of a model for a call to the native Array constructor
function. This function takes a variable number of arguments, and returns an array containing
those arguments; e.g., Array({p:1},{p:2}) returns [{p:1},{p:2}]. The arguments to this
call will be assigned to special escape variables via toNative. Then, a call to allocate will
allocate the result array, as it is a fresh object from a native call. Finally, via fromNative,
the elements of the result array will be set to the corresponding arguments of the call by
reading the corresponding special escape-variables.
14
Here, we only cover the important case where objects are exchanged with the native environment. Only
minor changes to these steps are needed to account for primitives.
REFERENCES
toNative (o , fromVar ) {
makeTraceWrite ( getEscapeVar ( o ) , fromVar ) ;
}
fromNative (o , toVar ) {
escapeVar = getEscapeVar ( o ) ;
if ( has N o tB e e nA l l oc a t ed ( o ) ) {
allocate (o , escapeVar ) ;
}
makeTraceWrite ( toVar , escapeVar ) ;
}
allocate (o , allocVar ) {
makeT r ac e A ll o c at i o n ( allocVar )
for ( var field in o ) {
fieldVar = makeVariable () ;
fromNative ( o [ field ] , fieldVar ) ;
make T r ac e F ie l d Wr i t e ( allocVar , field , fieldVar ) ;
}
}
Figure 15 The three main procedures for inferring a model for the native environment.
To handle cases where the automated inference described above would produce unsound
results, we allow tool users to specify that certain functions need to be modeled manually15 .
In such cases, the user needs to create a custom model that creates the appropriate trace
statements. This is the case for side effecting operations that may change the type of some
property, such as Array.prototype.sort which rearranges the elements of an array. Note
that, in such a rearrangement case, the model does not need to take care of the general
semantics of Array.prototype.sort, just the concrete invocations: The trace statements
that reproduce the rearrangement can be generated by inspecting shape of the array before
and after the call to Array.prototype.sort.
B
Full Results for six Type Systems
Figure 16 give the full results for the experiments in Section 6.
15
For the experiments in this paper, we modeled the behavior of 11 functions manually.
2327
2328
REFERENCES
escodegen
Base
subtyping
subtyping
subtyping
unions
unions
unions
minimist
Base
subtyping
subtyping
subtyping
unions
unions
unions
typescript
Base
subtyping
subtyping
subtyping
unions
unions
unions
Context
base
poly
intersect
base
poly
intersect
Err
151
136
40
121
106
25
Context
base
poly
intersect
base
poly
intersect
Err
162
155
64
126
122
64
Context
base
poly
intersect
base
poly
intersect
Err
290
223
144
283
217
128
esprima
Base
subtyping
subtyping
subtyping
unions
unions
unions
optparse
Base
subtyping
subtyping
subtyping
unions
unions
unions
underscore
Base
subtyping
subtyping
subtyping
unions
unions
unions
Context
base
poly
intersect
base
poly
intersect
Err
34
36
32
32
34
29
Context
base
poly
intersect
base
poly
intersect
Err
13
17
8
14
14
9
Context
base
poly
intersect
base
poly
intersect
Err
253
231
65
208
190
58
lazy.js
Base
subtyping
subtyping
subtyping
unions
unions
unions
qs
Base
subtyping
subtyping
subtyping
unions
unions
unions
xml2js
Base
subtyping
subtyping
subtyping
unions
unions
unions
Context
base
poly
intersect
base
poly
intersect
Err
78
57
51
67
56
51
Context
base
poly
intersect
base
poly
intersect
Err
94
89
43
78
73
40
Context
base
poly
intersect
base
poly
intersect
Err
47
39
22
45
39
27
Figure 16 Trend experiment results comparing type systems with and without unions, for various
models of polymorphism.
| 6cs.PL
|
Best Effort and Practice Activation Codes
Gerhard de Koning Gans1 and Eric R. Verheul1,2
arXiv:1101.0698v2 [cs.CR] 23 Jun 2011
1
Institute for Computing and Information Sciences
Radboud University Nijmegen
P.O. Box 9010, 6500 GL Nijmegen, The Netherlands
{gkoningg,eric.verheul}@cs.ru.nl
2
PricewaterhouseCoopers Advisory
P.O. Box 22735, 1100 DE Amsterdam, The Netherlands
[email protected]
Abstract. Activation Codes are used in many different digital services
and known by many different names including voucher, e-coupon and
discount code. In this paper we focus on a specific class of ACs that are
short, human-readable, fixed-length and represent value. Even though
this class of codes is extensively used there are no general guidelines for
the design of Activation Code schemes. We discuss different methods that
are used in practice and propose BEPAC, a new Activation Code scheme
that provides both authenticity and confidentiality. The small message
space of activation codes introduces some problems that are illustrated by
an adaptive chosen-plaintext attack (CPA-2) on a general 3-round Feistel network of size 22n . This attack recovers the complete permutation
from at most 2n+2 plaintext-ciphertext pairs. For this reason, BEPAC
is designed in such a way that authenticity and confidentiality are independent properties, i.e. loss of confidentiality does not imply loss of
authenticity.
Keywords: activation code, e-coupon, voucher, Feistel network, small
domain encryption, financial cryptography.
1
Introduction
This paper introduces Activation Codes (ACs) as a generic term for codes that
are used in many different digital services. They are known by many different
names including voucher, e-coupon and discount code. The common properties
of these codes are that they need to be short, human-readable, have a fixed
length and can be traded for economic benefit. There are schemes [4,5,8] that
include all kinds of property information in the code itself or include digital
signatures [14,12]. This makes the codes unsuitable for manual entry and thus
for printing on products, labels or receipts. The focus of this paper is on ACs
that can be printed and manually entered such as the AC that is printed on a
receipt in Figure 1. In this case the customer can enter the AC ‘TY5FJAHB’ on a
website to receive some product. We propose a scheme called BEPAC to generate
and verify this class of ACs. BEPAC is an acronym for Best Effort and Practice
Activation Codes. Here ‘best practice’ covers the use of a keyed hash function to
satisfy authenticity and ‘best effort ’ covers the use of a Feistel network to satisfy
confidentiality.
Security plays an important role in the
design of an AC system because of the
economic value it represents. A system
breach could result in big financial losses.
Nevertheless, to the best of our knowledge, there are no guidelines on the design of secure AC systems that consider
the previously mentioned properties. Despite the lack of general guidelines for
good practice, ACs are extensively used.
This underlines the need for a proper AC
Fig. 1: Activation Code
scheme that relies on elementary, wellstudied, cryptographic primitives to provide authenticity and confidentiality.
First, we discuss some examples that illustrate the need for a scheme that provides both confidentiality and authenticity. Then, we give a general definition of
an AC scheme and use it as a reference throughout this paper.
Our Contribution This paper addresses some known methods that are used to
generate ACs and proposes BEPAC, an AC scheme that combines best effort
with best practice. BEPAC is based on well-studied cryptographic primitives
to guarantee unique and authentic codes that provide a satisfactory level of
confidentiality. Confidentiality is obtained by a Feistel construction. The Feistel
construction has weak theoretical bounds when it is used on a small domain,
therefore we do not rely on it for authenticity. We use the work of Black and Rogaway [3] on small domain encryption and make some small changes to achieve
confidentiality. A practical attack on a general 3-round Feistel network is presented to demonstrate the weak bounds of the Feistel construction. For authenticity, we solely rely on a keyed-hash message authentication code (HMAC) of
the serial number, where the size of this HMAC determines the probability of
successfully guessing a valid AC. This separated design allows a separate analysis of both confidentiality and authenticity. An advantage of this approach is
that authenticity is not automatically compromised when confidentiality is broken. Finally, a BEPAC solution fits on a smart card and therefore allows AC
generation and clearance to be performed in a controlled environment.
Related Work Black and Rogaway [3] propose the Generalized Feistel Cipher
(GFC) as a solution to small domain encryption. This elegant solution can be
used to construct a permutation on any finite domain. In the BEPAC scheme we
use their method in a slightly adapted way and solely to provide confidentiality. Black and Rogaway provide an adapted proof of Luby [15] to prove secrecy
of a 3-round Feistel network. However, in their example configuration, the single DES round function does not give the 3 × 56-bit security since single DES
can be broken by exhaustive key search [25]. Moreover, in this setting it can
be broken round by round which actually means that we have 58-bit security.
Bellare et al. [2] propose 128-bit AES as a pseudo-random function which drastically increases the effort needed to break one round of the Feistel network
by brute-force. However, the adaptive chosen-plaintext attack (CPA-2) on a 3round Feistel network presented in this paper shows that the key length of the
pseudo-random function used in each round does not have any influence on the
attack complexity. Research on Feistel networks [22,20,19,15,13,21] has resulted
in theoretical security bounds. Feistel constructions of six or more rounds are secure against adaptive chosen plaintext and chosen ciphertext attacks (CPCA-2)
when the number of queries m ≪ 2n , see [22,13]. There is more related literature
on the design of ACs, but to the best of our knowledge there are no proposals
for the class of AC schemes that we discuss in this paper. Blundo et al. [4,5,8]
introduce an e-coupon which is 420 bytes in size. This scheme uses a message
authentication code (MAC) over some characterizing data like the identity of
the manufacturer, name of the promoted product, expiry date etc. The resulting
e-coupons contain valuable information but are too large to be entered manually
by a user. In the work of Kumar et al. [14] and Jakobssen et al. [12] a coupon is
basically a digital signature which also means that they describe relatively large
e-coupons. Chang et al. [6] recognize the problem of efficiency and describe a
scheme that is more suitable for mobile phones that have less processing power.
On the one hand, they circumvent the use of public key cryptography which
reduces the computational complexity, but on the other hand, their scheme describes relatively long codes. None of the schemes described previously satisfy
the requirement of short codes that can be entered manually.
Matsuyama and Fujimura [17] describe a digital ticket management scheme
that allows users to trade tickets. The authors discuss an account based and a
smart card based approach and try to treat different ticket types that are solely
electronically circulated. This in contrast with BEPAC which focuses on codes
that can easily be printed on product wrappings. Our intention is not to define
a trading system where ACs can be transfered from one person to another. Such
contextual requirements are defined in RFC 3506 as Voucher Trading Systems
(VTS). Terada et al. [28] come with a copy protection mechanism for a VTS and
use public key cryptography. This makes the vouchers only suitable for electronic
circulation. Furthermore, RFC 3506 and [17] do not discuss methods on how to
generate these vouchers securely. We propose the BEPAC scheme in order to fill
this gap.
2
AC Scheme Selection
This section first discusses some examples of AC systems. Then, two different
approaches to set up an AC scheme are discussed and their main drawbacks
are visited. After this, the Generalized Feistel Cipher of [3] is introduced in
Section 2.1 which has some useful concepts that we use in our scheme. The
focus of an AC scheme design lays on scalability, cost-efficiency and off-line use.
Finally, forgery of ACs should be hard, an adversary is only able to forge ACs
with a very small predefined probability.
Examples of Activation Codes First, we discuss some examples of ACs in real
life. A good first example is the scratch prepaid card that is used in the telecommunication industry. To use a prepaid card, the customer needs to remove some
foil and reveals a code that can be used to obtain mobile phone credits. Then we
have the e-coupon which is a widely used replacement for the conventional paper
coupon. An e-coupon represents value and is used to give financial discount or
rebate at the checkout of a web shop. The last example is a one-time password
that gives access to on-line content. This content should only be accessible to
authenticated people who possess this unique password; think of sneak previews
of new material or software distribution.
All the aforementioned examples use unique codes that should be easy to
handle, that is, people should be able to manually copy ACs without much
effort. At the same time, it should be impossible for an adversary to use an
AC more than once or autonomously generate a new valid AC. Altogether, ACs
are unique codes that have to guarantee authenticity. An AC system provides
authenticity when an adversary is not able to forge ACs. It is a misconception
that authenticity only is enough for an AC scheme. In the end, most AC systems
are used in a competitive environment. When vendor A starts a campaign where
ACs are used to promote a product and provide it for free to customers, then
it is not desirable that vendor B finds out details about this campaign like the
number of released ACs. Other sensitive details might be the value that different
ACs represent or the expiry date of ACs. It is for this reason that we need
confidentiality, which means that an adversary is not able to recognize patterns
or extract any information from a released AC.
Many systems use codes that need to provide the properties discussed above.
In this paper we refer to all these codes as ACs. By an AC scheme S we indicate
a tuple (A, N , P, λ) where A is the size of the used alphabet, N is the number
of desired ACs, P determines the probability P = P1 of an adversary guessing
an AC and λ is the length of the ACs.
Database Approach The database approach is very straightforward and consists of a database that contains all the released ACs and their current status.
The generation of new AC entries is done by a pseudo-random function. When
a customer redeems an AC, its status is set to ‘used’. An advantage of the
database approach is that the randomness of the ACs is directly related to the
randomness of the pseudo-random generator. So, it is important to select a good
pseudo-random generator, e.g. a FIPS certified one. On the other hand, the
protection of this valuable data is still a problem. For instance, if an attacker
manages to add entries to the database or is able to change the record status
to ‘unused’, it will be hard to detect this fraud in time. Also, it is necessary to
check any new AC against all existing entries since there might be a collision.
As a consequence, access to the complete set of ACs is needed on generation of
new ACs.
Block Cipher Approach Another approach is to use a block cipher that gives
a random permutation F : {0, 1}n → {0, 1}n from serial numbers to ACs. The
provider maintains a counter i to keep track of the number of generated ACs.
This way the authenticity of a serial number can be checked since only ACs
that decrypt to a serial < i are valid. A disadvantage of this method is the size
of the resulting AC which is 128-bits for AES and 64-bits for 3DES. For AES,
this results in a string of about 21 characters when we use numbers, upperand lowercase characters in our alphabet. For 3DES, this is about 11 characters.
Smaller block ciphers do exist, like Katan [10] which is 32-bits, but are not
very well-studied. Furthermore, block ciphers force ACs to have a length that
is a multiple of the block size b. An alternative could be the concept of elastic
block ciphers [9] which is an extended scheme where variable message sizes are
allowed as input. Moreover, this scheme uses well-studied block ciphers. Still,
the minimal size of a plaintext message is the block size b of the incorporated
block cipher. So, this does not give any advantage and is still too large for our
target, which is roughly 20 to 50-bit codes.
2.1
Small Domain Ciphers
Black and Rogaway introduce the Generalized Feistel Cipher (GFC) in [3]. The
GFC is designed to allow the construction of arbitrary domain ciphers. Here,
arbitrary domain means a domain space that is not necessarily {0, 1}n. For ACs
we want to use a small domain cipher where the domain size can be customized
to a certain extent, therefore we look into the proposed method in [3]. Before
we describe the Generalized Feistel Cipher, we briefly visit the basic Feistel
construction.
Feistel Network A Feistel network [16] is a permutation that
takes an input x of size 2n, then performs a number of rounds r
with round functions f1 , ..., fr , and finally delivers an output y of
size 2n. The input is split into two blocks hL, Ri ∈ {0, 1}2×n. As
shown in Figure 2, every right block is input to a round function
fi . The output of this function is combined with the left block
and becomes the new right block, e.g. L′ = f1 (R) + L for GFC.
The original right block becomes the new left block. For the ease
of decryption the last output blocks are swapped in case of an
odd number of rounds (which is the case in Fig. 2).
Fig. 2: Feistel
Generalized Feistel Cipher The GFC of Black and Rogaway [3] was introduced to handle flexible domain sizes. Take for example an encryption E : 514 →
514 which is not a domain that is captured by standard block cipher algorithms.
The BEPAC scheme borrows some of the ideas of GFC to be able to construct
arbitrarily sized AC configurations.
In GFC the left and right block of the Feistel network are “similarly sized”
which means that their domain size may deviate a little. For the particular
case of ACs we have looser restrictions on the arbitrariness of our domain and
we can increase the guessing probability P to influence the domain size. As a
consequence, the system parameters of BEPAC can be chosen such that the left
and right block are equally sized.
An obvious way to use the Feistel network is to create a pseudo-random
permutation F : K × M → M where K and M are the key space and message
space respectively. To generate ACs, we take as plaintext an index i and use the
resulting ciphertext as AC α. In order to check α the provider keeps track of the
last index i and considers a given α valid when F −1 (α) ≤ i. This construction
guarantees:
1. Collision-freeness, since F is a permutation.
2. Valid serial numbers, they cannot be predicted since F is a pseudo-random
permutation.
As Black and Rogaway already conclude in [3] the Generalized Feistel Cipher
has weak security bounds when used in applications where the message space is
roughly from k = 230 up to k = 260 . This suggests that our second argument
might not be that strong.
Also, the serial i is kept secret and one might argue that this presumes unforgeability. However, the way i is embedded allows an adversary to make useful
assumptions about i since the ACs are generated using consecutive numbers. In
the GFC, the left block L and right block R are initiated as follows:
L = i mod 2n ,
R = ⌊i/2n ⌋
Here, L represents the least significant bits of i, and the successor of every i
always causes a change in L. On the contrary, when i is sequentially incremented,
the value of R changes only once every 2n times. This way, the first 2n ACs are
generated with R = 0, the second 2n ACs with R = 1, etc. The problems that
this little example already points out are further explained in the next section.
3
Feistel Permutation Recovery using CPA-2
In this section we present a practical attack on a three-round Feistel construction
in order to illustrate the problem of choosing a small number of rounds and using
a serial embedding as suggested by Black and Rogaway [3].
Theorem 1. Consider a three-round 2n-bit Feistel construction. Then there exists an algorithm that needs at most 2n+2 adaptive chosen plain-/ciphertext pairs
to compute any ciphertext from any plaintext and vice versa without knowledge
of the secret round keys and regardless the used key length.
Proof. The two ciphertext blocks are defined in terms of the plaintext blocks as
follows:
R′ = f2 (f1 (R) + L) + R
L′′ = f3 (f2 (f1 (R) + L) + R) + f1 (R) + L
= f3 (R′ ) + f1 (R) + L
(1)
Note that L′′ uses R′ as input to f3 (·). With Li we denote L = i and similarly
′
Rj denotes R = j. The notation R(i,j)
means the value of R′ when Li and
Rj are used as input blocks. We first observe that several triples (f1 , f2 , f3 )
lead to the same permutation and show that it is always possible to find the
triple with f1 (0) = 0. To this end, if we replace the triple (f1 , f2 , f3 ) with the
triple (f1′ , f2′ , f3′ ) defined by Equation (2), this leads to the same permutation
(Equation (1)) with the desired property that f1′ (0) = 0.
f1′ (x) = f1 (x) − f1 (0),
f2′ (x) = f2 (x + f1 (0)),
f3′ (x) = f3 (x) + f1 (0) (2)
So, without loss of generality we may assume that f1 (0) = 0.
Next, we describe a method to find a triple (f1 , f2 , f3 ) with f1 (0) = 0. First,
we determine f2 , then f1 and finally f3 .
By Equation (1) we get f2 :
′
′
f2 (f1 (R0 ) + Li ) = R(i,0)
− R0 = R(i,0)
⇒
′
f2 (Li ) = R(i,0)
(3)
Now, to find f1 observe that f1 (j) is a solution for x in the equation f2 (x) =
′
R(0,j)
− Rj . However, this equation does not always have one unique solution
since f2 is a pseudo-random function. In case of multiple solutions we compare
the output of successive (wrt. x) function inputs with the values for f2 (x + i)
that were found using Equation (3). Then, the correct x is the unique solution
to:
′
f2 (x + i) = R(i,j)
− Rj
for i = 0, . . . , m
(4)
Sometimes m = 0 already gives a unique solution. At the end of this section we
show that with very high probability m = 1 defines a unique solution. We find:
f1 (j) = x
(5)
When f1 and f2 are both determined, L and R can be chosen such that every
value for R′ ∈ {0, . . . , 2n − 1} is visited. Since R′ functions as direct input to f3
it is possible to find all input-output pairs for f3 . To visit every possible input
′
value z = 0, . . . , 2n − 1 find a pair Li , Rj such that R(i,j)
= z. First, find an
index x such that f2 (x) = z − Rj . If such an x does not exist choose a different
value for Rj . There is always a solution for x since Rj covers the whole domain
of f2 . Second, derive Li :
f2 (f1 (Rj ) + Li ) = f2 (x)
⇒
Li = x − f1 (Rj )
(6)
Note that the determination of Li and Rj does not need any intermediate queries
since it is completely determined by f1 and f2 . Next, we query the system with
Li and Rj and use Equation (3) and (5) to compute f3 as follows:
f3 (x) = L′′(i,j) − f1 (Rj ) − Li
(7)
This completes the solution for a triple (f1 , f2 , f3 ) that results in the same permutation as the Feistel construction under attack.
Number of Queries The determination of f2 is given by Equation (1) and costs
2n queries. The determination of f1 is given by Equation (4). The probability p
that there exists an x′ 6= x for a preselected x such that
^
f2 (x + i) = f2 (x′ + i)
(f2 (x) = f2 (x′ )) ∧
(8)
i=1,...,m
can be split into two parts. First, we have the probability p1 that there is a
collision f2 (x) = f2 (x′ ) with x 6= x′ . Then, the second probability p2 covers
cases where a preselected position f2 (x + i) has the same value as some other
preselected position f2 (x′ + i). We take k = 2n and the two probabilities are
(k−1)
then given by p1 = 1 − ( k−1
and p2 = k1 . Now, p = p1 · p2 m because
k )
we need to multiply by p2 for every other successive match. To conclude, the
probability that there is an x′ 6= x for a Feistel construction of size 2 · k and
with m successive queries, i.e. the probability that there is no unique solution x
to Equation (4), is the probability
1
1
p= m − m ·
k
k
k−1
k
k−1
(9)
So, p < k1m and depending on the size k, m = 1 already gives p close to zero.
In practice one might sometimes need an additional query (m = 2) or only one
query (m = 0), but on average m = 1. This means that the cost for determination
of f1 is 2 · 2n queries on average. Then, the determination of f3 is given by
Equation (7) and costs at most 2n queries. As a result, the determination of
(f1 , f2 , f3 ) has an upper bound of 2n+2 queries.
4
BEPAC Scheme
In this section we propose our Activation Code Scheme called BEPAC. Its primary objective is to ensure authenticity and its secondary objective is to provide
confidentiality. Confidentiality is satisfied up to the security bounds given by
Black and Rogaway in [3]. In the BEPAC scheme, loss of confidentiality does
not affect the authenticity property.
The authenticity is achieved in an obvious way by the use of an HMAC which
is a keyed hash function. We take the truncated HMAC h of a sequence number
i and concatenate it to i itself. For this concatenation we use an embedding
m like the one used by Black and Rogaway in [3] and Spies in [2]. We rely on
the strength of the underlying hash function which covers the best practice part
of our solution: ACs are not forgeable. The length of an HMAC is usually too
long for the ease of use that is demanded for ACs. Therefore we introduce the
probability P = P1 that puts a lower bound on the success rate of guessing correct
ACs. We use this parameter to limit the length of the codes, i.e. P determines
the size of the HMAC. A lower success probability for an adversary is achieved
by concatenating a bigger part of the HMAC and thus results in a longer AC.
Our solution differs from encryption schemes for small domains [3,18] in the
sense that we make a clear separation between the part that provides authenticity and the part that provides confidentiality. The latter is added as an additional operation on the embedding m. We use a balanced Feistel construction
as proposed in [3] to create the necessary confusion and diffusion. This separation between authenticity and confidentiality is really different from an approach
where the sequence number i is directly fed into a Feistel construction and when
it solely depends on this construction for its authenticity. The attack in Section 3
demonstrates that we cannot rely on a Feistel construction for authenticity when
it is used on a small domain. These results form the basis of our design decision.
AC Scheme Setup
The BEPAC scheme setup is a construction
Serial i
(see Fig. 3) where an embedding m of an index i and a part of HMAC(i) are fed into a
s = HMAC(i) mod P
Feistel network. Since this is a balanced Feiswhen length is even:
h=s
tel network, m needs to be divided into two
when length is odd:
h= s/A
equally sized blocks. When this is not possible
h' = s mod A
a small part h′ of HMAC(i) bypasses the Feish
i
tel network and is embedded together with the
cryptogram c from the Feistel network to form
Feistel Network
AC α.
c
The BEPAC scheme S is a tuple
(A, N , P, λ, ω) where A is the size of the alActivation Code α
phabet, λ + ω is the length of the ACs where
λ is always even and ω is either 0 or 1. Then,
N is the number of ACs and P determines the
Fig. 3: BEPAC Scheme
probability P = P1 of obtaining a valid AC by
a random guess, e.g. P = 10.000. We assume A < P.
when length is odd
4.1
h'
h'
Definition 1 (Valid AC Scheme). An AC scheme S = (A, N , P, λ, ω) is
valid when Aλ > N × P × A−ω holds and λ is even.
A valid AC scheme S can be obtained as follows:
(a) The user chooses the alphabet size A, desired number of ACs N and some
minimal guess probability P1 .
(b) Nowthe minimal length
λ is calculated such that Aλ >= N × P by taking
A
λ = log(N × P)
(c) |Aλ − N × P| is minimized by taking P = ⌊Aλ /N ⌋
(d) The length λ can be either odd or even:
– When λ = 2k + 1 and A < P then we adjust P such that A is a divisor
of P. As a consequence, we might have a larger number of ACs N .
P = P − (P mod A),
N = ⌊Aλ /P⌋
After these operations we obtain the system S = (A, N , P, λ − 1, 1).
– When λ = 2k we obtain the system S = (A, N , P, λ, 0).
(e) The process is repeated from step (a) when no valid system S is found.
4.2
Generation
This section describes the generation of new ACs once a valid AC scheme is configured. Algorithm 1 contains the pseudo code for AC generation. The plaintext
is an embedding m of a part of HMAC(i) and i itself. In case of an odd AC
length (ω = 1) a small part h′ of HMAC(i) is excluded from this embedding.
The part of HMAC(i) that is used in m is determined by P.
s = HMAC(i) mod P,
h = ⌊s×A−ω ⌋,
h′ = s mod A,
m = h×N +i
The balanced Feistel construction is defined with:
k = A(λ/2) ,
L = m mod k,
R = ⌊m / k⌋
L and R are input blocks with size k of a balanced Feistel network. We denote
the output blocks after r rounds by L⋆ and R⋆ . When the number of rounds r
is even the cryptogram c is given by:
c = R ⋆ × k + L⋆
When r is odd, the left and right block are swapped and the cryptogram c is
given by:
c = L⋆ × k + R ⋆
This difference between odd and even is there to allow the same construction for
encoding and decoding. Finally, the activation code α is given by:
α = c × Aω + ωh′
4.3
Verification
This section describes the verification of previously generated ACs for a valid
AC scheme. Algorithm 2 contains the pseudo code for AC verification. Given an
AC α and an AC scheme S the validity can be checked as follows. First compute
c and h′ from α:
c = ⌊α/Aω ⌋ ,
h′ = α mod Aω
The balanced Feistel construction is defined with input block size k = A(λ/2) .
Now, the input blocks L and R are obtained from c as follows:
L = c mod k,
R = ⌊c/k⌋
L and R are input blocks with size k of a balanced Feistel network. We denote
the output blocks after r rounds by L⋆ and R⋆ . In this case we want to decrypt
and therefore use the round keys in reverse order. When the number of rounds
r is even the plaintext m is given by:
m = R ⋆ × k + L⋆
When r is odd, the left and right block are swapped and the plaintext m is given
by:
m = L⋆ × k + R ⋆
Now, we are able to obtain the partial HMAC h and index i from m by:
h = ⌊m/N ⌋ ,
i = m mod N
We calculate the partial HMAC ht and h′t like in the encoding, but now we
use the recovered index i. Finally, we say that α is a valid AC iff ht = h and
ωh′t = ωh′ .
Algorithm 1 Generate(i)
(λ/2)
k←A
s ← HMAC(i) mod P
h ← ⌊s × A−ω ⌋
h′ ← s mod A
m←h×N +i
L ← m mod k; R ← ⌊m/k⌋
for j ← 1 to r do
tmp ← (L + fj (R)) mod k
L ← R; R ← tmp
end for
if r is odd then
c ← L×k+R
else
c←R×k+L
end if
α ← c × Aω + ωh′
5
Algorithm 2 Verify(α)
k ← A(λ/2)
c ← ⌊α/Aω ⌋; h′ ← α mod Aω
L ← c mod k; R = ⌊c/k⌋
for j ← r to 1 do
tmp ← (L + fj (R)) mod k
L ← R; R ← tmp
end for
if r is odd then
m ← L×k+R
else
m←R×k+L
end if
h ← ⌊m/N ⌋; i ← m mod N
s ← HMAC(i) mod P
ht ← ⌊s × A−ω ⌋
h′t ← s mod A
if ht = h and ωh′t = ωh′ then
Valid
else
Invalid
end if
Example Application: Smart Card
In this section we want to give an example of an Activation Code System (ACS).
In an ACS there are a few things that need to be managed. The index i of
the latest generated AC and the ACs that have been used so far. Since this
information is highly valuable and represents financial value it must be well
protected. Think of an application where the ACs are printed on prepaid cards
covered by some scratch-off material. The production of these cards is a very
secured and well-defined process to ensure that activation codes are kept secret
during manufacturing. These cards need to have all kinds of physical properties,
e.g. the AC should not be readable when the card is partly peeled off from the
back. This can be achieved by printing a random pattern on top of the scratch-off
foil.
At some point there is a very critical task to be executed when the ACs
need to be delivered to the manufacturer. An obvious method to do this is to
encrypt the list of ACs with a secret key. Later on in the process, this list of
randomly generated codes needs to be maintained by the vendor who sells the
scratch cards. This induces a big security threat since leakage of this list or
Desired
A=8
N
P
N P(×103 )
101 103
101
26,214
102 103
102
20,968
103 103
103
16,777
4
3
10 10
10.004
13,416
105 103
100.003
10,737
106 103
1.000.006
68,719
107 103
10.001.379
54,968
8
3
10 10
100.001.057
43,98
109 103 1.000.010.575
35,184
A = 20
A = 31
l Bits N P(×103 ) l Bits
N P(×103 )
6 18 101
16 4 18
101
92,352
7 21 102
32 5 22
102 286,285
8 24 103
64 6 26
103
28,613
9 27 104
128 7 31
104
88,75
10 30 105
12,8 7 31
105 275,125
12 36 106
25,6 8 35
1.000.567
27,497
13 39 107
51,2 9 39
10.000.012
85,289
8
14 42 10
102,4 10 44
100.010.675 264,368
15 45 109
10,24 10 44 1.001.045.818
26,412
l Bits
4 20
5 25
5 25
6 30
7 35
7 35
8 40
9 45
9 45
Table 1: BEPAC Configurations
unauthorized modification results in financial loss. Especially when it directly
relates to the core business like in the telecommunications industry.
The use of a secure application module (SAM) significantly reduces this risk.
A SAM is typically a tamper-resistant device, often a smart card, which is in
most cases extensively tested and certified in accordance to a standard, e.g.
the Common Criteria3 . The elegance of the solution presented in this paper
is that it can be implemented using smart cards. The supplier determines the
probability P , number of codes N , size of character set A and the key K to
be used. An obvious approach is to use two smart cards since the production
and clearance of activation codes are very likely to happen at two different
locations. Both smart cards are initialized using the same AC scheme S and the
same key K. From that moment on the first one only gives out up to N new
activation codes. The second one is used at the clearance house to verify and
keep track of traded activation codes. This can be done by a sequence of bits
where the i-th bit determines whether the i-th activation code has been cleared.
For 1.000.000 activation codes approximately 122 kB of storage is needed. This
fits on a SmartMX card [27] which is available with 144 kB of EEPROM. Of
course, multiple cards can be used if more ACs are needed.
6
Analysis
In this section we discuss the system parameters of BEPAC and decide on some
minimal bounds and algorithms. We tested a 6-round BEPAC scheme for obvious
flaws using the NIST random number test [26]. This test implementation also
delivered the numbers in Table 1 which give a good indication of the length l of
the codes compared to different AC scheme configurations. In the left column the
desired values are given for the number of codes N and the guessing probability
P. We tested these different numbers for three different alphabet sizes A.
Number of Rounds We found good arguments to set the minimum number of
rounds to six for the BEPAC scheme. The literature shows that Feistel constructions of six or more rounds are secure against adaptive chosen plaintext
and chosen ciphertext attacks (CPCA-2) when the number of queries m ≪ 2n ,
3
http://www.commoncriteriaportal.org
see [22,13]. Patarin [21] shows that an adversary needs at least 23n/4 encryptions
to distinguish a six-round Feistel construction from a random permutation. A
six round Feistel network sufficiently covers the risk of leaking serial number
information, but this is of course a minimum.
Key Derivation In the BEPAC scheme we need different round keys for every
Feistel round and another different key for the calculation of the HMAC on
the serial. We propose to derive these keys from an initial randomly chosen
key [1] by a key derivation function (KDF). There are several definitions available
for KDFs and we propose to use KDF1 which is defined in ISO 18033-2 [11].
Recommendations for KDFs and their construction can also be found in [7]. The
first key that is derived is used as the key in the HMAC calculation of h and h′
(Section 4). After this the round keys for the Feistel construction are successively
derived.
Pseudo-random Functions Furthermore, we need to decide on the pseudo-random
functions (PRFs) that are used as round functions of the Feistel network. The
pseudo-randomness of the permutation defined by a Feistel network depends
on the chosen PRF in each round [15]. It is straightforward to use a cryptographic hash function since we already need a hash function for the HMAC [24]
calculation and it keeps our AC scheme simple.
Hash Function In the end, the BEPAC scheme is solely based on a single cryptographic hash function. We follow the secure hash standard FIPS 180-3 [23] and
propose to use an approved hash function like SHA-256.
7
Conclusions
In this paper we have introduced activation codes (ACs), short codes of fixed
length, that represent value. These ACs should be scalable, cost efficient and
forgery resistant. In the literature, several solutions [4,5,8,14,12,6,17,28,17] handle digital coupons or tickets that are somehow reminiscent to our notion of
ACs. The difference is that most solutions use public key cryptography or other
means that result in lengthy codes. In fact, these solutions come closer to some
extended notion of digital cash and are not meant to give a solution on the generation of ACs. To the best of our knowledge there is no scheme that focuses
on the class of ACs that are described in this paper (roughly think of 20 to
50-bit codes). Our proposed AC scheme for this class satisfies authenticity and
confidentiality in a way that when confidentiality is compromised it does not
automatically break authenticity and vice versa.
In order to allow a relatively small and arbitrary message space for our AC
scheme we use some of the ideas of Black and Rogaway [3] in their Generalized Feistel Cipher to satisfy the confidentiality in our scheme. Several studies [22,20,19,15,13,21] show that the security bounds of Feistel constructions are
not strong enough and thus make the use of Feistel constructions in small domains questionable. To illustrate this, we have demonstrated that CPA-2 allows
an adversary to recover the complete permutation from only 2n+2 plaintextciphertext pairs. Still, the Feistel construction is suitable for the purpose of
confidentiality in our AC scheme. Since confidentiality is a secondary goal, it
relaxes the demands on the security bounds. Furthermore, in BEPAC the plaintext cannot be predicted which counters the attacks from the literature. And
most important of all, a Feistel construction defines a permutation on the AC
domain, which means in practice that we do not have to store any additional
data in order to remember which ACs are already published and which are not.
To conclude, we found satisfactory system parameters for the minimum number of Feistel rounds and we defined a method to do the key derivation for the
round keys. Furthermore, we suggested a specific pseudo-random function (PRF)
and hash function for concrete implementations. Finally, we have implemented
the BEPAC scheme4 and performed some statistical tests using the NIST Random Number Test [26]. This test did not reveal any obvious flaws.
It might be interesting for future work to create a smart card implementation
of BEPAC as suggested in Section 5.
References
1. E. Barker and J. Kelsey. Recommendation for Random Number Generation Using
Deterministic Random Bit Generators (revised). NIST Special publication, 800:90,
2007.
2. M. Bellare, P. Rogaway, and T. Spies. Format-Preserving Feistel-Based Encryption
Mode. http://csrc.nist.gov/groups/ST/toolkit/BCM/modes_development.html,
November 2009.
3. J. Black and P. Rogaway. Ciphers with Arbitrary Finite Domains. Topics in
Cryptology–CT-RSA 2002, pages 185–203, 2002.
4. C. Blundo, S. Cimato, and A. De Bonis. A Lightweight Protocol for the Generation
and Distribution of Secure E-Coupons. In WWW ’02: Proceedings of the 11th
international conference on World Wide Web, pages 542–552, New York, NY, USA,
2002. ACM.
5. C. Blundo, S. Cimato, and A. De Bonis. Secure E-Coupons. Electronic Commerce
Research, 5(1):117–139, 2005.
6. C.C. Chang, C.C. Wu, and I.C. Lin. A Secure E-coupon System for Mobile Users.
International Journal of Computer Science and Network Security, 6(1):273, 2006.
7. L. Chen. Recommendation for Key Derivation Using Pseudorandom Functions.
NIST Special Publication, 800:108, 2009.
8. S. Cimato and A. De Bonis. Online Advertising: Secure E-Coupons. Theoretical
Computer Science, pages 370–383.
9. D. Cook, A. Keromytis, and M. Yung. Elastic Block Ciphers: The Basic Design.
In Proceedings of the 2nd ACM symposium on Information, computer and communications security, pages 350–352. ACM, 2007.
10. C. De Cannière, O. Dunkelman, and M. Knežević. KATAN and KTANTAN – A
Family of Small and Efficient Hardware-Oriented Block Ciphers. Cryptographic
Hardware and Embedded Systems-CHES 2009, pages 272–288, 2009.
4
http://www.cs.ru.nl/~ gkoningg/bepac
11. ISO/IEC 18033-2:2006. Information Technology - Security Techniques - Encryption
Algorithms - Part 2: Asymmetric Ciphers, 2006.
12. M. Jakobsson, P.D. MacKenzie, and J.P. Stern. Secure and Lightweight Advertising on the Web. Computer Networks-the International Journal of Computer and
Telecommunications Networkin, 31(11):1101–1110, 1999.
13. L.R. Knudsen. The Security of Feistel Ciphers with Six Rounds or Less. Journal
of Cryptology, 15(3):207–222, 2008.
14. M. Kumar, A. Rangachari, A. Jhingran, and R. Mohan. Sales Promotions on the
Internet. 3rd USENIX Workshop on Electronic Commerce, pages 167 – 176, 1998.
15. M. Luby. Pseudorandomness and Cryptographic Applications. Princeton University
Press, 1996.
16. Michael Luby and Charles Rackoff. How to Construct Pseudorandom Permutations
from Pseudorandom Functions. SIAM J. Comput., 17(2):373–386, 1988.
17. K. Matsuyama and K. Fujimura. Distributed Digital-Ticket Management for
Rights Trading System. In Proceedings of the 1st ACM conference on Electronic
commerce, page 118. ACM, 1999.
18. B. Morris, P. Rogaway, and T. Stegers. How to Encipher Messages on a Small
Domain. In Proceedings of the 29th Annual International Cryptology Conference
on Advances in Cryptology, page 302. Springer, 2009.
19. M. Naor and O. Reingold. On the Construction of Pseudorandom Permutations:
Luby-Rackoff Revisited. Journal of Cryptology, 12(1):29–66, 1999.
20. J. Patarin. New Results on Pseudorandom Permutation Generators Based on the
DES Scheme. In Advances in Cryptology-CRYPTO, volume 91, pages 301–312.
21. J. Patarin. About Feistel Schemes with Six (or More) Rounds. Lecture Notes in
Computer Science, 1372:103–121, 1998.
22. J. Patarin. Security of Random Feistel Schemes with 5 or More Rounds. Lecture
Notes in Computer Science, pages 106–122, 2004.
23. FIB PUB. FIPS 180-3: Secure HASH Standard (SHS). Federal Information Processing Standards Publication, 2008.
24. N.F. PUB. FIPS 198: Keyed-Hash Message Authentication Code (HMAC). Federal
Information Processing Standards Publication, 198.
25. J.J. Quisquater and F.X. Standaert. Exhaustive Key Search of the DES: Updates
and Refinements. Special-Purpose Hardware for Attacking Cryptographic Systems,
2005.
26. A. Rukhin, J. Soto, J. Nechvatal, M. Smid, E. Barker, S. Leigh, M. Levenson,
M. Vangel, D. Banks, A. Heckert, et al. A Statistical Test Suite for Random and
Pseudorandom Number Generators for Cryptographic Applications. NIST, 2000.
27. NXP Semiconductors. P5Cx012/02x/40/73/80/144 Family: Secure Dual Interface
and Contact PKI Smart Card Controller. NXP, June 2010.
28. M. Terada, M. Hanadate, and K. Fujimura. Copy Prevention Scheme for Rights
Trading Infrastructure. In Smart Card Research and Advanced Applications: IFIP
TC8/WG8, page 51. Springer, 2000.
| 5cs.CE
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.